Release: v2.2.0

Co-authored-by: wassname

.circleci/config.yml

@@ -9,7 +9,7 @@ jobs:
         steps:
             - checkout
             - run: sudo pip install torch
-            - run: sudo pip install tensorflow==2.0.0-rc0
+            - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
             - run: sudo pip install pytest codecov pytest-cov
             - run: sudo pip install tensorboardX scikit-learn
@@ -38,7 +38,7 @@ jobs:
         parallelism: 1
         steps:
             - checkout
-            - run: sudo pip install tensorflow==2.0.0-rc0
+            - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
             - run: sudo pip install pytest codecov pytest-cov
             - run: sudo pip install tensorboardX scikit-learn
@@ -65,7 +65,7 @@ jobs:
             - image: circleci/python:2.7
         steps:
             - checkout
-            - run: sudo pip install tensorflow==2.0.0-rc0
+            - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
             - run: sudo pip install pytest codecov pytest-cov
             - run: python -m pytest -sv ./transformers/tests/ --cov
@@ -81,7 +81,7 @@ jobs:
             - checkout
             - run: sudo pip install --progress-bar off -r docs/requirements.txt
             - run: sudo pip install --progress-bar off -r requirements.txt
-            - run: cd docs && make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+            - run: ./.circleci/deploy.sh
 workflow_filters: &workflow_filters
     filters:
         branches:
@@ -96,4 +96,4 @@ workflows:
             - build_py3_tf
             - build_py2_torch
             - build_py2_tf
-            - deploy_doc: *workflow_filters
+            - deploy_doc: *workflow_filters

.circleci/deploy.sh

@@ -0,0 +1,26 @@
+cd docs
+
+function deploy_doc(){
+	echo "Creating doc at commit $1 and pushing to folder $2"
+	git checkout $1
+	if [ ! -z "$2" ] 
+	then
+		if [ -d "$dir/$2" ]; then
+			echo "Directory" $2 "already exists"
+		else
+			echo "Pushing version" $2
+			make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
+		fi
+	else
+		echo "Pushing master"
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+	fi
+}
+
+deploy_doc "master" 
+deploy_doc "b33a385" v1.0.0
+deploy_doc "fe02e45" v1.1.0
+deploy_doc "89fd345" v1.2.0
+deploy_doc "fc9faa8" v2.0.0
+deploy_doc "3ddce1d" v2.1.1
+deploy_doc "f2f3294" v2.2.0

.github/ISSUE_TEMPLATE/---new-benchmark.md

@@ -0,0 +1,22 @@
+---
+name: "\U0001F5A5 New Benchmark"
+about: You benchmark a part of this library and would like to share your results
+title: "[Benchmark]"
+labels: ''
+assignees: ''
+
+---
+
+# Benchmarking Transformers
+
+## Benchmark
+
+Which part of Transformers did you benchmark?
+
+## Set-up
+
+What did you run your benchmarks on? Please include details, such as: CPU, GPU? If using multiple GPUs, which parallelization did you use?
+
+## Results
+
+Put your results here!

.github/ISSUE_TEMPLATE/--new-model-addition.md

@@ -17,6 +17,7 @@ assignees: ''
 
 * [ ] the model implementation is available: (give details)
 * [ ] the model weights are available: (give details)
+* [ ] who are the authors: (mention them)
 
 ## Additional context
 

.gitignore

@@ -137,4 +137,5 @@ examples/runs
 serialization_dir
 
 # emacs
-*.*~
+*.*~
+debug.env

CONTRIBUTING.md

@@ -62,6 +62,8 @@ Awesome! Please provide the following information:
 If you are willing to contribute the model yourself, let us know so we can best
 guide you.
 
+We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder.
+
 ### Do you want a new feature (that is not a model)?
 
 A world-class feature request addresses the following points:
@@ -81,6 +83,8 @@ A world-class feature request addresses the following points:
 If your issue is well written we're already 80% of the way there by the time you
 post it.
 
+We have added **templates** to guide you in the process of adding a new example script for training or testing the models in the library. You can find them in the [`templates`](./templates) folder.
+
 ## Start contributing! (Pull Requests)
 
 Before writing code, we strongly advise you to search through the exising PRs or

README.md

@@ -39,7 +39,7 @@ State-of-the-art NLP for everyone
 Lower compute costs, smaller carbon footprint
 - Researchers can share trained models instead of always retraining
 - Practitioners can reduce compute time and production costs
-- 8 architectures with over 30 pretrained models, some in more than 100 languages
+- 10 architectures with over 30 pretrained models, some in more than 100 languages
 
 Choose the right framework for every part of a model's lifetime
 - Train state-of-the-art models in 3 lines of code
@@ -58,7 +58,7 @@ Choose the right framework for every part of a model's lifetime
 | [Quick tour: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
 | [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-transformers to transformers |
 | [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
-| [Documentation](https://huggingface.co/transformers/) | Full API documentation and more |
+| [Documentation][(v2.2.0)](https://huggingface.co/transformers/v2.2.0) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) [(master](https://huggingface.co/transformers) | Full API documentation and more |
 
 ## Installation
 
@@ -86,6 +86,17 @@ When TensorFlow 2.0 and/or PyTorch has been installed, you can install from sour
 pip install [--editable] .
 ```
 
+### Run the examples
+
+Examples are included in the repository but are not shipped with the library.
+Therefore, in order to run the latest versions of the examples you also need to install from source. To do so, create a new virtual environment and follow these steps:
+
+```bash
+git clone git@github.com:huggingface/transformers
+cd transformers
+pip install [--editable] .
+```
+
 ### Tests
 
 A series of tests are included for the library and the example scripts. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples).
@@ -111,7 +122,7 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 
 ## Model architectures
 
-🤗 Transformers currently provides 8 NLU/NLG architectures:
+🤗 Transformers currently provides 10 NLU/NLG architectures:
 
 1. **[BERT](https://github.com/google-research/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
 2. **[GPT](https://github.com/openai/finetune-transformer-lm)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
@@ -122,6 +133,8 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 7. **[RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/roberta)** (from Facebook), released together with the paper a [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
 8. **[DistilBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation).
 9. **[CTRL](https://github.com/salesforce/ctrl/)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
+10. **[CamemBERT](https://camembert-model.fr)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
+11. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
 
 These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Peason R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
 
@@ -170,16 +183,16 @@ for model_class, tokenizer_class, pretrained_weights in MODELS:
 
 # Each architecture is provided with several class for fine-tuning on down-stream tasks, e.g.
 BERT_MODEL_CLASSES = [BertModel, BertForPreTraining, BertForMaskedLM, BertForNextSentencePrediction,
-                      BertForSequenceClassification, BertForMultipleChoice, BertForTokenClassification,
-                      BertForQuestionAnswering]
+                      BertForSequenceClassification, BertForTokenClassification, BertForQuestionAnswering]
 
 # All the classes for an architecture can be initiated from pretrained weights for this architecture
 # Note that additional weights added for fine-tuning are only initialized
 # and need to be trained on the down-stream task
-tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+pretrained_weights = 'bert-base-uncased'
+tokenizer = BertTokenizer.from_pretrained(pretrained_weights)
 for model_class in BERT_MODEL_CLASSES:
     # Load pretrained model/tokenizer
-    model = model_class.from_pretrained('bert-base-uncased')
+    model = model_class.from_pretrained(pretrained_weights)
 
     # Models can return full list of hidden-states & attentions weights at each layer
     model = model_class.from_pretrained(pretrained_weights,
@@ -242,14 +255,20 @@ sentence_2 = "His findings were not compatible with this research."
 inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
 inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
 
-pred_1 = pytorch_model(**inputs_1)[0].argmax().item()
-pred_2 = pytorch_model(**inputs_2)[0].argmax().item()
+pred_1 = pytorch_model(inputs_1['input_ids'], token_type_ids=inputs_1['token_type_ids'])[0].argmax().item()
+pred_2 = pytorch_model(inputs_2['input_ids'], token_type_ids=inputs_2['token_type_ids'])[0].argmax().item()
+
 print("sentence_1 is", "a paraphrase" if pred_1 else "not a paraphrase", "of sentence_0")
 print("sentence_2 is", "a paraphrase" if pred_2 else "not a paraphrase", "of sentence_0")
 ```
 
 ## Quick tour of the fine-tuning/usage scripts
 
+**Important**  
+Before running the fine-tuning scripts, please read the
+[instructions](#run-the-examples) on how to
+setup your environment to run the examples.
+
 The library comprises several example scripts with SOTA performances for NLU and NLG tasks:
 
 - `run_glue.py`: an example fine-tuning Bert, XLNet and XLM on nine different GLUE tasks (*sequence-level classification*)
@@ -411,7 +430,7 @@ and from the Salesforce CTRL model:
 python ./examples/run_generation.py \
     --model_type=ctrl \
     --length=20 \
-    --model_name_or_path=gpt2 \
+    --model_name_or_path=ctrl \
     --temperature=0 \
     --repetition_penalty=1.2 \
 ```
@@ -518,12 +537,12 @@ Here is a conversion examples from `BertAdam` with a linear warmup and decay sch
 # Parameters:
 lr = 1e-3
 max_grad_norm = 1.0
-num_total_steps = 1000
+num_training_steps = 1000
 num_warmup_steps = 100
-warmup_proportion = float(num_warmup_steps) / float(num_total_steps)  # 0.1
+warmup_proportion = float(num_warmup_steps) / float(num_training_steps)  # 0.1
 
 ### Previously BertAdam optimizer was instantiated like this:
-optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
+optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_training_steps)
 ### and used like this:
 for batch in train_data:
     loss = model(batch)
@@ -532,9 +551,10 @@ for batch in train_data:
 
 ### In Transformers, optimizer and schedules are splitted and instantiated like this:
 optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False)  # To reproduce BertAdam specific behavior set correct_bias=False
-scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_steps)  # PyTorch scheduler
+scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)  # PyTorch scheduler
 ### and used like this:
 for batch in train_data:
+    model.train()
     loss = model(batch)
     loss.backward()
     torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm)  # Gradient clipping is not in AdamW anymore (so you can use amp without issue)
@@ -545,4 +565,13 @@ for batch in train_data:
 
 ## Citation
 
-At the moment, there is no paper associated with Transformers but we are working on preparing one. In the meantime, please include a mention of the library and a link to the present repository if you use this work in a published or open-source project.
+We now have a paper you can cite for the 🤗 Transformers library:
+```
+@article{Wolf2019HuggingFacesTS,
+  title={HuggingFace's Transformers: State-of-the-art Natural Language Processing},
+  author={Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and R'emi Louf and Morgan Funtowicz and Jamie Brew},
+  journal={ArXiv},
+  year={2019},
+  volume={abs/1910.03771}
+}
+```

deploy_multi_version_doc.sh

@@ -0,0 +1,22 @@
+cd docs
+
+function deploy_doc(){
+	echo "Creating doc at commit $1 and pushing to folder $2"
+	git checkout $1
+	if [ ! -z "$2" ] 
+	then
+		echo "Pushing version" $2
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
+	else
+		echo "Pushing master"
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+	fi
+}
+
+deploy_doc "master" 
+deploy_doc "b33a385" v1.0.0
+deploy_doc "fe02e45" v1.1.0
+deploy_doc "89fd345" v1.2.0
+deploy_doc "fc9faa8" v2.0.0
+deploy_doc "3ddce1d" v2.1.1
+deploy_doc "f2f3294" v2.2.0

docs/source/_static/js/custom.js

@@ -1,5 +1,5 @@
 function addIcon() {
-    const huggingFaceLogo = "https://huggingface.co/assets/transformers-docs/huggingface_logo.svg";
+    const huggingFaceLogo = "https://huggingface.co/landing/assets/transformers-docs/huggingface_logo.svg";
     const image = document.createElement("img");
     image.setAttribute("src", huggingFaceLogo);
 
@@ -24,10 +24,10 @@ function addCustomFooter() {
     social.classList.add("footer__Social");
 
     const imageDetails = [
-        { link: "https://huggingface.co", imageLink: "https://huggingface.co/assets/transformers-docs/website.svg" },
-        { link: "https://twitter.com/huggingface", imageLink: "https://huggingface.co/assets/transformers-docs/twitter.svg" },
-        { link: "https://github.com/huggingface", imageLink: "https://huggingface.co/assets/transformers-docs/github.svg" },
-        { link: "https://www.linkedin.com/company/huggingface/", imageLink: "https://huggingface.co/assets/transformers-docs/linkedin.svg" }
+        { link: "https://huggingface.co", imageLink: "https://huggingface.co/landing/assets/transformers-docs/website.svg" },
+        { link: "https://twitter.com/huggingface", imageLink: "https://huggingface.co/landing/assets/transformers-docs/twitter.svg" },
+        { link: "https://github.com/huggingface", imageLink: "https://huggingface.co/landing/assets/transformers-docs/github.svg" },
+        { link: "https://www.linkedin.com/company/huggingface/", imageLink: "https://huggingface.co/landing/assets/transformers-docs/linkedin.svg" }
     ];
 
     imageDetails.forEach(imageLinks => {

docs/source/benchmarks.md

@@ -0,0 +1,54 @@
+# Benchmarks
+
+This section is dedicated to the Benchmarks done by the library, both by maintainers, contributors and users. These 
+benchmark will help keep track of the preformance improvements that are brought to our models across versions.
+
+## Benchmarking all models for inference
+
+As of version 2.1 we have benchmarked all models for inference, across many different settings: using PyTorch, with
+and without TorchScript, using TensorFlow, with and without XLA. All of those tests were done across CPUs (except for
+TensorFlow XLA) and GPUs.
+
+The approach is detailed in the [following blogpost](https://medium.com/huggingface/benchmarking-transformers-pytorch-and-tensorflow-e2917fb891c2)
+
+The results are available [here](https://docs.google.com/spreadsheets/d/1sryqufw2D0XlUH4sq3e9Wnxu5EAQkaohzrJbd5HdQ_w/edit?usp=sharing).
+
+## TF2 with mixed precision, XLA, Distribution (@tlkh)
+
+This work was done by [Timothy Liu](https://github.com/tlkh).
+
+There are very positive results to be gained from the various TensorFlow 2.0 features:
+
+- Automatic Mixed Precision (AMP)
+- XLA compiler
+- Distribution strategies (multi-GPU)
+
+The benefits are listed here (tested on CoLA, MRPC, SST-2):
+
+- AMP: Between 1.4x to 1.6x decrease in overall time without change in batch size
+- AMP+XLA: Up to 2.5x decrease in overall time on SST-2 (larger dataset)
+- Distribution: Between 1.4x to 3.4x decrease in overall time on 4xV100
+- Combined: Up to 5.7x decrease in overall training time, or 9.1x training throughput
+
+The model quality (measured by the validation accuracy) fluctuates slightly. Taking an average of 4 training runs 
+on a single GPU gives the following results:
+
+- CoLA: AMP results in slighter lower acc (0.820 vs 0.824)
+- MRPC: AMP results in lower acc (0.823 vs 0.835)
+- SST-2: AMP results in slighter lower acc (0.918 vs 0.922)
+
+However, in a distributed setting with 4xV100 (4x batch size), AMP can yield in better results:
+
+CoLA: AMP results in higher acc (0.828 vs 0.812)
+MRPC: AMP results in lower acc (0.817 vs 0.827)
+SST-2: AMP results in slightly lower acc (0.926 vs 0.929)
+
+The benchmark script is available [here](https://github.com/NVAITC/benchmarking/blob/master/tf2/bert_dist.py).
+
+Note: on some tasks (e.g. MRPC), the dataset is too small. The overhead due to the model compilation with XLA as well
+as the distribution strategy setup does not speed things up. The XLA compile time is also the reason why although throughput 
+can increase a lot (e.g. 2.7x for single GPU), overall (end-to-end) training speed-up is not as fast (as low as 1.4x)
+
+The benefits as seen on SST-2 (larger dataset) is much clear.
+
+All results can be seen on this [Google Sheet](https://docs.google.com/spreadsheets/d/1538MN224EzjbRL239sqSiUy6YY-rAjHyXhTzz_Zptls/edit#gid=960868445).

docs/source/conf.py

@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'2.1.0'
+release = u'2.2.0'
 
 
 # -- General configuration ---------------------------------------------------

docs/source/index.rst

@@ -63,6 +63,7 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
     bertology
     torchscript
     multilingual
+    benchmarks
 
 .. toctree::
     :maxdepth: 2

docs/source/main_classes/optimizer_schedules.rst

@@ -18,19 +18,17 @@ Schedules
 Learning Rate Schedules
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-.. autoclass:: transformers.ConstantLRSchedule
-    :members:
+.. autofunction:: transformers.get_constant_schedule
 
 
-.. autoclass:: transformers.WarmupConstantSchedule
-    :members:
+.. autofunction:: transformers.get_constant_schedule_with_warmup
 
 .. image:: /imgs/warmup_constant_schedule.png
     :target: /imgs/warmup_constant_schedule.png
     :alt:
 
 
-.. autoclass:: transformers.WarmupCosineSchedule
+.. autofunction:: transformers.get_cosine_schedule_with_warmup
     :members:
 
 .. image:: /imgs/warmup_cosine_schedule.png
@@ -38,8 +36,7 @@ Learning Rate Schedules
     :alt:
 
 
-.. autoclass:: transformers.WarmupCosineWithHardRestartsSchedule
-    :members:
+.. autofunction:: transformers.get_cosine_with_hard_restarts_schedule_with_warmup
 
 .. image:: /imgs/warmup_cosine_hard_restarts_schedule.png
     :target: /imgs/warmup_cosine_hard_restarts_schedule.png
@@ -47,8 +44,7 @@ Learning Rate Schedules
 
 
 
-.. autoclass:: transformers.WarmupLinearSchedule
-    :members:
+.. autofunction:: transformers.get_linear_schedule_with_warmup
 
 .. image:: /imgs/warmup_linear_schedule.png
     :target: /imgs/warmup_linear_schedule.png

docs/source/migration.md

@@ -84,12 +84,12 @@ Here is a conversion examples from `BertAdam` with a linear warmup and decay sch
 # Parameters:
 lr = 1e-3
 max_grad_norm = 1.0
-num_total_steps = 1000
+num_training_steps = 1000
 num_warmup_steps = 100
-warmup_proportion = float(num_warmup_steps) / float(num_total_steps)  # 0.1
+warmup_proportion = float(num_warmup_steps) / float(num_training_steps)  # 0.1
 
 ### Previously BertAdam optimizer was instantiated like this:
-optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
+optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, num_training_steps=num_training_steps)
 ### and used like this:
 for batch in train_data:
     loss = model(batch)
@@ -98,7 +98,7 @@ for batch in train_data:
 
 ### In Transformers, optimizer and schedules are splitted and instantiated like this:
 optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False)  # To reproduce BertAdam specific behavior set correct_bias=False
-scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_steps)  # PyTorch scheduler
+scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)  # PyTorch scheduler
 ### and used like this:
 for batch in train_data:
     loss = model(batch)

docs/source/model_doc/ctrl.rst

@@ -1,6 +1,11 @@
 CTRL
 ----------------------------------------------------
 
+Note: if you fine-tune a CTRL model using the Salesforce code (https://github.com/salesforce/ctrl),
+you'll be able to convert from TF to our HuggingFace/Transformers format using the 
+``convert_tf_to_huggingface_pytorch.py`` script (see `issue #1654 <https://github.com/huggingface/transformers/issues/1654>`_).
+
+
 ``CTRLConfig``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

docs/source/pretrained_models.rst

@@ -53,6 +53,14 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``bert-base-cased-finetuned-mrpc``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | The ``bert-base-cased`` model fine-tuned on MRPC                                                                                    |
 |                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/transformers/examples.html>`__)                           |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-german-dbmdz-cased``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on cased German text by DBMDZ                                                                                               |
+|                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-german-dbmdz-uncased``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on uncased German text by DBMDZ                                                                                             |
+|                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | GPT               | ``openai-gpt``                                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | OpenAI GPT English model                                                                                                            |
@@ -65,6 +73,9 @@ Here is the full list of the currently provided pretrained models together with
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``gpt2-large``                                             | | 36-layer, 1280-hidden, 20-heads, 774M parameters.                                                                                   |
 |                   |                                                            | | OpenAI's Large-sized GPT-2 English model                                                                                            |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``gpt2-xl``                                                | | 48-layer, 1600-hidden, 25-heads, 1558M parameters.                                                                                  |
+|                   |                                                            | | OpenAI's XL-sized GPT-2 English model                                                                                               |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | Transformer-XL    | ``transfo-xl-wt103``                                       | | 18-layer, 1024-hidden, 16-heads, 257M parameters.                                                                                   |
 |                   |                                                            | | English model trained on wikitext-103                                                                                               |
@@ -116,6 +127,14 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``roberta-large-mnli``                                     | | 24-layer, 1024-hidden, 16-heads, 355M parameters                                                                                    |
 |                   |                                                            | | ``roberta-large`` fine-tuned on `MNLI <http://www.nyu.edu/projects/bowman/multinli/>`__.                                            |
 |                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__)                                                   |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``roberta-base-openai-detector``                           | | 12-layer, 768-hidden, 12-heads, 125M parameters                                                                                     |
+|                   |                                                            | | ``roberta-base`` fine-tuned by OpenAI on the outputs of the 1.5B-parameter GPT-2 model.                                             |
+|                   |                                                            | (see `details <https://github.com/openai/gpt-2-output-dataset/tree/master/detector>`__)                                               |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``roberta-large-openai-detector``                          | | 24-layer, 1024-hidden, 16-heads, 355M parameters                                                                                    |
+|                   |                                                            | | ``roberta-large`` fine-tuned by OpenAI on the outputs of the 1.5B-parameter GPT-2 model.                                            |
+|                   |                                                            | (see `details <https://github.com/openai/gpt-2-output-dataset/tree/master/detector>`__)                                               |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | DistilBERT        | ``distilbert-base-uncased``                                | | 6-layer, 768-hidden, 12-heads, 66M parameters                                                                                       |
 |                   |                                                            | | The DistilBERT model distilled from the BERT model `bert-base-uncased` checkpoint                                                   |
@@ -128,9 +147,17 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``distilgpt2``                                             | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
 |                   |                                                            | | The DistilGPT2 model distilled from the GPT2 model `gpt2` checkpoint.                                                               |
 |                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilroberta-base``                                     | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
+|                   |                                                            | | The DistilRoBERTa model distilled from the RoBERTa model `roberta-base` checkpoint.                                                 |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | CTRL              | ``ctrl``                                                   | | 48-layer, 1280-hidden, 16-heads, 1.6B parameters                                                                                    |
 |                   |                                                            | | Salesforce's Large-sized CTRL English model                                                                                         |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| CamemBERT         | ``camembert-base``                                         | | 12-layer, 768-hidden, 12-heads, 110M parameters                                                                                     |
+|                   |                                                            | | CamemBERT using the BERT-base architecture                                                                                          |
+|                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/camembert>`__)                                                 |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 
-.. <https://huggingface.co/transformers/examples.html>`__
+.. <https://huggingface.co/transformers/examples.html>`__

docs/source/quickstart.md

@@ -188,3 +188,35 @@ assert predicted_text == 'Who was Jim Henson? Jim Henson was a man'
 ```
 
 Examples for each model class of each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the [documentation](#documentation).
+
+#### Using the past
+
+GPT-2 as well as some other models (GPT, XLNet, Transfo-XL, CTRL) make use of a `past` or `mems` attribute which can be used to prevent re-computing the key/value pairs when using sequential decoding. It is useful when generating sequences as a big part of the attention mechanism benefits from previous computations.
+
+Here is a fully-working example using the `past` with `GPT2LMHeadModel` and argmax decoding (which should only be used as an example, as argmax decoding introduces a lot of repetition):
+
+```python
+from transformers import GPT2LMHeadModel, GPT2Tokenizer
+import torch
+
+tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+model = GPT2LMHeadModel.from_pretrained('gpt2')
+
+generated = tokenizer.encode("The Manhattan bridge")
+context = torch.tensor([generated])
+past = None
+
+for i in range(100):
+    print(i)
+    output, past = model(context, past=past)
+    token = torch.argmax(output[0, :])
+
+    generated += [token.tolist()]
+    context = token.unsqueeze(0)
+
+sequence = tokenizer.decode(generated)
+
+print(sequence)
+```
+
+The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.

docs/source/serialization.rst

@@ -33,6 +33,8 @@ where
     * ``bert-large-uncased-whole-word-masking``: 24-layer, 1024-hidden, 16-heads, 340M parameters - Trained with Whole Word Masking (mask all of the the tokens corresponding to a word at once)
     * ``bert-large-cased-whole-word-masking``: 24-layer, 1024-hidden, 16-heads, 340M parameters - Trained with Whole Word Masking (mask all of the the tokens corresponding to a word at once)
     * ``bert-large-uncased-whole-word-masking-finetuned-squad``: The ``bert-large-uncased-whole-word-masking`` model finetuned on SQuAD (using the ``run_bert_squad.py`` examples). Results: *exact_match: 86.91579943235573, f1: 93.1532499015869*
+    * ``bert-base-german-dbmdz-cased``: Trained on German data only, 12-layer, 768-hidden, 12-heads, 110M parameters `Performance Evaluation <https://github.com/dbmdz/german-bert>`__
+    * ``bert-base-german-dbmdz-uncased``: Trained on (uncased) German data only, 12-layer, 768-hidden, 12-heads, 110M parameters `Performance Evaluation <https://github.com/dbmdz/german-bert>`__
     * ``openai-gpt``: OpenAI GPT English model, 12-layer, 768-hidden, 12-heads, 110M parameters
     * ``gpt2``: OpenAI GPT-2 English model, 12-layer, 768-hidden, 12-heads, 117M parameters
     * ``gpt2-medium``: OpenAI GPT-2 English model, 24-layer, 1024-hidden, 16-heads, 345M parameters
@@ -104,7 +106,7 @@ This section explain how you can save and re-load a fine-tuned model (BERT, GPT,
 There are three types of files you need to save to be able to reload a fine-tuned model:
 
 
-* the model it-self which should be saved following PyTorch serialization `best practices <https://pytorch.org/docs/stable/notes/serialization.html#best-practices>`__\ ,
+* the model itself which should be saved following PyTorch serialization `best practices <https://pytorch.org/docs/stable/notes/serialization.html#best-practices>`__\ ,
 * the configuration file of the model which is saved as a JSON file, and
 * the vocabulary (and the merges for the BPE-based models GPT and GPT-2).
 

examples/README.md

@@ -3,13 +3,47 @@
 In this section a few examples are put together. All of these examples work for several models, making use of the very
 similar API between the different models.
 
+**Important**  
+To run the latest versions of the examples, you have to install from source. Execute the following steps in a new virtual environment:
+
+```bash
+git clone git@github.com:huggingface/transformers
+cd transformers
+pip install [--editable] .
+```
+
 | Section                    | Description                                                                                                                                                |
 |----------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| [TensorFlow 2.0 models on GLUE](#TensorFlow-2.0-Bert-models-on-GLUE) | Examples running BERT TensorFlow 2.0 model on the GLUE tasks. 
 | [Language Model fine-tuning](#language-model-fine-tuning) | Fine-tuning the library models for language modeling on a text dataset. Causal language modeling for GPT/GPT-2, masked language modeling for BERT/RoBERTa. |
 | [Language Generation](#language-generation) | Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet.                                         |
 | [GLUE](#glue) | Examples running BERT/XLM/XLNet/RoBERTa on the 9 GLUE tasks. Examples feature distributed training as well as half-precision.                              |
-| [SQuAD](#squad) | Using BERT for question answering, examples with distributed training.                                                                                  |
+| [SQuAD](#squad) | Using BERT/RoBERTa/XLNet/XLM for question answering, examples with distributed training.                                                                                  |
 | [Multiple Choice](#multiple-choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks. 
+| [Named Entity Recognition](#named-entity-recognition) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training.                                                                                  |
+| [Abstractive summarization](#abstractive-summarization) | Fine-tuning the library models for abstractive summarization tasks on the CNN/Daily Mail dataset. |
+
+## TensorFlow 2.0 Bert models on GLUE
+
+Based on the script [`run_tf_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/run_tf_glue.py).
+
+Fine-tuning the library TensorFlow 2.0 Bert model for sequence classification on the  MRPC task of the GLUE benchmark: [General Language Understanding Evaluation](https://gluebenchmark.com/).
+
+This script has an option for mixed precision (Automatic Mixed Precision / AMP) to run models on Tensor Cores (NVIDIA Volta/Turing GPUs) and future hardware and an option for XLA, which uses the XLA compiler to reduce model runtime.
+Options are toggled using `USE_XLA` or `USE_AMP` variables in the script.
+These options and the below benchmark are provided by @tlkh.
+
+Quick benchmarks from the script (no other modifications):
+
+| GPU    | Mode | Time (2nd epoch) | Val Acc (3 runs) |
+| --------- | -------- | ----------------------- | ----------------------|
+| Titan V | FP32 | 41s | 0.8438/0.8281/0.8333 |
+| Titan V | AMP | 26s | 0.8281/0.8568/0.8411 |
+| V100    | FP32 | 35s | 0.8646/0.8359/0.8464 |
+| V100    | AMP | 22s | 0.8646/0.8385/0.8411 |
+| 1080 Ti | FP32 | 55s | - | 
+
+Mixed precision (AMP) reduces the training time considerably for the same hardware and hyper-parameters (same batch size was used).
 
 ## Language model fine-tuning
 
@@ -77,7 +111,7 @@ python run_lm_finetuning.py \
 
 Based on the script [`run_generation.py`](https://github.com/huggingface/transformers/blob/master/examples/run_generation.py).
 
-Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet.
+Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL, XLNet, CTRL.
 A similar script is used for our official demo [Write With Transfomer](https://transformer.huggingface.co), where you
 can try out the different models available in the library.
 
@@ -387,6 +421,182 @@ f1 = 93.15
 exact_match = 86.91
 ```
 
-This fine-tuneds model is available as a checkpoint under the reference
+This fine-tuned model is available as a checkpoint under the reference
 `bert-large-uncased-whole-word-masking-finetuned-squad`.
 
+#### Fine-tuning XLNet on SQuAD
+
+This example code fine-tunes XLNet on the SQuAD dataset. See above to download the data for SQuAD .
+
+```bash
+export SQUAD_DIR=/path/to/SQUAD
+
+python /data/home/hlu/transformers/examples/run_squad.py \
+    --model_type xlnet \
+    --model_name_or_path xlnet-large-cased \
+    --do_train \
+    --do_eval \
+    --do_lower_case \
+    --train_file /data/home/hlu/notebooks/NLP/examples/question_answering/train-v1.1.json \
+    --predict_file /data/home/hlu/notebooks/NLP/examples/question_answering/dev-v1.1.json \
+    --learning_rate 3e-5 \
+    --num_train_epochs 2 \
+    --max_seq_length 384 \
+    --doc_stride 128 \
+    --output_dir ./wwm_cased_finetuned_squad/ \
+    --per_gpu_eval_batch_size=4  \
+    --per_gpu_train_batch_size=4   \
+    --save_steps 5000
+```
+
+Training with the previously defined hyper-parameters yields the following results:
+
+```python
+{
+"exact": 85.45884578997162,
+"f1": 92.5974600601065,
+"total": 10570,
+"HasAns_exact": 85.45884578997162,
+"HasAns_f1": 92.59746006010651,
+"HasAns_total": 10570
+}
+```
+
+## Named Entity Recognition
+
+Based on the script [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py).
+This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
+Details and results for the fine-tuning provided by @stefan-it.
+
+### Data (Download and pre-processing steps)
+
+Data can be obtained from the [GermEval 2014](https://sites.google.com/site/germeval2014ner/data) shared task page.
+
+Here are the commands for downloading and pre-processing train, dev and test datasets. The original data format has four (tab-separated) columns, in a pre-processing step only the two relevant columns (token and outer span NER annotation) are extracted:
+
+```bash
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-train.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > train.txt.tmp
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-dev.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > dev.txt.tmp
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-test.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > test.txt.tmp
+```
+
+The GermEval 2014 dataset contains some strange "control character" tokens like `'\x96', '\u200e', '\x95', '\xad' or '\x80'`. One problem with these tokens is, that `BertTokenizer` returns an empty token for them, resulting in misaligned `InputExample`s. I wrote a script that a) filters these tokens and b) splits longer sentences into smaller ones (once the max. subtoken length is reached).
+
+```bash
+wget "https://raw.githubusercontent.com/stefan-it/fine-tuned-berts-seq/master/scripts/preprocess.py"
+```
+Let's define some variables that we need for further pre-processing steps and training the model:
+
+```bash
+export MAX_LENGTH=128
+export BERT_MODEL=bert-base-multilingual-cased
+```
+
+Run the pre-processing script on training, dev and test datasets:
+
+```bash
+python3 preprocess.py train.txt.tmp $BERT_MODEL $MAX_LENGTH > train.txt
+python3 preprocess.py dev.txt.tmp $BERT_MODEL $MAX_LENGTH > dev.txt
+python3 preprocess.py test.txt.tmp $BERT_MODEL $MAX_LENGTH > test.txt
+```
+
+The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so an own set of labels must be used:
+
+```bash
+cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
+```
+
+### Training
+
+Additional environment variables must be set:
+
+```bash
+export OUTPUT_DIR=germeval-model
+export BATCH_SIZE=32
+export NUM_EPOCHS=3
+export SAVE_STEPS=750
+export SEED=1
+```
+
+To start training, just run:
+
+```bash
+python3 run_ner.py --data_dir ./ \
+--model_type bert \
+--labels ./labels.txt \
+--model_name_or_path $BERT_MODEL \
+--output_dir $OUTPUT_DIR \
+--max_seq_length  $MAX_LENGTH \
+--num_train_epochs $NUM_EPOCHS \
+--per_gpu_train_batch_size $BATCH_SIZE \
+--save_steps $SAVE_STEPS \
+--seed $SEED \
+--do_train \
+--do_eval \
+--do_predict
+```
+
+If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
+
+### Evaluation
+
+Evaluation on development dataset outputs the following for our example:
+
+```bash
+10/04/2019 00:42:06 - INFO - __main__ -   ***** Eval results  *****
+10/04/2019 00:42:06 - INFO - __main__ -     f1 = 0.8623348017621146
+10/04/2019 00:42:06 - INFO - __main__ -     loss = 0.07183869666975543
+10/04/2019 00:42:06 - INFO - __main__ -     precision = 0.8467916366258111
+10/04/2019 00:42:06 - INFO - __main__ -     recall = 0.8784592370979806
+```
+
+On the test dataset the following results could be achieved:
+
+```bash
+10/04/2019 00:42:42 - INFO - __main__ -   ***** Eval results  *****
+10/04/2019 00:42:42 - INFO - __main__ -     f1 = 0.8614389652384803
+10/04/2019 00:42:42 - INFO - __main__ -     loss = 0.07064602487454782
+10/04/2019 00:42:42 - INFO - __main__ -     precision = 0.8604651162790697
+10/04/2019 00:42:42 - INFO - __main__ -     recall = 0.8624150210424085
+```
+
+### Comparing BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased)
+
+Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased) with the same hyperparameters as specified in the [example documentation](https://huggingface.co/transformers/examples.html#named-entity-recognition) (one run):
+
+| Model | F-Score Dev | F-Score Test
+| --------------------------------- | ------- | --------
+| `bert-large-cased`            | 95.59 | 91.70
+| `roberta-large`                  | 95.96 | 91.87
+| `distilbert-base-uncased` | 94.34 | 90.32
+
+## Abstractive summarization
+
+Based on the script
+[`run_summarization_finetuning.py`](https://github.com/huggingface/transformers/blob/master/examples/run_summarization_finetuning.py).
+
+Before running this script you should download **both** CNN and Daily Mail
+datasets from [Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/)  (the
+links next to "Stories") in the same folder. Then uncompress the archives by running:
+
+```bash
+tar -xvf cnn_stories.tgz && tar -xvf dailymail_stories.tgz
+```
+
+note that the finetuning script **will not work** if you do not download both
+datasets. We will refer as `$DATA_PATH` the path to where you uncompressed both
+archive.
+
+```bash
+export DATA_PATH=/path/to/dataset/
+
+python run_summarization_finetuning.py \
+    --output_dir=output \
+    --model_type=bert2bert \
+    --model_name_or_path=bert2bert \
+    --do_train \
+    --data_path=$DATA_PATH \
+```

examples/benchmarks.py

@@ -0,0 +1,477 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Benchmarking the library on inference and training """
+
+# If checking the tensors placement
+# tf.debugging.set_log_device_placement(True)
+
+from typing import List
+import timeit
+from transformers import is_tf_available, is_torch_available
+from time import time
+import argparse
+import csv
+
+if is_tf_available():
+    import tensorflow as tf
+    from transformers import TFAutoModel
+
+if is_torch_available():
+    import torch
+    from transformers import AutoModel
+
+from transformers import AutoConfig, AutoTokenizer
+
+input_text = """Bent over their instruments, three hundred Fertilizers were plunged, as 
+the Director of Hatcheries and Conditioning entered the room, in the 
+
+
+
+scarcely breathing silence, the absent-minded, soliloquizing hum or 
+whistle, of absorbed concentration. A troop of newly arrived students, 
+very young, pink and callow, followed nervously, rather abjectly, at the 
+Director's heels. Each of them carried a notebook, in which, whenever 
+the great man spoke, he desperately scribbled. Straight from the 
+horse's mouth. It was a rare privilege. The D. H. C. for Central London 
+always made a point of personally conducting his new students round 
+the various departments. 
+
+"Just to give you a general idea," he would explain to them. For of 
+course some sort of general idea they must have, if they were to do 
+their work intelligently-though as little of one, if they were to be good 
+and happy members of society, as possible. For particulars, as every 
+one knows, make for virtue and happiness; generalities are intellectu- 
+ally necessary evils. Not philosophers but fret-sawyers and stamp col- 
+lectors compose the backbone of society. 
+
+"To-morrow," he would add, smiling at them with a slightly menacing 
+geniality, "you'll be settling down to serious work. You won't have time 
+for generalities. Meanwhile ..." 
+
+Meanwhile, it was a privilege. Straight from the horse's mouth into the 
+notebook. The boys scribbled like mad. 
+
+Tall and rather thin but upright, the Director advanced into the room. 
+He had a long chin and big rather prominent teeth, just covered, when 
+he was not talking, by his full, floridly curved lips. Old, young? Thirty? 
+Fifty? Fifty-five? It was hard to say. And anyhow the question didn't 
+arise; in this year of stability, A. F. 632, it didn't occur to you to ask it. 
+
+"I shall begin at the beginning," said the D.H.C. and the more zealous 
+students recorded his intention in their notebooks: Begin at the begin- 
+ning. "These," he waved his hand, "are the incubators." And opening 
+an insulated door he showed them racks upon racks of numbered test- 
+tubes. "The week's supply of ova. Kept," he explained, "at blood heat; 
+whereas the male gametes," and here he opened another door, "they 
+have to be kept at thirty-five instead of thirty-seven. Full blood heat 
+sterilizes." Rams wrapped in theremogene beget no lambs. 
+
+Still leaning against the incubators he gave them, while the pencils 
+scurried illegibly across the pages, a brief description of the modern 
+
+
+
+fertilizing process; spoke first, of course, of its surgical introduc- 
+tion-"the operation undergone voluntarily for the good of Society, not 
+to mention the fact that it carries a bonus amounting to six months' 
+salary"; continued with some account of the technique for preserving 
+the excised ovary alive and actively developing; passed on to a consid- 
+eration of optimum temperature, salinity, viscosity; referred to the liq- 
+uor in which the detached and ripened eggs were kept; and, leading 
+his charges to the work tables, actually showed them how this liquor 
+was drawn off from the test-tubes; how it was let out drop by drop 
+onto the specially warmed slides of the microscopes; how the eggs 
+which it contained were inspected for abnormalities, counted and 
+transferred to a porous receptacle; how (and he now took them to 
+watch the operation) this receptacle was immersed in a warm bouillon 
+containing free-swimming spermatozoa-at a minimum concentration 
+of one hundred thousand per cubic centimetre, he insisted; and how, 
+after ten minutes, the container was lifted out of the liquor and its 
+contents re-examined; how, if any of the eggs remained unfertilized, it 
+was again immersed, and, if necessary, yet again; how the fertilized 
+ova went back to the incubators; where the Alphas and Betas re- 
+mained until definitely bottled; while the Gammas, Deltas and Epsilons 
+were brought out again, after only thirty-six hours, to undergo Bo- 
+kanovsky's Process. 
+
+"Bokanovsky's Process," repeated the Director, and the students un- 
+derlined the words in their little notebooks. 
+
+One egg, one embryo, one adult-normality. But a bokanovskified egg 
+will bud, will proliferate, will divide. From eight to ninety-six buds, and 
+every bud will grow into a perfectly formed embryo, and every embryo 
+into a full-sized adult. Making ninety-six human beings grow where 
+only one grew before. Progress. 
+
+"Essentially," the D.H.C. concluded, "bokanovskification consists of a 
+series of arrests of development. We check the normal growth and, 
+paradoxically enough, the egg responds by budding." 
+
+Responds by budding. The pencils were busy. 
+
+He pointed. On a very slowly moving band a rack-full of test-tubes was 
+entering a large metal box, another, rack-full was emerging. Machinery 
+faintly purred. It took eight minutes for the tubes to go through, he 
+
+
+
+told them. Eight minutes of hard X-rays being about as much as an 
+egg can stand. A few died; of the rest, the least susceptible divided 
+into two; most put out four buds; some eight; all were returned to the 
+incubators, where the buds began to develop; then, after two days, 
+were suddenly chilled, chilled and checked. Two, four, eight, the buds 
+in their turn budded; and having budded were dosed almost to death 
+with alcohol; consequently burgeoned again and having budded-bud 
+out of bud out of bud-were thereafter-further arrest being generally 
+fatal-left to develop in peace. By which time the original egg was in a 
+fair way to becoming anything from eight to ninety-six embryos- a 
+prodigious improvement, you will agree, on nature. Identical twins-but 
+not in piddling twos and threes as in the old viviparous days, when an 
+egg would sometimes accidentally divide; actually by dozens, by 
+scores at a time. 
+
+"Scores," the Director repeated and flung out his arms, as though he 
+were distributing largesse. "Scores." 
+
+But one of the students was fool enough to ask where the advantage 
+lay. 
+
+"My good boy!" The Director wheeled sharply round on him. "Can't you 
+see? Can't you see?" He raised a hand; his expression was solemn. 
+"Bokanovsky's Process is one of the major instruments of social stabil- 
+ity!" 
+
+Major instruments of social stability. 
+
+Standard men and women; in uniform batches. The whole of a small 
+factory staffed with the products of a single bokanovskified egg. 
+
+"Ninety-six identical twins working ninety-six identical machines!" The 
+voice was almost tremulous with enthusiasm. "You really know where 
+you are. For the first time in history." He quoted the planetary motto. 
+"Community, Identity, Stability." Grand words. "If we could bo- 
+kanovskify indefinitely the whole problem would be solved." 
+
+Solved by standard Gammas, unvarying Deltas, uniform Epsilons. Mil- 
+lions of identical twins. The principle of mass production at last applied 
+to biology. 
+
+
+
+"But, alas," the Director shook his head, "we can't bokanovskify indefi- 
+nitely." 
+
+Ninety-six seemed to be the limit; seventy-two a good average. From 
+the same ovary and with gametes of the same male to manufacture as 
+many batches of identical twins as possible-that was the best (sadly a 
+second best) that they could do. And even that was difficult. 
+
+"For in nature it takes thirty years for two hundred eggs to reach ma- 
+turity. But our business is to stabilize the population at this moment, 
+here and now. Dribbling out twins over a quarter of a century-what 
+would be the use of that?" 
+
+Obviously, no use at all. But Podsnap's Technique had immensely ac- 
+celerated the process of ripening. They could make sure of at least a 
+hundred and fifty mature eggs within two years. Fertilize and bo- 
+kanovskify-in other words, multiply by seventy-two-and you get an 
+average of nearly eleven thousand brothers and sisters in a hundred 
+and fifty batches of identical twins, all within two years of the same 
+age. 
+
+"And in exceptional cases we can make one ovary yield us over fifteen 
+thousand adult individuals." 
+
+Beckoning to a fair-haired, ruddy young man who happened to be 
+passing at the moment. "Mr. Foster," he called. The ruddy young man 
+approached. "Can you tell us the record for a single ovary, Mr. Foster?" 
+
+"Sixteen thousand and twelve in this Centre," Mr. Foster replied with- 
+out hesitation. He spoke very quickly, had a vivacious blue eye, and 
+took an evident pleasure in quoting figures. "Sixteen thousand and 
+twelve; in one hundred and eighty-nine batches of identicals. But of 
+course they've done much better," he rattled on, "in some of the tropi- 
+cal Centres. Singapore has often produced over sixteen thousand five 
+hundred; and Mombasa has actually touched the seventeen thousand 
+mark. But then they have unfair advantages. You should see the way a 
+negro ovary responds to pituitary! It's quite astonishing, when you're 
+used to working with European material. Still," he added, with a laugh 
+(but the light of combat was in his eyes and the lift of his chin was 
+challenging), "still, we mean to beat them if we can. I'm working on a 
+wonderful Delta-Minus ovary at this moment. Only just eighteen 
+
+
+
+months old. Over twelve thousand seven hundred children already, ei- 
+ther decanted or in embryo. And still going strong. We'll beat them 
+yet." 
+
+"That's the spirit I like!" cried the Director, and clapped Mr. Foster on 
+the shoulder. "Come along with us, and give these boys the benefit of 
+your expert knowledge." 
+
+Mr. Foster smiled modestly. "With pleasure." They went. 
+In the Bottling Room all was harmonious bustle and ordered activity. 
+Flaps of fresh sow's peritoneum ready cut to the proper size came 
+shooting up in little lifts from the Organ Store in the sub-basement. 
+Whizz and then, click! the lift-hatches hew open; the bottle-liner had 
+only to reach out a hand, take the flap, insert, smooth-down, and be- 
+fore the lined bottle had had time to travel out of reach along the end- 
+less band, whizz, click! another flap of peritoneum had shot up from 
+the depths, ready to be slipped into yet another bottle, the next of that 
+slow interminable procession on the band. 
+
+Next to the Liners stood the Matriculators. The procession advanced; 
+one by one the eggs were transferred from their test-tubes to the 
+larger containers; deftly the peritoneal lining was slit, the morula 
+dropped into place, the saline solution poured in ... and already the 
+bottle had passed, and it was the turn of the labellers. Heredity, date 
+of fertilization, membership of Bokanovsky Group-details were trans- 
+ferred from test-tube to bottle. No longer anonymous, but named, 
+identified, the procession marched slowly on; on through an opening in 
+the wall, slowly on into the Social Predestination Room. 
+"Eighty-eight cubic metres of card-index," said Mr. Foster with relish, 
+as they entered."""
+
+
+def create_setup_and_compute(model_names: List[str],
+                             gpu: bool = True,
+                             tensorflow: bool = False,
+                             average_over: int = 3,
+                             torchscript: bool = False,
+                             xla: bool = False,
+                             amp: bool = False,
+                             fp16: bool = False,
+                             save_to_csv: bool = False,
+                             csv_filename: str = f"results_{round(time())}.csv"):
+    if xla:
+        tf.config.optimizer.set_jit(True)
+    if amp:
+        tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})
+
+    if tensorflow:
+        dictionary = {model_name: {} for model_name in model_names}
+        results = _compute_tensorflow(model_names, dictionary, average_over, amp)
+    else:
+        device = 'cuda' if (gpu and torch.cuda.is_available()) else 'cpu'
+        dictionary = {model_name: {} for model_name in model_names}
+        results = _compute_pytorch(model_names, dictionary, average_over, device, torchscript, fp16)
+
+    print("=========== RESULTS ===========")
+    for model_name in model_names:
+        print("\t" + f"======= MODEL CHECKPOINT: {model_name} =======")
+        for batch_size in results[model_name]["bs"]:
+            print("\t\t" + f"===== BATCH SIZE: {batch_size} =====")
+            for slice_size in results[model_name]["ss"]:
+                result = results[model_name]['results'][batch_size][slice_size]
+                if isinstance(result, str):
+                    print(f"\t\t{model_name}/{batch_size}/{slice_size}: "
+                          f"{result}")
+                else:
+                    print(f"\t\t{model_name}/{batch_size}/{slice_size}: "
+                          f"{(round(1000 * result) / 1000)}"
+                          f"s")
+
+    if save_to_csv:
+        with open(csv_filename, mode='w') as csv_file:
+            fieldnames = ['model',
+                          '1x8', '1x64', '1x128', '1x256', '1x512', '1x1024',
+                          '2x8', '2x64', '2x128', '2x256', '2x512', '2x1024',
+                          '4x8', '4x64', '4x128', '4x256', '4x512', '4x1024',
+                          '8x8', '8x64', '8x128', '8x256', '8x512', '8x1024',
+                          ]
+
+            writer = csv.DictWriter(csv_file, fieldnames=fieldnames)
+            writer.writeheader()
+
+            for model_name in model_names:
+                model_results = {
+                    f'{bs}x{ss}': results[model_name]['results'][bs][ss]
+                    for bs in results[model_name]["results"]
+                    for ss in results[model_name]['results'][bs]
+                }
+                writer.writerow({'model': model_name, **model_results})
+
+
+def _compute_pytorch(model_names, dictionary, average_over, device, torchscript, fp16):
+    for c, model_name in enumerate(model_names):
+        print(f"{c + 1} / {len(model_names)}")
+        config = AutoConfig.from_pretrained(model_name, torchscript=torchscript)
+        model = AutoModel.from_pretrained(model_name, config=config)
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+        tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False)
+
+        max_input_size = tokenizer.max_model_input_sizes[model_name]
+        batch_sizes = [1, 2, 4, 8]
+        slice_sizes = [8, 64, 128, 256, 512, 1024]
+
+        dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}}
+        dictionary[model_name]["results"] = {i: {} for i in batch_sizes}
+
+        for batch_size in batch_sizes:
+            if fp16:
+                model.half()
+            model.to(device)
+            model.eval()
+            for slice_size in slice_sizes:
+                if max_input_size is not None and slice_size > max_input_size:
+                    dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+                else:
+                    sequence = torch.tensor(tokenized_sequence[:slice_size], device=device).repeat(batch_size, 1)
+                    try:
+                        if torchscript:
+                            print("Tracing model with sequence size", sequence.shape)
+                            inference = torch.jit.trace(model, sequence)
+                            inference(sequence)
+                        else:
+                            inference = model
+                            inference(sequence)
+
+                        print("Going through model with sequence of shape", sequence.shape)
+                        runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
+                        average_time = sum(runtimes)/float(len(runtimes)) / 3.0
+                        dictionary[model_name]["results"][batch_size][slice_size] = average_time
+                    except RuntimeError as e:
+                        print("Doesn't fit on GPU.", e)
+                        torch.cuda.empty_cache()
+                        dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+    return dictionary
+
+
+def _compute_tensorflow(model_names, dictionary, average_over, amp):
+    for c, model_name in enumerate(model_names):
+        print(f"{c + 1} / {len(model_names)}")
+        config = AutoConfig.from_pretrained(model_name)
+        model = TFAutoModel.from_pretrained(model_name, config=config)
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+        tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False)
+
+        max_input_size = tokenizer.max_model_input_sizes[model_name]
+        batch_sizes = [1, 2, 4, 8]
+        slice_sizes = [8, 64, 128, 256, 512, 1024]
+
+        dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}}
+        dictionary[model_name]["results"] = {i: {} for i in batch_sizes}
+
+        print("Using model", model)
+
+        @tf.function
+        def inference(inputs):
+            return model(inputs)
+
+        for batch_size in batch_sizes:
+            for slice_size in slice_sizes:
+                if max_input_size is not None and slice_size > max_input_size:
+                    dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+                else:
+                    sequence = tf.stack([tf.squeeze(tf.constant(tokenized_sequence[:slice_size])[None, :])] * batch_size)
+
+                    try:
+                        print("Going through model with sequence of shape", sequence.shape)
+                        # To make sure that the model is traced + that the tensors are on the appropriate device
+                        inference(sequence)
+
+                        runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
+                        average_time = sum(runtimes)/float(len(runtimes)) / 3.0
+                        dictionary[model_name]["results"][batch_size][slice_size] = average_time
+                    except tf.errors.ResourceExhaustedError as e:
+                        print("Doesn't fit on GPU.", e)
+                        torch.cuda.empty_cache()
+                        dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+    return dictionary
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--models", required=False, type=str, default='all', help="Model checkpoints to be provided "
+                                                                                  "to the AutoModel classes. Leave "
+                                                                                  "blank to benchmark the base version "
+                                                                                  "of all available model "
+                                                                                  "architectures.")
+    parser.add_argument("--torch", required=False, action="store_true", help="Benchmark the Pytorch version of the "
+                                                                             "models")
+    parser.add_argument("--torch_cuda", required=False, action="store_true", help="Pytorch only: run on available "
+                                                                                  "cuda devices")
+    parser.add_argument("--torchscript", required=False, action="store_true", help="Pytorch only: trace the models "
+                                                                                   "using torchscript")
+    parser.add_argument("--tensorflow", required=False, action="store_true", help="Benchmark the TensorFlow version "
+                                                                                  "of the models. Will run on GPU if "
+                                                                                  "the correct dependencies are "
+                                                                                  "installed")
+    parser.add_argument("--xla", required=False, action="store_true", help="TensorFlow only: use XLA acceleration.")
+    parser.add_argument("--amp", required=False, action="store_true", help="TensorFlow only: use automatic mixed precision acceleration.")
+    parser.add_argument("--fp16", required=False, action="store_true", help="PyTorch only: use FP16 to accelerate inference.")
+    parser.add_argument("--keras_predict", required=False, action="store_true", help="Whether to use model.predict "
+                                                                                     "instead of model() to do a "
+                                                                                     "forward pass.")
+    parser.add_argument("--save_to_csv", required=False, action="store_true", help="Save to a CSV file.")
+    parser.add_argument("--csv_filename", required=False, default=None, help="CSV filename used if saving results to csv.")
+    parser.add_argument("--average_over", required=False, default=30, type=int, help="Times an experiment will be run.")
+
+    args = parser.parse_args()
+    if args.models == 'all':
+        args.models = [
+            "gpt2",
+            "bert-base-cased",
+            "xlnet-base-cased",
+            "xlm-mlm-en-2048",
+            "transfo-xl-wt103",
+            "openai-gpt",
+            "distilbert-base-uncased",
+            "distilgpt2",
+            "roberta-base",
+            "ctrl"
+        ]
+    else:
+        args.models = args.models.split()
+
+    print("Running with arguments", args)
+
+    if args.torch:
+        if is_torch_available():
+            create_setup_and_compute(
+                model_names=args.models,
+                tensorflow=False,
+                gpu=args.torch_cuda,
+                torchscript=args.torchscript,
+                fp16=args.fp16,
+                save_to_csv=args.save_to_csv,
+                csv_filename=args.csv_filename,
+                average_over=args.average_over
+            )
+        else:
+            raise ImportError("Trying to run a PyTorch benchmark but PyTorch was not found in the environment.")
+
+    if args.tensorflow:
+        if is_tf_available():
+            create_setup_and_compute(
+                model_names=args.models,
+                tensorflow=True,
+                xla=args.xla,
+                amp=args.amp,
+                save_to_csv=args.save_to_csv,
+                csv_filename=args.csv_filename,
+                average_over=args.average_over
+            )
+        else:
+            raise ImportError("Trying to run a TensorFlow benchmark but TensorFlow was not found in the environment.")
+
+if __name__ == '__main__':
+    main()
+

examples/contrib/run_camembert.py

@@ -0,0 +1,48 @@
+from pathlib import Path
+import tarfile
+import urllib.request
+
+import torch
+
+from transformers.tokenization_camembert import CamembertTokenizer
+from transformers.modeling_camembert import CamembertForMaskedLM
+
+
+def fill_mask(masked_input, model, tokenizer, topk=5):
+    # Adapted from https://github.com/pytorch/fairseq/blob/master/fairseq/models/roberta/hub_interface.py
+    assert masked_input.count('<mask>') == 1
+    input_ids = torch.tensor(tokenizer.encode(masked_input, add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+    logits = model(input_ids)[0]  # The last hidden-state is the first element of the output tuple
+    masked_index = (input_ids.squeeze() == tokenizer.mask_token_id).nonzero().item()
+    logits = logits[0, masked_index, :]
+    prob = logits.softmax(dim=0)
+    values, indices = prob.topk(k=topk, dim=0)
+    topk_predicted_token_bpe = ' '.join([tokenizer.convert_ids_to_tokens(indices[i].item())
+                                         for i in range(len(indices))])
+    masked_token = tokenizer.mask_token
+    topk_filled_outputs = []
+    for index, predicted_token_bpe in enumerate(topk_predicted_token_bpe.split(' ')):
+        predicted_token = predicted_token_bpe.replace('\u2581', ' ')
+        if " {0}".format(masked_token) in masked_input:
+            topk_filled_outputs.append((
+                masked_input.replace(
+                    ' {0}'.format(masked_token), predicted_token
+                ),
+                values[index].item(),
+                predicted_token,
+            ))
+        else:
+            topk_filled_outputs.append((
+                masked_input.replace(masked_token, predicted_token),
+                values[index].item(),
+                predicted_token,
+            ))
+    return topk_filled_outputs
+
+
+tokenizer = CamembertTokenizer.from_pretrained('camembert-base')
+model = CamembertForMaskedLM.from_pretrained('camembert-base')
+model.eval()
+
+masked_input = "Le camembert est <mask> :)"
+print(fill_mask(masked_input, model, tokenizer, topk=3))

examples/contrib/run_openai_gpt.py

@@ -41,7 +41,7 @@ from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
 
 from transformers import (OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer,
                                      AdamW, cached_path, WEIGHTS_NAME, CONFIG_NAME,
-                                     WarmupLinearSchedule)
+                                     get_linear_schedule_with_warmup)
 
 ROCSTORIES_URL = "https://s3.amazonaws.com/datasets.huggingface.co/ROCStories.tar.gz"
 
@@ -211,7 +211,7 @@ def main():
             {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
             ]
         optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-        scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+        scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
 
     if args.do_train:
         nb_tr_steps, tr_loss, exp_average_loss = 0, 0, None
@@ -237,7 +237,7 @@ def main():
     # Save a trained model
     if args.do_train:
         # Save a trained model, configuration and tokenizer
-        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
+        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model itself
 
         # If we save using the predefined names, we can load using `from_pretrained`
         output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)

examples/contrib/run_swag.py

@@ -31,14 +31,18 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                               TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm, trange
 
-from tensorboardX import SummaryWriter
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
+from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
                                   BertForMultipleChoice, BertTokenizer)
 
-from transformers import AdamW, WarmupLinearSchedule
+from transformers import AdamW, get_linear_schedule_with_warmup
 
 logger = logging.getLogger(__name__)
 
@@ -318,7 +322,7 @@ def train(args, train_dataset, model, tokenizer):
         {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
     if args.fp16:
         try:
             from apex import amp

examples/distillation/README.md

@@ -1,31 +1,47 @@
 # Distil*
 
-This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT and DistilGPT2.
+This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT, DistilRoBERTa and DistilGPT2.
 
-**2019, October 3rd - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2.
+**October 23rd, 2019 - Update** We release **DistilRoBERTa**: 95% of `RoBERTa-base`'s performance on GLUE, twice as fast as RoBERTa while being 35% smaller.
+
+**October 3rd, 2019 - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2. **The paper superseeds our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances. Please use the paper as a reference when comparing/reporting results on DistilBERT.**
+
+**September 19th, 2019 - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 97% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
 
-**2019, September 19th - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 97% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
 
 ## What is Distil*
 
 Distil* is a class of compressed models that started with DistilBERT. DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
 
-We have applied the same method to GPT2 and release the weights of the compressed model. On the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 15.0 compared to 18.5 for DistilGPT2 (after fine-tuning on the train set).
+We have applied the same method to other Transformer architectures and released the weights:
+- GPT2: on the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 15.0 compared to 18.5 for **DistilGPT2** (after fine-tuning on the train set).
+- RoBERTa: **DistilRoBERTa** reaches 95% of `RoBERTa-base` performance on GLUE while being twice faster and 35% smaller.
+- and more to come! 🤗🤗🤗
 
-For more information on DistilBERT, please refer to our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108). The paper superseeds our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances.
+For more information on DistilBERT, please refer to our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108).
 
 Here are the results on the dev sets of GLUE:
 
-| Model      | Macro-score | CoLA | MNLI | MRPC | QNLI | QQP  | RTE  | SST-2| STS-B| WNLI |
-| :---:      |    :---:    | :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:|
-| BERT-base  |  **77.6**   | 48.9 | 84.3 | 88.6 | 89.3 | 89.5 | 71.3 | 91.7 | 91.2 | 43.7 |
-| DistilBERT |  **76.8**   | 49.1 | 81.8 | 90.2 | 90.2 | 89.2 | 62.9 | 92.7 | 90.7 | 44.4 |
+| Model                     | Macro-score                    | CoLA | MNLI | MRPC | QNLI | QQP  | RTE  | SST-2| STS-B| WNLI              |
+| :---:                     |    :---:                       | :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:             |
+| BERT-base                 |  **77.6**                      | 48.9 | 84.3 | 88.6 | 89.3 | 89.5 | 71.3 | 91.7 | 91.2 | 43.7              |
+| DistilBERT                |  **76.8**                      | 49.1 | 81.8 | 90.2 | 90.2 | 89.2 | 62.9 | 92.7 | 90.7 | 44.4              |
+| ---                       |    ---                         |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  ---              |
+| RoBERTa-base (reported)   |  **83.2**/**86.4**<sup>2</sup> | 63.6 | 87.6 | 90.2 | 92.8 | 91.9 | 78.7 | 94.8 | 91.2 | 57.7<sup>3</sup> |
+| DistilRoBERTa<sup>1</sup> |  **79.0**/**82.3**<sup>2</sup> | 59.4 | 83.9 | 86.6 | 90.8 | 89.4 | 67.9 | 92.5 | 88.3 | 52.1              |
+
+<sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directy perform transfer learning on the pre-trained DistilRoBERTa.
+
+<sup>2</sup> Macro-score computed without WNLI.
+
+<sup>3</sup> We compute this score ourselves for completeness.
 
 ## Setup
 
 This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`. 
 
-**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0). It is important to note that there is a small internal bug in the current version of PyTorch available on pip that causes a memory leak in our training/distillation. It has been recently fixed and will likely be integrated into the next release. For the moment, we recommend to [compile PyTorch from source](https://github.com/pytorch/pytorch#from-source). Please refer to [issue 1179](https://github.com/huggingface/transformers/issues/1179) for more details.
+**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0).
+
 
 ## How to use DistilBERT
 
@@ -33,7 +49,8 @@ Transformers includes two pre-trained Distil* models, currently only provided fo
 
 - `distilbert-base-uncased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-uncased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 66M parameters.
 - `distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.9 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
-- `distilgpt2`: DistilGPT2 English language model pretrained with the supervision of `gpt2` (the smallest version of GPT2) on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset and . The model has 6 layers, 768 dimension and 12 heads, totalizing 82M (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
+- `distilgpt2`: DistilGPT2 English language model pretrained with the supervision of `gpt2` (the smallest version of GPT2) on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset. The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
+- `distilroberta-base`: DistilRoBERTa English language model pretrained with the supervision of `roberta-base` solely on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset (it is ~4 times less training data than the teacher RoBERTa). The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 125M parameters for RoBERTa-base). On average DistilRoBERTa is twice as fast as Roberta-base.
 - and more to come! 🤗🤗🤗
 
 Using DistilBERT is very similar to using BERT. DistilBERT share the same tokenizer as BERT's `bert-base-uncased` even though we provide a link to this tokenizer under the `DistilBertTokenizer` name to have a consistent naming between the library models.
@@ -47,7 +64,10 @@ outputs = model(input_ids)
 last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 ```
 
-Similarly, using DistilGPT2 simply consists in calling the GPT2 classes from a different pretrained checkpoint: `model = GPT2Model.from_pretrained('distilgpt2')`.
+Similarly, using the other Distil* models simply consists in calling the base classes with a different pretrained checkpoint:
+- DistilGPT2: `model = GPT2Model.from_pretrained('distilgpt2')`
+- DistilRoBERTa: `model = RobertaModel.from_pretrained('distilroberta-base')`
+
 
 ## How to train Distil*
 
@@ -88,7 +108,7 @@ python train.py \
     --student_config training_configs/distilbert-base-uncased.json \
     --teacher_type bert \
     --teacher_name bert-base-uncased \
-    --alpha_ce 5.0 --alpha_mlm 2.0 --alpha_cos 1.0 --mlm \
+    --alpha_ce 5.0 --alpha_mlm 2.0 --alpha_cos 1.0 --alpha_clm 0.0 --mlm \
     --freeze_pos_embs \
     --dump_path serialization_dir/my_first_training \
     --data_file data/binarized_text.bert-base-uncased.pickle \
@@ -124,7 +144,7 @@ python -m torch.distributed.launch \
         --student_config training_configs/distilbert-base-uncased.json \
         --teacher_type bert \
         --teacher_name bert-base-uncased \
-        --alpha_ce 0.33 --alpha_mlm 0.33 --alpha_cos 0.33 --mlm \
+        --alpha_ce 0.33 --alpha_mlm 0.33 --alpha_cos 0.33 --alpha_clm 0.0 --mlm \
         --freeze_pos_embs \
         --dump_path serialization_dir/my_first_training \
         --data_file data/binarized_text.bert-base-uncased.pickle \
@@ -134,3 +154,16 @@ python -m torch.distributed.launch \
 **Tips:** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract.py` and `scripts/extract_distilbert.py` to create a valid initialization checkpoint and use `--student_pretrained_weights` argument to use this initialization for the distilled training!
 
 Happy distillation!
+
+## Citation
+
+If you find the ressource useful, you should cite the following paper:
+
+```
+@inproceedings{sanh2019distilbert,
+  title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
+  author={Sanh, Victor and Debut, Lysandre and Chaumond, Julien and Wolf, Thomas},
+  booktitle={NeurIPS EMC^2 Workshop},
+  year={2019}
+}
+```

examples/distillation/distiller.py

@@ -19,7 +19,6 @@ import os
 import math
 import psutil
 import time
-from tensorboardX import SummaryWriter
 from tqdm import trange, tqdm
 import numpy as np
 import psutil
@@ -31,7 +30,12 @@ from torch.optim import AdamW
 from torch.utils.data.distributed import DistributedSampler
 from torch.utils.data import RandomSampler, BatchSampler, DataLoader
 
-from transformers import WarmupLinearSchedule
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
+from transformers import get_linear_schedule_with_warmup
 
 from utils import logger
 from lm_seqs_dataset import LmSeqsDataset
@@ -133,9 +137,9 @@ class Distiller:
                                betas=(0.9, 0.98))
 
         warmup_steps = math.ceil(num_train_optimization_steps * params.warmup_prop)
-        self.scheduler = WarmupLinearSchedule(self.optimizer,
-                                                warmup_steps=warmup_steps,
-                                                t_total=num_train_optimization_steps)
+        self.scheduler = get_linear_schedule_with_warmup(self.optimizer,
+                                                num_warmup_steps=warmup_steps,
+                                                num_training_steps=num_train_optimization_steps)
 
         if self.fp16:
             try:

examples/distillation/run_squad_w_distillation.py

@@ -30,9 +30,13 @@ from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
 from torch.utils.data.distributed import DistributedSampler
 import torch.nn.functional as F
 import torch.nn as nn
-from tqdm import tqdm, trange
 
-from tensorboardX import SummaryWriter
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
+from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
                                   BertForQuestionAnswering, BertTokenizer,
@@ -42,7 +46,7 @@ from transformers import (WEIGHTS_NAME, BertConfig,
                                   XLNetTokenizer,
                                   DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
 
-from transformers import AdamW, WarmupLinearSchedule
+from transformers import AdamW, get_linear_schedule_with_warmup
 
 from ..utils_squad import (read_squad_examples, convert_examples_to_features,
                          RawResult, write_predictions,
@@ -97,7 +101,7 @@ def train(args, train_dataset, model, tokenizer, teacher=None):
         {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
     if args.fp16:
         try:
             from apex import amp
@@ -502,9 +506,15 @@ def main():
 
     args.model_type = args.model_type.lower()
     config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
-    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
+                                          cache_dir=args.cache_dir if args.cache_dir else None)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+                                                do_lower_case=args.do_lower_case,
+                                                cache_dir=args.cache_dir if args.cache_dir else None)
+    model = model_class.from_pretrained(args.model_name_or_path,
+                                        from_tf=bool('.ckpt' in args.model_name_or_path),
+                                        config=config,
+                                        cache_dir=args.cache_dir if args.cache_dir else None)
 
     if args.teacher_type is not None:
         assert args.teacher_name_or_path is not None
@@ -512,8 +522,11 @@ def main():
         assert args.alpha_ce + args.alpha_squad > 0.
         assert args.teacher_type != 'distilbert', "We constraint teachers not to be of type DistilBERT."
         teacher_config_class, teacher_model_class, _ = MODEL_CLASSES[args.teacher_type]
-        teacher_config = teacher_config_class.from_pretrained(args.teacher_name_or_path)
-        teacher = teacher_model_class.from_pretrained(args.teacher_name_or_path, config=teacher_config)
+        teacher_config = teacher_config_class.from_pretrained(args.teacher_name_or_path,
+                                                              cache_dir=args.cache_dir if args.cache_dir else None)
+        teacher = teacher_model_class.from_pretrained(args.teacher_name_or_path,
+                                                      config=teacher_config,
+                                                      cache_dir=args.cache_dir if args.cache_dir else None)
         teacher.to(args.device)
     else:
         teacher = None
@@ -549,8 +562,10 @@ def main():
         torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
 
         # Load a trained model and vocabulary that you have fine-tuned
-        model = model_class.from_pretrained(args.output_dir)
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model = model_class.from_pretrained(args.output_dir, cache_dir=args.cache_dir if args.cache_dir else None)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir,
+                                                    do_lower_case=args.do_lower_case,
+                                                    cache_dir=args.cache_dir if args.cache_dir else None)
         model.to(args.device)
 
 
@@ -567,7 +582,7 @@ def main():
         for checkpoint in checkpoints:
             # Reload the model
             global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
-            model = model_class.from_pretrained(checkpoint)
+            model = model_class.from_pretrained(checkpoint, cache_dir=args.cache_dir if args.cache_dir else None)
             model.to(args.device)
 
             # Evaluate

examples/distillation/scripts/binarized_data.py

@@ -68,7 +68,7 @@ def main():
     start = time.time()
     for text in data:
         text = f'{bos} {text.strip()} {sep}'
-        token_ids = tokenizer.encode(text)
+        token_ids = tokenizer.encode(text, add_special_tokens=False)
         rslt.append(token_ids)
 
         iter += 1

examples/requirements.txt

@@ -1,2 +1,4 @@
 tensorboardX
-scikit-learn
+tensorboard
+scikit-learn
+seqeval

examples/run_bertology.py

@@ -39,8 +39,9 @@ from transformers import (WEIGHTS_NAME,
 
 from run_glue import set_seed, load_and_cache_examples, ALL_MODELS, MODEL_CLASSES
 
-from utils_glue import (compute_metrics, convert_examples_to_features,
-                        output_modes, processors)
+from transformers import glue_compute_metrics as compute_metrics
+from transformers import glue_output_modes as output_modes
+from transformers import glue_processors as processors
 
 logger = logging.getLogger(__name__)
 
@@ -233,6 +234,8 @@ def main():
                         help="If > 0: limit the data to a subset of data_subset instances.")
     parser.add_argument("--overwrite_output_dir", action='store_true',
                         help="Whether to overwrite data in output directory")
+    parser.add_argument('--overwrite_cache', action='store_true',
+                        help="Overwrite the cached training and evaluation sets")
 
     parser.add_argument("--dont_normalize_importance_by_layer", action='store_true',
                         help="Don't normalize importance score by layers")
@@ -304,10 +307,16 @@ def main():
             break
     config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
     config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
-                                          num_labels=num_labels, finetuning_task=args.task_name,
-                                          output_attentions=True)
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path)
-    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
+                                          num_labels=num_labels,
+                                          finetuning_task=args.task_name,
+                                          output_attentions=True,
+                                          cache_dir=args.cache_dir if args.cache_dir else None)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+                                                cache_dir=args.cache_dir if args.cache_dir else None)
+    model = model_class.from_pretrained(args.model_name_or_path,
+                                        from_tf=bool('.ckpt' in args.model_name_or_path),
+                                        config=config,
+                                        cache_dir=args.cache_dir if args.cache_dir else None)
 
     if args.local_rank == 0:
         torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab

examples/run_generation.py

@@ -79,13 +79,12 @@ def set_seed(args):
 def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
     """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
         Args:
-            logits: logits distribution shape (vocabulary size)
+            logits: logits distribution shape (batch size x vocabulary size)
             top_k > 0: keep only top k tokens with highest probability (top-k filtering).
             top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
                 Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
         From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
     """
-    assert logits.dim() == 1  # batch size 1 for now - could be updated for more but the code would be less clear
     top_k = min(top_k, logits.size(-1))  # Safety check
     if top_k > 0:
         # Remove all tokens with a probability less than the last token of the top-k
@@ -102,12 +101,14 @@ def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')
         sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
         sorted_indices_to_remove[..., 0] = 0
 
-        indices_to_remove = sorted_indices[sorted_indices_to_remove]
+        # scatter sorted tensors to original indexing
+        indices_to_remove = sorted_indices_to_remove.scatter(dim=1, index=sorted_indices, src=sorted_indices_to_remove)
         logits[indices_to_remove] = filter_value
     return logits
 
 
-def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0, is_xlnet=False, xlm_lang=None, device='cpu'):
+def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0,
+                    is_xlnet=False, is_xlm_mlm=False, xlm_mask_token=None, xlm_lang=None, device='cpu'):
     context = torch.tensor(context, dtype=torch.long, device=device)
     context = context.unsqueeze(0).repeat(num_samples, 1)
     generated = context
@@ -125,22 +126,29 @@ def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=
                 target_mapping[0, 0, -1] = 1.0  # predict last token
                 inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping}
 
+            if is_xlm_mlm and xlm_mask_token:
+                # XLM MLM models are direct models (predict same token, not next token)
+                # => need one additional dummy token in the input (will be masked and guessed)
+                input_ids = torch.cat((generated, torch.full((1, 1), xlm_mask_token, dtype=torch.long, device=device)), dim=1)
+                inputs = {'input_ids': input_ids}
+
             if xlm_lang is not None:
                 inputs["langs"] = torch.tensor([xlm_lang] * inputs["input_ids"].shape[1], device=device).view(1, -1)
 
-            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet (cached hidden-states)
-            next_token_logits = outputs[0][0, -1, :] / (temperature if temperature > 0 else 1.)
+            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states)
+            next_token_logits = outputs[0][:, -1, :] / (temperature if temperature > 0 else 1.)
 
-            # reptition penalty from CTRL (https://arxiv.org/abs/1909.05858)
-            for _ in set(generated):
-                next_token_logits[_] /= repetition_penalty
+            # repetition penalty from CTRL (https://arxiv.org/abs/1909.05858)
+            for i in range(num_samples):
+                for _ in set(generated[i].tolist()):
+                    next_token_logits[i, _] /= repetition_penalty
                 
             filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
-            if temperature == 0: #greedy sampling:
-                next_token = torch.argmax(filtered_logits).unsqueeze(0)
+            if temperature == 0: # greedy sampling:
+                next_token = torch.argmax(filtered_logits, dim=-1).unsqueeze(-1)
             else:
                 next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
-            generated = torch.cat((generated, next_token.unsqueeze(0)), dim=1)
+            generated = torch.cat((generated, next_token), dim=1)
     return generated
 
 
@@ -154,6 +162,7 @@ def main():
     parser.add_argument("--padding_text", type=str, default="")
     parser.add_argument("--xlm_lang", type=str, default="", help="Optional language when used with the XLM model.")
     parser.add_argument("--length", type=int, default=20)
+    parser.add_argument("--num_samples", type=int, default=1)
     parser.add_argument("--temperature", type=float, default=1.0,
                         help="temperature of 0 implies greedy sampling")
     parser.add_argument("--repetition_penalty", type=float, default=1.0,
@@ -167,10 +176,7 @@ def main():
     parser.add_argument('--stop_token', type=str, default=None,
                         help="Token at which text generation is stopped")
     args = parser.parse_args()
-    if args.model_type in ["ctrl"]:
-        if args.temperature > 0.7 : 
-            print('CTRL typically works better with lower temperatures (and lower top_k).')
-        
+
     args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
     args.n_gpu = torch.cuda.device_count()
 
@@ -190,7 +196,11 @@ def main():
     elif args.length < 0:
         args.length = MAX_LENGTH  # avoid infinite loop
 
-    print(args)
+    logger.info(args)
+    if args.model_type in ["ctrl"]:
+        if args.temperature > 0.7:
+            logger.info('CTRL typically works better with lower temperatures (and lower top_k).')
+
     while True:
         xlm_lang = None
         # XLM Language usage detailed in the issues #1414
@@ -204,29 +214,43 @@ def main():
                     language = input("Using XLM. Select language in " + str(list(tokenizer.lang2id.keys())) + " >>> ")
             xlm_lang = tokenizer.lang2id[language]
 
+        # XLM masked-language modeling (MLM) models need masked token (see details in sample_sequence)
+        is_xlm_mlm = args.model_type in ["xlm"] and 'mlm' in args.model_name_or_path
+        if is_xlm_mlm:
+            xlm_mask_token = tokenizer.mask_token_id
+        else:
+            xlm_mask_token = None
+
         raw_text = args.prompt if args.prompt else input("Model prompt >>> ")
         if args.model_type in ["transfo-xl", "xlnet"]:
             # Models with memory likes to have a long prompt for short inputs.
             raw_text = (args.padding_text if args.padding_text else PADDING_TEXT) + raw_text
-        context_tokens = tokenizer.encode(raw_text)
+        context_tokens = tokenizer.encode(raw_text, add_special_tokens=False)
+        if args.model_type == "ctrl":
+            if not any(context_tokens[0] == x for x in tokenizer.control_codes.values()):
+                logger.info("WARNING! You are not starting your generation from a control code so you won't get good results")
         out = sample_sequence(
             model=model,
             context=context_tokens,
+            num_samples=args.num_samples,
             length=args.length,
             temperature=args.temperature,
             top_k=args.top_k,
             top_p=args.top_p,
             repetition_penalty=args.repetition_penalty,
             is_xlnet=bool(args.model_type == "xlnet"),
+            is_xlm_mlm=is_xlm_mlm,
+            xlm_mask_token=xlm_mask_token,
             xlm_lang=xlm_lang,
             device=args.device,
         )
-        out = out[0, len(context_tokens):].tolist()
+        out = out[:, len(context_tokens):].tolist()
+        for o in out:
+            text = tokenizer.decode(o, clean_up_tokenization_spaces=True)
+            text = text[: text.find(args.stop_token) if args.stop_token else None]
 
-        text = tokenizer.decode(out, clean_up_tokenization_spaces=True, skip_special_tokens=True)
-        text = text[: text.find(args.stop_token) if args.stop_token else None]
+            print(text)
 
-        print(text)
         if args.prompt:
             break
     return text

examples/run_glue.py

@@ -28,7 +28,12 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                               TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-from tensorboardX import SummaryWriter
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
 from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
@@ -42,9 +47,13 @@ from transformers import (WEIGHTS_NAME, BertConfig,
                                   XLNetTokenizer,
                                   DistilBertConfig,
                                   DistilBertForSequenceClassification,
-                                  DistilBertTokenizer)
+                                  DistilBertTokenizer,
+                                  AlbertConfig,
+                                  AlbertForSequenceClassification, 
+                                  AlbertTokenizer,
+                                )
 
-from transformers import AdamW, WarmupLinearSchedule
+from transformers import AdamW, get_linear_schedule_with_warmup
 
 from transformers import glue_compute_metrics as compute_metrics
 from transformers import glue_output_modes as output_modes
@@ -61,7 +70,8 @@ MODEL_CLASSES = {
     'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer),
     'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
     'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
-    'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
+    'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer),
+    'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer)
 }
 
 
@@ -94,8 +104,9 @@ def train(args, train_dataset, model, tokenizer):
         {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
         {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
+
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
     if args.fp16:
         try:
             from apex import amp
@@ -149,13 +160,16 @@ def train(args, train_dataset, model, tokenizer):
             if args.fp16:
                 with amp.scale_loss(loss, optimizer) as scaled_loss:
                     scaled_loss.backward()
-                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
             else:
                 loss.backward()
-                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
 
             tr_loss += loss.item()
             if (step + 1) % args.gradient_accumulation_steps == 0:
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+
                 optimizer.step()
                 scheduler.step()  # Update learning rate schedule
                 model.zero_grad()
@@ -211,6 +225,10 @@ def evaluate(args, model, tokenizer, prefix=""):
         eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
         eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
 
+        # multi-gpu eval
+        if args.n_gpu > 1:
+            model = torch.nn.DataParallel(model)
+
         # Eval!
         logger.info("***** Running evaluation {} *****".format(prefix))
         logger.info("  Num examples = %d", len(eval_dataset))
@@ -305,7 +323,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
         all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
     elif output_mode == "regression":
         all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
-
+ 
     dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
     return dataset
 
@@ -349,7 +367,7 @@ def main():
     parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
                         help="Batch size per GPU/CPU for evaluation.")
     parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
-                        help="Number of updates steps to accumulate before performing a backward/update pass.")
+                        help="Number of updates steps to accumulate before performing a backward/update pass.")     
     parser.add_argument("--learning_rate", default=5e-5, type=float,
                         help="The initial learning rate for Adam.")
     parser.add_argument("--weight_decay", default=0.0, type=float,
@@ -438,9 +456,17 @@ def main():
 
     args.model_type = args.model_type.lower()
     config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
-    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
+                                          num_labels=num_labels,
+                                          finetuning_task=args.task_name,
+                                          cache_dir=args.cache_dir if args.cache_dir else None)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+                                                do_lower_case=args.do_lower_case,
+                                                cache_dir=args.cache_dir if args.cache_dir else None)
+    model = model_class.from_pretrained(args.model_name_or_path,
+                                        from_tf=bool('.ckpt' in args.model_name_or_path),
+                                        config=config,
+                                        cache_dir=args.cache_dir if args.cache_dir else None)
 
     if args.local_rank == 0:
         torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
@@ -475,7 +501,7 @@ def main():
 
         # Load a trained model and vocabulary that you have fine-tuned
         model = model_class.from_pretrained(args.output_dir)
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
         model.to(args.device)
 
 

examples/run_lm_finetuning.py

@@ -34,10 +34,15 @@ import numpy as np
 import torch
 from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler
 from torch.utils.data.distributed import DistributedSampler
-from tensorboardX import SummaryWriter
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
 from tqdm import tqdm, trange
 
-from transformers import (WEIGHTS_NAME, AdamW, WarmupLinearSchedule,
+from transformers import (WEIGHTS_NAME, AdamW, get_linear_schedule_with_warmup,
                                   BertConfig, BertForMaskedLM, BertTokenizer,
                                   GPT2Config, GPT2LMHeadModel, GPT2Tokenizer,
                                   OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer,
@@ -58,12 +63,12 @@ MODEL_CLASSES = {
 
 
 class TextDataset(Dataset):
-    def __init__(self, tokenizer, file_path='train', block_size=512):
+    def __init__(self, tokenizer, args, file_path='train', block_size=512):
         assert os.path.isfile(file_path)
         directory, filename = os.path.split(file_path)
-        cached_features_file = os.path.join(directory, 'cached_lm_' + block_size + '_' + filename)
+        cached_features_file = os.path.join(directory, args.model_name_or_path + '_cached_lm_' + str(block_size) + '_' + filename)
 
-        if os.path.exists(cached_features_file):
+        if os.path.exists(cached_features_file) and not args.overwrite_cache:
             logger.info("Loading features from cached file %s", cached_features_file)
             with open(cached_features_file, 'rb') as handle:
                 self.examples = pickle.load(handle)
@@ -94,7 +99,7 @@ class TextDataset(Dataset):
 
 
 def load_and_cache_examples(args, tokenizer, evaluate=False):
-    dataset = TextDataset(tokenizer, file_path=args.eval_data_file if evaluate else args.train_data_file, block_size=args.block_size)
+    dataset = TextDataset(tokenizer, args, file_path=args.eval_data_file if evaluate else args.train_data_file, block_size=args.block_size)
     return dataset
 
 
@@ -180,7 +185,7 @@ def train(args, train_dataset, model, tokenizer):
         {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
     if args.fp16:
         try:
             from apex import amp
@@ -210,6 +215,7 @@ def train(args, train_dataset, model, tokenizer):
 
     global_step = 0
     tr_loss, logging_loss = 0.0, 0.0
+    model.resize_token_embeddings(len(tokenizer))
     model.zero_grad()
     train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
     set_seed(args)  # Added here for reproducibility (even between python 2 and 3)
@@ -295,6 +301,10 @@ def evaluate(args, model, tokenizer, prefix=""):
     eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
     eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
 
+    # multi-gpu evaluate
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
     # Eval!
     logger.info("***** Running evaluation {} *****".format(prefix))
     logger.info("  Num examples = %d", len(eval_dataset))
@@ -304,10 +314,12 @@ def evaluate(args, model, tokenizer, prefix=""):
     model.eval()
 
     for batch in tqdm(eval_dataloader, desc="Evaluating"):
-        batch = batch.to(args.device)
+        inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
+        inputs = inputs.to(args.device)
+        labels = labels.to(args.device)
 
         with torch.no_grad():
-            outputs = model(batch, masked_lm_labels=batch) if args.mlm else model(batch, labels=batch)
+            outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
             lm_loss = outputs[0]
             eval_loss += lm_loss.mean().item()
         nb_eval_steps += 1
@@ -464,12 +476,18 @@ def main():
         torch.distributed.barrier()  # Barrier to make sure only the first process in distributed training download model & vocab
 
     config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
+                                          cache_dir=args.cache_dir if args.cache_dir else None)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+                                                do_lower_case=args.do_lower_case,
+                                                cache_dir=args.cache_dir if args.cache_dir else None)
     if args.block_size <= 0:
         args.block_size = tokenizer.max_len_single_sentence  # Our input block size will be the max possible for the model
     args.block_size = min(args.block_size, tokenizer.max_len_single_sentence)
-    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
+    model = model_class.from_pretrained(args.model_name_or_path,
+                                        from_tf=bool('.ckpt' in args.model_name_or_path),
+                                        config=config,
+                                        cache_dir=args.cache_dir if args.cache_dir else None)
     model.to(args.device)
 
     if args.local_rank == 0:

examples/run_multiple_choice.py

@@ -29,7 +29,12 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                               TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-from tensorboardX import SummaryWriter
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
 from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
@@ -38,7 +43,7 @@ from transformers import (WEIGHTS_NAME, BertConfig,
                                   XLNetTokenizer, RobertaConfig,
                                   RobertaForMultipleChoice, RobertaTokenizer)
 
-from transformers import AdamW, WarmupLinearSchedule
+from transformers import AdamW, get_linear_schedule_with_warmup
 
 from utils_multiple_choice import (convert_examples_to_features, processors)
 
@@ -96,7 +101,7 @@ def train(args, train_dataset, model, tokenizer):
         {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
     if args.fp16:
         try:
             from apex import amp
@@ -224,6 +229,10 @@ def evaluate(args, model, tokenizer, prefix="", test=False):
         eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
         eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
 
+        # multi-gpu evaluate
+        if args.n_gpu > 1:
+            model = torch.nn.DataParallel(model)
+
         # Eval!
         logger.info("***** Running evaluation {} *****".format(prefix))
         logger.info("  Num examples = %d", len(eval_dataset))
@@ -459,9 +468,17 @@ def main():
 
     args.model_type = args.model_type.lower()
     config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
-    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
+                                          num_labels=num_labels,
+                                          finetuning_task=args.task_name,
+                                          cache_dir=args.cache_dir if args.cache_dir else None)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+                                                do_lower_case=args.do_lower_case,
+                                                cache_dir=args.cache_dir if args.cache_dir else None)
+    model = model_class.from_pretrained(args.model_name_or_path,
+                                        from_tf=bool('.ckpt' in args.model_name_or_path),
+                                        config=config,
+                                        cache_dir=args.cache_dir if args.cache_dir else None)
 
     if args.local_rank == 0:
         torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab

examples/run_ner.py

@@ -0,0 +1,532 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Fine-tuning the library models for named entity recognition on CoNLL-2003 (Bert or Roberta). """
+
+from __future__ import absolute_import, division, print_function
+
+import argparse
+import glob
+import logging
+import os
+import random
+
+import numpy as np
+import torch
+from seqeval.metrics import precision_score, recall_score, f1_score
+from tensorboardX import SummaryWriter
+from torch.nn import CrossEntropyLoss
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm, trange
+from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file
+
+from transformers import AdamW, get_linear_schedule_with_warmup
+from transformers import WEIGHTS_NAME, BertConfig, BertForTokenClassification, BertTokenizer
+from transformers import RobertaConfig, RobertaForTokenClassification, RobertaTokenizer
+from transformers import DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer
+from transformers import CamembertConfig, CamembertForTokenClassification, CamembertTokenizer
+
+logger = logging.getLogger(__name__)
+
+ALL_MODELS = sum(
+    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)),
+    ())
+
+MODEL_CLASSES = {
+    "bert": (BertConfig, BertForTokenClassification, BertTokenizer),
+    "roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer),
+    "distilbert": (DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer),
+    "camembert": (CamembertConfig, CamembertForTokenClassification, CamembertTokenizer),
+}
+
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+
+def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
+    """ Train the model """
+    if args.local_rank in [-1, 0]:
+        tb_writer = SummaryWriter()
+
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+    else:
+        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+
+    # Prepare optimizer and schedule (linear warmup and decay)
+    no_decay = ["bias", "LayerNorm.weight"]
+    optimizer_grouped_parameters = [
+        {"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
+         "weight_decay": args.weight_decay},
+        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0}
+    ]
+    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
+                                                          output_device=args.local_rank,
+                                                          find_unused_parameters=True)
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
+                args.train_batch_size * args.gradient_accumulation_steps * (
+                    torch.distributed.get_world_size() if args.local_rank != -1 else 1))
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+
+    global_step = 0
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+            model.train()
+            batch = tuple(t.to(args.device) for t in batch)
+            inputs = {"input_ids": batch[0],
+                      "attention_mask": batch[1],
+                      "labels": batch[3]}
+            if args.model_type != "distilbert":
+                inputs["token_type_ids"] = batch[2] if args.model_type in ["bert", "xlnet"] else None  # XLM and RoBERTa don"t use segment_ids
+
+            outputs = model(**inputs)
+            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)
+
+            if args.n_gpu > 1:
+                loss = loss.mean()  # mean() to average on multi-gpu parallel training
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+
+            if args.fp16:
+                with amp.scale_loss(loss, optimizer) as scaled_loss:
+                    scaled_loss.backward()
+            else:
+                loss.backward()
+
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+
+                scheduler.step()  # Update learning rate schedule
+                optimizer.step()
+                model.zero_grad()
+                global_step += 1
+
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    # Log metrics
+                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
+                        results, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev")
+                        for key, value in results.items():
+                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
+                    tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
+                    logging_loss = tr_loss
+
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    # Save model checkpoint
+                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
+                    model_to_save = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
+                    model_to_save.save_pretrained(output_dir)
+                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+
+    return global_step, tr_loss / global_step
+
+
+def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""):
+    eval_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode=mode)
+
+    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+    # Note that DistributedSampler samples randomly
+    eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
+    eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+    # multi-gpu evaluate
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Eval!
+    logger.info("***** Running evaluation %s *****", prefix)
+    logger.info("  Num examples = %d", len(eval_dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+    eval_loss = 0.0
+    nb_eval_steps = 0
+    preds = None
+    out_label_ids = None
+    model.eval()
+    for batch in tqdm(eval_dataloader, desc="Evaluating"):
+        batch = tuple(t.to(args.device) for t in batch)
+
+        with torch.no_grad():
+            inputs = {"input_ids": batch[0],
+                      "attention_mask": batch[1],
+                      "labels": batch[3]}
+            if args.model_type != "distilbert":
+                inputs["token_type_ids"] = batch[2] if args.model_type in ["bert", "xlnet"] else None  # XLM and RoBERTa don"t use segment_ids
+            outputs = model(**inputs)
+            tmp_eval_loss, logits = outputs[:2]
+
+            if args.n_gpu > 1:
+                tmp_eval_loss = tmp_eval_loss.mean()  # mean() to average on multi-gpu parallel evaluating
+
+            eval_loss += tmp_eval_loss.item()
+        nb_eval_steps += 1
+        if preds is None:
+            preds = logits.detach().cpu().numpy()
+            out_label_ids = inputs["labels"].detach().cpu().numpy()
+        else:
+            preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
+            out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
+
+    eval_loss = eval_loss / nb_eval_steps
+    preds = np.argmax(preds, axis=2)
+
+    label_map = {i: label for i, label in enumerate(labels)}
+
+    out_label_list = [[] for _ in range(out_label_ids.shape[0])]
+    preds_list = [[] for _ in range(out_label_ids.shape[0])]
+
+    for i in range(out_label_ids.shape[0]):
+        for j in range(out_label_ids.shape[1]):
+            if out_label_ids[i, j] != pad_token_label_id:
+                out_label_list[i].append(label_map[out_label_ids[i][j]])
+                preds_list[i].append(label_map[preds[i][j]])
+
+    results = {
+        "loss": eval_loss,
+        "precision": precision_score(out_label_list, preds_list),
+        "recall": recall_score(out_label_list, preds_list),
+        "f1": f1_score(out_label_list, preds_list)
+    }
+
+    logger.info("***** Eval results %s *****", prefix)
+    for key in sorted(results.keys()):
+        logger.info("  %s = %s", key, str(results[key]))
+
+    return results, preds_list
+
+
+def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode):
+    if args.local_rank not in [-1, 0] and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Load data features from cache or dataset file
+    cached_features_file = os.path.join(args.data_dir, "cached_{}_{}_{}".format(mode,
+        list(filter(None, args.model_name_or_path.split("/"))).pop(),
+        str(args.max_seq_length)))
+    if os.path.exists(cached_features_file) and not args.overwrite_cache:
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features = torch.load(cached_features_file)
+    else:
+        logger.info("Creating features from dataset file at %s", args.data_dir)
+        examples = read_examples_from_file(args.data_dir, mode)
+        features = convert_examples_to_features(examples, labels, args.max_seq_length, tokenizer,
+                                                cls_token_at_end=bool(args.model_type in ["xlnet"]),
+                                                # xlnet has a cls token at the end
+                                                cls_token=tokenizer.cls_token,
+                                                cls_token_segment_id=2 if args.model_type in ["xlnet"] else 0,
+                                                sep_token=tokenizer.sep_token,
+                                                sep_token_extra=bool(args.model_type in ["roberta"]),
+                                                # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
+                                                pad_on_left=bool(args.model_type in ["xlnet"]),
+                                                # pad on the left for xlnet
+                                                pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
+                                                pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
+                                                pad_token_label_id=pad_token_label_id
+                                                )
+        if args.local_rank in [-1, 0]:
+            logger.info("Saving features into cached file %s", cached_features_file)
+            torch.save(features, cached_features_file)
+
+    if args.local_rank == 0 and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Convert to Tensors and build dataset
+    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
+    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
+    all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
+    all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long)
+
+    dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
+    return dataset
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    ## Required parameters
+    parser.add_argument("--data_dir", default=None, type=str, required=True,
+                        help="The input data dir. Should contain the training files for the CoNLL-2003 NER task.")
+    parser.add_argument("--model_type", default=None, type=str, required=True,
+                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
+                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+    parser.add_argument("--output_dir", default=None, type=str, required=True,
+                        help="The output directory where the model predictions and checkpoints will be written.")
+
+    ## Other parameters
+    parser.add_argument("--labels", default="", type=str,
+                        help="Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.")
+    parser.add_argument("--config_name", default="", type=str,
+                        help="Pretrained config name or path if not the same as model_name")
+    parser.add_argument("--tokenizer_name", default="", type=str,
+                        help="Pretrained tokenizer name or path if not the same as model_name")
+    parser.add_argument("--cache_dir", default="", type=str,
+                        help="Where do you want to store the pre-trained models downloaded from s3")
+    parser.add_argument("--max_seq_length", default=128, type=int,
+                        help="The maximum total input sequence length after tokenization. Sequences longer "
+                             "than this will be truncated, sequences shorter will be padded.")
+    parser.add_argument("--do_train", action="store_true",
+                        help="Whether to run training.")
+    parser.add_argument("--do_eval", action="store_true",
+                        help="Whether to run eval on the dev set.")
+    parser.add_argument("--do_predict", action="store_true",
+                        help="Whether to run predictions on the test set.")
+    parser.add_argument("--evaluate_during_training", action="store_true",
+                        help="Whether to run evaluation during training at each logging step.")
+    parser.add_argument("--do_lower_case", action="store_true",
+                        help="Set this flag if you are using an uncased model.")
+
+    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
+                        help="Batch size per GPU/CPU for training.")
+    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
+                        help="Batch size per GPU/CPU for evaluation.")
+    parser.add_argument("--gradient_accumulation_steps", type=int, default=1,
+                        help="Number of updates steps to accumulate before performing a backward/update pass.")
+    parser.add_argument("--learning_rate", default=5e-5, type=float,
+                        help="The initial learning rate for Adam.")
+    parser.add_argument("--weight_decay", default=0.0, type=float,
+                        help="Weight decay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float,
+                        help="Max gradient norm.")
+    parser.add_argument("--num_train_epochs", default=3.0, type=float,
+                        help="Total number of training epochs to perform.")
+    parser.add_argument("--max_steps", default=-1, type=int,
+                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
+    parser.add_argument("--warmup_steps", default=0, type=int,
+                        help="Linear warmup over warmup_steps.")
+
+    parser.add_argument("--logging_steps", type=int, default=50,
+                        help="Log every X updates steps.")
+    parser.add_argument("--save_steps", type=int, default=50,
+                        help="Save checkpoint every X updates steps.")
+    parser.add_argument("--eval_all_checkpoints", action="store_true",
+                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
+    parser.add_argument("--no_cuda", action="store_true",
+                        help="Avoid using CUDA when available")
+    parser.add_argument("--overwrite_output_dir", action="store_true",
+                        help="Overwrite the content of the output directory")
+    parser.add_argument("--overwrite_cache", action="store_true",
+                        help="Overwrite the cached training and evaluation sets")
+    parser.add_argument("--seed", type=int, default=42,
+                        help="random seed for initialization")
+
+    parser.add_argument("--fp16", action="store_true",
+                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
+    parser.add_argument("--fp16_opt_level", type=str, default="O1",
+                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+                             "See details at https://nvidia.github.io/apex/amp.html")
+    parser.add_argument("--local_rank", type=int, default=-1,
+                        help="For distributed training: local_rank")
+    parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
+    parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
+    args = parser.parse_args()
+
+    if os.path.exists(args.output_dir) and os.listdir(
+            args.output_dir) and args.do_train and not args.overwrite_output_dir:
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
+                args.output_dir))
+
+    # Setup distant debugging if needed
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    # Setup CUDA, GPU & distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        args.n_gpu = torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend="nccl")
+        args.n_gpu = 1
+    args.device = device
+
+    # Setup logging
+    logging.basicConfig(format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+                        datefmt="%m/%d/%Y %H:%M:%S",
+                        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
+    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+                   args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
+
+    # Set seed
+    set_seed(args)
+
+    # Prepare CONLL-2003 task
+    labels = get_labels(args.labels)
+    num_labels = len(labels)
+    # Use cross entropy ignore index as padding label id so that only real label ids contribute to the loss later
+    pad_token_label_id = CrossEntropyLoss().ignore_index
+
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    args.model_type = args.model_type.lower()
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
+                                          num_labels=num_labels,
+                                          cache_dir=args.cache_dir if args.cache_dir else None)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+                                                do_lower_case=args.do_lower_case,
+                                                cache_dir=args.cache_dir if args.cache_dir else None)
+    model = model_class.from_pretrained(args.model_name_or_path,
+                                        from_tf=bool(".ckpt" in args.model_name_or_path),
+                                        config=config,
+                                        cache_dir=args.cache_dir if args.cache_dir else None)
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    model.to(args.device)
+
+    logger.info("Training/evaluation parameters %s", args)
+
+    # Training
+    if args.do_train:
+        train_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode="train")
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer, labels, pad_token_label_id)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+
+    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+
+        # Good practice: save your training arguments together with the trained model
+        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
+
+    # Evaluation
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        checkpoints = [args.output_dir]
+        if args.eval_all_checkpoints:
+            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True)))
+            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+        for checkpoint in checkpoints:
+            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
+            model = model_class.from_pretrained(checkpoint)
+            model.to(args.device)
+            result, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev", prefix=global_step)
+            if global_step:
+                result = {"{}_{}".format(global_step, k): v for k, v in result.items()}
+            results.update(result)
+        output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
+        with open(output_eval_file, "w") as writer:
+            for key in sorted(results.keys()):
+                writer.write("{} = {}\n".format(key, str(results[key])))
+
+    if args.do_predict and args.local_rank in [-1, 0]:
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model = model_class.from_pretrained(args.output_dir)
+        model.to(args.device)
+        result, predictions = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="test")
+        # Save results
+        output_test_results_file = os.path.join(args.output_dir, "test_results.txt")
+        with open(output_test_results_file, "w") as writer:
+            for key in sorted(result.keys()):
+                writer.write("{} = {}\n".format(key, str(result[key])))
+        # Save predictions
+        output_test_predictions_file = os.path.join(args.output_dir, "test_predictions.txt")
+        with open(output_test_predictions_file, "w") as writer:
+            with open(os.path.join(args.data_dir, "test.txt"), "r") as f:
+                example_id = 0
+                for line in f:
+                    if line.startswith("-DOCSTART-") or line == "" or line == "\n":
+                        writer.write(line)
+                        if not predictions[example_id]:
+                            example_id += 1
+                    elif predictions[example_id]:
+                        output_line = line.split()[0] + " " + predictions[example_id].pop(0) + "\n"
+                        writer.write(output_line)
+                    else:
+                        logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
+
+    return results
+
+
+if __name__ == "__main__":
+    main()
+

examples/run_squad.py

@@ -22,15 +22,20 @@ import logging
 import os
 import random
 import glob
+import timeit
 
 import numpy as np
 import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                               TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm, trange
 
-from tensorboardX import SummaryWriter
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
+from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
                                   BertForQuestionAnswering, BertTokenizer,
@@ -38,9 +43,10 @@ from transformers import (WEIGHTS_NAME, BertConfig,
                                   XLMTokenizer, XLNetConfig,
                                   XLNetForQuestionAnswering,
                                   XLNetTokenizer,
-                                  DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
+                                  DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer,
+                                  AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer)
 
-from transformers import AdamW, WarmupLinearSchedule
+from transformers import AdamW, get_linear_schedule_with_warmup
 
 from utils_squad import (read_squad_examples, convert_examples_to_features,
                          RawResult, write_predictions,
@@ -60,7 +66,8 @@ MODEL_CLASSES = {
     'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
     'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
     'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
-    'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
+    'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
+    'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer)
 }
 
 def set_seed(args):
@@ -95,7 +102,7 @@ def train(args, train_dataset, model, tokenizer):
         {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
     if args.fp16:
         try:
             from apex import amp
@@ -123,7 +130,7 @@ def train(args, train_dataset, model, tokenizer):
     logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
     logger.info("  Total optimization steps = %d", t_total)
 
-    global_step = 0
+    global_step = 1
     tr_loss, logging_loss = 0.0, 0.0
     model.zero_grad()
     train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
@@ -134,8 +141,8 @@ def train(args, train_dataset, model, tokenizer):
             model.train()
             batch = tuple(t.to(args.device) for t in batch)
             inputs = {'input_ids':       batch[0],
-                      'attention_mask':  batch[1], 
-                      'start_positions': batch[3], 
+                      'attention_mask':  batch[1],
+                      'start_positions': batch[3],
                       'end_positions':   batch[4]}
             if args.model_type != 'distilbert':
                 inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
@@ -153,13 +160,16 @@ def train(args, train_dataset, model, tokenizer):
             if args.fp16:
                 with amp.scale_loss(loss, optimizer) as scaled_loss:
                     scaled_loss.backward()
-                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
             else:
                 loss.backward()
-                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
 
             tr_loss += loss.item()
             if (step + 1) % args.gradient_accumulation_steps == 0:
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+
                 optimizer.step()
                 scheduler.step()  # Update learning rate schedule
                 model.zero_grad()
@@ -209,11 +219,16 @@ def evaluate(args, model, tokenizer, prefix=""):
     eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
     eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
 
+    # multi-gpu evaluate
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
     # Eval!
     logger.info("***** Running evaluation {} *****".format(prefix))
     logger.info("  Num examples = %d", len(dataset))
     logger.info("  Batch size = %d", args.eval_batch_size)
     all_results = []
+    start_time = timeit.default_timer()
     for batch in tqdm(eval_dataloader, desc="Evaluating"):
         model.eval()
         batch = tuple(t.to(args.device) for t in batch)
@@ -246,6 +261,9 @@ def evaluate(args, model, tokenizer, prefix=""):
                                    end_logits   = to_list(outputs[1][i]))
             all_results.append(result)
 
+    evalTime = timeit.default_timer() - start_time
+    logger.info("  Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))
+
     # Compute predictions
     output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
     output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
@@ -298,7 +316,11 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
                                                 max_seq_length=args.max_seq_length,
                                                 doc_stride=args.doc_stride,
                                                 max_query_length=args.max_query_length,
-                                                is_training=not evaluate)
+                                                is_training=not evaluate,
+                                                cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
+                                                pad_token_segment_id=3 if args.model_type in ['xlnet'] else 0,
+                                                cls_token_at_end=True if args.model_type in ['xlnet'] else False,
+                                                sequence_a_is_doc=True if args.model_type in ['xlnet'] else False)
         if args.local_rank in [-1, 0]:
             logger.info("Saving features into cached file %s", cached_features_file)
             torch.save(features, cached_features_file)
@@ -382,7 +404,7 @@ def main():
     parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
                         help="Number of updates steps to accumulate before performing a backward/update pass.")
     parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight deay if we apply some.")
+                        help="Weight decay if we apply some.")
     parser.add_argument("--adam_epsilon", default=1e-8, type=float,
                         help="Epsilon for Adam optimizer.")
     parser.add_argument("--max_grad_norm", default=1.0, type=float,
@@ -466,9 +488,15 @@ def main():
 
     args.model_type = args.model_type.lower()
     config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
-    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
+                                          cache_dir=args.cache_dir if args.cache_dir else None)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+                                                do_lower_case=args.do_lower_case,
+                                                cache_dir=args.cache_dir if args.cache_dir else None)
+    model = model_class.from_pretrained(args.model_name_or_path,
+                                        from_tf=bool('.ckpt' in args.model_name_or_path),
+                                        config=config,
+                                        cache_dir=args.cache_dir if args.cache_dir else None)
 
     if args.local_rank == 0:
         torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
@@ -477,6 +505,16 @@ def main():
 
     logger.info("Training/evaluation parameters %s", args)
 
+    # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
+    # Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
+    # remove the need for this code, but it is still valid.
+    if args.fp16:
+        try:
+            import apex
+            apex.amp.register_half_function(torch, 'einsum')
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+
     # Training
     if args.do_train:
         train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
@@ -501,7 +539,7 @@ def main():
         torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
 
         # Load a trained model and vocabulary that you have fine-tuned
-        model = model_class.from_pretrained(args.output_dir)
+        model = model_class.from_pretrained(args.output_dir, force_download=True)
         tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
         model.to(args.device)
 
@@ -519,7 +557,7 @@ def main():
         for checkpoint in checkpoints:
             # Reload the model
             global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
-            model = model_class.from_pretrained(checkpoint)
+            model = model_class.from_pretrained(checkpoint, force_download=True)
             model.to(args.device)
 
             # Evaluate

examples/run_summarization_finetuning.py

@@ -0,0 +1,492 @@
+# coding=utf-8
+# Copyright 2019 The HuggingFace Inc. team.
+# Copyright (c) 2019 The HuggingFace Inc.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Finetuning seq2seq models for sequence generation."""
+
+import argparse
+import functools
+import logging
+import os
+import random
+import sys
+
+import numpy as np
+from tqdm import tqdm, trange
+import torch
+from torch.optim import Adam
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
+
+from transformers import (
+    AutoTokenizer,
+    BertForMaskedLM,
+    BertConfig,
+    PreTrainedEncoderDecoder,
+    Model2Model,
+)
+
+from utils_summarization import (
+    CNNDailyMailDataset,
+    encode_for_summarization,
+    fit_to_block_size,
+    build_lm_labels,
+    build_mask,
+    compute_token_type_ids,
+)
+
+logger = logging.getLogger(__name__)
+logging.basicConfig(stream=sys.stdout, level=logging.INFO)
+
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+
+
+# ------------
+# Load dataset
+# ------------
+
+
+def load_and_cache_examples(args, tokenizer):
+    dataset = CNNDailyMailDataset(tokenizer, data_dir=args.data_dir)
+    return dataset
+
+
+def collate(data, tokenizer, block_size):
+    """ List of tuple as an input. """
+    # remove the files with empty an story/summary, encode and fit to block
+    data = filter(lambda x: not (len(x[0]) == 0 or len(x[1]) == 0), data)
+    data = [
+        encode_for_summarization(story, summary, tokenizer) for story, summary in data
+    ]
+    data = [
+        (
+            fit_to_block_size(story, block_size, tokenizer.pad_token_id),
+            fit_to_block_size(summary, block_size, tokenizer.pad_token_id),
+        )
+        for story, summary in data
+    ]
+
+    stories = torch.tensor([story for story, summary in data])
+    summaries = torch.tensor([summary for story, summary in data])
+    encoder_token_type_ids = compute_token_type_ids(stories, tokenizer.cls_token_id)
+    encoder_mask = build_mask(stories, tokenizer.pad_token_id)
+    decoder_mask = build_mask(summaries, tokenizer.pad_token_id)
+    lm_labels = build_lm_labels(summaries, tokenizer.pad_token_id)
+
+    return (
+        stories,
+        summaries,
+        encoder_token_type_ids,
+        encoder_mask,
+        decoder_mask,
+        lm_labels,
+    )
+
+
+# ----------
+# Optimizers
+# ----------
+
+
+class BertSumOptimizer(object):
+    """ Specific optimizer for BertSum.
+
+    As described in [1], the authors fine-tune BertSum for abstractive
+    summarization using two Adam Optimizers with different warm-up steps and
+    learning rate. They also use a custom learning rate scheduler.
+
+    [1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders."
+        arXiv preprint arXiv:1908.08345 (2019).
+    """
+
+    def __init__(self, model, lr, warmup_steps, beta_1=0.99, beta_2=0.999, eps=1e-8):
+        self.encoder = model.encoder
+        self.decoder = model.decoder
+        self.lr = lr
+        self.warmup_steps = warmup_steps
+
+        self.optimizers = {
+            "encoder": Adam(
+                model.encoder.parameters(),
+                lr=lr["encoder"],
+                betas=(beta_1, beta_2),
+                eps=eps,
+            ),
+            "decoder": Adam(
+                model.decoder.parameters(),
+                lr=lr["decoder"],
+                betas=(beta_1, beta_2),
+                eps=eps,
+            ),
+        }
+
+        self._step = 0
+
+    def _update_rate(self, stack):
+        return self.lr[stack] * min(
+            self._step ** (-0.5), self._step * self.warmup_steps[stack] ** (-0.5)
+        )
+
+    def zero_grad(self):
+        self.optimizer_decoder.zero_grad()
+        self.optimizer_encoder.zero_grad()
+
+    def step(self):
+        self._step += 1
+        for stack, optimizer in self.optimizers.items():
+            new_rate = self._update_rate(stack)
+            for param_group in optimizer.param_groups:
+                param_group["lr"] = new_rate
+            optimizer.step()
+
+
+# ------------
+# Train
+# ------------
+
+
+def train(args, model, tokenizer):
+    """ Fine-tune the pretrained model on the corpus. """
+    set_seed(args)
+
+    # Load the data
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_dataset = load_and_cache_examples(args, tokenizer)
+    train_sampler = RandomSampler(train_dataset)
+    model_collate_fn = functools.partial(collate, tokenizer=tokenizer, block_size=512)
+    train_dataloader = DataLoader(
+        train_dataset,
+        sampler=train_sampler,
+        batch_size=args.train_batch_size,
+        collate_fn=model_collate_fn,
+    )
+
+    # Training schedule
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = t_total // (
+            len(train_dataloader) // args.gradient_accumulation_steps + 1
+        )
+    else:
+        t_total = (
+            len(train_dataloader)
+            // args.gradient_accumulation_steps
+            * args.num_train_epochs
+        )
+
+    # Prepare the optimizer
+    lr = {"encoder": 0.002, "decoder": 0.2}
+    warmup_steps = {"encoder": 20000, "decoder": 10000}
+    optimizer = BertSumOptimizer(model, lr, warmup_steps)
+
+    # Train
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info(
+        "  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size
+    )
+    logger.info(
+        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
+        args.train_batch_size * args.gradient_accumulation_steps
+        # * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
+    )
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+
+    model.zero_grad()
+    train_iterator = trange(args.num_train_epochs, desc="Epoch", disable=True)
+
+    global_step = 0
+    tr_loss = 0.0
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=True)
+        for step, batch in enumerate(epoch_iterator):
+            source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch
+
+            source = source.to(args.device)
+            target = target.to(args.device)
+            encoder_token_type_ids = encoder_token_type_ids.to(args.device)
+            encoder_mask = encoder_mask.to(args.device)
+            decoder_mask = decoder_mask.to(args.device)
+            lm_labels = lm_labels.to(args.device)
+
+            model.train()
+            outputs = model(
+                source,
+                target,
+                encoder_token_type_ids=encoder_token_type_ids,
+                encoder_attention_mask=encoder_mask,
+                decoder_attention_mask=decoder_mask,
+                decoder_lm_labels=lm_labels,
+            )
+
+            loss = outputs[0]
+            print(loss)
+            if args.gradient_accumulation_steps > 1:
+                loss /= args.gradient_accumulation_steps
+
+            loss.backward()
+
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+                optimizer.step()
+                model.zero_grad()
+                global_step += 1
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    return global_step, tr_loss / global_step
+
+
+# ------------
+# Train
+# ------------
+
+
+def evaluate(args, model, tokenizer, prefix=""):
+    set_seed(args)
+
+    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+    eval_dataset = load_and_cache_examples(args, tokenizer, evaluate=True)
+    eval_sampler = SequentialSampler(eval_dataset)
+    eval_dataloader = DataLoader(
+        eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size
+    )
+
+    # multi-gpu evaluate
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+        
+    logger.info("***** Running evaluation {} *****".format(prefix))
+    logger.info("  Num examples = %d", len(eval_dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+    eval_loss = 0.0
+    nb_eval_steps = 0
+    model.eval()
+
+    for batch in tqdm(eval_dataloader, desc="Evaluating"):
+        source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch
+
+        source = source.to(args.device)
+        target = target.to(args.device)
+        encoder_token_type_ids = encoder_token_type_ids.to(args.device)
+        encoder_mask = encoder_mask.to(args.device)
+        decoder_mask = decoder_mask.to(args.device)
+        lm_labels = lm_labels.to(args.device)
+
+        with torch.no_grad():
+            outputs = model(
+                source,
+                target,
+                encoder_token_type_ids=encoder_token_type_ids,
+                encoder_attention_mask=encoder_mask,
+                decoder_attention_mask=decoder_mask,
+                decoder_lm_labels=lm_labels,
+            )
+            lm_loss = outputs[0]
+            eval_loss += lm_loss.mean().item()
+        nb_eval_steps += 1
+
+    eval_loss = eval_loss / nb_eval_steps
+    perplexity = torch.exp(torch.tensor(eval_loss))
+
+    result = {"perplexity": perplexity}
+
+    # Save the evaluation's results
+    output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
+    if not os.path.exists(args.output_dir):
+        os.makedirs(args.output_dir)
+
+    with open(output_eval_file, "w") as writer:
+        logger.info("***** Eval results {} *****".format(prefix))
+        for key in sorted(result.keys()):
+            logger.info("  %s = %s", key, str(result[key]))
+            writer.write("%s = %s\n" % (key, str(result[key])))
+
+    return result
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    # Required parameters
+    parser.add_argument(
+        "--data_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The input training data file (a text file).",
+    )
+    parser.add_argument(
+        "--output_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+
+    # Optional parameters
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--do_evaluate",
+        type=bool,
+        default=False,
+        help="Run model evaluation on out-of-sample data.",
+    )
+    parser.add_argument("--do_train", type=bool, default=False, help="Run training.")
+    parser.add_argument(
+        "--do_overwrite_output_dir",
+        type=bool,
+        default=False,
+        help="Whether to overwrite the output dir.",
+    )
+    parser.add_argument(
+        "--model_name_or_path",
+        default="bert-base-cased",
+        type=str,
+        help="The model checkpoint to initialize the encoder and decoder's weights with.",
+    )
+    parser.add_argument(
+        "--model_type",
+        default="bert",
+        type=str,
+        help="The decoder architecture to be fine-tuned.",
+    )
+    parser.add_argument(
+        "--max_grad_norm", default=1.0, type=float, help="Max gradient norm."
+    )
+    parser.add_argument(
+        "--max_steps",
+        default=-1,
+        type=int,
+        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
+    )
+    parser.add_argument(
+        "--to_cpu", default=False, type=bool, help="Whether to force training on CPU."
+    )
+    parser.add_argument(
+        "--num_train_epochs",
+        default=10,
+        type=int,
+        help="Total number of training epochs to perform.",
+    )
+    parser.add_argument(
+        "--per_gpu_train_batch_size",
+        default=4,
+        type=int,
+        help="Batch size per GPU/CPU for training.",
+    )
+    parser.add_argument("--seed", default=42, type=int)
+    args = parser.parse_args()
+
+    if (
+        os.path.exists(args.output_dir)
+        and os.listdir(args.output_dir)
+        and args.do_train
+        and not args.do_overwrite_output_dir
+    ):
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --do_overwrite_output_dir to overwrite.".format(
+                args.output_dir
+            )
+        )
+
+    # Set up training device
+    if args.to_cpu or not torch.cuda.is_available():
+        args.device = torch.device("cpu")
+        args.n_gpu = 0
+    else:
+        args.device = torch.device("cuda")
+        args.n_gpu = torch.cuda.device_count()
+
+    # Load pretrained model and tokenizer. The decoder's weights are randomly initialized.
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
+    config = BertConfig.from_pretrained(args.model_name_or_path)
+    decoder_model = BertForMaskedLM(config)
+    model = Model2Model.from_pretrained(
+        args.model_name_or_path, decoder_model=decoder_model
+    )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.warning(
+        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+        0,
+        args.device,
+        args.n_gpu,
+        False,
+        False,
+    )
+
+    logger.info("Training/evaluation parameters %s", args)
+
+    # Train the model
+    model.to(args.device)
+    if args.do_train:
+        global_step, tr_loss = train(args, model, tokenizer)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+
+        if not os.path.exists(args.output_dir):
+            os.makedirs(args.output_dir)
+
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = (
+            model.module if hasattr(model, "module") else model
+        )  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+        torch.save(args, os.path.join(args.output_dir, "training_arguments.bin"))
+
+    # Evaluate the model
+    results = {}
+    if args.do_evaluate:
+        checkpoints = []
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+        for checkpoint in checkpoints:
+            encoder_checkpoint = os.path.join(checkpoint, "encoder")
+            decoder_checkpoint = os.path.join(checkpoint, "decoder")
+            model = PreTrainedEncoderDecoder.from_pretrained(
+                encoder_checkpoint, decoder_checkpoint
+            )
+            model.to(args.device)
+            results = "placeholder"
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

examples/run_tf_glue.py

@@ -1,40 +1,93 @@
+import os
 import tensorflow as tf
 import tensorflow_datasets
-from transformers import BertTokenizer, TFBertForSequenceClassification, glue_convert_examples_to_features, BertForSequenceClassification
+from transformers import BertTokenizer, TFBertForSequenceClassification, BertConfig, glue_convert_examples_to_features, BertForSequenceClassification, glue_processors
 
-# Load dataset, tokenizer, model from pretrained model/vocabulary
+# script parameters
+BATCH_SIZE = 32
+EVAL_BATCH_SIZE = BATCH_SIZE * 2
+USE_XLA = False
+USE_AMP = False
+EPOCHS = 3
+
+TASK = "mrpc"
+
+if TASK == "sst-2":
+    TFDS_TASK = "sst2"
+elif TASK == "sts-b":
+    TFDS_TASK = "stsb"
+else: 
+    TFDS_TASK = TASK
+
+num_labels = len(glue_processors[TASK]().get_labels())
+print(num_labels)
+
+tf.config.optimizer.set_jit(USE_XLA)
+tf.config.optimizer.set_experimental_options({"auto_mixed_precision": USE_AMP})
+
+# Load tokenizer and model from pretrained model/vocabulary. Specify the number of labels to classify (2+: classification, 1: regression)
+config = BertConfig.from_pretrained("bert-base-cased", num_labels=num_labels)
 tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
-model = TFBertForSequenceClassification.from_pretrained('bert-base-cased')
-data = tensorflow_datasets.load('glue/mrpc')
+model = TFBertForSequenceClassification.from_pretrained('bert-base-cased', config=config)
+
+# Load dataset via TensorFlow Datasets
+data, info = tensorflow_datasets.load(f'glue/{TFDS_TASK}', with_info=True)
+train_examples = info.splits['train'].num_examples
+
+# MNLI expects either validation_matched or validation_mismatched
+valid_examples = info.splits['validation'].num_examples
 
 # Prepare dataset for GLUE as a tf.data.Dataset instance
-train_dataset = glue_convert_examples_to_features(data['train'], tokenizer, 128, 'mrpc')
-valid_dataset = glue_convert_examples_to_features(data['validation'], tokenizer, 128, 'mrpc')
-train_dataset = train_dataset.shuffle(100).batch(32).repeat(2)
-valid_dataset = valid_dataset.batch(64)
+train_dataset = glue_convert_examples_to_features(data['train'], tokenizer, 128, TASK)
+
+# MNLI expects either validation_matched or validation_mismatched
+valid_dataset = glue_convert_examples_to_features(data['validation'], tokenizer, 128, TASK)
+train_dataset = train_dataset.shuffle(128).batch(BATCH_SIZE).repeat(-1)
+valid_dataset = valid_dataset.batch(EVAL_BATCH_SIZE)
 
 # Prepare training: Compile tf.keras model with optimizer, loss and learning rate schedule 
-optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
-loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
+opt = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08)
+if USE_AMP:
+    # loss scaling is currently required when using mixed precision
+    opt = tf.keras.mixed_precision.experimental.LossScaleOptimizer(opt, 'dynamic')
+
+
+if num_labels == 1:
+    loss = tf.keras.losses.MeanSquaredError()
+else:
+    loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
+
 metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')
-model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
+model.compile(optimizer=opt, loss=loss, metrics=[metric])
 
 # Train and evaluate using tf.keras.Model.fit()
-history = model.fit(train_dataset, epochs=2, steps_per_epoch=115,
-                    validation_data=valid_dataset, validation_steps=7)
+train_steps = train_examples//BATCH_SIZE
+valid_steps = valid_examples//EVAL_BATCH_SIZE
 
-# Load the TensorFlow model in PyTorch for inspection
+history = model.fit(train_dataset, epochs=EPOCHS, steps_per_epoch=train_steps,
+                    validation_data=valid_dataset, validation_steps=valid_steps)
+
+# Save TF2 model
+os.makedirs('./save/', exist_ok=True)
 model.save_pretrained('./save/')
-pytorch_model = BertForSequenceClassification.from_pretrained('./save/', from_tf=True)
 
-# Quickly test a few predictions - MRPC is a paraphrasing task, let's see if our model learned the task
-sentence_0 = "This research was consistent with his findings."
-sentence_1 = "His findings were compatible with this research."
-sentence_2 = "His findings were not compatible with this research."
-inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
-inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
+if TASK == "mrpc":
+    # Load the TensorFlow model in PyTorch for inspection
+    # This is to demo the interoperability between the two frameworks, you don't have to 
+    # do this in real life (you can run the inference on the TF model).
+    pytorch_model = BertForSequenceClassification.from_pretrained('./save/', from_tf=True)
 
-pred_1 = pytorch_model(**inputs_1)[0].argmax().item()
-pred_2 = pytorch_model(**inputs_2)[0].argmax().item()
-print("sentence_1 is", "a paraphrase" if pred_1 else "not a paraphrase", "of sentence_0")
-print("sentence_2 is", "a paraphrase" if pred_2 else "not a paraphrase", "of sentence_0")
+    # Quickly test a few predictions - MRPC is a paraphrasing task, let's see if our model learned the task
+    sentence_0 = 'This research was consistent with his findings.'
+    sentence_1 = 'His findings were compatible with this research.'
+    sentence_2 = 'His findings were not compatible with this research.'
+    inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
+    inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
+
+    del inputs_1["special_tokens_mask"]
+    del inputs_2["special_tokens_mask"]
+
+    pred_1 = pytorch_model(**inputs_1)[0].argmax().item()
+    pred_2 = pytorch_model(**inputs_2)[0].argmax().item()
+    print('sentence_1 is', 'a paraphrase' if pred_1 else 'not a paraphrase', 'of sentence_0')
+    print('sentence_2 is', 'a paraphrase' if pred_2 else 'not a paraphrase', 'of sentence_0')

examples/utils_ner.py

@@ -0,0 +1,212 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Named entity recognition fine-tuning: utilities to work with CoNLL-2003 task. """
+
+from __future__ import absolute_import, division, print_function
+
+import logging
+import os
+from io import open
+
+logger = logging.getLogger(__name__)
+
+
+class InputExample(object):
+    """A single training/test example for token classification."""
+
+    def __init__(self, guid, words, labels):
+        """Constructs a InputExample.
+
+        Args:
+            guid: Unique id for the example.
+            words: list. The words of the sequence.
+            labels: (Optional) list. The labels for each word of the sequence. This should be
+            specified for train and dev examples, but not for test examples.
+        """
+        self.guid = guid
+        self.words = words
+        self.labels = labels
+
+
+class InputFeatures(object):
+    """A single set of features of data."""
+
+    def __init__(self, input_ids, input_mask, segment_ids, label_ids):
+        self.input_ids = input_ids
+        self.input_mask = input_mask
+        self.segment_ids = segment_ids
+        self.label_ids = label_ids
+
+
+def read_examples_from_file(data_dir, mode):
+    file_path = os.path.join(data_dir, "{}.txt".format(mode))
+    guid_index = 1
+    examples = []
+    with open(file_path, encoding="utf-8") as f:
+        words = []
+        labels = []
+        for line in f:
+            if line.startswith("-DOCSTART-") or line == "" or line == "\n":
+                if words:
+                    examples.append(InputExample(guid="{}-{}".format(mode, guid_index),
+                                                 words=words,
+                                                 labels=labels))
+                    guid_index += 1
+                    words = []
+                    labels = []
+            else:
+                splits = line.split(" ")
+                words.append(splits[0])
+                if len(splits) > 1:
+                    labels.append(splits[-1].replace("\n", ""))
+                else:
+                    # Examples could have no label for mode = "test"
+                    labels.append("O")
+        if words:
+            examples.append(InputExample(guid="%s-%d".format(mode, guid_index),
+                                         words=words,
+                                         labels=labels))
+    return examples
+
+
+def convert_examples_to_features(examples,
+                                 label_list,
+                                 max_seq_length,
+                                 tokenizer,
+                                 cls_token_at_end=False,
+                                 cls_token="[CLS]",
+                                 cls_token_segment_id=1,
+                                 sep_token="[SEP]",
+                                 sep_token_extra=False,
+                                 pad_on_left=False,
+                                 pad_token=0,
+                                 pad_token_segment_id=0,
+                                 pad_token_label_id=-1,
+                                 sequence_a_segment_id=0,
+                                 mask_padding_with_zero=True):
+    """ Loads a data file into a list of `InputBatch`s
+        `cls_token_at_end` define the location of the CLS token:
+            - False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]
+            - True (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS]
+        `cls_token_segment_id` define the segment id associated to the CLS token (0 for BERT, 2 for XLNet)
+    """
+
+    label_map = {label: i for i, label in enumerate(label_list)}
+
+    features = []
+    for (ex_index, example) in enumerate(examples):
+        if ex_index % 10000 == 0:
+            logger.info("Writing example %d of %d", ex_index, len(examples))
+
+        tokens = []
+        label_ids = []
+        for word, label in zip(example.words, example.labels):
+            word_tokens = tokenizer.tokenize(word)
+            tokens.extend(word_tokens)
+            # Use the real label id for the first token of the word, and padding ids for the remaining tokens
+            label_ids.extend([label_map[label]] + [pad_token_label_id] * (len(word_tokens) - 1))
+
+        # Account for [CLS] and [SEP] with "- 2" and with "- 3" for RoBERTa.
+        special_tokens_count = 3 if sep_token_extra else 2
+        if len(tokens) > max_seq_length - special_tokens_count:
+            tokens = tokens[:(max_seq_length - special_tokens_count)]
+            label_ids = label_ids[:(max_seq_length - special_tokens_count)]
+
+        # The convention in BERT is:
+        # (a) For sequence pairs:
+        #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
+        #  type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1
+        # (b) For single sequences:
+        #  tokens:   [CLS] the dog is hairy . [SEP]
+        #  type_ids:   0   0   0   0  0     0   0
+        #
+        # Where "type_ids" are used to indicate whether this is the first
+        # sequence or the second sequence. The embedding vectors for `type=0` and
+        # `type=1` were learned during pre-training and are added to the wordpiece
+        # embedding vector (and position vector). This is not *strictly* necessary
+        # since the [SEP] token unambiguously separates the sequences, but it makes
+        # it easier for the model to learn the concept of sequences.
+        #
+        # For classification tasks, the first vector (corresponding to [CLS]) is
+        # used as as the "sentence vector". Note that this only makes sense because
+        # the entire model is fine-tuned.
+        tokens += [sep_token]
+        label_ids += [pad_token_label_id]
+        if sep_token_extra:
+            # roberta uses an extra separator b/w pairs of sentences
+            tokens += [sep_token]
+            label_ids += [pad_token_label_id]
+        segment_ids = [sequence_a_segment_id] * len(tokens)
+
+        if cls_token_at_end:
+            tokens += [cls_token]
+            label_ids += [pad_token_label_id]
+            segment_ids += [cls_token_segment_id]
+        else:
+            tokens = [cls_token] + tokens
+            label_ids = [pad_token_label_id] + label_ids
+            segment_ids = [cls_token_segment_id] + segment_ids
+
+        input_ids = tokenizer.convert_tokens_to_ids(tokens)
+
+        # The mask has 1 for real tokens and 0 for padding tokens. Only real
+        # tokens are attended to.
+        input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
+
+        # Zero-pad up to the sequence length.
+        padding_length = max_seq_length - len(input_ids)
+        if pad_on_left:
+            input_ids = ([pad_token] * padding_length) + input_ids
+            input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask
+            segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids
+            label_ids = ([pad_token_label_id] * padding_length) + label_ids
+        else:
+            input_ids += ([pad_token] * padding_length)
+            input_mask += ([0 if mask_padding_with_zero else 1] * padding_length)
+            segment_ids += ([pad_token_segment_id] * padding_length)
+            label_ids += ([pad_token_label_id] * padding_length)
+
+        assert len(input_ids) == max_seq_length
+        assert len(input_mask) == max_seq_length
+        assert len(segment_ids) == max_seq_length
+        assert len(label_ids) == max_seq_length
+
+        if ex_index < 5:
+            logger.info("*** Example ***")
+            logger.info("guid: %s", example.guid)
+            logger.info("tokens: %s", " ".join([str(x) for x in tokens]))
+            logger.info("input_ids: %s", " ".join([str(x) for x in input_ids]))
+            logger.info("input_mask: %s", " ".join([str(x) for x in input_mask]))
+            logger.info("segment_ids: %s", " ".join([str(x) for x in segment_ids]))
+            logger.info("label_ids: %s", " ".join([str(x) for x in label_ids]))
+
+        features.append(
+                InputFeatures(input_ids=input_ids,
+                              input_mask=input_mask,
+                              segment_ids=segment_ids,
+                              label_ids=label_ids))
+    return features
+
+
+def get_labels(path):
+    if path:
+        with open(path, "r") as f:
+            labels = f.read().splitlines()
+        if "O" not in labels:
+            labels = ["O"] + labels
+        return labels
+    else:
+        return ["O", "B-MISC", "I-MISC",  "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"]

examples/utils_squad.py

@@ -23,6 +23,7 @@ import logging
 import math
 import collections
 from io import open
+from tqdm import tqdm
 
 from transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize
 
@@ -192,7 +193,8 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
                                  cls_token='[CLS]', sep_token='[SEP]', pad_token=0,
                                  sequence_a_segment_id=0, sequence_b_segment_id=1,
                                  cls_token_segment_id=0, pad_token_segment_id=0,
-                                 mask_padding_with_zero=True):
+                                 mask_padding_with_zero=True,
+                                 sequence_a_is_doc=False):
     """Loads a data file into a list of `InputBatch`s."""
 
     unique_id = 1000000000
@@ -201,7 +203,7 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
     # f = np.zeros((max_N, max_M), dtype=np.float32)
 
     features = []
-    for (example_index, example) in enumerate(examples):
+    for (example_index, example) in enumerate(tqdm(examples)):
 
         # if example_index % 100 == 0:
         #     logger.info('Converting %s/%s pos %s neg %s', example_index, len(examples), cnt_pos, cnt_neg)
@@ -238,6 +240,7 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
 
         # The -3 accounts for [CLS], [SEP] and [SEP]
         max_tokens_for_doc = max_seq_length - len(query_tokens) - 3
+        assert max_tokens_for_doc > 0
 
         # We can have documents that are longer than the maximum sequence length.
         # To deal with this we do a sliding window approach, where we take chunks
@@ -272,17 +275,19 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
                 p_mask.append(0)
                 cls_index = 0
 
-            # Query
-            for token in query_tokens:
-                tokens.append(token)
+            # XLNet: P SEP Q SEP CLS
+            # Others: CLS Q SEP P SEP
+            if not sequence_a_is_doc:
+                # Query
+                tokens += query_tokens
+                segment_ids += [sequence_a_segment_id] * len(query_tokens)
+                p_mask += [1] * len(query_tokens)
+
+                # SEP token
+                tokens.append(sep_token)
                 segment_ids.append(sequence_a_segment_id)
                 p_mask.append(1)
 
-            # SEP token
-            tokens.append(sep_token)
-            segment_ids.append(sequence_a_segment_id)
-            p_mask.append(1)
-
             # Paragraph
             for i in range(doc_span.length):
                 split_token_index = doc_span.start + i
@@ -292,10 +297,23 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
                                                        split_token_index)
                 token_is_max_context[len(tokens)] = is_max_context
                 tokens.append(all_doc_tokens[split_token_index])
-                segment_ids.append(sequence_b_segment_id)
+                if not sequence_a_is_doc:
+                    segment_ids.append(sequence_b_segment_id)
+                else:
+                    segment_ids.append(sequence_a_segment_id)
                 p_mask.append(0)
             paragraph_len = doc_span.length
 
+            if sequence_a_is_doc:
+                # SEP token
+                tokens.append(sep_token)
+                segment_ids.append(sequence_a_segment_id)
+                p_mask.append(1)
+
+                tokens += query_tokens
+                segment_ids += [sequence_b_segment_id] * len(query_tokens)
+                p_mask += [1] * len(query_tokens)
+
             # SEP token
             tokens.append(sep_token)
             segment_ids.append(sequence_b_segment_id)
@@ -342,7 +360,10 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
                     end_position = 0
                     span_is_impossible = True
                 else:
-                    doc_offset = len(query_tokens) + 2
+                    if sequence_a_is_doc:
+                        doc_offset = 0
+                    else:
+                        doc_offset = len(query_tokens) + 2
                     start_position = tok_start_position - doc_start + doc_offset
                     end_position = tok_end_position - doc_start + doc_offset
 

examples/utils_summarization.py

@@ -0,0 +1,184 @@
+from collections import deque
+import os
+
+import torch
+from torch.utils.data import Dataset
+
+
+# ------------
+# Data loading
+# ------------
+
+
+class CNNDailyMailDataset(Dataset):
+    """ Abstracts the dataset used to train seq2seq models.
+
+    CNN/Daily News:
+
+    The CNN/Daily News raw datasets are downloaded from [1]. The stories are
+    stored in different files; the summary appears at the end of the story as
+    sentences that are prefixed by the special `@highlight` line. To process
+    the data, untar both datasets in the same folder, and pass the path to this
+    folder as the "data_dir argument. The formatting code was inspired by [2].
+
+    [1] https://cs.nyu.edu/~kcho/
+    [2] https://github.com/abisee/cnn-dailymail/
+    """
+
+    def __init__(self, tokenizer, prefix="train", data_dir=""):
+        assert os.path.isdir(data_dir)
+        self.tokenizer = tokenizer
+
+        # We initialize the class by listing all the files that contain
+        # stories and summaries. Files are not read in memory given
+        # the size of the corpus.
+        self.stories_path = []
+        datasets = ("cnn", "dailymail")
+        for dataset in datasets:
+            path_to_stories = os.path.join(data_dir, dataset, "stories")
+            story_filenames_list = os.listdir(path_to_stories)
+            for story_filename in story_filenames_list:
+                path_to_story = os.path.join(path_to_stories, story_filename)
+                if not os.path.isfile(path_to_story):
+                    continue
+                self.stories_path.append(path_to_story)
+
+    def __len__(self):
+        return len(self.stories_path)
+
+    def __getitem__(self, idx):
+        story_path = self.stories_path[idx]
+        with open(story_path, encoding="utf-8") as source:
+            raw_story = source.read()
+            story_lines, summary_lines = process_story(raw_story)
+        return story_lines, summary_lines
+
+
+def process_story(raw_story):
+    """ Extract the story and summary from a story file.
+
+    Attributes:
+        raw_story (str): content of the story file as an utf-8 encoded string.
+
+    Raises:
+        IndexError: If the stoy is empty or contains no highlights.
+    """
+    nonempty_lines = list(
+        filter(lambda x: len(x) != 0, [line.strip() for line in raw_story.split("\n")])
+    )
+
+    # for some unknown reason some lines miss a period, add it
+    nonempty_lines = [_add_missing_period(line) for line in nonempty_lines]
+
+    # gather article lines
+    story_lines = []
+    lines = deque(nonempty_lines)
+    while True:
+        try:
+            element = lines.popleft()
+            if element.startswith("@highlight"):
+                break
+            story_lines.append(element)
+        except IndexError:
+            # if "@highlight" is absent from the file we pop
+            # all elements until there is None.
+            return story_lines, []
+
+    # gather summary lines
+    summary_lines = list(filter(lambda t: not t.startswith("@highlight"), lines))
+
+    return story_lines, summary_lines
+
+
+def _add_missing_period(line):
+    END_TOKENS = [".", "!", "?", "...", "'", "`", '"', u"\u2019", u"\u2019", ")"]
+    if line.startswith("@highlight"):
+        return line
+    if line[-1] in END_TOKENS:
+        return line
+    return line + "."
+
+
+# --------------------------
+# Encoding and preprocessing
+# --------------------------
+
+
+def fit_to_block_size(sequence, block_size, pad_token):
+    """ Adapt the source and target sequences' lengths to the block size.
+    If the sequence is shorter than the block size we pad it with -1 ids
+    which correspond to padding tokens.
+    """
+    if len(sequence) > block_size:
+        return sequence[:block_size]
+    else:
+        sequence.extend([pad_token] * (block_size - len(sequence)))
+        return sequence
+
+
+def build_lm_labels(sequence, pad_token):
+    """ Padding token, encoded as 0, are represented by the value -1 so they
+    are not taken into account in the loss computation. """
+    padded = sequence.clone()
+    padded[padded == pad_token] = -1
+    return padded
+
+
+def build_mask(sequence, pad_token):
+    """ Builds the mask. The attention mechanism will only attend to positions
+    with value 1. """
+    mask = torch.ones_like(sequence)
+    idx_pad_tokens = sequence == pad_token
+    mask[idx_pad_tokens] = 0
+    return mask
+
+
+def encode_for_summarization(story_lines, summary_lines, tokenizer):
+    """ Encode the story and summary lines, and join them
+    as specified in [1] by using `[SEP] [CLS]` tokens to separate
+    sentences.
+    """
+    story_lines_token_ids = [
+        tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line))
+        for line in story_lines
+    ]
+    summary_lines_token_ids = [
+        tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line))
+        for line in summary_lines
+    ]
+
+    story_token_ids = [
+        token for sentence in story_lines_token_ids for token in sentence
+    ]
+    summary_token_ids = [
+        token for sentence in summary_lines_token_ids for token in sentence
+    ]
+
+    return story_token_ids, summary_token_ids
+
+
+def compute_token_type_ids(batch, separator_token_id):
+    """ Segment embeddings as described in [1]
+
+    The values {0,1} were found in the repository [2].
+
+    Attributes:
+        batch: torch.Tensor, size [batch_size, block_size]
+            Batch of input.
+        separator_token_id: int
+            The value of the token that separates the segments.
+
+    [1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders."
+        arXiv preprint arXiv:1908.08345 (2019).
+    [2] https://github.com/nlpyang/PreSumm (/src/prepro/data_builder.py, commit fac1217)
+    """
+    batch_embeddings = []
+    for sequence in batch:
+        sentence_num = 0
+        embeddings = []
+        for s in sequence:
+            if s == separator_token_id:
+                sentence_num += 1
+            embeddings.append(sentence_num % 2)
+        batch_embeddings.append(embeddings)
+    return torch.tensor(batch_embeddings)

examples/utils_summarization_test.py

@@ -0,0 +1,136 @@
+# coding=utf-8
+# Copyright 2019 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import unittest
+
+import numpy as np
+import torch
+
+from utils_summarization import (
+    compute_token_type_ids,
+    fit_to_block_size,
+    build_mask,
+    build_lm_labels,
+    process_story,
+)
+
+
+class SummarizationDataProcessingTest(unittest.TestCase):
+    def setUp(self):
+        self.block_size = 10
+
+    def test_fit_to_block_sequence_too_small(self):
+        """ Pad the sequence with 0 if the sequence is smaller than the block size."""
+        sequence = [1, 2, 3, 4]
+        expected_output = [1, 2, 3, 4, 0, 0, 0, 0, 0, 0]
+        self.assertEqual(
+            fit_to_block_size(sequence, self.block_size, 0), expected_output
+        )
+
+    def test_fit_to_block_sequence_fit_exactly(self):
+        """ Do nothing if the sequence is the right size. """
+        sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
+        expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
+        self.assertEqual(
+            fit_to_block_size(sequence, self.block_size, 0), expected_output
+        )
+
+    def test_fit_to_block_sequence_too_big(self):
+        """ Truncate the sequence if it is too long. """
+        sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
+        expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
+        self.assertEqual(
+            fit_to_block_size(sequence, self.block_size, 0), expected_output
+        )
+
+    def test_process_story_no_highlights(self):
+        """ Processing a story with no highlights returns an empty list for the summary.
+        """
+        raw_story = """It was the year of Our Lord one thousand seven hundred and
+        seventy-five.\n\nSpiritual revelations were conceded to England at that
+        favoured period, as at this."""
+        _, summary_lines = process_story(raw_story)
+        self.assertEqual(summary_lines, [])
+
+    def test_process_empty_story(self):
+        """ An empty story returns an empty collection of lines.
+        """
+        raw_story = ""
+        story_lines, summary_lines = process_story(raw_story)
+        self.assertEqual(story_lines, [])
+        self.assertEqual(summary_lines, [])
+
+    def test_process_story_with_missing_period(self):
+        raw_story = (
+            "It was the year of Our Lord one thousand seven hundred and "
+            "seventy-five\n\nSpiritual revelations were conceded to England "
+            "at that favoured period, as at this.\n@highlight\n\nIt was the best of times"
+        )
+        story_lines, summary_lines = process_story(raw_story)
+
+        expected_story_lines = [
+            "It was the year of Our Lord one thousand seven hundred and seventy-five.",
+            "Spiritual revelations were conceded to England at that favoured period, as at this.",
+        ]
+        self.assertEqual(expected_story_lines, story_lines)
+
+        expected_summary_lines = ["It was the best of times."]
+        self.assertEqual(expected_summary_lines, summary_lines)
+
+    def test_build_lm_labels_no_padding(self):
+        sequence = torch.tensor([1, 2, 3, 4])
+        expected = sequence
+        np.testing.assert_array_equal(
+            build_lm_labels(sequence, 0).numpy(), expected.numpy()
+        )
+
+    def test_build_lm_labels(self):
+        sequence = torch.tensor([1, 2, 3, 4, 0, 0, 0])
+        expected = torch.tensor([1, 2, 3, 4, -1, -1, -1])
+        np.testing.assert_array_equal(
+            build_lm_labels(sequence, 0).numpy(), expected.numpy()
+        )
+
+    def test_build_mask_no_padding(self):
+        sequence = torch.tensor([1, 2, 3, 4])
+        expected = torch.tensor([1, 1, 1, 1])
+        np.testing.assert_array_equal(build_mask(sequence, 0).numpy(), expected.numpy())
+
+    def test_build_mask(self):
+        sequence = torch.tensor([1, 2, 3, 4, 23, 23, 23])
+        expected = torch.tensor([1, 1, 1, 1, 0, 0, 0])
+        np.testing.assert_array_equal(
+            build_mask(sequence, 23).numpy(), expected.numpy()
+        )
+
+    def test_build_mask_with_padding_equal_to_one(self):
+        sequence = torch.tensor([8, 2, 3, 4, 1, 1, 1])
+        expected = torch.tensor([1, 1, 1, 1, 0, 0, 0])
+        np.testing.assert_array_equal(build_mask(sequence, 1).numpy(), expected.numpy())
+
+    def test_compute_token_type_ids(self):
+        separator = 101
+        batch = torch.tensor(
+            [[1, 2, 3, 4, 5, 6], [1, 2, 3, 101, 5, 6], [1, 101, 3, 4, 101, 6]]
+        )
+        expected = torch.tensor(
+            [[0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1], [0, 1, 1, 1, 0, 0]]
+        )
+
+        result = compute_token_type_ids(batch, separator)
+        np.testing.assert_array_equal(result, expected)
+
+
+if __name__ == "__main__":
+    unittest.main()

setup.py

@@ -38,7 +38,7 @@ from setuptools import find_packages, setup
 
 setup(
     name="transformers",
-    version="2.1.0",
+    version="2.2.0",
     author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
     author_email="thomas@huggingface.co",
     description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch",

templates/adding_a_new_example_script/README.md

@@ -0,0 +1,5 @@
+# How to add a new example script in 🤗Transformers
+
+This folder provide a template for adding a new example script implementing a training or inference task with the models in the  🤗Transformers library.
+
+Currently only examples for PyTorch are provided which are adaptations of the library's SQuAD examples which implement single-GPU and distributed training with gradient accumulation and mixed-precision (using NVIDIA's apex library) to cover a reasonable range of use cases.

templates/adding_a_new_example_script/run_xxx.py

@@ -0,0 +1,559 @@
+# coding=utf-8
+# Copyright 2018 XXX.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Finetuning the library models for task XXX."""
+
+from __future__ import absolute_import, division, print_function
+
+import argparse
+import logging
+import os
+import random
+import glob
+
+import numpy as np
+import torch
+from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
+                              TensorDataset)
+from torch.utils.data.distributed import DistributedSampler
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
+from tqdm import tqdm, trange
+
+from transformers import (WEIGHTS_NAME, BertConfig,
+                                  BertForQuestionAnswering, BertTokenizer,
+                                  XLMConfig, XLMForQuestionAnswering,
+                                  XLMTokenizer, XLNetConfig,
+                                  XLNetForQuestionAnswering,
+                                  XLNetTokenizer,
+                                  DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
+
+from transformers import AdamW, get_linear_schedule_with_warmup
+
+from utils_squad import (read_squad_examples, convert_examples_to_features,
+                         RawResult, write_predictions,
+                         RawResultExtended, write_predictions_extended)
+
+# The follwing import is the official SQuAD evaluation script (2.0).
+# You can remove it from the dependencies if you are using this script outside of the library
+# We've added it here for automated tests (see examples/test_examples.py file)
+from utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad
+
+logger = logging.getLogger(__name__)
+
+ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
+                  for conf in (BertConfig, XLNetConfig, XLMConfig)), ())
+
+MODEL_CLASSES = {
+    'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
+    'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
+    'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
+    'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
+}
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+def to_list(tensor):
+    return tensor.detach().cpu().tolist()
+
+def train(args, train_dataset, model, tokenizer):
+    """ Train the model """
+    if args.local_rank in [-1, 0]:
+        tb_writer = SummaryWriter()
+
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+    else:
+        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+
+    # Prepare optimizer and schedule (linear warmup and decay)
+    no_decay = ['bias', 'LayerNorm.weight']
+    optimizer_grouped_parameters = [
+        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
+        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
+        ]
+    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
+                                                          output_device=args.local_rank,
+                                                          find_unused_parameters=True)
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
+                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+
+    global_step = 0
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+            model.train()
+            batch = tuple(t.to(args.device) for t in batch)
+            inputs = {'input_ids':       batch[0],
+                      'attention_mask':  batch[1],
+                      'start_positions': batch[3],
+                      'end_positions':   batch[4]}
+            if args.model_type != 'distilbert':
+                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
+            if args.model_type in ['xlnet', 'xlm']:
+                inputs.update({'cls_index': batch[5],
+                               'p_mask':       batch[6]})
+            outputs = model(**inputs)
+            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
+
+            if args.n_gpu > 1:
+                loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+
+            if args.fp16:
+                with amp.scale_loss(loss, optimizer) as scaled_loss:
+                    scaled_loss.backward()
+            else:
+                loss.backward()
+
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+
+                optimizer.step()
+                scheduler.step()  # Update learning rate schedule
+                model.zero_grad()
+                global_step += 1
+
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    # Log metrics
+                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
+                        results = evaluate(args, model, tokenizer)
+                        for key, value in results.items():
+                            tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
+                    tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
+                    logging_loss = tr_loss
+
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    # Save model checkpoint
+                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
+                    model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+                    model_to_save.save_pretrained(output_dir)
+                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+
+    return global_step, tr_loss / global_step
+
+
+def evaluate(args, model, tokenizer, prefix=""):
+    dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
+
+    if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+        os.makedirs(args.output_dir)
+
+    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+    # Note that DistributedSampler samples randomly
+    eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
+    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+    # Eval!
+    logger.info("***** Running evaluation {} *****".format(prefix))
+    logger.info("  Num examples = %d", len(dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+    all_results = []
+    for batch in tqdm(eval_dataloader, desc="Evaluating"):
+        model.eval()
+        batch = tuple(t.to(args.device) for t in batch)
+        with torch.no_grad():
+            inputs = {'input_ids':      batch[0],
+                      'attention_mask': batch[1]
+                      }
+            if args.model_type != 'distilbert':
+                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]  # XLM don't use segment_ids
+            example_indices = batch[3]
+            if args.model_type in ['xlnet', 'xlm']:
+                inputs.update({'cls_index': batch[4],
+                               'p_mask':    batch[5]})
+            outputs = model(**inputs)
+
+        for i, example_index in enumerate(example_indices):
+            eval_feature = features[example_index.item()]
+            unique_id = int(eval_feature.unique_id)
+            if args.model_type in ['xlnet', 'xlm']:
+                # XLNet uses a more complex post-processing procedure
+                result = RawResultExtended(unique_id            = unique_id,
+                                           start_top_log_probs  = to_list(outputs[0][i]),
+                                           start_top_index      = to_list(outputs[1][i]),
+                                           end_top_log_probs    = to_list(outputs[2][i]),
+                                           end_top_index        = to_list(outputs[3][i]),
+                                           cls_logits           = to_list(outputs[4][i]))
+            else:
+                result = RawResult(unique_id    = unique_id,
+                                   start_logits = to_list(outputs[0][i]),
+                                   end_logits   = to_list(outputs[1][i]))
+            all_results.append(result)
+
+    # Compute predictions
+    output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
+    output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
+    if args.version_2_with_negative:
+        output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
+    else:
+        output_null_log_odds_file = None
+
+    if args.model_type in ['xlnet', 'xlm']:
+        # XLNet uses a more complex post-processing procedure
+        write_predictions_extended(examples, features, all_results, args.n_best_size,
+                        args.max_answer_length, output_prediction_file,
+                        output_nbest_file, output_null_log_odds_file, args.predict_file,
+                        model.config.start_n_top, model.config.end_n_top,
+                        args.version_2_with_negative, tokenizer, args.verbose_logging)
+    else:
+        write_predictions(examples, features, all_results, args.n_best_size,
+                        args.max_answer_length, args.do_lower_case, output_prediction_file,
+                        output_nbest_file, output_null_log_odds_file, args.verbose_logging,
+                        args.version_2_with_negative, args.null_score_diff_threshold)
+
+    # Evaluate with the official SQuAD script
+    evaluate_options = EVAL_OPTS(data_file=args.predict_file,
+                                 pred_file=output_prediction_file,
+                                 na_prob_file=output_null_log_odds_file)
+    results = evaluate_on_squad(evaluate_options)
+    return results
+
+
+def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
+    if args.local_rank not in [-1, 0] and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Load data features from cache or dataset file
+    input_file = args.predict_file if evaluate else args.train_file
+    cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
+        'dev' if evaluate else 'train',
+        list(filter(None, args.model_name_or_path.split('/'))).pop(),
+        str(args.max_seq_length)))
+    if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features = torch.load(cached_features_file)
+    else:
+        logger.info("Creating features from dataset file at %s", input_file)
+        examples = read_squad_examples(input_file=input_file,
+                                                is_training=not evaluate,
+                                                version_2_with_negative=args.version_2_with_negative)
+        features = convert_examples_to_features(examples=examples,
+                                                tokenizer=tokenizer,
+                                                max_seq_length=args.max_seq_length,
+                                                doc_stride=args.doc_stride,
+                                                max_query_length=args.max_query_length,
+                                                is_training=not evaluate)
+        if args.local_rank in [-1, 0]:
+            logger.info("Saving features into cached file %s", cached_features_file)
+            torch.save(features, cached_features_file)
+
+    if args.local_rank == 0 and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Convert to Tensors and build dataset
+    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
+    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
+    all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
+    all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
+    all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
+    if evaluate:
+        all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
+        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
+                                all_example_index, all_cls_index, all_p_mask)
+    else:
+        all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
+        all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
+        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
+                                all_start_positions, all_end_positions,
+                                all_cls_index, all_p_mask)
+
+    if output_examples:
+        return dataset, examples, features
+    return dataset
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    ## Required parameters
+    parser.add_argument("--train_file", default=None, type=str, required=True,
+                        help="SQuAD json for training. E.g., train-v1.1.json")
+    parser.add_argument("--predict_file", default=None, type=str, required=True,
+                        help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
+    parser.add_argument("--model_type", default=None, type=str, required=True,
+                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
+                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+    parser.add_argument("--output_dir", default=None, type=str, required=True,
+                        help="The output directory where the model checkpoints and predictions will be written.")
+
+    ## Other parameters
+    parser.add_argument("--config_name", default="", type=str,
+                        help="Pretrained config name or path if not the same as model_name")
+    parser.add_argument("--tokenizer_name", default="", type=str,
+                        help="Pretrained tokenizer name or path if not the same as model_name")
+    parser.add_argument("--cache_dir", default="", type=str,
+                        help="Where do you want to store the pre-trained models downloaded from s3")
+
+    parser.add_argument('--version_2_with_negative', action='store_true',
+                        help='If true, the SQuAD examples contain some that do not have an answer.')
+    parser.add_argument('--null_score_diff_threshold', type=float, default=0.0,
+                        help="If null_score - best_non_null is greater than the threshold predict null.")
+
+    parser.add_argument("--max_seq_length", default=384, type=int,
+                        help="The maximum total input sequence length after WordPiece tokenization. Sequences "
+                             "longer than this will be truncated, and sequences shorter than this will be padded.")
+    parser.add_argument("--doc_stride", default=128, type=int,
+                        help="When splitting up a long document into chunks, how much stride to take between chunks.")
+    parser.add_argument("--max_query_length", default=64, type=int,
+                        help="The maximum number of tokens for the question. Questions longer than this will "
+                             "be truncated to this length.")
+    parser.add_argument("--do_train", action='store_true',
+                        help="Whether to run training.")
+    parser.add_argument("--do_eval", action='store_true',
+                        help="Whether to run eval on the dev set.")
+    parser.add_argument("--evaluate_during_training", action='store_true',
+                        help="Rul evaluation during training at each logging step.")
+    parser.add_argument("--do_lower_case", action='store_true',
+                        help="Set this flag if you are using an uncased model.")
+
+    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
+                        help="Batch size per GPU/CPU for training.")
+    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
+                        help="Batch size per GPU/CPU for evaluation.")
+    parser.add_argument("--learning_rate", default=5e-5, type=float,
+                        help="The initial learning rate for Adam.")
+    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
+                        help="Number of updates steps to accumulate before performing a backward/update pass.")
+    parser.add_argument("--weight_decay", default=0.0, type=float,
+                        help="Weight deay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float,
+                        help="Max gradient norm.")
+    parser.add_argument("--num_train_epochs", default=3.0, type=float,
+                        help="Total number of training epochs to perform.")
+    parser.add_argument("--max_steps", default=-1, type=int,
+                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
+    parser.add_argument("--warmup_steps", default=0, type=int,
+                        help="Linear warmup over warmup_steps.")
+    parser.add_argument("--n_best_size", default=20, type=int,
+                        help="The total number of n-best predictions to generate in the nbest_predictions.json output file.")
+    parser.add_argument("--max_answer_length", default=30, type=int,
+                        help="The maximum length of an answer that can be generated. This is needed because the start "
+                             "and end predictions are not conditioned on one another.")
+    parser.add_argument("--verbose_logging", action='store_true',
+                        help="If true, all of the warnings related to data processing will be printed. "
+                             "A number of warnings are expected for a normal SQuAD evaluation.")
+
+    parser.add_argument('--logging_steps', type=int, default=50,
+                        help="Log every X updates steps.")
+    parser.add_argument('--save_steps', type=int, default=50,
+                        help="Save checkpoint every X updates steps.")
+    parser.add_argument("--eval_all_checkpoints", action='store_true',
+                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
+    parser.add_argument("--no_cuda", action='store_true',
+                        help="Whether not to use CUDA when available")
+    parser.add_argument('--overwrite_output_dir', action='store_true',
+                        help="Overwrite the content of the output directory")
+    parser.add_argument('--overwrite_cache', action='store_true',
+                        help="Overwrite the cached training and evaluation sets")
+    parser.add_argument('--seed', type=int, default=42,
+                        help="random seed for initialization")
+
+    parser.add_argument("--local_rank", type=int, default=-1,
+                        help="local_rank for distributed training on gpus")
+    parser.add_argument('--fp16', action='store_true',
+                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
+    parser.add_argument('--fp16_opt_level', type=str, default='O1',
+                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+                             "See details at https://nvidia.github.io/apex/amp.html")
+    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
+    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
+    args = parser.parse_args()
+
+    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
+        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
+
+    # Setup distant debugging if needed
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    # Setup CUDA, GPU & distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        args.n_gpu = torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend='nccl')
+        args.n_gpu = 1
+    args.device = device
+
+    # Setup logging
+    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+                        datefmt = '%m/%d/%Y %H:%M:%S',
+                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
+    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
+
+    # Set seed
+    set_seed(args)
+
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    args.model_type = args.model_type.lower()
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
+                                          cache_dir=args.cache_dir if args.cache_dir else None)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+                                                do_lower_case=args.do_lower_case,
+                                                cache_dir=args.cache_dir if args.cache_dir else None)
+    model = model_class.from_pretrained(args.model_name_or_path,
+                                        from_tf=bool('.ckpt' in args.model_name_or_path),
+                                        config=config,
+                                        cache_dir=args.cache_dir if args.cache_dir else None)
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    model.to(args.device)
+
+    logger.info("Training/evaluation parameters %s", args)
+
+    # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
+    # Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
+    # remove the need for this code, but it is still valid.
+    if args.fp16:
+        try:
+            import apex
+            apex.amp.register_half_function(torch, 'einsum')
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+
+    # Training
+    if args.do_train:
+        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+
+
+    # Save the trained model and the tokenizer
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+
+        # Good practice: save your training arguments together with the trained model
+        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
+
+        # Load a trained model and vocabulary that you have fine-tuned
+        model = model_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model.to(args.device)
+
+
+    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        checkpoints = [args.output_dir]
+        if args.eval_all_checkpoints:
+            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
+            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs
+
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+
+        for checkpoint in checkpoints:
+            # Reload the model
+            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
+            model = model_class.from_pretrained(checkpoint)
+            model.to(args.device)
+
+            # Evaluate
+            result = evaluate(args, model, tokenizer, prefix=global_step)
+
+            result = dict((k + ('_{}'.format(global_step) if global_step else ''), v) for k, v in result.items())
+            results.update(result)
+
+    logger.info("Results: {}".format(results))
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

templates/adding_a_new_example_script/utils_xxx.py

@@ -0,0 +1,995 @@
+
+# coding=utf-8
+# Copyright 2018 XXX.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Load XXX dataset. """
+
+from __future__ import absolute_import, division, print_function
+
+import json
+import logging
+import math
+import collections
+from io import open
+
+from transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize
+
+# Required by XLNet evaluation method to compute optimal threshold (see write_predictions_extended() method)
+from utils_squad_evaluate import find_all_best_thresh_v2, make_qid_to_has_ans, get_raw_scores
+
+logger = logging.getLogger(__name__)
+
+
+class SquadExample(object):
+    """
+    A single training/test example for the Squad dataset.
+    For examples without an answer, the start and end position are -1.
+    """
+
+    def __init__(self,
+                 qas_id,
+                 question_text,
+                 doc_tokens,
+                 orig_answer_text=None,
+                 start_position=None,
+                 end_position=None,
+                 is_impossible=None):
+        self.qas_id = qas_id
+        self.question_text = question_text
+        self.doc_tokens = doc_tokens
+        self.orig_answer_text = orig_answer_text
+        self.start_position = start_position
+        self.end_position = end_position
+        self.is_impossible = is_impossible
+
+    def __str__(self):
+        return self.__repr__()
+
+    def __repr__(self):
+        s = ""
+        s += "qas_id: %s" % (self.qas_id)
+        s += ", question_text: %s" % (
+            self.question_text)
+        s += ", doc_tokens: [%s]" % (" ".join(self.doc_tokens))
+        if self.start_position:
+            s += ", start_position: %d" % (self.start_position)
+        if self.end_position:
+            s += ", end_position: %d" % (self.end_position)
+        if self.is_impossible:
+            s += ", is_impossible: %r" % (self.is_impossible)
+        return s
+
+
+class InputFeatures(object):
+    """A single set of features of data."""
+
+    def __init__(self,
+                 unique_id,
+                 example_index,
+                 doc_span_index,
+                 tokens,
+                 token_to_orig_map,
+                 token_is_max_context,
+                 input_ids,
+                 input_mask,
+                 segment_ids,
+                 cls_index,
+                 p_mask,
+                 paragraph_len,
+                 start_position=None,
+                 end_position=None,
+                 is_impossible=None):
+        self.unique_id = unique_id
+        self.example_index = example_index
+        self.doc_span_index = doc_span_index
+        self.tokens = tokens
+        self.token_to_orig_map = token_to_orig_map
+        self.token_is_max_context = token_is_max_context
+        self.input_ids = input_ids
+        self.input_mask = input_mask
+        self.segment_ids = segment_ids
+        self.cls_index = cls_index
+        self.p_mask = p_mask
+        self.paragraph_len = paragraph_len
+        self.start_position = start_position
+        self.end_position = end_position
+        self.is_impossible = is_impossible
+
+
+def read_squad_examples(input_file, is_training, version_2_with_negative):
+    """Read a SQuAD json file into a list of SquadExample."""
+    with open(input_file, "r", encoding='utf-8') as reader:
+        input_data = json.load(reader)["data"]
+
+    def is_whitespace(c):
+        if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F:
+            return True
+        return False
+
+    examples = []
+    for entry in input_data:
+        for paragraph in entry["paragraphs"]:
+            paragraph_text = paragraph["context"]
+            doc_tokens = []
+            char_to_word_offset = []
+            prev_is_whitespace = True
+            for c in paragraph_text:
+                if is_whitespace(c):
+                    prev_is_whitespace = True
+                else:
+                    if prev_is_whitespace:
+                        doc_tokens.append(c)
+                    else:
+                        doc_tokens[-1] += c
+                    prev_is_whitespace = False
+                char_to_word_offset.append(len(doc_tokens) - 1)
+
+            for qa in paragraph["qas"]:
+                qas_id = qa["id"]
+                question_text = qa["question"]
+                start_position = None
+                end_position = None
+                orig_answer_text = None
+                is_impossible = False
+                if is_training:
+                    if version_2_with_negative:
+                        is_impossible = qa["is_impossible"]
+                    if (len(qa["answers"]) != 1) and (not is_impossible):
+                        raise ValueError(
+                            "For training, each question should have exactly 1 answer.")
+                    if not is_impossible:
+                        answer = qa["answers"][0]
+                        orig_answer_text = answer["text"]
+                        answer_offset = answer["answer_start"]
+                        answer_length = len(orig_answer_text)
+                        start_position = char_to_word_offset[answer_offset]
+                        end_position = char_to_word_offset[answer_offset + answer_length - 1]
+                        # Only add answers where the text can be exactly recovered from the
+                        # document. If this CAN'T happen it's likely due to weird Unicode
+                        # stuff so we will just skip the example.
+                        #
+                        # Note that this means for training mode, every example is NOT
+                        # guaranteed to be preserved.
+                        actual_text = " ".join(doc_tokens[start_position:(end_position + 1)])
+                        cleaned_answer_text = " ".join(
+                            whitespace_tokenize(orig_answer_text))
+                        if actual_text.find(cleaned_answer_text) == -1:
+                            logger.warning("Could not find answer: '%s' vs. '%s'",
+                                           actual_text, cleaned_answer_text)
+                            continue
+                    else:
+                        start_position = -1
+                        end_position = -1
+                        orig_answer_text = ""
+
+                example = SquadExample(
+                    qas_id=qas_id,
+                    question_text=question_text,
+                    doc_tokens=doc_tokens,
+                    orig_answer_text=orig_answer_text,
+                    start_position=start_position,
+                    end_position=end_position,
+                    is_impossible=is_impossible)
+                examples.append(example)
+    return examples
+
+
+def convert_examples_to_features(examples, tokenizer, max_seq_length,
+                                 doc_stride, max_query_length, is_training,
+                                 cls_token_at_end=False,
+                                 cls_token='[CLS]', sep_token='[SEP]', pad_token=0,
+                                 sequence_a_segment_id=0, sequence_b_segment_id=1,
+                                 cls_token_segment_id=0, pad_token_segment_id=0,
+                                 mask_padding_with_zero=True):
+    """Loads a data file into a list of `InputBatch`s."""
+
+    unique_id = 1000000000
+    # cnt_pos, cnt_neg = 0, 0
+    # max_N, max_M = 1024, 1024
+    # f = np.zeros((max_N, max_M), dtype=np.float32)
+
+    features = []
+    for (example_index, example) in enumerate(examples):
+
+        # if example_index % 100 == 0:
+        #     logger.info('Converting %s/%s pos %s neg %s', example_index, len(examples), cnt_pos, cnt_neg)
+
+        query_tokens = tokenizer.tokenize(example.question_text)
+
+        if len(query_tokens) > max_query_length:
+            query_tokens = query_tokens[0:max_query_length]
+
+        tok_to_orig_index = []
+        orig_to_tok_index = []
+        all_doc_tokens = []
+        for (i, token) in enumerate(example.doc_tokens):
+            orig_to_tok_index.append(len(all_doc_tokens))
+            sub_tokens = tokenizer.tokenize(token)
+            for sub_token in sub_tokens:
+                tok_to_orig_index.append(i)
+                all_doc_tokens.append(sub_token)
+
+        tok_start_position = None
+        tok_end_position = None
+        if is_training and example.is_impossible:
+            tok_start_position = -1
+            tok_end_position = -1
+        if is_training and not example.is_impossible:
+            tok_start_position = orig_to_tok_index[example.start_position]
+            if example.end_position < len(example.doc_tokens) - 1:
+                tok_end_position = orig_to_tok_index[example.end_position + 1] - 1
+            else:
+                tok_end_position = len(all_doc_tokens) - 1
+            (tok_start_position, tok_end_position) = _improve_answer_span(
+                all_doc_tokens, tok_start_position, tok_end_position, tokenizer,
+                example.orig_answer_text)
+
+        # The -3 accounts for [CLS], [SEP] and [SEP]
+        max_tokens_for_doc = max_seq_length - len(query_tokens) - 3
+
+        # We can have documents that are longer than the maximum sequence length.
+        # To deal with this we do a sliding window approach, where we take chunks
+        # of the up to our max length with a stride of `doc_stride`.
+        _DocSpan = collections.namedtuple(  # pylint: disable=invalid-name
+            "DocSpan", ["start", "length"])
+        doc_spans = []
+        start_offset = 0
+        while start_offset < len(all_doc_tokens):
+            length = len(all_doc_tokens) - start_offset
+            if length > max_tokens_for_doc:
+                length = max_tokens_for_doc
+            doc_spans.append(_DocSpan(start=start_offset, length=length))
+            if start_offset + length == len(all_doc_tokens):
+                break
+            start_offset += min(length, doc_stride)
+
+        for (doc_span_index, doc_span) in enumerate(doc_spans):
+            tokens = []
+            token_to_orig_map = {}
+            token_is_max_context = {}
+            segment_ids = []
+
+            # p_mask: mask with 1 for token than cannot be in the answer (0 for token which can be in an answer)
+            # Original TF implem also keep the classification token (set to 0) (not sure why...)
+            p_mask = []
+
+            # CLS token at the beginning
+            if not cls_token_at_end:
+                tokens.append(cls_token)
+                segment_ids.append(cls_token_segment_id)
+                p_mask.append(0)
+                cls_index = 0
+
+            # Query
+            for token in query_tokens:
+                tokens.append(token)
+                segment_ids.append(sequence_a_segment_id)
+                p_mask.append(1)
+
+            # SEP token
+            tokens.append(sep_token)
+            segment_ids.append(sequence_a_segment_id)
+            p_mask.append(1)
+
+            # Paragraph
+            for i in range(doc_span.length):
+                split_token_index = doc_span.start + i
+                token_to_orig_map[len(tokens)] = tok_to_orig_index[split_token_index]
+
+                is_max_context = _check_is_max_context(doc_spans, doc_span_index,
+                                                       split_token_index)
+                token_is_max_context[len(tokens)] = is_max_context
+                tokens.append(all_doc_tokens[split_token_index])
+                segment_ids.append(sequence_b_segment_id)
+                p_mask.append(0)
+            paragraph_len = doc_span.length
+
+            # SEP token
+            tokens.append(sep_token)
+            segment_ids.append(sequence_b_segment_id)
+            p_mask.append(1)
+
+            # CLS token at the end
+            if cls_token_at_end:
+                tokens.append(cls_token)
+                segment_ids.append(cls_token_segment_id)
+                p_mask.append(0)
+                cls_index = len(tokens) - 1  # Index of classification token
+
+            input_ids = tokenizer.convert_tokens_to_ids(tokens)
+
+            # The mask has 1 for real tokens and 0 for padding tokens. Only real
+            # tokens are attended to.
+            input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
+
+            # Zero-pad up to the sequence length.
+            while len(input_ids) < max_seq_length:
+                input_ids.append(pad_token)
+                input_mask.append(0 if mask_padding_with_zero else 1)
+                segment_ids.append(pad_token_segment_id)
+                p_mask.append(1)
+
+            assert len(input_ids) == max_seq_length
+            assert len(input_mask) == max_seq_length
+            assert len(segment_ids) == max_seq_length
+
+            span_is_impossible = example.is_impossible
+            start_position = None
+            end_position = None
+            if is_training and not span_is_impossible:
+                # For training, if our document chunk does not contain an annotation
+                # we throw it out, since there is nothing to predict.
+                doc_start = doc_span.start
+                doc_end = doc_span.start + doc_span.length - 1
+                out_of_span = False
+                if not (tok_start_position >= doc_start and
+                        tok_end_position <= doc_end):
+                    out_of_span = True
+                if out_of_span:
+                    start_position = 0
+                    end_position = 0
+                    span_is_impossible = True
+                else:
+                    doc_offset = len(query_tokens) + 2
+                    start_position = tok_start_position - doc_start + doc_offset
+                    end_position = tok_end_position - doc_start + doc_offset
+
+            if is_training and span_is_impossible:
+                start_position = cls_index
+                end_position = cls_index
+
+            if example_index < 20:
+                logger.info("*** Example ***")
+                logger.info("unique_id: %s" % (unique_id))
+                logger.info("example_index: %s" % (example_index))
+                logger.info("doc_span_index: %s" % (doc_span_index))
+                logger.info("tokens: %s" % " ".join(tokens))
+                logger.info("token_to_orig_map: %s" % " ".join([
+                    "%d:%d" % (x, y) for (x, y) in token_to_orig_map.items()]))
+                logger.info("token_is_max_context: %s" % " ".join([
+                    "%d:%s" % (x, y) for (x, y) in token_is_max_context.items()
+                ]))
+                logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
+                logger.info(
+                    "input_mask: %s" % " ".join([str(x) for x in input_mask]))
+                logger.info(
+                    "segment_ids: %s" % " ".join([str(x) for x in segment_ids]))
+                if is_training and span_is_impossible:
+                    logger.info("impossible example")
+                if is_training and not span_is_impossible:
+                    answer_text = " ".join(tokens[start_position:(end_position + 1)])
+                    logger.info("start_position: %d" % (start_position))
+                    logger.info("end_position: %d" % (end_position))
+                    logger.info(
+                        "answer: %s" % (answer_text))
+
+            features.append(
+                InputFeatures(
+                    unique_id=unique_id,
+                    example_index=example_index,
+                    doc_span_index=doc_span_index,
+                    tokens=tokens,
+                    token_to_orig_map=token_to_orig_map,
+                    token_is_max_context=token_is_max_context,
+                    input_ids=input_ids,
+                    input_mask=input_mask,
+                    segment_ids=segment_ids,
+                    cls_index=cls_index,
+                    p_mask=p_mask,
+                    paragraph_len=paragraph_len,
+                    start_position=start_position,
+                    end_position=end_position,
+                    is_impossible=span_is_impossible))
+            unique_id += 1
+
+    return features
+
+
+def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer,
+                         orig_answer_text):
+    """Returns tokenized answer spans that better match the annotated answer."""
+
+    # The SQuAD annotations are character based. We first project them to
+    # whitespace-tokenized words. But then after WordPiece tokenization, we can
+    # often find a "better match". For example:
+    #
+    #   Question: What year was John Smith born?
+    #   Context: The leader was John Smith (1895-1943).
+    #   Answer: 1895
+    #
+    # The original whitespace-tokenized answer will be "(1895-1943).". However
+    # after tokenization, our tokens will be "( 1895 - 1943 ) .". So we can match
+    # the exact answer, 1895.
+    #
+    # However, this is not always possible. Consider the following:
+    #
+    #   Question: What country is the top exporter of electornics?
+    #   Context: The Japanese electronics industry is the lagest in the world.
+    #   Answer: Japan
+    #
+    # In this case, the annotator chose "Japan" as a character sub-span of
+    # the word "Japanese". Since our WordPiece tokenizer does not split
+    # "Japanese", we just use "Japanese" as the annotation. This is fairly rare
+    # in SQuAD, but does happen.
+    tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text))
+
+    for new_start in range(input_start, input_end + 1):
+        for new_end in range(input_end, new_start - 1, -1):
+            text_span = " ".join(doc_tokens[new_start:(new_end + 1)])
+            if text_span == tok_answer_text:
+                return (new_start, new_end)
+
+    return (input_start, input_end)
+
+
+def _check_is_max_context(doc_spans, cur_span_index, position):
+    """Check if this is the 'max context' doc span for the token."""
+
+    # Because of the sliding window approach taken to scoring documents, a single
+    # token can appear in multiple documents. E.g.
+    #  Doc: the man went to the store and bought a gallon of milk
+    #  Span A: the man went to the
+    #  Span B: to the store and bought
+    #  Span C: and bought a gallon of
+    #  ...
+    #
+    # Now the word 'bought' will have two scores from spans B and C. We only
+    # want to consider the score with "maximum context", which we define as
+    # the *minimum* of its left and right context (the *sum* of left and
+    # right context will always be the same, of course).
+    #
+    # In the example the maximum context for 'bought' would be span C since
+    # it has 1 left context and 3 right context, while span B has 4 left context
+    # and 0 right context.
+    best_score = None
+    best_span_index = None
+    for (span_index, doc_span) in enumerate(doc_spans):
+        end = doc_span.start + doc_span.length - 1
+        if position < doc_span.start:
+            continue
+        if position > end:
+            continue
+        num_left_context = position - doc_span.start
+        num_right_context = end - position
+        score = min(num_left_context, num_right_context) + 0.01 * doc_span.length
+        if best_score is None or score > best_score:
+            best_score = score
+            best_span_index = span_index
+
+    return cur_span_index == best_span_index
+
+
+RawResult = collections.namedtuple("RawResult",
+                                   ["unique_id", "start_logits", "end_logits"])
+
+def write_predictions(all_examples, all_features, all_results, n_best_size,
+                      max_answer_length, do_lower_case, output_prediction_file,
+                      output_nbest_file, output_null_log_odds_file, verbose_logging,
+                      version_2_with_negative, null_score_diff_threshold):
+    """Write final predictions to the json file and log-odds of null if needed."""
+    logger.info("Writing predictions to: %s" % (output_prediction_file))
+    logger.info("Writing nbest to: %s" % (output_nbest_file))
+
+    example_index_to_features = collections.defaultdict(list)
+    for feature in all_features:
+        example_index_to_features[feature.example_index].append(feature)
+
+    unique_id_to_result = {}
+    for result in all_results:
+        unique_id_to_result[result.unique_id] = result
+
+    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "PrelimPrediction",
+        ["feature_index", "start_index", "end_index", "start_logit", "end_logit"])
+
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict()
+
+    for (example_index, example) in enumerate(all_examples):
+        features = example_index_to_features[example_index]
+
+        prelim_predictions = []
+        # keep track of the minimum score of null start+end of position 0
+        score_null = 1000000  # large and positive
+        min_null_feature_index = 0  # the paragraph slice with min null score
+        null_start_logit = 0  # the start logit at the slice with min null score
+        null_end_logit = 0  # the end logit at the slice with min null score
+        for (feature_index, feature) in enumerate(features):
+            result = unique_id_to_result[feature.unique_id]
+            start_indexes = _get_best_indexes(result.start_logits, n_best_size)
+            end_indexes = _get_best_indexes(result.end_logits, n_best_size)
+            # if we could have irrelevant answers, get the min score of irrelevant
+            if version_2_with_negative:
+                feature_null_score = result.start_logits[0] + result.end_logits[0]
+                if feature_null_score < score_null:
+                    score_null = feature_null_score
+                    min_null_feature_index = feature_index
+                    null_start_logit = result.start_logits[0]
+                    null_end_logit = result.end_logits[0]
+            for start_index in start_indexes:
+                for end_index in end_indexes:
+                    # We could hypothetically create invalid predictions, e.g., predict
+                    # that the start of the span is in the question. We throw out all
+                    # invalid predictions.
+                    if start_index >= len(feature.tokens):
+                        continue
+                    if end_index >= len(feature.tokens):
+                        continue
+                    if start_index not in feature.token_to_orig_map:
+                        continue
+                    if end_index not in feature.token_to_orig_map:
+                        continue
+                    if not feature.token_is_max_context.get(start_index, False):
+                        continue
+                    if end_index < start_index:
+                        continue
+                    length = end_index - start_index + 1
+                    if length > max_answer_length:
+                        continue
+                    prelim_predictions.append(
+                        _PrelimPrediction(
+                            feature_index=feature_index,
+                            start_index=start_index,
+                            end_index=end_index,
+                            start_logit=result.start_logits[start_index],
+                            end_logit=result.end_logits[end_index]))
+        if version_2_with_negative:
+            prelim_predictions.append(
+                _PrelimPrediction(
+                    feature_index=min_null_feature_index,
+                    start_index=0,
+                    end_index=0,
+                    start_logit=null_start_logit,
+                    end_logit=null_end_logit))
+        prelim_predictions = sorted(
+            prelim_predictions,
+            key=lambda x: (x.start_logit + x.end_logit),
+            reverse=True)
+
+        _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+            "NbestPrediction", ["text", "start_logit", "end_logit"])
+
+        seen_predictions = {}
+        nbest = []
+        for pred in prelim_predictions:
+            if len(nbest) >= n_best_size:
+                break
+            feature = features[pred.feature_index]
+            if pred.start_index > 0:  # this is a non-null prediction
+                tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
+                orig_doc_start = feature.token_to_orig_map[pred.start_index]
+                orig_doc_end = feature.token_to_orig_map[pred.end_index]
+                orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
+                tok_text = " ".join(tok_tokens)
+
+                # De-tokenize WordPieces that have been split off.
+                tok_text = tok_text.replace(" ##", "")
+                tok_text = tok_text.replace("##", "")
+
+                # Clean whitespace
+                tok_text = tok_text.strip()
+                tok_text = " ".join(tok_text.split())
+                orig_text = " ".join(orig_tokens)
+
+                final_text = get_final_text(tok_text, orig_text, do_lower_case, verbose_logging)
+                if final_text in seen_predictions:
+                    continue
+
+                seen_predictions[final_text] = True
+            else:
+                final_text = ""
+                seen_predictions[final_text] = True
+
+            nbest.append(
+                _NbestPrediction(
+                    text=final_text,
+                    start_logit=pred.start_logit,
+                    end_logit=pred.end_logit))
+        # if we didn't include the empty option in the n-best, include it
+        if version_2_with_negative:
+            if "" not in seen_predictions:
+                nbest.append(
+                    _NbestPrediction(
+                        text="",
+                        start_logit=null_start_logit,
+                        end_logit=null_end_logit))
+                
+            # In very rare edge cases we could only have single null prediction.
+            # So we just create a nonce prediction in this case to avoid failure.
+            if len(nbest)==1:
+                nbest.insert(0,
+                    _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
+
+        # In very rare edge cases we could have no valid predictions. So we
+        # just create a nonce prediction in this case to avoid failure.
+        if not nbest:
+            nbest.append(
+                _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
+
+        assert len(nbest) >= 1
+
+        total_scores = []
+        best_non_null_entry = None
+        for entry in nbest:
+            total_scores.append(entry.start_logit + entry.end_logit)
+            if not best_non_null_entry:
+                if entry.text:
+                    best_non_null_entry = entry
+
+        probs = _compute_softmax(total_scores)
+
+        nbest_json = []
+        for (i, entry) in enumerate(nbest):
+            output = collections.OrderedDict()
+            output["text"] = entry.text
+            output["probability"] = probs[i]
+            output["start_logit"] = entry.start_logit
+            output["end_logit"] = entry.end_logit
+            nbest_json.append(output)
+
+        assert len(nbest_json) >= 1
+
+        if not version_2_with_negative:
+            all_predictions[example.qas_id] = nbest_json[0]["text"]
+        else:
+            # predict "" iff the null score - the score of best non-null > threshold
+            score_diff = score_null - best_non_null_entry.start_logit - (
+                best_non_null_entry.end_logit)
+            scores_diff_json[example.qas_id] = score_diff
+            if score_diff > null_score_diff_threshold:
+                all_predictions[example.qas_id] = ""
+            else:
+                all_predictions[example.qas_id] = best_non_null_entry.text
+        all_nbest_json[example.qas_id] = nbest_json
+
+    with open(output_prediction_file, "w") as writer:
+        writer.write(json.dumps(all_predictions, indent=4) + "\n")
+
+    with open(output_nbest_file, "w") as writer:
+        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+
+    if version_2_with_negative:
+        with open(output_null_log_odds_file, "w") as writer:
+            writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    return all_predictions
+
+
+# For XLNet (and XLM which uses the same head)
+RawResultExtended = collections.namedtuple("RawResultExtended",
+    ["unique_id", "start_top_log_probs", "start_top_index",
+     "end_top_log_probs", "end_top_index", "cls_logits"])
+
+
+def write_predictions_extended(all_examples, all_features, all_results, n_best_size,
+                                max_answer_length, output_prediction_file,
+                                output_nbest_file,
+                                output_null_log_odds_file, orig_data_file,
+                                start_n_top, end_n_top, version_2_with_negative,
+                                tokenizer, verbose_logging):
+    """ XLNet write prediction logic (more complex than Bert's).
+        Write final predictions to the json file and log-odds of null if needed.
+
+        Requires utils_squad_evaluate.py
+    """
+    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "PrelimPrediction",
+        ["feature_index", "start_index", "end_index",
+        "start_log_prob", "end_log_prob"])
+
+    _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "NbestPrediction", ["text", "start_log_prob", "end_log_prob"])
+
+    logger.info("Writing predictions to: %s", output_prediction_file)
+    # logger.info("Writing nbest to: %s" % (output_nbest_file))
+
+    example_index_to_features = collections.defaultdict(list)
+    for feature in all_features:
+        example_index_to_features[feature.example_index].append(feature)
+
+    unique_id_to_result = {}
+    for result in all_results:
+        unique_id_to_result[result.unique_id] = result
+
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict()
+
+    for (example_index, example) in enumerate(all_examples):
+        features = example_index_to_features[example_index]
+
+        prelim_predictions = []
+        # keep track of the minimum score of null start+end of position 0
+        score_null = 1000000  # large and positive
+
+        for (feature_index, feature) in enumerate(features):
+            result = unique_id_to_result[feature.unique_id]
+
+            cur_null_score = result.cls_logits
+
+            # if we could have irrelevant answers, get the min score of irrelevant
+            score_null = min(score_null, cur_null_score)
+
+            for i in range(start_n_top):
+                for j in range(end_n_top):
+                    start_log_prob = result.start_top_log_probs[i]
+                    start_index = result.start_top_index[i]
+
+                    j_index = i * end_n_top + j
+
+                    end_log_prob = result.end_top_log_probs[j_index]
+                    end_index = result.end_top_index[j_index]
+
+                    # We could hypothetically create invalid predictions, e.g., predict
+                    # that the start of the span is in the question. We throw out all
+                    # invalid predictions.
+                    if start_index >= feature.paragraph_len - 1:
+                        continue
+                    if end_index >= feature.paragraph_len - 1:
+                        continue
+
+                    if not feature.token_is_max_context.get(start_index, False):
+                        continue
+                    if end_index < start_index:
+                        continue
+                    length = end_index - start_index + 1
+                    if length > max_answer_length:
+                        continue
+
+                    prelim_predictions.append(
+                        _PrelimPrediction(
+                            feature_index=feature_index,
+                            start_index=start_index,
+                            end_index=end_index,
+                            start_log_prob=start_log_prob,
+                            end_log_prob=end_log_prob))
+
+        prelim_predictions = sorted(
+            prelim_predictions,
+            key=lambda x: (x.start_log_prob + x.end_log_prob),
+            reverse=True)
+
+        seen_predictions = {}
+        nbest = []
+        for pred in prelim_predictions:
+            if len(nbest) >= n_best_size:
+                break
+            feature = features[pred.feature_index]
+
+            # XLNet un-tokenizer
+            # Let's keep it simple for now and see if we need all this later.
+            # 
+            # tok_start_to_orig_index = feature.tok_start_to_orig_index
+            # tok_end_to_orig_index = feature.tok_end_to_orig_index
+            # start_orig_pos = tok_start_to_orig_index[pred.start_index]
+            # end_orig_pos = tok_end_to_orig_index[pred.end_index]
+            # paragraph_text = example.paragraph_text
+            # final_text = paragraph_text[start_orig_pos: end_orig_pos + 1].strip()
+
+            # Previously used Bert untokenizer
+            tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
+            orig_doc_start = feature.token_to_orig_map[pred.start_index]
+            orig_doc_end = feature.token_to_orig_map[pred.end_index]
+            orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
+            tok_text = tokenizer.convert_tokens_to_string(tok_tokens)
+
+            # Clean whitespace
+            tok_text = tok_text.strip()
+            tok_text = " ".join(tok_text.split())
+            orig_text = " ".join(orig_tokens)
+
+            final_text = get_final_text(tok_text, orig_text, tokenizer.do_lower_case,
+                                        verbose_logging)
+
+            if final_text in seen_predictions:
+                continue
+
+            seen_predictions[final_text] = True
+
+            nbest.append(
+                _NbestPrediction(
+                    text=final_text,
+                    start_log_prob=pred.start_log_prob,
+                    end_log_prob=pred.end_log_prob))
+
+        # In very rare edge cases we could have no valid predictions. So we
+        # just create a nonce prediction in this case to avoid failure.
+        if not nbest:
+            nbest.append(
+                _NbestPrediction(text="", start_log_prob=-1e6,
+                end_log_prob=-1e6))
+
+        total_scores = []
+        best_non_null_entry = None
+        for entry in nbest:
+            total_scores.append(entry.start_log_prob + entry.end_log_prob)
+            if not best_non_null_entry:
+                best_non_null_entry = entry
+
+        probs = _compute_softmax(total_scores)
+
+        nbest_json = []
+        for (i, entry) in enumerate(nbest):
+            output = collections.OrderedDict()
+            output["text"] = entry.text
+            output["probability"] = probs[i]
+            output["start_log_prob"] = entry.start_log_prob
+            output["end_log_prob"] = entry.end_log_prob
+            nbest_json.append(output)
+
+        assert len(nbest_json) >= 1
+        assert best_non_null_entry is not None
+
+        score_diff = score_null
+        scores_diff_json[example.qas_id] = score_diff
+        # note(zhiliny): always predict best_non_null_entry
+        # and the evaluation script will search for the best threshold
+        all_predictions[example.qas_id] = best_non_null_entry.text
+
+        all_nbest_json[example.qas_id] = nbest_json
+
+    with open(output_prediction_file, "w") as writer:
+        writer.write(json.dumps(all_predictions, indent=4) + "\n")
+
+    with open(output_nbest_file, "w") as writer:
+        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+
+    if version_2_with_negative:
+        with open(output_null_log_odds_file, "w") as writer:
+            writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    with open(orig_data_file, "r", encoding='utf-8') as reader:
+        orig_data = json.load(reader)["data"]
+
+    qid_to_has_ans = make_qid_to_has_ans(orig_data)
+    has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
+    no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
+    exact_raw, f1_raw = get_raw_scores(orig_data, all_predictions)
+    out_eval = {}
+
+    find_all_best_thresh_v2(out_eval, all_predictions, exact_raw, f1_raw, scores_diff_json, qid_to_has_ans)
+
+    return out_eval
+
+
+def get_final_text(pred_text, orig_text, do_lower_case, verbose_logging=False):
+    """Project the tokenized prediction back to the original text."""
+
+    # When we created the data, we kept track of the alignment between original
+    # (whitespace tokenized) tokens and our WordPiece tokenized tokens. So
+    # now `orig_text` contains the span of our original text corresponding to the
+    # span that we predicted.
+    #
+    # However, `orig_text` may contain extra characters that we don't want in
+    # our prediction.
+    #
+    # For example, let's say:
+    #   pred_text = steve smith
+    #   orig_text = Steve Smith's
+    #
+    # We don't want to return `orig_text` because it contains the extra "'s".
+    #
+    # We don't want to return `pred_text` because it's already been normalized
+    # (the SQuAD eval script also does punctuation stripping/lower casing but
+    # our tokenizer does additional normalization like stripping accent
+    # characters).
+    #
+    # What we really want to return is "Steve Smith".
+    #
+    # Therefore, we have to apply a semi-complicated alignment heuristic between
+    # `pred_text` and `orig_text` to get a character-to-character alignment. This
+    # can fail in certain cases in which case we just return `orig_text`.
+
+    def _strip_spaces(text):
+        ns_chars = []
+        ns_to_s_map = collections.OrderedDict()
+        for (i, c) in enumerate(text):
+            if c == " ":
+                continue
+            ns_to_s_map[len(ns_chars)] = i
+            ns_chars.append(c)
+        ns_text = "".join(ns_chars)
+        return (ns_text, ns_to_s_map)
+
+    # We first tokenize `orig_text`, strip whitespace from the result
+    # and `pred_text`, and check if they are the same length. If they are
+    # NOT the same length, the heuristic has failed. If they are the same
+    # length, we assume the characters are one-to-one aligned.
+    tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
+
+    tok_text = " ".join(tokenizer.tokenize(orig_text))
+
+    start_position = tok_text.find(pred_text)
+    if start_position == -1:
+        if verbose_logging:
+            logger.info(
+                "Unable to find text: '%s' in '%s'" % (pred_text, orig_text))
+        return orig_text
+    end_position = start_position + len(pred_text) - 1
+
+    (orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text)
+    (tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text)
+
+    if len(orig_ns_text) != len(tok_ns_text):
+        if verbose_logging:
+            logger.info("Length not equal after stripping spaces: '%s' vs '%s'",
+                        orig_ns_text, tok_ns_text)
+        return orig_text
+
+    # We then project the characters in `pred_text` back to `orig_text` using
+    # the character-to-character alignment.
+    tok_s_to_ns_map = {}
+    for (i, tok_index) in tok_ns_to_s_map.items():
+        tok_s_to_ns_map[tok_index] = i
+
+    orig_start_position = None
+    if start_position in tok_s_to_ns_map:
+        ns_start_position = tok_s_to_ns_map[start_position]
+        if ns_start_position in orig_ns_to_s_map:
+            orig_start_position = orig_ns_to_s_map[ns_start_position]
+
+    if orig_start_position is None:
+        if verbose_logging:
+            logger.info("Couldn't map start position")
+        return orig_text
+
+    orig_end_position = None
+    if end_position in tok_s_to_ns_map:
+        ns_end_position = tok_s_to_ns_map[end_position]
+        if ns_end_position in orig_ns_to_s_map:
+            orig_end_position = orig_ns_to_s_map[ns_end_position]
+
+    if orig_end_position is None:
+        if verbose_logging:
+            logger.info("Couldn't map end position")
+        return orig_text
+
+    output_text = orig_text[orig_start_position:(orig_end_position + 1)]
+    return output_text
+
+
+def _get_best_indexes(logits, n_best_size):
+    """Get the n-best logits from a list."""
+    index_and_score = sorted(enumerate(logits), key=lambda x: x[1], reverse=True)
+
+    best_indexes = []
+    for i in range(len(index_and_score)):
+        if i >= n_best_size:
+            break
+        best_indexes.append(index_and_score[i][0])
+    return best_indexes
+
+
+def _compute_softmax(scores):
+    """Compute softmax probability over raw logits."""
+    if not scores:
+        return []
+
+    max_score = None
+    for score in scores:
+        if max_score is None or score > max_score:
+            max_score = score
+
+    exp_scores = []
+    total_sum = 0.0
+    for score in scores:
+        x = math.exp(score - max_score)
+        exp_scores.append(x)
+        total_sum += x
+
+    probs = []
+    for score in exp_scores:
+        probs.append(score / total_sum)
+    return probs

templates/adding_a_new_model/README.md

@@ -0,0 +1,62 @@
+# How to add a new model in 🤗Transformers
+
+This folder describes the process to add a new model in 🤗Transformers and provide templates for the required files.
+
+The library is designed to incorporate a variety of models and code bases. As such the process for adding a new model usually mostly consists in copy-pasting to relevant original code in the various sections of the templates included in the present repository.
+
+One important point though is that the library has the following goals impacting the way models are incorporated:
+
+- one specific feature of the API is the capability to run the model and tokenizer inline. The tokenization code thus often have to be slightly adapted to allow for running in the python interpreter.
+- the package is also designed to be as self-consistent and with a small and reliable set of packages dependencies. In consequence, additional dependencies are usually not allowed when adding a model but can be allowed for the inclusion of a new tokenizer (recent examples of dependencies added for tokenizer specificities include `sentencepiece` and `sacremoses`). Please make sure to check the existing dependencies when possible before adding a new one.
+
+For a quick overview of the library organization, please check the [QuickStart section of the documentation](https://huggingface.co/transformers/quickstart.html).
+
+# Typical workflow for including a model
+
+Here an overview of the general workflow: 
+
+- [ ] add model/configuration/tokenization classes
+- [ ] add conversion scripts
+- [ ] add tests
+- [ ] finalize
+
+Let's detail what should be done at each step
+
+## Adding model/configuration/tokenization classes
+
+Here is the workflow for adding model/configuration/tokenization classes:
+
+- [ ] copy the python files from the present folder to the main folder and rename them, replacing `xxx` with your model name,
+- [ ] edit the files to replace `XXX` (with various casing) with your model name
+- [ ] copy-paste or create a simple configuration class for your model in the `configuration_...` file
+- [ ] copy-paste or create the code for your model in the `modeling_...` files (PyTorch and TF 2.0)
+- [ ] copy-paste or create a tokenizer class for your model in the `tokenization_...` file
+
+# Adding conversion scripts
+
+Here is the workflow for the conversion scripts:
+
+- [ ] copy the conversion script (`convert_...`) from the present folder to the main folder.
+- [ ] edit this script to convert your original checkpoint weights to the current pytorch ones.
+
+# Adding tests:
+
+Here is the workflow for the adding tests:
+
+- [ ] copy the python files from the `tests` sub-folder of the present folder to the `tests` subfolder of the main folder and rename them, replacing `xxx` with your model name,
+- [ ] edit the tests files to replace `XXX` (with various casing) with your model name
+- [ ] edit the tests code as needed
+
+# Final steps
+
+You can then finish the addition step by adding imports for your classes in the common files:
+
+- [ ] add import for all the relevant classes in `__init__.py`
+- [ ] add your configuration in `configuration_auto.py`
+- [ ] add your PyTorch and TF 2.0 model respectively in `modeling_auto.py` and `modeling_tf_auto.py`
+- [ ] add your tokenizer in `tokenization_auto.py`
+- [ ] add your models and tokenizer to `pipeline.py`
+- [ ] add a link to your conversion script in the main conversion utility (currently in `__main__` but will be moved to the `commands` subfolder in the near future)
+- [ ] edit the PyTorch to TF 2.0 conversion script to add your model in the `convert_pytorch_checkpoint_to_tf2.py` file
+- [ ] add a mention of your model in the doc: `README.md` and the documentation itself at `docs/source/pretrained_models.rst`.
+- [ ] upload the pretrained weigths, configurations and vocabulary files.

templates/adding_a_new_model/configuration_xxx.py

@@ -0,0 +1,130 @@
+# coding=utf-8
+# Copyright 2010, XXX authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" XXX model configuration """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import sys
+import six
+from io import open
+
+from .configuration_utils import PretrainedConfig
+
+logger = logging.getLogger(__name__)
+
+XXX_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    'xxx-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-base-uncased-config.json",
+    'xxx-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-large-uncased-config.json",
+}
+
+
+class XxxConfig(PretrainedConfig):
+    r"""
+        :class:`~transformers.XxxConfig` is the configuration class to store the configuration of a
+        `XxxModel`.
+
+
+        Arguments:
+            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `XxxModel`.
+            hidden_size: Size of the encoder layers and the pooler layer.
+            num_hidden_layers: Number of hidden layers in the Transformer encoder.
+            num_attention_heads: Number of attention heads for each attention layer in
+                the Transformer encoder.
+            intermediate_size: The size of the "intermediate" (i.e., feed-forward)
+                layer in the Transformer encoder.
+            hidden_act: The non-linear activation function (function or string) in the
+                encoder and pooler. If string, "gelu", "relu", "swish" and "gelu_new" are supported.
+            hidden_dropout_prob: The dropout probabilitiy for all fully connected
+                layers in the embeddings, encoder, and pooler.
+            attention_probs_dropout_prob: The dropout ratio for the attention
+                probabilities.
+            max_position_embeddings: The maximum sequence length that this model might
+                ever be used with. Typically set this to something large just in case
+                (e.g., 512 or 1024 or 2048).
+            type_vocab_size: The vocabulary size of the `token_type_ids` passed into
+                `XxxModel`.
+            initializer_range: The sttdev of the truncated_normal_initializer for
+                initializing all weight matrices.
+            layer_norm_eps: The epsilon used by LayerNorm.
+    """
+    pretrained_config_archive_map = XXX_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+    def __init__(self,
+                 vocab_size_or_config_json_file=50257,
+                 n_positions=1024,
+                 n_ctx=1024,
+                 n_embd=768,
+                 n_layer=12,
+                 n_head=12,
+                 resid_pdrop=0.1,
+                 embd_pdrop=0.1,
+                 attn_pdrop=0.1,
+                 layer_norm_epsilon=1e-5,
+                 initializer_range=0.02,
+
+                 num_labels=1,
+                 summary_type='cls_index',
+                 summary_use_proj=True,
+                 summary_activation=None,
+                 summary_proj_to_labels=True,
+                 summary_first_dropout=0.1,
+                 **kwargs):
+        super(XxxConfig, self).__init__(**kwargs)
+        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, six.string_types) else -1
+        self.n_ctx = n_ctx
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+
+        self.num_labels = num_labels
+        self.summary_type = summary_type
+        self.summary_use_proj = summary_use_proj
+        self.summary_activation = summary_activation
+        self.summary_first_dropout = summary_first_dropout
+        self.summary_proj_to_labels = summary_proj_to_labels
+        if isinstance(vocab_size_or_config_json_file, six.string_types):
+            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
+                json_config = json.loads(reader.read())
+            for key, value in json_config.items():
+                self.__dict__[key] = value
+        elif not isinstance(vocab_size_or_config_json_file, int):
+            raise ValueError(
+                "First argument must be either a vocabulary size (int)"
+                "or the path to a pretrained model config file (str)"
+            )
+
+    @property
+    def max_position_embeddings(self):
+        return self.n_positions
+
+    @property
+    def hidden_size(self):
+        return self.n_embd
+
+    @property
+    def num_attention_heads(self):
+        return self.n_head
+
+    @property
+    def num_hidden_layers(self):
+        return self.n_layer

templates/adding_a_new_model/convert_xxx_original_tf_checkpoint_to_pytorch.py

@@ -0,0 +1,65 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert XXX checkpoint."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import torch
+
+from transformers import XxxConfig, XxxForPreTraining, load_tf_weights_in_xxx
+
+import logging
+logging.basicConfig(level=logging.INFO)
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, xxx_config_file, pytorch_dump_path):
+    # Initialise PyTorch model
+    config = XxxConfig.from_json_file(xxx_config_file)
+    print("Building PyTorch model from configuration: {}".format(str(config)))
+    model = XxxForPreTraining(config)
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_xxx(model, config, tf_checkpoint_path)
+
+    # Save pytorch-model
+    print("Save PyTorch model to {}".format(pytorch_dump_path))
+    torch.save(model.state_dict(), pytorch_dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    ## Required parameters
+    parser.add_argument("--tf_checkpoint_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the TensorFlow checkpoint path.")
+    parser.add_argument("--xxx_config_file",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "The config json file corresponding to the pre-trained XXX model. \n"
+                            "This specifies the model architecture.")
+    parser.add_argument("--pytorch_dump_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the output PyTorch model.")
+    args = parser.parse_args()
+    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path,
+                                     args.xxx_config_file,
+                                     args.pytorch_dump_path)

templates/adding_a_new_model/modeling_tf_xxx.py

@@ -0,0 +1,504 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" TF 2.0 XXX model. """
+
+####################################################
+# In this template, replace all the XXX (various casings) with your model name
+####################################################
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import math
+import os
+import sys
+from io import open
+
+import numpy as np
+import tensorflow as tf
+
+from .configuration_xxx import XxxConfig
+from .modeling_tf_utils import TFPreTrainedModel, get_initializer
+from .file_utils import add_start_docstrings
+
+logger = logging.getLogger(__name__)
+
+####################################################
+# This dict contrains shortcut names and associated url
+# for the pretrained weights provided with the models
+####################################################
+TF_XXX_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'xxx-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-base-uncased-tf_model.h5",
+    'xxx-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-large-uncased-tf_model.h5",
+}
+
+####################################################
+# TF 2.0 Models are constructed using Keras imperative API by sub-classing
+# - tf.keras.layers.Layer for the layers and
+# - TFPreTrainedModel for the models (itself a sub-class of tf.keras.Model)
+####################################################
+
+####################################################
+# Here is an example of typical layer in a TF 2.0 model of the library
+# The classes are usually identical to the PyTorch ones and prefixed with 'TF'.
+#
+# Note that class __init__ parameters includes **kwargs (send to 'super').
+# This let us have a control on class scope and variable names:
+# More precisely, we set the names of the class attributes (lower level layers) to
+# to the equivalent attributes names in the PyTorch model so we can have equivalent
+# class and scope structure between PyTorch and TF 2.0 models and easily load one in the other.
+#
+# See the conversion methods in modeling_tf_pytorch_utils.py for more details
+####################################################
+class TFXxxLayer(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFXxxLayer, self).__init__(**kwargs)
+        self.attention = TFXxxAttention(config, name='attention')
+        self.intermediate = TFXxxIntermediate(config, name='intermediate')
+        self.transformer_output = TFXxxOutput(config, name='output')
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        attention_outputs = self.attention([hidden_states, attention_mask, head_mask], training=training)
+        attention_output = attention_outputs[0]
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.transformer_output([intermediate_output, attention_output], training=training)
+        outputs = (layer_output,) + attention_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+####################################################
+# The full model without a specific pretrained or finetuning head is
+# provided as a tf.keras.layers.Layer usually called "TFXxxMainLayer"
+####################################################
+class TFXxxMainLayer(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFXxxMainLayer, self).__init__(**kwargs)
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        raise NotImplementedError  # Not implemented yet in the library fr TF 2.0 models
+
+    def _prune_heads(self, heads_to_prune):
+        raise NotImplementedError  # Not implemented yet in the library fr TF 2.0 models
+
+    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, training=False):
+        # We allow three types of multi-inputs:
+        # - traditional keyword arguments in the call method
+        # - all the arguments provided as a dict in the first positional argument of call
+        # - all the arguments provided as a list/tuple (ordered) in the first positional argument of call
+        # The last two options are useful to use the tf.keras fit() method.
+
+        if isinstance(inputs, (tuple, list)):
+            input_ids = inputs[0]
+            attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
+            token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
+            position_ids = inputs[3] if len(inputs) > 3 else position_ids
+            head_mask = inputs[4] if len(inputs) > 4 else head_mask
+            assert len(inputs) <= 5, "Too many inputs."
+        elif isinstance(inputs, dict):
+            input_ids = inputs.get('input_ids')
+            attention_mask = inputs.get('attention_mask', attention_mask)
+            token_type_ids = inputs.get('token_type_ids', token_type_ids)
+            position_ids = inputs.get('position_ids', position_ids)
+            head_mask = inputs.get('head_mask', head_mask)
+            assert len(inputs) <= 5, "Too many inputs."
+        else:
+            input_ids = inputs
+
+        if attention_mask is None:
+            attention_mask = tf.fill(tf.shape(input_ids), 1)
+        if token_type_ids is None:
+            token_type_ids = tf.fill(tf.shape(input_ids), 0)
+
+        # We create a 3D attention mask from a 2D tensor mask.
+        # Sizes are [batch_size, 1, 1, to_seq_length]
+        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+        # this attention mask is more simple than the triangular masking of causal attention
+        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+        extended_attention_mask = attention_mask[:, tf.newaxis, tf.newaxis, :]
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+
+        extended_attention_mask = tf.cast(extended_attention_mask, tf.float32)
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if not head_mask is None:
+            raise NotImplementedError
+        else:
+            head_mask = [None] * self.num_hidden_layers
+            # head_mask = tf.constant([0] * self.num_hidden_layers)
+
+        ##################################
+        # Replace this with your model code
+        embedding_output = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids)
+        encoder_outputs = self.encoder([embedding_output, extended_attention_mask, head_mask], training=training)
+        sequence_output = encoder_outputs[0]
+        outputs = (sequence_output,) + encoder_outputs[1:]  # add hidden_states and attentions if they are here
+
+        return outputs  # sequence_output, (hidden_states), (attentions)
+
+
+####################################################
+# TFXxxPreTrainedModel is a sub-class of tf.keras.Model
+# which take care of loading and saving pretrained weights
+# and various common utilities.
+# Here you just need to specify a few (self-explanatory)
+# pointers for your model.
+####################################################
+class TFXxxPreTrainedModel(TFPreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = XxxConfig
+    pretrained_model_archive_map = TF_XXX_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "transformer"
+
+
+XXX_START_DOCSTRING = r"""    The XXX model was proposed in
+    `XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`_
+    by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova. It's a bidirectional transformer
+    pre-trained using a combination of masked language modeling objective and next sentence prediction
+    on a large corpus comprising the Toronto Book Corpus and Wikipedia.
+
+    This model is a tf.keras.Model `tf.keras.Model`_ sub-class. Use it as a regular TF 2.0 Keras Model and
+    refer to the TF 2.0 documentation for all matter related to general usage and behavior.
+
+    .. _`XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`:
+        https://arxiv.org/abs/1810.04805
+
+    .. _`tf.keras.Model`:
+        https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/Model
+
+    Note on the model inputs:
+        TF 2.0 models accepts two formats as inputs:
+
+            - having all inputs as keyword arguments (like PyTorch models), or
+            - having all inputs as a list, tuple or dict in the first positional arguments.
+
+        This second option is usefull when using `tf.keras.Model.fit()` method which currently requires having all the tensors in the first argument of the model call function: `model(inputs)`.
+
+        If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument :
+
+        - a single Tensor with input_ids only and nothing else: `model(inputs_ids)
+        - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
+            `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
+        - a dictionary with one or several input Tensors associaed to the input names given in the docstring:
+            `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
+
+    Parameters:
+        config (:class:`~transformers.XxxConfig`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+XXX_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, XXX input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+                
+                ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+                
+                ``token_type_ids:   0   0   0   0  0     0   0``
+
+            Xxx is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
+            Indices can be obtained using :class:`transformers.XxxTokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **token_type_ids**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Segment token indices to indicate first and second portions of the inputs.
+            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
+            corresponds to a `sentence B` token
+            (see `XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
+        **position_ids**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
+        **head_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+"""
+
+@add_start_docstrings("The bare Xxx Model transformer outputing raw hidden-states without any specific head on top.",
+                      XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
+class TFXxxModel(TFXxxPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``tf.Tensor`` of shape ``(batch_size, hidden_size)``
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during Xxx pretraining. This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import XxxTokenizer, TFXxxModel
+
+        tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
+        model = TFXxxModel.from_pretrained('xxx-base-uncased')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFXxxModel, self).__init__(config, *inputs, **kwargs)
+        self.transformer = TFXxxMainLayer(config, name='transformer')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.transformer(inputs, **kwargs)
+        return outputs
+
+
+@add_start_docstrings("""Xxx Model with a `language modeling` head on top. """,
+    XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
+class TFXxxForMaskedLM(TFXxxPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **prediction_scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import XxxTokenizer, TFXxxForMaskedLM
+
+        tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
+        model = TFXxxForMaskedLM.from_pretrained('xxx-base-uncased')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        prediction_scores = outputs[0]
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFXxxForMaskedLM, self).__init__(config, *inputs, **kwargs)
+
+        self.transformer = TFXxxMainLayer(config, name='transformer')
+        self.mlm = TFXxxMLMHead(config, self.transformer.embeddings, name='mlm')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.transformer(inputs, **kwargs)
+
+        sequence_output = outputs[0]
+        prediction_scores = self.mlm(sequence_output, training=kwargs.get('training', False))
+
+        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+
+        return outputs  # prediction_scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Xxx Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
+class TFXxxForSequenceClassification(TFXxxPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **logits**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import XxxTokenizer, TFXxxForSequenceClassification
+
+        tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
+        model = TFXxxForSequenceClassification.from_pretrained('xxx-base-uncased')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        logits = outputs[0]
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFXxxForSequenceClassification, self).__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
+        self.transformer = TFXxxMainLayer(config, name='transformer')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(config.num_labels,
+                                                kernel_initializer=get_initializer(config.initializer_range),
+                                                name='classifier')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.transformer(inputs, **kwargs)
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output, training=kwargs.get('training', False))
+        logits = self.classifier(pooled_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        return outputs  # logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Xxx Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
+class TFXxxForTokenClassification(TFXxxPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import XxxTokenizer, TFXxxForTokenClassification
+
+        tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
+        model = TFXxxForTokenClassification.from_pretrained('xxx-base-uncased')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        scores = outputs[0]
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFXxxForTokenClassification, self).__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
+        self.transformer = TFXxxMainLayer(config, name='transformer')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(config.num_labels,
+                                                kernel_initializer=get_initializer(config.initializer_range),
+                                                name='classifier')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.transformer(inputs, **kwargs)
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output, training=kwargs.get('training', False))
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        return outputs  # scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Xxx Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
+    the hidden-states output to compute `span start logits` and `span end logits`). """,
+    XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
+class TFXxxForQuestionAnswering(TFXxxPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **start_scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length,)``
+            Span-start scores (before SoftMax).
+        **end_scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length,)``
+            Span-end scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import XxxTokenizer, TFXxxForQuestionAnswering
+
+        tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
+        model = TFXxxForQuestionAnswering.from_pretrained('xxx-base-uncased')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        start_scores, end_scores = outputs[:2]
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFXxxForQuestionAnswering, self).__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
+        self.transformer = TFXxxMainLayer(config, name='transformer')
+        self.qa_outputs = tf.keras.layers.Dense(config.num_labels,
+                                                kernel_initializer=get_initializer(config.initializer_range),
+                                                name='qa_outputs')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.transformer(inputs, **kwargs)
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = tf.split(logits, 2, axis=-1)
+        start_logits = tf.squeeze(start_logits, axis=-1)
+        end_logits = tf.squeeze(end_logits, axis=-1)
+
+        outputs = (start_logits, end_logits,) + outputs[2:]
+
+        return outputs  # start_logits, end_logits, (hidden_states), (attentions)

templates/adding_a_new_model/modeling_xxx.py

@@ -0,0 +1,658 @@
+# coding=utf-8
+# Copyright 2018 XXX Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch XXX model. """
+
+####################################################
+# In this template, replace all the XXX (various casings) with your model name
+####################################################
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import math
+import os
+import sys
+from io import open
+
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from .modeling_utils import PreTrainedModel, prune_linear_layer
+from .configuration_xxx import XxxConfig
+from .file_utils import add_start_docstrings
+
+logger = logging.getLogger(__name__)
+
+####################################################
+# This dict contrains shortcut names and associated url
+# for the pretrained weights provided with the models
+####################################################
+XXX_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'xxx-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-base-uncased-pytorch_model.bin",
+    'xxx-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-large-uncased-pytorch_model.bin",
+}
+
+####################################################
+# This is a conversion method from TF 1.0 to PyTorch
+# More details: https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28
+####################################################
+def load_tf_weights_in_xxx(model, config, tf_checkpoint_path):
+    """ Load tf checkpoints in a pytorch model.
+    """
+    try:
+        import re
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions.")
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        name = name.split('/')
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(n in ["adam_v", "adam_m", "global_step"] for n in name):
+            logger.info("Skipping {}".format("/".join(name)))
+            continue
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
+                l = re.split(r'_(\d+)', m_name)
+            else:
+                l = [m_name]
+            if l[0] == 'kernel' or l[0] == 'gamma':
+                pointer = getattr(pointer, 'weight')
+            elif l[0] == 'output_bias' or l[0] == 'beta':
+                pointer = getattr(pointer, 'bias')
+            elif l[0] == 'output_weights':
+                pointer = getattr(pointer, 'weight')
+            elif l[0] == 'squad':
+                pointer = getattr(pointer, 'classifier')
+            else:
+                try:
+                    pointer = getattr(pointer, l[0])
+                except AttributeError:
+                    logger.info("Skipping {}".format("/".join(name)))
+                    continue
+            if len(l) >= 2:
+                num = int(l[1])
+                pointer = pointer[num]
+        if m_name[-11:] == '_embeddings':
+            pointer = getattr(pointer, 'weight')
+        elif m_name == 'kernel':
+            array = np.transpose(array)
+        try:
+            assert pointer.shape == array.shape
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info("Initialize PyTorch weight {}".format(name))
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+####################################################
+# PyTorch Models are constructed by sub-classing
+# - torch.nn.Module for the layers and
+# - PreTrainedModel for the models (itself a sub-class of torch.nn.Module)
+####################################################
+
+####################################################
+# Here is an example of typical layer in a PyTorch model of the library
+# The classes are usually identical to the TF 2.0 ones without the 'TF' prefix.
+#
+# See the conversion methods in modeling_tf_pytorch_utils.py for more details
+####################################################
+class XxxLayer(nn.Module):
+    def __init__(self, config):
+        super(XxxLayer, self).__init__()
+        self.attention = XxxAttention(config)
+        self.intermediate = XxxIntermediate(config)
+        self.output = XxxOutput(config)
+
+    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+        attention_outputs = self.attention(hidden_states, attention_mask, head_mask)
+        attention_output = attention_outputs[0]
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        outputs = (layer_output,) + attention_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+
+####################################################
+# PreTrainedModel is a sub-class of torch.nn.Module
+# which take care of loading and saving pretrained weights
+# and various common utilities.
+#
+# Here you just need to specify a few (self-explanatory)
+# pointers for your model and the weights initialization
+# method if its not fully covered by PreTrainedModel's default method
+####################################################
+class XxxPreTrainedModel(PreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = XxxConfig
+    pretrained_model_archive_map = XXX_PRETRAINED_MODEL_ARCHIVE_MAP
+    load_tf_weights = load_tf_weights_in_xxx
+    base_model_prefix = "transformer"
+
+    def _init_weights(self, module):
+        """ Initialize the weights """
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, XxxLayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+XXX_START_DOCSTRING = r"""    The XXX model was proposed in
+    `XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`_
+    by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova. It's a bidirectional transformer
+    pre-trained using a combination of masked language modeling objective and next sentence prediction
+    on a large corpus comprising the Toronto Book Corpus and Wikipedia.
+
+    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+    refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+    .. _`XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`:
+        https://arxiv.org/abs/1810.04805
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~transformers.XxxConfig`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+XXX_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, XXX input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+                
+                ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+                
+                ``token_type_ids:   0   0   0   0  0     0   0``
+
+            Xxx is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
+            Indices can be obtained using :class:`transformers.XxxTokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Segment token indices to indicate first and second portions of the inputs.
+            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
+            corresponds to a `sentence B` token
+            (see `XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
+        **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+"""
+
+@add_start_docstrings("The bare Xxx Model transformer outputting raw hidden-states without any specific head on top.",
+                      XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
+class XxxModel(XxxPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during Xxx pretraining. This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
+        model = XxxModel.from_pretrained('xxx-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config):
+        super(XxxModel, self).__init__(config)
+
+        self.embeddings = XxxEmbeddings(config)
+        self.encoder = XxxEncoder(config)
+        self.pooler = XxxPooler(config)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings.word_embeddings = new_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None):
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We create a 3D attention mask from a 2D tensor mask.
+        # Sizes are [batch_size, 1, 1, to_seq_length]
+        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+        # this attention mask is more simple than the triangular masking of causal attention
+        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.num_hidden_layers
+
+        ##################################
+        # Replace this with your model code
+        embedding_output = self.embeddings(input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds)
+        encoder_outputs = self.encoder(embedding_output, extended_attention_mask, head_mask=head_mask)
+        sequence_output = encoder_outputs[0]
+        outputs = (sequence_output,) + encoder_outputs[1:]  # add hidden_states and attentions if they are here
+
+        return outputs  # sequence_output, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Xxx Model with a `language modeling` head on top. """,
+    XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
+class XxxForMaskedLM(XxxPreTrainedModel):
+    r"""
+        **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
+        model = XxxForMaskedLM.from_pretrained('xxx-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, masked_lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(XxxForMaskedLM, self).__init__(config)
+
+        self.transformer = XxxModel(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
+                masked_lm_labels=None):
+
+        outputs = self.transformer(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+
+        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+        if masked_lm_labels is not None:
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
+            outputs = (masked_lm_loss,) + outputs
+
+        return outputs  # (masked_lm_loss), prediction_scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Xxx Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
+class XxxForSequenceClassification(XxxPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
+            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification (or regression if config.num_labels==1) loss.
+        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
+        model = XxxForSequenceClassification.from_pretrained('xxx-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(XxxForSequenceClassification, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.transformer = XxxModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
+
+        outputs = self.transformer(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Xxx Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
+class XxxForTokenClassification(XxxPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the token classification loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss.
+        **scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
+        model = XxxForTokenClassification.from_pretrained('xxx-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(XxxForTokenClassification, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.transformer = XxxModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
+
+        outputs = self.transformer(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)[active_loss]
+                active_labels = labels.view(-1)[active_loss]
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Xxx Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
+    the hidden-states output to compute `span start logits` and `span end logits`). """,
+    XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
+class XxxForQuestionAnswering(XxxPreTrainedModel):
+    r"""
+        **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+        **end_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        **start_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-start scores (before SoftMax).
+        **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-end scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
+        model = XxxForQuestionAnswering.from_pretrained('xxx-large-uncased-whole-word-masking-finetuned-squad')
+        question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        input_text = "[CLS] " + question + " [SEP] " + text + " [SEP]"
+        input_ids = tokenizer.encode(input_text)
+        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))] 
+        start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))
+        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)  
+        print(' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]))
+        # a nice puppet
+
+
+    """
+    def __init__(self, config):
+        super(XxxForQuestionAnswering, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.transformer = XxxModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
+                start_positions=None, end_positions=None):
+
+        outputs = self.transformer(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        outputs = (start_logits, end_logits,) + outputs[2:]
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions.clamp_(0, ignored_index)
+            end_positions.clamp_(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+            outputs = (total_loss,) + outputs
+
+        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)

templates/adding_a_new_model/tests/modeling_tf_xxx_test.py

@@ -0,0 +1,256 @@
+# coding=utf-8
+# Copyright 2018 XXX Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import pytest
+import sys
+
+from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
+from .configuration_common_test import ConfigTester
+
+from transformers import XxxConfig, is_tf_available
+
+if is_tf_available():
+    import tensorflow as tf
+    from transformers.modeling_tf_xxx import (TFXxxModel, TFXxxForMaskedLM,
+                                               TFXxxForSequenceClassification,
+                                               TFXxxForTokenClassification,
+                                               TFXxxForQuestionAnswering,
+                                               TF_XXX_PRETRAINED_MODEL_ARCHIVE_MAP)
+else:
+    pytestmark = pytest.mark.skip("Require TensorFlow")
+
+
+class TFXxxModelTest(TFCommonTestCases.TFCommonModelTester):
+
+    all_model_classes = (TFXxxModel, TFXxxForMaskedLM, TFXxxForQuestionAnswering,
+                         TFXxxForSequenceClassification,
+                         TFXxxForTokenClassification) if is_tf_available() else ()
+
+    class TFXxxModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_input_mask=True,
+                     use_token_type_ids=True,
+                     use_labels=True,
+                     vocab_size=99,
+                     hidden_size=32,
+                     num_hidden_layers=5,
+                     num_attention_heads=4,
+                     intermediate_size=37,
+                     hidden_act="gelu",
+                     hidden_dropout_prob=0.1,
+                     attention_probs_dropout_prob=0.1,
+                     max_position_embeddings=512,
+                     type_vocab_size=16,
+                     type_sequence_label_size=2,
+                     initializer_range=0.02,
+                     num_labels=3,
+                     num_choices=4,
+                     scope=None,
+                    ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_input_mask = use_input_mask
+            self.use_token_type_ids = use_token_type_ids
+            self.use_labels = use_labels
+            self.vocab_size = vocab_size
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.intermediate_size = intermediate_size
+            self.hidden_act = hidden_act
+            self.hidden_dropout_prob = hidden_dropout_prob
+            self.attention_probs_dropout_prob = attention_probs_dropout_prob
+            self.max_position_embeddings = max_position_embeddings
+            self.type_vocab_size = type_vocab_size
+            self.type_sequence_label_size = type_sequence_label_size
+            self.initializer_range = initializer_range
+            self.num_labels = num_labels
+            self.num_choices = num_choices
+            self.scope = scope
+
+        def prepare_config_and_inputs(self):
+            input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+            input_mask = None
+            if self.use_input_mask:
+                input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+            token_type_ids = None
+            if self.use_token_type_ids:
+                token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+            sequence_labels = None
+            token_labels = None
+            choice_labels = None
+            if self.use_labels:
+                sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+                choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+            config = XxxConfig(
+                vocab_size_or_config_json_file=self.vocab_size,
+                hidden_size=self.hidden_size,
+                num_hidden_layers=self.num_hidden_layers,
+                num_attention_heads=self.num_attention_heads,
+                intermediate_size=self.intermediate_size,
+                hidden_act=self.hidden_act,
+                hidden_dropout_prob=self.hidden_dropout_prob,
+                attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+                max_position_embeddings=self.max_position_embeddings,
+                type_vocab_size=self.type_vocab_size,
+                initializer_range=self.initializer_range)
+
+            return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+        def create_and_check_xxx_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = TFXxxModel(config=config)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            sequence_output, pooled_output = model(inputs)
+
+            inputs = [input_ids, input_mask]
+            sequence_output, pooled_output = model(inputs)
+
+            sequence_output, pooled_output = model(input_ids)
+
+            result = {
+                "sequence_output": sequence_output.numpy(),
+                "pooled_output": pooled_output.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["sequence_output"].shape),
+                [self.batch_size, self.seq_length, self.hidden_size])
+            self.parent.assertListEqual(list(result["pooled_output"].shape), [self.batch_size, self.hidden_size])
+
+
+        def create_and_check_xxx_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = TFXxxForMaskedLM(config=config)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            prediction_scores, = model(inputs)
+            result = {
+                "prediction_scores": prediction_scores.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["prediction_scores"].shape),
+                [self.batch_size, self.seq_length, self.vocab_size])
+
+
+        def create_and_check_xxx_for_sequence_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            config.num_labels = self.num_labels
+            model = TFXxxForSequenceClassification(config=config)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            logits, = model(inputs)
+            result = {
+                "logits": logits.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].shape),
+                [self.batch_size, self.num_labels])
+
+
+        def create_and_check_xxx_for_token_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            config.num_labels = self.num_labels
+            model = TFXxxForTokenClassification(config=config)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            logits, = model(inputs)
+            result = {
+                "logits": logits.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].shape),
+                [self.batch_size, self.seq_length, self.num_labels])
+
+
+        def create_and_check_xxx_for_question_answering(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = TFXxxForQuestionAnswering(config=config)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            start_logits, end_logits = model(inputs)
+            result = {
+                "start_logits": start_logits.numpy(),
+                "end_logits": end_logits.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["start_logits"].shape),
+                [self.batch_size, self.seq_length])
+            self.parent.assertListEqual(
+                list(result["end_logits"].shape),
+                [self.batch_size, self.seq_length])
+
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids, token_type_ids, input_mask,
+             sequence_labels, token_labels, choice_labels) = config_and_inputs
+            inputs_dict = {'input_ids': input_ids, 'token_type_ids': token_type_ids, 'attention_mask': input_mask}
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = TFXxxModelTest.TFXxxModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=XxxConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_xxx_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xxx_model(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xxx_for_masked_lm(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xxx_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xxx_for_sequence_classification(*config_and_inputs)
+
+    def test_for_token_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xxx_for_token_classification(*config_and_inputs)
+
+    @pytest.mark.slow
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/transformers_test/"
+        for model_name in ['xxx-base-uncased']:
+            model = TFXxxModel.from_pretrained(model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
+if __name__ == "__main__":
+    unittest.main()

templates/adding_a_new_model/tests/modeling_xxx_test.py

@@ -0,0 +1,255 @@
+# coding=utf-8
+# Copyright 2018 XXX Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import pytest
+
+from transformers import is_torch_available
+
+from .modeling_common_test import (CommonTestCases, ids_tensor)
+from .configuration_common_test import ConfigTester
+
+if is_torch_available():
+    from transformers import (XxxConfig, XxxModel, XxxForMaskedLM,
+                                        XxxForNextSentencePrediction, XxxForPreTraining,
+                                        XxxForQuestionAnswering, XxxForSequenceClassification,
+                                        XxxForTokenClassification, XxxForMultipleChoice)
+    from transformers.modeling_xxx import XXX_PRETRAINED_MODEL_ARCHIVE_MAP
+else:
+    pytestmark = pytest.mark.skip("Require Torch")
+
+
+class XxxModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes = (XxxModel, XxxForMaskedLM, XxxForQuestionAnswering,
+                         XxxForSequenceClassification,
+                         XxxForTokenClassification) if is_torch_available() else ()
+
+    class XxxModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_input_mask=True,
+                     use_token_type_ids=True,
+                     use_labels=True,
+                     vocab_size=99,
+                     hidden_size=32,
+                     num_hidden_layers=5,
+                     num_attention_heads=4,
+                     intermediate_size=37,
+                     hidden_act="gelu",
+                     hidden_dropout_prob=0.1,
+                     attention_probs_dropout_prob=0.1,
+                     max_position_embeddings=512,
+                     type_vocab_size=16,
+                     type_sequence_label_size=2,
+                     initializer_range=0.02,
+                     num_labels=3,
+                     num_choices=4,
+                     scope=None,
+                    ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_input_mask = use_input_mask
+            self.use_token_type_ids = use_token_type_ids
+            self.use_labels = use_labels
+            self.vocab_size = vocab_size
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.intermediate_size = intermediate_size
+            self.hidden_act = hidden_act
+            self.hidden_dropout_prob = hidden_dropout_prob
+            self.attention_probs_dropout_prob = attention_probs_dropout_prob
+            self.max_position_embeddings = max_position_embeddings
+            self.type_vocab_size = type_vocab_size
+            self.type_sequence_label_size = type_sequence_label_size
+            self.initializer_range = initializer_range
+            self.num_labels = num_labels
+            self.num_choices = num_choices
+            self.scope = scope
+
+        def prepare_config_and_inputs(self):
+            input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+            input_mask = None
+            if self.use_input_mask:
+                input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+            token_type_ids = None
+            if self.use_token_type_ids:
+                token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+            sequence_labels = None
+            token_labels = None
+            choice_labels = None
+            if self.use_labels:
+                sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+                choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+            config = XxxConfig(
+                vocab_size_or_config_json_file=self.vocab_size,
+                hidden_size=self.hidden_size,
+                num_hidden_layers=self.num_hidden_layers,
+                num_attention_heads=self.num_attention_heads,
+                intermediate_size=self.intermediate_size,
+                hidden_act=self.hidden_act,
+                hidden_dropout_prob=self.hidden_dropout_prob,
+                attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+                max_position_embeddings=self.max_position_embeddings,
+                type_vocab_size=self.type_vocab_size,
+                initializer_range=self.initializer_range)
+
+            return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+        def check_loss_output(self, result):
+            self.parent.assertListEqual(
+                list(result["loss"].size()),
+                [])
+
+        def create_and_check_xxx_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = XxxModel(config=config)
+            model.eval()
+            sequence_output, pooled_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+            sequence_output, pooled_output = model(input_ids, token_type_ids=token_type_ids)
+            sequence_output, pooled_output = model(input_ids)
+
+            result = {
+                "sequence_output": sequence_output,
+                "pooled_output": pooled_output,
+            }
+            self.parent.assertListEqual(
+                list(result["sequence_output"].size()),
+                [self.batch_size, self.seq_length, self.hidden_size])
+            self.parent.assertListEqual(list(result["pooled_output"].size()), [self.batch_size, self.hidden_size])
+
+
+        def create_and_check_xxx_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = XxxForMaskedLM(config=config)
+            model.eval()
+            loss, prediction_scores = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, masked_lm_labels=token_labels)
+            result = {
+                "loss": loss,
+                "prediction_scores": prediction_scores,
+            }
+            self.parent.assertListEqual(
+                list(result["prediction_scores"].size()),
+                [self.batch_size, self.seq_length, self.vocab_size])
+            self.check_loss_output(result)
+
+
+        def create_and_check_xxx_for_question_answering(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = XxxForQuestionAnswering(config=config)
+            model.eval()
+            loss, start_logits, end_logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids,
+                                                   start_positions=sequence_labels, end_positions=sequence_labels)
+            result = {
+                "loss": loss,
+                "start_logits": start_logits,
+                "end_logits": end_logits,
+            }
+            self.parent.assertListEqual(
+                list(result["start_logits"].size()),
+                [self.batch_size, self.seq_length])
+            self.parent.assertListEqual(
+                list(result["end_logits"].size()),
+                [self.batch_size, self.seq_length])
+            self.check_loss_output(result)
+
+
+        def create_and_check_xxx_for_sequence_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            config.num_labels = self.num_labels
+            model = XxxForSequenceClassification(config)
+            model.eval()
+            loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
+            result = {
+                "loss": loss,
+                "logits": logits,
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].size()),
+                [self.batch_size, self.num_labels])
+            self.check_loss_output(result)
+
+
+        def create_and_check_xxx_for_token_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            config.num_labels = self.num_labels
+            model = XxxForTokenClassification(config=config)
+            model.eval()
+            loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+            result = {
+                "loss": loss,
+                "logits": logits,
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].size()),
+                [self.batch_size, self.seq_length, self.num_labels])
+            self.check_loss_output(result)
+
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids, token_type_ids, input_mask,
+             sequence_labels, token_labels, choice_labels) = config_and_inputs
+            inputs_dict = {'input_ids': input_ids, 'token_type_ids': token_type_ids, 'attention_mask': input_mask}
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = XxxModelTest.XxxModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=XxxConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_xxx_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xxx_model(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xxx_for_masked_lm(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xxx_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xxx_for_sequence_classification(*config_and_inputs)
+
+    def test_for_token_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xxx_for_token_classification(*config_and_inputs)
+
+    @pytest.mark.slow
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/transformers_test/"
+        for model_name in list(XXX_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            model = XxxModel.from_pretrained(model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
+if __name__ == "__main__":
+    unittest.main()

templates/adding_a_new_model/tests/tokenization_xxx_test.py

@@ -0,0 +1,57 @@
+# coding=utf-8
+# Copyright 2018 XXX Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import os
+import unittest
+from io import open
+
+from transformers.tokenization_bert import (XxxTokenizer, VOCAB_FILES_NAMES)
+
+from .tokenization_tests_commons import CommonTestCases
+
+class XxxTokenizationTest(CommonTestCases.CommonTokenizerTester):
+
+    tokenizer_class = XxxTokenizer
+
+    def setUp(self):
+        super(XxxTokenizationTest, self).setUp()
+
+        vocab_tokens = [
+            "[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn",
+            "##ing", ",", "low", "lowest",
+        ]
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES['vocab_file'])
+        with open(self.vocab_file, "w", encoding='utf-8') as vocab_writer:
+            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
+
+    def get_tokenizer(self, **kwargs):
+        return XxxTokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_input_output_texts(self):
+        input_text = u"UNwant\u00E9d,running"
+        output_text = u"unwanted, running"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = self.tokenizer_class(self.vocab_file)
+
+        tokens = tokenizer.tokenize(u"UNwant\u00E9d,running")
+        self.assertListEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"])
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [7, 4, 5, 10, 8, 9])
+
+
+if __name__ == '__main__':
+    unittest.main()

templates/adding_a_new_model/tokenization_xxx.py

@@ -0,0 +1,218 @@
+# coding=utf-8
+# Copyright 2018 XXX Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Tokenization class for model XXX."""
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import collections
+import logging
+import os
+import unicodedata
+from io import open
+
+from .tokenization_utils import PreTrainedTokenizer
+
+logger = logging.getLogger(__name__)
+
+####################################################
+# In this template, replace all the XXX (various casings) with your model name
+####################################################
+
+####################################################
+# Mapping from the keyword arguments names of Tokenizer `__init__`
+# to file names for serializing Tokenizer instances
+####################################################
+VOCAB_FILES_NAMES = {'vocab_file': 'vocab.txt'}
+
+####################################################
+# Mapping from the keyword arguments names of Tokenizer `__init__`
+# to pretrained vocabulary URL for all the model shortcut names.
+####################################################
+PRETRAINED_VOCAB_FILES_MAP = {
+    'vocab_file':
+    {
+        'xxx-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-base-uncased-vocab.txt",
+        'xxx-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-large-uncased-vocab.txt",
+    }
+}
+
+####################################################
+# Mapping from model shortcut names to max length of inputs
+####################################################
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    'xxx-base-uncased': 512,
+    'xxx-large-uncased': 512,
+}
+
+####################################################
+# Mapping from model shortcut names to a dictionary of additional
+# keyword arguments for Tokenizer `__init__`.
+# To be used for checkpoint specific configurations.
+####################################################
+PRETRAINED_INIT_CONFIGURATION = {
+    'xxx-base-uncased': {'do_lower_case': True},
+    'xxx-large-uncased': {'do_lower_case': True},
+}
+
+
+def load_vocab(vocab_file):
+    """Loads a vocabulary file into a dictionary."""
+    vocab = collections.OrderedDict()
+    with open(vocab_file, "r", encoding="utf-8") as reader:
+        tokens = reader.readlines()
+    for index, token in enumerate(tokens):
+        token = token.rstrip('\n')
+        vocab[token] = index
+    return vocab
+
+
+class XxxTokenizer(PreTrainedTokenizer):
+    r"""
+    Constructs a XxxTokenizer.
+    :class:`~transformers.XxxTokenizer` runs end-to-end tokenization: punctuation splitting + wordpiece
+
+    Args:
+        vocab_file: Path to a one-wordpiece-per-line vocabulary file
+        do_lower_case: Whether to lower case the input. Only has an effect when do_wordpiece_only=False
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(self, vocab_file, do_lower_case=True,
+                 unk_token="[UNK]", sep_token="[SEP]", pad_token="[PAD]", cls_token="[CLS]",
+                 mask_token="[MASK]", **kwargs):
+        """Constructs a XxxTokenizer.
+
+        Args:
+            **vocab_file**: Path to a one-wordpiece-per-line vocabulary file
+            **do_lower_case**: (`optional`) boolean (default True)
+                Whether to lower case the input
+                Only has an effect when do_basic_tokenize=True
+        """
+        super(XxxTokenizer, self).__init__(unk_token=unk_token, sep_token=sep_token,
+                                           pad_token=pad_token, cls_token=cls_token,
+                                           mask_token=mask_token, **kwargs)
+        self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
+        self.max_len_sentences_pair = self.max_len - 3  # take into account special tokens
+
+        if not os.path.isfile(vocab_file):
+            raise ValueError(
+                "Can't find a vocabulary file at path '{}'. To load the vocabulary from a Google pretrained "
+                "model use `tokenizer = XxxTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`".format(vocab_file))
+        self.vocab = load_vocab(vocab_file)
+
+    @property
+    def vocab_size(self):
+        return len(self.vocab)
+
+    def _tokenize(self, text):
+        """ Take as input a string and return a list of strings (tokens) for words/sub-words
+        """
+        split_tokens = []
+        if self.do_basic_tokenize:
+            for token in self.basic_tokenizer.tokenize(text, never_split=self.all_special_tokens):
+                for sub_token in self.wordpiece_tokenizer.tokenize(token):
+                    split_tokens.append(sub_token)
+        else:
+            split_tokens = self.wordpiece_tokenizer.tokenize(text)
+        return split_tokens
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str/unicode) in an id using the vocab. """
+        return self.vocab.get(token, self.vocab.get(self.unk_token))
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (string/unicode) using the vocab."""
+        return self.ids_to_tokens.get(index, self.unk_token)
+
+    def convert_tokens_to_string(self, tokens):
+        """ Converts a sequence of tokens (string) in a single string. """
+        out_string = ' '.join(tokens).replace(' ##', '').strip()
+        return out_string
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks
+        by concatenating and adding special tokens.
+        A BERT sequence has the following format:
+            single sequence: [CLS] X [SEP]
+            pair of sequences: [CLS] A [SEP] B [SEP]
+        """
+        if token_ids_1 is None:
+            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+        cls = [self.cls_token_id]
+        sep = [self.sep_token_id]
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(self, token_ids_0, token_ids_1=None, already_has_special_tokens=False):
+        """
+        Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer ``prepare_for_model`` or ``encode_plus`` methods.
+
+        Args:
+            token_ids_0: list of ids (must not contain special tokens)
+            token_ids_1: Optional list of ids (must not contain special tokens), necessary when fetching sequence ids
+                for sequence pairs
+            already_has_special_tokens: (default False) Set to True if the token list is already formated with
+                special tokens for the model
+
+        Returns:
+            A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            if token_ids_1 is not None:
+                raise ValueError("You should not supply a second sequence if the provided sequence of "
+                                 "ids is already formated with special tokens for the model.")
+            return list(map(lambda x: 1 if x in [self.sep_token_id, self.cls_token_id] else 0, token_ids_0))
+
+        if token_ids_1 is not None:
+            return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1]
+
+    def create_token_type_ids_from_sequences(self, token_ids_0, token_ids_1=None):
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task.
+        A BERT sequence pair mask has the following format:
+        0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence
+
+        if token_ids_1 is None, only returns the first portion of the mask (0's).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, vocab_path):
+        """Save the tokenizer vocabulary to a directory or file."""
+        index = 0
+        if os.path.isdir(vocab_path):
+            vocab_file = os.path.join(vocab_path, VOCAB_FILES_NAMES['vocab_file'])
+        else:
+            vocab_file = vocab_path
+        with open(vocab_file, "w", encoding="utf-8") as writer:
+            for token, token_index in sorted(self.vocab.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning("Saving vocabulary to {}: vocabulary indices are not consecutive."
+                                   " Please check that the vocabulary is not corrupted!".format(vocab_file))
+                    index = token_index
+                writer.write(token + u'\n')
+                index += 1
+        return (vocab_file,)

transformers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "2.1.0"
+__version__ = "2.2.0"
 
 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -42,6 +42,8 @@ from .tokenization_xlnet import XLNetTokenizer, SPIECE_UNDERLINE
 from .tokenization_xlm import XLMTokenizer
 from .tokenization_roberta import RobertaTokenizer
 from .tokenization_distilbert import DistilBertTokenizer
+from .tokenization_albert import AlbertTokenizer
+from .tokenization_camembert import CamembertTokenizer
 
 # Configurations
 from .configuration_utils import PretrainedConfig
@@ -56,6 +58,8 @@ from .configuration_ctrl import CTRLConfig, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_xlm import XLMConfig, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_roberta import RobertaConfig, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_distilbert import DistilBertConfig, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+from .configuration_albert import AlbertConfig, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+from .configuration_camembert import CamembertConfig, CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 
 # Modeling
 if is_torch_available():
@@ -72,6 +76,7 @@ if is_torch_available():
                                 OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
                                 load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_transfo_xl import (TransfoXLPreTrainedModel, TransfoXLModel, TransfoXLLMHeadModel,
+                                    AdaptiveEmbedding,
                                     load_tf_weights_in_transfo_xl, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_gpt2 import (GPT2PreTrainedModel, GPT2Model,
                                 GPT2LMHeadModel, GPT2DoubleHeadsModel,
@@ -89,14 +94,25 @@ if is_torch_available():
                             XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_roberta import (RobertaForMaskedLM, RobertaModel,
                                 RobertaForSequenceClassification, RobertaForMultipleChoice,
+                                RobertaForTokenClassification,
                                 ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_distilbert import (DistilBertForMaskedLM, DistilBertModel,
                                 DistilBertForSequenceClassification, DistilBertForQuestionAnswering,
+                                DistilBertForTokenClassification,
                                 DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
+    from .modeling_camembert import (CamembertForMaskedLM, CamembertModel,
+                                CamembertForSequenceClassification, CamembertForMultipleChoice,
+                                CamembertForTokenClassification,
+                                CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
+    from .modeling_encoder_decoder import PreTrainedEncoderDecoder, Model2Model
+
+    from .modeling_albert import (AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification,
+                                AlbertForQuestionAnswering,
+                                load_tf_weights_in_albert, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     # Optimization
-    from .optimization import (AdamW, ConstantLRSchedule, WarmupConstantSchedule, WarmupCosineSchedule,
-                               WarmupCosineWithHardRestartsSchedule, WarmupLinearSchedule)
+    from .optimization import (AdamW, get_constant_schedule, get_constant_schedule_with_warmup, get_cosine_schedule_with_warmup,
+                               get_cosine_with_hard_restarts_schedule_with_warmup, get_linear_schedule_with_warmup)
 
 
 # TensorFlow
@@ -110,65 +126,60 @@ if is_tf_available():
                                    TFBertForMaskedLM, TFBertForNextSentencePrediction,
                                    TFBertForSequenceClassification, TFBertForMultipleChoice,
                                    TFBertForTokenClassification, TFBertForQuestionAnswering,
-                                   load_bert_pt_weights_in_tf2,
                                    TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_gpt2 import (TFGPT2PreTrainedModel, TFGPT2MainLayer,
                                    TFGPT2Model, TFGPT2LMHeadModel, TFGPT2DoubleHeadsModel,
-                                   load_gpt2_pt_weights_in_tf2,
                                    TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_openai import (TFOpenAIGPTPreTrainedModel, TFOpenAIGPTMainLayer,
                                      TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel, TFOpenAIGPTDoubleHeadsModel,
-                                     load_openai_gpt_pt_weights_in_tf2,
                                      TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_transfo_xl import (TFTransfoXLPreTrainedModel, TFTransfoXLMainLayer,
                                          TFTransfoXLModel, TFTransfoXLLMHeadModel,
-                                         load_transfo_xl_pt_weights_in_tf2,
                                          TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_xlnet import (TFXLNetPreTrainedModel, TFXLNetMainLayer,
                                     TFXLNetModel, TFXLNetLMHeadModel,
                                     TFXLNetForSequenceClassification,
                                     TFXLNetForQuestionAnsweringSimple,
-                                    load_xlnet_pt_weights_in_tf2,
                                     TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_xlm import (TFXLMPreTrainedModel, TFXLMMainLayer,
                                   TFXLMModel, TFXLMWithLMHeadModel,
                                   TFXLMForSequenceClassification,
                                   TFXLMForQuestionAnsweringSimple,
-                                  load_xlm_pt_weights_in_tf2,
                                   TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_roberta import (TFRobertaPreTrainedModel, TFRobertaMainLayer,
                                       TFRobertaModel, TFRobertaForMaskedLM,
                                       TFRobertaForSequenceClassification,
-                                      load_roberta_pt_weights_in_tf2,
+                                      TFRobertaForTokenClassification,
                                       TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_distilbert import (TFDistilBertPreTrainedModel, TFDistilBertMainLayer,
                                          TFDistilBertModel, TFDistilBertForMaskedLM,
                                          TFDistilBertForSequenceClassification,
                                          TFDistilBertForQuestionAnswering,
-                                         load_distilbert_pt_weights_in_tf2,
                                          TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_ctrl import (TFCTRLPreTrainedModel, TFCTRLModel,
                                     TFCTRLLMHeadModel,
-                                    load_ctrl_pt_weights_in_tf2,
                                     TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP)
 
+    from .modeling_tf_albert import (TFAlbertPreTrainedModel, TFAlbertModel, TFAlbertForMaskedLM,
+                                     TFAlbertForSequenceClassification,
+                                    TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
+
 # TF 2.0 <=> PyTorch conversion utilities
-if is_tf_available() and is_torch_available():
-    from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name,
-                                            load_pytorch_checkpoint_in_tf2_model,
-                                            load_pytorch_weights_in_tf2_model,
-                                            load_pytorch_model_in_tf2_model,
-                                            load_tf2_checkpoint_in_pytorch_model,
-                                            load_tf2_weights_in_pytorch_model,
-                                            load_tf2_model_in_pytorch_model)
+from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name,
+                                        load_pytorch_checkpoint_in_tf2_model,
+                                        load_pytorch_weights_in_tf2_model,
+                                        load_pytorch_model_in_tf2_model,
+                                        load_tf2_checkpoint_in_pytorch_model,
+                                        load_tf2_weights_in_pytorch_model,
+                                        load_tf2_model_in_pytorch_model)
 
 if not is_tf_available() and not is_torch_available():
     logger.warning("Neither PyTorch nor TensorFlow >= 2.0 have been found."

transformers/configuration_albert.py

@@ -0,0 +1,100 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" ALBERT model configuration """
+
+from .configuration_utils import PretrainedConfig
+
+ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    'albert-base-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-config.json",
+    'albert-large-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-config.json",
+    'albert-xlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-config.json",
+    'albert-xxlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-config.json",
+    'albert-base-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v2-config.json",
+    'albert-large-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-v2-config.json",
+    'albert-xlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-v2-config.json",
+    'albert-xxlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-v2-config.json",
+}
+
+class AlbertConfig(PretrainedConfig):
+    """Configuration for `AlbertModel`.
+
+    The default settings match the configuration of model `albert_xxlarge`.
+    """
+
+    pretrained_config_archive_map = ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+    def __init__(self,
+                 vocab_size_or_config_json_file=30000,
+                 embedding_size=128,
+                 hidden_size=4096,
+                 num_hidden_layers=12,
+                 num_hidden_groups=1,
+                 num_attention_heads=64,
+                 intermediate_size=16384,
+                 inner_group_num=1,
+                 hidden_act="gelu_new",
+                 hidden_dropout_prob=0,
+                 attention_probs_dropout_prob=0,
+                 max_position_embeddings=512,
+                 type_vocab_size=2,
+                 initializer_range=0.02,
+                 layer_norm_eps=1e-12, **kwargs):
+        """Constructs AlbertConfig.
+
+        Args:
+            vocab_size: Vocabulary size of `inputs_ids` in `AlbertModel`.
+            embedding_size: size of voc embeddings.
+            hidden_size: Size of the encoder layers and the pooler layer.
+            num_hidden_layers: Number of hidden layers in the Transformer encoder.
+            num_hidden_groups: Number of group for the hidden layers, parameters in
+                the same group are shared.
+            num_attention_heads: Number of attention heads for each attention layer in
+                the Transformer encoder.
+            intermediate_size: The size of the "intermediate" (i.e., feed-forward)
+                layer in the Transformer encoder.
+            inner_group_num: int, number of inner repetition of attention and ffn.
+            down_scale_factor: float, the scale to apply
+            hidden_act: The non-linear activation function (function or string) in the
+                encoder and pooler.
+            hidden_dropout_prob: The dropout probability for all fully connected
+                layers in the embeddings, encoder, and pooler.
+            attention_probs_dropout_prob: The dropout ratio for the attention
+                probabilities.
+            max_position_embeddings: The maximum sequence length that this model might
+                ever be used with. Typically set this to something large just in case
+                (e.g., 512 or 1024 or 2048).
+            type_vocab_size: The vocabulary size of the `token_type_ids` passed into
+                `AlbertModel`.
+            initializer_range: The stdev of the truncated_normal_initializer for
+                initializing all weight matrices.
+        """
+        super(AlbertConfig, self).__init__(**kwargs)
+
+        self.vocab_size = vocab_size_or_config_json_file
+        self.embedding_size = embedding_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_hidden_groups = num_hidden_groups
+        self.num_attention_heads = num_attention_heads
+        self.inner_group_num = inner_group_num
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps

transformers/configuration_auto.py

@@ -27,6 +27,7 @@ from .configuration_xlm import XLMConfig
 from .configuration_roberta import RobertaConfig
 from .configuration_distilbert import DistilBertConfig
 from .configuration_ctrl import CTRLConfig
+from .configuration_camembert import CamembertConfig
 
 logger = logging.getLogger(__name__)
 
@@ -50,6 +51,7 @@ class AutoConfig(object):
             - contains `xlnet`: XLNetConfig (XLNet model)
             - contains `xlm`: XLMConfig (XLM model)
             - contains `roberta`: RobertaConfig (RoBERTa model)
+            - contains `camembert`: CamembertConfig (CamemBERT model)
             - contains `ctrl` : CTRLConfig (CTRL model)
         This class cannot be instantiated using `__init__()` (throw an error).
     """
@@ -72,6 +74,7 @@ class AutoConfig(object):
             - contains `xlnet`: XLNetConfig (XLNet model)
             - contains `xlm`: XLMConfig (XLM model)
             - contains `roberta`: RobertaConfig (RoBERTa model)
+            - contains `camembert`: CamembertConfig (CamemBERT model)
             - contains `ctrl` : CTRLConfig (CTRL model)
         Params:
             pretrained_model_name_or_path: either:
@@ -116,6 +119,8 @@ class AutoConfig(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
+        elif 'camembert' in pretrained_model_name_or_path:
+            return CamembertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -134,4 +139,4 @@ class AutoConfig(object):
             return CTRLConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                          "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
-                         "'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))
+                         "'xlm', 'roberta', 'camembert', 'ctrl'".format(pretrained_model_name_or_path))

transformers/configuration_bert.py

@@ -40,6 +40,8 @@ BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-config.json",
     'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-config.json",
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-config.json",
+    'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-config.json",
+    'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-config.json",
 }
 
 

transformers/configuration_camembert.py

@@ -0,0 +1,33 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" CamemBERT configuration """
+
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import logging
+
+from .configuration_roberta import RobertaConfig
+
+logger = logging.getLogger(__name__)
+
+CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    'camembert-base': "https://s3.amazonaws.com/models.huggingface.co/bert/camembert-base-config.json",
+}
+
+
+class CamembertConfig(RobertaConfig):
+    pretrained_config_archive_map = CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP

transformers/configuration_gpt2.py

@@ -29,6 +29,7 @@ logger = logging.getLogger(__name__)
 GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-config.json",
                                       "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-config.json",
                                       "gpt2-large": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-large-config.json",
+                                      "gpt2-xl": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-xl-config.json",
                                       "distilgpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/distilgpt2-config.json",}
 
 class GPT2Config(PretrainedConfig):

transformers/configuration_roberta.py

@@ -28,6 +28,9 @@ ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     'roberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-base-config.json",
     'roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-config.json",
     'roberta-large-mnli': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-mnli-config.json",
+    'distilroberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/distilroberta-base-config.json",
+    'roberta-base-openai-detector': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-base-openai-detector-config.json",
+    'roberta-large-openai-detector': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-openai-detector-config.json",
 }
 
 

transformers/configuration_utils.py

@@ -53,9 +53,11 @@ class PretrainedConfig(object):
         self.num_labels = kwargs.pop('num_labels', 2)
         self.output_attentions = kwargs.pop('output_attentions', False)
         self.output_hidden_states = kwargs.pop('output_hidden_states', False)
-        self.torchscript = kwargs.pop('torchscript', False)
+        self.output_past = kwargs.pop('output_past', True)  # Not used by all models
+        self.torchscript = kwargs.pop('torchscript', False)  # Only used by PyTorch models
         self.use_bfloat16 = kwargs.pop('use_bfloat16', False)
         self.pruned_heads = kwargs.pop('pruned_heads', {})
+        self.is_decoder = kwargs.pop('is_decoder', False)
 
     def save_pretrained(self, save_directory):
         """ Save a configuration object to the directory `save_directory`, so that it
@@ -130,20 +132,19 @@ class PretrainedConfig(object):
         # redirect to the cache, if necessary
         try:
             resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
-        except EnvironmentError as e:
+        except EnvironmentError:
             if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
-                logger.error(
-                    "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
-                        config_file))
+                msg = "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
+                        config_file)
             else:
-                logger.error(
-                    "Model name '{}' was not found in model name list ({}). "
-                    "We assumed '{}' was a path or url but couldn't find any file "
-                    "associated to this path or url.".format(
+                msg = "Model name '{}' was not found in model name list ({}). " \
+                      "We assumed '{}' was a path or url to a configuration file named {} or " \
+                      "a directory containing such a file but couldn't find any such file at this path or url.".format(
                         pretrained_model_name_or_path,
                         ', '.join(cls.pretrained_config_archive_map.keys()),
-                        config_file))
-            raise e
+                        config_file, CONFIG_NAME)
+            raise EnvironmentError(msg)
+
         if resolved_config_file == config_file:
             logger.info("loading configuration file {}".format(config_file))
         else:
@@ -154,7 +155,7 @@ class PretrainedConfig(object):
         config = cls.from_json_file(resolved_config_file)
 
         if hasattr(config, 'pruned_heads'):
-            config.pruned_heads = dict((int(key), set(value)) for key, value in config.pruned_heads.items())
+            config.pruned_heads = dict((int(key), value) for key, value in config.pruned_heads.items())
 
         # Update config with kwargs if needed
         to_remove = []
@@ -165,7 +166,7 @@ class PretrainedConfig(object):
         for key in to_remove:
             kwargs.pop(key, None)
 
-        logger.info("Model config %s", config)
+        logger.info("Model config %s", str(config))
         if return_unused_kwargs:
             return config, kwargs
         else:

transformers/convert_albert_original_tf_checkpoint_to_pytorch.py

@@ -0,0 +1,67 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert ALBERT checkpoint."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import torch
+
+from transformers import AlbertConfig, AlbertForMaskedLM, load_tf_weights_in_albert
+
+import logging
+logging.basicConfig(level=logging.INFO)
+
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, albert_config_file, pytorch_dump_path):
+    # Initialise PyTorch model
+    config = AlbertConfig.from_json_file(albert_config_file)
+    print("Building PyTorch model from configuration: {}".format(str(config)))
+    model = AlbertForMaskedLM(config)
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_albert(model, config, tf_checkpoint_path)
+
+    # Save pytorch-model
+    print("Save PyTorch model to {}".format(pytorch_dump_path))
+    torch.save(model.state_dict(), pytorch_dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    ## Required parameters
+    parser.add_argument("--tf_checkpoint_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the TensorFlow checkpoint path.")
+    parser.add_argument("--albert_config_file",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "The config json file corresponding to the pre-trained ALBERT model. \n"
+                            "This specifies the model architecture.")
+    parser.add_argument("--pytorch_dump_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the output PyTorch model.")
+    args = parser.parse_args()
+    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path,
+                                     args.albert_config_file,
+                                     args.pytorch_dump_path)
+ 

transformers/convert_pytorch_checkpoint_to_tf2.py

@@ -24,15 +24,17 @@ import tensorflow as tf
 
 from transformers import is_torch_available, cached_path
 
-from transformers import (BertConfig, TFBertForPreTraining, TFBertForQuestionAnswering, TFBertForSequenceClassification, load_bert_pt_weights_in_tf2, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  GPT2Config, TFGPT2LMHeadModel, load_gpt2_pt_weights_in_tf2, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  XLNetConfig, TFXLNetLMHeadModel, load_xlnet_pt_weights_in_tf2, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  XLMConfig, TFXLMWithLMHeadModel, load_xlm_pt_weights_in_tf2, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  TransfoXLConfig, TFTransfoXLLMHeadModel, load_transfo_xl_pt_weights_in_tf2, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  OpenAIGPTConfig, TFOpenAIGPTLMHeadModel, load_openai_gpt_pt_weights_in_tf2, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  RobertaConfig, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, load_roberta_pt_weights_in_tf2, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  DistilBertConfig, TFDistilBertForMaskedLM, TFDistilBertForQuestionAnswering, load_distilbert_pt_weights_in_tf2, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  CTRLConfig, TFCTRLLMHeadModel, load_ctrl_pt_weights_in_tf2, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP)
+from transformers import (load_pytorch_checkpoint_in_tf2_model,
+    BertConfig, TFBertForPreTraining, TFBertForQuestionAnswering, TFBertForSequenceClassification, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  GPT2Config, TFGPT2LMHeadModel, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  XLNetConfig, TFXLNetLMHeadModel, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  XLMConfig, TFXLMWithLMHeadModel, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  TransfoXLConfig, TFTransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  OpenAIGPTConfig, TFOpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  RobertaConfig, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  DistilBertConfig, TFDistilBertForMaskedLM, TFDistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  CTRLConfig, TFCTRLLMHeadModel, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  AlbertConfig, TFAlbertForMaskedLM, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
 
 if is_torch_available():
     import torch
@@ -45,7 +47,8 @@ if is_torch_available():
                                       OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                       RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
                                       DistilBertForMaskedLM, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
-                                      CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP)
+                                      CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+                                      AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 else:
     (BertForPreTraining, BertForQuestionAnswering, BertForSequenceClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
@@ -55,7 +58,8 @@ else:
     OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
     RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
     DistilBertForMaskedLM, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
-    CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP) = (
+    CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+    AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP) = (
         None, None, None, None,
         None, None,
         None, None,
@@ -64,6 +68,7 @@ else:
         None, None,
         None, None, None,
         None, None, None,
+        None, None,
         None, None)
 
 
@@ -71,27 +76,28 @@ import logging
 logging.basicConfig(level=logging.INFO)
 
 MODEL_CLASSES = {
-    'bert': (BertConfig, TFBertForPreTraining, load_bert_pt_weights_in_tf2, BertForPreTraining, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'bert-large-uncased-whole-word-masking-finetuned-squad': (BertConfig, TFBertForQuestionAnswering, load_bert_pt_weights_in_tf2, BertForQuestionAnswering, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'bert-large-cased-whole-word-masking-finetuned-squad': (BertConfig, TFBertForQuestionAnswering, load_bert_pt_weights_in_tf2, BertForQuestionAnswering, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'bert-base-cased-finetuned-mrpc': (BertConfig, TFBertForSequenceClassification, load_bert_pt_weights_in_tf2, BertForSequenceClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'gpt2': (GPT2Config, TFGPT2LMHeadModel, load_gpt2_pt_weights_in_tf2, GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'xlnet': (XLNetConfig, TFXLNetLMHeadModel, load_xlnet_pt_weights_in_tf2, XLNetLMHeadModel, XLNET_PRETRAINED_MODEL_ARCHIVE_MAP, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'xlm': (XLMConfig, TFXLMWithLMHeadModel, load_xlm_pt_weights_in_tf2, XLMWithLMHeadModel, XLM_PRETRAINED_MODEL_ARCHIVE_MAP, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'transfo-xl': (TransfoXLConfig, TFTransfoXLLMHeadModel, load_transfo_xl_pt_weights_in_tf2, TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'openai-gpt': (OpenAIGPTConfig, TFOpenAIGPTLMHeadModel, load_openai_gpt_pt_weights_in_tf2, OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'roberta': (RobertaConfig, TFRobertaForMaskedLM, load_roberta_pt_weights_in_tf2, RobertaForMaskedLM, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'roberta-large-mnli': (RobertaConfig, TFRobertaForSequenceClassification, load_roberta_pt_weights_in_tf2, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'distilbert': (DistilBertConfig, TFDistilBertForMaskedLM, load_distilbert_pt_weights_in_tf2, DistilBertForMaskedLM, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'distilbert-base-uncased-distilled-squad': (DistilBertConfig, TFDistilBertForQuestionAnswering, load_distilbert_pt_weights_in_tf2, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'ctrl': (CTRLConfig, TFCTRLLMHeadModel, load_ctrl_pt_weights_in_tf2, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP)
+    'bert': (BertConfig, TFBertForPreTraining, BertForPreTraining, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'bert-large-uncased-whole-word-masking-finetuned-squad': (BertConfig, TFBertForQuestionAnswering, BertForQuestionAnswering, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'bert-large-cased-whole-word-masking-finetuned-squad': (BertConfig, TFBertForQuestionAnswering, BertForQuestionAnswering, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'bert-base-cased-finetuned-mrpc': (BertConfig, TFBertForSequenceClassification, BertForSequenceClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'gpt2': (GPT2Config, TFGPT2LMHeadModel, GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'xlnet': (XLNetConfig, TFXLNetLMHeadModel, XLNetLMHeadModel, XLNET_PRETRAINED_MODEL_ARCHIVE_MAP, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'xlm': (XLMConfig, TFXLMWithLMHeadModel, XLMWithLMHeadModel, XLM_PRETRAINED_MODEL_ARCHIVE_MAP, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'transfo-xl': (TransfoXLConfig, TFTransfoXLLMHeadModel, TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'openai-gpt': (OpenAIGPTConfig, TFOpenAIGPTLMHeadModel, OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'roberta': (RobertaConfig, TFRobertaForMaskedLM, RobertaForMaskedLM, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'roberta-large-mnli': (RobertaConfig, TFRobertaForSequenceClassification, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'distilbert': (DistilBertConfig, TFDistilBertForMaskedLM, DistilBertForMaskedLM, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'distilbert-base-uncased-distilled-squad': (DistilBertConfig, TFDistilBertForQuestionAnswering, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'ctrl': (CTRLConfig, TFCTRLLMHeadModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'albert': (AlbertConfig, TFAlbertForMaskedLM, AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
 }
 
 def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file, tf_dump_path, compare_with_pt_model=False, use_cached_models=True):
     if model_type not in MODEL_CLASSES:
         raise ValueError("Unrecognized model type, should be one of {}.".format(list(MODEL_CLASSES.keys())))
 
-    config_class, model_class, loading_fct, pt_model_class, aws_model_maps, aws_config_map = MODEL_CLASSES[model_type]
+    config_class, model_class, pt_model_class, aws_model_maps, aws_config_map = MODEL_CLASSES[model_type]
 
     # Initialise TF model
     if config_file in aws_config_map:
@@ -105,7 +111,8 @@ def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file
     # Load weights from tf checkpoint
     if pytorch_checkpoint_path in aws_model_maps:
         pytorch_checkpoint_path = cached_path(aws_model_maps[pytorch_checkpoint_path], force_download=not use_cached_models)
-    tf_model = loading_fct(tf_model, pytorch_checkpoint_path)
+    # Load PyTorch checkpoint in tf2 model:
+    tf_model = load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path)
 
     if compare_with_pt_model:
         inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
@@ -147,7 +154,7 @@ def convert_all_pt_checkpoints_to_tf(args_model_type, tf_dump_path, model_shortc
         if model_type not in MODEL_CLASSES:
             raise ValueError("Unrecognized model type {}, should be one of {}.".format(model_type, list(MODEL_CLASSES.keys())))
 
-        config_class, model_class, loading_fct, pt_model_class, aws_model_maps, aws_config_map = MODEL_CLASSES[model_type]
+        config_class, model_class, pt_model_class, aws_model_maps, aws_config_map = MODEL_CLASSES[model_type]
 
         if model_shortcut_names_or_path is None:
             model_shortcut_names_or_path = list(aws_model_maps.keys())

transformers/convert_roberta_original_pytorch_checkpoint_to_pytorch.py

@@ -23,15 +23,15 @@ import torch
 
 from fairseq.models.roberta import RobertaModel as FairseqRobertaModel
 from fairseq.modules import TransformerSentenceEncoderLayer
-from transformers import (BertConfig, BertEncoder,
-                                                BertIntermediate, BertLayer,
-                                                BertModel, BertOutput,
-                                                BertSelfAttention,
-                                                BertSelfOutput)
-from transformers import (RobertaEmbeddings,
-                                                   RobertaForMaskedLM,
-                                                   RobertaForSequenceClassification,
-                                                   RobertaModel)
+from transformers.modeling_bert import (BertConfig, BertEncoder,
+                                        BertIntermediate, BertLayer,
+                                        BertModel, BertOutput,
+                                        BertSelfAttention,
+                                        BertSelfOutput)
+from transformers.modeling_roberta import (RobertaEmbeddings,
+                                           RobertaForMaskedLM,
+                                           RobertaForSequenceClassification,
+                                           RobertaModel)
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)

transformers/data/processors/glue.py

@@ -80,6 +80,7 @@ def glue_convert_examples_to_features(examples, tokenizer,
             logger.info("Writing example %d" % (ex_index))
         if is_tf_dataset:
             example = processor.get_example_from_tensor_dict(example)
+            example = processor.tfds_map(example)
 
         inputs = tokenizer.encode_plus(
             example.text_a,

transformers/data/processors/utils.py

@@ -107,6 +107,13 @@ class DataProcessor(object):
         """Gets the list of labels for this data set."""
         raise NotImplementedError()
 
+    def tfds_map(self, example):
+        """Some tensorflow_datasets datasets are not formatted the same way the GLUE datasets are. 
+        This method converts examples to the correct format."""
+        if len(self.get_labels()) > 1:
+            example.label = self.get_labels()[int(example.label)]
+        return example
+
     @classmethod
     def _read_tsv(cls, input_file, quotechar=None):
         """Reads a tab separated value file."""

transformers/file_utils.py

@@ -246,7 +246,7 @@ def http_get(url, temp_file, proxies=None):
     progress.close()
 
 
-def get_from_cache(url, cache_dir=None, force_download=False, proxies=None):
+def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag_timeout=10):
     """
     Given a URL, look for the corresponding dataset in the local cache.
     If it's not there, download it. Then return the path to the cached file.
@@ -266,12 +266,12 @@ def get_from_cache(url, cache_dir=None, force_download=False, proxies=None):
         etag = s3_etag(url, proxies=proxies)
     else:
         try:
-            response = requests.head(url, allow_redirects=True, proxies=proxies)
+            response = requests.head(url, allow_redirects=True, proxies=proxies, timeout=etag_timeout)
             if response.status_code != 200:
                 etag = None
             else:
                 etag = response.headers.get("ETag")
-        except EnvironmentError:
+        except (EnvironmentError, requests.exceptions.Timeout):
             etag = None
 
     if sys.version_info[0] == 2 and etag is not None:

transformers/modeling_albert.py

@@ -0,0 +1,764 @@
+
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch ALBERT model. """
+
+import os
+import math
+import logging
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss, MSELoss
+from transformers.modeling_utils import PreTrainedModel
+from transformers.configuration_albert import AlbertConfig
+from transformers.modeling_bert import BertEmbeddings, BertSelfAttention, prune_linear_layer, ACT2FN
+from .file_utils import add_start_docstrings
+
+logger = logging.getLogger(__name__)
+
+
+ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'albert-base-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-pytorch_model.bin",
+    'albert-large-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-pytorch_model.bin",
+    'albert-xlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-pytorch_model.bin",
+    'albert-xxlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-pytorch_model.bin",
+    'albert-base-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v2-pytorch_model.bin",
+    'albert-large-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-v2-pytorch_model.bin",
+    'albert-xlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-v2-pytorch_model.bin",
+    'albert-xxlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-v2-pytorch_model.bin",
+}
+
+
+def load_tf_weights_in_albert(model, config, tf_checkpoint_path):
+    """ Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions.")
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        print(name)
+    
+    for name, array in zip(names, arrays):
+        original_name = name
+        name = name.replace("ffn_1", "ffn")
+        name = name.replace("/bert/", "/albert/")
+        name = name.replace("ffn/intermediate/output", "ffn_output")
+        name = name.replace("attention_1", "attention")   
+        name = name.replace("cls/predictions", "predictions")
+        name = name.replace("transform/", "")
+        name = name.replace("LayerNorm_1", "full_layer_layer_norm")    
+        name = name.replace("LayerNorm", "attention/LayerNorm")    
+        name = name.replace("inner_group_", "albert_layers/") 
+        name = name.replace("group_", "albert_layer_groups/")    
+        name = name.split('/')
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
+                l = re.split(r'_(\d+)', m_name)
+            else:
+                l = [m_name]
+
+            if l[0] == 'kernel' or l[0] == 'gamma':
+                pointer = getattr(pointer, 'weight')
+            elif l[0] == 'output_bias' or l[0] == 'beta':
+                pointer = getattr(pointer, 'bias')
+            elif l[0] == 'output_weights':
+                pointer = getattr(pointer, 'weight')
+            elif l[0] == 'squad':
+                pointer = getattr(pointer, 'classifier')
+            else:
+                try:
+                    pointer = getattr(pointer, l[0])
+                except AttributeError:
+                    logger.info("Skipping {}".format("/".join(name)))
+                    continue
+            if len(l) >= 2:
+                num = int(l[1])
+                pointer = pointer[num]
+
+        if m_name[-11:] == '_embeddings':
+            pointer = getattr(pointer, 'weight')
+        elif m_name == 'kernel':
+            array = np.transpose(array)
+        try:
+            assert pointer.shape == array.shape
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        print("Initialize PyTorch weight {} from {}".format(name, original_name))
+        pointer.data = torch.from_numpy(array)
+
+    return model
+
+
+class AlbertEmbeddings(BertEmbeddings):
+    """
+    Construct the embeddings from word, position and token_type embeddings.
+    """
+    def __init__(self, config):
+        super(AlbertEmbeddings, self).__init__(config)
+
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=0)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.embedding_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_size)
+        self.LayerNorm = torch.nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
+
+
+class AlbertAttention(BertSelfAttention):
+    def __init__(self, config):
+        super(AlbertAttention, self).__init__(config)
+
+        self.output_attentions = config.output_attentions
+        self.num_attention_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size 
+        self.attention_head_size = config.hidden_size // config.num_attention_heads
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        mask = torch.ones(self.num_attention_heads, self.attention_head_size)
+        heads = set(heads) - self.pruned_heads  # Convert to set and emove already pruned heads
+        for head in heads:
+            # Compute how many pruned heads are before the head and move the index accordingly
+            head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
+            mask[head] = 0
+        mask = mask.view(-1).contiguous().eq(1)
+        index = torch.arange(len(mask))[mask].long()
+
+        # Prune linear layers
+        self.query = prune_linear_layer(self.query, index)
+        self.key = prune_linear_layer(self.key, index)
+        self.value = prune_linear_layer(self.value, index)
+        self.dense = prune_linear_layer(self.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.num_attention_heads = self.num_attention_heads - len(heads)
+        self.all_head_size = self.attention_head_size * self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(self, input_ids, attention_mask=None, head_mask=None):
+        mixed_query_layer = self.query(input_ids)
+        mixed_key_layer = self.key(input_ids)
+        mixed_value_layer = self.value(input_ids)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        reshaped_context_layer = context_layer.view(*new_context_layer_shape)
+        
+
+        # Should find a better way to do this
+        w = self.dense.weight.t().view(self.num_attention_heads, self.attention_head_size, self.hidden_size).to(context_layer.dtype)
+        b = self.dense.bias.to(context_layer.dtype)
+
+        projected_context_layer = torch.einsum("bfnd,ndh->bfh", context_layer, w) + b
+        projected_context_layer_dropout = self.dropout(projected_context_layer)
+        layernormed_context_layer = self.LayerNorm(input_ids + projected_context_layer_dropout)
+        return (layernormed_context_layer, attention_probs) if self.output_attentions else (layernormed_context_layer,)
+
+
+class AlbertLayer(nn.Module):
+    def __init__(self, config):
+        super(AlbertLayer, self).__init__()
+        
+        self.config = config
+        self.full_layer_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attention = AlbertAttention(config)
+        self.ffn = nn.Linear(config.hidden_size, config.intermediate_size) 
+        self.ffn_output = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.activation = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+        attention_output = self.attention(hidden_states, attention_mask, head_mask)
+        ffn_output = self.ffn(attention_output[0])
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(ffn_output)
+        hidden_states = self.full_layer_layer_norm(ffn_output + attention_output[0])
+
+        return (hidden_states,) + attention_output[1:]  # add attentions if we output them
+
+
+class AlbertLayerGroup(nn.Module):
+    def __init__(self, config):
+        super(AlbertLayerGroup, self).__init__()
+        
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.albert_layers = nn.ModuleList([AlbertLayer(config) for _ in range(config.inner_group_num)])
+
+    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+        layer_hidden_states = ()
+        layer_attentions = ()
+
+        for layer_index, albert_layer in enumerate(self.albert_layers):
+            layer_output = albert_layer(hidden_states, attention_mask, head_mask[layer_index])
+            hidden_states = layer_output[0]
+
+            if self.output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+            if self.output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (layer_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (layer_attentions,)
+        return outputs  # last-layer hidden state, (layer hidden states), (layer attentions)
+
+
+class AlbertTransformer(nn.Module):
+    def __init__(self, config):
+        super(AlbertTransformer, self).__init__()
+        
+        self.config = config
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.embedding_hidden_mapping_in = nn.Linear(config.embedding_size, config.hidden_size)
+        self.albert_layer_groups = nn.ModuleList([AlbertLayerGroup(config) for _ in range(config.num_hidden_groups)])
+
+    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+        hidden_states = self.embedding_hidden_mapping_in(hidden_states)
+
+        all_attentions = ()
+
+        if self.output_hidden_states:
+            all_hidden_states = (hidden_states,)
+
+        for i in range(self.config.num_hidden_layers):
+            # Number of layers in a hidden group
+            layers_per_group = int(self.config.num_hidden_layers / self.config.num_hidden_groups)
+
+            # Index of the hidden group
+            group_idx = int(i / (self.config.num_hidden_layers / self.config.num_hidden_groups))
+
+            # Index of the layer inside the group
+            layer_idx = int(i - group_idx * layers_per_group)
+            
+            layer_group_output = self.albert_layer_groups[group_idx](hidden_states, attention_mask, head_mask[group_idx*layers_per_group:(group_idx+1)*layers_per_group])  
+            hidden_states = layer_group_output[0]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + layer_group_output[-1]
+
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
+
+
+
+class AlbertPreTrainedModel(PreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = AlbertConfig
+    pretrained_model_archive_map = ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "albert"
+
+    def _init_weights(self, module):
+        """ Initialize the weights.
+        """
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if isinstance(module, (nn.Linear)) and module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+ALBERT_START_DOCSTRING = r"""    The ALBERT model was proposed in
+    `ALBERT: A Lite BERT for Self-supervised Learning of Language Representations`_
+    by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut. It presents
+    two parameter-reduction techniques to lower memory consumption and increase the trainig speed of BERT.
+
+    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+    refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+    .. _`ALBERT: A Lite BERT for Self-supervised Learning of Language Representations`:
+        https://arxiv.org/abs/1909.11942
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~transformers.AlbertConfig`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, BERT input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+                
+                ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+                
+                ``token_type_ids:   0   0   0   0  0     0   0``
+
+            Albert is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
+            Indices can be obtained using :class:`transformers.AlbertTokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Segment token indices to indicate first and second portions of the inputs.
+            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
+            corresponds to a `sentence B` token
+            (see `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
+        **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+"""
+
+@add_start_docstrings("The bare ALBERT Model transformer outputting raw hidden-states without any specific head on top.",
+                      ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class AlbertModel(AlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during Bert pretraining. This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+    """
+
+    config_class = AlbertConfig
+    pretrained_model_archive_map = ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    load_tf_weights = load_tf_weights_in_albert
+    base_model_prefix = "albert"
+
+    def __init__(self, config):
+        super(AlbertModel, self).__init__(config)
+
+        self.config = config
+        self.embeddings = AlbertEmbeddings(config)
+        self.encoder = AlbertTransformer(config)
+        self.pooler = nn.Linear(config.hidden_size, config.hidden_size)
+        self.pooler_activation = nn.Tanh()
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        old_embeddings = self.embeddings.word_embeddings
+        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
+        self.embeddings.word_embeddings = new_embeddings
+        return self.embeddings.word_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            ALBERT has a different architecture in that its layers are shared across groups, which then has inner groups.
+            If an ALBERT model has 12 hidden layers and 2 hidden groups, with two inner groups, there
+            is a total of 4 different layers.
+
+            These layers are flattened: the indices [0,1] correspond to the two inner groups of the first hidden layer,
+            while [2,3] correspond to the two inner groups of the second hidden layer.
+
+            Any layer with in index other than [0,1,2,3] will result in an error.
+            See base class PreTrainedModel for more information about head pruning
+        """
+        for layer, heads in heads_to_prune.items():
+            group_idx = int(layer / self.config.inner_group_num)
+            inner_group_idx = int(layer - group_idx * self.config.inner_group_num)
+            self.encoder.albert_layer_groups[group_idx].albert_layers[inner_group_idx].attention.prune_heads(heads)
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                inputs_embeds=None):
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.num_hidden_layers
+
+        embedding_output = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids,
+                                           inputs_embeds=inputs_embeds)
+        encoder_outputs = self.encoder(embedding_output,
+                                       extended_attention_mask,
+                                       head_mask=head_mask)
+
+        sequence_output = encoder_outputs[0]
+
+        pooled_output = self.pooler_activation(self.pooler(sequence_output[:, 0]))
+
+        outputs = (sequence_output, pooled_output) + encoder_outputs[1:]  # add hidden_states and attentions if they are here
+        return outputs
+
+class AlbertMLMHead(nn.Module):
+    def __init__(self, config):
+        super(AlbertMLMHead, self).__init__()
+
+        self.LayerNorm = nn.LayerNorm(config.embedding_size)
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+        self.dense = nn.Linear(config.hidden_size, config.embedding_size)
+        self.decoder = nn.Linear(config.embedding_size, config.vocab_size)
+        self.activation = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+
+        prediction_scores = hidden_states + self.bias
+
+        return prediction_scores
+
+
+@add_start_docstrings("Bert Model with a `language modeling` head on top.", ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class AlbertForMaskedLM(AlbertPreTrainedModel):
+    r"""
+        **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+    """
+
+    def __init__(self, config):
+        super(AlbertForMaskedLM, self).__init__(config)
+
+        self.albert = AlbertModel(config)
+        self.predictions = AlbertMLMHead(config)
+
+        self.init_weights()
+        self.tie_weights()
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.predictions.decoder,
+                                   self.albert.embeddings.word_embeddings)
+
+    def get_output_embeddings(self):
+        return self.predictions.decoder
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
+                masked_lm_labels=None):
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds
+        )
+        sequence_outputs = outputs[0]
+
+        prediction_scores = self.predictions(sequence_outputs)
+
+        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+        if masked_lm_labels is not None:
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
+            outputs = (masked_lm_loss,) + outputs
+
+        return outputs
+
+
+@add_start_docstrings("""Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class AlbertForSequenceClassification(AlbertPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
+            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification (or regression if config.num_labels==1) loss.
+        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = AlbertForSequenceClassification.from_pretrained('albert-base-v2')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(AlbertForSequenceClassification, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.albert = AlbertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), logits, (hidden_states), (attentions)
+
+
+
+@add_start_docstrings("""Albert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
+    the hidden-states output to compute `span start logits` and `span end logits`). """,
+    ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class AlbertForQuestionAnswering(AlbertPreTrainedModel):
+    r"""
+        **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+        **end_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        **start_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-start scores (before SoftMax).
+        **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-end scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = AlbertForQuestionAnswering.from_pretrained('albert-base-v2')
+        question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        input_text = "[CLS] " + question + " [SEP] " + text + " [SEP]"
+        input_ids = tokenizer.encode(input_text)
+        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))] 
+        start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))
+        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)  
+        print(' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]))
+        # a nice puppet
+
+
+    """
+    def __init__(self, config):
+        super(AlbertForQuestionAnswering, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.albert = AlbertModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                inputs_embeds=None, start_positions=None, end_positions=None):
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        outputs = (start_logits, end_logits,) + outputs[2:]
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions.clamp_(0, ignored_index)
+            end_positions.clamp_(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+            outputs = (total_loss,) + outputs
+
+        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)

transformers/modeling_auto.py

@@ -27,6 +27,7 @@ from .modeling_xlnet import XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassi
 from .modeling_xlm import XLMModel, XLMWithLMHeadModel, XLMForSequenceClassification, XLMForQuestionAnswering
 from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification
 from .modeling_distilbert import DistilBertModel, DistilBertForQuestionAnswering, DistilBertForMaskedLM, DistilBertForSequenceClassification
+from .modeling_camembert import CamembertModel, CamembertForMaskedLM, CamembertForSequenceClassification, CamembertForMultipleChoice
 
 from .modeling_utils import PreTrainedModel, SequenceSummary
 
@@ -48,6 +49,7 @@ class AutoModel(object):
         The base model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertModel (DistilBERT model)
+            - contains `camembert`: CamembertModel (CamemBERT model)
             - contains `roberta`: RobertaModel (RoBERTa model)
             - contains `bert`: BertModel (Bert model)
             - contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
@@ -71,6 +73,7 @@ class AutoModel(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertModel (DistilBERT model)
+            - contains `camembert`: CamembertModel (CamemBERT model)
             - contains `roberta`: RobertaModel (RoBERTa model)
             - contains `bert`: BertModel (Bert model)
             - contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
@@ -138,6 +141,8 @@ class AutoModel(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'camembert' in pretrained_model_name_or_path:
+            return CamembertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -172,6 +177,7 @@ class AutoModelWithLMHead(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
+            - contains `camembert`: CamembertForMaskedLM (CamemBERT model)
             - contains `roberta`: RobertaForMaskedLM (RoBERTa model)
             - contains `bert`: BertForMaskedLM (Bert model)
             - contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -198,6 +204,7 @@ class AutoModelWithLMHead(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
+            - contains `camembert`: CamembertForMaskedLM (CamemBERT model)
             - contains `roberta`: RobertaForMaskedLM (RoBERTa model)
             - contains `bert`: BertForMaskedLM (Bert model)
             - contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -264,6 +271,8 @@ class AutoModelWithLMHead(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'camembert' in pretrained_model_name_or_path:
+            return CamembertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -298,6 +307,7 @@ class AutoModelForSequenceClassification(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertForSequenceClassification (DistilBERT model)
+            - contains `camembert`: CamembertForSequenceClassification (CamemBERT model)
             - contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
             - contains `bert`: BertForSequenceClassification (Bert model)
             - contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -320,6 +330,7 @@ class AutoModelForSequenceClassification(object):
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
             - contains `distilbert`: DistilBertForSequenceClassification (DistilBERT model)
+            - contains `camembert`: CamembertForSequenceClassification (CamemBERT model)
             - contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
             - contains `bert`: BertForSequenceClassification (Bert model)
             - contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -383,6 +394,8 @@ class AutoModelForSequenceClassification(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'camembert' in pretrained_model_name_or_path:
+            return CamembertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:

transformers/modeling_beam_search.py

@@ -0,0 +1,271 @@
+# coding=utf-8
+# Copyright (c) 2019 Yang Liu
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+"""
+A general wrapper around models with LM heads to generate sequences
+using beam search.
+"""
+import torch
+from torch import nn
+
+
+class TransformerBeamSearch(nn.Module):
+    def __init__(
+        self,
+        model,
+        tokenizer,
+        batch_size,
+        beam_size,
+        min_length,
+        max_length,
+        alpha=0,
+        block_repeating_trigram=True,
+    ):
+        """
+        Attributes:
+            mask_word_id: token id that corresponds to the mask
+        """
+        super(TransformerBeamSearch, self).__init__()
+        self.model = model
+        self.tokenizer = tokenizer
+
+        self.start_token_id = tokenizer.start_token_id
+        self.end_token_id = tokenizer.end_token_id
+        self.pad_token_id = tokenizer.pad_token_id
+
+        self.beam_size = beam_size
+        self.min_length = min_length
+        self.max_length = max_length
+
+        self.block_repeating_trigram = block_repeating_trigram
+        self.apply_length_penalty = False if alpha == 0 else True
+        self.alpha = alpha
+
+        # State of the beam
+        self.hypotheses = [[] for _ in range(batch_size)]
+        self.batch_offset = torch.arange(batch_size, dtype=torch.long)
+        self.beam_offset = torch.arange(
+            0, batch_size * self.beam_size, step=self.beam_size, dtype=torch.long
+        )
+        self.growing_beam = torch.full(
+            (batch_size * self.beam_size, 1), self.start_token_id, dtype=torch.long
+        )
+        self.topk_log_probabilities = torch.tensor(
+            [0.0] + [float("-inf")] * (self.beam_size - 1), dtype=torch.float
+        ).repeat(batch_size)
+        self.results = {
+            "prediction": [[] for _ in batch_size],
+            "scores": [[] for _ in batch_size],
+        }
+        self._step = 0
+        self.is_done = False
+
+    def step(self, log_probabilities):
+        """ Grows the beam by one step. """
+        self._step += 1
+
+        # The batch size changes as some beams finish so we define _B
+        vocab_size = log_probabilities.size(-1)
+        _B = log_probabilities.size(0) // self.beam_size
+
+        # Multiply each beam probability with the probability of the
+        # next token (conditioned on the words in the beam).
+        log_probabilities += self.topk_log_probabilities.view(-1, 1)
+
+        self.enforce_min_length(log_probabilities)
+        if self.block_repeating_trigram:
+            self.remove_repeating_trigrams(log_probabilities, _B)
+
+        # Find the `beam_size` (previous_beam + token) combinations with
+        # the highest score
+        topk_log_probabilities, topk_ids = log_probabilities.topk(
+            log_probabilities.view(_B, self.beam_size * vocab_size),
+            self.beam_size,
+            dim=1,
+        )
+
+        # Apply the length penalty. The +1 accounts for the [EOS] token
+        # that will be added if the beam ends.
+        topk_scores = topk_log_probabilities / self.length_penalty()
+
+        # Retrieve the corresponding respective beam and token id
+        # topk_token_ids[i] will be added to topk_beam_ids[i]
+        topk_beam_ids = topk_ids.div(vocab_size)
+        topk_token_ids = topk_ids.fmod(vocab_size)
+
+        # Retrieve the row index of the surviving beams in the original
+        # view of the log_probabilities tensor
+        surviving_beams_rows = (topk_beam_ids + self.beam_offset[:_B].view(-1, 1)).view(
+            -1
+        )
+
+        # Append the last predictions
+        self.growing_beam = torch.cat(
+            [
+                self.growing_beam.index_select(0, surviving_beams_rows),
+                topk_token_ids.view(-1, 1),
+            ],
+            1,
+        )
+
+        # Check if any of the beam searches has ended during this
+        # growth step. Also if top beam (most probable) has ended
+        # for one element of the batch.
+        is_finished = topk_token_ids.eq(self.end_token_id)
+        self.enforce_max_length()
+        is_top_beam_finished = is_finished[:, 0].eq(1)
+
+        # Save the finished searches
+        if is_finished.any():
+            predictions = self.growing_beam.view(
+                -1, self.beam_size, self.growing_beam.size(1)
+            )
+            for i in range(is_finished.size(0)):
+                if is_top_beam_finished[i]:
+                    is_finished[i].fill_(1)
+                finished_hyp = is_finished[i].nonzero().view(-1)
+
+                # Store finished hypotheses for this batch.
+                b = self.batch_offset[i]
+                for j in finished_hyp:
+                    self.hypotheses[b].append((topk_scores[i, j], predictions[i, j, :]))
+
+                # If the batch reached the end, save the best hypotheses
+                # in terms of length-penalized score.
+                if is_top_beam_finished[i]:
+                    best_hyp = sorted(
+                        self.hypotheses[b], key=lambda x: x[0], reverse=True
+                    )
+                    best_score, best_prediction = best_hyp[0]
+                    self.results["scores"][b].append(best_score)
+                    self.results["predictions"][b].append(best_prediction)
+
+            non_finished = is_top_beam_finished.eq(0).nonzero().view(-1)
+            if len(non_finished) == 0:
+                self.is_done = True
+
+            # Remove finished batches for the next step.
+            topk_log_probabilities = topk_log_probabilities.index_select(
+                0, non_finished
+            )
+            self.batch_offset = self.batch_offset.index_select(0, non_finished)
+            self.growing_beam = predictions.index_select(0, non_finished).view(
+                -1, self.growing_beam.size(-1)
+            )
+
+            surviving_beams_rows = surviving_beams_rows.index_select(0, non_finished)
+
+        return surviving_beams_rows
+
+    def forward(self, encoder_input_ids, **kwargs):
+        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
+        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
+        # that apply to the model as whole.
+        # We let the specific kwargs override the common ones in case of conflict.
+        kwargs_encoder = {
+            argument[len("encoder_"):]: value
+            for argument, value in kwargs.items()
+            if argument.startswith("encoder_")
+        }
+        kwargs_decoder = {
+            argument[len("decoder_"):]: value
+            for argument, value in kwargs.items()
+            if argument.startswith("decoder_")
+        }
+        kwargs_common = {
+            argument: value
+            for argument, value in kwargs.items()
+            if not (argument.startswith("encoder_") or argument.startswith("decoder_"))
+        }
+        kwargs_decoder = dict(kwargs_common, **kwargs_decoder)
+        kwargs_encoder = dict(kwargs_common, **kwargs_encoder)
+
+        # forward pass on the encoder
+        encoder_outputs = self.model.encoder.forward(encoder_input_ids, kwargs_encoder)
+        kwargs_decoder["encoder_hidden_states"] = tile(
+            encoder_outputs, self.beam_size, dim=0
+        )
+
+        # grow the beam by generating sequences in an autoregressive way
+        self.growing_beam = torch.full(
+            (self.batch_size * self.beam_size, 1), self.start_token_id, dtype=torch.long
+        )
+        for step in range(self.max_length):
+            decoder_input = self.growing_beam[:, -1]
+            outputs = self.model.decoder(decoder_input, kwargs_decoder)
+            log_probabilities = torch.nn.functional.log_softmax(outputs[1])
+            surviving_beams_rows = self.step(log_probabilities)
+            if self.is_done:
+                break
+
+            kwargs_decoder["encoder_hidden_states"] = kwargs_decoder[
+                "encoder_hidden_states"
+            ].index_select(0, surviving_beams_rows)
+
+        return self.results
+
+    def remove_repeating_trigrams(self, log_probabilities, _B):
+        if(self._step + 1 > 3):
+            for i in range(_B * self.beam_size):
+                tokens = [t for t in self.growing_beam[i]]
+                trigrams = [(tokens[i-1], tokens[i], tokens[i+1]) for i in range(1, len(words) - 1)]
+                last_trigram = tuple(trigrams[-1])
+                if last_trigram in trigrams[:-1]:
+                    log_probabilities[i] = -1e20
+
+    def enforce_min_length(self):
+        if self._step < self.min_length:
+            self.log_probabilities[self.end_token_id] = -1e20
+
+    def enforce_max_length(self):
+        if self._step + 1 == self.max_length:
+            self.is_finished.fill_(1)
+
+    def length_penalty(self):
+        return ((5.0 + (self._step + 1)) / 6.0) ** self.alpha
+
+
+def tile(x, count, dim=0):
+    """
+    Tiles `x` along dimension `dim` `count` times.
+
+    Example:
+        >> ex = torch.tensor([1,2],[3,4])
+        >> tile(ex, 2, 0)
+        torch.Tensor([[1,2],[1,2],[3,4],[3,4]])
+    """
+    perm = list(range(len(x.size())))
+    if dim != 0:
+        perm[0], perm[dim] = perm[dim], perm[0]
+        x = x.permute(perm).contiguous()
+    out_size = list(x.size())
+    out_size[0] *= count
+    batch = x.size(0)
+    x = (
+        x.view(batch, -1)
+        .transpose(0, 1)
+        .repeat(count, 1)
+        .transpose(0, 1)
+        .contiguous()
+        .view(*out_size)
+    )
+    if dim != 0:
+        x = x.permute(perm).contiguous()
+    return x

transformers/modeling_bert.py

@@ -17,12 +17,10 @@
 
 from __future__ import absolute_import, division, print_function, unicode_literals
 
-import json
 import logging
 import math
 import os
 import sys
-from io import open
 
 import torch
 from torch import nn
@@ -48,8 +46,11 @@ BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-pytorch_model.bin",
     'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-pytorch_model.bin",
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-pytorch_model.bin",
+    'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-pytorch_model.bin",
+    'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-pytorch_model.bin",
 }
 
+
 def load_tf_weights_in_bert(model, config, tf_checkpoint_path):
     """ Load tf checkpoints in a pytorch model.
     """
@@ -125,12 +126,14 @@ def gelu(x):
     """
     return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
 
+
 def gelu_new(x):
     """ Implementation of the gelu activation function currently in Google Bert repo (identical to OpenAI GPT).
         Also see https://arxiv.org/abs/1606.08415
     """
     return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
 
+
 def swish(x):
     return x * torch.sigmoid(x)
 
@@ -140,6 +143,7 @@ ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish, "gelu_
 
 BertLayerNorm = torch.nn.LayerNorm
 
+
 class BertEmbeddings(nn.Module):
     """Construct the embeddings from word, position and token_type embeddings.
     """
@@ -154,19 +158,26 @@ class BertEmbeddings(nn.Module):
         self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
-    def forward(self, input_ids, token_type_ids=None, position_ids=None):
-        seq_length = input_ids.size(1)
-        if position_ids is None:
-            position_ids = torch.arange(seq_length, dtype=torch.long, device=input_ids.device)
-            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
-        if token_type_ids is None:
-            token_type_ids = torch.zeros_like(input_ids)
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
 
-        words_embeddings = self.word_embeddings(input_ids)
+        seq_length = input_shape[1]
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+        if position_ids is None:
+            position_ids = torch.arange(seq_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).expand(input_shape)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
         position_embeddings = self.position_embeddings(position_ids)
         token_type_embeddings = self.token_type_embeddings(token_type_ids)
 
-        embeddings = words_embeddings + position_embeddings + token_type_embeddings
+        embeddings = inputs_embeds + position_embeddings + token_type_embeddings
         embeddings = self.LayerNorm(embeddings)
         embeddings = self.dropout(embeddings)
         return embeddings
@@ -196,10 +207,19 @@ class BertSelfAttention(nn.Module):
         x = x.view(*new_x_shape)
         return x.permute(0, 2, 1, 3)
 
-    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, encoder_hidden_states=None, encoder_attention_mask=None):
         mixed_query_layer = self.query(hidden_states)
-        mixed_key_layer = self.key(hidden_states)
-        mixed_value_layer = self.value(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        if encoder_hidden_states is not None:
+            mixed_key_layer = self.key(encoder_hidden_states)
+            mixed_value_layer = self.value(encoder_hidden_states)
+            attention_mask = encoder_attention_mask
+        else:
+            mixed_key_layer = self.key(hidden_states)
+            mixed_value_layer = self.value(hidden_states)
 
         query_layer = self.transpose_for_scores(mixed_query_layer)
         key_layer = self.transpose_for_scores(mixed_key_layer)
@@ -258,7 +278,7 @@ class BertAttention(nn.Module):
         if len(heads) == 0:
             return
         mask = torch.ones(self.self.num_attention_heads, self.self.attention_head_size)
-        heads = set(heads) - self.pruned_heads  # Convert to set and emove already pruned heads
+        heads = set(heads) - self.pruned_heads  # Convert to set and remove already pruned heads
         for head in heads:
             # Compute how many pruned heads are before the head and move the index accordingly
             head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
@@ -277,9 +297,9 @@ class BertAttention(nn.Module):
         self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
         self.pruned_heads = self.pruned_heads.union(heads)
 
-    def forward(self, input_tensor, attention_mask=None, head_mask=None):
-        self_outputs = self.self(input_tensor, attention_mask, head_mask)
-        attention_output = self.output(self_outputs[0], input_tensor)
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, encoder_hidden_states=None, encoder_attention_mask=None):
+        self_outputs = self.self(hidden_states, attention_mask, head_mask, encoder_hidden_states, encoder_attention_mask)
+        attention_output = self.output(self_outputs[0], hidden_states)
         outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
         return outputs
 
@@ -317,15 +337,25 @@ class BertLayer(nn.Module):
     def __init__(self, config):
         super(BertLayer, self).__init__()
         self.attention = BertAttention(config)
+        self.is_decoder = config.is_decoder
+        if self.is_decoder:
+            self.crossattention = BertAttention(config)
         self.intermediate = BertIntermediate(config)
         self.output = BertOutput(config)
 
-    def forward(self, hidden_states, attention_mask=None, head_mask=None):
-        attention_outputs = self.attention(hidden_states, attention_mask, head_mask)
-        attention_output = attention_outputs[0]
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, encoder_hidden_states=None, encoder_attention_mask=None):
+        self_attention_outputs = self.attention(hidden_states, attention_mask, head_mask)
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        if self.is_decoder and encoder_hidden_states is not None:
+            cross_attention_outputs = self.crossattention(attention_output, attention_mask, head_mask, encoder_hidden_states, encoder_attention_mask)
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:]  # add cross attentions if we output attention weights
+
         intermediate_output = self.intermediate(attention_output)
         layer_output = self.output(intermediate_output, attention_output)
-        outputs = (layer_output,) + attention_outputs[1:]  # add attentions if we output them
+        outputs = (layer_output,) + outputs
         return outputs
 
 
@@ -336,14 +366,14 @@ class BertEncoder(nn.Module):
         self.output_hidden_states = config.output_hidden_states
         self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
 
-    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, encoder_hidden_states=None, encoder_attention_mask=None):
         all_hidden_states = ()
         all_attentions = ()
         for i, layer_module in enumerate(self.layer):
             if self.output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
-            layer_outputs = layer_module(hidden_states, attention_mask, head_mask[i])
+            layer_outputs = layer_module(hidden_states, attention_mask, head_mask[i], encoder_hidden_states, encoder_attention_mask)
             hidden_states = layer_outputs[0]
 
             if self.output_attentions:
@@ -482,7 +512,7 @@ BERT_START_DOCSTRING = r"""    The BERT model was proposed in
         https://pytorch.org/docs/stable/nn.html#module
 
     Parameters:
-        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model. 
+        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model.
             Initializing with a config file does not load the weights associated with the model, only the configuration.
             Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
@@ -496,13 +526,13 @@ BERT_INPUTS_DOCSTRING = r"""
             (a) For sequence pairs:
 
                 ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
-                
+
                 ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
 
             (b) For single sequences:
 
                 ``tokens:         [CLS] the dog is hairy . [SEP]``
-                
+
                 ``token_type_ids:   0   0   0   0  0     0   0``
 
             Bert is a model with absolute position embeddings so it's usually advised to pad the inputs on
@@ -527,6 +557,18 @@ BERT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        **encoder_hidden_states**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``:
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model
+            is configured as a decoder.
+        **encoder_attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask
+            is used in the cross-attention if the model is configured as a decoder.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
 """
 
 @add_start_docstrings("The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
@@ -562,6 +604,7 @@ class BertModel(BertPreTrainedModel):
     """
     def __init__(self, config):
         super(BertModel, self).__init__(config)
+        self.config = config
 
         self.embeddings = BertEmbeddings(config)
         self.encoder = BertEncoder(config)
@@ -569,12 +612,12 @@ class BertModel(BertPreTrainedModel):
 
         self.init_weights()
 
-    def _resize_token_embeddings(self, new_num_tokens):
-        old_embeddings = self.embeddings.word_embeddings
-        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
-        self.embeddings.word_embeddings = new_embeddings
+    def get_input_embeddings(self):
         return self.embeddings.word_embeddings
 
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
             heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
@@ -583,27 +626,76 @@ class BertModel(BertPreTrainedModel):
         for layer, heads in heads_to_prune.items():
             self.encoder.layer[layer].attention.prune_heads(heads)
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
-        if attention_mask is None:
-            attention_mask = torch.ones_like(input_ids)
-        if token_type_ids is None:
-            token_type_ids = torch.zeros_like(input_ids)
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None,
+                head_mask=None, inputs_embeds=None, encoder_hidden_states=None, encoder_attention_mask=None):
+        """ Forward pass on the Model.
 
-        # We create a 3D attention mask from a 2D tensor mask.
-        # Sizes are [batch_size, 1, 1, to_seq_length]
-        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
-        # this attention mask is more simple than the triangular masking of causal attention
-        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
-        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+        The model can behave as an encoder (with only self-attention) as well
+        as a decoder, in which case a layer of cross-attention is added between
+        the self-attention layers, following the architecture described in `Attention is all you need`_ by Ashish Vaswani,
+        Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+
+        To behave as an decoder the model needs to be initialized with the
+        `is_decoder` argument of the configuration set to `True`; an
+        `encoder_hidden_states` is expected as an input to the forward pass.
+
+        .. _`Attention is all you need`:
+            https://arxiv.org/abs/1706.03762
+
+        """
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if encoder_attention_mask is None:
+            encoder_attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+
+        # Provided a padding mask of dimensions [batch_size, seq_length]
+        # - if the model is a decoder, apply a causal mask in addition to the padding mask
+        # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if attention_mask.dim() == 2:
+            if self.config.is_decoder:
+                batch_size, seq_length = input_shape
+                seq_ids = torch.arange(seq_length, device=device)
+                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
 
         # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
         # masked positions, this operation will create a tensor which is 0.0 for
         # positions we want to attend and -10000.0 for masked positions.
         # Since we are adding it to the raw scores before the softmax, this is
         # effectively the same as removing these entirely.
-        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
         extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
 
+        # If a 2D ou 3D attention mask is provided for the cross-attention
+        # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+        if encoder_attention_mask.dim() == 3:
+            encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+        if encoder_attention_mask.dim() == 2:
+            encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+
+        encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
+        encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
+
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head
         # attention_probs has shape bsz x n_heads x N x N
@@ -615,14 +707,16 @@ class BertModel(BertPreTrainedModel):
                 head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
             elif head_mask.dim() == 2:
                 head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
-            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype)  # switch to fload if need + fp16 compatibility
         else:
             head_mask = [None] * self.config.num_hidden_layers
 
-        embedding_output = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids)
+        embedding_output = self.embeddings(input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds)
         encoder_outputs = self.encoder(embedding_output,
-                                       extended_attention_mask,
-                                       head_mask=head_mask)
+                                       attention_mask=extended_attention_mask,
+                                       head_mask=head_mask,
+                                       encoder_hidden_states=encoder_hidden_states,
+                                       encoder_attention_mask=encoder_extended_attention_mask)
         sequence_output = encoder_outputs[0]
         pooled_output = self.pooler(sequence_output)
 
@@ -631,8 +725,9 @@ class BertModel(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model with two heads on top as done during the pre-training:
-    a `masked language modeling` head and a `next sentence prediction (classification)` head. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                       a `masked language modeling` head and a `next sentence prediction (classification)` head. """,
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForPreTraining(BertPreTrainedModel):
     r"""
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
@@ -677,23 +772,19 @@ class BertForPreTraining(BertPreTrainedModel):
         self.cls = BertPreTrainingHeads(config)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.cls.predictions.decoder,
-                                   self.bert.embeddings.word_embeddings)
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 masked_lm_labels=None, next_sentence_label=None):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         sequence_output, pooled_output = outputs[:2]
         prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)
@@ -711,7 +802,8 @@ class BertForPreTraining(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model with a `language modeling` head on top. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForMaskedLM(BertPreTrainedModel):
     r"""
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
@@ -719,10 +811,17 @@ class BertForMaskedLM(BertPreTrainedModel):
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
             Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
+        **lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the left-to-right language modeling loss (next word prediction).
+            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+        **masked_lm_loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
             Masked language modeling loss.
+        **ltr_lm_loss**: (`optional`, returned when ``lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Next token prediction loss.
         **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
@@ -749,38 +848,52 @@ class BertForMaskedLM(BertPreTrainedModel):
         self.cls = BertOnlyMLMHead(config)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.cls.predictions.decoder,
-                                   self.bert.embeddings.word_embeddings)
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
-                masked_lm_labels=None):
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
+                masked_lm_labels=None, encoder_hidden_states=None, encoder_attention_mask=None, lm_labels=None, ):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds,
+                            encoder_hidden_states=encoder_hidden_states,
+                            encoder_attention_mask=encoder_attention_mask)
 
         sequence_output = outputs[0]
         prediction_scores = self.cls(sequence_output)
 
         outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+
+        # Although this may seem awkward, BertForMaskedLM supports two scenarios:
+        # 1. If a tensor that contains the indices of masked labels is provided,
+        #    the cross-entropy is the MLM cross-entropy that measures the likelihood
+        #    of predictions for masked words.
+        # 2. If `lm_labels` is provided we are in a causal scenario where we
+        #    try to predict the next token for each input in the decoder.
         if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss(ignore_index=-1)  # -1 index = padding token
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             outputs = (masked_lm_loss,) + outputs
 
-        return outputs  # (masked_lm_loss), prediction_scores, (hidden_states), (attentions)
+        if lm_labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            prediction_scores = prediction_scores[:, :-1, :].contiguous()
+            lm_labels = lm_labels[:, 1:].contiguous()
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            ltr_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), lm_labels.view(-1))
+            outputs = (ltr_lm_loss,) + outputs
+
+        return outputs  # (masked_lm_loss), (ltr_lm_loss), prediction_scores, (hidden_states), (attentions)
 
 
 @add_start_docstrings("""Bert Model with a `next sentence prediction (classification)` head on top. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForNextSentencePrediction(BertPreTrainedModel):
     r"""
         **next_sentence_label**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
@@ -819,14 +932,15 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 next_sentence_label=None):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         pooled_output = outputs[1]
 
@@ -842,8 +956,9 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model transformer with a sequence classification/regression head on top (a linear layer on top of
-    the pooled output) e.g. for GLUE tasks. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      the pooled output) e.g. for GLUE tasks. """,
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForSequenceClassification(BertPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
@@ -885,14 +1000,15 @@ class BertForSequenceClassification(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
-                position_ids=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         pooled_output = outputs[1]
 
@@ -915,8 +1031,9 @@ class BertForSequenceClassification(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model with a multiple choice classification head on top (a linear layer on top of
-    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForMultipleChoice(BertPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
@@ -958,8 +1075,8 @@ class BertForMultipleChoice(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
-                position_ids=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
         num_choices = input_ids.shape[1]
 
         input_ids = input_ids.view(-1, input_ids.size(-1))
@@ -971,7 +1088,8 @@ class BertForMultipleChoice(BertPreTrainedModel):
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
                             position_ids=position_ids,
-                            head_mask=head_mask)
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         pooled_output = outputs[1]
 
@@ -990,8 +1108,9 @@ class BertForMultipleChoice(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model with a token classification head on top (a linear layer on top of
-    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForTokenClassification(BertPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
@@ -1031,14 +1150,15 @@ class BertForTokenClassification(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
-                position_ids=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         sequence_output = outputs[0]
 
@@ -1062,8 +1182,9 @@ class BertForTokenClassification(BertPreTrainedModel):
 
 
 @add_start_docstrings("""Bert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
-    the hidden-states output to compute `span start logits` and `span end logits`). """,
-    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+                      the hidden-states output to compute `span start logits` and `span end logits`). """,
+                      BERT_START_DOCSTRING,
+                      BERT_INPUTS_DOCSTRING)
 class BertForQuestionAnswering(BertPreTrainedModel):
     r"""
         **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
@@ -1093,12 +1214,16 @@ class BertForQuestionAnswering(BertPreTrainedModel):
     Examples::
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-        model = BertForQuestionAnswering.from_pretrained('bert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        start_positions = torch.tensor([1])
-        end_positions = torch.tensor([3])
-        outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
-        loss, start_scores, end_scores = outputs[:2]
+        model = BertForQuestionAnswering.from_pretrained('bert-large-uncased-whole-word-masking-finetuned-squad')
+        question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        input_text = "[CLS] " + question + " [SEP] " + text + " [SEP]"
+        input_ids = tokenizer.encode(input_text)
+        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))] 
+        start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))
+        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)  
+        print(' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]))
+        # a nice puppet
+
 
     """
     def __init__(self, config):
@@ -1110,14 +1235,15 @@ class BertForQuestionAnswering(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 start_positions=None, end_positions=None):
 
         outputs = self.bert(input_ids,
                             attention_mask=attention_mask,
                             token_type_ids=token_type_ids,
-                            position_ids=position_ids, 
-                            head_mask=head_mask)
+                            position_ids=position_ids,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
 
         sequence_output = outputs[0]
 

transformers/modeling_camembert.py

@@ -0,0 +1,293 @@
+# coding=utf-8
+# Copyright 2019 Inria, Facebook AI Research and the HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch CamemBERT model. """
+
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import logging
+
+from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification, RobertaForMultipleChoice, RobertaForTokenClassification
+from .configuration_camembert import CamembertConfig
+from .file_utils import add_start_docstrings
+
+logger = logging.getLogger(__name__)
+
+CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'camembert-base': "https://s3.amazonaws.com/models.huggingface.co/bert/camembert-base-pytorch_model.bin",
+}
+
+
+CAMEMBERT_START_DOCSTRING = r"""    The CamemBERT model was proposed in
+    `CamemBERT: a Tasty French Language Model`_
+    by Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suárez, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah, and Benoît Sagot. It is based on Facebook's RoBERTa model released in 2019.
+    
+    It is a model trained on 138GB of French text.
+    
+    This implementation is the same as RoBERTa.
+
+    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+    refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+    .. _`CamemBERT: a Tasty French Language Model`:
+        https://arxiv.org/abs/1911.03894
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~transformers.CamembertConfig`): Model configuration class with all the parameters of the 
+            model. Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+CAMEMBERT_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, CamemBERT input sequence should be formatted with <s> and </s> tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         <s> Is this Jacksonville ? </s> </s> No it is not . </s>``
+
+            (b) For single sequences:
+
+                ``tokens:         <s> the dog is hairy . </s>``
+
+            Fully encoded sequences or sequence pairs can be obtained using the CamembertTokenizer.encode function with 
+            the ``add_special_tokens`` parameter set to ``True``.
+
+            CamemBERT is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **token_type_ids**: (`optional` need to be trained) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Optional segment token indices to indicate first and second portions of the inputs.
+            This embedding matrice is not trained (not pretrained during CamemBERT pretraining), you will have to train it
+            during finetuning.
+            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
+            corresponds to a `sentence B` token
+            (see `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
+        **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1[``.
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+"""
+
+@add_start_docstrings("The bare CamemBERT Model transformer outputting raw hidden-states without any specific head on top.",
+                      CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertModel(RobertaModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            eo match pre-training, CamemBERT input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] [SEP] no it is not . [SEP]``
+
+                ``token_type_ids:   0   0  0    0    0     0       0   0   0     1  1  1  1   1   1``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+
+                ``token_type_ids:   0   0   0   0  0     0   0``
+
+            objective during Bert pretraining. This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertModel.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model with a `language modeling` head on top. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForMaskedLM(RobertaForMaskedLM):
+    r"""
+        **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForMaskedLM.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, masked_lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model transformer with a sequence classification/regression head on top (a linear layer 
+    on top of the pooled output) e.g. for GLUE tasks. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForSequenceClassification(RobertaForSequenceClassification):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in ``[0, ..., config.num_labels]``.
+            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
+            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification (or regression if config.num_labels==1) loss.
+        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForSequenceClassification.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model with a multiple choice classification head on top (a linear layer on top of
+    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForMultipleChoice(RobertaForMultipleChoice):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss.
+        **classification_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)`` where `num_choices` is the size of the second dimension
+            of the input tensors. (see `input_ids` above).
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForMultipleChoice.from_pretrained('camembert-base')
+        choices = ["J'aime le camembert !", "Je deteste le camembert !"]
+        input_ids = torch.tensor([tokenizer.encode(s, add_special_tokens=True) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
+        labels = torch.tensor(1).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, classification_scores = outputs[:2]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForTokenClassification(RobertaForTokenClassification):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the token classification loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss.
+        **scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForTokenClassification.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, scores = outputs[:2]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP

transformers/modeling_ctrl.py

@@ -220,7 +220,8 @@ CTRL_INPUTS_DOCSTRING = r"""    Inputs:
         **past**:
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
-            (see `past` output below). Can be used to speed up sequential decoding.
+            (see `past` output below). Can be used to speed up sequential decoding. The token ids which have their past given to this model 
+            should not be passed as input ids as they have already been computed.
         **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -236,6 +237,10 @@ CTRL_INPUTS_DOCSTRING = r"""    Inputs:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare CTRL Model transformer outputting raw hidden-states without any specific head on top.",
@@ -248,7 +253,8 @@ class CTRLModel(CTRLPreTrainedModel):
         **past**:
             list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding.
+            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
+            should not be passed as input ids as they have already been computed.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -269,16 +275,16 @@ class CTRLModel(CTRLPreTrainedModel):
     def __init__(self, config):
         super(CTRLModel, self).__init__(config)
         self.output_hidden_states = config.output_hidden_states
+        self.output_attentions = config.output_attentions
+        self.output_past = config.output_past
+
         self.d_model_size = config.n_embd
         self.num_layers = config.n_layer
-        
+
         self.pos_encoding = positional_encoding(config.n_positions, self.d_model_size, torch.float)
-            
-        self.output_attentions = config.output_attentions
 
         self.w = nn.Embedding(config.vocab_size, config.n_embd)
 
-        
         self.dropout = nn.Dropout(config.embd_pdrop)
         self.h = nn.ModuleList([EncoderLayer(config.n_embd,
                                              config.n_head,
@@ -289,10 +295,12 @@ class CTRLModel(CTRLPreTrainedModel):
 
         self.init_weights()
 
-    def _resize_token_embeddings(self, new_num_tokens):
-        self.w = self._get_resized_embeddings(self.w, new_num_tokens)
+    def get_input_embeddings(self):
         return self.w
 
+    def set_input_embeddings(self, new_embeddings):
+        self.w = new_embeddings
+
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
                 heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
@@ -300,17 +308,26 @@ class CTRLModel(CTRLPreTrainedModel):
         for layer, heads in heads_to_prune.items():
             self.h[layer].attn.prune_heads(heads)
 
-    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
-        input_shape = input_ids.size()
-        input_ids = input_ids.view(-1, input_shape[-1])
+    def forward(self, input_ids=None, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None):
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         if past is None:
             past_length = 0
             past = [None] * len(self.h)
         else:
             past_length = past[0][0].size(-2)
         if position_ids is None:
-            position_ids = torch.arange(past_length, input_ids.size(-1) + past_length, dtype=torch.long, device=input_ids.device)
-            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
 
         # Attention mask.
         if attention_mask is not None:
@@ -352,9 +369,10 @@ class CTRLModel(CTRLPreTrainedModel):
             token_type_embeds = 0
         position_ids = position_ids.view(-1, input_shape[-1])
 
-        inputs_embeds = self.w(input_ids)
+        if inputs_embeds is None:
+            inputs_embeds = self.w(input_ids)
         # inputs_embeds = embedded.unsqueeze(0) if len(input_ids.shape)<2 else embedded
-        seq_len = input_ids.shape[-1]
+        seq_len = input_shape[-1]
         mask = torch.triu(torch.ones(seq_len, seq_len), 1).to(inputs_embeds.device)
 
         inputs_embeds *= np.sqrt(self.d_model_size)
@@ -378,7 +396,8 @@ class CTRLModel(CTRLPreTrainedModel):
                         attention_mask=attention_mask,
                         head_mask=head_mask[i])
             hidden_states, present = outputs[:2]
-            presents = presents + (present,)
+            if self.output_past:
+                presents = presents + (present,)
 
             if self.output_attentions:
                 all_attentions.append(outputs[2])
@@ -388,7 +407,9 @@ class CTRLModel(CTRLPreTrainedModel):
         if self.output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states, presents)
+        outputs = (hidden_states,)
+        if self.output_past:
+            outputs = outputs + (presents,)
         if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if self.output_attentions:
@@ -418,7 +439,8 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         **past**:
             list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding.
+            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
+            should not be passed as input ids as they have already been computed.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -446,22 +468,19 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=True)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-                Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.lm_head, self.transformer.w)
+    def get_output_embeddings(self):
+        return self.lm_head
 
-    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 labels=None):
         transformer_outputs = self.transformer(input_ids,
                                                past=past,
                                                attention_mask=attention_mask,
                                                token_type_ids=token_type_ids,
                                                position_ids=position_ids,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
 
         hidden_states = transformer_outputs[0]
 

transformers/modeling_distilbert.py

@@ -30,6 +30,7 @@ import numpy as np
 
 import torch
 import torch.nn as nn
+from torch.nn import CrossEntropyLoss
 
 from .modeling_utils import PreTrainedModel, prune_linear_layer
 from .configuration_distilbert import DistilBertConfig
@@ -334,9 +335,6 @@ class DistilBertPreTrainedModel(PreTrainedModel):
     load_tf_weights = None
     base_model_prefix = "distilbert"
 
-    def __init__(self, *inputs, **kwargs):
-        super(DistilBertPreTrainedModel, self).__init__(*inputs, **kwargs)
-    
     def _init_weights(self, module):
         """ Initialize the weights.
         """
@@ -390,6 +388,10 @@ DISTILBERT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare DistilBERT encoder/transformer outputting raw hidden-states without any specific head on top.",
@@ -424,12 +426,12 @@ class DistilBertModel(DistilBertPreTrainedModel):
 
         self.init_weights()
 
-    def _resize_token_embeddings(self, new_num_tokens):
-        old_embeddings = self.embeddings.word_embeddings
-        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
-        self.embeddings.word_embeddings = new_embeddings
+    def get_input_embeddings(self):
         return self.embeddings.word_embeddings
 
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings.word_embeddings = new_embeddings
+
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
             heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
@@ -439,9 +441,20 @@ class DistilBertModel(DistilBertPreTrainedModel):
             self.transformer.layer[layer].attention.prune_heads(heads)
 
     def forward(self,
-                input_ids, attention_mask=None, head_mask=None):
+                input_ids=None, attention_mask=None, head_mask=None, inputs_embeds=None):
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
         if attention_mask is None:
-            attention_mask = torch.ones_like(input_ids) # (bs, seq_length)
+            attention_mask = torch.ones(input_shape, device=device) # (bs, seq_length)
 
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head
@@ -458,8 +471,9 @@ class DistilBertModel(DistilBertPreTrainedModel):
         else:
             head_mask = [None] * self.config.num_hidden_layers
 
-        embedding_output = self.embeddings(input_ids)   # (bs, seq_length, dim)
-        tfmr_output = self.transformer(x=embedding_output,
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)   # (bs, seq_length, dim)
+        tfmr_output = self.transformer(x=inputs_embeds,
                                        attn_mask=attention_mask,
                                        head_mask=head_mask)
         hidden_state = tfmr_output[0]
@@ -511,21 +525,17 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
         self.vocab_projector = nn.Linear(config.dim, config.vocab_size)
 
         self.init_weights()
-        self.tie_weights()
 
         self.mlm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.vocab_projector,
-                                   self.distilbert.embeddings.word_embeddings)
+    def get_output_embeddings(self):
+        return self.vocab_projector
 
-    def forward(self, input_ids, attention_mask=None, head_mask=None, masked_lm_labels=None):
+    def forward(self, input_ids=None, attention_mask=None, head_mask=None, inputs_embeds=None, masked_lm_labels=None):
         dlbrt_output = self.distilbert(input_ids=input_ids,
                                        attention_mask=attention_mask,
-                                       head_mask=head_mask)
+                                       head_mask=head_mask,
+                                       inputs_embeds=inputs_embeds)
         hidden_states = dlbrt_output[0]                              # (bs, seq_length, dim)
         prediction_logits = self.vocab_transform(hidden_states)      # (bs, seq_length, dim)
         prediction_logits = gelu(prediction_logits)                  # (bs, seq_length, dim)
@@ -586,10 +596,11 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids,  attention_mask=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, head_mask=None, inputs_embeds=None, labels=None):
         distilbert_output = self.distilbert(input_ids=input_ids,
                                             attention_mask=attention_mask,
-                                            head_mask=head_mask)
+                                            head_mask=head_mask,
+                                            inputs_embeds=inputs_embeds)
         hidden_state = distilbert_output[0]                    # (bs, seq_len, dim)
         pooled_output = hidden_state[:, 0]                    # (bs, dim)
         pooled_output = self.pre_classifier(pooled_output)   # (bs, dim)
@@ -660,10 +671,11 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
 
         self.init_weights()
         
-    def forward(self, input_ids, attention_mask=None, head_mask=None, start_positions=None, end_positions=None):
+    def forward(self, input_ids=None, attention_mask=None, head_mask=None, inputs_embeds=None, start_positions=None, end_positions=None):
         distilbert_output = self.distilbert(input_ids=input_ids,
                                             attention_mask=attention_mask,
-                                            head_mask=head_mask)
+                                            head_mask=head_mask,
+                                            inputs_embeds=inputs_embeds)
         hidden_states = distilbert_output[0]                                 # (bs, max_query_len, dim)
 
         hidden_states = self.dropout(hidden_states)                       # (bs, max_query_len, dim)
@@ -691,3 +703,75 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
             outputs = (total_loss,) + outputs
 
         return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""DistilBert Model with a token classification head on top (a linear layer on top of
+                      the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+                      DISTILBERT_START_DOCSTRING,
+                      DISTILBERT_INPUTS_DOCSTRING)
+class DistilBertForTokenClassification(DistilBertPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the token classification loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss.
+        **scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
+        model = DistilBertForTokenClassification.from_pretrained('distilbert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(DistilBertForTokenClassification, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.distilbert = DistilBertModel(config)
+        self.dropout = nn.Dropout(config.dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids=None, attention_mask=None, head_mask=None,
+                inputs_embeds=None, labels=None):
+
+        outputs = self.distilbert(input_ids,
+                            attention_mask=attention_mask,
+                            head_mask=head_mask,
+                            inputs_embeds=inputs_embeds)
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)[active_loss]
+                active_labels = labels.view(-1)[active_loss]
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), scores, (hidden_states), (attentions)

transformers/modeling_encoder_decoder.py

@@ -0,0 +1,310 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Classes to support Encoder-Decoder architectures """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import logging
+import os
+
+import torch
+from torch import nn
+
+from .modeling_auto import AutoModel, AutoModelWithLMHead
+
+logger = logging.getLogger(__name__)
+
+
+class PreTrainedEncoderDecoder(nn.Module):
+    r"""
+        :class:`~transformers.PreTrainedEncoderDecoder` is a generic model class that will be
+        instantiated as a transformer architecture with one of the base model
+        classes of the library as encoder and (optionally) another one as
+        decoder when created with the `AutoModel.from_pretrained(pretrained_model_name_or_path)`
+        class method.
+    """
+
+    def __init__(self, encoder, decoder):
+        super(PreTrainedEncoderDecoder, self).__init__()
+        self.encoder = encoder
+        self.decoder = decoder
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        encoder_pretrained_model_name_or_path=None,
+        decoder_pretrained_model_name_or_path=None,
+        *model_args,
+        **kwargs
+    ):
+        r""" Instantiates an encoder and a decoder from one or two base classes of the library from pre-trained model checkpoints.
+
+
+        The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
+        To train the model, you need to first set it back in training mode with `model.train()`
+
+        Params:
+            encoder_pretrained_model_name_or_path: information necessary to initiate the encoder. Either:
+
+                - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/encoder``.
+                - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
+
+            decoder_pretrained_model_name_or_path: information necessary to initiate the decoder. Either:
+
+                - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/decoder``.
+                - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
+
+            model_args: (`optional`) Sequence of positional arguments:
+                All remaning positional arguments will be passed to the underlying model's ``__init__`` method
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:
+
+                - the model is a model provided by the library (loaded with the ``shortcut-name`` string of a pretrained model), or
+                - the model was saved using :func:`~transformers.PreTrainedModel.save_pretrained` and is reloaded by suppling the save directory.
+                - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
+
+            state_dict: (`optional`) dict:
+                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                This option can be used if you want to create a model from a pretrained configuration but load your own weights.
+                In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
+
+            cache_dir: (`optional`) string:
+                Path to a directory in which a downloaded pre-trained model
+                configuration should be cached if the standard cache should not be used.
+
+            force_download: (`optional`) boolean, default False:
+                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            output_loading_info: (`optional`) boolean:
+                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+
+            kwargs: (`optional`) Remaining dictionary of keyword arguments.
+                Can be used to update the configuration object (after it being loaded) and initiate the model. (e.g. ``output_attention=True``). Behave differently depending on whether a `config` is provided or automatically loaded:
+
+                - If a configuration is provided with ``config``, ``**kwargs`` will be directly passed to the underlying model's ``__init__`` method (we assume all relevant updates to the configuration have already been done)
+                - If a configuration is not provided, ``kwargs`` will be first passed to the configuration class initialization function (:func:`~transformers.PretrainedConfig.from_pretrained`). Each key of ``kwargs`` that corresponds to a configuration attribute will be used to override said attribute with the supplied ``kwargs`` value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model's ``__init__`` function.
+
+                You can specify kwargs sepcific for the encoder and decoder by prefixing the key with `encoder_` and `decoder_` respectively. (e.g. ``decoder_output_attention=True``). The remaining kwargs will be passed to both encoders and decoders.
+
+        Examples::
+
+            model = PreTrainedEncoderDecoder.from_pretained('bert-base-uncased', 'bert-base-uncased') # initialize Bert2Bert
+        """
+
+        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
+        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
+        # that apply to the model as a whole.
+        # We let the specific kwargs override the common ones in case of conflict.
+        kwargs_common = {
+            argument: value
+            for argument, value in kwargs.items()
+            if not argument.startswith("encoder_")
+            and not argument.startswith("decoder_")
+        }
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_encoder.update(
+            {
+                argument[len("encoder_") :]: value
+                for argument, value in kwargs.items()
+                if argument.startswith("encoder_")
+            }
+        )
+        kwargs_decoder.update(
+            {
+                argument[len("decoder_") :]: value
+                for argument, value in kwargs.items()
+                if argument.startswith("decoder_")
+            }
+        )
+
+        # Load and initialize the encoder and decoder
+        # The distinction between encoder and decoder at the model level is made
+        # by the value of the flag `is_decoder` that we need to set correctly.
+        encoder = kwargs_encoder.pop("model", None)
+        if encoder is None:
+            encoder = AutoModel.from_pretrained(
+                encoder_pretrained_model_name_or_path, *model_args, **kwargs_encoder
+            )
+        encoder.config.is_decoder = False
+
+        decoder = kwargs_decoder.pop("model", None)
+        if decoder is None:
+            decoder = AutoModelWithLMHead.from_pretrained(
+                decoder_pretrained_model_name_or_path, **kwargs_decoder
+            )
+        decoder.config.is_decoder = True
+
+        model = cls(encoder, decoder)
+
+        return model
+
+    def save_pretrained(self, save_directory):
+        """ Save a Seq2Seq model and its configuration file in a format such
+        that it can be loaded using `:func:`~transformers.PreTrainedEncoderDecoder.from_pretrained`
+
+        We save the encoder' and decoder's parameters in two separate directories.
+        """
+        self.encoder.save_pretrained(os.path.join(save_directory, "encoder"))
+        self.decoder.save_pretrained(os.path.join(save_directory, "decoder"))
+
+    def forward(self, encoder_input_ids, decoder_input_ids, **kwargs):
+        """ The forward pass on a seq2eq depends what we are performing:
+
+        - During training we perform one forward pass through both the encoder
+          and decoder;
+        - During prediction, we perform one forward pass through the encoder,
+          and then perform several forward passes with the encoder's hidden
+          state through the decoder to decode a full sequence.
+
+        Therefore, we skip the forward pass on the encoder if an argument named
+        `encoder_hidden_state` is passed to this function.
+
+        Params:
+            encoder_input_ids: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``
+                Indices of encoder input sequence tokens in the vocabulary.
+            decoder_input_ids: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``
+                Indices of decoder input sequence tokens in the vocabulary.
+            kwargs: (`optional`) Remaining dictionary of keyword arguments.
+        """
+        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
+        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
+        # that apply to the model as whole.
+        # We let the specific kwargs override the common ones in case of conflict.
+        kwargs_common = {
+            argument: value
+            for argument, value in kwargs.items()
+            if not argument.startswith("encoder_")
+            and not argument.startswith("decoder_")
+        }
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_encoder.update(
+            {
+                argument[len("encoder_") :]: value
+                for argument, value in kwargs.items()
+                if argument.startswith("encoder_")
+            }
+        )
+        kwargs_decoder.update(
+            {
+                argument[len("decoder_") :]: value
+                for argument, value in kwargs.items()
+                if argument.startswith("decoder_")
+            }
+        )
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        if encoder_hidden_states is None:
+            encoder_outputs = self.encoder(encoder_input_ids, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[
+                0
+            ]  # output the last layer hidden state
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get(
+            "attention_mask", None
+        )
+        decoder_outputs = self.decoder(decoder_input_ids, **kwargs_decoder)
+
+        return decoder_outputs + encoder_outputs
+
+
+class Model2Model(PreTrainedEncoderDecoder):
+    r"""
+        :class:`~transformers.Model2Model` instantiates a Seq2Seq2 model
+        where both of the encoder and decoder are of the same family. If the
+        name of or that path to a pretrained model is specified the encoder and
+        the decoder will be initialized with the pretrained weight (the
+        cross-attention will be intialized randomly if its weights are not
+        present).
+
+        It is possible to override this behavior and initialize, say, the decoder randomly
+        by creating it beforehand as follows
+
+            config = BertConfig.from_pretrained()
+            decoder = BertForMaskedLM(config)
+            model = Model2Model.from_pretrained('bert-base-uncased', decoder_model=decoder)
+    """
+
+    def __init__(self, *args, **kwargs):
+        super(Model2Model, self).__init__(*args, **kwargs)
+        self.tie_weights()
+
+    def tie_weights(self):
+        """ Tying the encoder and decoders' embeddings together.
+
+       We need for each to get down to the embedding weights. However the
+        different model classes are inconsistent to that respect:
+        - BertModel: embeddings.word_embeddings
+        - RoBERTa: embeddings.word_embeddings
+        - XLMModel: embeddings
+        - GPT2: wte
+        - BertForMaskedLM: bert.embeddings.word_embeddings
+        - RobertaForMaskedLM: roberta.embeddings.word_embeddings
+
+        argument of the XEmbedding layer for each model, but it is "blocked"
+        by a model-specific keyword (bert, )...
+        """
+        # self._tie_or_clone_weights(self.encoder, self.decoder)
+        pass
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
+
+        if (
+            "bert" not in pretrained_model_name_or_path
+            or "roberta" in pretrained_model_name_or_path
+            or "distilbert" in pretrained_model_name_or_path
+        ):
+            raise ValueError("Only the Bert model is currently supported.")
+
+        model = super(Model2Model, cls).from_pretrained(
+            encoder_pretrained_model_name_or_path=pretrained_model_name_or_path,
+            decoder_pretrained_model_name_or_path=pretrained_model_name_or_path,
+            *args,
+            **kwargs
+        )
+
+        return model
+
+
+class Model2LSTM(PreTrainedEncoderDecoder):
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        if kwargs.get("decoder_model", None) is None:
+            # We will create a randomly initilized LSTM model as decoder
+            if "decoder_config" not in kwargs:
+                raise ValueError(
+                    "To load an LSTM in Encoder-Decoder model, please supply either: "
+                    "    - a torch.nn.LSTM model as `decoder_model` parameter (`decoder_model=lstm_model`), or"
+                    "    - a dictionary of configuration parameters that will be used to initialize a"
+                    "      torch.nn.LSTM model as `decoder_config` keyword argument. "
+                    "      E.g. `decoder_config={'input_size': 768, 'hidden_size': 768, 'num_layers': 2}`"
+                )
+            kwargs["decoder_model"] = torch.nn.LSTM(kwargs.pop("decoder_config"))
+        model = super(Model2LSTM, cls).from_pretrained(*args, **kwargs)
+        return model

transformers/modeling_gpt2.py

@@ -39,6 +39,7 @@ logger = logging.getLogger(__name__)
 GPT2_PRETRAINED_MODEL_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-pytorch_model.bin",
                                      "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-pytorch_model.bin",
                                      "gpt2-large": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-large-pytorch_model.bin",
+                                     "gpt2-xl": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-xl-pytorch_model.bin",
                                      "distilgpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/distilgpt2-pytorch_model.bin",}
 
 def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
@@ -297,7 +298,8 @@ GPT2_INPUTS_DOCSTRING = r"""    Inputs:
         **past**:
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
-            (see `past` output below). Can be used to speed up sequential decoding.
+            (see `past` output below). Can be used to speed up sequential decoding. The token ids which have their past given to this model 
+            should not be passed as input ids as they have already been computed.
         **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -313,6 +315,10 @@ GPT2_INPUTS_DOCSTRING = r"""    Inputs:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
@@ -325,7 +331,8 @@ class GPT2Model(GPT2PreTrainedModel):
         **past**:
             list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding.
+            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
+            should not be passed as input ids as they have already been computed.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -347,6 +354,7 @@ class GPT2Model(GPT2PreTrainedModel):
         super(GPT2Model, self).__init__(config)
         self.output_hidden_states = config.output_hidden_states
         self.output_attentions = config.output_attentions
+        self.output_past = config.output_past
 
         self.wte = nn.Embedding(config.vocab_size, config.n_embd)
         self.wpe = nn.Embedding(config.n_positions, config.n_embd)
@@ -356,10 +364,12 @@ class GPT2Model(GPT2PreTrainedModel):
 
         self.init_weights()
 
-    def _resize_token_embeddings(self, new_num_tokens):
-        self.wte = self._get_resized_embeddings(self.wte, new_num_tokens)
+    def get_input_embeddings(self):
         return self.wte
 
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
             heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
@@ -367,9 +377,17 @@ class GPT2Model(GPT2PreTrainedModel):
         for layer, heads in heads_to_prune.items():
             self.h[layer].attn.prune_heads(heads)
 
-    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
-        input_shape = input_ids.size()
-        input_ids = input_ids.view(-1, input_shape[-1])
+    def forward(self, input_ids=None, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None):
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         if token_type_ids is not None:
             token_type_ids = token_type_ids.view(-1, input_shape[-1])
         if position_ids is not None:
@@ -381,8 +399,9 @@ class GPT2Model(GPT2PreTrainedModel):
         else:
             past_length = past[0][0].size(-2)
         if position_ids is None:
-            position_ids = torch.arange(past_length, input_ids.size(-1) + past_length, dtype=torch.long, device=input_ids.device)
-            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
 
         # Attention mask.
         if attention_mask is not None:
@@ -416,7 +435,8 @@ class GPT2Model(GPT2PreTrainedModel):
         else:
             head_mask = [None] * self.config.n_layer
 
-        inputs_embeds = self.wte(input_ids)
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
         position_embeds = self.wpe(position_ids)
         if token_type_ids is not None:
             token_type_embeds = self.wte(token_type_ids)
@@ -440,7 +460,8 @@ class GPT2Model(GPT2PreTrainedModel):
                             head_mask=head_mask[i])
 
             hidden_states, present = outputs[:2]
-            presents = presents + (present,)
+            if self.output_past:
+                presents = presents + (present,)
 
             if self.output_attentions:
                 all_attentions.append(outputs[2])
@@ -452,7 +473,9 @@ class GPT2Model(GPT2PreTrainedModel):
         if self.output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states, presents)
+        outputs = (hidden_states,)
+        if self.output_past:
+            outputs = outputs + (presents,)
         if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if self.output_attentions:
@@ -460,7 +483,7 @@ class GPT2Model(GPT2PreTrainedModel):
             attention_output_shape = input_shape[:-1] + (-1,) + all_attentions[0].shape[-2:]
             all_attentions = tuple(t.view(*attention_output_shape) for t in all_attentions)
             outputs = outputs + (all_attentions,)
-        return outputs  # last hidden state, presents, (all hidden_states), (attentions)
+        return outputs  # last hidden state, (presents), (all hidden_states), (attentions)
 
 
 @add_start_docstrings("""The GPT2 Model transformer with a language modeling head on top
@@ -482,7 +505,8 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         **past**:
             list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding.
+            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
+            should not be passed as input ids as they have already been computed.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -510,23 +534,19 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.lm_head,
-                                   self.transformer.wte)
+    def get_output_embeddings(self):
+        return self.lm_head
 
-    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 labels=None):
         transformer_outputs = self.transformer(input_ids,
                                                past=past,
                                                attention_mask=attention_mask,
                                                token_type_ids=token_type_ids,
                                                position_ids=position_ids,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
         hidden_states = transformer_outputs[0]
 
         lm_logits = self.lm_head(hidden_states)
@@ -578,7 +598,8 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         **past**:
             list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding.
+            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
+            should not be passed as input ids as they have already been computed.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -618,23 +639,19 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         self.multiple_choice_head = SequenceSummary(config)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.lm_head,
-                                   self.transformer.wte)
+    def get_output_embeddings(self):
+        return self.lm_head
 
-    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 mc_token_ids=None, lm_labels=None, mc_labels=None):
         transformer_outputs = self.transformer(input_ids,
                                                past=past,
                                                attention_mask=attention_mask,
                                                token_type_ids=token_type_ids,
                                                position_ids=position_ids,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
 
         hidden_states = transformer_outputs[0]
 

transformers/modeling_openai.py

@@ -322,6 +322,10 @@ OPENAI_GPT_INPUTS_DOCSTRING = r"""    Inputs:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare OpenAI GPT transformer model outputting raw hidden-states without any specific head on top.",
@@ -360,10 +364,12 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
 
         self.init_weights()
 
-    def _resize_token_embeddings(self, new_num_tokens):
-        self.tokens_embed = self._get_resized_embeddings(self.tokens_embed, new_num_tokens)
+    def get_input_embeddings(self):
         return self.tokens_embed
 
+    def set_input_embeddings(self, new_embeddings):
+        self.tokens_embed = new_embeddings
+
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
             heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
@@ -371,14 +377,22 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
         for layer, heads in heads_to_prune.items():
             self.h[layer].attn.prune_heads(heads)
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None):
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         if position_ids is None:
-            # This was used when we had a single embedding matrice from position and token embeddings
-            # start = self.config.vocab_size + self.config.n_special
-            # end = start + input_ids.size(-1)
-            # position_ids = torch.arange(start, end, dtype=torch.long, device=input_ids.device)
-            position_ids = torch.arange(input_ids.size(-1), dtype=torch.long, device=input_ids.device)
-            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+            # Code is different from when we had a single embedding matrice from position and token embeddings
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(input_shape[-1], dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
 
         # Attention mask.
         if attention_mask is not None:
@@ -411,11 +425,8 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
         else:
             head_mask = [None] * self.config.n_layer
 
-        input_shape = input_ids.size()
-        input_ids = input_ids.view(-1, input_ids.size(-1))
-        position_ids = position_ids.view(-1, position_ids.size(-1))
-
-        inputs_embeds = self.tokens_embed(input_ids)
+        if inputs_embeds is None:
+            inputs_embeds = self.tokens_embed(input_ids)
         position_embeds = self.positions_embed(position_ids)
         if token_type_ids is not None:
             token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1))
@@ -489,22 +500,18 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.lm_head,
-                                   self.transformer.tokens_embed)
+    def get_output_embeddings(self):
+        return self.lm_head
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 labels=None):
         transformer_outputs = self.transformer(input_ids,
                                                attention_mask=attention_mask,
                                                token_type_ids=token_type_ids,
                                                position_ids=position_ids,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
         hidden_states = transformer_outputs[0]
         lm_logits = self.lm_head(hidden_states)
 
@@ -568,9 +575,12 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = OpenAIGPTDoubleHeadsModel.from_pretrained('openai-gpt')
         tokenizer.add_special_tokens({'cls_token': '[CLS]'})  # Add a [CLS] to the vocabulary (we should train it also!)
+        model.resize_token_embeddings(len(tokenizer))
+
         choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
         input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
-        mc_token_ids = torch.tensor([input_ids.size(-1), input_ids.size(-1)]).unsqueeze(0)  # Batch size 1
+        mc_token_ids = torch.tensor([input_ids.size(-1)-1, input_ids.size(-1)-1]).unsqueeze(0)  # Batch size 1
+
         outputs = model(input_ids, mc_token_ids=mc_token_ids)
         lm_prediction_scores, mc_prediction_scores = outputs[:2]
 
@@ -583,22 +593,18 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
         self.multiple_choice_head = SequenceSummary(config)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.lm_head,
-                                   self.transformer.tokens_embed)
+    def get_output_embeddings(self):
+        return self.lm_head
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 mc_token_ids=None, lm_labels=None, mc_labels=None):
         transformer_outputs = self.transformer(input_ids,
                                                attention_mask=attention_mask,
                                                token_type_ids=token_type_ids,
                                                position_ids=position_ids,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
         hidden_states = transformer_outputs[0]
 
         lm_logits = self.lm_head(hidden_states)

transformers/modeling_roberta.py

@@ -34,6 +34,9 @@ ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'roberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-base-pytorch_model.bin",
     'roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-pytorch_model.bin",
     'roberta-large-mnli': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-mnli-pytorch_model.bin",
+    'distilroberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/distilroberta-base-pytorch_model.bin",
+    'roberta-base-openai-detector': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-base-openai-detector-pytorch_model.bin",
+    'roberta-large-openai-detector': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-openai-detector-pytorch_model.bin",
 }
 
 class RobertaEmbeddings(BertEmbeddings):
@@ -47,16 +50,24 @@ class RobertaEmbeddings(BertEmbeddings):
         self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size,
                                                 padding_idx=self.padding_idx)
 
-    def forward(self, input_ids, token_type_ids=None, position_ids=None):
-        seq_length = input_ids.size(1)
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
         if position_ids is None:
             # Position numbers begin at padding_idx+1. Padding symbols are ignored.
             # cf. fairseq's `utils.make_positions`
-            position_ids = torch.arange(self.padding_idx+1, seq_length+self.padding_idx+1, dtype=torch.long, device=input_ids.device)
-            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+            position_ids = torch.arange(self.padding_idx+1, seq_length+self.padding_idx+1, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).expand(input_shape)
         return super(RobertaEmbeddings, self).forward(input_ids,
                                                       token_type_ids=token_type_ids,
-                                                      position_ids=position_ids)
+                                                      position_ids=position_ids,
+                                                      inputs_embeds=inputs_embeds)
 
 
 ROBERTA_START_DOCSTRING = r"""    The RoBERTa model was proposed in
@@ -125,6 +136,10 @@ ROBERTA_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare RoBERTa Model transformer outputting raw hidden-states without any specific head on top.",
@@ -168,18 +183,11 @@ class RobertaModel(BertModel):
         self.embeddings = RobertaEmbeddings(config)
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
-        if input_ids[:, 0].sum().item() != 0:
-            logger.warning("A sequence with no special tokens has been passed to the RoBERTa model. "
-                           "This model requires special tokens in order to work. "
-                           "Please specify add_special_tokens=True in your tokenize.encode()"
-                           "or tokenizer.convert_tokens_to_ids().")
-        return super(RobertaModel, self).forward(input_ids,
-                                                 attention_mask=attention_mask,
-                                                 token_type_ids=token_type_ids,
-                                                 position_ids=position_ids,
-                                                 head_mask=head_mask)
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
 
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
 
 @add_start_docstrings("""RoBERTa Model with a `language modeling` head on top. """,
     ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
@@ -224,21 +232,18 @@ class RobertaForMaskedLM(BertPreTrainedModel):
         self.lm_head = RobertaLMHead(config)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the input and output embeddings.
-            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
-        """
-        self._tie_or_clone_weights(self.lm_head.decoder, self.roberta.embeddings.word_embeddings)
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
 
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 masked_lm_labels=None):
         outputs = self.roberta(input_ids,
                                attention_mask=attention_mask,
                                token_type_ids=token_type_ids,
                                position_ids=position_ids,
-                               head_mask=head_mask)
+                               head_mask=head_mask,
+                               inputs_embeds=inputs_embeds)
         sequence_output = outputs[0]
         prediction_scores = self.lm_head(sequence_output)
 
@@ -319,13 +324,14 @@ class RobertaForSequenceClassification(BertPreTrainedModel):
         self.roberta = RobertaModel(config)
         self.classifier = RobertaClassificationHead(config)
     
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
                 labels=None):
         outputs = self.roberta(input_ids,
                                attention_mask=attention_mask,
                                token_type_ids=token_type_ids,
                                position_ids=position_ids,
-                               head_mask=head_mask)
+                               head_mask=head_mask,
+                               inputs_embeds=inputs_embeds)
         sequence_output = outputs[0]
         logits = self.classifier(sequence_output)
 
@@ -342,6 +348,7 @@ class RobertaForSequenceClassification(BertPreTrainedModel):
 
         return outputs  # (loss), logits, (hidden_states), (attentions)
 
+
 @add_start_docstrings("""Roberta Model with a multiple choice classification head on top (a linear layer on top of
     the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
     ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
@@ -381,6 +388,10 @@ class RobertaForMultipleChoice(BertPreTrainedModel):
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
@@ -424,8 +435,8 @@ class RobertaForMultipleChoice(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None,
-                position_ids=None, head_mask=None):
+    def forward(self, input_ids=None, token_type_ids=None, attention_mask=None, labels=None,
+                position_ids=None, head_mask=None, inputs_embeds=None):
         num_choices = input_ids.shape[1]
 
         flat_input_ids = input_ids.view(-1, input_ids.size(-1))
@@ -450,6 +461,82 @@ class RobertaForMultipleChoice(BertPreTrainedModel):
         return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
 
 
+@add_start_docstrings("""Roberta Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
+class RobertaForTokenClassification(BertPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the token classification loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss.
+        **scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
+        model = RobertaForTokenClassification.from_pretrained('roberta-base')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, scores = outputs[:2]
+
+    """
+    config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "roberta"
+
+    def __init__(self, config):
+        super(RobertaForTokenClassification, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.roberta = RobertaModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
+
+        outputs = self.roberta(input_ids,
+                               attention_mask=attention_mask,
+                               token_type_ids=token_type_ids,
+                               position_ids=position_ids,
+                               head_mask=head_mask,
+                               inputs_embeds=inputs_embeds)
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)[active_loss]
+                active_labels = labels.view(-1)[active_loss]
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), scores, (hidden_states), (attentions)
+
 
 class RobertaClassificationHead(nn.Module):
     """Head for sentence-level classification tasks."""

transformers/modeling_tf_albert.py

@@ -0,0 +1,799 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" TF 2.0 ALBERT model. """
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import math
+import os
+import sys
+from io import open
+
+import numpy as np
+import tensorflow as tf
+
+from .configuration_albert import AlbertConfig
+from .modeling_tf_utils import TFPreTrainedModel, get_initializer
+from .modeling_tf_bert import ACT2FN, TFBertSelfAttention
+from .file_utils import add_start_docstrings
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'albert-base-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-tf_model.h5",
+    'albert-large-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-tf_model.h5",
+    'albert-xlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-tf_model.h5",
+    'albert-xxlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-tf_model.h5",
+    'albert-base-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v2-tf_model.h5",
+    'albert-large-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-v2-tf_model.h5",
+    'albert-xlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-v2-tf_model.h5",
+    'albert-xxlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-v2-tf_model.h5",
+}
+
+
+class TFAlbertEmbeddings(tf.keras.layers.Layer):
+    """Construct the embeddings from word, position and token_type embeddings.
+    """
+
+    def __init__(self, config, **kwargs):
+        super(TFAlbertEmbeddings, self).__init__(**kwargs)
+
+        self.config = config
+        self.position_embeddings = tf.keras.layers.Embedding(config.max_position_embeddings,
+                                                             config.embedding_size,
+                                                             embeddings_initializer=get_initializer(
+                                                                 self.config.initializer_range),
+                                                             name='position_embeddings')
+        self.token_type_embeddings = tf.keras.layers.Embedding(config.type_vocab_size,
+                                                               config.embedding_size,
+                                                               embeddings_initializer=get_initializer(
+                                                                   self.config.initializer_range),
+                                                               name='token_type_embeddings')
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+
+    def build(self, input_shape):
+        """Build shared word embedding layer """
+        with tf.name_scope("word_embeddings"):
+            # Create and initialize weights. The random normal initializer was chosen
+            # arbitrarily, and works well.
+            self.word_embeddings = self.add_weight(
+                "weight",
+                shape=[self.config.vocab_size, self.config.embedding_size],
+                initializer=get_initializer(self.config.initializer_range))
+        super(TFAlbertEmbeddings, self).build(input_shape)
+
+    def call(self, inputs, mode="embedding", training=False):
+        """Get token embeddings of inputs.
+        Args:
+            inputs: list of three int64 tensors with shape [batch_size, length]: (input_ids, position_ids, token_type_ids)
+            mode: string, a valid value is one of "embedding" and "linear".
+        Returns:
+            outputs: (1) If mode == "embedding", output embedding tensor, float32 with
+                shape [batch_size, length, embedding_size]; (2) mode == "linear", output
+                linear tensor, float32 with shape [batch_size, length, vocab_size].
+        Raises:
+            ValueError: if mode is not valid.
+
+        Shared weights logic adapted from
+            https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24
+        """
+        if mode == "embedding":
+            return self._embedding(inputs, training=training)
+        elif mode == "linear":
+            return self._linear(inputs)
+        else:
+            raise ValueError("mode {} is not valid.".format(mode))
+
+    def _embedding(self, inputs, training=False):
+        """Applies embedding based on inputs tensor."""
+        input_ids, position_ids, token_type_ids, inputs_embeds = inputs
+
+        if input_ids is not None:
+            input_shape = tf.shape(input_ids)
+        else:
+            input_shape = tf.shape(inputs_embeds)[:-1]
+
+        seq_length = input_shape[1]
+        if position_ids is None:
+            position_ids = tf.range(seq_length, dtype=tf.int32)[tf.newaxis, :]
+        if token_type_ids is None:
+            token_type_ids = tf.fill(input_shape, 0)
+
+        if inputs_embeds is None:
+            inputs_embeds = tf.gather(self.word_embeddings, input_ids)
+        position_embeddings = self.position_embeddings(position_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + position_embeddings + token_type_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings, training=training)
+        return embeddings
+
+    def _linear(self, inputs):
+        """Computes logits by running inputs through a linear layer.
+            Args:
+                inputs: A float32 tensor with shape [batch_size, length, embedding_size]
+            Returns:
+                float32 tensor with shape [batch_size, length, vocab_size].
+        """
+        batch_size = tf.shape(inputs)[0]
+        length = tf.shape(inputs)[1]
+        x = tf.reshape(inputs, [-1, self.config.embedding_size])
+        logits = tf.matmul(x, self.word_embeddings, transpose_b=True)
+        return tf.reshape(logits, [batch_size, length, self.config.vocab_size])
+
+
+class TFAlbertSelfAttention(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertSelfAttention, self).__init__(**kwargs)
+        if config.hidden_size % config.num_attention_heads != 0:
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads))
+        self.output_attentions = config.output_attentions
+
+        self.num_attention_heads = config.num_attention_heads
+        assert config.hidden_size % config.num_attention_heads == 0
+        self.attention_head_size = int(
+            config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = tf.keras.layers.Dense(self.all_head_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='query')
+        self.key = tf.keras.layers.Dense(self.all_head_size,
+                                         kernel_initializer=get_initializer(
+                                             config.initializer_range),
+                                         name='key')
+        self.value = tf.keras.layers.Dense(self.all_head_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='value')
+
+        self.dropout = tf.keras.layers.Dropout(
+            config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x, batch_size):
+        x = tf.reshape(
+            x, (batch_size, -1, self.num_attention_heads, self.attention_head_size))
+        return tf.transpose(x, perm=[0, 2, 1, 3])
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        batch_size = tf.shape(hidden_states)[0]
+        mixed_query_layer = self.query(hidden_states)
+        mixed_key_layer = self.key(hidden_states)
+        mixed_value_layer = self.value(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
+        key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
+        value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        # (batch size, num_heads, seq_len_q, seq_len_k)
+        attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
+        # scale attention_scores
+        dk = tf.cast(tf.shape(key_layer)[-1], tf.float32)
+        attention_scores = attention_scores / tf.math.sqrt(dk)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in TFAlbertModel call() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = tf.nn.softmax(attention_scores, axis=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs, training=training)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = tf.matmul(attention_probs, value_layer)
+
+        context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
+        context_layer = tf.reshape(context_layer,
+                                   (batch_size, -1, self.all_head_size))  # (batch_size, seq_len_q, all_head_size)
+
+        outputs = (context_layer, attention_probs) if self.output_attentions else (
+            context_layer,)
+        return outputs
+
+
+class TFAlbertSelfOutput(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertSelfOutput, self).__init__(**kwargs)
+        self.dense = tf.keras.layers.Dense(config.hidden_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='dense')
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+
+    def call(self, inputs, training=False):
+        hidden_states, input_tensor = inputs
+
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class TFAlbertAttention(TFBertSelfAttention):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertAttention, self).__init__(config, **kwargs)
+
+        self.hidden_size = config.hidden_size
+        self.dense = tf.keras.layers.Dense(config.hidden_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='dense')
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        raise NotImplementedError
+
+    def call(self, inputs, training=False):
+        input_tensor, attention_mask, head_mask = inputs
+
+        batch_size = tf.shape(input_tensor)[0]
+        mixed_query_layer = self.query(input_tensor)
+        mixed_key_layer = self.key(input_tensor)
+        mixed_value_layer = self.value(input_tensor)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
+        key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
+        value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        # (batch size, num_heads, seq_len_q, seq_len_k)
+        attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
+        # scale attention_scores
+        dk = tf.cast(tf.shape(key_layer)[-1], tf.float32)
+        attention_scores = attention_scores / tf.math.sqrt(dk)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in TFBertModel call() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = tf.nn.softmax(attention_scores, axis=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs, training=training)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = tf.matmul(attention_probs, value_layer)
+
+        context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
+        context_layer = tf.reshape(context_layer,
+                                   (batch_size, -1, self.all_head_size))  # (batch_size, seq_len_q, all_head_size)
+
+        self_outputs = (context_layer, attention_probs) if self.output_attentions else (
+            context_layer,)
+
+        hidden_states = self_outputs[0]
+
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        attention_output = self.LayerNorm(hidden_states + input_tensor)
+
+        # add attentions if we output them
+        outputs = (attention_output,) + self_outputs[1:]
+        return outputs
+
+
+class TFAlbertLayer(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertLayer, self).__init__(**kwargs)
+        self.attention = TFAlbertAttention(config, name='attention')
+
+        self.ffn = tf.keras.layers.Dense(config.intermediate_size, kernel_initializer=get_initializer(
+            config.initializer_range), name='ffn')
+
+        if isinstance(config.hidden_act, str) or (sys.version_info[0] == 2 and isinstance(config.hidden_act, unicode)):
+            self.activation = ACT2FN[config.hidden_act]
+        else:
+            self.activation = config.hidden_act
+
+        self.ffn_output = tf.keras.layers.Dense(config.hidden_size, kernel_initializer=get_initializer(
+            config.initializer_range), name='ffn_output')
+        self.full_layer_layer_norm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='full_layer_layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        attention_outputs = self.attention(
+            [hidden_states, attention_mask, head_mask], training=training)
+        ffn_output = self.ffn(attention_outputs[0])
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(ffn_output)
+
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = self.full_layer_layer_norm(
+            ffn_output + attention_outputs[0])
+
+        # add attentions if we output them
+        outputs = (hidden_states,) + attention_outputs[1:]
+        return outputs
+
+
+class TFAlbertLayerGroup(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertLayerGroup, self).__init__(**kwargs)
+
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.albert_layers = [TFAlbertLayer(config, name="albert_layers_._{}".format(
+            i)) for i in range(config.inner_group_num)]
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        layer_hidden_states = ()
+        layer_attentions = ()
+
+        for layer_index, albert_layer in enumerate(self.albert_layers):
+            layer_output = albert_layer(
+                [hidden_states, attention_mask, head_mask[layer_index]], training=training)
+            hidden_states = layer_output[0]
+
+            if self.output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+            if self.output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (layer_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (layer_attentions,)
+        # last-layer hidden state, (layer hidden states), (layer attentions)
+        return outputs
+
+
+class TFAlbertTransformer(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertTransformer, self).__init__(**kwargs)
+
+        self.config = config
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.embedding_hidden_mapping_in = tf.keras.layers.Dense(config.hidden_size, kernel_initializer=get_initializer(
+            config.initializer_range), name='embedding_hidden_mapping_in')
+        self.albert_layer_groups = [TFAlbertLayerGroup(
+            config, name="albert_layer_groups_._{}".format(i)) for i in range(config.num_hidden_groups)]
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        hidden_states = self.embedding_hidden_mapping_in(hidden_states)
+        all_attentions = ()
+
+        if self.output_hidden_states:
+            all_hidden_states = (hidden_states,)
+
+        for i in range(self.config.num_hidden_layers):
+            # Number of layers in a hidden group
+            layers_per_group = int(
+                self.config.num_hidden_layers / self.config.num_hidden_groups)
+
+            # Index of the hidden group
+            group_idx = int(
+                i / (self.config.num_hidden_layers / self.config.num_hidden_groups))
+
+            layer_group_output = self.albert_layer_groups[group_idx](
+                [hidden_states, attention_mask, head_mask[group_idx*layers_per_group:(group_idx+1)*layers_per_group]], training=training)
+            hidden_states = layer_group_output[0]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + layer_group_output[-1]
+
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+
+        # last-layer hidden state, (all hidden states), (all attentions)
+        return outputs
+
+
+class TFAlbertPreTrainedModel(TFPreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = AlbertConfig
+    pretrained_model_archive_map = TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "albert"
+
+
+class TFAlbertMLMHead(tf.keras.layers.Layer):
+    def __init__(self, config, input_embeddings, **kwargs):
+        super(TFAlbertMLMHead, self).__init__(**kwargs)
+        self.vocab_size = config.vocab_size
+
+        self.dense = tf.keras.layers.Dense(config.embedding_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='dense')
+        if isinstance(config.hidden_act, str) or (sys.version_info[0] == 2 and isinstance(config.hidden_act, unicode)):
+            self.activation = ACT2FN[config.hidden_act]
+        else:
+            self.activation = config.hidden_act
+
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = input_embeddings
+
+    def build(self, input_shape):
+        self.bias = self.add_weight(shape=(self.vocab_size,),
+                                    initializer='zeros',
+                                    trainable=True,
+                                    name='bias')
+        self.decoder_bias = self.add_weight(shape=(self.vocab_size,),
+                                    initializer='zeros',
+                                    trainable=True,
+                                    name='decoder/bias')
+        super(TFAlbertMLMHead, self).build(input_shape)
+
+    def call(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        hidden_states = self.decoder(hidden_states, mode="linear") + self.decoder_bias
+        hidden_states = hidden_states + self.bias
+        return hidden_states
+
+
+ALBERT_START_DOCSTRING = r"""    The ALBERT model was proposed in
+    `ALBERT: A Lite BERT for Self-supervised Learning of Language Representations`_
+    by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut. It presents
+    two parameter-reduction techniques to lower memory consumption and increase the trainig speed of BERT.
+
+    This model is a tf.keras.Model `tf.keras.Model`_ sub-class. Use it as a regular TF 2.0 Keras Model and
+    refer to the TF 2.0 documentation for all matter related to general usage and behavior.
+
+    .. _`ALBERT: A Lite BERT for Self-supervised Learning of Language Representations`:
+        https://arxiv.org/abs/1909.11942
+
+    .. _`tf.keras.Model`:
+        https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/Model
+
+    Note on the model inputs:
+        TF 2.0 models accepts two formats as inputs:
+
+            - having all inputs as keyword arguments (like PyTorch models), or
+            - having all inputs as a list, tuple or dict in the first positional arguments.
+
+        This second option is usefull when using `tf.keras.Model.fit()` method which currently requires having all the tensors in the first argument of the model call function: `model(inputs)`.
+
+        If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument :
+
+        - a single Tensor with input_ids only and nothing else: `model(inputs_ids)
+        - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
+            `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
+        - a dictionary with one or several input Tensors associaed to the input names given in the docstring:
+            `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
+
+    Parameters:
+        config (:class:`~transformers.AlbertConfig`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, ALBERT input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+                
+                ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+                
+                ``token_type_ids:   0   0   0   0  0     0   0``
+
+            Albert is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
+            Indices can be obtained using :class:`transformers.AlbertTokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **token_type_ids**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Segment token indices to indicate first and second portions of the inputs.
+            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
+            corresponds to a `sentence B` token
+            (see `ALBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
+        **position_ids**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
+        **head_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+"""
+
+@add_start_docstrings("The bare Albert Model transformer outputing raw hidden-states without any specific head on top.",
+                      ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class TFAlbertModel(TFAlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``tf.Tensor`` of shape ``(batch_size, hidden_size)``
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during Albert pretraining. This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import AlbertTokenizer, TFAlbertModel
+
+        tokenizer = AlbertTokenizer.from_pretrained('bert-base-uncased')
+        model = TFAlbertModel.from_pretrained('bert-base-uncased')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+
+    def __init__(self, config, **kwargs):
+        super(TFAlbertModel, self).__init__(config, **kwargs)
+        self.num_hidden_layers = config.num_hidden_layers
+
+        self.embeddings = TFAlbertEmbeddings(config, name="embeddings")
+        self.encoder = TFAlbertTransformer(config, name="encoder")
+        self.pooler = tf.keras.layers.Dense(config.hidden_size, kernel_initializer=get_initializer(
+            config.initializer_range), activation='tanh', name='pooler')
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        raise NotImplementedError
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
+        """
+        raise NotImplementedError
+
+    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, training=False):
+        if isinstance(inputs, (tuple, list)):
+            input_ids = inputs[0]
+            attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
+            token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
+            position_ids = inputs[3] if len(inputs) > 3 else position_ids
+            head_mask = inputs[4] if len(inputs) > 4 else head_mask
+            inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
+            assert len(inputs) <= 6, "Too many inputs."
+        elif isinstance(inputs, dict):
+            input_ids = inputs.get('input_ids')
+            attention_mask = inputs.get('attention_mask', attention_mask)
+            token_type_ids = inputs.get('token_type_ids', token_type_ids)
+            position_ids = inputs.get('position_ids', position_ids)
+            head_mask = inputs.get('head_mask', head_mask)
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 6, "Too many inputs."
+        else:
+            input_ids = inputs
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.shape
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.shape[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if attention_mask is None:
+            attention_mask = tf.fill(input_shape, 1)
+        if token_type_ids is None:
+            token_type_ids = tf.fill(input_shape, 0)
+
+        # We create a 3D attention mask from a 2D tensor mask.
+        # Sizes are [batch_size, 1, 1, to_seq_length]
+        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+        # this attention mask is more simple than the triangular masking of causal attention
+        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+        extended_attention_mask = attention_mask[:, tf.newaxis, tf.newaxis, :]
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+
+        extended_attention_mask = tf.cast(extended_attention_mask, tf.float32)
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if not head_mask is None:
+            raise NotImplementedError
+        else:
+            head_mask = [None] * self.num_hidden_layers
+            # head_mask = tf.constant([0] * self.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            [input_ids, position_ids, token_type_ids, inputs_embeds], training=training)
+        encoder_outputs = self.encoder(
+            [embedding_output, extended_attention_mask, head_mask], training=training)
+
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output[:, 0])
+
+        # add hidden_states and attentions if they are here
+        outputs = (sequence_output, pooled_output,) + encoder_outputs[1:]
+        # sequence_output, pooled_output, (hidden_states), (attentions)
+        return outputs
+
+
+@add_start_docstrings("""Albert Model with a `language modeling` head on top. """,
+                      ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class TFAlbertForMaskedLM(TFAlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **prediction_scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import AlbertTokenizer, TFAlbertForMaskedLM
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = TFAlbertForMaskedLM.from_pretrained('albert-base-v2')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        prediction_scores = outputs[0]
+
+    """
+
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFAlbertForMaskedLM, self).__init__(config, *inputs, **kwargs)
+
+        self.albert = TFAlbertModel(config, name='albert')
+        self.predictions = TFAlbertMLMHead(
+            config, self.albert.embeddings, name='predictions')
+
+    def get_output_embeddings(self):
+        return self.albert.embeddings
+
+    def call(self, inputs, **kwargs):
+        outputs = self.albert(inputs, **kwargs)
+
+        sequence_output = outputs[0]
+        prediction_scores = self.predictions(
+            sequence_output, training=kwargs.get('training', False))
+
+        # Add hidden states and attention if they are here
+        outputs = (prediction_scores,) + outputs[2:]
+
+        return outputs  # prediction_scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **logits**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import AlbertTokenizer, TFAlbertForSequenceClassification
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = TFAlbertForSequenceClassification.from_pretrained('albert-base-v2')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        logits = outputs[0]
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFAlbertForSequenceClassification, self).__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertModel(config, name='albert')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(config.num_labels,
+                                                kernel_initializer=get_initializer(config.initializer_range),
+                                                name='classifier')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.albert(inputs, **kwargs)
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output, training=kwargs.get('training', False))
+        logits = self.classifier(pooled_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        return outputs  # logits, (hidden_states), (attentions)

transformers/modeling_tf_auto.py

@@ -26,6 +26,7 @@ from .modeling_tf_xlnet import TFXLNetModel, TFXLNetLMHeadModel, TFXLNetForSeque
 from .modeling_tf_xlm import TFXLMModel, TFXLMWithLMHeadModel, TFXLMForSequenceClassification, TFXLMForQuestionAnsweringSimple
 from .modeling_tf_roberta import TFRobertaModel, TFRobertaForMaskedLM, TFRobertaForSequenceClassification
 from .modeling_tf_distilbert import TFDistilBertModel, TFDistilBertForQuestionAnswering, TFDistilBertForMaskedLM, TFDistilBertForSequenceClassification
+from .modeling_tf_ctrl import TFCTRLModel, TFCTRLLMHeadModel
 
 from .file_utils import add_start_docstrings
 
@@ -52,6 +53,7 @@ class TFAutoModel(object):
             - contains `transfo-xl`: TFTransfoXLModel (Transformer-XL model)
             - contains `xlnet`: TFXLNetModel (XLNet model)
             - contains `xlm`: TFXLMModel (XLM model)
+            - contains `ctrl`: TFCTRLModel (CTRL model)
 
         This class cannot be instantiated using `__init__()` (throws an error).
     """
@@ -73,7 +75,7 @@ class TFAutoModel(object):
             - contains `gpt2`: TFGPT2Model (OpenAI GPT-2 model)
             - contains `transfo-xl`: TFTransfoXLModel (Transformer-XL model)
             - contains `xlnet`: TFXLNetModel (XLNet model)
-            - contains `xlm`: TFXLMModel (XLM model)
+            - contains `ctrl`: TFCTRLModel (CTRL model)
 
         Params:
             pretrained_model_name_or_path: either:
@@ -147,10 +149,12 @@ class TFAutoModel(object):
             return TFXLNetModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'xlm' in pretrained_model_name_or_path:
             return TFXLMModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'ctrl' in pretrained_model_name_or_path:
+            return TFCTRLModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                          "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
-                         "'xlm', 'roberta'".format(pretrained_model_name_or_path))
+                         "'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))
 
 
 class TFAutoModelWithLMHead(object):
@@ -173,6 +177,7 @@ class TFAutoModelWithLMHead(object):
             - contains `transfo-xl`: TFTransfoXLLMHeadModel (Transformer-XL model)
             - contains `xlnet`: TFXLNetLMHeadModel (XLNet model)
             - contains `xlm`: TFXLMWithLMHeadModel (XLM model)
+            - contains `ctrl`: TFCTRLLMHeadModel (CTRL model)
 
         This class cannot be instantiated using `__init__()` (throws an error).
     """
@@ -198,6 +203,7 @@ class TFAutoModelWithLMHead(object):
             - contains `transfo-xl`: TFTransfoXLLMHeadModel (Transformer-XL model)
             - contains `xlnet`: TFXLNetLMHeadModel (XLNet model)
             - contains `xlm`: TFXLMWithLMHeadModel (XLM model)
+            - contains `ctrl`: TFCTRLLMHeadModel (CTRL model)
 
         Params:
             pretrained_model_name_or_path: either:
@@ -271,10 +277,12 @@ class TFAutoModelWithLMHead(object):
             return TFXLNetLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'xlm' in pretrained_model_name_or_path:
             return TFXLMWithLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'ctrl' in pretrained_model_name_or_path:
+            return TFCTRLLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                          "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
-                         "'xlm', 'roberta'".format(pretrained_model_name_or_path))
+                         "'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))
 
 
 class TFAutoModelForSequenceClassification(object):

transformers/modeling_tf_bert.py

@@ -30,7 +30,6 @@ import tensorflow as tf
 from .configuration_bert import BertConfig
 from .modeling_tf_utils import TFPreTrainedModel, get_initializer
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
@@ -52,14 +51,6 @@ TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
 }
 
 
-def load_bert_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def gelu(x):
     """ Gaussian Error Linear Unit.
     Original Implementation of the gelu activation function in Google Bert repo when initially created.
@@ -151,19 +142,25 @@ class TFBertEmbeddings(tf.keras.layers.Layer):
 
     def _embedding(self, inputs, training=False):
         """Applies embedding based on inputs tensor."""
-        input_ids, position_ids, token_type_ids = inputs
+        input_ids, position_ids, token_type_ids, inputs_embeds = inputs
 
-        seq_length = tf.shape(input_ids)[1]
+        if input_ids is not None:
+            input_shape = tf.shape(input_ids)
+        else:
+            input_shape = tf.shape(inputs_embeds)[:-1]
+        
+        seq_length = input_shape[1]
         if position_ids is None:
             position_ids = tf.range(seq_length, dtype=tf.int32)[tf.newaxis, :]
         if token_type_ids is None:
-            token_type_ids = tf.fill(tf.shape(input_ids), 0)
+            token_type_ids = tf.fill(input_shape, 0)
 
-        words_embeddings = tf.gather(self.word_embeddings, input_ids)
+        if inputs_embeds is None:
+            inputs_embeds = tf.gather(self.word_embeddings, input_ids)
         position_embeddings = self.position_embeddings(position_ids)
         token_type_embeddings = self.token_type_embeddings(token_type_ids)
 
-        embeddings = words_embeddings + position_embeddings + token_type_embeddings
+        embeddings = inputs_embeds + position_embeddings + token_type_embeddings
         embeddings = self.LayerNorm(embeddings)
         embeddings = self.dropout(embeddings, training=training)
         return embeddings
@@ -469,6 +466,9 @@ class TFBertMainLayer(tf.keras.layers.Layer):
         self.encoder = TFBertEncoder(config, name='encoder')
         self.pooler = TFBertPooler(config, name='pooler')
 
+    def get_input_embeddings(self):
+        return self.embeddings
+
     def _resize_token_embeddings(self, new_num_tokens):
         raise NotImplementedError
 
@@ -479,28 +479,39 @@ class TFBertMainLayer(tf.keras.layers.Layer):
         """
         raise NotImplementedError
 
-    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, training=False):
+    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
             token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
             position_ids = inputs[3] if len(inputs) > 3 else position_ids
             head_mask = inputs[4] if len(inputs) > 4 else head_mask
-            assert len(inputs) <= 5, "Too many inputs."
+            inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
+            assert len(inputs) <= 6, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             attention_mask = inputs.get('attention_mask', attention_mask)
             token_type_ids = inputs.get('token_type_ids', token_type_ids)
             position_ids = inputs.get('position_ids', position_ids)
             head_mask = inputs.get('head_mask', head_mask)
-            assert len(inputs) <= 5, "Too many inputs."
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 6, "Too many inputs."
         else:
             input_ids = inputs
 
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.shape
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.shape[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         if attention_mask is None:
-            attention_mask = tf.fill(tf.shape(input_ids), 1)
+            attention_mask = tf.fill(input_shape, 1)
         if token_type_ids is None:
-            token_type_ids = tf.fill(tf.shape(input_ids), 0)
+            token_type_ids = tf.fill(input_shape, 0)
 
         # We create a 3D attention mask from a 2D tensor mask.
         # Sizes are [batch_size, 1, 1, to_seq_length]
@@ -529,7 +540,7 @@ class TFBertMainLayer(tf.keras.layers.Layer):
             head_mask = [None] * self.num_hidden_layers
             # head_mask = tf.constant([0] * self.num_hidden_layers)
 
-        embedding_output = self.embeddings([input_ids, position_ids, token_type_ids], training=training)
+        embedding_output = self.embeddings([input_ids, position_ids, token_type_ids, inputs_embeds], training=training)
         encoder_outputs = self.encoder([embedding_output, extended_attention_mask, head_mask], training=training)
 
         sequence_output = encoder_outputs[0]
@@ -545,7 +556,6 @@ class TFBertPreTrainedModel(TFPreTrainedModel):
     """
     config_class = BertConfig
     pretrained_model_archive_map = TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_bert_pt_weights_in_tf2
     base_model_prefix = "bert"
 
 
@@ -626,6 +636,10 @@ BERT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare Bert Model transformer outputing raw hidden-states without any specific head on top.",
@@ -708,6 +722,9 @@ class TFBertForPreTraining(TFBertPreTrainedModel):
         self.nsp = TFBertNSPHead(config, name='nsp___cls')
         self.mlm = TFBertMLMHead(config, self.bert.embeddings, name='mlm___cls')
 
+    def get_output_embeddings(self):
+        return self.bert.embeddings
+
     def call(self, inputs, **kwargs):
         outputs = self.bert(inputs, **kwargs)
 
@@ -753,6 +770,9 @@ class TFBertForMaskedLM(TFBertPreTrainedModel):
         self.bert = TFBertMainLayer(config, name='bert')
         self.mlm = TFBertMLMHead(config, self.bert.embeddings, name='mlm___cls')
 
+    def get_output_embeddings(self):
+        return self.bert.embeddings
+
     def call(self, inputs, **kwargs):
         outputs = self.bert(inputs, **kwargs)
 
@@ -898,33 +918,39 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel):
                                                 kernel_initializer=get_initializer(config.initializer_range),
                                                 name='classifier')
 
-    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, training=False):
+    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
             token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
             position_ids = inputs[3] if len(inputs) > 3 else position_ids
             head_mask = inputs[4] if len(inputs) > 4 else head_mask
-            assert len(inputs) <= 5, "Too many inputs."
+            inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
+            assert len(inputs) <= 6, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             attention_mask = inputs.get('attention_mask', attention_mask)
             token_type_ids = inputs.get('token_type_ids', token_type_ids)
             position_ids = inputs.get('position_ids', position_ids)
             head_mask = inputs.get('head_mask', head_mask)
-            assert len(inputs) <= 5, "Too many inputs."
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 6, "Too many inputs."
         else:
             input_ids = inputs
 
-        num_choices = tf.shape(input_ids)[1]
-        seq_length = tf.shape(input_ids)[2]
+        if input_ids is not None:
+            num_choices = tf.shape(input_ids)[1]
+            seq_length = tf.shape(input_ids)[2]
+        else:
+            num_choices = tf.shape(inputs_embeds)[1]
+            seq_length = tf.shape(inputs_embeds)[2]
 
-        flat_input_ids = tf.reshape(input_ids, (-1, seq_length))
+        flat_input_ids = tf.reshape(input_ids, (-1, seq_length)) if input_ids is not None else None
         flat_attention_mask = tf.reshape(attention_mask, (-1, seq_length)) if attention_mask is not None else None
         flat_token_type_ids = tf.reshape(token_type_ids, (-1, seq_length)) if token_type_ids is not None else None
         flat_position_ids = tf.reshape(position_ids, (-1, seq_length)) if position_ids is not None else None
 
-        flat_inputs = [flat_input_ids, flat_attention_mask, flat_token_type_ids, flat_position_ids, head_mask]
+        flat_inputs = [flat_input_ids, flat_attention_mask, flat_token_type_ids, flat_position_ids, head_mask, inputs_embeds]
 
         outputs = self.bert(flat_inputs, training=training)
 

transformers/modeling_tf_ctrl.py

@@ -27,20 +27,11 @@ import tensorflow as tf
 from .configuration_ctrl import CTRLConfig
 from .modeling_tf_utils import TFPreTrainedModel, get_initializer, shape_list, TFSharedEmbeddings
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
 TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP = {"ctrl": "https://s3.amazonaws.com/models.huggingface.co/bert/ctrl-tf_model.h5"}
 
-def load_ctrl_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def angle_defn(pos, i, d_model_size):
     angle_rates = 1 / np.power(10000, (2 * (i//2)) / np.float32(d_model_size))
     return pos * angle_rates
@@ -177,12 +168,14 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
     def __init__(self, config, **kwargs):
         super(TFCTRLMainLayer, self).__init__(**kwargs)
         self.output_hidden_states = config.output_hidden_states
+        self.output_attentions = config.output_attentions
+        self.output_past = config.output_past
+
         self.d_model_size = config.n_embd
         self.num_layers = config.n_layer
 
         self.pos_encoding = positional_encoding(config.n_positions, self.d_model_size)
 
-        self.output_attentions = config.output_attentions
 
         self.w = TFSharedEmbeddings(config.vocab_size,
                                     config.n_embd,
@@ -199,6 +192,9 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
                                  name='h_._{}'.format(i)) for i in range(config.n_layer)]
         self.layernorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_epsilon, name="layernorm")
 
+    def get_input_embeddings(self):
+        return self.w
+
     def _resize_token_embeddings(self, new_num_tokens):
         raise NotImplementedError
 
@@ -208,7 +204,7 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
         """
         raise NotImplementedError
 
-    def call(self, inputs, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, training=False):
+    def call(self, inputs, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             past = inputs[1] if len(inputs) > 1 else past
@@ -216,7 +212,8 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
             token_type_ids = inputs[3] if len(inputs) > 3 else token_type_ids
             position_ids = inputs[4] if len(inputs) > 4 else position_ids
             head_mask = inputs[5] if len(inputs) > 5 else head_mask
-            assert len(inputs) <= 6, "Too many inputs."
+            inputs_embeds = inputs[6] if len(inputs) > 6 else inputs_embeds
+            assert len(inputs) <= 7, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             past = inputs.get('past', past)
@@ -224,12 +221,20 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
             token_type_ids = inputs.get('token_type_ids', token_type_ids)
             position_ids = inputs.get('position_ids', position_ids)
             head_mask = inputs.get('head_mask', head_mask)
-            assert len(inputs) <= 6, "Too many inputs."
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 7, "Too many inputs."
         else:
             input_ids = inputs
 
-        input_shape = shape_list(input_ids)
-        input_ids = tf.reshape(input_ids, [-1, input_shape[-1]])
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = shape_list(input_ids)
+            input_ids = tf.reshape(input_ids, [-1, input_shape[-1]])
+        elif inputs_embeds is not None:
+            input_shape = shape_list(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
 
         if past is None:
             past_length = 0
@@ -237,8 +242,8 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
         else:
             past_length = shape_list(past[0][0])[-2]
         if position_ids is None:
-            position_ids = tf.range(past_length, shape_list(input_ids)[-1] + past_length, dtype=tf.int32)[tf.newaxis, :]
-            position_ids = tf.tile(position_ids, [shape_list(input_ids)[0], 1])
+            position_ids = tf.range(past_length, input_shape[-1] + past_length, dtype=tf.int32)[tf.newaxis, :]
+            position_ids = tf.tile(position_ids, [input_shape[0], 1])
 
         # Attention mask.
         if attention_mask is not None:
@@ -277,8 +282,8 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
             token_type_embeds = 0
         position_ids = tf.reshape(position_ids, [-1, shape_list(position_ids)[-1]])
 
-        inputs_embeds = self.w(input_ids, mode='embedding')
-        # x = embedded.unsqueeze(0) if len(input_ids.shape)<2 else embedded
+        if inputs_embeds is None:
+            inputs_embeds = self.w(input_ids, mode='embedding')
         seq_len = input_shape[-1]
         mask = 1 - tf.linalg.band_part(tf.ones((seq_len, seq_len)), -1, 0)
 
@@ -299,7 +304,9 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
                 all_hidden_states = all_hidden_states + (tf.reshape(hidden_states, output_shape),)
             outputs = h([hidden_states, mask, layer_past, attention_mask, head_mask[i]], training=training)
             hidden_states, present = outputs[:2]
-            presents = presents + (present,)
+
+            if self.output_past:
+                presents = presents + (present,)
 
             if self.output_attentions:
                 all_attentions.append(outputs[2])
@@ -309,7 +316,9 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
         if self.output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states, presents)
+        outputs = (hidden_states,)
+        if self.output_past:
+            outputs = outputs + (presents,)
         if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if self.output_attentions:
@@ -327,7 +336,6 @@ class TFCTRLPreTrainedModel(TFPreTrainedModel):
     config_class = CTRLConfig
     pretrained_model_archive_map = TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "transformer"
-    load_pt_weights = load_ctrl_pt_weights_in_tf2
 
 
 CTRL_START_DOCSTRING = r"""    CTRL model was proposed in 
@@ -378,6 +386,10 @@ CTRL_INPUTS_DOCSTRING = r"""    Inputs:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare CTRL Model transformer outputting raw hidden-states without any specific head on top.",
@@ -480,6 +492,9 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel):
 
         self.lm_head = TFCTRLLMHead(config, self.transformer.w, name="lm_head")
 
+    def get_output_embeddings(self):
+        return self.lm_head.input_embeddings
+
     def call(self, inputs, **kwargs):
         transformer_outputs = self.transformer(inputs, **kwargs)
         hidden_states = transformer_outputs[0]

transformers/modeling_tf_distilbert.py

@@ -31,7 +31,6 @@ import tensorflow as tf
 from .configuration_distilbert import DistilBertConfig
 from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, shape_list, get_initializer
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
@@ -66,14 +65,6 @@ def gelu_new(x):
         (np.sqrt(2 / np.pi) * (x + 0.044715 * tf.pow(x, 3)))))
     return x * cdf
 
-def load_distilbert_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
-    attns_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
-    tf_inputs = [inputs_list, attns_list]
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
 class TFEmbeddings(tf.keras.layers.Layer):
     def __init__(self, config, **kwargs):
         super(TFEmbeddings, self).__init__(**kwargs)
@@ -105,7 +96,7 @@ class TFEmbeddings(tf.keras.layers.Layer):
                 initializer=get_initializer(self.initializer_range))
         super(TFEmbeddings, self).build(input_shape)
 
-    def call(self, inputs, mode="embedding", training=False):
+    def call(self, inputs, inputs_embeds=None, mode="embedding", training=False):
         """Get token embeddings of inputs.
         Args:
             inputs: list of three int64 tensors with shape [batch_size, length]: (input_ids, position_ids, token_type_ids)
@@ -121,13 +112,13 @@ class TFEmbeddings(tf.keras.layers.Layer):
             https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24
         """
         if mode == "embedding":
-            return self._embedding(inputs, training=training)
+            return self._embedding(inputs, inputs_embeds=inputs_embeds, training=training)
         elif mode == "linear":
             return self._linear(inputs)
         else:
             raise ValueError("mode {} is not valid.".format(mode))
 
-    def _embedding(self, inputs, training=False):
+    def _embedding(self, inputs, inputs_embeds=None, training=False):
         """
         Parameters
         ----------
@@ -145,14 +136,19 @@ class TFEmbeddings(tf.keras.layers.Layer):
         else:
             input_ids, position_ids = inputs
 
-        seq_length = tf.shape(input_ids)[1]
+        if input_ids is not None:
+            seq_length = tf.shape(input_ids)[1]
+        else:
+            seq_length = tf.shape(inputs_embeds)[1]
+
         if position_ids is None:
             position_ids = tf.range(seq_length, dtype=tf.int32)[tf.newaxis, :]
 
-        word_embeddings = tf.gather(self.word_embeddings, input_ids)
+        if inputs_embeds is None:
+            inputs_embeds = tf.gather(self.word_embeddings, input_ids)
         position_embeddings = self.position_embeddings(position_ids)  # (bs, max_seq_length, dim)
 
-        embeddings = word_embeddings + position_embeddings            # (bs, max_seq_length, dim)
+        embeddings = inputs_embeds + position_embeddings              # (bs, max_seq_length, dim)
         embeddings = self.LayerNorm(embeddings)                       # (bs, max_seq_length, dim)
         embeddings = self.dropout(embeddings, training=training)      # (bs, max_seq_length, dim)
         return embeddings
@@ -407,28 +403,42 @@ class TFDistilBertMainLayer(tf.keras.layers.Layer):
         self.embeddings = TFEmbeddings(config, name="embeddings")   # Embeddings
         self.transformer = TFTransformer(config, name="transformer") # Encoder
 
+    def get_input_embeddings(self):
+        return self.embeddings
+
     def _resize_token_embeddings(self, new_num_tokens):
         raise NotImplementedError
 
     def _prune_heads(self, heads_to_prune):
         raise NotImplementedError
 
-    def call(self, inputs, attention_mask=None, head_mask=None, training=False):
+    def call(self, inputs, attention_mask=None, head_mask=None, inputs_embeds=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
             head_mask = inputs[2] if len(inputs) > 2 else head_mask
-            assert len(inputs) <= 3, "Too many inputs."
+            inputs_embeds = inputs[3] if len(inputs) > 3 else inputs_embeds
+            assert len(inputs) <= 4, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             attention_mask = inputs.get('attention_mask', attention_mask)
             head_mask = inputs.get('head_mask', head_mask)
-            assert len(inputs) <= 3, "Too many inputs."
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 4, "Too many inputs."
         else:
             input_ids = inputs
 
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = shape_list(input_ids)
+        elif inputs_embeds is not None:
+            input_shape = shape_list(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         if attention_mask is None:
-            attention_mask = tf.ones(shape_list(input_ids)) # (bs, seq_length)
+            attention_mask = tf.ones(input_shape) # (bs, seq_length)
         attention_mask = tf.cast(attention_mask, dtype=tf.float32)
 
         # Prepare head mask if needed
@@ -441,7 +451,7 @@ class TFDistilBertMainLayer(tf.keras.layers.Layer):
         else:
             head_mask = [None] * self.num_hidden_layers
 
-        embedding_output = self.embeddings(input_ids)   # (bs, seq_length, dim)
+        embedding_output = self.embeddings(input_ids, inputs_embeds=inputs_embeds)   # (bs, seq_length, dim)
         tfmr_output = self.transformer([embedding_output, attention_mask, head_mask], training=training)
 
         return tfmr_output # last-layer hidden-state, (all hidden_states), (all attentions)
@@ -454,7 +464,6 @@ class TFDistilBertPreTrainedModel(TFPreTrainedModel):
     """
     config_class = DistilBertConfig
     pretrained_model_archive_map = TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_distilbert_pt_weights_in_tf2
     base_model_prefix = "distilbert"
 
 
@@ -518,6 +527,10 @@ DISTILBERT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare DistilBERT encoder/transformer outputing raw hidden-states without any specific head on top.",
@@ -619,6 +632,9 @@ class TFDistilBertForMaskedLM(TFDistilBertPreTrainedModel):
         self.vocab_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-12, name="vocab_layer_norm")
         self.vocab_projector = TFDistilBertLMHead(config, self.distilbert.embeddings, name="vocab_projector")
 
+    def get_output_embeddings(self):
+        return self.vocab_projector.input_embeddings
+
     def call(self, inputs, **kwargs):
         distilbert_output = self.distilbert(inputs, **kwargs)
 

transformers/modeling_tf_gpt2.py

@@ -32,7 +32,6 @@ from .modeling_tf_utils import (TFPreTrainedModel, TFConv1D, TFSharedEmbeddings,
                                 TFSequenceSummary, shape_list, get_initializer)
 from .configuration_gpt2 import GPT2Config
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
@@ -42,14 +41,6 @@ TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models
                                      "distilgpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/distilgpt2-tf_model.h5",}
 
 
-def load_gpt2_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def gelu(x):
     """Gaussian Error Linear Unit.
     This is a smoother version of the RELU.
@@ -228,6 +219,9 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
                           name='h_._{}'.format(i)) for i in range(config.n_layer)]
         self.ln_f = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_epsilon, name='ln_f')
 
+    def get_input_embeddings(self):
+        return self.wte
+
     def _resize_token_embeddings(self, new_num_tokens):
         raise NotImplementedError
 
@@ -237,7 +231,7 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
         """
         raise NotImplementedError
 
-    def call(self, inputs, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, training=False):
+    def call(self, inputs, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             past = inputs[1] if len(inputs) > 1 else past
@@ -245,7 +239,8 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
             token_type_ids = inputs[3] if len(inputs) > 3 else token_type_ids
             position_ids = inputs[4] if len(inputs) > 4 else position_ids
             head_mask = inputs[5] if len(inputs) > 5 else head_mask
-            assert len(inputs) <= 6, "Too many inputs."
+            inputs_embeds = inputs[6] if len(inputs) > 6 else inputs_embeds
+            assert len(inputs) <= 7, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             past = inputs.get('past', past)
@@ -253,17 +248,28 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
             token_type_ids = inputs.get('token_type_ids', token_type_ids)
             position_ids = inputs.get('position_ids', position_ids)
             head_mask = inputs.get('head_mask', head_mask)
-            assert len(inputs) <= 6, "Too many inputs."
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 7, "Too many inputs."
         else:
             input_ids = inputs
 
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = shape_list(input_ids)
+            input_ids = tf.reshape(input_ids, [-1, input_shape[-1]])
+        elif inputs_embeds is not None:
+            input_shape = shape_list(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         if past is None:
             past_length = 0
             past = [None] * len(self.h)
         else:
             past_length = shape_list(past[0][0])[-2]
         if position_ids is None:
-            position_ids = tf.range(past_length, shape_list(input_ids)[-1] + past_length, dtype=tf.int32)[tf.newaxis, :]
+            position_ids = tf.range(past_length, input_shape[-1] + past_length, dtype=tf.int32)[tf.newaxis, :]
 
         if attention_mask is not None:
             # We create a 3D attention mask from a 2D tensor mask.
@@ -295,11 +301,10 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
             head_mask = [None] * self.num_hidden_layers
             # head_mask = tf.constant([0] * self.num_hidden_layers)
 
-        input_shape = shape_list(input_ids)
-        input_ids = tf.reshape(input_ids, [-1, input_shape[-1]])
         position_ids = tf.reshape(position_ids, [-1, shape_list(position_ids)[-1]])
 
-        inputs_embeds = self.wte(input_ids, mode='embedding')
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids, mode='embedding')
         position_embeds = self.wpe(position_ids)
         if token_type_ids is not None:
             token_type_ids = tf.reshape(token_type_ids, [-1, shape_list(token_type_ids)[-1]])
@@ -350,7 +355,6 @@ class TFGPT2PreTrainedModel(TFPreTrainedModel):
     """
     config_class = GPT2Config
     pretrained_model_archive_map = TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_gpt2_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
 
@@ -418,6 +422,10 @@ GPT2_INPUTS_DOCSTRING = r"""    Inputs:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare GPT2 Model transformer outputing raw hidden-states without any specific head on top.",
@@ -496,6 +504,9 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel):
         super(TFGPT2LMHeadModel, self).__init__(config, *inputs, **kwargs)
         self.transformer = TFGPT2MainLayer(config, name='transformer')
 
+    def get_output_embeddings(self):
+        return self.transformer.wte
+
     def call(self, inputs, **kwargs):
         transformer_outputs = self.transformer(inputs, **kwargs)
         hidden_states = transformer_outputs[0]
@@ -566,7 +577,10 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
         self.transformer = TFGPT2MainLayer(config, name='transformer')
         self.multiple_choice_head = TFSequenceSummary(config, initializer_range=config.initializer_range, name='multiple_choice_head')
 
-    def call(self, inputs, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, mc_token_ids=None, training=False):
+    def get_output_embeddings(self):
+        return self.transformer.wte
+
+    def call(self, inputs, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, mc_token_ids=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             past = inputs[1] if len(inputs) > 1 else past
@@ -574,8 +588,9 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
             token_type_ids = inputs[3] if len(inputs) > 3 else token_type_ids
             position_ids = inputs[4] if len(inputs) > 4 else position_ids
             head_mask = inputs[5] if len(inputs) > 5 else head_mask
-            mc_token_ids = inputs[6] if len(inputs) > 6 else mc_token_ids
-            assert len(inputs) <= 7, "Too many inputs."
+            inputs_embeds = inputs[6] if len(inputs) > 6 else inputs_embeds
+            mc_token_ids = inputs[7] if len(inputs) > 7 else mc_token_ids
+            assert len(inputs) <= 8, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             past = inputs.get('past', past)
@@ -583,21 +598,25 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
             token_type_ids = inputs.get('token_type_ids', token_type_ids)
             position_ids = inputs.get('position_ids', position_ids)
             head_mask = inputs.get('head_mask', head_mask)
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
             mc_token_ids = inputs.get('mc_token_ids', mc_token_ids)
-            assert len(inputs) <= 7, "Too many inputs."
+            assert len(inputs) <= 8, "Too many inputs."
         else:
             input_ids = inputs
 
-        input_shapes = shape_list(input_ids)
+        if input_ids is not None:
+            input_shapes = shape_list(input_ids)
+        else:
+            input_shapes = shape_list(inputs_embeds)[:-1]
 
         seq_length = input_shapes[-1]
 
-        flat_input_ids = tf.reshape(input_ids, (-1, seq_length))
+        flat_input_ids = tf.reshape(input_ids, (-1, seq_length)) if input_ids is not None else None
         flat_attention_mask = tf.reshape(attention_mask, (-1, seq_length)) if attention_mask is not None else None
         flat_token_type_ids = tf.reshape(token_type_ids, (-1, seq_length)) if token_type_ids is not None else None
         flat_position_ids = tf.reshape(position_ids, (-1, seq_length)) if position_ids is not None else None
 
-        flat_inputs = [flat_input_ids, past, flat_attention_mask, flat_token_type_ids, flat_position_ids, head_mask]
+        flat_inputs = [flat_input_ids, past, flat_attention_mask, flat_token_type_ids, flat_position_ids, head_mask, inputs_embeds]
 
         transformer_outputs = self.transformer(flat_inputs, training=training)
         hidden_states = transformer_outputs[0]

transformers/modeling_tf_openai.py

@@ -32,21 +32,12 @@ from .modeling_tf_utils import (TFPreTrainedModel, TFConv1D, TFSharedEmbeddings,
                                 TFSequenceSummary, shape_list, get_initializer)
 from .configuration_openai import OpenAIGPTConfig
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
 TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP = {"openai-gpt": "https://s3.amazonaws.com/models.huggingface.co/bert/openai-gpt-tf_model.h5"}
 
 
-def load_openai_gpt_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def gelu(x):
     """Gaussian Error Linear Unit.
     This is a smoother version of the RELU.
@@ -226,6 +217,9 @@ class TFOpenAIGPTMainLayer(tf.keras.layers.Layer):
                           scale=True,
                           name='h_._{}'.format(i)) for i in range(config.n_layer)]
 
+    def get_input_embeddings(self):
+        return self.tokens_embed
+
     def _resize_token_embeddings(self, new_num_tokens):
         raise NotImplementedError
 
@@ -235,26 +229,38 @@ class TFOpenAIGPTMainLayer(tf.keras.layers.Layer):
         """
         raise NotImplementedError
 
-    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, training=False):
+    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
             token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
             position_ids = inputs[3] if len(inputs) > 3 else position_ids
             head_mask = inputs[4] if len(inputs) > 4 else head_mask
-            assert len(inputs) <= 5, "Too many inputs."
+            inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
+            assert len(inputs) <= 6, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             attention_mask = inputs.get('attention_mask', attention_mask)
             token_type_ids = inputs.get('token_type_ids', token_type_ids)
             position_ids = inputs.get('position_ids', position_ids)
             head_mask = inputs.get('head_mask', head_mask)
-            assert len(inputs) <= 5, "Too many inputs."
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 6, "Too many inputs."
         else:
             input_ids = inputs
 
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = shape_list(input_ids)
+            input_ids = tf.reshape(input_ids, [-1, input_shape[-1]])
+        elif inputs_embeds is not None:
+            input_shape = shape_list(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         if position_ids is None:
-            position_ids = tf.range(shape_list(input_ids)[-1], dtype=tf.int32)[tf.newaxis, :]
+            position_ids = tf.range(input_shape[-1], dtype=tf.int32)[tf.newaxis, :]
 
         if attention_mask is not None:
             # We create a 3D attention mask from a 2D tensor mask.
@@ -286,11 +292,10 @@ class TFOpenAIGPTMainLayer(tf.keras.layers.Layer):
             head_mask = [None] * self.num_hidden_layers
             # head_mask = tf.constant([0] * self.num_hidden_layers)
 
-        input_shape = shape_list(input_ids)
-        input_ids = tf.reshape(input_ids, [-1, input_shape[-1]])
         position_ids = tf.reshape(position_ids, [-1, shape_list(position_ids)[-1]])
 
-        inputs_embeds = self.tokens_embed(input_ids, mode='embedding')
+        if inputs_embeds is None:
+            inputs_embeds = self.tokens_embed(input_ids, mode='embedding')
         position_embeds = self.positions_embed(position_ids)
         if token_type_ids is not None:
             token_type_ids = tf.reshape(token_type_ids, [-1, shape_list(token_type_ids)[-1]])
@@ -335,7 +340,6 @@ class TFOpenAIGPTPreTrainedModel(TFPreTrainedModel):
     """
     config_class = OpenAIGPTConfig
     pretrained_model_archive_map = TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_openai_gpt_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
 
@@ -399,6 +403,10 @@ OPENAI_GPT_INPUTS_DOCSTRING = r"""    Inputs:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare OpenAI GPT transformer model outputing raw hidden-states without any specific head on top.",
@@ -468,6 +476,9 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel):
         super(TFOpenAIGPTLMHeadModel, self).__init__(config, *inputs, **kwargs)
         self.transformer = TFOpenAIGPTMainLayer(config, name='transformer')
 
+    def get_output_embeddings(self):
+        return self.transformer.tokens_embed
+
     def call(self, inputs, **kwargs):
         transformer_outputs = self.transformer(inputs, **kwargs)
         hidden_states = transformer_outputs[0]
@@ -530,36 +541,44 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
         self.transformer = TFOpenAIGPTMainLayer(config, name='transformer')
         self.multiple_choice_head = TFSequenceSummary(config, initializer_range=config.initializer_range, name='multiple_choice_head')
 
-    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, mc_token_ids=None, training=False):
+    def get_output_embeddings(self):
+        return self.transformer.tokens_embed
+
+    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, mc_token_ids=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
             token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
             position_ids = inputs[3] if len(inputs) > 3 else position_ids
             head_mask = inputs[4] if len(inputs) > 4 else head_mask
-            mc_token_ids = inputs[5] if len(inputs) > 5 else mc_token_ids
-            assert len(inputs) <= 6, "Too many inputs."
+            inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
+            mc_token_ids = inputs[6] if len(inputs) > 6 else mc_token_ids
+            assert len(inputs) <= 7, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             attention_mask = inputs.get('attention_mask', attention_mask)
             token_type_ids = inputs.get('token_type_ids', token_type_ids)
             position_ids = inputs.get('position_ids', position_ids)
             head_mask = inputs.get('head_mask', head_mask)
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
             mc_token_ids = inputs.get('mc_token_ids', mc_token_ids)
-            assert len(inputs) <= 6, "Too many inputs."
+            assert len(inputs) <= 7, "Too many inputs."
         else:
             input_ids = inputs
 
-        input_shapes = shape_list(input_ids)
+        if input_ids is not None:
+            input_shapes = shape_list(input_ids)
+        else:
+            input_shapes = shape_list(inputs_embeds)[:-1]
 
         seq_length = input_shapes[-1]
 
-        flat_input_ids = tf.reshape(input_ids, (-1, seq_length))
+        flat_input_ids = tf.reshape(input_ids, (-1, seq_length)) if input_ids is not None else None
         flat_attention_mask = tf.reshape(attention_mask, (-1, seq_length)) if attention_mask is not None else None
         flat_token_type_ids = tf.reshape(token_type_ids, (-1, seq_length)) if token_type_ids is not None else None
         flat_position_ids = tf.reshape(position_ids, (-1, seq_length)) if position_ids is not None else None
 
-        flat_inputs = [flat_input_ids, flat_attention_mask, flat_token_type_ids, flat_position_ids, head_mask]
+        flat_inputs = [flat_input_ids, flat_attention_mask, flat_token_type_ids, flat_position_ids, head_mask, inputs_embeds]
 
         transformer_outputs = self.transformer(flat_inputs, training=training)
         hidden_states = transformer_outputs[0]

transformers/modeling_tf_pytorch_utils.py

@@ -25,8 +25,6 @@ import numpy
 
 logger = logging.getLogger(__name__)
 
-DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-
 def convert_tf_weight_name_to_pt_weight_name(tf_name, start_prefix_to_remove=''):
     """ Convert a TF 2.0 model variable name in a pytorch model weight name.
 
@@ -105,7 +103,7 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
         raise e
 
     if tf_inputs is None:
-        tf_inputs = tf.constant(DUMMY_INPUTS)
+        tf_inputs = tf_model.dummy_inputs
 
     if tf_inputs is not None:
         tfo = tf_model(tf_inputs, training=False)  # Make sure model is built
@@ -200,7 +198,7 @@ def load_tf2_checkpoint_in_pytorch_model(pt_model, tf_checkpoint_path, tf_inputs
     tf_model = tf_model_class(pt_model.config)
 
     if tf_inputs is None:
-        tf_inputs = tf.constant(DUMMY_INPUTS)
+        tf_inputs = tf_model.dummy_inputs
 
     if tf_inputs is not None:
         tfo = tf_model(tf_inputs, training=False)  # Make sure model is built

transformers/modeling_tf_roberta.py

@@ -26,7 +26,6 @@ import tensorflow as tf
 from .configuration_roberta import RobertaConfig
 from .modeling_tf_utils import TFPreTrainedModel, get_initializer
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 from .modeling_tf_bert import TFBertEmbeddings, TFBertMainLayer, gelu, gelu_new
 
@@ -36,16 +35,9 @@ TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'roberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-base-tf_model.h5",
     'roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-tf_model.h5",
     'roberta-large-mnli': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-mnli-tf_model.h5",
+    'distilroberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/distilroberta-base-tf_model.h5",
 }
 
-def load_roberta_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 class TFRobertaEmbeddings(TFBertEmbeddings):
     """
     Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
@@ -56,13 +48,17 @@ class TFRobertaEmbeddings(TFBertEmbeddings):
 
     def _embedding(self, inputs, training=False):
         """Applies embedding based on inputs tensor."""
-        input_ids, position_ids, token_type_ids = inputs
+        input_ids, position_ids, token_type_ids, inputs_embeds = inputs
+
+        if input_ids is not None:
+            seq_length = tf.shape(input_ids)[1]
+        else:
+            seq_length = tf.shape(inputs_embeds)[1]
 
-        seq_length = tf.shape(input_ids)[1]
         if position_ids is None:
             position_ids = tf.range(self.padding_idx+1, seq_length+self.padding_idx+1, dtype=tf.int32)[tf.newaxis, :]
 
-        return super(TFRobertaEmbeddings, self)._embedding([input_ids, position_ids, token_type_ids], training=training)
+        return super(TFRobertaEmbeddings, self)._embedding([input_ids, position_ids, token_type_ids, inputs_embeds], training=training)
 
 
 class TFRobertaMainLayer(TFBertMainLayer):
@@ -73,21 +69,8 @@ class TFRobertaMainLayer(TFBertMainLayer):
         super(TFRobertaMainLayer, self).__init__(config, **kwargs)
         self.embeddings = TFRobertaEmbeddings(config, name='embeddings')
 
-    def call(self, inputs, **kwargs):
-        # Check that input_ids starts with control token
-        if isinstance(inputs, (tuple, list)):
-            input_ids = inputs[0]
-        elif isinstance(inputs, dict):
-            input_ids = inputs.get('input_ids')
-        else:
-            input_ids = inputs
-
-        if tf.not_equal(tf.reduce_sum(input_ids[:, 0]), 0):
-            logger.warning("A sequence with no special tokens has been passed to the RoBERTa model. "
-                           "This model requires special tokens in order to work. "
-                           "Please specify add_special_tokens=True in your encoding.")
-
-        return super(TFRobertaMainLayer, self).call(inputs, **kwargs)
+    def get_input_embeddings(self):
+        return self.embeddings
 
 
 class TFRobertaPreTrainedModel(TFPreTrainedModel):
@@ -96,7 +79,6 @@ class TFRobertaPreTrainedModel(TFPreTrainedModel):
     """
     config_class = RobertaConfig
     pretrained_model_archive_map = TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_roberta_pt_weights_in_tf2
     base_model_prefix = "roberta"
 
 
@@ -182,6 +164,10 @@ ROBERTA_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare RoBERTa Model transformer outputing raw hidden-states without any specific head on top.",
@@ -301,6 +287,9 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel):
         self.roberta = TFRobertaMainLayer(config, name="roberta")
         self.lm_head = TFRobertaLMHead(config, self.roberta.embeddings, name="lm_head")
 
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
+
     def call(self, inputs, **kwargs):
         outputs = self.roberta(inputs, **kwargs)
 
@@ -380,3 +369,54 @@ class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel):
         outputs = (logits,) + outputs[2:]
 
         return outputs  # logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""RoBERTa Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
+class TFRobertaForTokenClassification(TFRobertaPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import RobertaTokenizer, TFRobertaForTokenClassification
+
+        tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
+        model = TFRobertaForTokenClassification.from_pretrained('roberta-base')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        scores = outputs[0]
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFRobertaForTokenClassification, self).__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
+        self.roberta = TFRobertaMainLayer(config, name='roberta')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(config.num_labels,
+                                                kernel_initializer=get_initializer(config.initializer_range),
+                                                name='classifier')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.roberta(inputs, **kwargs)
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output, training=kwargs.get('training', False))
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        return outputs  # scores, (hidden_states), (attentions)

transformers/modeling_tf_transfo_xl.py

@@ -33,7 +33,6 @@ from .configuration_transfo_xl import TransfoXLConfig
 from .modeling_tf_utils import TFPreTrainedModel, TFConv1D, TFSequenceSummary, shape_list, get_initializer
 from .modeling_tf_transfo_xl_utilities import TFAdaptiveSoftmaxMask
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
@@ -41,14 +40,6 @@ TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'transfo-xl-wt103': "https://s3.amazonaws.com/models.huggingface.co/bert/transfo-xl-wt103-tf_model.h5",
 }
 
-def load_transfo_xl_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 class TFPositionalEmbedding(tf.keras.layers.Layer):
     def __init__(self, demb, **kwargs):
         super(TFPositionalEmbedding, self).__init__(**kwargs)
@@ -422,6 +413,9 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
                                             name='r_r_bias')
         super(TFTransfoXLMainLayer, self).build(input_shape)
 
+    def get_input_embeddings(self):
+        return self.word_emb
+
     def _resize_token_embeddings(self, new_num_tokens):
         return self.word_emb
 
@@ -436,11 +430,11 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
     def _prune_heads(self, heads):
         raise NotImplementedError
 
-    def init_mems(self, data):
+    def init_mems(self, bsz):
         if self.mem_len > 0:
             mems = []
             for i in range(self.n_layer):
-                empty = tf.zeros([self.mem_len, shape_list(data)[1], self.d_model])
+                empty = tf.zeros([self.mem_len, bsz, self.d_model])
                 mems.append(empty)
 
             return mems
@@ -470,28 +464,37 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
 
         return new_mems
 
-    def call(self, inputs, mems=None, head_mask=None, training=False):
+    def call(self, inputs, mems=None, head_mask=None, inputs_embeds=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             mems = inputs[1] if len(inputs) > 1 else mems
             head_mask = inputs[2] if len(inputs) > 2 else head_mask
-            assert len(inputs) <= 3, "Too many inputs."
+            inputs_embeds = inputs[3] if len(inputs) > 3 else inputs_embeds
+            assert len(inputs) <= 4, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             mems = inputs.get('mems', mems)
             head_mask = inputs.get('head_mask', head_mask)
-            assert len(inputs) <= 3, "Too many inputs."
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 4, "Too many inputs."
         else:
             input_ids = inputs
 
         # the original code for Transformer-XL used shapes [len, bsz] but we want a unified interface in the library
         # so we transpose here from shape [bsz, len] to shape [len, bsz]
-        input_ids = tf.transpose(input_ids, perm=(1, 0))
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_ids = tf.transpose(input_ids, perm=(1, 0))
+            qlen, bsz = shape_list(input_ids)
+        elif inputs_embeds is not None:
+            inputs_embeds = tf.transpose(inputs_embeds, perm=(1, 0, 2))
+            qlen, bsz = shape_list(inputs_embeds)[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
 
         if mems is None:
-            mems = self.init_mems(input_ids)
-
-        qlen, bsz = shape_list(input_ids)
+            mems = self.init_mems(bsz)
 
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head
@@ -503,7 +506,10 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
         else:
             head_mask = [None] * self.n_layer
 
-        word_emb = self.word_emb(input_ids)
+        if inputs_embeds is not None:
+            word_emb = inputs_embeds
+        else:
+            word_emb = self.word_emb(input_ids)
 
         mlen = shape_list(mems[0])[0] if mems is not None else 0
         klen = mlen + qlen
@@ -577,7 +583,6 @@ class TFTransfoXLPreTrainedModel(TFPreTrainedModel):
     """
     config_class = TransfoXLConfig
     pretrained_model_archive_map = TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_transfo_xl_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
 
@@ -636,6 +641,10 @@ TRANSFO_XL_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare Bert Model transformer outputing raw hidden-states without any specific head on top.",
@@ -726,28 +735,33 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
     def reset_length(self, tgt_len, ext_len, mem_len):
         self.transformer.reset_length(tgt_len, ext_len, mem_len)
 
-    def init_mems(self, data):
-        return self.transformer.init_mems(data)
+    def init_mems(self, bsz):
+        return self.transformer.init_mems(bsz)
 
-    def call(self, inputs, mems=None, head_mask=None, labels=None, training=False):
+    def call(self, inputs, mems=None, head_mask=None, inputs_embeds=None, labels=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             mems = inputs[1] if len(inputs) > 1 else mems
             head_mask = inputs[2] if len(inputs) > 2 else head_mask
-            labels = inputs[3] if len(inputs) > 3 else labels
-            assert len(inputs) <= 4, "Too many inputs."
+            inputs_embeds = inputs[3] if len(inputs) > 3 else inputs_embeds
+            labels = inputs[4] if len(inputs) > 4 else labels
+            assert len(inputs) <= 5, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             mems = inputs.get('mems', mems)
             head_mask = inputs.get('head_mask', head_mask)
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
             labels = inputs.get('labels', labels)
-            assert len(inputs) <= 4, "Too many inputs."
+            assert len(inputs) <= 5, "Too many inputs."
         else:
             input_ids = inputs
 
-        bsz, tgt_len = shape_list(input_ids)[:2]
+        if input_ids is not None:
+            bsz, tgt_len = shape_list(input_ids)[:2]
+        else:
+            bsz, tgt_len = shape_list(inputs_embeds)[:2]
 
-        transformer_outputs = self.transformer([input_ids, mems, head_mask], training=training)
+        transformer_outputs = self.transformer([input_ids, mems, head_mask, inputs_embeds], training=training)
 
         last_hidden = transformer_outputs[0]
         pred_hid = last_hidden[:, -tgt_len:]

transformers/modeling_tf_utils.py

@@ -25,15 +25,17 @@ import tensorflow as tf
 
 from .configuration_utils import PretrainedConfig
 from .file_utils import cached_path, WEIGHTS_NAME, TF_WEIGHTS_NAME, TF2_WEIGHTS_NAME
+from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
+DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
 
 class TFPreTrainedModel(tf.keras.Model):
     r""" Base class for all TF models.
 
         :class:`~transformers.TFPreTrainedModel` takes care of storing the configuration of the models and handles methods for loading/downloading/saving models
-        as well as a few methods commons to all models to (i) resize the input embeddings and (ii) prune heads in the self-attention heads.
+        as well as a few methods common to all models to (i) resize the input embeddings and (ii) prune heads in the self-attention heads.
 
         Class attributes (overridden by derived classes):
             - ``config_class``: a class derived from :class:`~transformers.PretrainedConfig` to use as configuration class for this model architecture.
@@ -48,8 +50,8 @@ class TFPreTrainedModel(tf.keras.Model):
     """
     config_class = None
     pretrained_model_archive_map = {}
-    load_pt_weights = lambda model, config, path: None
     base_model_prefix = ""
+    dummy_inputs = tf.constant(DUMMY_INPUTS)  # dummy inputs to build the network
 
     def __init__(self, config, *inputs, **kwargs):
         super(TFPreTrainedModel, self).__init__(*inputs, **kwargs)
@@ -63,6 +65,21 @@ class TFPreTrainedModel(tf.keras.Model):
         # Save config in model
         self.config = config
 
+    def get_input_embeddings(self):
+        """ Get model's input embeddings
+        """
+        base_model = getattr(self, self.base_model_prefix, self)
+        if base_model is not self:
+            return base_model.get_input_embeddings()
+        else:
+            raise NotImplementedError
+
+    def get_output_embeddings(self):
+        """ Get model's output embeddings
+            Return None if the model doesn't have output embeddings
+        """
+        return None  # Overwrite for models with output embeddings
+
     def _get_resized_embeddings(self, old_embeddings, new_num_tokens=None):
         """ Build a resized Embedding Variable from a provided token Embedding Module.
             Increasing the size will add newly initialized vectors at the end
@@ -262,17 +279,16 @@ class TFPreTrainedModel(tf.keras.Model):
 
         if from_pt:
             # Load from a PyTorch checkpoint
-            return cls.load_pt_weights(model, resolved_archive_file)
+            return load_pytorch_checkpoint_in_tf2_model(model, resolved_archive_file)
 
-        inputs = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
-        ret = model(inputs, training=False)  # build the network with dummy inputs
+        ret = model(model.dummy_inputs, training=False)  # build the network with dummy inputs
 
         assert os.path.isfile(resolved_archive_file), "Error retrieving file {}".format(resolved_archive_file)
         # 'by_name' allow us to do transfer learning by skipping/adding layers
         # see https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/tensorflow/python/keras/engine/network.py#L1339-L1357
         model.load_weights(resolved_archive_file, by_name=True)
 
-        ret = model(inputs, training=False)  # Make sure restore ops are run
+        ret = model(model.dummy_inputs, training=False)  # Make sure restore ops are run
 
         return model
 
@@ -393,26 +409,26 @@ class TFSequenceSummary(tf.keras.layers.Layer):
             # We can probably just use the multi-head attention module of PyTorch >=1.1.0
             raise NotImplementedError
 
-        self.summary = None
-        if hasattr(config, 'summary_use_proj') and config.summary_use_proj:
+        self.has_summary = hasattr(config, 'summary_use_proj') and config.summary_use_proj
+        if self.has_summary:
             if hasattr(config, 'summary_proj_to_labels') and config.summary_proj_to_labels and config.num_labels > 0:
                 num_classes = config.num_labels
             else:
                 num_classes = config.hidden_size
             self.summary = tf.keras.layers.Dense(num_classes,
-                                                 kernel_initializer=get_initializer(initializer_range),
-                                                 name='summary')
+                                                    kernel_initializer=get_initializer(initializer_range),
+                                                    name='summary')
 
-        self.activation = None
-        if hasattr(config, 'summary_activation') and config.summary_activation == 'tanh':
+        self.has_activation = hasattr(config, 'summary_activation') and config.summary_activation == 'tanh'
+        if self.has_activation:
             self.activation = tf.keras.activations.tanh
 
-        self.first_dropout = None
-        if hasattr(config, 'summary_first_dropout') and config.summary_first_dropout > 0:
+        self.has_first_dropout = hasattr(config, 'summary_first_dropout') and config.summary_first_dropout > 0
+        if self.has_first_dropout:
             self.first_dropout = tf.keras.layers.Dropout(config.summary_first_dropout)
 
-        self.last_dropout = None
-        if hasattr(config, 'summary_last_dropout') and config.summary_last_dropout > 0:
+        self.has_last_dropout = hasattr(config, 'summary_last_dropout') and config.summary_last_dropout > 0
+        if self.has_last_dropout:
             self.last_dropout = tf.keras.layers.Dropout(config.summary_last_dropout)
 
     def call(self, inputs, training=False):
@@ -455,17 +471,17 @@ class TFSequenceSummary(tf.keras.layers.Layer):
         elif self.summary_type == 'attn':
             raise NotImplementedError
 
-        if training and self.first_dropout is not None:
-            output = self.first_dropout(output)
+        if self.has_first_dropout:
+            output = self.first_dropout(output, training=training)
 
-        if self.summary is not None:
+        if self.has_summary:
             output = self.summary(output)
 
-        if self.activation is not None:
+        if self.has_activation:
             output = self.activation(output)
 
-        if training and self.last_dropout is not None:
-            output = self.last_dropout(output)
+        if self.has_last_dropout:
+            output = self.last_dropout(output, training=training)
 
         return output
 
@@ -476,10 +492,10 @@ def shape_list(x):
     return [dynamic[i] if s is None else s for i, s in enumerate(static)]
 
 def get_initializer(initializer_range=0.02):
-  """Creates a `tf.initializers.truncated_normal` with the given range.
-  Args:
-    initializer_range: float, initializer range for stddev.
-  Returns:
-    TruncatedNormal initializer with stddev = `initializer_range`.
-  """
-  return tf.keras.initializers.TruncatedNormal(stddev=initializer_range)
+    """Creates a `tf.initializers.truncated_normal` with the given range.
+    Args:
+        initializer_range: float, initializer range for stddev.
+    Returns:
+        TruncatedNormal initializer with stddev = `initializer_range`.
+    """
+    return tf.keras.initializers.TruncatedNormal(stddev=initializer_range)

transformers/modeling_tf_xlm.py

@@ -25,9 +25,8 @@ import numpy as np
 import tensorflow as tf
 
 from .configuration_xlm import XLMConfig
-from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list, get_initializer
+from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list, get_initializer, DUMMY_INPUTS
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
@@ -45,19 +44,6 @@ TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP = {
 }
 
 
-def load_xlm_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
-    attns_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
-    if tf_model.config.use_lang_emb and tf_model.config.n_langs > 1:
-        langs_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
-    else:
-        langs_list = None
-    tf_inputs = [inputs_list, attns_list, langs_list]
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def create_sinusoidal_embeddings(n_pos, dim, out):
     position_enc = np.array([
         [pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)]
@@ -98,7 +84,8 @@ def get_masks(slen, lengths, causal, padding_mask=None, dtype=tf.float32):
         attn_mask = mask
 
     # sanity check
-    assert shape_list(mask) == [bs, slen]
+    # assert shape_list(mask) == [bs, slen]
+    tf.debugging.assert_equal(shape_list(mask), [bs, slen])
     assert causal is False or shape_list(attn_mask) == [bs, slen, slen]
 
     mask = tf.cast(mask, dtype=dtype)
@@ -290,6 +277,9 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
                     self.prune_heads({int(layer): list(map(int, heads))})
 
 
+    def get_input_embeddings(self):
+        return self.embeddings
+
     def _resize_token_embeddings(self, new_num_tokens):
         raise NotImplementedError
 
@@ -301,7 +291,7 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
         raise NotImplementedError
 
     def call(self, inputs, attention_mask=None, langs=None, token_type_ids=None,
-             position_ids=None, lengths=None, cache=None, head_mask=None,
+             position_ids=None, lengths=None, cache=None, head_mask=None, inputs_embeds=None,
              training=False):  # removed: src_enc=None, src_len=None
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
@@ -312,7 +302,8 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
             lengths = inputs[5] if len(inputs) > 5 else lengths
             cache = inputs[6] if len(inputs) > 6 else cache
             head_mask = inputs[7] if len(inputs) > 7 else head_mask
-            assert len(inputs) <= 8, "Too many inputs."
+            inputs_embeds = inputs[8] if len(inputs) > 8 else inputs_embeds
+            assert len(inputs) <= 9, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             attention_mask = inputs.get('attention_mask', attention_mask)
@@ -322,17 +313,30 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
             lengths = inputs.get('lengths', lengths)
             cache = inputs.get('cache', cache)
             head_mask = inputs.get('head_mask', head_mask)
-            assert len(inputs) <= 8, "Too many inputs."
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 9, "Too many inputs."
         else:
             input_ids = inputs
 
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            bs, slen = shape_list(input_ids)
+        elif inputs_embeds is not None:
+            bs, slen = shape_list(inputs_embeds)[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         if lengths is None:
-            lengths = tf.reduce_sum(tf.cast(tf.not_equal(input_ids, self.pad_index), dtype=tf.int32), axis=1)
+            if input_ids is not None:
+                lengths = tf.reduce_sum(tf.cast(tf.not_equal(input_ids, self.pad_index), dtype=tf.int32), axis=1)
+            else:
+                lengths = tf.convert_to_tensor([slen]*bs, tf.int32)
         # mask = input_ids != self.pad_index
 
         # check inputs
-        bs, slen = shape_list(input_ids)
-        assert shape_list(lengths)[0] == bs
+        # assert shape_list(lengths)[0] == bs
+        tf.debugging.assert_equal(shape_list(lengths)[0], bs)
         # assert lengths.max().item() <= slen
         # input_ids = input_ids.transpose(0, 1)  # batch size as dimension 0
         # assert (src_enc is None) == (src_len is None)
@@ -349,12 +353,14 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
         if position_ids is None:
             position_ids = tf.expand_dims(tf.range(slen), axis=0)
         else:
-            assert shape_list(position_ids) == [bs, slen]  # (slen, bs)
+            # assert shape_list(position_ids) == [bs, slen]  # (slen, bs)
+            tf.debugging.assert_equal(shape_list(position_ids), [bs, slen])
             # position_ids = position_ids.transpose(0, 1)
 
         # langs
         if langs is not None:
-            assert shape_list(langs) == [bs, slen]  # (slen, bs)
+            # assert shape_list(langs) == [bs, slen]  # (slen, bs)
+            tf.debugging.assert_equal(shape_list(langs), [bs, slen])
             # langs = langs.transpose(0, 1)
 
         # Prepare head mask if needed
@@ -368,7 +374,7 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
             head_mask = [None] * self.n_layers
 
         # do not recompute cached elements
-        if cache is not None:
+        if cache is not None and input_ids is not None:
             _slen = slen - cache['slen']
             input_ids = input_ids[:, -_slen:]
             position_ids = position_ids[:, -_slen:]
@@ -378,8 +384,10 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
             attn_mask = attn_mask[:, -_slen:]
 
         # embeddings
-        tensor = self.embeddings(input_ids)
-        tensor = tensor + self.position_embeddings(position_ids)
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        tensor = inputs_embeds + self.position_embeddings(position_ids)
         if langs is not None and self.use_lang_emb:
             tensor = tensor + self.lang_embeddings(langs)
         if token_type_ids is not None:
@@ -441,9 +449,19 @@ class TFXLMPreTrainedModel(TFPreTrainedModel):
     """
     config_class = XLMConfig
     pretrained_model_archive_map = TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_xlm_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
+    @property
+    def dummy_inputs(self):
+        # Sometimes XLM has language embeddings so don't forget to build them as well if needed
+        inputs_list = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
+        attns_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
+        if self.config.use_lang_emb and self.config.n_langs > 1:
+            langs_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
+        else:
+            langs_list = None
+        return [inputs_list, attns_list, langs_list]
+
 
 XLM_START_DOCSTRING = r"""    The XLM model was proposed in
     `Cross-lingual Language Model Pretraining`_
@@ -530,6 +548,10 @@ XLM_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare XLM Model transformer outputing raw hidden-states without any specific head on top.",
@@ -637,6 +659,8 @@ class TFXLMWithLMHeadModel(TFXLMPreTrainedModel):
         self.transformer = TFXLMMainLayer(config, name='transformer')
         self.pred_layer = TFXLMPredLayer(config, self.transformer.embeddings, name='pred_layer_._proj')
 
+    def get_output_embeddings(self):
+        return self.pred_layer.input_embeddings
 
     def call(self, inputs, **kwargs):
         transformer_outputs = self.transformer(inputs, **kwargs)

transformers/modeling_tf_xlnet.py

@@ -30,7 +30,6 @@ import tensorflow as tf
 from .configuration_xlnet import XLNetConfig
 from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list, get_initializer
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 
 logger = logging.getLogger(__name__)
@@ -41,13 +40,6 @@ TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP = {
 }
 
 
-def load_xlnet_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)  # build the network
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def gelu(x):
     """ Implementation of the gelu activation function.
         XLNet is using OpenAI GPT's gelu
@@ -362,6 +354,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         super(TFXLNetMainLayer, self).__init__(**kwargs)
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.output_past = config.output_past
 
         self.mem_len = config.mem_len
         self.reuse_len = config.reuse_len
@@ -378,6 +371,9 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         self.layer = [TFXLNetLayer(config, name='layer_._{}'.format(i)) for i in range(config.n_layer)]
         self.dropout = tf.keras.layers.Dropout(config.dropout)
 
+    def get_input_embeddings(self):
+        return self.word_embedding
+
     def build(self, input_shape):
         initializer = get_initializer(self.initializer_range)
         self.mask_emb = self.add_weight(shape=(1, 1, self.d_model),
@@ -421,16 +417,13 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
 
     def cache_mem(self, curr_out, prev_mem):
         """cache hidden states into memory."""
-        if self.mem_len is None or self.mem_len == 0:
-            return None
-        else:
-            if self.reuse_len is not None and self.reuse_len > 0:
-                curr_out = curr_out[:self.reuse_len]
+        if self.reuse_len is not None and self.reuse_len > 0:
+            curr_out = curr_out[:self.reuse_len]
 
-            if prev_mem is None:
-                new_mem = curr_out[-self.mem_len:]
-            else:
-                new_mem = tf.concat([prev_mem, curr_out], 0)[-self.mem_len:]
+        if prev_mem is None:
+            new_mem = curr_out[-self.mem_len:]
+        else:
+            new_mem = tf.concat([prev_mem, curr_out], 0)[-self.mem_len:]
 
         return tf.stop_gradient(new_mem)
 
@@ -494,7 +487,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         return pos_emb
 
     def call(self, inputs, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
-            token_type_ids=None, input_mask=None, head_mask=None, training=False):
+            token_type_ids=None, input_mask=None, head_mask=None, inputs_embeds=None, training=False):
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
@@ -504,7 +497,8 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             token_type_ids = inputs[5] if len(inputs) > 5 else token_type_ids
             input_mask = inputs[6] if len(inputs) > 6 else input_mask
             head_mask = inputs[7] if len(inputs) > 7 else head_mask
-            assert len(inputs) <= 8, "Too many inputs."
+            inputs_embeds = inputs[8] if len(inputs) > 8 else inputs_embeds
+            assert len(inputs) <= 9, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get('input_ids')
             attention_mask = inputs.get('attention_mask', attention_mask)
@@ -514,7 +508,8 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             token_type_ids = inputs.get('token_type_ids', token_type_ids)
             input_mask = inputs.get('input_mask', input_mask)
             head_mask = inputs.get('head_mask', head_mask)
-            assert len(inputs) <= 8, "Too many inputs."
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 9, "Too many inputs."
         else:
             input_ids = inputs
 
@@ -522,14 +517,23 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         # but we want a unified interface in the library with the batch size on the first dimension
         # so we move here the first dimension (batch) to the end
 
-        input_ids = tf.transpose(input_ids, perm=(1, 0))
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_ids = tf.transpose(input_ids, perm=(1, 0))
+            qlen, bsz = shape_list(input_ids)[:2]
+        elif inputs_embeds is not None:
+            inputs_embeds = tf.transpose(inputs_embeds, perm=(1, 0, 2))
+            qlen, bsz = shape_list(inputs_embeds)[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         token_type_ids = tf.transpose(token_type_ids, perm=(1, 0)) if token_type_ids is not None else None
         input_mask = tf.transpose(input_mask, perm=(1, 0)) if input_mask is not None else None
         attention_mask = tf.transpose(attention_mask, perm=(1, 0)) if attention_mask is not None else None
         perm_mask = tf.transpose(perm_mask, perm=(1, 2, 0)) if perm_mask is not None else None
         target_mapping = tf.transpose(target_mapping, perm=(1, 2, 0)) if target_mapping is not None else None
 
-        qlen, bsz = shape_list(input_ids)[:2]
         mlen = shape_list(mems[0])[0] if mems is not None and mems[0] is not None else 0
         klen = mlen + qlen
 
@@ -546,8 +550,8 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             raise ValueError('Unsupported attention type: {}'.format(self.attn_type))
 
         # data mask: input mask & perm mask
-        assert input_mask is None or attention_mask is None, "You can only use one of input_mask (uses 1 for padding) "
-        "or attention_mask (uses 0 for padding, added for compatbility with BERT). Please choose one."
+        assert input_mask is None or attention_mask is None, "You can only use one of input_mask (uses 1 for padding) " \
+            "or attention_mask (uses 0 for padding, added for compatbility with BERT). Please choose one."
         if input_mask is None and attention_mask is not None:
             input_mask = 1.0 - attention_mask
         if input_mask is not None and perm_mask is not None:
@@ -580,7 +584,10 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             non_tgt_mask = None
 
         ##### Word embeddings and prepare h & g hidden states
-        word_emb_k = self.word_embedding(input_ids)
+        if inputs_embeds is not None:
+            word_emb_k = inputs_embeds
+        else:
+            word_emb_k = self.word_embedding(input_ids)
         output_h = self.dropout(word_emb_k, training=training)
         if target_mapping is not None:
             word_emb_q = tf.tile(self.mask_emb, [tf.shape(target_mapping)[0], bsz, 1])
@@ -632,7 +639,8 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         hidden_states = []
         for i, layer_module in enumerate(self.layer):
             # cache new mems
-            new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
+            if self.mem_len is not None and self.mem_len > 0 and self.output_past:
+                new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
             if self.output_hidden_states:
                 hidden_states.append((output_h, output_g) if output_g is not None else output_h)
 
@@ -650,7 +658,11 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         output = self.dropout(output_g if output_g is not None else output_h, training=training)
 
         # Prepare outputs, we transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
-        outputs = (tf.transpose(output, perm=(1, 0, 2)), new_mems)
+        outputs = (tf.transpose(output, perm=(1, 0, 2)),)
+
+        if self.mem_len is not None and self.mem_len > 0 and self.output_past:
+            outputs = outputs + (new_mems,)
+
         if self.output_hidden_states:
             if output_g is not None:
                 hidden_states = tuple(tf.transpose(h, perm=(1, 0, 2)) for hs in hidden_states for h in hs)
@@ -661,7 +673,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
             outputs = outputs + (attentions,)
 
-        return outputs  # outputs, new_mems, (hidden_states), (attentions)
+        return outputs  # outputs, (new_mems), (hidden_states), (attentions)
 
 
 class TFXLNetPreTrainedModel(TFPreTrainedModel):
@@ -670,7 +682,6 @@ class TFXLNetPreTrainedModel(TFPreTrainedModel):
     """
     config_class = XLNetConfig
     pretrained_model_archive_map = TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_xlnet_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
 
@@ -768,6 +779,10 @@ XLNET_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare XLNet Model transformer outputing raw hidden-states without any specific head on top.",
@@ -777,7 +792,7 @@ class TFXLNetModel(TFXLNetPreTrainedModel):
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``tf.Tensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -819,7 +834,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **prediction_scores**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``tf.Tensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -856,6 +871,9 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
         self.transformer = TFXLNetMainLayer(config, name='transformer')
         self.lm_loss = TFXLNetLMHead(config, self.transformer.word_embedding, name='lm_loss')
 
+    def get_output_embeddings(self):
+        return self.lm_loss.input_embeddings
+
     def call(self, inputs, **kwargs):
         transformer_outputs = self.transformer(inputs, **kwargs)
         hidden_state = transformer_outputs[0]
@@ -863,7 +881,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
 
         outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
 
-        return outputs  # return logits, mems, (hidden states), (attentions)
+        return outputs  # return logits, (mems), (hidden states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a sequence classification/regression head on top (a linear layer on top of
@@ -874,7 +892,7 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel):
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **logits**: ``tf.Tensor`` of shape ``(batch_size, config.num_labels)``
             Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``tf.Tensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -918,7 +936,7 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel):
 
         outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
 
-        return outputs  # return logits, mems, (hidden states), (attentions)
+        return outputs  # return logits, (mems), (hidden states), (attentions)
 
 
 # @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
@@ -932,6 +950,11 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
             Span-start scores (before SoftMax).
         **end_scores**: ``tf.Tensor`` of shape ``(batch_size, sequence_length,)``
             Span-end scores (before SoftMax).
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
+            list of ``tf.Tensor`` (one for each layer):
+            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
+            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
+            See details in the docstring of the `mems` input above.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -971,7 +994,7 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
 
         outputs = (start_logits, end_logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
 
-        return outputs  # start_logits, end_logits, (hidden_states), (attentions)
+        return outputs  # start_logits, end_logits, (mems), (hidden_states), (attentions)
 
 # @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
 #     the hidden-states output to compute `span start logits` and `span end logits`). """,

transformers/modeling_transfo_xl.py

@@ -553,6 +553,10 @@ TRANSFO_XL_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
@@ -639,9 +643,12 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
 
         self.init_weights()
 
-    def _resize_token_embeddings(self, new_num_tokens):
+    def get_input_embeddings(self):
         return self.word_emb
 
+    def set_input_embeddings(self, new_embeddings):
+        self.word_emb = new_embeddings
+
     def backward_compatible(self):
         self.sample_softmax = -1
 
@@ -654,12 +661,12 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
         logger.info("Head pruning is not implemented for Transformer-XL model")
         pass
 
-    def init_mems(self, data):
+    def init_mems(self, bsz):
         if self.mem_len > 0:
             mems = []
             param = next(self.parameters())
             for i in range(self.n_layer):
-                empty = torch.zeros(self.mem_len, data.size(1), self.config.d_model,
+                empty = torch.zeros(self.mem_len, bsz, self.config.d_model,
                                     dtype=param.dtype, device=param.device)
                 mems.append(empty)
 
@@ -690,15 +697,22 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
 
         return new_mems
 
-    def forward(self, input_ids, mems=None, head_mask=None):
+    def forward(self, input_ids=None, mems=None, head_mask=None, inputs_embeds=None):
         # the original code for Transformer-XL used shapes [len, bsz] but we want a unified interface in the library
         # so we transpose here from shape [bsz, len] to shape [len, bsz]
-        input_ids = input_ids.transpose(0, 1).contiguous()
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_ids = input_ids.transpose(0, 1).contiguous()
+            qlen, bsz = input_ids.size()
+        elif inputs_embeds is not None:
+            inputs_embeds = inputs_embeds.transpose(0, 1).contiguous()
+            qlen, bsz = inputs_embeds.shape[0], inputs_embeds.shape[1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
 
         if mems is None:
-            mems = self.init_mems(input_ids)
-
-        qlen, bsz = input_ids.size()
+            mems = self.init_mems(bsz)
 
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head
@@ -715,7 +729,10 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
         else:
             head_mask = [None] * self.n_layer
 
-        word_emb = self.word_emb(input_ids)
+        if inputs_embeds is not None:
+            word_emb = inputs_embeds
+        else:
+            word_emb = self.word_emb(input_ids)
 
         mlen = mems[0].size(0) if mems is not None else 0
         klen = mlen + qlen
@@ -826,7 +843,6 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
             self.crit = ProjectedAdaptiveLogSoftmax(config.n_token, config.d_embed, config.d_model,
                                                     config.cutoffs, div_val=config.div_val)
         self.init_weights()
-        self.tie_weights()
 
     def tie_weights(self):
         """
@@ -858,14 +874,18 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
     def reset_length(self, tgt_len, ext_len, mem_len):
         self.transformer.reset_length(tgt_len, ext_len, mem_len)
 
-    def init_mems(self, data):
-        return self.transformer.init_mems(data)
+    def init_mems(self, bsz):
+        return self.transformer.init_mems(bsz)
 
-    def forward(self, input_ids, mems=None, head_mask=None, labels=None):
-        bsz = input_ids.size(0)
-        tgt_len = input_ids.size(1)
+    def forward(self, input_ids=None, mems=None, head_mask=None, inputs_embeds=None, labels=None):
+        if input_ids is not None:
+            bsz, tgt_len = input_ids.size(0), input_ids.size(1)
+        elif inputs_embeds is not None:
+            bsz, tgt_len = inputs_embeds.size(0), inputs_embeds.size(1)
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
 
-        transformer_outputs = self.transformer(input_ids, mems=mems, head_mask=head_mask)
+        transformer_outputs = self.transformer(input_ids, mems=mems, head_mask=head_mask, inputs_embeds=inputs_embeds)
 
         last_hidden = transformer_outputs[0]
         pred_hid = last_hidden[:, -tgt_len:]

transformers/modeling_utils.py

@@ -53,7 +53,7 @@ class PreTrainedModel(nn.Module):
     r""" Base class for all models.
 
         :class:`~transformers.PreTrainedModel` takes care of storing the configuration of the models and handles methods for loading/downloading/saving models
-        as well as a few methods commons to all models to (i) resize the input embeddings and (ii) prune heads in the self-attention heads.
+        as well as a few methods common to all models to (i) resize the input embeddings and (ii) prune heads in the self-attention heads.
 
         Class attributes (overridden by derived classes):
             - ``config_class``: a class derived from :class:`~transformers.PretrainedConfig` to use as configuration class for this model architecture.
@@ -83,6 +83,94 @@ class PreTrainedModel(nn.Module):
         # Save config in model
         self.config = config
 
+    @property
+    def base_model(self):
+        return getattr(self, self.base_model_prefix, self)
+
+    def get_input_embeddings(self):
+        """ Get model's input embeddings
+        """
+        base_model = getattr(self, self.base_model_prefix, self)
+        if base_model is not self:
+            return base_model.get_input_embeddings()
+        else:
+            raise NotImplementedError
+
+    def set_input_embeddings(self, value):
+        """ Set model's input embeddings
+        """
+        base_model = getattr(self, self.base_model_prefix, self)
+        if base_model is not self:
+            base_model.set_input_embeddings(value)
+        else:
+            raise NotImplementedError
+
+    def get_output_embeddings(self):
+        """ Get model's output embeddings
+            Return None if the model doesn't have output embeddings
+        """
+        return None  # Overwrite for models with output embeddings
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        output_embeddings = self.get_output_embeddings()
+        if output_embeddings is not None:
+            self._tie_or_clone_weights(output_embeddings, self.get_input_embeddings())
+
+    def _tie_or_clone_weights(self, output_embeddings, input_embeddings):
+        """ Tie or clone module weights depending of weither we are using TorchScript or not
+        """
+        if self.config.torchscript:
+            output_embeddings.weight = nn.Parameter(input_embeddings.weight.clone())
+        else:
+            output_embeddings.weight = input_embeddings.weight
+
+        if hasattr(output_embeddings, 'bias') and output_embeddings.bias is not None:
+            output_embeddings.bias.data = torch.nn.functional.pad(
+                output_embeddings.bias.data,
+                (0, output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0]),
+                'constant',
+                0
+            )
+        if hasattr(output_embeddings, 'out_features') and hasattr(input_embeddings, 'num_embeddings'):
+            output_embeddings.out_features = input_embeddings.num_embeddings
+
+    def resize_token_embeddings(self, new_num_tokens=None):
+        """ Resize input token embeddings matrix of the model if new_num_tokens != config.vocab_size.
+        Take care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
+
+        Arguments:
+
+            new_num_tokens: (`optional`) int:
+                New number of tokens in the embedding matrix. Increasing the size will add newly initialized vectors at the end. Reducing the size will remove vectors from the end.
+                If not provided or None: does nothing and just returns a pointer to the input tokens ``torch.nn.Embeddings`` Module of the model.
+
+        Return: ``torch.nn.Embeddings``
+            Pointer to the input tokens Embeddings Module of the model
+        """
+        base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
+        model_embeds = base_model._resize_token_embeddings(new_num_tokens)
+        if new_num_tokens is None:
+            return model_embeds
+
+        # Update base model and current model config
+        self.config.vocab_size = new_num_tokens
+        base_model.vocab_size = new_num_tokens
+
+        # Tie weights again if needed
+        if hasattr(self, 'tie_weights'):
+            self.tie_weights()
+
+        return model_embeds
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        old_embeddings = self.get_input_embeddings()
+        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
+        self.set_input_embeddings(new_embeddings)
+        return self.get_input_embeddings()
+
     def _get_resized_embeddings(self, old_embeddings, new_num_tokens=None):
         """ Build a resized Embedding Module from a provided token Embedding Module.
             Increasing the size will add newly initialized vectors at the end
@@ -117,50 +205,6 @@ class PreTrainedModel(nn.Module):
 
         return new_embeddings
 
-    def _tie_or_clone_weights(self, first_module, second_module):
-        """ Tie or clone module weights depending of weither we are using TorchScript or not
-        """
-        if self.config.torchscript:
-            first_module.weight = nn.Parameter(second_module.weight.clone())
-        else:
-            first_module.weight = second_module.weight
-
-        if hasattr(first_module, 'bias') and first_module.bias is not None:
-            first_module.bias.data = torch.nn.functional.pad(
-                first_module.bias.data,
-                (0, first_module.weight.shape[0] - first_module.bias.shape[0]),
-                'constant',
-                0
-            )
-
-    def resize_token_embeddings(self, new_num_tokens=None):
-        """ Resize input token embeddings matrix of the model if new_num_tokens != config.vocab_size.
-        Take care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
-
-        Arguments:
-
-            new_num_tokens: (`optional`) int:
-                New number of tokens in the embedding matrix. Increasing the size will add newly initialized vectors at the end. Reducing the size will remove vectors from the end.
-                If not provided or None: does nothing and just returns a pointer to the input tokens ``torch.nn.Embeddings`` Module of the model.
-
-        Return: ``torch.nn.Embeddings``
-            Pointer to the input tokens Embeddings Module of the model
-        """
-        base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
-        model_embeds = base_model._resize_token_embeddings(new_num_tokens)
-        if new_num_tokens is None:
-            return model_embeds
-
-        # Update base model and current model config
-        self.config.vocab_size = new_num_tokens
-        base_model.vocab_size = new_num_tokens
-
-        # Tie weights again if needed
-        if hasattr(self, 'tie_weights'):
-            self.tie_weights()
-
-        return model_embeds
-
     def init_weights(self):
         """ Initialize and prunes weights if needed. """
         # Initialize weights
@@ -170,6 +214,9 @@ class PreTrainedModel(nn.Module):
         if self.config.pruned_heads:
             self.prune_heads(self.config.pruned_heads)
 
+        # Tie weights if needed
+        self.tie_weights()
+
     def prune_heads(self, heads_to_prune):
         """ Prunes heads of the base model.
 
@@ -178,14 +225,12 @@ class PreTrainedModel(nn.Module):
                 heads_to_prune: dict with keys being selected layer indices (`int`) and associated values being the list of heads to prune in said layer (list of `int`).
                 E.g. {1: [0, 2], 2: [2, 3]} will prune heads 0 and 2 on layer 1 and heads 2 and 3 on layer 2.
         """
-        base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
-
         # save new sets of pruned heads as union of previously stored pruned heads and newly pruned heads
         for layer, heads in heads_to_prune.items():
             union_heads = set(self.config.pruned_heads.get(layer, [])) | set(heads)
             self.config.pruned_heads[layer] = list(union_heads)  # Unfortunately we have to store it as list for JSON
 
-        base_model._prune_heads(heads_to_prune)
+        self.base_model._prune_heads(heads_to_prune)
 
     def save_pretrained(self, save_directory):
         """ Save a model and its configuration file to a directory, so that it
@@ -193,7 +238,7 @@ class PreTrainedModel(nn.Module):
         """
         assert os.path.isdir(save_directory), "Saving path should be a directory where the model and configuration can be saved"
 
-        # Only save the model it-self if we are using distributed training
+        # Only save the model itself if we are using distributed training
         model_to_save = self.module if hasattr(self, 'module') else self
 
         # Save configuration file
@@ -270,6 +315,10 @@ class PreTrainedModel(nn.Module):
             model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
+        if "albert" in pretrained_model_name_or_path and "v2" in pretrained_model_name_or_path:
+            logger.warning("There is currently an upstream reproducibility issue with ALBERT v2 models. Please see " +
+                           "https://github.com/google-research/google-research/issues/119 for more information.")
+
         config = kwargs.pop('config', None)
         state_dict = kwargs.pop('state_dict', None)
         cache_dir = kwargs.pop('cache_dir', None)
@@ -284,6 +333,7 @@ class PreTrainedModel(nn.Module):
                 pretrained_model_name_or_path, *model_args,
                 cache_dir=cache_dir, return_unused_kwargs=True,
                 force_download=force_download,
+                proxies=proxies,
                 **kwargs
             )
         else:
@@ -316,20 +366,20 @@ class PreTrainedModel(nn.Module):
             # redirect to the cache, if necessary
             try:
                 resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
-            except EnvironmentError as e:
+            except EnvironmentError:
                 if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
-                    logger.error(
-                        "Couldn't reach server at '{}' to download pretrained weights.".format(
-                            archive_file))
+                    msg = "Couldn't reach server at '{}' to download pretrained weights.".format(
+                            archive_file)
                 else:
-                    logger.error(
-                        "Model name '{}' was not found in model name list ({}). "
-                        "We assumed '{}' was a path or url but couldn't find any file "
-                        "associated to this path or url.".format(
+                    msg = "Model name '{}' was not found in model name list ({}). " \
+                        "We assumed '{}' was a path or url to model weight files named one of {} but " \
+                        "couldn't find any such file at this path or url.".format(
                             pretrained_model_name_or_path,
                             ', '.join(cls.pretrained_model_archive_map.keys()),
-                            archive_file))
-                raise e
+                            archive_file,
+                            [WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME])
+                raise EnvironmentError(msg)
+
             if resolved_archive_file == archive_file:
                 logger.info("loading weights file {}".format(archive_file))
             else:
@@ -383,6 +433,8 @@ class PreTrainedModel(nn.Module):
             if metadata is not None:
                 state_dict._metadata = metadata
 
+            # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants
+            # so we need to apply the function recursively.
             def load(module, prefix=''):
                 local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
                 module._load_from_state_dict(

transformers/modeling_xlm.py

@@ -73,15 +73,15 @@ def get_masks(slen, lengths, causal, padding_mask=None):
     """
     Generate hidden states mask, and optionally an attention mask.
     """
-    bs = lengths.size(0)
+    alen = torch.arange(slen, dtype=torch.long, device=lengths.device)
     if padding_mask is not None:
         mask = padding_mask
     else:
         assert lengths.max().item() <= slen
-        alen = torch.arange(slen, dtype=torch.long, device=lengths.device)
         mask = alen < lengths[:, None]
 
     # attention mask is the same as mask, or triangular inferior attention (causal)
+    bs = lengths.size(0)
     if causal:
         attn_mask = alen[None, None, :].repeat(bs, slen, 1) <= alen[None, :, None]
     else:
@@ -311,6 +311,10 @@ XLM_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare XLM Model transformer outputting raw hidden-states without any specific head on top.",
@@ -407,10 +411,12 @@ class XLMModel(XLMPreTrainedModel):
 
         self.init_weights()
 
-    def _resize_token_embeddings(self, new_num_tokens):
-        self.embeddings = self._get_resized_embeddings(self.embeddings, new_num_tokens)
+    def get_input_embeddings(self):
         return self.embeddings
 
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings = new_embeddings
+
     def _prune_heads(self, heads_to_prune):
         """ Prunes heads of the model.
             heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
@@ -419,14 +425,21 @@ class XLMModel(XLMPreTrainedModel):
         for layer, heads in heads_to_prune.items():
             self.attentions[layer].prune_heads(heads)
 
-    def forward(self, input_ids, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
-                lengths=None, cache=None, head_mask=None):  # removed: src_enc=None, src_len=None
+    def forward(self, input_ids=None, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
+                lengths=None, cache=None, head_mask=None, inputs_embeds=None):  # removed: src_enc=None, src_len=None
+        if input_ids is not None:
+            bs, slen = input_ids.size()
+        else:
+            bs, slen = inputs_embeds.size()[:-1]
+
         if lengths is None:
-            lengths = (input_ids != self.pad_index).sum(dim=1).long()
+            if input_ids is not None:
+                lengths = (input_ids != self.pad_index).sum(dim=1).long()
+            else:
+                lengths = torch.LongTensor([slen]*bs)
         # mask = input_ids != self.pad_index
 
         # check inputs
-        bs, slen = input_ids.size()
         assert lengths.size(0) == bs
         assert lengths.max().item() <= slen
         # input_ids = input_ids.transpose(0, 1)  # batch size as dimension 0
@@ -440,10 +453,12 @@ class XLMModel(XLMPreTrainedModel):
         # if self.is_decoder and src_enc is not None:
         #     src_mask = torch.arange(src_len.max(), dtype=torch.long, device=lengths.device) < src_len[:, None]
 
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
         # position_ids
         if position_ids is None:
-            position_ids = input_ids.new((slen,)).long()
-            position_ids = torch.arange(slen, out=position_ids).unsqueeze(0)
+            position_ids = torch.arange(slen, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).expand((bs, slen))
         else:
             assert position_ids.size() == (bs, slen)  # (slen, bs)
             # position_ids = position_ids.transpose(0, 1)
@@ -469,7 +484,7 @@ class XLMModel(XLMPreTrainedModel):
             head_mask = [None] * self.n_layers
 
         # do not recompute cached elements
-        if cache is not None:
+        if cache is not None and input_ids is not None:
             _slen = slen - cache['slen']
             input_ids = input_ids[:, -_slen:]
             position_ids = position_ids[:, -_slen:]
@@ -479,8 +494,10 @@ class XLMModel(XLMPreTrainedModel):
             attn_mask = attn_mask[:, -_slen:]
 
         # embeddings
-        tensor = self.embeddings(input_ids)
-        tensor = tensor + self.position_embeddings(position_ids).expand_as(tensor)
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        tensor = inputs_embeds + self.position_embeddings(position_ids).expand_as(inputs_embeds)
         if langs is not None and self.use_lang_emb:
             tensor = tensor + self.lang_embeddings(langs)
         if token_type_ids is not None:
@@ -618,15 +635,12 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
         self.pred_layer = XLMPredLayer(config)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the embeddings
-        """
-        self._tie_or_clone_weights(self.pred_layer.proj, self.transformer.embeddings)
+    def get_output_embeddings(self):
+        return self.pred_layer.proj
 
-    def forward(self, input_ids, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
-                lengths=None, cache=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
+                lengths=None, cache=None, head_mask=None, inputs_embeds=None, labels=None):
         transformer_outputs = self.transformer(input_ids,
                                                attention_mask=attention_mask,
                                                langs=langs,
@@ -634,7 +648,8 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
                                                position_ids=position_ids,
                                                lengths=lengths, 
                                                cache=cache,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
 
         output = transformer_outputs[0]
         outputs = self.pred_layer(output, labels)
@@ -686,8 +701,8 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
-                lengths=None, cache=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
+                lengths=None, cache=None, head_mask=None, inputs_embeds=None, labels=None):
         transformer_outputs = self.transformer(input_ids,
                                                attention_mask=attention_mask,
                                                langs=langs,
@@ -695,7 +710,8 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
                                                position_ids=position_ids,
                                                lengths=lengths, 
                                                cache=cache,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
 
         output = transformer_outputs[0]
         logits = self.sequence_summary(output)
@@ -769,8 +785,8 @@ class XLMForQuestionAnsweringSimple(XLMPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
-                lengths=None, cache=None, head_mask=None, start_positions=None, end_positions=None):
+    def forward(self, input_ids=None, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
+                lengths=None, cache=None, head_mask=None, inputs_embeds=None, start_positions=None, end_positions=None):
         transformer_outputs = self.transformer(input_ids,
                                                attention_mask=attention_mask,
                                                langs=langs,
@@ -778,7 +794,8 @@ class XLMForQuestionAnsweringSimple(XLMPreTrainedModel):
                                                position_ids=position_ids,
                                                lengths=lengths, 
                                                cache=cache,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
 
         sequence_output = transformer_outputs[0]
 
@@ -864,8 +881,8 @@ class XLMForQuestionAnswering(XLMPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
-                lengths=None, cache=None, head_mask=None, start_positions=None, end_positions=None,
+    def forward(self, input_ids=None, attention_mask=None, langs=None, token_type_ids=None, position_ids=None,
+                lengths=None, cache=None, head_mask=None, inputs_embeds=None, start_positions=None, end_positions=None,
                 is_impossible=None, cls_index=None, p_mask=None):
         transformer_outputs = self.transformer(input_ids,
                                                attention_mask=attention_mask,
@@ -874,7 +891,8 @@ class XLMForQuestionAnswering(XLMPreTrainedModel):
                                                position_ids=position_ids,
                                                lengths=lengths, 
                                                cache=cache,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
 
         output = transformer_outputs[0]
 

transformers/modeling_xlnet.py

@@ -188,11 +188,8 @@ def swish(x):
 ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish}
 
 
-try:
-    from apex.normalization.fused_layer_norm import FusedLayerNorm as XLNetLayerNorm
-except (ImportError, AttributeError) as e:
-    logger.info("Better speed can be achieved with apex installed from https://www.github.com/nvidia/apex .")
-    from torch.nn import LayerNorm as XLNetLayerNorm
+XLNetLayerNorm = nn.LayerNorm
+
 
 class XLNetRelativeAttention(nn.Module):
     def __init__(self, config):
@@ -239,45 +236,60 @@ class XLNetRelativeAttention(nn.Module):
 
         return x
 
+    @staticmethod
+    def rel_shift_bnij(x, klen=-1):
+        x_size = x.shape
+
+        x = x.reshape(x_size[0], x_size[1], x_size[3], x_size[2])
+        x = x[:, :, 1:, :]
+        x = x.reshape(x_size[0], x_size[1], x_size[2], x_size[3]-1)
+        # Note: the tensor-slice form was faster in my testing than torch.index_select
+        #       However, tracing doesn't like the nature of the slice, and if klen changes
+        #       during the run then it'll fail, whereas index_select will be fine.
+        x = torch.index_select(x, 3, torch.arange(klen, device=x.device, dtype=torch.long))
+        # x = x[:, :, :, :klen]
+
+        return x
+
     def rel_attn_core(self, q_head, k_head_h, v_head_h, k_head_r, seg_mat=None, attn_mask=None, head_mask=None):
         """Core relative positional attention operations."""
 
         # content based attention score
-        ac = torch.einsum('ibnd,jbnd->ijbn', q_head + self.r_w_bias, k_head_h)
+        ac = torch.einsum('ibnd,jbnd->bnij', q_head + self.r_w_bias, k_head_h)
 
         # position based attention score
-        bd = torch.einsum('ibnd,jbnd->ijbn', q_head + self.r_r_bias, k_head_r)
-        bd = self.rel_shift(bd, klen=ac.shape[1])
+        bd = torch.einsum('ibnd,jbnd->bnij', q_head + self.r_r_bias, k_head_r)
+        bd = self.rel_shift_bnij(bd, klen=ac.shape[3])
 
         # segment based attention score
         if seg_mat is None:
             ef = 0
         else:
             ef = torch.einsum('ibnd,snd->ibns', q_head + self.r_s_bias, self.seg_embed)
-            ef = torch.einsum('ijbs,ibns->ijbn', seg_mat, ef)
+            ef = torch.einsum('ijbs,ibns->bnij', seg_mat, ef)
 
         # merge attention scores and perform masking
         attn_score = (ac + bd + ef) * self.scale
         if attn_mask is not None:
             # attn_score = attn_score * (1 - attn_mask) - 1e30 * attn_mask
             if attn_mask.dtype == torch.float16:
-                attn_score = attn_score - 65500 * attn_mask
+                attn_score = attn_score - 65500 * torch.einsum('ijbn->bnij', attn_mask)
             else:
-                attn_score = attn_score - 1e30 * attn_mask
+                attn_score = attn_score - 1e30 * torch.einsum('ijbn->bnij', attn_mask)
 
         # attention probability
-        attn_prob = F.softmax(attn_score, dim=1)
+        attn_prob = F.softmax(attn_score, dim=3)
         attn_prob = self.dropout(attn_prob)
 
         # Mask heads if we want to
         if head_mask is not None:
-            attn_prob = attn_prob * head_mask
+            attn_prob = attn_prob * torch.einsum('ijbn->bnij', head_mask)
 
         # attention output
-        attn_vec = torch.einsum('ijbn,jbnd->ibnd', attn_prob, v_head_h)
+        attn_vec = torch.einsum('bnij,jbnd->ibnd', attn_prob, v_head_h)
 
         if self.output_attentions:
-            return attn_vec, attn_prob
+            return attn_vec, torch.einsum('bnij->ijbn', attn_prob)
 
         return attn_vec
 
@@ -546,6 +558,10 @@ XLNET_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
 """
 
 @add_start_docstrings("The bare XLNet Model transformer outputting raw hidden-states without any specific head on top.",
@@ -555,7 +571,7 @@ class XLNetModel(XLNetPreTrainedModel):
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -581,6 +597,7 @@ class XLNetModel(XLNetPreTrainedModel):
         super(XLNetModel, self).__init__(config)
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.output_past = config.output_past
 
         self.mem_len = config.mem_len
         self.reuse_len = config.reuse_len
@@ -598,10 +615,12 @@ class XLNetModel(XLNetPreTrainedModel):
 
         self.init_weights()
 
-    def _resize_token_embeddings(self, new_num_tokens):
-        self.word_embedding = self._get_resized_embeddings(self.word_embedding, new_num_tokens)
+    def get_input_embeddings(self):
         return self.word_embedding
 
+    def set_input_embeddings(self, new_embeddings):
+        self.word_embedding = new_embeddings
+
     def _prune_heads(self, heads_to_prune):
         raise NotImplementedError
 
@@ -637,16 +656,13 @@ class XLNetModel(XLNetPreTrainedModel):
 
     def cache_mem(self, curr_out, prev_mem):
         """cache hidden states into memory."""
-        if self.mem_len is None or self.mem_len == 0:
-            return None
-        else:
-            if self.reuse_len is not None and self.reuse_len > 0:
-                curr_out = curr_out[:self.reuse_len]
+        if self.reuse_len is not None and self.reuse_len > 0:
+            curr_out = curr_out[:self.reuse_len]
 
-            if prev_mem is None:
-                new_mem = curr_out[-self.mem_len:]
-            else:
-                new_mem = torch.cat([prev_mem, curr_out], dim=0)[-self.mem_len:]
+        if prev_mem is None:
+            new_mem = curr_out[-self.mem_len:]
+        else:
+            new_mem = torch.cat([prev_mem, curr_out], dim=0)[-self.mem_len:]
 
         return new_mem.detach()
 
@@ -700,19 +716,29 @@ class XLNetModel(XLNetPreTrainedModel):
         pos_emb = pos_emb.to(next(self.parameters()))
         return pos_emb
 
-    def forward(self, input_ids, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
-                token_type_ids=None, input_mask=None, head_mask=None):
+    def forward(self, input_ids=None, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
+                token_type_ids=None, input_mask=None, head_mask=None, inputs_embeds=None):
         # the original code for XLNet uses shapes [len, bsz] with the batch dimension at the end
         # but we want a unified interface in the library with the batch size on the first dimension
         # so we move here the first dimension (batch) to the end
-        input_ids = input_ids.transpose(0, 1).contiguous()
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_ids = input_ids.transpose(0, 1).contiguous()
+            qlen, bsz = input_ids.shape[0], input_ids.shape[1]
+        elif inputs_embeds is not None:
+            inputs_embeds.transpose(0, 1).contiguous()
+            qlen, bsz = inputs_embeds.shape[0], inputs_embeds.shape[1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         token_type_ids = token_type_ids.transpose(0, 1).contiguous() if token_type_ids is not None else None
         input_mask = input_mask.transpose(0, 1).contiguous() if input_mask is not None else None
         attention_mask = attention_mask.transpose(0, 1).contiguous() if attention_mask is not None else None
         perm_mask = perm_mask.permute(1, 2, 0).contiguous() if perm_mask is not None else None
         target_mapping = target_mapping.permute(1, 2, 0).contiguous() if target_mapping is not None else None
 
-        qlen, bsz = input_ids.shape[0], input_ids.shape[1]
+
         mlen = mems[0].shape[0] if mems is not None and mems[0] is not None else 0
         klen = mlen + qlen
 
@@ -765,7 +791,10 @@ class XLNetModel(XLNetPreTrainedModel):
             non_tgt_mask = None
 
         ##### Word embeddings and prepare h & g hidden states
-        word_emb_k = self.word_embedding(input_ids)
+        if inputs_embeds is not None:
+            word_emb_k = inputs_embeds
+        else:
+            word_emb_k = self.word_embedding(input_ids)
         output_h = self.dropout(word_emb_k)
         if target_mapping is not None:
             word_emb_q = self.mask_emb.expand(target_mapping.shape[0], bsz, -1)
@@ -817,8 +846,9 @@ class XLNetModel(XLNetPreTrainedModel):
         attentions = []
         hidden_states = []
         for i, layer_module in enumerate(self.layer):
-            # cache new mems
-            new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
+            if self.mem_len is not None and self.mem_len > 0 and self.output_past:
+                # cache new mems
+                new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
             if self.output_hidden_states:
                 hidden_states.append((output_h, output_g) if output_g is not None else output_h)
 
@@ -836,7 +866,11 @@ class XLNetModel(XLNetPreTrainedModel):
         output = self.dropout(output_g if output_g is not None else output_h)
 
         # Prepare outputs, we transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
-        outputs = (output.permute(1, 0, 2).contiguous(), new_mems)
+        outputs = (output.permute(1, 0, 2).contiguous(),)
+
+        if self.mem_len is not None and self.mem_len > 0 and self.output_past:
+            outputs = outputs + (new_mems,)
+
         if self.output_hidden_states:
             if output_g is not None:
                 hidden_states = tuple(h.permute(1, 0, 2).contiguous() for hs in hidden_states for h in hs)
@@ -847,7 +881,7 @@ class XLNetModel(XLNetPreTrainedModel):
             attentions = tuple(t.permute(2, 3, 0, 1).contiguous() for t in attentions)
             outputs = outputs + (attentions,)
 
-        return outputs  # outputs, new_mems, (hidden_states), (attentions)
+        return outputs  # outputs, (new_mems), (hidden_states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a language modeling head on top
@@ -867,7 +901,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
             Language modeling loss.
         **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -903,23 +937,21 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
         self.lm_loss = nn.Linear(config.d_model, config.n_token, bias=True)
 
         self.init_weights()
-        self.tie_weights()
 
-    def tie_weights(self):
-        """ Make sure we are sharing the embeddings
-        """
-        self._tie_or_clone_weights(self.lm_loss, self.transformer.word_embedding)
+    def get_output_embeddings(self):
+        return self.lm_loss
 
-    def forward(self, input_ids, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
-                token_type_ids=None, input_mask=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
+                token_type_ids=None, input_mask=None, head_mask=None, inputs_embeds=None, labels=None):
         transformer_outputs = self.transformer(input_ids,
                                                attention_mask=attention_mask,
                                                mems=mems,
                                                perm_mask=perm_mask,
                                                target_mapping=target_mapping,
                                                token_type_ids=token_type_ids,
-                                               input_mask=input_mask, 
-                                               head_mask=head_mask)
+                                               input_mask=input_mask,
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
 
         logits = self.lm_loss(transformer_outputs[0])
 
@@ -932,7 +964,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
                             labels.view(-1))
             outputs = (loss,) + outputs
 
-        return outputs  # return (loss), logits, mems, (hidden states), (attentions)
+        return outputs  # return (loss), logits, (mems), (hidden states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a sequence classification/regression head on top (a linear layer on top of
@@ -951,7 +983,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
             Classification (or regression if config.num_labels==1) loss.
         **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
             Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -984,16 +1016,17 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
-                token_type_ids=None, input_mask=None, head_mask=None, labels=None):
+    def forward(self, input_ids=None, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
+                token_type_ids=None, input_mask=None, head_mask=None, inputs_embeds=None, labels=None):
         transformer_outputs = self.transformer(input_ids,
                                                attention_mask=attention_mask,
                                                mems=mems,
                                                perm_mask=perm_mask,
                                                target_mapping=target_mapping,
                                                token_type_ids=token_type_ids,
-                                               input_mask=input_mask, 
-                                               head_mask=head_mask)
+                                               input_mask=input_mask,
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
         output = transformer_outputs[0]
 
         output = self.sequence_summary(output)
@@ -1011,7 +1044,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
             outputs = (loss,) + outputs
 
-        return outputs  # return (loss), logits, mems, (hidden states), (attentions)
+        return outputs  # return (loss), logits, (mems), (hidden states), (attentions)
 
 @add_start_docstrings("""XLNet Model with a multiple choice classification head on top (a linear layer on top of
     the pooled output and a softmax) e.g. for RACE/SWAG tasks. """,
@@ -1035,6 +1068,10 @@ class XLNetForMultipleChoice(XLNetPreTrainedModel):
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
@@ -1046,6 +1083,11 @@ class XLNetForMultipleChoice(XLNetPreTrainedModel):
         **classification_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above).
             Classification scores (before SoftMax).
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
+            list of ``torch.FloatTensor`` (one for each layer):
+            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
+            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
+            See details in the docstring of the `mems` input above.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -1074,9 +1116,9 @@ class XLNetForMultipleChoice(XLNetPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, token_type_ids=None, input_mask=None, attention_mask=None,
+    def forward(self, input_ids=None, token_type_ids=None, input_mask=None, attention_mask=None,
                 mems=None, perm_mask=None, target_mapping=None,
-                labels=None, head_mask=None):
+                labels=None, head_mask=None, inputs_embeds=None):
         num_choices = input_ids.shape[1]
 
         flat_input_ids = input_ids.view(-1, input_ids.size(-1))
@@ -1087,7 +1129,7 @@ class XLNetForMultipleChoice(XLNetPreTrainedModel):
         transformer_outputs = self.transformer(flat_input_ids, token_type_ids=flat_token_type_ids,
                                                input_mask=flat_input_mask, attention_mask=flat_attention_mask,
                                                mems=mems, perm_mask=perm_mask, target_mapping=target_mapping,
-                                               head_mask=head_mask)
+                                               head_mask=head_mask, inputs_embeds=inputs_embeds)
 
 
         output = transformer_outputs[0]
@@ -1102,7 +1144,7 @@ class XLNetForMultipleChoice(XLNetPreTrainedModel):
             loss = loss_fct(reshaped_logits, labels.view(-1))
             outputs = (loss,) + outputs
 
-        return outputs  # return (loss), logits, mems, (hidden states), (attentions)
+        return outputs  # return (loss), logits, (mems), (hidden states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
@@ -1126,7 +1168,7 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
             Span-start scores (before SoftMax).
         **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
             Span-end scores (before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -1159,8 +1201,8 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
-                token_type_ids=None, input_mask=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
+                token_type_ids=None, input_mask=None, head_mask=None, inputs_embeds=None,
                 start_positions=None, end_positions=None):
 
         outputs = self.transformer(input_ids,
@@ -1169,8 +1211,9 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
                                     perm_mask=perm_mask,
                                     target_mapping=target_mapping,
                                     token_type_ids=token_type_ids,
-                                    input_mask=input_mask, 
-                                    head_mask=head_mask)
+                                    input_mask=input_mask,
+                                    head_mask=head_mask,
+                                    inputs_embeds=inputs_embeds)
 
         sequence_output = outputs[0]
 
@@ -1197,7 +1240,7 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
             total_loss = (start_loss + end_loss) / 2
             outputs = (total_loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        return outputs  # (loss), start_logits, end_logits, (mems), (hidden_states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
@@ -1239,7 +1282,7 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
         **cls_logits**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
             ``torch.FloatTensor`` of shape ``(batch_size,)``
             Log probabilities for the ``is_impossible`` label of the answers.
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -1275,8 +1318,8 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
-                token_type_ids=None, input_mask=None, head_mask=None,
+    def forward(self, input_ids=None, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
+                token_type_ids=None, input_mask=None, head_mask=None, inputs_embeds=None,
                 start_positions=None, end_positions=None, is_impossible=None, cls_index=None, p_mask=None,):
         transformer_outputs = self.transformer(input_ids,
                                                attention_mask=attention_mask,
@@ -1284,8 +1327,9 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
                                                perm_mask=perm_mask,
                                                target_mapping=target_mapping,
                                                token_type_ids=token_type_ids,
-                                               input_mask=input_mask, 
-                                               head_mask=head_mask)
+                                               input_mask=input_mask,
+                                               head_mask=head_mask,
+                                               inputs_embeds=inputs_embeds)
         hidden_states = transformer_outputs[0]
         start_logits = self.start_logits(hidden_states, p_mask=p_mask)
 

transformers/optimization.py

@@ -23,85 +23,65 @@ from torch.optim.lr_scheduler import LambdaLR
 
 logger = logging.getLogger(__name__)
 
-class ConstantLRSchedule(LambdaLR):
-    """ Constant learning rate schedule.
+
+def get_constant_schedule(optimizer, last_epoch=-1):
+    """ Create a schedule with a constant learning rate.
     """
-    def __init__(self, optimizer, last_epoch=-1):
-        super(ConstantLRSchedule, self).__init__(optimizer, lambda _: 1.0, last_epoch=last_epoch)
+    return LambdaLR(optimizer, lambda _: 1, last_epoch=last_epoch)
 
 
-class WarmupConstantSchedule(LambdaLR):
-    """ Linear warmup and then constant.
-        Linearly increases learning rate schedule from 0 to 1 over `warmup_steps` training steps.
-        Keeps learning rate schedule equal to 1. after warmup_steps.
+def get_constant_schedule_with_warmup(optimizer, num_warmup_steps, last_epoch=-1):
+    """ Create a schedule with a constant learning rate preceded by a warmup
+    period during which the learning rate increases linearly between 0 and 1.
     """
-    def __init__(self, optimizer, warmup_steps, last_epoch=-1):
-        self.warmup_steps = warmup_steps
-        super(WarmupConstantSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch)
-
-    def lr_lambda(self, step):
-        if step < self.warmup_steps:
-            return float(step) / float(max(1.0, self.warmup_steps))
+    def lr_lambda(current_step):
+        if current_step < num_warmup_steps:
+            return float(current_step) / float(max(1.0, num_warmup_steps))
         return 1.
 
+    return LambdaLR(optimizer, lr_lambda, last_epoch=last_epoch)
 
-class WarmupLinearSchedule(LambdaLR):
-    """ Linear warmup and then linear decay.
-        Linearly increases learning rate from 0 to 1 over `warmup_steps` training steps.
-        Linearly decreases learning rate from 1. to 0. over remaining `t_total - warmup_steps` steps.
+
+def get_linear_schedule_with_warmup(optimizer, num_warmup_steps, num_training_steps, last_epoch=-1):
+    """ Create a schedule with a learning rate that decreases linearly after
+    linearly increasing during a warmup period.
     """
-    def __init__(self, optimizer, warmup_steps, t_total, last_epoch=-1):
-        self.warmup_steps = warmup_steps
-        self.t_total = t_total
-        super(WarmupLinearSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch)
+    def lr_lambda(current_step):
+        if current_step < num_warmup_steps:
+            return float(current_step) / float(max(1, num_warmup_steps))
+        return max(0.0, float(num_training_steps - current_step) / float(max(1, num_training_steps - num_warmup_steps)))
 
-    def lr_lambda(self, step):
-        if step < self.warmup_steps:
-            return float(step) / float(max(1, self.warmup_steps))
-        return max(0.0, float(self.t_total - step) / float(max(1.0, self.t_total - self.warmup_steps)))
+    return LambdaLR(optimizer, lr_lambda, last_epoch)
 
 
-class WarmupCosineSchedule(LambdaLR):
-    """ Linear warmup and then cosine decay.
-        Linearly increases learning rate from 0 to 1 over `warmup_steps` training steps.
-        Decreases learning rate from 1. to 0. over remaining `t_total - warmup_steps` steps following a cosine curve.
-        If `cycles` (default=0.5) is different from default, learning rate follows cosine function after warmup.
+def get_cosine_schedule_with_warmup(optimizer, num_warmup_steps, num_training_steps, num_cycles=.5, last_epoch=-1):
+    """ Create a schedule with a learning rate that decreases following the
+    values of the cosine function between 0 and `pi * cycles` after a warmup
+    period during which it increases linearly between 0 and 1.
     """
-    def __init__(self, optimizer, warmup_steps, t_total, cycles=.5, last_epoch=-1):
-        self.warmup_steps = warmup_steps
-        self.t_total = t_total
-        self.cycles = cycles
-        super(WarmupCosineSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch)
+    def lr_lambda(current_step):
+        if current_step < num_warmup_steps:
+            return float(current_step) / float(max(1, num_warmup_steps))
+        progress = float(current_step - num_warmup_steps) / float(max(1, num_training_steps - num_warmup_steps))
+        return max(0., 0.5 * (1. + math.cos(math.pi * float(num_cycles) * 2. * progress)))
 
-    def lr_lambda(self, step):
-        if step < self.warmup_steps:
-            return float(step) / float(max(1.0, self.warmup_steps))
-        # progress after warmup
-        progress = float(step - self.warmup_steps) / float(max(1, self.t_total - self.warmup_steps))
-        return max(0.0, 0.5 * (1. + math.cos(math.pi * float(self.cycles) * 2.0 * progress)))
+    return LambdaLR(optimizer, lr_lambda, last_epoch)
 
 
-class WarmupCosineWithHardRestartsSchedule(LambdaLR):
-    """ Linear warmup and then cosine cycles with hard restarts.
-        Linearly increases learning rate from 0 to 1 over `warmup_steps` training steps.
-        If `cycles` (default=1.) is different from default, learning rate follows `cycles` times a cosine decaying
-        learning rate (with hard restarts).
+def get_cosine_with_hard_restarts_schedule_with_warmup(optimizer, num_warmup_steps, num_training_steps, num_cycles=1., last_epoch=-1):
+    """ Create a schedule with a learning rate that decreases following the
+    values of the cosine function with several hard restarts, after a warmup
+    period during which it increases linearly between 0 and 1.
     """
-    def __init__(self, optimizer, warmup_steps, t_total, cycles=1., last_epoch=-1):
-        self.warmup_steps = warmup_steps
-        self.t_total = t_total
-        self.cycles = cycles
-        super(WarmupCosineWithHardRestartsSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch)
-
-    def lr_lambda(self, step):
-        if step < self.warmup_steps:
-            return float(step) / float(max(1, self.warmup_steps))
-        # progress after warmup
-        progress = float(step - self.warmup_steps) / float(max(1, self.t_total - self.warmup_steps))
-        if progress >= 1.0:
-            return 0.0
-        return max(0.0, 0.5 * (1. + math.cos(math.pi * ((float(self.cycles) * progress) % 1.0))))
+    def lr_lambda(current_step):
+        if current_step < num_warmup_steps:
+            return float(current_step) / float(max(1, num_warmup_steps))
+        progress = float(current_step - num_warmup_steps) / float(max(1, num_training_steps - num_warmup_steps))
+        if progress >= 1.:
+            return 0.
+        return max(0., 0.5 * (1. + math.cos(math.pi * ((float(num_cycles) * progress) % 1.))))
 
+    return LambdaLR(optimizer, lr_lambda, last_epoch)
 
 
 class AdamW(Optimizer):

transformers/tests/conftest.py

@@ -7,6 +7,13 @@ def pytest_addoption(parser):
     parser.addoption(
         "--runslow", action="store_true", default=False, help="run slow tests"
     )
+    parser.addoption(
+        "--use_cuda", action="store_true", default=False, help="run tests on gpu"
+    )
+
+
+def pytest_configure(config):
+    config.addinivalue_line("markers", "slow: mark test as slow to run")
 
 
 def pytest_collection_modifyitems(config, items):
@@ -17,3 +24,8 @@ def pytest_collection_modifyitems(config, items):
     for item in items:
         if "slow" in item.keywords:
             item.add_marker(skip_slow)
+
+@pytest.fixture
+def use_cuda(request):
+    """ Run test on gpu """
+    return request.config.getoption("--use_cuda")

transformers/tests/modeling_albert_test.py

@@ -0,0 +1,237 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import pytest
+
+from transformers import is_torch_available
+
+from .modeling_common_test import (CommonTestCases, ids_tensor)
+from .configuration_common_test import ConfigTester
+
+if is_torch_available():
+    from transformers import (AlbertConfig, AlbertModel, AlbertForMaskedLM,
+                              AlbertForSequenceClassification, AlbertForQuestionAnswering,
+                              )
+    from transformers.modeling_albert import ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+else:
+    pytestmark = pytest.mark.skip("Require Torch")
+
+
+class AlbertModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes = (AlbertModel, AlbertForMaskedLM) if is_torch_available() else ()
+
+    class AlbertModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_input_mask=True,
+                     use_token_type_ids=True,
+                     use_labels=True,
+                     vocab_size=99,
+                     embedding_size=16,
+                     hidden_size=36,
+                     num_hidden_layers=6,
+                     num_hidden_groups=6,
+                     num_attention_heads=6,
+                     intermediate_size=37,
+                     hidden_act="gelu",
+                     hidden_dropout_prob=0.1,
+                     attention_probs_dropout_prob=0.1,
+                     max_position_embeddings=512,
+                     type_vocab_size=16,
+                     type_sequence_label_size=2,
+                     initializer_range=0.02,
+                     num_labels=3,
+                     num_choices=4,
+                     scope=None,
+                    ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_input_mask = use_input_mask
+            self.use_token_type_ids = use_token_type_ids
+            self.use_labels = use_labels
+            self.vocab_size = vocab_size
+            self.embedding_size = embedding_size
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.intermediate_size = intermediate_size
+            self.hidden_act = hidden_act
+            self.hidden_dropout_prob = hidden_dropout_prob
+            self.attention_probs_dropout_prob = attention_probs_dropout_prob
+            self.max_position_embeddings = max_position_embeddings
+            self.type_vocab_size = type_vocab_size
+            self.type_sequence_label_size = type_sequence_label_size
+            self.initializer_range = initializer_range
+            self.num_labels = num_labels
+            self.num_choices = num_choices
+            self.scope = scope
+            self.num_hidden_groups = num_hidden_groups
+
+        def prepare_config_and_inputs(self):
+            input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+            input_mask = None
+            if self.use_input_mask:
+                input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+            token_type_ids = None
+            if self.use_token_type_ids:
+                token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+            sequence_labels = None
+            token_labels = None
+            choice_labels = None
+            if self.use_labels:
+                sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+                choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+            config = AlbertConfig(
+                vocab_size_or_config_json_file=self.vocab_size,
+                hidden_size=self.hidden_size,
+                num_hidden_layers=self.num_hidden_layers,
+                num_attention_heads=self.num_attention_heads,
+                intermediate_size=self.intermediate_size,
+                hidden_act=self.hidden_act,
+                hidden_dropout_prob=self.hidden_dropout_prob,
+                attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+                max_position_embeddings=self.max_position_embeddings,
+                type_vocab_size=self.type_vocab_size,
+                initializer_range=self.initializer_range,
+                num_hidden_groups=self.num_hidden_groups)
+
+            return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+        def check_loss_output(self, result):
+            self.parent.assertListEqual(
+                list(result["loss"].size()),
+                [])
+
+        def create_and_check_albert_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = AlbertModel(config=config)
+            model.eval()
+            sequence_output, pooled_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+            sequence_output, pooled_output = model(input_ids, token_type_ids=token_type_ids)
+            sequence_output, pooled_output = model(input_ids)
+
+            result = {
+                "sequence_output": sequence_output,
+                "pooled_output": pooled_output,
+            }
+            self.parent.assertListEqual(
+                list(result["sequence_output"].size()),
+                [self.batch_size, self.seq_length, self.hidden_size])
+            self.parent.assertListEqual(list(result["pooled_output"].size()), [self.batch_size, self.hidden_size])
+
+
+        def create_and_check_albert_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = AlbertForMaskedLM(config=config)
+            model.eval()
+            loss, prediction_scores = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, masked_lm_labels=token_labels)
+            result = {
+                "loss": loss,
+                "prediction_scores": prediction_scores,
+            }
+            self.parent.assertListEqual(
+                list(result["prediction_scores"].size()),
+                [self.batch_size, self.seq_length, self.vocab_size])
+            self.check_loss_output(result)
+
+        def create_and_check_albert_for_question_answering(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = AlbertForQuestionAnswering(config=config)
+            model.eval()
+            loss, start_logits, end_logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids,
+                                                   start_positions=sequence_labels, end_positions=sequence_labels)
+            result = {
+                "loss": loss,
+                "start_logits": start_logits,
+                "end_logits": end_logits,
+            }
+            self.parent.assertListEqual(
+                list(result["start_logits"].size()),
+                [self.batch_size, self.seq_length])
+            self.parent.assertListEqual(
+                list(result["end_logits"].size()),
+                [self.batch_size, self.seq_length])
+            self.check_loss_output(result)
+
+
+        def create_and_check_albert_for_sequence_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            config.num_labels = self.num_labels
+            model = AlbertForSequenceClassification(config)
+            model.eval()
+            loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
+            result = {
+                "loss": loss,
+                "logits": logits,
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].size()),
+                [self.batch_size, self.num_labels])
+            self.check_loss_output(result)
+
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids, token_type_ids, input_mask,
+             sequence_labels, token_labels, choice_labels) = config_and_inputs
+            inputs_dict = {'input_ids': input_ids, 'token_type_ids': token_type_ids, 'attention_mask': input_mask}
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = AlbertModelTest.AlbertModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=AlbertConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_albert_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_model(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_for_masked_lm(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_for_sequence_classification(*config_and_inputs)
+
+    @pytest.mark.slow
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/transformers_test/"
+        for model_name in list(ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            model = AlbertModel.from_pretrained(model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
+if __name__ == "__main__":
+    unittest.main()