SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Merge branch 'master' into saving-and-resuming

Browse Source
This commit is contained in:
Thomas Wolf
2019-12-21 14:29:59 +01:00
committed by GitHub
parent 6aa919469d ed9b84816e
commit eeb70cdd77
135 changed files with 9041 additions and 1529 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
.circleci/config.yml
View File

@@ -1,9 +1,11 @@
version: 2
jobs:
build_py3_torch_and_tf:
run_tests_py3_torch_and_tf:
working_directory: ~/transformers
docker:
- image: circleci/python:3.5
environment:
OMP_NUM_THREADS: 1
resource_class: xlarge
parallelism: 1
steps:
@@ -11,65 +13,67 @@ jobs:
- run: sudo pip install torch
- run: sudo pip install tensorflow
- run: sudo pip install --progress-bar off .
- run: sudo pip install pytest codecov pytest-cov
- run: sudo pip install pytest codecov pytest-cov pytest-xdist
- run: sudo pip install tensorboardX scikit-learn
- run: python -m pytest -sv ./transformers/tests/ --cov
- run: python -m pytest -n 8 --dist=loadfile -s -v ./transformers/tests/ --cov
- run: codecov
build_py3_torch:
run_tests_py3_torch:
working_directory: ~/transformers
docker:
- image: circleci/python:3.5
environment:
OMP_NUM_THREADS: 1
resource_class: xlarge
parallelism: 1
steps:
- checkout
- run: sudo pip install torch
- run: sudo pip install --progress-bar off .
- run: sudo pip install pytest codecov pytest-cov
- run: sudo pip install pytest codecov pytest-cov pytest-xdist
- run: sudo pip install tensorboardX scikit-learn
- run: python -m pytest -sv ./transformers/tests/ --cov
- run: python -m pytest -sv ./examples/
- run: python -m pytest -n 8 --dist=loadfile -s -v ./transformers/tests/ --cov
- run: codecov
build_py3_tf:
run_tests_py3_tf:
working_directory: ~/transformers
docker:
- image: circleci/python:3.5
environment:
OMP_NUM_THREADS: 1
resource_class: xlarge
parallelism: 1
steps:
- checkout
- run: sudo pip install tensorflow
- run: sudo pip install --progress-bar off .
- run: sudo pip install pytest codecov pytest-cov
- run: sudo pip install pytest codecov pytest-cov pytest-xdist
- run: sudo pip install tensorboardX scikit-learn
- run: python -m pytest -sv ./transformers/tests/ --cov
- run: python -m pytest -n 8 --dist=loadfile -s -v ./transformers/tests/ --cov
- run: codecov
build_py2_torch:
run_tests_py3_custom_tokenizers:
working_directory: ~/transformers
resource_class: large
parallelism: 1
docker:
- image: circleci/python:2.7
- image: circleci/python:3.5
steps:
- checkout
- run: sudo pip install --progress-bar off .
- run: sudo pip install pytest pytest-xdist
- run: sudo pip install mecab-python3
- run: RUN_CUSTOM_TOKENIZERS=1 python -m pytest -sv ./transformers/tests/tokenization_bert_japanese_test.py
run_examples_py3_torch:
working_directory: ~/transformers
docker:
- image: circleci/python:3.5
environment:
OMP_NUM_THREADS: 1
resource_class: xlarge
parallelism: 1
steps:
- checkout
- run: sudo pip install torch
- run: sudo pip install --progress-bar off .
- run: sudo pip install pytest codecov pytest-cov
- run: python -m pytest -sv ./transformers/tests/ --cov
- run: codecov
build_py2_tf:
working_directory: ~/transformers
resource_class: large
parallelism: 1
docker:
- image: circleci/python:2.7
steps:
- checkout
- 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
- run: codecov
- run: sudo pip install pytest pytest-xdist
- run: sudo pip install tensorboardX scikit-learn
- run: python -m pytest -n 8 --dist=loadfile -s -v ./examples/
deploy_doc:
working_directory: ~/transformers
docker:
@@ -82,6 +86,16 @@ jobs:
- run: sudo pip install --progress-bar off -r docs/requirements.txt
- run: sudo pip install --progress-bar off -r requirements.txt
- run: ./.circleci/deploy.sh
check_repository_consistency:
working_directory: ~/transformers
docker:
- image: circleci/python:3.5
resource_class: small
parallelism: 1
steps:
- checkout
- run: sudo pip install requests
- run: python ./utils/link_tester.py
workflow_filters: &workflow_filters
filters:
branches:
@@ -91,9 +105,10 @@ workflows:
version: 2
build_and_test:
jobs:
- build_py3_torch_and_tf
- build_py3_torch
- build_py3_tf
- build_py2_torch
- build_py2_tf
- check_repository_consistency
- run_examples_py3_torch
- run_tests_py3_custom_tokenizers
- run_tests_py3_torch_and_tf
- run_tests_py3_torch
- run_tests_py3_tf
- deploy_doc: *workflow_filters

2
CONTRIBUTING.md
View File

@@ -168,7 +168,7 @@ Follow these steps to start contributing:
to be merged;
4. Make sure pre-existing tests still pass;
5. Add high-coverage tests. No quality test, no merge;
6. All public methods must have informative doctrings;
6. All public methods must have informative docstrings;
### Style guide

84
README.md
View File

@@ -55,10 +55,12 @@ Choose the right framework for every part of a model's lifetime
| [Online demo](#online-demo) | Experimenting with this repo’s text generation capabilities |
| [Quick tour: Usage](#quick-tour) | Tokenizers & models usage: Bert and GPT-2 |
| [Quick tour: TF 2.0 and PyTorch ](#Quick-tour-TF-20-training-and-PyTorch-interoperability) | Train a TF 2.0 model in 10 lines of code, load it in PyTorch |
| [Quick tour: pipelines](#quick-tour-of-pipelines) | Using Pipelines: Wrapper around tokenizer and models to use finetuned models |
| [Quick tour: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
| [Quick tour: Share your models ](#Quick-tour-of-model-sharing) | Upload and share your fine-tuned models with the community |
| [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][(v2.2.0/v2.2.1)](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 |
| [Documentation][(v2.3.0)](https://huggingface.co/transformers/v2.3.0)[(v2.2.0/v2.2.1/v2.2.2)](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
@@ -144,7 +146,9 @@ At some point in the future, you'll be able to seamlessly move from pre-training
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. **[ALBERT](https://github.com/google-research/ALBERT)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
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.
12. **[T5](https://github.com/google-research/text-to-text-transfer-transformer)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
13. **[XLM-RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/xlmr)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
14. 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).
@@ -166,7 +170,7 @@ import torch
from transformers import *
# Transformers has a unified API
# for 8 transformer architectures and 30 pretrained weights.
# for 10 transformer architectures and 30 pretrained weights.
# Model | Tokenizer | Pretrained weights shortcut
MODELS = [(BertModel, BertTokenizer, 'bert-base-uncased'),
(OpenAIGPTModel, OpenAIGPTTokenizer, 'openai-gpt'),
@@ -176,7 +180,9 @@ MODELS = [(BertModel, BertTokenizer, 'bert-base-uncased'),
(XLNetModel, XLNetTokenizer, 'xlnet-base-cased'),
(XLMModel, XLMTokenizer, 'xlm-mlm-enfr-1024'),
(DistilBertModel, DistilBertTokenizer, 'distilbert-base-uncased'),
(RobertaModel, RobertaTokenizer, 'roberta-base')]
(RobertaModel, RobertaTokenizer, 'roberta-base'),
(XLMRobertaModel, XLMRobertaTokenizer, 'xlm-roberta-base'),
]
# To use TensorFlow 2.0 versions of the models, simply prefix the class names with 'TF', e.g. `TFRobertaModel` is the TF 2.0 counterpart of the PyTorch model `RobertaModel`
@@ -445,6 +451,76 @@ python ./examples/run_generation.py \
--repetition_penalty=1.2 \
```
## Quick tour of model sharing
New in `v2.2.2`: you can now upload and share your fine-tuned models with the community, using the <abbr title="Command-line interface">CLI</abbr> that's built-in to the library.
**First, create an account on [https://huggingface.co/join](https://huggingface.co/join)**. Then:
```shell
transformers-cli login
# log in using the same credentials as on huggingface.co
```
Upload your model:
```shell
transformers-cli upload ./path/to/pretrained_model/
# ^^ Upload folder containing weights/tokenizer/config
# saved via `.save_pretrained()`
transformers-cli upload ./config.json [--filename folder/foobar.json]
# ^^ Upload a single file
# (you can optionally override its filename, which can be nested inside a folder)
```
Your model will then be accessible through its identifier, a concatenation of your username and the folder name above:
```python
"username/model_name"
```
Anyone can load it from code:
```python
tokenizer = AutoTokenizer.from_pretrained("username/pretrained_model")
model = AutoModel.from_pretrained("username/pretrained_model")
```
Finally, list all your files on S3:
```shell
transformers-cli ls
# List all your S3 objects.
```
## Quick tour of pipelines
New in version `v2.3`: `Pipeline` are high-level objects which automatically handle tokenization, running your data through a transformers model
and outputting the result in a structured object.
You can create `Pipeline` objects for the following down-stream tasks:
- `feature-extraction`: Generates a tensor representation for the input sequence
- `ner`: Generates named entity mapping for each word in the input sequence.
- `sentiment-analysis`: Gives the polarity (positive / negative) of the whole input sequence.
- `question-answering`: Provided some context and a question refering to the context, it will extract the answer to the question
in the context.
```python
from transformers import pipeline
# Allocate a pipeline for sentiment-analysis
nlp = pipeline('sentiment-analysis')
nlp('We are very happy to include pipeline into the transformers repository.')
>>> {'label': 'POSITIVE', 'score': 0.99893874}
# Allocate a pipeline for question-answering
nlp = pipeline('question-answering')
nlp({
'question': 'What is the name of the repository ?',
'context': 'Pipeline have been included in the huggingface/transformers repository'
})
>>> {'score': 0.28756016668193496, 'start': 35, 'end': 59, 'answer': 'huggingface/transformers'}
```
## Migrating from pytorch-transformers to transformers
Here is a quick summary of what you should take care of when migrating from `pytorch-transformers` to `transformers`.

2
docs/source/conf.py
View File

@@ -26,7 +26,7 @@ author = u'huggingface'
# The short X.Y version
version = u''
# The full version, including alpha/beta/rc tags
release = u'2.2.1'
release = u'2.3.0'
# -- General configuration ---------------------------------------------------

2
docs/source/index.rst
View File

@@ -50,6 +50,7 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
9. `CTRL <https://github.com/pytorch/fairseq/tree/master/examples/ctrl>`_ (from Salesforce), released together with the paper `CTRL: A Conditional Transformer Language Model for Controllable Generation <https://www.github.com/salesforce/ctrl>`_ by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
10. `CamemBERT <https://huggingface.co/transformers/model_doc/camembert.html>`_ (from FAIR, Inria, Sorbonne Université) released together with the paper `CamemBERT: a Tasty French Language Model <https://arxiv.org/abs/1911.03894>`_ by Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suarez, Yoann Dupont, Laurent Romary, Eric Villemonte de la Clergerie, Djame Seddah, and Benoît Sagot.
11. `ALBERT <https://github.com/google-research/ALBERT>`_ (from Google Research), released together with the paper a `ALBERT: A Lite BERT for Self-supervised Learning of Language Representations <https://arxiv.org/abs/1909.11942>`_ by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
12. `XLM-RoBERTa <https://github.com/pytorch/fairseq/tree/master/examples/xlmr>`_ (from Facebook AI), released together with the paper `Unsupervised Cross-lingual Representation Learning at Scale <https://arxiv.org/abs/1911.02116>`_ by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
.. toctree::
:maxdepth: 2
@@ -58,6 +59,7 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
installation
quickstart
pretrained_models
model_sharing
examples
notebooks
serialization

40
docs/source/model_sharing.md Normal file
View File

@@ -0,0 +1,40 @@
# Model upload and sharing
Starting with `v2.2.2`, you can now upload and share your fine-tuned models with the community, using the <abbr title="Command-line interface">CLI</abbr> that's built-in to the library.
**First, create an account on [https://huggingface.co/join](https://huggingface.co/join)**. Then:
```shell
transformers-cli login
# log in using the same credentials as on huggingface.co
```
Upload your model:
```shell
transformers-cli upload ./path/to/pretrained_model/
# ^^ Upload folder containing weights/tokenizer/config
# saved via `.save_pretrained()`
transformers-cli upload ./config.json [--filename folder/foobar.json]
# ^^ Upload a single file
# (you can optionally override its filename, which can be nested inside a folder)
```
Your model will then be accessible through its identifier, a concatenation of your username and the folder name above:
```python
"username/pretrained_model"
```
Anyone can load it from code:
```python
tokenizer = AutoTokenizer.from_pretrained("username/pretrained_model")
model = AutoModel.from_pretrained("username/pretrained_model")
```
Finally, list all your files on S3:
```shell
transformers-cli ls
# List all your S3 objects.
```

56
docs/source/pretrained_models.rst
View File

@@ -3,6 +3,7 @@ Pretrained models
Here is the full list of the currently provided pretrained models together with a short presentation of each model.
For a list that includes community-uploaded models, refer to `https://huggingface.co/models <https://huggingface.co/models>`__.
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| Architecture | Shortcut name | Details of the model |
@@ -61,6 +62,32 @@ Here is the full list of the currently provided pretrained models together with
| | ``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>`__). |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``bert-base-japanese`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece. |
| | | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization. |
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``bert-base-japanese-whole-word-masking`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized with MeCab and WordPiece. |
| | | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization. |
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``bert-base-japanese-char`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on Japanese text. Text is tokenized into characters. |
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``bert-base-japanese-char-whole-word-masking`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized into characters. |
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``bert-base-finnish-cased-v1`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on cased Finnish text. |
| | | (see `details on turkunlp.org <http://turkunlp.org/FinBERT/>`__). |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``bert-base-finnish-uncased-v1`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on uncased Finnish text. |
| | | (see `details on turkunlp.org <http://turkunlp.org/FinBERT/>`__). |
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| GPT | ``openai-gpt`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | OpenAI GPT English model |
@@ -128,6 +155,10 @@ Here is the full list of the currently provided pretrained models together with
| | | | ``roberta-large`` fine-tuned on `MNLI <http://www.nyu.edu/projects/bowman/multinli/>`__. |
| | | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``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>`__) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``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>`__) |
@@ -148,10 +179,6 @@ Here is the full list of the currently provided pretrained models together with
| | | | 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>`__) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``distilbert-base-german-cased`` | | 6-layer, 768-hidden, 12-heads, 66M parameters |
| | | | The German DistilBERT model distilled from the German DBMDZ BERT model `bert-base-german-dbmdz-cased` checkpoint. |
| | | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__) |
@@ -199,6 +226,27 @@ Here is the full list of the currently provided pretrained models together with
| | | | ALBERT xxlarge model with no dropout, additional training data and longer training |
| | | (see `details <https://github.com/google-research/ALBERT>`__) |
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| T5 | ``t5-small`` | | ~60M parameters with 6-layers, 512-hidden-state, 2048 feed-forward hidden-state, 8-heads, |
| | | | Trained on English text: the Colossal Clean Crawled Corpus (C4) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``t5-base`` | | ~220M parameters with 12-layers, 768-hidden-state, 3072 feed-forward hidden-state, 12-heads, |
| | | | Trained on English text: the Colossal Clean Crawled Corpus (C4) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``t5-large`` | | ~770M parameters with 24-layers, 1024-hidden-state, 4096 feed-forward hidden-state, 16-heads, |
| | | | Trained on English text: the Colossal Clean Crawled Corpus (C4) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``t5-3B`` | | ~2.8B parameters with 24-layers, 1024-hidden-state, 16384 feed-forward hidden-state, 32-heads, |
| | | | Trained on English text: the Colossal Clean Crawled Corpus (C4) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``t5-11B`` | | ~11B parameters with 24-layers, 1024-hidden-state, 65536 feed-forward hidden-state, 128-heads, |
| | | | Trained on English text: the Colossal Clean Crawled Corpus (C4) |
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| XLM-RoBERTa | ``xlm-roberta-base`` | | ~125M parameters with 12-layers, 768-hidden-state, 3072 feed-forward hidden-state, 8-heads, |
| | | | Trained on on 2.5 TB of newly created clean CommonCrawl data in 100 languages |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``xlm-roberta-large`` | | ~355M parameters with 24-layers, 1027-hidden-state, 4096 feed-forward hidden-state, 16-heads, |
| | | | Trained on 2.5 TB of newly created clean CommonCrawl data in 100 languages |
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
.. <https://huggingface.co/transformers/examples.html>`__

93
docs/source/quickstart.md
View File

@@ -220,3 +220,96 @@ print(sequence)
```
The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.
### Model2Model example
Encoder-decoder architectures require two tokenized inputs: one for the encoder and the other one for the decoder. Let's assume that we want to use `Model2Model` for generative question answering, and start by tokenizing the question and answer that will be fed to the model.
```python
import torch
from transformers import BertTokenizer, Model2Model
# OPTIONAL: if you want to have more information on what's happening under the hood, activate the logger as follows
import logging
logging.basicConfig(level=logging.INFO)
# Load pre-trained model tokenizer (vocabulary)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
# Encode the input to the encoder (the question)
question = "Who was Jim Henson?"
encoded_question = tokenizer.encode(question)
# Encode the input to the decoder (the answer)
answer = "Jim Henson was a puppeteer"
encoded_answer = tokenizer.encode(answer)
# Convert inputs to PyTorch tensors
question_tensor = torch.tensor([encoded_question])
answer_tensor = torch.tensor([encoded_answer])
```
Let's see how we can use `Model2Model` to get the value of the loss associated with this (question, answer) pair:
```python
# In order to compute the loss we need to provide language model
# labels (the token ids that the model should have produced) to
# the decoder.
lm_labels = encoded_answer
labels_tensor = torch.tensor([lm_labels])
# Load pre-trained model (weights)
model = Model2Model.from_pretrained('bert-base-uncased')
# Set the model in evaluation mode to deactivate the DropOut modules
# This is IMPORTANT to have reproducible results during evaluation!
model.eval()
# If you have a GPU, put everything on cuda
question_tensor = question_tensor.to('cuda')
answer_tensor = answer_tensor.to('cuda')
labels_tensor = labels_tensor.to('cuda')
model.to('cuda')
# Predict hidden states features for each layer
with torch.no_grad():
# See the models docstrings for the detail of the inputs
outputs = model(question_tensor, answer_tensor, decoder_lm_labels=labels_tensor)
# Transformers models always output tuples.
# See the models docstrings for the detail of all the outputs
# In our case, the first element is the value of the LM loss
lm_loss = outputs[0]
```
This loss can be used to fine-tune `Model2Model` on the question answering task. Assuming that we fine-tuned the model, let us now see how to generate an answer:
```python
# Let's re-use the previous question
question = "Who was Jim Henson?"
encoded_question = tokenizer.encode(question)
question_tensor = torch.tensor([encoded_question])
# This time we try to generate the answer, so we start with an empty sequence
answer = "[CLS]"
encoded_answer = tokenizer.encode(answer, add_special_tokens=False)
answer_tensor = torch.tensor([encoded_answer])
# Load pre-trained model (weights)
model = Model2Model.from_pretrained('fine-tuned-weights')
model.eval()
# If you have a GPU, put everything on cuda
question_tensor = encoded_question.to('cuda')
answer_tensor = encoded_answer.to('cuda')
model.to('cuda')
# Predict all tokens
with torch.no_grad():
outputs = model(question_tensor, answer_tensor)
predictions = outputs[0]
# confirm we were able to predict 'jim'
predicted_index = torch.argmax(predictions[0, -1]).item()
predicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]
assert predicted_token == 'jim'
```

63
examples/README.md
View File

@@ -357,9 +357,9 @@ eval_loss = 0.44457291918821606
Based on the script [`run_squad.py`](https://github.com/huggingface/transformers/blob/master/examples/run_squad.py).
#### Fine-tuning on SQuAD
#### Fine-tuning BERT on SQuAD1.0
This example code fine-tunes BERT on the SQuAD dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large)
This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large)
on a single tesla V100 16GB. The data for SQuAD can be downloaded with the following links and should be saved in a
$SQUAD_DIR directory.
@@ -367,6 +367,12 @@ $SQUAD_DIR directory.
* [dev-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json)
* [evaluate-v1.1.py](https://github.com/allenai/bi-att-flow/blob/master/squad/evaluate-v1.1.py)
And for SQuAD2.0, you need to download:
- [train-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json)
- [dev-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json)
- [evaluate-v2.0.py](https://worksheets.codalab.org/rest/bundles/0x6b567e1cf2e041ec80d7098f031c5c9e/contents/blob/)
```bash
export SQUAD_DIR=/path/to/SQUAD
@@ -396,7 +402,7 @@ exact_match = 81.22
#### Distributed training
Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD:
Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD1.0:
```bash
python -m torch.distributed.launch --nproc_per_node=8 run_squad.py \
@@ -428,7 +434,9 @@ This fine-tuned model is available as a checkpoint under the reference
#### Fine-tuning XLNet on SQuAD
This example code fine-tunes XLNet on the SQuAD dataset. See above to download the data for SQuAD .
This example code fine-tunes XLNet on both SQuAD1.0 and SQuAD2.0 dataset. See above to download the data for SQuAD .
##### Command for SQuAD1.0:
```bash
export SQUAD_DIR=/path/to/SQUAD
@@ -451,7 +459,32 @@ python /data/home/hlu/transformers/examples/run_squad.py \
--save_steps 5000
```
Training with the previously defined hyper-parameters yields the following results:
##### Command for SQuAD2.0:
```bash
export SQUAD_DIR=/path/to/SQUAD
python run_squad.py \
--model_type xlnet \
--model_name_or_path xlnet-large-cased \
--do_train \
--do_eval \
--version_2_with_negative \
--train_file $SQUAD_DIR/train-v2.0.json \
--predict_file $SQUAD_DIR/dev-v2.0.json \
--learning_rate 3e-5 \
--num_train_epochs 4 \
--max_seq_length 384 \
--doc_stride 128 \
--output_dir ./wwm_cased_finetuned_squad/ \
--per_gpu_eval_batch_size=2 \
--per_gpu_train_batch_size=2 \
--save_steps 5000
```
Larger batch size may improve the performance while costing more memory.
##### Results for SQuAD1.0 with the previously defined hyper-parameters:
```python
{
@@ -464,10 +497,28 @@ Training with the previously defined hyper-parameters yields the following resul
}
```
##### Results for SQuAD2.0 with the previously defined hyper-parameters:
```python
{
"exact": 80.4177545691906,
"f1": 84.07154997729623,
"total": 11873,
"HasAns_exact": 76.73751686909581,
"HasAns_f1": 84.05558584352873,
"HasAns_total": 5928,
"NoAns_exact": 84.0874684608915,
"NoAns_f1": 84.0874684608915,
"NoAns_total": 5945
}
```
## Named Entity Recognition
Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py) for Pytorch and
[`run_tf_ner.py`(https://github.com/huggingface/transformers/blob/master/examples/run_tf_ner.py)] for Tensorflow 2.
[`run_tf_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_tf_ner.py) for Tensorflow 2.
This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
Details and results for the fine-tuning provided by @stefan-it.

290
examples/run_generation.py
View File

@@ -20,14 +20,10 @@ from __future__ import absolute_import, division, print_function, unicode_litera
import argparse
import logging
from tqdm import trange
import torch
import torch.nn.functional as F
import numpy as np
from transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig
from transformers import GPT2LMHeadModel, GPT2Tokenizer
from transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer
from transformers import XLNetLMHeadModel, XLNetTokenizer
@@ -36,22 +32,22 @@ from transformers import CTRLLMHeadModel, CTRLTokenizer
from transformers import XLMWithLMHeadModel, XLMTokenizer
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt = '%m/%d/%Y %H:%M:%S',
level = logging.INFO)
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger = logging.getLogger(__name__)
MAX_LENGTH = int(10000) # Hardcoded max length to avoid infinite loop
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig)), ())
MODEL_CLASSES = {
'gpt2': (GPT2LMHeadModel, GPT2Tokenizer),
'ctrl': (CTRLLMHeadModel, CTRLTokenizer),
'openai-gpt': (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
'xlnet': (XLNetLMHeadModel, XLNetTokenizer),
'transfo-xl': (TransfoXLLMHeadModel, TransfoXLTokenizer),
'xlm': (XLMWithLMHeadModel, XLMTokenizer),
"gpt2": (GPT2LMHeadModel, GPT2Tokenizer),
"ctrl": (CTRLLMHeadModel, CTRLTokenizer),
"openai-gpt": (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
"xlnet": (XLNetLMHeadModel, XLNetTokenizer),
"transfo-xl": (TransfoXLLMHeadModel, TransfoXLTokenizer),
"xlm": (XLMWithLMHeadModel, XLMTokenizer),
}
# Padding text to help Transformer-XL and XLNet with short prompts as proposed by Aman Rusia
@@ -75,81 +71,79 @@ def set_seed(args):
if args.n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
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 (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
"""
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
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
logits[indices_to_remove] = filter_value
if top_p > 0.0:
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
# Remove tokens with cumulative probability above the threshold
sorted_indices_to_remove = cumulative_probs > top_p
# Shift the indices to the right to keep also the first token above the threshold
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
# 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
#
# Functions to prepare models' input
#
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
with torch.no_grad():
for _ in trange(length):
def prepare_ctrl_input(args, _, tokenizer, prompt_text):
if args.temperature > 0.7:
logger.info(
"CTRL typically works better with lower temperatures (and lower top_k)."
)
inputs = {'input_ids': generated}
if is_xlnet:
# XLNet is a direct (predict same token, not next token) and bi-directional model by default
# => need one additional dummy token in the input (will be masked), attention mask and target mapping (see model docstring)
input_ids = torch.cat((generated, torch.zeros((1, 1), dtype=torch.long, device=device)), dim=1)
perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float, device=device)
perm_mask[:, :, -1] = 1.0 # Previous tokens don't see last token
target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float, device=device)
target_mapping[0, 0, -1] = 1.0 # predict last token
inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping}
encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False)
if not any(encoded_prompt[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"
)
return prompt_text
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)
def prepare_xlm_input(args, model, tokenizer, prompt_text):
# kwargs = {"language": None, "mask_token_id": None}
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.)
# 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, dim=-1).unsqueeze(-1)
# Set the language
use_lang_emb = hasattr(model.config, "use_lang_emb") and model.config.use_lang_emb
if hasattr(model.config, "lang2id") and use_lang_emb:
available_languages = model.config.lang2id.keys()
if args.xlm_language in available_languages:
language = args.xlm_language
else:
next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
generated = torch.cat((generated, next_token), dim=1)
return generated
language = None
while language not in available_languages:
language = input(
"Using XLM. Select language in "
+ str(list(available_languages))
+ " >>> "
)
# kwargs["language"] = tokenizer.lang2id[language]
# TODO fix mask_token_id setup when configurations will be synchronized between models and tokenizers
# XLM masked-language modeling (MLM) models need masked token
# is_xlm_mlm = "mlm" in args.model_name_or_path
# if is_xlm_mlm:
# kwargs["mask_token_id"] = tokenizer.mask_token_id
return prompt_text
def prepare_xlnet_input(args, _, tokenizer, prompt_text):
prompt_text = (args.padding_text if args.padding_text else PADDING_TEXT) + prompt_text
return prompt_text, {}
def prepare_transfoxl_input(args, _, tokenizer, prompt_text):
prompt_text = (args.padding_text if args.padding_text else PADDING_TEXT) + prompt_text
return prompt_text, {}
PREPROCESSING_FUNCTIONS = {
"ctrl": prepare_ctrl_input,
"xlm": prepare_xlm_input,
"xlnet": prepare_xlnet_input,
"transfo-xl": prepare_transfoxl_input,
}
def adjust_length_to_model(length, max_sequence_length):
if length < 0 and max_sequence_length > 0:
length = max_sequence_length
elif 0 < max_sequence_length < length:
length = max_sequence_length # No generation bigger than model size
elif length < 0:
length = MAX_LENGTH # avoid infinite loop
return length
def main():
@@ -157,104 +151,76 @@ def main():
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))
help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
parser.add_argument("--prompt", type=str, default="")
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,
help="primarily useful for CTRL model; in that case, use 1.2")
parser.add_argument("--top_k", type=int, default=0)
parser.add_argument("--top_p", type=float, default=0.9)
parser.add_argument("--no_cuda", action='store_true',
help="Avoid using CUDA when available")
parser.add_argument('--seed', type=int, default=42,
help="random seed for initialization")
parser.add_argument('--stop_token', type=str, default=None,
help="Token at which text generation is stopped")
parser.add_argument("--stop_token", type=str, default=None, help="Token at which text generation is stopped")
parser.add_argument("--temperature", type=float, default=1.0, help="temperature of 1.0 has no effect, lower tend toward greedy sampling")
parser.add_argument("--repetition_penalty", type=float, default=1.0, help="primarily useful for CTRL model; in that case, use 1.2")
parser.add_argument("--k", type=int, default=0)
parser.add_argument("--p", type=float, default=0.9)
parser.add_argument("--padding_text", type=str, default="", help="Padding text for Transfo-XL and XLNet.")
parser.add_argument("--xlm_language", type=str, default="", help="Optional language when used with the XLM model.")
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
args = parser.parse_args()
args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu"
)
args.n_gpu = torch.cuda.device_count()
set_seed(args)
# Initialize the model and tokenizer
try:
args.model_type = args.model_type.lower()
model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
except KeyError:
raise KeyError(
"the model {} you specified is not supported. You are welcome to add it and open a PR :)"
)
tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
model = model_class.from_pretrained(args.model_name_or_path)
model.to(args.device)
model.eval()
if args.length < 0 and model.config.max_position_embeddings > 0:
args.length = model.config.max_position_embeddings
elif 0 < model.config.max_position_embeddings < args.length:
args.length = model.config.max_position_embeddings # No generation bigger than model size
elif args.length < 0:
args.length = MAX_LENGTH # avoid infinite loop
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
if args.model_type in ["xlm"] and hasattr(tokenizer, 'lang2id') and hasattr(model.config, 'use_lang_emb') \
and model.config.use_lang_emb:
if args.xlm_lang:
language = args.xlm_lang
else:
language = None
while language not in tokenizer.lang2id.keys():
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, 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,
args.length = adjust_length_to_model(
args.length, max_sequence_length=model.config.max_position_embeddings
)
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]
logger.info(args)
prompt_text = args.prompt if args.prompt else input("Model prompt >>> ")
# Different models need different input formatting and/or extra arguments
requires_preprocessing = args.model_type in PREPROCESSING_FUNCTIONS.keys()
if requires_preprocessing:
prepare_input = PREPROCESSING_FUNCTIONS.get(args.model_type)
prompt_text = prepare_input(args, model, tokenizer, prompt_text)
encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False, return_tensors='pt')
output_sequences = model.generate(
input_ids=encoded_prompt,
max_length=args.length,
temperature=args.temperature,
top_k=args.k,
top_p=args.p,
repetition_penalty=args.repetition_penalty,
)
# Batch size == 1. to add more examples please use num_return_sequences > 1
generated_sequence = output_sequences[0].tolist()
text = tokenizer.decode(generated_sequence, clean_up_tokenization_spaces=True)
text = text[: t.find(args.stop_token) if args.stop_token else None]
print(text)
if args.prompt:
break
return text
if __name__ == '__main__':
if __name__ == "__main__":
main()

10
examples/run_glue.py
View File

@@ -52,6 +52,9 @@ from transformers import (WEIGHTS_NAME, BertConfig,
AlbertConfig,
AlbertForSequenceClassification,
AlbertTokenizer,
XLMRobertaConfig,
XLMRobertaForSequenceClassification,
XLMRobertaTokenizer,
)
from transformers import AdamW, get_linear_schedule_with_warmup
@@ -72,7 +75,8 @@ MODEL_CLASSES = {
'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer),
'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer)
'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer),
'xlmroberta': (XLMRobertaConfig, XLMRobertaForSequenceClassification, XLMRobertaTokenizer),
}
@@ -337,7 +341,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
else:
logger.info("Creating features from dataset file at %s", args.data_dir)
label_list = processor.get_labels()
if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta']:
if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta', 'xlmroberta']:
# HACK(label indices are swapped in RoBERTa pretrained model)
label_list[1], label_list[2] = label_list[2], label_list[1]
examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
@@ -413,7 +417,7 @@ def main():
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 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,

2
examples/run_lm_finetuning.py
View File

@@ -430,7 +430,7 @@ def main():
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 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,

5
examples/run_ner.py
View File

@@ -38,11 +38,13 @@ from transformers import WEIGHTS_NAME, BertConfig, BertForTokenClassification, B
from transformers import RobertaConfig, RobertaForTokenClassification, RobertaTokenizer
from transformers import DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer
from transformers import CamembertConfig, CamembertForTokenClassification, CamembertTokenizer
from transformers import XLMRobertaConfig, XLMRobertaForTokenClassification, XLMRobertaTokenizer
logger = logging.getLogger(__name__)
ALL_MODELS = sum(
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)),
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig,
CamembertConfig, XLMRobertaConfig)),
())
MODEL_CLASSES = {
@@ -50,6 +52,7 @@ MODEL_CLASSES = {
"roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer),
"distilbert": (DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer),
"camembert": (CamembertConfig, CamembertForTokenClassification, CamembertTokenizer),
"xlmroberta": (XLMRobertaConfig, XLMRobertaForTokenClassification, XLMRobertaTokenizer),
}

42
examples/run_squad.py
View File

@@ -62,7 +62,6 @@ MODEL_CLASSES = {
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer),
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer)
}
@@ -200,8 +199,10 @@ def train(args, train_dataset, model, tokenizer):
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]})
inputs.update({'cls_index': batch[5],
'p_mask': batch[6]})
if args.version_2_with_negative:
inputs.update({'is_impossible': batch[7]})
outputs = model(**inputs)
# model outputs are always tuple in transformers (see doc)
loss = outputs[0]
@@ -296,7 +297,7 @@ def evaluate(args, model, tokenizer, prefix=""):
dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
# multi-gpu evaluate
if args.n_gpu > 1:
if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
model = torch.nn.DataParallel(model)
# Eval!
@@ -420,7 +421,7 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
else:
logger.info("Creating features from dataset file at %s", input_dir)
if not args.data_dir:
if not args.data_dir and ((evaluate and not args.predict_file) or (not evaluate and not args.train_file)):
try:
import tensorflow_datasets as tfds
except ImportError:
@@ -436,8 +437,10 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
tfds_examples, evaluate=evaluate)
else:
processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
examples = processor.get_dev_examples(
args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
if evaluate:
examples = processor.get_dev_examples(args.data_dir, filename=args.predict_file)
else:
examples = processor.get_train_examples(args.data_dir, filename=args.train_file)
features, dataset = squad_convert_examples_to_features(
examples=examples,
@@ -477,7 +480,14 @@ def main():
# Other parameters
parser.add_argument("--data_dir", default=None, type=str,
help="The input data dir. Should contain the .json files for the task. If not specified, will run with tensorflow_datasets.")
help="The input data dir. Should contain the .json files for the task." +
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
parser.add_argument("--train_file", default=None, type=str,
help="The input training file. If a data dir is specified, will look for the file there" +
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
parser.add_argument("--predict_file", default=None, type=str,
help="The input evaluation file. If a data dir is specified, will look for the file there" +
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
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,
@@ -564,11 +574,6 @@ def main():
help="Can be used for distant debugging.")
args = parser.parse_args()
args.predict_file = os.path.join(args.output_dir, 'predictions_{}_{}.txt'.format(
list(filter(None, args.model_name_or_path.split('/'))).pop(),
str(args.max_seq_length))
)
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))
@@ -676,12 +681,15 @@ def main():
# 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]:
if args.do_train:
logger.info("Loading checkpoints saved during training for evaluation")
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
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
else:
logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
checkpoints = [args.model_name_or_path]
logger.info("Evaluate the following checkpoints: %s", checkpoints)

6
examples/summarization/README.md
View File

@@ -29,7 +29,7 @@ And move all the stories to the same folder. We will refer as `$DATA_PATH` the p
python run_summarization.py \
--documents_dir $DATA_PATH \
--summaries_output_dir $SUMMARIES_PATH \ # optional
--to_cpu false \
--no_cuda false \
--batch_size 4 \
--min_length 50 \
--max_length 200 \
@@ -39,7 +39,7 @@ python run_summarization.py \
--compute_rouge true
```
The scripts executes on GPU if one is available and if `to_cpu` is not set to `true`. Inference on multiple GPUs is not suported yet. The ROUGE scores will be displayed in the console at the end of evaluation and written in a `rouge_scores.txt` file. The script takes 30 hours to compute with a single Tesla V100 GPU and a batch size of 10 (300,000 texts to summarize).
The scripts executes on GPU if one is available and if `no_cuda` is not set to `true`. Inference on multiple GPUs is not suported yet. The ROUGE scores will be displayed in the console at the end of evaluation and written in a `rouge_scores.txt` file. The script takes 30 hours to compute with a single Tesla V100 GPU and a batch size of 10 (300,000 texts to summarize).
## Summarize any text
@@ -49,7 +49,7 @@ Put the documents that you would like to summarize in a folder (the path to whic
python run_summarization.py \
--documents_dir $DATA_PATH \
--summaries_output_dir $SUMMARIES_PATH \ # optional
--to_cpu false \
--no_cuda false \
--batch_size 4 \
--min_length 50 \
--max_length 200 \

28
examples/summarization/configuration_bertabs.py
View File

@@ -33,6 +33,8 @@ class BertAbsConfig(PretrainedConfig):
r""" Class to store the configuration of the BertAbs model.
Arguments:
vocab_size: int
Number of tokens in the vocabulary.
max_pos: int
The maximum sequence length that this model will be used with.
enc_layer: int
@@ -65,7 +67,7 @@ class BertAbsConfig(PretrainedConfig):
def __init__(
self,
vocab_size_or_config_json_file=30522,
vocab_size=30522,
max_pos=512,
enc_layers=6,
enc_hidden_size=512,
@@ -81,14 +83,7 @@ class BertAbsConfig(PretrainedConfig):
):
super(BertAbsConfig, self).__init__(**kwargs)
if self._input_is_path_to_json(vocab_size_or_config_json_file):
path_to_json = vocab_size_or_config_json_file
with open(path_to_json, "r", encoding="utf-8") as reader:
json_config = json.loads(reader.read())
for key, value in json_config.items():
self.__dict__[key] = value
elif isinstance(vocab_size_or_config_json_file, int):
self.vocab_size = vocab_size_or_config_json_file
self.vocab_size = vocab_size
self.max_pos = max_pos
self.enc_layers = enc_layers
@@ -102,18 +97,3 @@ class BertAbsConfig(PretrainedConfig):
self.dec_heads = dec_heads
self.dec_ff_size = dec_ff_size
self.dec_dropout = dec_dropout
else:
raise ValueError(
"First argument must be either a vocabulary size (int)"
"or the path to a pretrained model config file (str)"
)
def _input_is_path_to_json(self, first_argument):
""" Checks whether the first argument passed to config
is the path to a JSON file that contains the config.
"""
is_python_2 = sys.version_info[0] == 2
if is_python_2:
return isinstance(first_argument, unicode)
else:
return isinstance(first_argument, str)

2
requirements.txt
View File

@@ -5,7 +5,7 @@ boto3
# Used for downloading models over HTTP
requests
# For OpenAI GPT
regex
regex != 2019.12.17
# For XLNet
sentencepiece
# For XLM

14
setup.py
View File

@@ -38,13 +38,15 @@ from setuptools import find_packages, setup
extras = {
'serving': ['uvicorn', 'fastapi']
'serving': ['pydantic', 'uvicorn', 'fastapi'],
'serving-tf': ['pydantic', 'uvicorn', 'fastapi', 'tensorflow'],
'serving-torch': ['pydantic', 'uvicorn', 'fastapi', 'torch']
}
extras['all'] = [package for package in extras.values()]
setup(
name="transformers",
version="2.2.1",
version="2.3.0",
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",
@@ -57,16 +59,12 @@ setup(
"tests.*", "tests"]),
install_requires=['numpy',
'boto3',
'filelock',
'requests',
'tqdm',
'regex',
'regex != 2019.12.17',
'sentencepiece',
'sacremoses'],
entry_points={
'console_scripts': [
"transformers=transformers.__main__:main",
]
},
extras_require=extras,
scripts=[
'transformers-cli'

20
templates/adding_a_new_model/configuration_xxx.py
View File

@@ -39,7 +39,7 @@ class XxxConfig(PretrainedConfig):
Arguments:
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `XxxModel`.
vocab_size: 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
@@ -64,7 +64,7 @@ class XxxConfig(PretrainedConfig):
pretrained_config_archive_map = XXX_PRETRAINED_CONFIG_ARCHIVE_MAP
def __init__(self,
vocab_size_or_config_json_file=50257,
vocab_size=50257,
n_positions=1024,
n_ctx=1024,
n_embd=768,
@@ -75,8 +75,6 @@ class XxxConfig(PretrainedConfig):
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,
@@ -84,7 +82,7 @@ class XxxConfig(PretrainedConfig):
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.vocab_size = vocab_size
self.n_ctx = n_ctx
self.n_positions = n_positions
self.n_embd = n_embd
@@ -95,23 +93,11 @@ class XxxConfig(PretrainedConfig):
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):

10
templates/adding_a_new_model/convert_xxx_original_tf_checkpoint_to_pytorch.py
View File

@@ -26,9 +26,9 @@ 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):
def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file, pytorch_dump_path):
# Initialise PyTorch model
config = XxxConfig.from_json_file(xxx_config_file)
config = XxxConfig.from_json_file(config_file)
print("Building PyTorch model from configuration: {}".format(str(config)))
model = XxxForPreTraining(config)
@@ -48,11 +48,11 @@ if __name__ == "__main__":
type = str,
required = True,
help = "Path to the TensorFlow checkpoint path.")
parser.add_argument("--xxx_config_file",
parser.add_argument("--config_file",
default = None,
type = str,
required = True,
help = "The config json file corresponding to the pre-trained XXX model. \n"
help = "The config json file corresponding to the pre-trained model. \n"
"This specifies the model architecture.")
parser.add_argument("--pytorch_dump_path",
default = None,
@@ -61,5 +61,5 @@ if __name__ == "__main__":
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.config_file,
args.pytorch_dump_path)

2
templates/adding_a_new_model/modeling_tf_xxx.py
View File

@@ -26,6 +26,8 @@ import logging
import math
import os
import sys
import copy
import itertools
from io import open
import numpy as np

2
templates/adding_a_new_model/modeling_xxx.py
View File

@@ -25,6 +25,8 @@ import logging
import math
import os
import sys
import copy
import itertools
from io import open
import torch

9
templates/adding_a_new_model/tests/modeling_tf_xxx_test.py
View File

@@ -17,12 +17,11 @@ from __future__ import division
from __future__ import print_function
import unittest
import shutil
import sys
from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
from .configuration_common_test import ConfigTester
from .utils import require_tf, slow
from .utils import CACHE_DIR, require_tf, slow
from transformers import XxxConfig, is_tf_available
@@ -111,7 +110,7 @@ class TFXxxModelTest(TFCommonTestCases.TFCommonModelTester):
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = XxxConfig(
vocab_size_or_config_json_file=self.vocab_size,
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
@@ -245,10 +244,8 @@ class TFXxxModelTest(TFCommonTestCases.TFCommonModelTester):
@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)
model = TFXxxModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model)
if __name__ == "__main__":

9
templates/adding_a_new_model/tests/modeling_xxx_test.py
View File

@@ -17,13 +17,12 @@ from __future__ import division
from __future__ import print_function
import unittest
import shutil
from transformers import is_torch_available
from .modeling_common_test import (CommonTestCases, ids_tensor)
from .configuration_common_test import ConfigTester
from .utils import require_torch, slow, torch_device
from .utils import CACHE_DIR, require_torch, slow, torch_device
if is_torch_available():
from transformers import (XxxConfig, XxxModel, XxxForMaskedLM,
@@ -109,7 +108,7 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = XxxConfig(
vocab_size_or_config_json_file=self.vocab_size,
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
@@ -249,10 +248,8 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
@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)
model = XxxModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model)
if __name__ == "__main__":

2
templates/adding_a_new_model/tokenization_xxx.py
View File

@@ -85,7 +85,7 @@ class XxxTokenizer(PreTrainedTokenizer):
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
do_lower_case: Whether to lower case the input. Only has an effect when do_basic_tokenize=True
"""
vocab_files_names = VOCAB_FILES_NAMES

11
transformers-cli Normal file → Executable file
View File

@@ -1,14 +1,21 @@
#!/usr/bin/env python
from argparse import ArgumentParser
from transformers.commands.download import DownloadCommand
from transformers.commands.run import RunCommand
from transformers.commands.user import UserCommands
from transformers.commands.convert import ConvertCommand
from transformers.commands.serving import ServeCommand
if __name__ == '__main__':
parser = ArgumentParser(description='Transformers CLI tool', usage='transformers-cli <command> [<args>]')
parser = ArgumentParser('Transformers CLI tool', usage='transformers-cli <command> [<args>]')
commands_parser = parser.add_subparsers(help='transformers-cli command helpers')
# Register commands
ConvertCommand.register_subcommand(commands_parser)
DownloadCommand.register_subcommand(commands_parser)
RunCommand.register_subcommand(commands_parser)
ServeCommand.register_subcommand(commands_parser)
UserCommands.register_subcommand(commands_parser)
# Let's go

34
transformers/__init__.py Normal file → Executable file
View File

@@ -1,4 +1,4 @@
__version__ = "2.2.1"
__version__ = "2.3.0"
# Work around to update TensorFlow's absl.logging threshold which alters the
# default Python logging output behavior when present.
@@ -19,11 +19,12 @@ logger = logging.getLogger(__name__) # pylint: disable=invalid-name
# Files and general utilities
from .file_utils import (TRANSFORMERS_CACHE, PYTORCH_TRANSFORMERS_CACHE, PYTORCH_PRETRAINED_BERT_CACHE,
cached_path, add_start_docstrings, add_end_docstrings,
WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, CONFIG_NAME,
WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, CONFIG_NAME, MODEL_CARD_NAME,
is_tf_available, is_torch_available)
from .data import (is_sklearn_available,
InputExample, InputFeatures, DataProcessor,
SingleSentenceClassificationProcessor,
glue_output_modes, glue_convert_examples_to_features,
glue_processors, glue_tasks_num_labels,
xnli_output_modes, xnli_processors, xnli_tasks_num_labels,
@@ -33,10 +34,14 @@ from .data import (is_sklearn_available,
if is_sklearn_available():
from .data import glue_compute_metrics, xnli_compute_metrics
# Model Cards
from .modelcard import ModelCard
# Tokenizers
from .tokenization_utils import (PreTrainedTokenizer)
from .tokenization_auto import AutoTokenizer
from .tokenization_bert import BertTokenizer, BasicTokenizer, WordpieceTokenizer
from .tokenization_bert_japanese import BertJapaneseTokenizer, MecabTokenizer, CharacterTokenizer
from .tokenization_openai import OpenAIGPTTokenizer
from .tokenization_transfo_xl import (TransfoXLTokenizer, TransfoXLCorpus)
from .tokenization_gpt2 import GPT2Tokenizer
@@ -47,28 +52,31 @@ from .tokenization_roberta import RobertaTokenizer
from .tokenization_distilbert import DistilBertTokenizer
from .tokenization_albert import AlbertTokenizer
from .tokenization_camembert import CamembertTokenizer
from .tokenization_t5 import T5Tokenizer
from .tokenization_xlm_roberta import XLMRobertaTokenizer
# Configurations
from .configuration_utils import PretrainedConfig
from .configuration_auto import AutoConfig
from .configuration_auto import AutoConfig, ALL_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_bert import BertConfig, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_openai import OpenAIGPTConfig, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_transfo_xl import TransfoXLConfig, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_gpt2 import GPT2Config, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_ctrl import CTRLConfig, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_xlnet import XLNetConfig, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP
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
from .configuration_t5 import T5Config, T5_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_xlm_roberta import XLMRobertaConfig, XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
# Modeling
if is_torch_available():
from .modeling_utils import (PreTrainedModel, prune_layer, Conv1D)
from .modeling_auto import (AutoModel, AutoModelForSequenceClassification, AutoModelForQuestionAnswering,
AutoModelWithLMHead)
AutoModelWithLMHead, AutoModelForTokenClassification, ALL_PRETRAINED_MODEL_ARCHIVE_MAP)
from .modeling_bert import (BertPreTrainedModel, BertModel, BertForPreTraining,
BertForMaskedLM, BertForNextSentencePrediction,
@@ -109,11 +117,17 @@ if is_torch_available():
CamembertForTokenClassification,
CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
from .modeling_encoder_decoder import PreTrainedEncoderDecoder, Model2Model
from .modeling_t5 import (T5PreTrainedModel, T5Model, T5WithLMHeadModel,
load_tf_weights_in_t5,
T5_PRETRAINED_MODEL_ARCHIVE_MAP)
from .modeling_albert import (AlbertPreTrainedModel, AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification,
AlbertForQuestionAnswering,
load_tf_weights_in_albert, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
from .modeling_xlm_roberta import (XLMRobertaForMaskedLM, XLMRobertaModel, XLMRobertaForMultipleChoice,
XLMRobertaForSequenceClassification, XLMRobertaForTokenClassification)
# Optimization
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)
@@ -123,7 +137,7 @@ if is_torch_available():
if is_tf_available():
from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list
from .modeling_tf_auto import (TFAutoModel, TFAutoModelForSequenceClassification, TFAutoModelForQuestionAnswering,
TFAutoModelWithLMHead)
TFAutoModelWithLMHead, TFAutoModelForTokenClassification, TF_ALL_PRETRAINED_MODEL_ARCHIVE_MAP)
from .modeling_tf_bert import (TFBertPreTrainedModel, TFBertMainLayer, TFBertEmbeddings,
TFBertModel, TFBertForPreTraining,
@@ -177,6 +191,10 @@ if is_tf_available():
from .modeling_tf_albert import (TFAlbertPreTrainedModel, TFAlbertModel, TFAlbertForMaskedLM,
TFAlbertForSequenceClassification,
TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
from .modeling_tf_t5 import (TFT5PreTrainedModel, TFT5Model, TFT5WithLMHeadModel,
TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP)
# Optimization
from .optimization_tf import (WarmUp, create_optimizer, AdamWeightDecay, GradientAccumulator)
@@ -189,6 +207,10 @@ from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name
load_tf2_weights_in_pytorch_model,
load_tf2_model_in_pytorch_model)
# Pipelines
from .pipelines import pipeline, PipelineDataFormat, CsvPipelineDataFormat, JsonPipelineDataFormat, PipedPipelineDataFormat, \
Pipeline, FeatureExtractionPipeline, QuestionAnsweringPipeline, NerPipeline, TextClassificationPipeline
if not is_tf_available() and not is_torch_available():
logger.warning("Neither PyTorch nor TensorFlow >= 2.0 have been found."
"Models won't be available and only tokenizers, configuration"

144
transformers/__main__.py
View File

@@ -1,129 +1,37 @@
# coding: utf8
def main():
import sys
if (len(sys.argv) < 4 or len(sys.argv) > 6) or sys.argv[1] not in ["bert", "gpt", "transfo_xl", "gpt2", "xlnet", "xlm"]:
if len(sys.argv) < 2 or sys.argv[1] not in ["convert", "train", "predict", "serve"]:
print(
"This command line utility let you convert original (author released) model checkpoint to pytorch.\n"
"It should be used as one of: \n"
">> transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT, \n"
">> transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG], \n"
">> transformers transfo_xl TF_CHECKPOINT_OR_DATASET PYTORCH_DUMP_OUTPUT [TF_CONFIG] or \n"
">> transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [GPT2_CONFIG] or \n"
">> transformers xlnet TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT [FINETUNING_TASK_NAME] or \n"
">> transformers xlm XLM_CHECKPOINT_PATH PYTORCH_DUMP_OUTPUT")
else:
if sys.argv[1] == "bert":
try:
from .convert_bert_original_tf_checkpoint_to_pytorch import convert_tf_checkpoint_to_pytorch
except ImportError:
print("transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
"In that case, it requires TensorFlow to be installed. Please see "
"https://www.tensorflow.org/install/ for installation instructions.")
raise
"First argument to `transformers` command line interface should be one of: \n"
">> convert serve train predict")
if sys.argv[1] == "convert":
from transformers.commands import convert
convert(sys.argv)
elif sys.argv[1] == "train":
from transformers.commands import train
train(sys.argv)
elif sys.argv[1] == "serve":
pass
# from argparse import ArgumentParser
# from transformers.commands.serving import ServeCommand
# parser = ArgumentParser('Transformers CLI tool', usage='transformers serve <command> [<args>]')
# commands_parser = parser.add_subparsers(help='transformers-cli command helpers')
if len(sys.argv) != 5:
# pylint: disable=line-too-long
print("Should be used as `transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")
else:
PYTORCH_DUMP_OUTPUT = sys.argv.pop()
TF_CONFIG = sys.argv.pop()
TF_CHECKPOINT = sys.argv.pop()
convert_tf_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)
elif sys.argv[1] == "gpt":
from .convert_openai_original_tf_checkpoint_to_pytorch import convert_openai_checkpoint_to_pytorch
if len(sys.argv) < 4 or len(sys.argv) > 5:
# pylint: disable=line-too-long
print("Should be used as `transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG]`")
else:
OPENAI_GPT_CHECKPOINT_FOLDER_PATH = sys.argv[2]
PYTORCH_DUMP_OUTPUT = sys.argv[3]
if len(sys.argv) == 5:
OPENAI_GPT_CONFIG = sys.argv[4]
else:
OPENAI_GPT_CONFIG = ""
convert_openai_checkpoint_to_pytorch(OPENAI_GPT_CHECKPOINT_FOLDER_PATH,
OPENAI_GPT_CONFIG,
PYTORCH_DUMP_OUTPUT)
elif sys.argv[1] == "transfo_xl":
try:
from .convert_transfo_xl_original_tf_checkpoint_to_pytorch import convert_transfo_xl_checkpoint_to_pytorch
except ImportError:
print("transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
"In that case, it requires TensorFlow to be installed. Please see "
"https://www.tensorflow.org/install/ for installation instructions.")
raise
if len(sys.argv) < 4 or len(sys.argv) > 5:
# pylint: disable=line-too-long
print("Should be used as `transformers transfo_xl TF_CHECKPOINT/TF_DATASET_FILE PYTORCH_DUMP_OUTPUT [TF_CONFIG]`")
else:
if 'ckpt' in sys.argv[2].lower():
TF_CHECKPOINT = sys.argv[2]
TF_DATASET_FILE = ""
else:
TF_DATASET_FILE = sys.argv[2]
TF_CHECKPOINT = ""
PYTORCH_DUMP_OUTPUT = sys.argv[3]
if len(sys.argv) == 5:
TF_CONFIG = sys.argv[4]
else:
TF_CONFIG = ""
convert_transfo_xl_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT, TF_DATASET_FILE)
elif sys.argv[1] == "gpt2":
try:
from .convert_gpt2_original_tf_checkpoint_to_pytorch import convert_gpt2_checkpoint_to_pytorch
except ImportError:
print("transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
"In that case, it requires TensorFlow to be installed. Please see "
"https://www.tensorflow.org/install/ for installation instructions.")
raise
if len(sys.argv) < 4 or len(sys.argv) > 5:
# pylint: disable=line-too-long
print("Should be used as `transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [TF_CONFIG]`")
else:
TF_CHECKPOINT = sys.argv[2]
PYTORCH_DUMP_OUTPUT = sys.argv[3]
if len(sys.argv) == 5:
TF_CONFIG = sys.argv[4]
else:
TF_CONFIG = ""
convert_gpt2_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)
elif sys.argv[1] == "xlnet":
try:
from .convert_xlnet_original_tf_checkpoint_to_pytorch import convert_xlnet_checkpoint_to_pytorch
except ImportError:
print("transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
"In that case, it requires TensorFlow to be installed. Please see "
"https://www.tensorflow.org/install/ for installation instructions.")
raise
# # Register commands
# ServeCommand.register_subcommand(commands_parser)
if len(sys.argv) < 5 or len(sys.argv) > 6:
# pylint: disable=line-too-long
print("Should be used as `transformers xlnet TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT [FINETUNING_TASK_NAME]`")
else:
TF_CHECKPOINT = sys.argv[2]
TF_CONFIG = sys.argv[3]
PYTORCH_DUMP_OUTPUT = sys.argv[4]
if len(sys.argv) == 6:
FINETUNING_TASK = sys.argv[5]
else:
FINETUNING_TASK = None
# # Let's go
# args = parser.parse_args()
convert_xlnet_checkpoint_to_pytorch(TF_CHECKPOINT,
TF_CONFIG,
PYTORCH_DUMP_OUTPUT,
FINETUNING_TASK)
elif sys.argv[1] == "xlm":
from .convert_xlm_original_pytorch_checkpoint_to_pytorch import convert_xlm_checkpoint_to_pytorch
if len(sys.argv) != 4:
# pylint: disable=line-too-long
print("Should be used as `transformers xlm XLM_CHECKPOINT_PATH PYTORCH_DUMP_OUTPUT`")
else:
XLM_CHECKPOINT_PATH = sys.argv[2]
PYTORCH_DUMP_OUTPUT = sys.argv[3]
convert_xlm_checkpoint_to_pytorch(XLM_CHECKPOINT_PATH, PYTORCH_DUMP_OUTPUT)
# if not hasattr(args, 'func'):
# parser.print_help()
# exit(1)
# # Run
# service = args.func(args)
# service.run()
if __name__ == '__main__':
main()

115
transformers/commands/convert.py Normal file
View File

@@ -0,0 +1,115 @@
from argparse import ArgumentParser, Namespace
from logging import getLogger
from transformers import AutoModel, AutoTokenizer
from transformers.commands import BaseTransformersCLICommand
def convert_command_factory(args: Namespace):
"""
Factory function used to convert a model TF 1.0 checkpoint in a PyTorch checkpoint.
:return: ServeCommand
"""
return ConvertCommand(args.model_type, args.tf_checkpoint, args.pytorch_dump_output,
args.config, args.finetuning_task_name)
class ConvertCommand(BaseTransformersCLICommand):
@staticmethod
def register_subcommand(parser: ArgumentParser):
"""
Register this command to argparse so it's available for the transformer-cli
:param parser: Root parser to register command-specific arguments
:return:
"""
train_parser = parser.add_parser('convert', help="CLI tool to run convert model from original "
"author checkpoints to Transformesr PyTorch checkpoints.")
train_parser.add_argument('--model_type', type=str, required=True,
help='Model\'s type.')
train_parser.add_argument('--tf_checkpoint', type=str, required=True,
help='TensorFlow checkpoint path or folder.')
train_parser.add_argument('--pytorch_dump_output', type=str, required=True,
help='Path to the PyTorch savd model output.')
train_parser.add_argument('--config', type=str, default="",
help='Configuration file path or folder.')
train_parser.add_argument('--finetuning_task_name', type=str, default=None,
help='Optional fine-tuning task name if the TF model was a finetuned model.')
train_parser.set_defaults(func=convert_command_factory)
def __init__(self, model_type: str, tf_checkpoint: str, pytorch_dump_output: str,
config: str, finetuning_task_name: str, *args):
self._logger = getLogger('transformers-cli/converting')
self._logger.info('Loading model {}'.format(model_type))
self._model_type = model_type
self._tf_checkpoint = tf_checkpoint
self._pytorch_dump_output = pytorch_dump_output
self._config = config
self._finetuning_task_name = finetuning_task_name
def run(self):
if self._model_type == "bert":
try:
from transformers.convert_bert_original_tf_checkpoint_to_pytorch import convert_tf_checkpoint_to_pytorch
except ImportError:
msg = "transformers can only be used from the commandline to convert TensorFlow models in PyTorch, " \
"In that case, it requires TensorFlow to be installed. Please see " \
"https://www.tensorflow.org/install/ for installation instructions."
raise ImportError(msg)
convert_tf_checkpoint_to_pytorch(self._tf_checkpoint, self._config, self._pytorch_dump_output)
elif self._model_type == "gpt":
from transformers.convert_openai_original_tf_checkpoint_to_pytorch import convert_openai_checkpoint_to_pytorch
convert_openai_checkpoint_to_pytorch(self._tf_checkpoint,
self._config,
self._pytorch_dump_output)
elif self._model_type == "transfo_xl":
try:
from transformers.convert_transfo_xl_original_tf_checkpoint_to_pytorch import convert_transfo_xl_checkpoint_to_pytorch
except ImportError:
msg = "transformers can only be used from the commandline to convert TensorFlow models in PyTorch, " \
"In that case, it requires TensorFlow to be installed. Please see " \
"https://www.tensorflow.org/install/ for installation instructions."
raise ImportError(msg)
if 'ckpt' in self._tf_checkpoint.lower():
TF_CHECKPOINT = self._tf_checkpoint
TF_DATASET_FILE = ""
else:
TF_DATASET_FILE = self._tf_checkpoint
TF_CHECKPOINT = ""
convert_transfo_xl_checkpoint_to_pytorch(TF_CHECKPOINT,
self._config,
self._pytorch_dump_output,
TF_DATASET_FILE)
elif self._model_type == "gpt2":
try:
from transformers.convert_gpt2_original_tf_checkpoint_to_pytorch import convert_gpt2_checkpoint_to_pytorch
except ImportError:
msg = "transformers can only be used from the commandline to convert TensorFlow models in PyTorch, " \
"In that case, it requires TensorFlow to be installed. Please see " \
"https://www.tensorflow.org/install/ for installation instructions."
raise ImportError(msg)
convert_gpt2_checkpoint_to_pytorch(self._tf_checkpoint, self._config, self._pytorch_dump_output)
elif self._model_type == "xlnet":
try:
from transformers.convert_xlnet_original_tf_checkpoint_to_pytorch import convert_xlnet_checkpoint_to_pytorch
except ImportError:
msg = "transformers can only be used from the commandline to convert TensorFlow models in PyTorch, " \
"In that case, it requires TensorFlow to be installed. Please see " \
"https://www.tensorflow.org/install/ for installation instructions."
raise ImportError(msg)
convert_xlnet_checkpoint_to_pytorch(self._tf_checkpoint,
self._config,
self._pytorch_dump_output,
self._finetuning_task_name)
elif self._model_type == "xlm":
from transformers.convert_xlm_original_pytorch_checkpoint_to_pytorch import convert_xlm_checkpoint_to_pytorch
convert_xlm_checkpoint_to_pytorch(self._tf_checkpoint, self._pytorch_dump_output)
else:
raise ValueError("--model_type should be selected in the list [bert, gpt, gpt2, transfo_xl, xlnet, xlm]")

29
transformers/commands/download.py Normal file
View File

@@ -0,0 +1,29 @@
from argparse import ArgumentParser
from transformers.commands import BaseTransformersCLICommand
def download_command_factory(args):
return DownloadCommand(args.model, args.cache_dir, args.force)
class DownloadCommand(BaseTransformersCLICommand):
@staticmethod
def register_subcommand(parser: ArgumentParser):
download_parser = parser.add_parser('download')
download_parser.add_argument('--cache-dir', type=str, default=None, help='Path to location to store the models')
download_parser.add_argument('--force', action='store_true', help='Force the model to be download even if already in cache-dir')
download_parser.add_argument('model', type=str, help='Name of the model to download')
download_parser.set_defaults(func=download_command_factory)
def __init__(self, model: str, cache: str, force: bool):
self._model = model
self._cache = cache
self._force = force
def run(self):
from transformers import AutoModel, AutoTokenizer
AutoModel.from_pretrained(self._model, cache_dir=self._cache, force_download=self._force)
AutoTokenizer.from_pretrained(self._model, cache_dir=self._cache, force_download=self._force)

79
transformers/commands/run.py Normal file
View File

@@ -0,0 +1,79 @@
import logging
from argparse import ArgumentParser
from transformers.commands import BaseTransformersCLICommand
from transformers.pipelines import pipeline, Pipeline, PipelineDataFormat, SUPPORTED_TASKS
logger = logging.getLogger(__name__) # pylint: disable=invalid-name
def try_infer_format_from_ext(path: str):
if not path:
return 'pipe'
for ext in PipelineDataFormat.SUPPORTED_FORMATS:
if path.endswith(ext):
return ext
raise Exception(
'Unable to determine file format from file extension {}. '
'Please provide the format through --format {}'.format(path, PipelineDataFormat.SUPPORTED_FORMATS)
)
def run_command_factory(args):
nlp = pipeline(task=args.task,
model=args.model if args.model else None,
config=args.config,
tokenizer=args.tokenizer,
device=args.device)
format = try_infer_format_from_ext(args.input) if args.format == 'infer' else args.format
reader = PipelineDataFormat.from_str(format=format,
output_path=args.output,
input_path=args.input,
column=args.column if args.column else nlp.default_input_names,
overwrite=args.overwrite)
return RunCommand(nlp, reader)
class RunCommand(BaseTransformersCLICommand):
def __init__(self, nlp: Pipeline, reader: PipelineDataFormat):
self._nlp = nlp
self._reader = reader
@staticmethod
def register_subcommand(parser: ArgumentParser):
run_parser = parser.add_parser('run', help="Run a pipeline through the CLI")
run_parser.add_argument('--task', choices=SUPPORTED_TASKS.keys(), help='Task to run')
run_parser.add_argument('--input', type=str, help='Path to the file to use for inference')
run_parser.add_argument('--output', type=str, help='Path to the file that will be used post to write results.')
run_parser.add_argument('--model', type=str, help='Name or path to the model to instantiate.')
run_parser.add_argument('--config', type=str, help='Name or path to the model\'s config to instantiate.')
run_parser.add_argument('--tokenizer', type=str, help='Name of the tokenizer to use. (default: same as the model name)')
run_parser.add_argument('--column', type=str, help='Name of the column to use as input. (For multi columns input as QA use column1,columns2)')
run_parser.add_argument('--format', type=str, default='infer', choices=PipelineDataFormat.SUPPORTED_FORMATS, help='Input format to read from')
run_parser.add_argument('--device', type=int, default=-1, help='Indicate the device to run onto, -1 indicates CPU, >= 0 indicates GPU (default: -1)')
run_parser.add_argument('--overwrite', action='store_true', help='Allow overwriting the output file.')
run_parser.set_defaults(func=run_command_factory)
def run(self):
nlp, outputs = self._nlp, []
for entry in self._reader:
output = nlp(**entry) if self._reader.is_multi_columns else nlp(entry)
if isinstance(output, dict):
outputs.append(output)
else:
outputs += output
# Saving data
if self._nlp.binary_output:
binary_path = self._reader.save_binary(outputs)
logger.warning('Current pipeline requires output to be in binary format, saving at {}'.format(binary_path))
else:
self._reader.save(outputs)

158
transformers/commands/serving.py Normal file
View File

@@ -0,0 +1,158 @@
from argparse import ArgumentParser, Namespace
from typing import List, Optional, Union, Any
import logging
try:
from uvicorn import run
from fastapi import FastAPI, HTTPException, Body
from pydantic import BaseModel
_serve_dependancies_installed = True
except (ImportError, AttributeError):
BaseModel = object
Body = lambda *x, **y: None
_serve_dependancies_installed = False
from transformers import Pipeline
from transformers.commands import BaseTransformersCLICommand
from transformers.pipelines import SUPPORTED_TASKS, pipeline
logger = logging.getLogger('transformers-cli/serving')
def serve_command_factory(args: Namespace):
"""
Factory function used to instantiate serving server from provided command line arguments.
:return: ServeCommand
"""
nlp = pipeline(task=args.task,
model=args.model if args.model else None,
config=args.config,
tokenizer=args.tokenizer,
device=args.device)
return ServeCommand(nlp, args.host, args.port)
class ServeModelInfoResult(BaseModel):
"""
Expose model information
"""
infos: dict
class ServeTokenizeResult(BaseModel):
"""
Tokenize result model
"""
tokens: List[str]
tokens_ids: Optional[List[int]]
class ServeDeTokenizeResult(BaseModel):
"""
DeTokenize result model
"""
text: str
class ServeForwardResult(BaseModel):
"""
Forward result model
"""
output: Any
class ServeCommand(BaseTransformersCLICommand):
@staticmethod
def register_subcommand(parser: ArgumentParser):
"""
Register this command to argparse so it's available for the transformer-cli
:param parser: Root parser to register command-specific arguments
:return:
"""
serve_parser = parser.add_parser('serve', help='CLI tool to run inference requests through REST and GraphQL endpoints.')
serve_parser.add_argument('--task', type=str, choices=SUPPORTED_TASKS.keys(), help='The task to run the pipeline on')
serve_parser.add_argument('--host', type=str, default='localhost', help='Interface the server will listen on.')
serve_parser.add_argument('--port', type=int, default=8888, help='Port the serving will listen to.')
serve_parser.add_argument('--model', type=str, help='Model\'s name or path to stored model.')
serve_parser.add_argument('--config', type=str, help='Model\'s config name or path to stored model.')
serve_parser.add_argument('--tokenizer', type=str, help='Tokenizer name to use.')
serve_parser.add_argument('--device', type=int, default=-1, help='Indicate the device to run onto, -1 indicates CPU, >= 0 indicates GPU (default: -1)')
serve_parser.set_defaults(func=serve_command_factory)
def __init__(self, pipeline: Pipeline, host: str, port: int):
self._pipeline = pipeline
self._host = host
self._port = port
if not _serve_dependancies_installed:
raise ImportError("Using serve command requires FastAPI and unicorn. "
"Please install transformers with [serving]: pip install transformers[serving]."
"Or install FastAPI and unicorn separatly.")
else:
logger.info('Serving model over {}:{}'.format(host, port))
self._app = FastAPI()
# Register routes
self._app.add_api_route('/', self.model_info, response_model=ServeModelInfoResult, methods=['GET'])
self._app.add_api_route('/tokenize', self.tokenize, response_model=ServeTokenizeResult, methods=['POST'])
self._app.add_api_route('/detokenize', self.detokenize, response_model=ServeDeTokenizeResult, methods=['POST'])
self._app.add_api_route('/forward', self.forward, response_model=ServeForwardResult, methods=['POST'])
def run(self):
run(self._app, host=self._host, port=self._port)
def model_info(self):
return ServeModelInfoResult(infos=vars(self._pipeline.model.config))
def tokenize(self, text_input: str = Body(None, embed=True), return_ids: bool = Body(False, embed=True)):
"""
Tokenize the provided input and eventually returns corresponding tokens id:
- **text_input**: String to tokenize
- **return_ids**: Boolean flags indicating if the tokens have to be converted to their integer mapping.
"""
try:
tokens_txt = self._pipeline.tokenizer.tokenize(text_input)
if return_ids:
tokens_ids = self._pipeline.tokenizer.convert_tokens_to_ids(tokens_txt)
return ServeTokenizeResult(tokens=tokens_txt, tokens_ids=tokens_ids)
else:
return ServeTokenizeResult(tokens=tokens_txt)
except Exception as e:
raise HTTPException(status_code=500, detail={"model": '', "error": str(e)})
def detokenize(self, tokens_ids: List[int] = Body(None, embed=True),
skip_special_tokens: bool = Body(False, embed=True),
cleanup_tokenization_spaces: bool = Body(True, embed=True)):
"""
Detokenize the provided tokens ids to readable text:
- **tokens_ids**: List of tokens ids
- **skip_special_tokens**: Flag indicating to not try to decode special tokens
- **cleanup_tokenization_spaces**: Flag indicating to remove all leading/trailing spaces and intermediate ones.
"""
try:
decoded_str = self._pipeline.tokenizer.decode(tokens_ids, skip_special_tokens, cleanup_tokenization_spaces)
return ServeDeTokenizeResult(model='', text=decoded_str)
except Exception as e:
raise HTTPException(status_code=500, detail={"model": '', "error": str(e)})
def forward(self, inputs: Union[str, dict, List[str], List[int], List[dict]] = Body(None, embed=True)):
"""
**inputs**:
**attention_mask**:
**tokens_type_ids**:
"""
# Check we don't have empty string
if len(inputs) == 0:
return ServeForwardResult(output=[], attention=[])
try:
# Forward through the model
output = self._pipeline(inputs)
return ServeForwardResult(output=output)
except Exception as e:
raise HTTPException(500, {"error": str(e)})

131
transformers/commands/train.py Normal file
View File

@@ -0,0 +1,131 @@
import os
from argparse import ArgumentParser, Namespace
from logging import getLogger
from transformers.commands import BaseTransformersCLICommand
from transformers import (is_tf_available, is_torch_available,
TextClassificationPipeline,
SingleSentenceClassificationProcessor as Processor)
if not is_tf_available() and not is_torch_available():
raise ImportError("At least one of PyTorch or TensorFlow 2.0+ should be installed to use CLI training")
# TF training parameters
USE_XLA = False
USE_AMP = False
def train_command_factory(args: Namespace):
"""
Factory function used to instantiate serving server from provided command line arguments.
:return: ServeCommand
"""
return TrainCommand(args)
class TrainCommand(BaseTransformersCLICommand):
@staticmethod
def register_subcommand(parser: ArgumentParser):
"""
Register this command to argparse so it's available for the transformer-cli
:param parser: Root parser to register command-specific arguments
:return:
"""
train_parser = parser.add_parser('train', help='CLI tool to train a model on a task.')
train_parser.add_argument('--train_data', type=str, required=True,
help="path to train (and optionally evaluation) dataset as a csv with "
"tab separated labels and sentences.")
train_parser.add_argument('--column_label', type=int, default=0,
help='Column of the dataset csv file with example labels.')
train_parser.add_argument('--column_text', type=int, default=1,
help='Column of the dataset csv file with example texts.')
train_parser.add_argument('--column_id', type=int, default=2,
help='Column of the dataset csv file with example ids.')
train_parser.add_argument('--skip_first_row', action='store_true',
help='Skip the first row of the csv file (headers).')
train_parser.add_argument('--validation_data', type=str, default='',
help='path to validation dataset.')
train_parser.add_argument('--validation_split', type=float, default=0.1,
help="if validation dataset is not provided, fraction of train dataset "
"to use as validation dataset.")
train_parser.add_argument('--output', type=str, default='./',
help='path to saved the trained model.')
train_parser.add_argument('--task', type=str, default='text_classification',
help='Task to train the model on.')
train_parser.add_argument('--model', type=str, default='bert-base-uncased',
help='Model\'s name or path to stored model.')
train_parser.add_argument('--train_batch_size', type=int, default=32,
help='Batch size for training.')
train_parser.add_argument('--valid_batch_size', type=int, default=64,
help='Batch size for validation.')
train_parser.add_argument('--learning_rate', type=float, default=3e-5,
help="Learning rate.")
train_parser.add_argument('--adam_epsilon', type=float, default=1e-08,
help="Epsilon for Adam optimizer.")
train_parser.set_defaults(func=train_command_factory)
def __init__(self, args: Namespace):
self.logger = getLogger('transformers-cli/training')
self.framework = 'tf' if is_tf_available() else 'torch'
os.makedirs(args.output, exist_ok=True)
assert os.path.isdir(args.output)
self.output = args.output
self.column_label = args.column_label
self.column_text = args.column_text
self.column_id = args.column_id
self.logger.info('Loading {} pipeline for {}'.format(args.task, args.model))
if args.task == 'text_classification':
self.pipeline = TextClassificationPipeline.from_pretrained(args.model)
elif args.task == 'token_classification':
raise NotImplementedError
elif args.task == 'question_answering':
raise NotImplementedError
self.logger.info('Loading dataset from {}'.format(args.train_data))
self.train_dataset = Processor.create_from_csv(args.train_data,
column_label=args.column_label,
column_text=args.column_text,
column_id=args.column_id,
skip_first_row=args.skip_first_row)
self.valid_dataset = None
if args.validation_data:
self.logger.info('Loading validation dataset from {}'.format(args.validation_data))
self.valid_dataset = Processor.create_from_csv(args.validation_data,
column_label=args.column_label,
column_text=args.column_text,
column_id=args.column_id,
skip_first_row=args.skip_first_row)
self.validation_split = args.validation_split
self.train_batch_size = args.train_batch_size
self.valid_batch_size = args.valid_batch_size
self.learning_rate = args.learning_rate
self.adam_epsilon = args.adam_epsilon
def run(self):
if self.framework == 'tf':
return self.run_tf()
return self.run_torch()
def run_torch(self):
raise NotImplementedError
def run_tf(self):
self.pipeline.fit(self.train_dataset,
validation_data=self.valid_dataset,
validation_split=self.validation_split,
learning_rate=self.learning_rate,
adam_epsilon=self.adam_epsilon,
train_batch_size=self.train_batch_size,
valid_batch_size=self.valid_batch_size)
# Save trained pipeline
self.pipeline.save_pretrained(self.output)

39
transformers/commands/user.py
View File

@@ -19,8 +19,8 @@ class UserCommands(BaseTransformersCLICommand):
list_parser.set_defaults(func=lambda args: ListObjsCommand(args))
# upload
upload_parser = parser.add_parser('upload')
upload_parser.add_argument('file', type=str, help='Local filepath of the file to upload.')
upload_parser.add_argument('--filename', type=str, default=None, help='Optional: override object filename on S3.')
upload_parser.add_argument('path', type=str, help='Local path of the folder or individual file to upload.')
upload_parser.add_argument('--filename', type=str, default=None, help='Optional: override individual object filename on S3.')
upload_parser.set_defaults(func=lambda args: UploadCommand(args))
@@ -138,13 +138,41 @@ class ListObjsCommand(BaseUserCommand):
class UploadCommand(BaseUserCommand):
def walk_dir(self, rel_path):
"""
Recursively list all files in a folder.
"""
entries: List[os.DirEntry] = list(os.scandir(rel_path))
files = [
(
os.path.join(os.getcwd(), f.path), # filepath
f.path # filename
)
for f in entries if f.is_file()
]
for f in entries:
if f.is_dir():
files += self.walk_dir(f.path)
return files
def run(self):
token = HfFolder.get_token()
if token is None:
print("Not logged in")
exit(1)
filepath = os.path.join(os.getcwd(), self.args.file)
filename = self.args.filename if self.args.filename is not None else os.path.basename(filepath)
local_path = os.path.abspath(self.args.path)
if os.path.isdir(local_path):
if self.args.filename is not None:
raise ValueError("Cannot specify a filename override when uploading a folder.")
rel_path = os.path.basename(local_path)
files = self.walk_dir(rel_path)
elif os.path.isfile(local_path):
filename = self.args.filename if self.args.filename is not None else os.path.basename(local_path)
files = [(local_path, filename)]
else:
raise ValueError("Not a valid file or directory: {}".format(local_path))
for filepath, filename in files:
print(
"About to upload file {} to S3 under filename {}".format(
ANSI.bold(filepath), ANSI.bold(filename)
@@ -156,8 +184,9 @@ class UploadCommand(BaseUserCommand):
print("Abort")
exit()
print(
ANSI.bold("Uploading... This might take a while if file is large")
ANSI.bold("Uploading... This might take a while if files are large")
)
for filepath, filename in files:
access_url = self._api.presign_and_upload(
token=token, filename=filename, filepath=filepath
)

4
transformers/configuration_albert.py
View File

@@ -37,7 +37,7 @@ class AlbertConfig(PretrainedConfig):
pretrained_config_archive_map = ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
def __init__(self,
vocab_size_or_config_json_file=30000,
vocab_size=30000,
embedding_size=128,
hidden_size=4096,
num_hidden_layers=12,
@@ -83,7 +83,7 @@ class AlbertConfig(PretrainedConfig):
"""
super(AlbertConfig, self).__init__(**kwargs)
self.vocab_size = vocab_size_or_config_json_file
self.vocab_size = vocab_size
self.embedding_size = embedding_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers

83
transformers/configuration_auto.py
View File

@@ -18,21 +18,42 @@ from __future__ import absolute_import, division, print_function, unicode_litera
import logging
from .configuration_bert import BertConfig
from .configuration_openai import OpenAIGPTConfig
from .configuration_gpt2 import GPT2Config
from .configuration_transfo_xl import TransfoXLConfig
from .configuration_xlnet import XLNetConfig
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
from .configuration_albert import AlbertConfig
from .configuration_bert import BertConfig, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_openai import OpenAIGPTConfig, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_transfo_xl import TransfoXLConfig, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_gpt2 import GPT2Config, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_ctrl import CTRLConfig, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_xlnet import XLNetConfig, XLNET_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
from .configuration_t5 import T5Config, T5_PRETRAINED_CONFIG_ARCHIVE_MAP
from .configuration_xlm_roberta import XLMRobertaConfig, XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
logger = logging.getLogger(__name__)
ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict((key, value)
for pretrained_map in [
BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,
XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
T5_PRETRAINED_CONFIG_ARCHIVE_MAP,
XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
]
for key, value, in pretrained_map.items())
class AutoConfig(object):
r""":class:`~transformers.AutoConfig` is a generic configuration class
that will be instantiated as one of the configuration classes of the library
@@ -47,6 +68,7 @@ class AutoConfig(object):
- contains `distilbert`: DistilBertConfig (DistilBERT model)
- contains `albert`: AlbertConfig (ALBERT model)
- contains `camembert`: CamembertConfig (CamemBERT model)
- contains `xlm-roberta`: XLMRobertaConfig (XLM-RoBERTa model)
- contains `roberta`: RobertaConfig (RoBERTa model)
- contains `bert`: BertConfig (Bert model)
- contains `openai-gpt`: OpenAIGPTConfig (OpenAI GPT model)
@@ -61,6 +83,34 @@ class AutoConfig(object):
raise EnvironmentError("AutoConfig is designed to be instantiated "
"using the `AutoConfig.from_pretrained(pretrained_model_name_or_path)` method.")
@classmethod
def for_model(cls, model_type, *args, **kwargs):
if 'distilbert' in model_type:
return DistilBertConfig(*args, **kwargs)
elif 'roberta' in model_type:
return RobertaConfig(*args, **kwargs)
elif 'bert' in model_type:
return BertConfig(*args, **kwargs)
elif 'openai-gpt' in model_type:
return OpenAIGPTConfig(*args, **kwargs)
elif 'gpt2' in model_type:
return GPT2Config(*args, **kwargs)
elif 'transfo-xl' in model_type:
return TransfoXLConfig(*args, **kwargs)
elif 'xlnet' in model_type:
return XLNetConfig(*args, **kwargs)
elif 'xlm' in model_type:
return XLMConfig(*args, **kwargs)
elif 'ctrl' in model_type:
return CTRLConfig(*args, **kwargs)
elif 'albert' in model_type:
return AlbertConfig(*args, **kwargs)
elif 'camembert' in model_type:
return CamembertConfig(*args, **kwargs)
raise ValueError("Unrecognized model identifier in {}. Should contains one of "
"'distilbert', 'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
"'xlm', 'roberta', 'ctrl', 'camembert', 'albert'".format(model_type))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
r""" Instantiate a one of the configuration classes of the library
@@ -68,9 +118,11 @@ class AutoConfig(object):
The configuration class to instantiate is selected as the first pattern matching
in the `pretrained_model_name_or_path` string (in the following order):
- contains `t5`: T5Config (T5 model)
- contains `distilbert`: DistilBertConfig (DistilBERT model)
- contains `albert`: AlbertConfig (ALBERT model)
- contains `camembert`: CamembertConfig (CamemBERT model)
- contains `xlm-roberta`: XLMRobertaConfig (XLM-RoBERTa model)
- contains `roberta`: RobertaConfig (RoBERTa model)
- contains `bert`: BertConfig (Bert model)
- contains `openai-gpt`: OpenAIGPTConfig (OpenAI GPT model)
@@ -83,6 +135,7 @@ class AutoConfig(object):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model configuration to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model configuration that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing a configuration file saved using the :func:`~transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
- a path or url to a saved configuration JSON `file`, e.g.: ``./my_model_directory/configuration.json``.
@@ -123,12 +176,16 @@ class AutoConfig(object):
assert unused_kwargs == {'foo': False}
"""
if 'distilbert' in pretrained_model_name_or_path:
if 't5' in pretrained_model_name_or_path:
return T5Config.from_pretrained(pretrained_model_name_or_path, **kwargs)
elif 'distilbert' in pretrained_model_name_or_path:
return DistilBertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
elif 'albert' in pretrained_model_name_or_path:
return AlbertConfig.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 'xlm-roberta' in pretrained_model_name_or_path:
return XLMRobertaConfig.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:
@@ -147,4 +204,4 @@ 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', 'distilbert', 'camembert', 'ctrl', 'albert'".format(pretrained_model_name_or_path))
"'xlm-roberta', 'xlm', 'roberta', 'distilbert', 'camembert', 'ctrl', 'albert'".format(pretrained_model_name_or_path))

22
transformers/configuration_bert.py
View File

@@ -42,6 +42,12 @@ BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
'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",
'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-config.json",
'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-config.json",
'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-config.json",
'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-config.json",
'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/config.json",
'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/config.json",
}
@@ -52,7 +58,7 @@ class BertConfig(PretrainedConfig):
Arguments:
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `BertModel`.
vocab_size: Vocabulary size of `inputs_ids` in `BertModel`.
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
@@ -77,7 +83,7 @@ class BertConfig(PretrainedConfig):
pretrained_config_archive_map = BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
def __init__(self,
vocab_size_or_config_json_file=30522,
vocab_size=30522,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
@@ -91,14 +97,7 @@ class BertConfig(PretrainedConfig):
layer_norm_eps=1e-12,
**kwargs):
super(BertConfig, self).__init__(**kwargs)
if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
and isinstance(vocab_size_or_config_json_file, unicode)):
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 isinstance(vocab_size_or_config_json_file, int):
self.vocab_size = vocab_size_or_config_json_file
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
@@ -110,6 +109,3 @@ class BertConfig(PretrainedConfig):
self.type_vocab_size = type_vocab_size
self.initializer_range = initializer_range
self.layer_norm_eps = layer_norm_eps
else:
raise ValueError("First argument must be either a vocabulary size (int)"
" or the path to a pretrained model config file (str)")

23
transformers/configuration_ctrl.py
View File

@@ -31,7 +31,7 @@ class CTRLConfig(PretrainedConfig):
"""Configuration class to store the configuration of a `CTRLModel`.
Args:
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
vocab_size: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
n_positions: Number of positional embeddings.
n_ctx: Size of the causal mask (usually same as n_positions).
dff: Size of the inner dimension of the FFN.
@@ -52,7 +52,7 @@ class CTRLConfig(PretrainedConfig):
def __init__(
self,
vocab_size_or_config_json_file=246534,
vocab_size=246534,
n_positions=256,
n_ctx=256,
n_embd=1280,
@@ -64,8 +64,6 @@ class CTRLConfig(PretrainedConfig):
attn_pdrop=0.1,
layer_norm_epsilon=1e-6,
initializer_range=0.02,
num_labels=1,
summary_type='cls_index',
summary_use_proj=True,
summary_activation=None,
@@ -76,7 +74,7 @@ class CTRLConfig(PretrainedConfig):
"""Constructs CTRLConfig.
Args:
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
vocab_size: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
n_positions: Number of positional embeddings.
n_ctx: Size of the causal mask (usually same as n_positions).
dff: Size of the inner dimension of the FFN.
@@ -94,8 +92,7 @@ class CTRLConfig(PretrainedConfig):
initializing all weight matrices.
"""
super(CTRLConfig, self).__init__(**kwargs)
self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
self.vocab_size = vocab_size
self.n_ctx = n_ctx
self.n_positions = n_positions
self.n_embd = n_embd
@@ -108,23 +105,11 @@ class CTRLConfig(PretrainedConfig):
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, str) or (sys.version_info[0] == 2
and isinstance(vocab_size_or_config_json_file, unicode)):
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):

16
transformers/configuration_distilbert.py
View File

@@ -37,7 +37,7 @@ class DistilBertConfig(PretrainedConfig):
pretrained_config_archive_map = DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
def __init__(self,
vocab_size_or_config_json_file=30522,
vocab_size=30522,
max_position_embeddings=512,
sinusoidal_pos_embds=False,
n_layers=6,
@@ -53,15 +53,7 @@ class DistilBertConfig(PretrainedConfig):
seq_classif_dropout=0.2,
**kwargs):
super(DistilBertConfig, self).__init__(**kwargs)
if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
and isinstance(vocab_size_or_config_json_file, unicode)):
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 isinstance(vocab_size_or_config_json_file, int):
self.vocab_size = vocab_size_or_config_json_file
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.sinusoidal_pos_embds = sinusoidal_pos_embds
self.n_layers = n_layers
@@ -75,9 +67,7 @@ class DistilBertConfig(PretrainedConfig):
self.tie_weights_ = tie_weights_
self.qa_dropout = qa_dropout
self.seq_classif_dropout = seq_classif_dropout
else:
raise ValueError("First argument must be either a vocabulary size (int)"
" or the path to a pretrained model config file (str)")
@property
def hidden_size(self):
return self.dim

25
transformers/configuration_gpt2.py
View File

@@ -36,7 +36,7 @@ class GPT2Config(PretrainedConfig):
"""Configuration class to store the configuration of a `GPT2Model`.
Args:
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
vocab_size: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
n_positions: Number of positional embeddings.
n_ctx: Size of the causal mask (usually same as n_positions).
n_embd: Dimensionality of the embeddings and hidden states.
@@ -56,7 +56,7 @@ class GPT2Config(PretrainedConfig):
def __init__(
self,
vocab_size_or_config_json_file=50257,
vocab_size=50257,
n_positions=1024,
n_ctx=1024,
n_embd=768,
@@ -67,8 +67,6 @@ class GPT2Config(PretrainedConfig):
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,
@@ -79,7 +77,7 @@ class GPT2Config(PretrainedConfig):
"""Constructs GPT2Config.
Args:
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
vocab_size: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
n_positions: Number of positional embeddings.
n_ctx: Size of the causal mask (usually same as n_positions).
n_embd: Dimensionality of the embeddings and hidden states.
@@ -96,15 +94,7 @@ class GPT2Config(PretrainedConfig):
initializing all weight matrices.
"""
super(GPT2Config, self).__init__(**kwargs)
if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
and isinstance(vocab_size_or_config_json_file, unicode)):
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 isinstance(vocab_size_or_config_json_file, int):
self.vocab_size = vocab_size_or_config_json_file
self.vocab_size = vocab_size
self.n_ctx = n_ctx
self.n_positions = n_positions
self.n_embd = n_embd
@@ -115,18 +105,11 @@ class GPT2Config(PretrainedConfig):
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
else:
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):

23
transformers/configuration_openai.py
View File

@@ -35,7 +35,7 @@ class OpenAIGPTConfig(PretrainedConfig):
Configuration class to store the configuration of a `OpenAIGPTModel`.
Args:
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `OpenAIGPTModel` or a configuration json file.
vocab_size: Vocabulary size of `inputs_ids` in `OpenAIGPTModel` or a configuration json file.
n_positions: Number of positional embeddings.
n_ctx: Size of the causal mask (usually same as n_positions).
n_embd: Dimensionality of the embeddings and hidden states.
@@ -58,7 +58,7 @@ class OpenAIGPTConfig(PretrainedConfig):
def __init__(
self,
vocab_size_or_config_json_file=40478,
vocab_size=40478,
n_positions=512,
n_ctx=512,
n_embd=768,
@@ -71,8 +71,6 @@ class OpenAIGPTConfig(PretrainedConfig):
layer_norm_epsilon=1e-5,
initializer_range=0.02,
predict_special_tokens=True,
num_labels=1,
summary_type='cls_index',
summary_use_proj=True,
summary_activation=None,
@@ -83,15 +81,7 @@ class OpenAIGPTConfig(PretrainedConfig):
"""Constructs OpenAIGPTConfig.
"""
super(OpenAIGPTConfig, self).__init__(**kwargs)
if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
and isinstance(vocab_size_or_config_json_file, unicode)):
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 isinstance(vocab_size_or_config_json_file, int):
self.vocab_size = vocab_size_or_config_json_file
self.vocab_size = vocab_size
self.n_ctx = n_ctx
self.n_positions = n_positions
self.n_embd = n_embd
@@ -104,18 +94,11 @@ class OpenAIGPTConfig(PretrainedConfig):
self.layer_norm_epsilon = layer_norm_epsilon
self.initializer_range = initializer_range
self.predict_special_tokens = predict_special_tokens
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
else:
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):

108
transformers/configuration_t5.py Normal file
View File

@@ -0,0 +1,108 @@
# coding=utf-8
# Copyright 2010, The T5 Authors and 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.
""" T5 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__)
T5_PRETRAINED_CONFIG_ARCHIVE_MAP = {
't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-small-config.json",
't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-base-config.json",
't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-large-config.json",
't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-3b-config.json",
't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-11b-config.json",
}
class T5Config(PretrainedConfig):
r"""
:class:`~transformers.T5Config` is the configuration class to store the configuration of a
`T5Model`.
Arguments:
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `T5Model`.
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
`T5Model`.
initializer_factor: A factor for initializing all weight matrices (should be kept to 1.0, used for initialization testing).
layer_norm_eps: The epsilon used by LayerNorm.
"""
pretrained_config_archive_map = T5_PRETRAINED_CONFIG_ARCHIVE_MAP
def __init__(self,
vocab_size=32128,
n_positions=512,
d_model=512,
d_kv=64,
d_ff=2048,
num_layers=6,
num_heads=8,
relative_attention_num_buckets=32,
dropout_rate=0.1,
layer_norm_epsilon=1e-6,
initializer_factor=1.0,
**kwargs):
super(T5Config, self).__init__(**kwargs)
self.vocab_size = vocab_size
self.n_positions = n_positions
self.d_model = d_model
self.d_kv = d_kv
self.d_ff = d_ff
self.num_layers = num_layers
self.num_heads = num_heads
self.relative_attention_num_buckets = relative_attention_num_buckets
self.dropout_rate = dropout_rate
self.layer_norm_epsilon = layer_norm_epsilon
self.initializer_factor = initializer_factor
@property
def max_position_embeddings(self):
return self.n_positions
@property
def hidden_size(self):
return self.d_model
@property
def num_attention_heads(self):
return self.num_heads
@property
def num_hidden_layers(self):
return self.num_layers

26
transformers/configuration_transfo_xl.py
View File

@@ -34,7 +34,7 @@ class TransfoXLConfig(PretrainedConfig):
"""Configuration class to store the configuration of a `TransfoXLModel`.
Args:
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `TransfoXLModel` or a configuration json file.
vocab_size: Vocabulary size of `inputs_ids` in `TransfoXLModel` or a configuration json file.
cutoffs: cutoffs for the adaptive softmax
d_model: Dimensionality of the model's hidden states.
d_embed: Dimensionality of the embeddings
@@ -68,7 +68,7 @@ class TransfoXLConfig(PretrainedConfig):
pretrained_config_archive_map = TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
def __init__(self,
vocab_size_or_config_json_file=267735,
vocab_size=267735,
cutoffs=[20000, 40000, 200000],
d_model=1024,
d_embed=1024,
@@ -100,7 +100,7 @@ class TransfoXLConfig(PretrainedConfig):
"""Constructs TransfoXLConfig.
"""
super(TransfoXLConfig, self).__init__(**kwargs)
self.n_token = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
self.vocab_size = vocab_size
self.cutoffs = []
self.cutoffs.extend(cutoffs)
self.tie_weight = tie_weight
@@ -133,27 +133,17 @@ class TransfoXLConfig(PretrainedConfig):
self.init_std = init_std
self.layer_norm_epsilon = layer_norm_epsilon
if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
and isinstance(vocab_size_or_config_json_file, unicode)):
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.tgt_len + self.ext_len + self.mem_len
@property
def vocab_size(self):
return self.n_token
def n_token(self): # Backward compatibility
return self.vocab_size
@vocab_size.setter
def vocab_size(self, value):
self.n_token = value
@n_token.setter
def n_token(self, value): # Backward compatibility
self.vocab_size = value
@property
def hidden_size(self):

67
transformers/configuration_utils.py
View File

@@ -24,7 +24,7 @@ import logging
import os
from io import open
from .file_utils import cached_path, CONFIG_NAME
from .file_utils import CONFIG_NAME, cached_path, is_remote_url, hf_bucket_url
logger = logging.getLogger(__name__)
@@ -49,16 +49,47 @@ class PretrainedConfig(object):
pretrained_config_archive_map = {}
def __init__(self, **kwargs):
self.finetuning_task = kwargs.pop('finetuning_task', None)
self.num_labels = kwargs.pop('num_labels', 2)
# Attributes with defaults
self.output_attentions = kwargs.pop('output_attentions', False)
self.output_hidden_states = kwargs.pop('output_hidden_states', 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', {})
# Is decoder is used in encoder-decoder models to differentiate encoder from decoder
self.is_decoder = kwargs.pop('is_decoder', False)
# Parameters for sequence generation
self.max_length = kwargs.pop('max_length', 20)
self.do_sample = kwargs.pop('do_sample', False)
self.num_beams = kwargs.pop('num_beams', 1)
self.temperature = kwargs.pop('temperature', 1.0)
self.top_k = kwargs.pop('top_k', 50)
self.top_p = kwargs.pop('top_p', 1.0)
self.repetition_penalty = kwargs.pop('repetition_penalty', 1.0)
self.bos_token_id = kwargs.pop('bos_token_id', 0)
self.pad_token_id = kwargs.pop('pad_token_id', 0)
self.eos_token_ids = kwargs.pop('eos_token_ids', 0)
self.length_penalty = kwargs.pop('length_penalty', 1.)
self.num_return_sequences = kwargs.pop('num_return_sequences', 1)
# Fine-tuning task arguments
self.finetuning_task = kwargs.pop('finetuning_task', None)
self.num_labels = kwargs.pop('num_labels', 2)
self.id2label = kwargs.pop('id2label', {i: 'LABEL_{}'.format(i) for i in range(self.num_labels)})
self.id2label = dict((int(key), value) for key, value in self.id2label.items())
self.label2id = kwargs.pop('label2id', dict(zip(self.id2label.values(), self.id2label.keys())))
self.label2id = dict((key, int(value)) for key, value in self.label2id.items())
# Additional attributes without default values
for key, value in kwargs.items():
try:
setattr(self, key, value)
except AttributeError as err:
logger.error("Can't set {} with value {} for {}".format(key, value, self))
raise err
def save_pretrained(self, save_directory):
""" Save a configuration object to the directory `save_directory`, so that it
can be re-loaded using the :func:`~transformers.PretrainedConfig.from_pretrained` class method.
@@ -79,6 +110,7 @@ class PretrainedConfig(object):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model configuration to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model configuration that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing a configuration file saved using the :func:`~transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
- a path or url to a saved configuration JSON `file`, e.g.: ``./my_model_directory/configuration.json``.
@@ -131,12 +163,18 @@ class PretrainedConfig(object):
config_file = cls.pretrained_config_archive_map[pretrained_model_name_or_path]
elif os.path.isdir(pretrained_model_name_or_path):
config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
else:
elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
config_file = pretrained_model_name_or_path
# redirect to the cache, if necessary
else:
config_file = hf_bucket_url(pretrained_model_name_or_path, postfix=CONFIG_NAME)
try:
# Load from URL or cache if already cached
resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download,
proxies=proxies, resume_download=resume_download)
# Load config
config = cls.from_json_file(resolved_config_file)
except EnvironmentError:
if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
msg = "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
@@ -150,15 +188,18 @@ class PretrainedConfig(object):
config_file, CONFIG_NAME)
raise EnvironmentError(msg)
except json.JSONDecodeError:
msg = "Couldn't reach server at '{}' to download configuration file or " \
"configuration file is not a valid JSON file. " \
"Please check network or file content here: {}.".format(config_file, resolved_config_file)
raise EnvironmentError(msg)
if resolved_config_file == config_file:
logger.info("loading configuration file {}".format(config_file))
else:
logger.info("loading configuration file {} from cache at {}".format(
config_file, resolved_config_file))
# Load config
config = cls.from_json_file(resolved_config_file)
if hasattr(config, 'pruned_heads'):
config.pruned_heads = dict((int(key), value) for key, value in config.pruned_heads.items())
@@ -180,17 +221,15 @@ class PretrainedConfig(object):
@classmethod
def from_dict(cls, json_object):
"""Constructs a `Config` from a Python dictionary of parameters."""
config = cls(vocab_size_or_config_json_file=-1)
for key, value in json_object.items():
setattr(config, key, value)
return config
return cls(**json_object)
@classmethod
def from_json_file(cls, json_file):
"""Constructs a `BertConfig` from a json file of parameters."""
"""Constructs a `Config` from a json file of parameters."""
with open(json_file, "r", encoding='utf-8') as reader:
text = reader.read()
return cls.from_dict(json.loads(text))
dict_obj = json.loads(text)
return cls(**dict_obj)
def __eq__(self, other):
return self.__dict__ == other.__dict__

39
transformers/configuration_xlm.py
View File

@@ -42,7 +42,7 @@ class XLMConfig(PretrainedConfig):
"""Configuration class to store the configuration of a `XLMModel`.
Args:
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `XLMModel`.
vocab_size: Vocabulary size of `inputs_ids` in `XLMModel`.
d_model: Size of the encoder layers and the pooler layer.
n_layer: Number of hidden layers in the Transformer encoder.
n_head: Number of attention heads for each attention layer in
@@ -81,7 +81,7 @@ class XLMConfig(PretrainedConfig):
pretrained_config_archive_map = XLM_PRETRAINED_CONFIG_ARCHIVE_MAP
def __init__(self,
vocab_size_or_config_json_file=30145,
vocab_size=30145,
emb_dim=2048,
n_layers=12,
n_heads=16,
@@ -103,9 +103,6 @@ class XLMConfig(PretrainedConfig):
unk_index=3,
mask_index=5,
is_encoder=True,
finetuning_task=None,
num_labels=2,
summary_type='first',
summary_use_proj=True,
summary_activation=None,
@@ -113,19 +110,13 @@ class XLMConfig(PretrainedConfig):
summary_first_dropout=0.1,
start_n_top=5,
end_n_top=5,
mask_token_id=0,
lang_id=0,
**kwargs):
"""Constructs XLMConfig.
"""
super(XLMConfig, self).__init__(**kwargs)
if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
and isinstance(vocab_size_or_config_json_file, unicode)):
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 isinstance(vocab_size_or_config_json_file, int):
self.n_words = vocab_size_or_config_json_file
self.vocab_size = vocab_size
self.emb_dim = emb_dim
self.n_layers = n_layers
self.n_heads = n_heads
@@ -147,8 +138,6 @@ class XLMConfig(PretrainedConfig):
self.max_position_embeddings = max_position_embeddings
self.embed_init_std = embed_init_std
self.init_std = init_std
self.finetuning_task = finetuning_task
self.num_labels = num_labels
self.summary_type = summary_type
self.summary_use_proj = summary_use_proj
self.summary_activation = summary_activation
@@ -156,17 +145,19 @@ class XLMConfig(PretrainedConfig):
self.summary_first_dropout = summary_first_dropout
self.start_n_top = start_n_top
self.end_n_top = end_n_top
else:
raise ValueError("First argument must be either a vocabulary size (int)"
" or the path to a pretrained model config file (str)")
self.mask_token_id = mask_token_id
self.lang_id = lang_id
if "n_words" in kwargs:
self.n_words = kwargs["n_words"]
@property
def vocab_size(self):
return self.n_words
def n_words(self): # For backward compatibility
return self.vocab_size
@vocab_size.setter
def vocab_size(self, value):
self.n_words = value
@n_words.setter
def n_words(self, value): # For backward compatibility
self.vocab_size = value
@property
def hidden_size(self):

38
transformers/configuration_xlm_roberta.py Normal file
View File

@@ -0,0 +1,38 @@
# 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.
""" XLM-RoBERTa configuration """
from __future__ import (absolute_import, division, print_function,
unicode_literals)
import logging
from .configuration_roberta import RobertaConfig
logger = logging.getLogger(__name__)
XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
'xlm-roberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-base-config.json",
'xlm-roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-config.json",
'xlm-roberta-large-finetuned-conll02-dutch': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll02-dutch-config.json",
'xlm-roberta-large-finetuned-conll02-spanish': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll02-spanish-config.json",
'xlm-roberta-large-finetuned-conll03-english': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll03-english-config.json",
'xlm-roberta-large-finetuned-conll03-german': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll03-german-config.json",
}
class XLMRobertaConfig(RobertaConfig):
pretrained_config_archive_map = XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP

35
transformers/configuration_xlnet.py
View File

@@ -35,7 +35,7 @@ class XLNetConfig(PretrainedConfig):
"""Configuration class to store the configuration of a ``XLNetModel``.
Args:
vocab_size_or_config_json_file: Vocabulary size of ``inputs_ids`` in ``XLNetModel``.
vocab_size: Vocabulary size of ``inputs_ids`` in ``XLNetModel``.
d_model: Size of the encoder layers and the pooler layer.
n_layer: Number of hidden layers in the Transformer encoder.
n_head: Number of attention heads for each attention layer in
@@ -72,28 +72,22 @@ class XLNetConfig(PretrainedConfig):
pretrained_config_archive_map = XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP
def __init__(self,
vocab_size_or_config_json_file=32000,
vocab_size=32000,
d_model=1024,
n_layer=24,
n_head=16,
d_inner=4096,
max_position_embeddings=512,
ff_activation="gelu",
untie_r=True,
attn_type="bi",
initializer_range=0.02,
layer_norm_eps=1e-12,
dropout=0.1,
mem_len=None,
reuse_len=None,
bi_data=False,
clamp_len=-1,
same_length=False,
finetuning_task=None,
num_labels=2,
summary_type='last',
summary_use_proj=True,
summary_activation='tanh',
@@ -104,15 +98,7 @@ class XLNetConfig(PretrainedConfig):
"""Constructs XLNetConfig.
"""
super(XLNetConfig, self).__init__(**kwargs)
if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
and isinstance(vocab_size_or_config_json_file, unicode)):
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():
setattr(config, key, value)
elif isinstance(vocab_size_or_config_json_file, int):
self.n_token = vocab_size_or_config_json_file
self.vocab_size = vocab_size
self.d_model = d_model
self.n_layer = n_layer
self.n_head = n_head
@@ -133,29 +119,24 @@ class XLNetConfig(PretrainedConfig):
self.clamp_len = clamp_len
self.same_length = same_length
self.finetuning_task = finetuning_task
self.num_labels = num_labels
self.summary_type = summary_type
self.summary_use_proj = summary_use_proj
self.summary_activation = summary_activation
self.summary_last_dropout = summary_last_dropout
self.start_n_top = start_n_top
self.end_n_top = end_n_top
else:
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 -1
@property
def vocab_size(self):
return self.n_token
def n_token(self): # Backward compatibility
return self.vocab_size
@vocab_size.setter
def vocab_size(self, value):
self.n_token = value
@n_token.setter
def n_token(self, value): # Backward compatibility
self.vocab_size = value
@property
def hidden_size(self):

33
transformers/convert_pytorch_checkpoint_to_tf2.py
View File

@@ -32,9 +32,10 @@ from transformers import (load_pytorch_checkpoint_in_tf2_model,
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,
DistilBertConfig, TFDistilBertForMaskedLM, TFDistilBertForQuestionAnswering, TFDistilBertForSequenceClassification, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
CTRLConfig, TFCTRLLMHeadModel, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,
AlbertConfig, TFAlbertForMaskedLM, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
AlbertConfig, TFAlbertForMaskedLM, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
T5Config, TFT5WithLMHeadModel, T5_PRETRAINED_CONFIG_ARCHIVE_MAP)
if is_torch_available():
import torch
@@ -46,9 +47,10 @@ if is_torch_available():
TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
DistilBertForMaskedLM, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
DistilBertForMaskedLM, DistilBertForQuestionAnswering, DistilBertForSequenceClassification, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP)
else:
(BertForPreTraining, BertForQuestionAnswering, BertForSequenceClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
@@ -57,9 +59,10 @@ else:
TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
DistilBertForMaskedLM, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
DistilBertForMaskedLM, DistilBertForSequenceClassification, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP) = (
AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP) = (
None, None, None, None,
None, None,
None, None,
@@ -67,7 +70,8 @@ else:
None, None,
None, None,
None, None, None,
None, None, None,
None, None, None, None,
None, None,
None, None,
None, None)
@@ -89,8 +93,10 @@ MODEL_CLASSES = {
'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),
'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)
'albert': (AlbertConfig, TFAlbertForMaskedLM, AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
't5': (T5Config, TFT5WithLMHeadModel, T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP, T5_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):
@@ -115,24 +121,21 @@ def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file
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]]
tf_inputs = tf.constant(inputs_list)
tfo = tf_model(tf_inputs, training=False) # build the network
tfo = tf_model(tf_model.dummy_inputs, training=False) # build the network
state_dict = torch.load(pytorch_checkpoint_path, map_location='cpu')
pt_model = pt_model_class.from_pretrained(pretrained_model_name_or_path=None,
config=config,
state_dict=state_dict)
pt_inputs = torch.tensor(inputs_list)
with torch.no_grad():
pto = pt_model(pt_inputs)
pto = pt_model(**pt_model.dummy_inputs)
np_pt = pto[0].detach().numpy()
np_pt = pto[0].numpy()
np_tf = tfo[0].numpy()
diff = np.amax(np.abs(np_pt - np_tf))
print("Max absolute difference between models outputs {}".format(diff))
assert diff <= 2e-2, "Error, model absolute difference is >2e-2"
assert diff <= 2e-2, "Error, model absolute difference is >2e-2: {}".format(diff)
# Save pytorch-model
print("Save TensorFlow model to {}".format(tf_dump_path))

31
transformers/convert_roberta_original_pytorch_checkpoint_to_pytorch.py
View File

@@ -20,6 +20,13 @@ import argparse
import logging
import numpy as np
import torch
import pathlib
import fairseq
from packaging import version
if version.parse(fairseq.__version__) < version.parse("0.9.0"):
raise Exception("requires fairseq >= 0.9.0")
from fairseq.models.roberta import RobertaModel as FairseqRobertaModel
from fairseq.modules import TransformerSentenceEncoderLayer
@@ -45,8 +52,9 @@ def convert_roberta_checkpoint_to_pytorch(roberta_checkpoint_path, pytorch_dump_
"""
roberta = FairseqRobertaModel.from_pretrained(roberta_checkpoint_path)
roberta.eval() # disable dropout
roberta_sent_encoder = roberta.model.decoder.sentence_encoder
config = BertConfig(
vocab_size_or_config_json_file=50265,
vocab_size=roberta_sent_encoder.embed_tokens.num_embeddings,
hidden_size=roberta.args.encoder_embed_dim,
num_hidden_layers=roberta.args.encoder_layers,
num_attention_heads=roberta.args.encoder_attention_heads,
@@ -64,7 +72,6 @@ def convert_roberta_checkpoint_to_pytorch(roberta_checkpoint_path, pytorch_dump_
# Now let's copy all the weights.
# Embeddings
roberta_sent_encoder = roberta.model.decoder.sentence_encoder
model.roberta.embeddings.word_embeddings.weight = roberta_sent_encoder.embed_tokens.weight
model.roberta.embeddings.position_embeddings.weight = roberta_sent_encoder.embed_positions.weight
model.roberta.embeddings.token_type_embeddings.weight.data = torch.zeros_like(model.roberta.embeddings.token_type_embeddings.weight) # just zero them out b/c RoBERTa doesn't use them.
@@ -79,15 +86,18 @@ def convert_roberta_checkpoint_to_pytorch(roberta_checkpoint_path, pytorch_dump_
### self attention
self_attn: BertSelfAttention = layer.attention.self
assert(
roberta_layer.self_attn.in_proj_weight.shape == torch.Size((3 * config.hidden_size, config.hidden_size))
roberta_layer.self_attn.k_proj.weight.data.shape == \
roberta_layer.self_attn.q_proj.weight.data.shape == \
roberta_layer.self_attn.v_proj.weight.data.shape == \
torch.Size((config.hidden_size, config.hidden_size))
)
# we use three distinct linear layers so we split the source layer here.
self_attn.query.weight.data = roberta_layer.self_attn.in_proj_weight[:config.hidden_size, :]
self_attn.query.bias.data = roberta_layer.self_attn.in_proj_bias[:config.hidden_size]
self_attn.key.weight.data = roberta_layer.self_attn.in_proj_weight[config.hidden_size:2*config.hidden_size, :]
self_attn.key.bias.data = roberta_layer.self_attn.in_proj_bias[config.hidden_size:2*config.hidden_size]
self_attn.value.weight.data = roberta_layer.self_attn.in_proj_weight[2*config.hidden_size:, :]
self_attn.value.bias.data = roberta_layer.self_attn.in_proj_bias[2*config.hidden_size:]
self_attn.query.weight.data = roberta_layer.self_attn.q_proj.weight
self_attn.query.bias.data = roberta_layer.self_attn.q_proj.bias
self_attn.key.weight.data = roberta_layer.self_attn.k_proj.weight
self_attn.key.bias.data = roberta_layer.self_attn.k_proj.bias
self_attn.value.weight.data = roberta_layer.self_attn.v_proj.weight
self_attn.value.bias.data = roberta_layer.self_attn.v_proj.bias
### self-attention output
self_output: BertSelfOutput = layer.attention.output
@@ -151,6 +161,7 @@ def convert_roberta_checkpoint_to_pytorch(roberta_checkpoint_path, pytorch_dump_
if not success:
raise Exception("Something went wRoNg")
pathlib.Path(pytorch_dump_folder_path).mkdir(parents=True, exist_ok=True)
print(f"Saving model to {pytorch_dump_folder_path}")
model.save_pretrained(pytorch_dump_folder_path)

65
transformers/convert_t5_original_tf_checkpoint_to_pytorch.py Executable file
View File

@@ -0,0 +1,65 @@
# coding=utf-8
# Copyright 2018 The T5 authors and 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 T5 checkpoint."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import argparse
import torch
from transformers import T5Config, T5Model, load_tf_weights_in_t5
import logging
logging.basicConfig(level=logging.INFO)
def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file, pytorch_dump_path):
# Initialise PyTorch model
config = T5Config.from_json_file(config_file)
print("Building PyTorch model from configuration: {}".format(str(config)))
model = T5Model(config)
# Load weights from tf checkpoint
load_tf_weights_in_t5(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("--config_file",
default = None,
type = str,
required = True,
help = "The config json file corresponding to the pre-trained T5 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.config_file,
args.pytorch_dump_path)

2
transformers/data/__init__.py
View File

@@ -1,4 +1,4 @@
from .processors import InputExample, InputFeatures, DataProcessor, SquadFeatures
from .processors import InputExample, InputFeatures, DataProcessor, SquadFeatures, SingleSentenceClassificationProcessor
from .processors import glue_output_modes, glue_processors, glue_tasks_num_labels, glue_convert_examples_to_features
from .processors import squad_convert_examples_to_features, SquadExample, SquadV1Processor, SquadV2Processor
from .processors import xnli_output_modes, xnli_processors, xnli_tasks_num_labels

2
transformers/data/processors/__init__.py
View File

@@ -1,4 +1,4 @@
from .utils import InputExample, InputFeatures, DataProcessor
from .utils import InputExample, InputFeatures, DataProcessor, SingleSentenceClassificationProcessor
from .glue import glue_output_modes, glue_processors, glue_tasks_num_labels, glue_convert_examples_to_features
from .squad import squad_convert_examples_to_features, SquadFeatures, SquadExample, SquadV1Processor, SquadV2Processor
from .xnli import xnli_output_modes, xnli_processors, xnli_tasks_num_labels

190
transformers/data/processors/squad.py
View File

@@ -18,8 +18,8 @@ if is_tf_available():
logger = logging.getLogger(__name__)
def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer,
orig_answer_text):
def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer, orig_answer_text):
"""Returns tokenized answer spans that better match the annotated answer."""
tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text))
@@ -31,6 +31,7 @@ def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer,
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."""
best_score = None
@@ -50,6 +51,7 @@ def _check_is_max_context(doc_spans, cur_span_index, position):
return cur_span_index == best_span_index
def _new_check_is_max_context(doc_spans, cur_span_index, position):
"""Check if this is the 'max context' doc span for the token."""
# if len(doc_spans) == 1:
@@ -71,14 +73,16 @@ def _new_check_is_max_context(doc_spans, cur_span_index, position):
return cur_span_index == best_span_index
def _is_whitespace(c):
if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F:
return True
return False
def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
doc_stride, max_query_length, is_training,
return_dataset=False):
def squad_convert_examples_to_features(
examples, tokenizer, max_seq_length, doc_stride, max_query_length, is_training, return_dataset=False
):
"""
Converts a list of examples into a list of features that can be directly given as input to a model.
It is model-dependant and takes advantage of many of the tokenizer's features to create the model's inputs.
@@ -116,7 +120,7 @@ def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
unique_id = 1000000000
features = []
for (example_index, example) in enumerate(tqdm(examples)):
for (example_index, example) in enumerate(tqdm(examples, desc="Converting examples to features")):
if is_training and not example.is_impossible:
# Get start and end position
start_position = example.start_position
@@ -129,7 +133,6 @@ def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
logger.warning("Could not find answer: '%s' vs. '%s'", actual_text, cleaned_answer_text)
continue
tok_to_orig_index = []
orig_to_tok_index = []
all_doc_tokens = []
@@ -140,7 +143,6 @@ def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
tok_to_orig_index.append(i)
all_doc_tokens.append(sub_token)
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:
@@ -154,7 +156,9 @@ def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
spans = []
truncated_query = tokenizer.encode(example.question_text, add_special_tokens=False, max_length=max_query_length)
truncated_query = tokenizer.encode(
example.question_text, add_special_tokens=False, max_length=max_query_length
)
sequence_added_tokens = tokenizer.max_len - tokenizer.max_len_single_sentence
sequence_pair_added_tokens = tokenizer.max_len - tokenizer.max_len_sentences_pair
@@ -168,15 +172,18 @@ def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
return_overflowing_tokens=True,
pad_to_max_length=True,
stride=max_seq_length - doc_stride - len(truncated_query) - sequence_pair_added_tokens,
truncation_strategy='only_second' if tokenizer.padding_side == "right" else 'only_first'
truncation_strategy="only_second" if tokenizer.padding_side == "right" else "only_first",
)
paragraph_len = min(len(all_doc_tokens) - len(spans) * doc_stride, max_seq_length - len(truncated_query) - sequence_pair_added_tokens)
paragraph_len = min(
len(all_doc_tokens) - len(spans) * doc_stride,
max_seq_length - len(truncated_query) - sequence_pair_added_tokens,
)
if tokenizer.pad_token_id in encoded_dict['input_ids']:
non_padded_ids = encoded_dict['input_ids'][:encoded_dict['input_ids'].index(tokenizer.pad_token_id)]
if tokenizer.pad_token_id in encoded_dict["input_ids"]:
non_padded_ids = encoded_dict["input_ids"][: encoded_dict["input_ids"].index(tokenizer.pad_token_id)]
else:
non_padded_ids = encoded_dict['input_ids']
non_padded_ids = encoded_dict["input_ids"]
tokens = tokenizer.convert_ids_to_tokens(non_padded_ids)
@@ -202,16 +209,20 @@ def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
for doc_span_index in range(len(spans)):
for j in range(spans[doc_span_index]["paragraph_len"]):
is_max_context = _new_check_is_max_context(spans, doc_span_index, doc_span_index * doc_stride + j)
index = j if tokenizer.padding_side == "left" else spans[doc_span_index]["truncated_query_with_special_tokens_length"] + j
index = (
j
if tokenizer.padding_side == "left"
else spans[doc_span_index]["truncated_query_with_special_tokens_length"] + j
)
spans[doc_span_index]["token_is_max_context"][index] = is_max_context
for span in spans:
# Identify the position of the CLS token
cls_index = span['input_ids'].index(tokenizer.cls_token_id)
cls_index = span["input_ids"].index(tokenizer.cls_token_id)
# 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 = np.array(span['token_type_ids'])
p_mask = np.array(span["token_type_ids"])
p_mask = np.minimum(p_mask, 1)
@@ -224,7 +235,6 @@ def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
# Set the CLS index to '0'
p_mask[cls_index] = 0
span_is_impossible = example.is_impossible
start_position = 0
end_position = 0
@@ -251,51 +261,95 @@ def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
start_position = tok_start_position - doc_start + doc_offset
end_position = tok_end_position - doc_start + doc_offset
features.append(SquadFeatures(
span['input_ids'],
span['attention_mask'],
span['token_type_ids'],
features.append(
SquadFeatures(
span["input_ids"],
span["attention_mask"],
span["token_type_ids"],
cls_index,
p_mask.tolist(),
example_index=example_index,
unique_id=unique_id,
paragraph_len=span['paragraph_len'],
paragraph_len=span["paragraph_len"],
token_is_max_context=span["token_is_max_context"],
tokens=span["tokens"],
token_to_orig_map=span["token_to_orig_map"],
start_position=start_position,
end_position=end_position
))
end_position=end_position,
)
)
unique_id += 1
if return_dataset == 'pt':
if return_dataset == "pt":
if not is_torch_available():
raise ImportError("Pytorch must be installed to return a pytorch dataset.")
# 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.attention_mask for f in features], dtype=torch.long)
all_segment_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
all_attention_masks = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
all_token_type_ids = torch.tensor([f.token_type_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 not is_training:
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)
dataset = TensorDataset(
all_input_ids, all_attention_masks, all_token_type_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)
dataset = TensorDataset(
all_input_ids,
all_attention_masks,
all_token_type_ids,
all_start_positions,
all_end_positions,
all_cls_index,
all_p_mask,
)
return features, dataset
elif return_dataset == "tf":
if not is_tf_available():
raise ImportError("TensorFlow must be installed to return a TensorFlow dataset.")
def gen():
for ex in features:
yield (
{
"input_ids": ex.input_ids,
"attention_mask": ex.attention_mask,
"token_type_ids": ex.token_type_ids,
}, {
"start_position": ex.start_position,
"end_position": ex.end_position,
"cls_index": ex.cls_index,
"p_mask": ex.p_mask,
}
)
return tf.data.Dataset.from_generator(
gen,
(
{"input_ids": tf.int32, "attention_mask": tf.int32, "token_type_ids": tf.int32},
{"start_position": tf.int64, "end_position": tf.int64, "cls_index": tf.int64, "p_mask": tf.int32},
),
(
{
"input_ids": tf.TensorShape([None]),
"attention_mask": tf.TensorShape([None]),
"token_type_ids": tf.TensorShape([None]),
},
{
"start_position": tf.TensorShape([]),
"end_position": tf.TensorShape([]),
"cls_index": tf.TensorShape([]),
"p_mask": tf.TensorShape([None]),
},
),
)
return features
@@ -305,31 +359,32 @@ class SquadProcessor(DataProcessor):
Processor for the SQuAD data set.
Overriden by SquadV1Processor and SquadV2Processor, used by the version 1.1 and version 2.0 of SQuAD, respectively.
"""
train_file = None
dev_file = None
def _get_example_from_tensor_dict(self, tensor_dict, evaluate=False):
if not evaluate:
answer = tensor_dict['answers']['text'][0].numpy().decode('utf-8')
answer_start = tensor_dict['answers']['answer_start'][0].numpy()
answer = tensor_dict["answers"]["text"][0].numpy().decode("utf-8")
answer_start = tensor_dict["answers"]["answer_start"][0].numpy()
answers = []
else:
answers = [{
"answer_start": start.numpy(),
"text": text.numpy().decode('utf-8')
} for start, text in zip(tensor_dict['answers']["answer_start"], tensor_dict['answers']["text"])]
answers = [
{"answer_start": start.numpy(), "text": text.numpy().decode("utf-8")}
for start, text in zip(tensor_dict["answers"]["answer_start"], tensor_dict["answers"]["text"])
]
answer = None
answer_start = None
return SquadExample(
qas_id=tensor_dict['id'].numpy().decode("utf-8"),
question_text=tensor_dict['question'].numpy().decode('utf-8'),
context_text=tensor_dict['context'].numpy().decode('utf-8'),
qas_id=tensor_dict["id"].numpy().decode("utf-8"),
question_text=tensor_dict["question"].numpy().decode("utf-8"),
context_text=tensor_dict["context"].numpy().decode("utf-8"),
answer_text=answer,
start_position_character=answer_start,
title=tensor_dict['title'].numpy().decode('utf-8'),
answers=answers
title=tensor_dict["title"].numpy().decode("utf-8"),
answers=answers,
)
def get_examples_from_dataset(self, dataset, evaluate=False):
@@ -373,10 +428,15 @@ class SquadProcessor(DataProcessor):
which is `train-v1.1.json` and `train-v2.0.json` for squad versions 1.1 and 2.0 respectively.
"""
if data_dir is None:
data_dir = ""
if self.train_file is None:
raise ValueError("SquadProcessor should be instantiated via SquadV1Processor or SquadV2Processor")
with open(os.path.join(data_dir, self.train_file if filename is None else filename), "r", encoding='utf-8') as reader:
with open(
os.path.join(data_dir, self.train_file if filename is None else filename), "r", encoding="utf-8"
) as reader:
input_data = json.load(reader)["data"]
return self._create_examples(input_data, "train")
@@ -389,10 +449,15 @@ class SquadProcessor(DataProcessor):
filename: None by default, specify this if the evaluation file has a different name than the original one
which is `train-v1.1.json` and `train-v2.0.json` for squad versions 1.1 and 2.0 respectively.
"""
if data_dir is None:
data_dir = ""
if self.dev_file is None:
raise ValueError("SquadProcessor should be instantiated via SquadV1Processor or SquadV2Processor")
with open(os.path.join(data_dir, self.dev_file if filename is None else filename), "r", encoding='utf-8') as reader:
with open(
os.path.join(data_dir, self.dev_file if filename is None else filename), "r", encoding="utf-8"
) as reader:
input_data = json.load(reader)["data"]
return self._create_examples(input_data, "dev")
@@ -400,7 +465,7 @@ class SquadProcessor(DataProcessor):
is_training = set_type == "train"
examples = []
for entry in tqdm(input_data):
title = entry['title']
title = entry["title"]
for paragraph in entry["paragraphs"]:
context_text = paragraph["context"]
for qa in paragraph["qas"]:
@@ -418,8 +483,8 @@ class SquadProcessor(DataProcessor):
if not is_impossible:
if is_training:
answer = qa["answers"][0]
answer_text = answer['text']
start_position_character = answer['answer_start']
answer_text = answer["text"]
start_position_character = answer["answer_start"]
else:
answers = qa["answers"]
@@ -431,12 +496,13 @@ class SquadProcessor(DataProcessor):
start_position_character=start_position_character,
title=title,
is_impossible=is_impossible,
answers=answers
answers=answers,
)
examples.append(example)
return examples
class SquadV1Processor(SquadProcessor):
train_file = "train-v1.1.json"
dev_file = "dev-v1.1.json"
@@ -462,7 +528,8 @@ class SquadExample(object):
is_impossible: False by default, set to True if the example has no possible answer.
"""
def __init__(self,
def __init__(
self,
qas_id,
question_text,
context_text,
@@ -470,7 +537,8 @@ class SquadExample(object):
start_position_character,
title,
answers=[],
is_impossible=False):
is_impossible=False,
):
self.qas_id = qas_id
self.question_text = question_text
self.context_text = context_text
@@ -503,7 +571,9 @@ class SquadExample(object):
# Start end end positions only has a value during evaluation.
if start_position_character is not None and not is_impossible:
self.start_position = char_to_word_offset[start_position_character]
self.end_position = char_to_word_offset[start_position_character + len(answer_text) - 1]
self.end_position = char_to_word_offset[
min(start_position_character + len(answer_text) - 1, len(char_to_word_offset) - 1)
]
class SquadFeatures(object):
@@ -531,22 +601,21 @@ class SquadFeatures(object):
end_position: end of the answer token index
"""
def __init__(self,
def __init__(
self,
input_ids,
attention_mask,
token_type_ids,
cls_index,
p_mask,
example_index,
unique_id,
paragraph_len,
token_is_max_context,
tokens,
token_to_orig_map,
start_position,
end_position
end_position,
):
self.input_ids = input_ids
self.attention_mask = attention_mask
@@ -574,6 +643,7 @@ class SquadResult(object):
start_logits: The logits corresponding to the start of the answer
end_logits: The logits corresponding to the end of the answer
"""
def __init__(self, unique_id, start_logits, end_logits, start_top_index=None, end_top_index=None, cls_logits=None):
self.start_logits = start_logits
self.end_logits = end_logits

247
transformers/data/processors/utils.py
View File

@@ -18,6 +18,11 @@ import csv
import sys
import copy
import json
import logging
from ...file_utils import is_tf_available, is_torch_available
logger = logging.getLogger(__name__)
class InputExample(object):
"""
@@ -64,7 +69,7 @@ class InputFeatures(object):
label: Label corresponding to the input
"""
def __init__(self, input_ids, attention_mask, token_type_ids, label):
def __init__(self, input_ids, attention_mask=None, token_type_ids=None, label=None):
self.input_ids = input_ids
self.attention_mask = attention_mask
self.token_type_ids = token_type_ids
@@ -86,34 +91,6 @@ class InputFeatures(object):
class DataProcessor(object):
"""Base class for data converters for sequence classification data sets."""
def get_example_from_tensor_dict(self, tensor_dict):
"""Gets an example from a dict with tensorflow tensors
Args:
tensor_dict: Keys and values should match the corresponding Glue
tensorflow_dataset examples.
"""
raise NotImplementedError()
def get_train_examples(self, data_dir):
"""Gets a collection of `InputExample`s for the train set."""
raise NotImplementedError()
def get_dev_examples(self, data_dir):
"""Gets a collection of `InputExample`s for the dev set."""
raise NotImplementedError()
def get_labels(self):
"""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."""
@@ -125,3 +102,215 @@ class DataProcessor(object):
line = list(unicode(cell, 'utf-8') for cell in line)
lines.append(line)
return lines
class SingleSentenceClassificationProcessor(DataProcessor):
""" Generic processor for a single sentence classification data set."""
def __init__(self, labels=None, examples=None, mode='classification', verbose=False):
self.labels = [] if labels is None else labels
self.examples = [] if examples is None else examples
self.mode = mode
self.verbose = verbose
def __len__(self):
return len(self.examples)
def __getitem__(self, idx):
if isinstance(idx, slice):
return SingleSentenceClassificationProcessor(labels=self.labels,
examples=self.examples[idx])
return self.examples[idx]
@classmethod
def create_from_csv(cls, file_name, split_name='', column_label=0, column_text=1,
column_id=None, skip_first_row=False, **kwargs):
processor = cls(**kwargs)
processor.add_examples_from_csv(file_name,
split_name=split_name,
column_label=column_label,
column_text=column_text,
column_id=column_id,
skip_first_row=skip_first_row,
overwrite_labels=True,
overwrite_examples=True)
return processor
@classmethod
def create_from_examples(cls, texts_or_text_and_labels, labels=None, **kwargs):
processor = cls(**kwargs)
processor.add_examples(texts_or_text_and_labels, labels=labels)
return processor
def add_examples_from_csv(self, file_name, split_name='', column_label=0, column_text=1, column_id=None,
skip_first_row=False, overwrite_labels=False, overwrite_examples=False):
lines = self._read_tsv(file_name)
if skip_first_row:
lines = lines[1:]
texts = []
labels = []
ids = []
for (i, line) in enumerate(lines):
texts.append(line[column_text])
labels.append(line[column_label])
if column_id is not None:
ids.append(line[column_id])
else:
guid = "%s-%s" % (split_name, i) if split_name else "%s" % i
ids.append(guid)
return self.add_examples(texts, labels, ids, overwrite_labels=overwrite_labels, overwrite_examples=overwrite_examples)
def add_examples(self, texts_or_text_and_labels, labels=None, ids=None,
overwrite_labels=False, overwrite_examples=False):
assert labels is None or len(texts_or_text_and_labels) == len(labels)
assert ids is None or len(texts_or_text_and_labels) == len(ids)
if ids is None:
ids = [None] * len(texts_or_text_and_labels)
if labels is None:
labels = [None] * len(texts_or_text_and_labels)
examples = []
added_labels = set()
for (text_or_text_and_label, label, guid) in zip(texts_or_text_and_labels, labels, ids):
if isinstance(text_or_text_and_label, (tuple, list)) and label is None:
text, label = text_or_text_and_label
else:
text = text_or_text_and_label
added_labels.add(label)
examples.append(InputExample(guid=guid, text_a=text, text_b=None, label=label))
# Update examples
if overwrite_examples:
self.examples = examples
else:
self.examples.extend(examples)
# Update labels
if overwrite_labels:
self.labels = list(added_labels)
else:
self.labels = list(set(self.labels).union(added_labels))
return self.examples
def get_features(self,
tokenizer,
max_length=None,
pad_on_left=False,
pad_token=0,
mask_padding_with_zero=True,
return_tensors=None):
"""
Convert examples in a list of ``InputFeatures``
Args:
tokenizer: Instance of a tokenizer that will tokenize the examples
max_length: Maximum example length
task: GLUE task
label_list: List of labels. Can be obtained from the processor using the ``processor.get_labels()`` method
output_mode: String indicating the output mode. Either ``regression`` or ``classification``
pad_on_left: If set to ``True``, the examples will be padded on the left rather than on the right (default)
pad_token: Padding token
mask_padding_with_zero: If set to ``True``, the attention mask will be filled by ``1`` for actual values
and by ``0`` for padded values. If set to ``False``, inverts it (``1`` for padded values, ``0`` for
actual values)
Returns:
If the ``examples`` input is a ``tf.data.Dataset``, will return a ``tf.data.Dataset``
containing the task-specific features. If the input is a list of ``InputExamples``, will return
a list of task-specific ``InputFeatures`` which can be fed to the model.
"""
if max_length is None:
max_length = tokenizer.max_len
label_map = {label: i for i, label in enumerate(self.labels)}
all_input_ids = []
for (ex_index, example) in enumerate(self.examples):
if ex_index % 10000 == 0:
logger.info("Tokenizing example %d", ex_index)
input_ids = tokenizer.encode(
example.text_a,
add_special_tokens=True,
max_length=min(max_length, tokenizer.max_len),
)
all_input_ids.append(input_ids)
batch_length = max(len(input_ids) for input_ids in all_input_ids)
features = []
for (ex_index, (input_ids, example)) in enumerate(zip(all_input_ids, self.examples)):
if ex_index % 10000 == 0:
logger.info("Writing example %d", ex_index)
# The mask has 1 for real tokens and 0 for padding tokens. Only real
# tokens are attended to.
attention_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
# Zero-pad up to the sequence length.
padding_length = batch_length - len(input_ids)
if pad_on_left:
input_ids = ([pad_token] * padding_length) + input_ids
attention_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + attention_mask
else:
input_ids = input_ids + ([pad_token] * padding_length)
attention_mask = attention_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
assert len(input_ids) == batch_length, "Error with input length {} vs {}".format(len(input_ids), batch_length)
assert len(attention_mask) == batch_length, "Error with input length {} vs {}".format(len(attention_mask), batch_length)
if self.mode == "classification":
label = label_map[example.label]
elif self.mode == "regression":
label = float(example.label)
else:
raise ValueError(self.mode)
if ex_index < 5 and self.verbose:
logger.info("*** Example ***")
logger.info("guid: %s" % (example.guid))
logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
logger.info("attention_mask: %s" % " ".join([str(x) for x in attention_mask]))
logger.info("label: %s (id = %d)" % (example.label, label))
features.append(
InputFeatures(input_ids=input_ids,
attention_mask=attention_mask,
label=label))
if return_tensors is None:
return features
elif return_tensors == 'tf':
if not is_tf_available():
raise ImportError("return_tensors set to 'tf' but TensorFlow 2.0 can't be imported")
import tensorflow as tf
def gen():
for ex in features:
yield ({'input_ids': ex.input_ids,
'attention_mask': ex.attention_mask},
ex.label)
dataset = tf.data.Dataset.from_generator(gen,
({'input_ids': tf.int32,
'attention_mask': tf.int32},
tf.int64),
({'input_ids': tf.TensorShape([None]),
'attention_mask': tf.TensorShape([None])},
tf.TensorShape([])))
return dataset
elif return_tensors == 'pt':
if not is_torch_available():
raise ImportError("return_tensors set to 'pt' but PyTorch can't be imported")
import torch
from torch.utils.data import TensorDataset
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
if self.mode == "classification":
all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
elif self.mode == "regression":
all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
dataset = TensorDataset(all_input_ids, all_attention_mask, all_labels)
return dataset
else:
raise ValueError("return_tensors should be one of 'tf' or 'pt'")

113
transformers/file_utils.py
View File

@@ -10,10 +10,9 @@ import json
import logging
import os
import six
import shutil
import tempfile
import fnmatch
from functools import wraps
from functools import partial, wraps
from hashlib import sha256
from io import open
@@ -23,25 +22,37 @@ from botocore.exceptions import ClientError
import requests
from tqdm.auto import tqdm
from contextlib import contextmanager
from . import __version__
from filelock import FileLock
logger = logging.getLogger(__name__) # pylint: disable=invalid-name
try:
os.environ.setdefault('USE_TORCH', 'YES')
if os.environ['USE_TORCH'].upper() in ('1', 'ON', 'YES'):
import torch
_torch_available = True # pylint: disable=invalid-name
logger.info("PyTorch version {} available.".format(torch.__version__))
else:
logger.info("USE_TORCH override through env variable, disabling PyTorch")
_torch_available = False
except ImportError:
_torch_available = False # pylint: disable=invalid-name
try:
os.environ.setdefault('USE_TF', 'YES')
if os.environ['USE_TF'].upper() in ('1', 'ON', 'YES'):
import tensorflow as tf
assert hasattr(tf, '__version__') and int(tf.__version__[0]) >= 2
_tf_available = True # pylint: disable=invalid-name
logger.info("TensorFlow version {} available.".format(tf.__version__))
else:
logger.info("USE_TF override through env variable, disabling Tensorflow")
_tf_available = False
except (ImportError, AssertionError):
_tf_available = False # pylint: disable=invalid-name
try:
import torch
_torch_available = True # pylint: disable=invalid-name
logger.info("PyTorch version {} available.".format(torch.__version__))
except ImportError:
_torch_available = False # pylint: disable=invalid-name
try:
from torch.hub import _get_torch_home
torch_cache_home = _get_torch_home()
@@ -72,11 +83,20 @@ WEIGHTS_NAME = "pytorch_model.bin"
TF2_WEIGHTS_NAME = 'tf_model.h5'
TF_WEIGHTS_NAME = 'model.ckpt'
CONFIG_NAME = "config.json"
MODEL_CARD_NAME = "modelcard.json"
DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
DUMMY_MASK = [[1, 1, 1, 1, 1], [1, 1, 1, 0, 0], [0, 0, 0, 1, 1]]
S3_BUCKET_PREFIX = "https://s3.amazonaws.com/models.huggingface.co/bert"
CLOUDFRONT_DISTRIB_PREFIX = "https://d2ws9o8vfrpkyk.cloudfront.net"
def is_torch_available():
return _torch_available
def is_tf_available():
return _tf_available
if not six.PY2:
@@ -103,12 +123,25 @@ else:
return fn
return docstring_decorator
def is_remote_url(url_or_filename):
parsed = urlparse(url_or_filename)
return parsed.scheme in ('http', 'https', 's3')
def hf_bucket_url(identifier, postfix=None, cdn=False):
endpoint = CLOUDFRONT_DISTRIB_PREFIX if cdn else S3_BUCKET_PREFIX
if postfix is None:
return "/".join((endpoint, identifier))
else:
return "/".join((endpoint, identifier, postfix))
def url_to_filename(url, etag=None):
"""
Convert `url` into a hashed filename in a repeatable way.
If `etag` is specified, append its hash to the url's, delimited
by a period.
If the url ends with .h5 (Keras HDF5 weights) ands '.h5' to the name
If the url ends with .h5 (Keras HDF5 weights) adds '.h5' to the name
so that TF 2.0 can identify it as a HDF5 file
(see https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/tensorflow/python/keras/engine/network.py#L1380)
"""
@@ -153,7 +186,7 @@ def filename_to_url(filename, cache_dir=None):
return url, etag
def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=None, resume_download=False):
def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=None, resume_download=False, user_agent=None):
"""
Given something that might be a URL (or might be a local path),
determine which. If it's a URL, download the file and cache it, and
@@ -163,6 +196,7 @@ def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=N
cache_dir: specify a cache directory to save the file to (overwrite the default cache dir).
force_download: if True, re-dowload the file even if it's already cached in the cache dir.
resume_download: if True, resume the download if incompletly recieved file is found.
user_agent: Optional string or dict that will be appended to the user-agent on remote requests.
"""
if cache_dir is None:
cache_dir = TRANSFORMERS_CACHE
@@ -171,17 +205,15 @@ def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=N
if sys.version_info[0] == 3 and isinstance(cache_dir, Path):
cache_dir = str(cache_dir)
parsed = urlparse(url_or_filename)
if parsed.scheme in ('http', 'https', 's3'):
if is_remote_url(url_or_filename):
# URL, so get it from the cache (downloading if necessary)
return get_from_cache(url_or_filename, cache_dir=cache_dir,
force_download=force_download, proxies=proxies,
resume_download=resume_download)
resume_download=resume_download, user_agent=user_agent)
elif os.path.exists(url_or_filename):
# File, and it exists.
return url_or_filename
elif parsed.scheme == '':
elif urlparse(url_or_filename).scheme == '':
# File, but it doesn't exist.
raise EnvironmentError("file {} not found".format(url_or_filename))
else:
@@ -238,14 +270,26 @@ def s3_get(url, temp_file, proxies=None):
s3_resource.Bucket(bucket_name).download_fileobj(s3_path, temp_file)
def http_get(url, temp_file, proxies=None, resume_size=0):
headers={'Range':'bytes=%d-'%(resume_size,)} if resume_size > 0 else None
def http_get(url, temp_file, proxies=None, resume_size=0, user_agent=None):
ua = "transformers/{}; python/{}".format(__version__, sys.version.split()[0])
if isinstance(user_agent, dict):
ua += "; " + "; ".join(
"{}/{}".format(k, v) for k, v in user_agent.items()
)
elif isinstance(user_agent, six.string_types):
ua += "; "+ user_agent
headers = {
"user-agent": ua
}
if resume_size > 0:
headers['Range'] = 'bytes=%d-' % (resume_size,)
response = requests.get(url, stream=True, proxies=proxies, headers=headers)
if response.status_code == 416: # Range not satisfiable
return
content_length = response.headers.get('Content-Length')
total = resume_size + int(content_length) if content_length is not None else None
progress = tqdm(unit="B", unit_scale=True, total=total, initial=resume_size, desc="Downloading")
progress = tqdm(unit="B", unit_scale=True, total=total, initial=resume_size,
desc="Downloading", disable=bool(logger.level<=logging.INFO))
for chunk in response.iter_content(chunk_size=1024):
if chunk: # filter out keep-alive new chunks
progress.update(len(chunk))
@@ -253,7 +297,7 @@ def http_get(url, temp_file, proxies=None, resume_size=0):
progress.close()
def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag_timeout=10, resume_download=False):
def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag_timeout=10, resume_download=False, user_agent=None):
"""
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.
@@ -291,28 +335,34 @@ def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag
# If we don't have a connection (etag is None) and can't identify the file
# try to get the last downloaded one
if not os.path.exists(cache_path) and etag is None:
matching_files = fnmatch.filter(os.listdir(cache_dir), filename + '.*')
matching_files = list(filter(lambda s: not s.endswith('.json'), matching_files))
matching_files = [
file
for file in fnmatch.filter(os.listdir(cache_dir), filename + '.*')
if not file.endswith('.json') and not file.endswith('.lock')
]
if matching_files:
cache_path = os.path.join(cache_dir, matching_files[-1])
# Prevent parallel downloads of the same file with a lock.
lock_path = cache_path + '.lock'
with FileLock(lock_path):
if resume_download:
incomplete_path = cache_path + '.incomplete'
@contextmanager
def _resumable_file_manager():
with open(incomplete_path,'a+b') as f:
yield f
os.remove(incomplete_path)
temp_file_manager = _resumable_file_manager
if os.path.exists(incomplete_path):
resume_size = os.stat(incomplete_path).st_size
else:
resume_size = 0
else:
temp_file_manager = tempfile.NamedTemporaryFile
temp_file_manager = partial(tempfile.NamedTemporaryFile, dir=cache_dir, delete=False)
resume_size = 0
if not os.path.exists(cache_path) or force_download:
if etag is not None and (not os.path.exists(cache_path) or force_download):
# Download to temporary file, then copy to cache dir once finished.
# Otherwise you get corrupt cache entries if the download gets interrupted.
with temp_file_manager() as temp_file:
@@ -324,16 +374,13 @@ def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag
logger.warn('Warning: resumable downloads are not implemented for "s3://" urls')
s3_get(url, temp_file, proxies=proxies)
else:
http_get(url, temp_file, proxies=proxies, resume_size=resume_size)
http_get(url, temp_file, proxies=proxies, resume_size=resume_size, user_agent=user_agent)
# we are copying the file before closing it, so flush to avoid truncation
temp_file.flush()
# shutil.copyfileobj() starts at the current position, so go to the start
temp_file.seek(0)
logger.info("copying %s to cache at %s", temp_file.name, cache_path)
with open(cache_path, 'wb') as cache_file:
shutil.copyfileobj(temp_file, cache_file)
logger.info("storing %s in cache at %s", url, cache_path)
os.rename(temp_file.name, cache_path)
logger.info("creating metadata file for %s", cache_path)
meta = {'url': url, 'etag': etag}
@@ -344,6 +391,4 @@ def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag
output_string = unicode(output_string, 'utf-8') # The beauty of python 2
meta_file.write(output_string)
logger.info("removing temp file %s", temp_file.name)
return cache_path

3
transformers/hf_api.py
View File

@@ -131,8 +131,9 @@ class HfApi:
# the client still has to specify it when uploading the file.
with open(filepath, "rb") as f:
pf = TqdmProgressFileReader(f)
data = f if pf.total_size > 0 else ""
r = requests.put(urls.write, data=f, headers={
r = requests.put(urls.write, data=data, headers={
"content-type": urls.type,
})
r.raise_for_status()

229
transformers/modelcard.py Normal file
View File

@@ -0,0 +1,229 @@
# 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.
""" Configuration base class and utilities."""
from __future__ import (absolute_import, division, print_function,
unicode_literals)
import copy
import json
import logging
import os
from io import open
from .configuration_auto import ALL_PRETRAINED_CONFIG_ARCHIVE_MAP
from .file_utils import CONFIG_NAME, MODEL_CARD_NAME, WEIGHTS_NAME, TF2_WEIGHTS_NAME, \
cached_path, is_remote_url, hf_bucket_url
logger = logging.getLogger(__name__)
class ModelCard(object):
r""" Model Card class.
Store model card as well as methods for loading/downloading/saving model cards.
Please read the following paper for details and explanation on the sections:
"Model Cards for Model Reporting"
by Margaret Mitchell, Simone Wu,
Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer,
Inioluwa Deborah Raji and Timnit Gebru for the proposal behind model cards.
Link: https://arxiv.org/abs/1810.03993
Note:
A model card can be loaded and saved to disk.
Parameters:
"""
def __init__(self, **kwargs):
# Recomended attributes from https://arxiv.org/abs/1810.03993 (see papers)
self.model_details = kwargs.pop('model_details', {})
self.intended_use = kwargs.pop('intended_use', {})
self.factors = kwargs.pop('factors', {})
self.metrics = kwargs.pop('metrics', {})
self.evaluation_data = kwargs.pop('evaluation_data', {})
self.training_data = kwargs.pop('training_data', {})
self.quantitative_analyses = kwargs.pop('quantitative_analyses', {})
self.ethical_considerations = kwargs.pop('ethical_considerations', {})
self.caveats_and_recommendations = kwargs.pop('caveats_and_recommendations', {})
# Open additional attributes
for key, value in kwargs.items():
try:
setattr(self, key, value)
except AttributeError as err:
logger.error("Can't set {} with value {} for {}".format(key, value, self))
raise err
def save_pretrained(self, save_directory_or_file):
""" Save a model card object to the directory or file `save_directory_or_file`.
"""
if os.path.isdir(save_directory_or_file):
# If we save using the predefined names, we can load using `from_pretrained`
output_model_card_file = os.path.join(save_directory_or_file, MODEL_CARD_NAME)
else:
output_model_card_file = save_directory_or_file
self.to_json_file(output_model_card_file)
logger.info("Model card saved in {}".format(output_model_card_file))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
r""" Instantiate a :class:`~transformers.ModelCard` from a pre-trained model model card.
Parameters:
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model card to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model card that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing a mode card file saved using the :func:`~transformers.ModelCard.save_pretrained` method, e.g.: ``./my_model_directory/``.
- a path or url to a saved model card JSON `file`, e.g.: ``./my_model_directory/modelcard.json``.
cache_dir: (`optional`) string:
Path to a directory in which a downloaded pre-trained model
card should be cached if the standard cache should not be used.
kwargs: (`optional`) dict: key/value pairs with which to update the ModelCard object after loading.
- The values in kwargs of any keys which are model card attributes will be used to override the loaded values.
- Behavior concerning key/value pairs whose keys are *not* model card attributes is controlled by the `return_unused_kwargs` keyword parameter.
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.
find_from_standard_name: (`optional`) boolean, default True:
If the pretrained_model_name_or_path ends with our standard model or config filenames, replace them with our standard modelcard filename.
Can be used to directly feed a model/config url and access the colocated modelcard.
return_unused_kwargs: (`optional`) bool:
- If False, then this function returns just the final model card object.
- If True, then this functions returns a tuple `(model card, unused_kwargs)` where `unused_kwargs` is a dictionary consisting of the key/value pairs whose keys are not model card attributes: ie the part of kwargs which has not been used to update `ModelCard` and is otherwise ignored.
Examples::
modelcard = ModelCard.from_pretrained('bert-base-uncased') # Download model card from S3 and cache.
modelcard = ModelCard.from_pretrained('./test/saved_model/') # E.g. model card was saved using `save_pretrained('./test/saved_model/')`
modelcard = ModelCard.from_pretrained('./test/saved_model/modelcard.json')
modelcard = ModelCard.from_pretrained('bert-base-uncased', output_attention=True, foo=False)
"""
cache_dir = kwargs.pop('cache_dir', None)
proxies = kwargs.pop('proxies', None)
find_from_standard_name = kwargs.pop('find_from_standard_name', True)
return_unused_kwargs = kwargs.pop('return_unused_kwargs', False)
if pretrained_model_name_or_path in ALL_PRETRAINED_CONFIG_ARCHIVE_MAP:
# For simplicity we use the same pretrained url than the configuration files
# but with a different suffix (modelcard.json). This suffix is replaced below.
model_card_file = ALL_PRETRAINED_CONFIG_ARCHIVE_MAP[pretrained_model_name_or_path]
elif os.path.isdir(pretrained_model_name_or_path):
model_card_file = os.path.join(pretrained_model_name_or_path, MODEL_CARD_NAME)
elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
model_card_file = pretrained_model_name_or_path
else:
model_card_file = hf_bucket_url(pretrained_model_name_or_path, postfix=MODEL_CARD_NAME)
if find_from_standard_name or pretrained_model_name_or_path in ALL_PRETRAINED_CONFIG_ARCHIVE_MAP:
model_card_file = model_card_file.replace(CONFIG_NAME, MODEL_CARD_NAME)
model_card_file = model_card_file.replace(WEIGHTS_NAME, MODEL_CARD_NAME)
model_card_file = model_card_file.replace(TF2_WEIGHTS_NAME, MODEL_CARD_NAME)
try:
# Load from URL or cache if already cached
resolved_model_card_file = cached_path(model_card_file, cache_dir=cache_dir, force_download=True,
proxies=proxies, resume_download=False)
if resolved_model_card_file == model_card_file:
logger.info("loading model card file {}".format(model_card_file))
else:
logger.info("loading model card file {} from cache at {}".format(
model_card_file, resolved_model_card_file))
# Load model card
modelcard = cls.from_json_file(resolved_model_card_file)
except EnvironmentError:
if pretrained_model_name_or_path in ALL_PRETRAINED_CONFIG_ARCHIVE_MAP:
logger.warning("Couldn't reach server at '{}' to download model card file.".format(
model_card_file))
else:
logger.warning("Model name '{}' was not found in model name list ({}). " \
"We assumed '{}' was a path or url to a model card 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(ALL_PRETRAINED_CONFIG_ARCHIVE_MAP.keys()),
model_card_file, MODEL_CARD_NAME))
logger.warning("Creating an empty model card.")
# We fall back on creating an empty model card
modelcard = cls()
except json.JSONDecodeError:
logger.warning("Couldn't reach server at '{}' to download model card file or "
"model card file is not a valid JSON file. "
"Please check network or file content here: {}.".format(model_card_file, resolved_model_card_file))
logger.warning("Creating an empty model card.")
# We fall back on creating an empty model card
modelcard = cls()
# Update model card with kwargs if needed
to_remove = []
for key, value in kwargs.items():
if hasattr(modelcard, key):
setattr(modelcard, key, value)
to_remove.append(key)
for key in to_remove:
kwargs.pop(key, None)
logger.info("Model card: %s", str(modelcard))
if return_unused_kwargs:
return modelcard, kwargs
else:
return modelcard
@classmethod
def from_dict(cls, json_object):
"""Constructs a `ModelCard` from a Python dictionary of parameters."""
return cls(**json_object)
@classmethod
def from_json_file(cls, json_file):
"""Constructs a `ModelCard` from a json file of parameters."""
with open(json_file, "r", encoding='utf-8') as reader:
text = reader.read()
dict_obj = json.loads(text)
return cls(**dict_obj)
def __eq__(self, other):
return self.__dict__ == other.__dict__
def __repr__(self):
return str(self.to_json_string())
def to_dict(self):
"""Serializes this instance to a Python dictionary."""
output = copy.deepcopy(self.__dict__)
return output
def to_json_string(self):
"""Serializes this instance to a JSON string."""
return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
def to_json_file(self, json_file_path):
""" Save this instance to a json file."""
with open(json_file_path, "w", encoding='utf-8') as writer:
writer.write(self.to_json_string())

392
transformers/modeling_auto.py
View File

@@ -18,18 +18,31 @@ from __future__ import absolute_import, division, print_function, unicode_litera
import logging
from .modeling_bert import BertModel, BertForMaskedLM, BertForSequenceClassification, BertForQuestionAnswering
from .modeling_openai import OpenAIGPTModel, OpenAIGPTLMHeadModel
from .modeling_gpt2 import GPT2Model, GPT2LMHeadModel
from .modeling_ctrl import CTRLModel, CTRLLMHeadModel
from .modeling_transfo_xl import TransfoXLModel, TransfoXLLMHeadModel
from .modeling_xlnet import XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification, XLNetForQuestionAnswering
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_camembert import CamembertModel, CamembertForMaskedLM, CamembertForSequenceClassification, CamembertForMultipleChoice
from .modeling_albert import AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification, AlbertForQuestionAnswering
from .configuration_auto import (AlbertConfig, BertConfig, CamembertConfig, CTRLConfig,
DistilBertConfig, GPT2Config, OpenAIGPTConfig, RobertaConfig,
TransfoXLConfig, XLMConfig, XLNetConfig, XLMRobertaConfig)
from .modeling_bert import BertModel, BertForMaskedLM, BertForSequenceClassification, BertForQuestionAnswering, \
BertForTokenClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_openai import OpenAIGPTModel, OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_gpt2 import GPT2Model, GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_ctrl import CTRLModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_transfo_xl import TransfoXLModel, TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_xlnet import XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification, XLNetForQuestionAnswering, \
XLNetForTokenClassification, XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_xlm import XLMModel, XLMWithLMHeadModel, XLMForSequenceClassification, XLMForQuestionAnswering, \
XLM_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification, \
RobertaForTokenClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_distilbert import DistilBertModel, DistilBertForQuestionAnswering, DistilBertForMaskedLM, \
DistilBertForSequenceClassification, DistilBertForTokenClassification, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_camembert import CamembertModel, CamembertForMaskedLM, CamembertForSequenceClassification, \
CamembertForMultipleChoice, CamembertForTokenClassification, CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_albert import AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification, \
AlbertForQuestionAnswering, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_t5 import T5Model, T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_xlm_roberta import XLMRobertaModel, XLMRobertaForMaskedLM, XLMRobertaForSequenceClassification, \
XLMRobertaForMultipleChoice, XLMRobertaForTokenClassification, XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_utils import PreTrainedModel, SequenceSummary
@@ -38,21 +51,42 @@ from .file_utils import add_start_docstrings
logger = logging.getLogger(__name__)
ALL_PRETRAINED_MODEL_ARCHIVE_MAP = dict((key, value)
for pretrained_map in [
BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
T5_PRETRAINED_MODEL_ARCHIVE_MAP,
XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
]
for key, value, in pretrained_map.items())
class AutoModel(object):
r"""
:class:`~transformers.AutoModel` is a generic model class
that will be instantiated as one of the base model classes of the library
when created with the `AutoModel.from_pretrained(pretrained_model_name_or_path)`
class method.
or the `AutoModel.from_config(config)` class methods.
The `from_pretrained()` method takes care of returning the correct model class instance
using pattern matching on the `pretrained_model_name_or_path` string.
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 `t5`: T5Model (T5 model)
- contains `distilbert`: DistilBertModel (DistilBERT model)
- contains `albert`: AlbertModel (ALBERT model)
- contains `camembert`: CamembertModel (CamemBERT model)
- contains `xlm-roberta`: XLMRobertaModel (XLM-RoBERTa model)
- contains `roberta`: RobertaModel (RoBERTa model)
- contains `bert`: BertModel (Bert model)
- contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
@@ -66,7 +100,56 @@ class AutoModel(object):
"""
def __init__(self):
raise EnvironmentError("AutoModel is designed to be instantiated "
"using the `AutoModel.from_pretrained(pretrained_model_name_or_path)` method.")
"using the `AutoModel.from_pretrained(pretrained_model_name_or_path)` or "
"`AutoModel.from_config(config)` methods.")
@classmethod
def from_config(cls, config):
r""" Instantiates one of the base model classes of the library
from a configuration.
config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
The model class to instantiate is selected based on the configuration class:
- isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
- isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
- isInstance of `bert` configuration class: BertModel (Bert model)
- isInstance of `openai-gpt` configuration class: OpenAIGPTModel (OpenAI GPT model)
- isInstance of `gpt2` configuration class: GPT2Model (OpenAI GPT-2 model)
- isInstance of `ctrl` configuration class: CTRLModel (Salesforce CTRL model)
- isInstance of `transfo-xl` configuration class: TransfoXLModel (Transformer-XL model)
- isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
- isInstance of `xlm` configuration class: XLMModel (XLM model)
Examples::
config = BertConfig.from_pretrained('bert-base-uncased') # Download configuration from S3 and cache.
model = AutoModel.from_config(config) # E.g. model was saved using `save_pretrained('./test/saved_model/')`
"""
if isinstance(config, DistilBertConfig):
return DistilBertModel(config)
elif isinstance(config, RobertaConfig):
return RobertaModel(config)
elif isinstance(config, BertConfig):
return BertModel(config)
elif isinstance(config, OpenAIGPTConfig):
return OpenAIGPTModel(config)
elif isinstance(config, GPT2Config):
return GPT2Model(config)
elif isinstance(config, TransfoXLConfig):
return TransfoXLModel(config)
elif isinstance(config, XLNetConfig):
return XLNetModel(config)
elif isinstance(config, XLMConfig):
return XLMModel(config)
elif isinstance(config, CTRLConfig):
return CTRLModel(config)
elif isinstance(config, AlbertConfig):
return AlbertModel(config)
elif isinstance(config, CamembertConfig):
return CamembertModel(config)
elif isinstance(config, XLMRobertaConfig):
return XLMRobertaModel(config)
raise ValueError("Unrecognized configuration class {}".format(config))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -75,9 +158,11 @@ 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 `t5`: T5Model (T5 model)
- contains `distilbert`: DistilBertModel (DistilBERT model)
- contains `albert`: AlbertModel (ALBERT model)
- contains `camembert`: CamembertModel (CamemBERT model)
- contains `xlm-roberta`: XLMRobertaModel (XLM-RoBERTa model)
- contains `roberta`: RobertaModel (RoBERTa model)
- contains `bert`: BertModel (Bert model)
- contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
@@ -94,6 +179,7 @@ class AutoModel(object):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
- 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.
@@ -146,12 +232,16 @@ class AutoModel(object):
model = AutoModel.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
"""
if 'distilbert' in pretrained_model_name_or_path:
if 't5' in pretrained_model_name_or_path:
return T5Model.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'distilbert' in pretrained_model_name_or_path:
return DistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'albert' in pretrained_model_name_or_path:
return AlbertModel.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 'xlm-roberta' in pretrained_model_name_or_path:
return XLMRobertaModel.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:
@@ -170,7 +260,7 @@ class AutoModel(object):
return CTRLModel.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, 'ctrl', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
"'xlm-roberta', 'xlm', 'roberta, 'ctrl', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
class AutoModelWithLMHead(object):
@@ -185,9 +275,11 @@ 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 `t5`: T5ModelWithLMHead (T5 model)
- contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
- contains `albert`: AlbertForMaskedLM (ALBERT model)
- contains `camembert`: CamembertForMaskedLM (CamemBERT model)
- contains `xlm-roberta`: XLMRobertaForMaskedLM (XLM-RoBERTa model)
- contains `roberta`: RobertaForMaskedLM (RoBERTa model)
- contains `bert`: BertForMaskedLM (Bert model)
- contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -201,7 +293,52 @@ class AutoModelWithLMHead(object):
"""
def __init__(self):
raise EnvironmentError("AutoModelWithLMHead is designed to be instantiated "
"using the `AutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
"using the `AutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` or "
"`AutoModelWithLMHead.from_config(config)` methods.")
@classmethod
def from_config(cls, config):
r""" Instantiates one of the base model classes of the library
from a configuration.
config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
The model class to instantiate is selected based on the configuration class:
- isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
- isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
- isInstance of `bert` configuration class: BertModel (Bert model)
- isInstance of `openai-gpt` configuration class: OpenAIGPTModel (OpenAI GPT model)
- isInstance of `gpt2` configuration class: GPT2Model (OpenAI GPT-2 model)
- isInstance of `ctrl` configuration class: CTRLModel (Salesforce CTRL model)
- isInstance of `transfo-xl` configuration class: TransfoXLModel (Transformer-XL model)
- isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
- isInstance of `xlm` configuration class: XLMModel (XLM model)
Examples::
config = BertConfig.from_pretrained('bert-base-uncased') # Download configuration from S3 and cache.
model = AutoModelWithLMHead.from_config(config) # E.g. model was saved using `save_pretrained('./test/saved_model/')`
"""
if isinstance(config, DistilBertConfig):
return DistilBertForMaskedLM(config)
elif isinstance(config, RobertaConfig):
return RobertaForMaskedLM(config)
elif isinstance(config, BertConfig):
return BertForMaskedLM(config)
elif isinstance(config, OpenAIGPTConfig):
return OpenAIGPTLMHeadModel(config)
elif isinstance(config, GPT2Config):
return GPT2LMHeadModel(config)
elif isinstance(config, TransfoXLConfig):
return TransfoXLLMHeadModel(config)
elif isinstance(config, XLNetConfig):
return XLNetLMHeadModel(config)
elif isinstance(config, XLMConfig):
return XLMWithLMHeadModel(config)
elif isinstance(config, CTRLConfig):
return CTRLLMHeadModel(config)
elif isinstance(config, XLMRobertaConfig):
return XLMRobertaForMaskedLM(config)
raise ValueError("Unrecognized configuration class {}".format(config))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -213,9 +350,11 @@ 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 `t5`: T5ModelWithLMHead (T5 model)
- contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
- contains `albert`: AlbertForMaskedLM (ALBERT model)
- contains `camembert`: CamembertForMaskedLM (CamemBERT model)
- contains `xlm-roberta`: XLMRobertaForMaskedLM (XLM-RoBERTa model)
- contains `roberta`: RobertaForMaskedLM (RoBERTa model)
- contains `bert`: BertForMaskedLM (Bert model)
- contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -232,6 +371,7 @@ class AutoModelWithLMHead(object):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
- 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.
@@ -283,12 +423,16 @@ class AutoModelWithLMHead(object):
model = AutoModelWithLMHead.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
"""
if 'distilbert' in pretrained_model_name_or_path:
if 't5' in pretrained_model_name_or_path:
return T5WithLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'distilbert' in pretrained_model_name_or_path:
return DistilBertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'albert' in pretrained_model_name_or_path:
return AlbertForMaskedLM.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 'xlm-roberta' in pretrained_model_name_or_path:
return XLMRobertaForMaskedLM.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:
@@ -307,7 +451,7 @@ class AutoModelWithLMHead(object):
return CTRLLMHeadModel.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','ctrl', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
"'xlm-roberta', 'xlm', 'roberta','ctrl', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
class AutoModelForSequenceClassification(object):
@@ -325,6 +469,7 @@ class AutoModelForSequenceClassification(object):
- contains `distilbert`: DistilBertForSequenceClassification (DistilBERT model)
- contains `albert`: AlbertForSequenceClassification (ALBERT model)
- contains `camembert`: CamembertForSequenceClassification (CamemBERT model)
- contains `xlm-roberta`: XLMRobertaForSequenceClassification (XLM-RoBERTa model)
- contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
- contains `bert`: BertForSequenceClassification (Bert model)
- contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -333,8 +478,45 @@ class AutoModelForSequenceClassification(object):
This class cannot be instantiated using `__init__()` (throws an error).
"""
def __init__(self):
raise EnvironmentError("AutoModelWithLMHead is designed to be instantiated "
"using the `AutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
raise EnvironmentError("AutoModelForSequenceClassification is designed to be instantiated "
"using the `AutoModelForSequenceClassification.from_pretrained(pretrained_model_name_or_path)` or "
"`AutoModelForSequenceClassification.from_config(config)` methods.")
@classmethod
def from_config(cls, config):
r""" Instantiates one of the base model classes of the library
from a configuration.
config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
The model class to instantiate is selected based on the configuration class:
- isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
- isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
- isInstance of `bert` configuration class: BertModel (Bert model)
- isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
- isInstance of `xlm` configuration class: XLMModel (XLM model)
Examples::
config = BertConfig.from_pretrained('bert-base-uncased') # Download configuration from S3 and cache.
model = AutoModelForSequenceClassification.from_config(config) # E.g. model was saved using `save_pretrained('./test/saved_model/')`
"""
if isinstance(config, AlbertConfig):
return AlbertForSequenceClassification(config)
elif isinstance(config, CamembertConfig):
return CamembertForSequenceClassification(config)
elif isinstance(config, DistilBertConfig):
return DistilBertForSequenceClassification(config)
elif isinstance(config, RobertaConfig):
return RobertaForSequenceClassification(config)
elif isinstance(config, BertConfig):
return BertForSequenceClassification(config)
elif isinstance(config, XLNetConfig):
return XLNetForSequenceClassification(config)
elif isinstance(config, XLMConfig):
return XLMForSequenceClassification(config)
elif isinstance(config, XLMRobertaConfig):
return XLMRobertaForSequenceClassification(config)
raise ValueError("Unrecognized configuration class {}".format(config))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -349,6 +531,7 @@ class AutoModelForSequenceClassification(object):
- contains `distilbert`: DistilBertForSequenceClassification (DistilBERT model)
- contains `albert`: AlbertForSequenceClassification (ALBERT model)
- contains `camembert`: CamembertForSequenceClassification (CamemBERT model)
- contains `xlm-roberta`: XLMRobertaForSequenceClassification (XLM-RoBERTa model)
- contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
- contains `bert`: BertForSequenceClassification (Bert model)
- contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -361,6 +544,7 @@ class AutoModelForSequenceClassification(object):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
- 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.
@@ -419,6 +603,8 @@ class AutoModelForSequenceClassification(object):
return AlbertForSequenceClassification.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 'xlm-roberta' in pretrained_model_name_or_path:
return XLMRobertaForSequenceClassification.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:
@@ -429,7 +615,7 @@ class AutoModelForSequenceClassification(object):
return XLMForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
raise ValueError("Unrecognized model identifier in {}. Should contains one of "
"'bert', 'xlnet', 'xlm', 'roberta', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
"'bert', 'xlnet', 'xlm-roberta', 'xlm', 'roberta', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
class AutoModelForQuestionAnswering(object):
@@ -453,8 +639,38 @@ class AutoModelForQuestionAnswering(object):
This class cannot be instantiated using `__init__()` (throws an error).
"""
def __init__(self):
raise EnvironmentError("AutoModelWithLMHead is designed to be instantiated "
"using the `AutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
raise EnvironmentError("AutoModelForQuestionAnswering is designed to be instantiated "
"using the `AutoModelForQuestionAnswering.from_pretrained(pretrained_model_name_or_path)` or "
"`AutoModelForQuestionAnswering.from_config(config)` methods.")
@classmethod
def from_config(cls, config):
r""" Instantiates one of the base model classes of the library
from a configuration.
config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
The model class to instantiate is selected based on the configuration class:
- isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
- isInstance of `bert` configuration class: BertModel (Bert model)
- isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
- isInstance of `xlm` configuration class: XLMModel (XLM model)
Examples::
config = BertConfig.from_pretrained('bert-base-uncased') # Download configuration from S3 and cache.
model = AutoModelForSequenceClassification.from_config(config) # E.g. model was saved using `save_pretrained('./test/saved_model/')`
"""
if isinstance(config, AlbertConfig):
return AlbertForQuestionAnswering(config)
elif isinstance(config, DistilBertConfig):
return DistilBertForQuestionAnswering(config)
elif isinstance(config, BertConfig):
return BertForQuestionAnswering(config)
elif isinstance(config, XLNetConfig):
return XLNetForQuestionAnswering(config)
elif isinstance(config, XLMConfig):
return XLMForQuestionAnswering(config)
raise ValueError("Unrecognized configuration class {}".format(config))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -479,6 +695,7 @@ class AutoModelForQuestionAnswering(object):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
- 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.
@@ -541,3 +758,130 @@ class AutoModelForQuestionAnswering(object):
raise ValueError("Unrecognized model identifier in {}. Should contains one of "
"'bert', 'xlnet', 'xlm', 'distilbert', 'albert'".format(pretrained_model_name_or_path))
class AutoModelForTokenClassification:
def __init__(self):
raise EnvironmentError("AutoModelForTokenClassification is designed to be instantiated "
"using the `AutoModelForTokenClassification.from_pretrained(pretrained_model_name_or_path)` or "
"`AutoModelForTokenClassification.from_config(config)` methods.")
@classmethod
def from_config(cls, config):
r""" Instantiates one of the base model classes of the library
from a configuration.
config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
The model class to instantiate is selected based on the configuration class:
- isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
- isInstance of `bert` configuration class: BertModel (Bert model)
- isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
- isInstance of `camembert` configuration class: CamembertModel (Camembert model)
- isInstance of `roberta` configuration class: RobertaModel (Roberta model)
Examples::
config = BertConfig.from_pretrained('bert-base-uncased') # Download configuration from S3 and cache.
model = AutoModelForTokenClassification.from_config(config) # E.g. model was saved using `save_pretrained('./test/saved_model/')`
"""
if isinstance(config, CamembertConfig):
return CamembertForTokenClassification(config)
elif isinstance(config, DistilBertConfig):
return DistilBertForTokenClassification(config)
elif isinstance(config, BertConfig):
return BertForTokenClassification(config)
elif isinstance(config, XLNetConfig):
return XLNetForTokenClassification(config)
elif isinstance(config, RobertaConfig):
return RobertaForTokenClassification(config)
elif isinstance(config, XLMRobertaConfig):
return XLMRobertaForTokenClassification(config)
raise ValueError("Unrecognized configuration class {}".format(config))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
r""" Instantiates one of the question answering model classes of the library
from a pre-trained model configuration.
The `from_pretrained()` method takes care of returning the correct model class instance
using pattern matching on the `pretrained_model_name_or_path` string.
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`: DistilBertForTokenClassification (DistilBERT model)
- contains `camembert`: CamembertForTokenClassification (Camembert model)
- contains `bert`: BertForTokenClassification (Bert model)
- contains `xlnet`: XLNetForTokenClassification (XLNet model)
- contains `roberta`: RobertaForTokenClassification (Roberta model)
The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
To train the model, you should first set it back in training mode with `model.train()`
Params:
pretrained_model_name_or_path: 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/``.
- 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.
Examples::
model = AutoModelForTokenClassification.from_pretrained('bert-base-uncased') # Download model and configuration from S3 and cache.
model = AutoModelForTokenClassification.from_pretrained('./test/bert_model/') # E.g. model was saved using `save_pretrained('./test/saved_model/')`
model = AutoModelForTokenClassification.from_pretrained('bert-base-uncased', output_attention=True) # Update configuration during loading
assert model.config.output_attention == True
# Loading from a TF checkpoint file instead of a PyTorch model (slower)
config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
model = AutoModelForTokenClassification.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
"""
if 'camembert' in pretrained_model_name_or_path:
return CamembertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'distilbert' in pretrained_model_name_or_path:
return DistilBertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'xlm-roberta' in pretrained_model_name_or_path:
return XLMRobertaForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'roberta' in pretrained_model_name_or_path:
return RobertaForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'bert' in pretrained_model_name_or_path:
return BertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'xlnet' in pretrained_model_name_or_path:
return XLNetForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
raise ValueError("Unrecognized model identifier in {}. Should contains one of "
"'bert', 'xlnet', 'camembert', 'distilbert', 'xlm-roberta', 'roberta'".format(pretrained_model_name_or_path))

6
transformers/modeling_bert.py
View File

@@ -48,6 +48,12 @@ BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
'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",
'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-pytorch_model.bin",
'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-pytorch_model.bin",
'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-pytorch_model.bin",
'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-pytorch_model.bin",
'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/pytorch_model.bin",
'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/pytorch_model.bin",
}

4
transformers/modeling_ctrl.py
View File

@@ -268,7 +268,7 @@ class CTRLModel(CTRLPreTrainedModel):
tokenizer = CTRLTokenizer.from_pretrained('ctrl')
model = CTRLModel.from_pretrained('ctrl')
input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).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
@@ -458,7 +458,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
tokenizer = CTRLTokenizer.from_pretrained('ctrl')
model = CTRLLMHeadModel.from_pretrained('ctrl')
input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
outputs = model(input_ids, labels=input_ids)
loss, logits = outputs[:2]

8
transformers/modeling_distilbert.py
View File

@@ -415,7 +415,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
model = DistilBertModel.from_pretrained('distilbert-base-uncased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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
@@ -511,7 +511,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
model = DistilBertForMaskedLM.from_pretrained('distilbert-base-uncased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
outputs = model(input_ids, masked_lm_labels=input_ids)
loss, prediction_scores = outputs[:2]
@@ -581,7 +581,7 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
model = DistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
labels = torch.tensor([1]).unsqueeze(0) # Batch size 1
outputs = model(input_ids, labels=labels)
loss, logits = outputs[:2]
@@ -656,7 +656,7 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
model = DistilBertForQuestionAnswering.from_pretrained('distilbert-base-uncased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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)

102
transformers/modeling_encoder_decoder.py
View File

@@ -18,9 +18,11 @@ from __future__ import absolute_import, division, print_function, unicode_litera
import logging
import os
import warnings
import torch
from torch import nn
from tqdm import trange
from .modeling_auto import AutoModel, AutoModelWithLMHead
@@ -59,12 +61,14 @@ class PreTrainedEncoderDecoder(nn.Module):
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 string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- 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 string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- 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.
@@ -117,8 +121,7 @@ class PreTrainedEncoderDecoder(nn.Module):
kwargs_common = {
argument: value
for argument, value in kwargs.items()
if not argument.startswith("encoder_")
and not argument.startswith("decoder_")
if not argument.startswith("encoder_") and not argument.startswith("decoder_")
}
kwargs_decoder = kwargs_common.copy()
kwargs_encoder = kwargs_common.copy()
@@ -164,7 +167,39 @@ class PreTrainedEncoderDecoder(nn.Module):
We save the encoder' and decoder's parameters in two separate directories.
"""
# If the root output directory does not exist, create it
if not os.path.exists(save_directory):
os.mkdir(save_directory)
# Check whether the output directory is empty or not
sub_directories = [directory for directory in os.listdir(save_directory)
if os.path.isdir(os.path.join(save_directory, directory))]
if len(sub_directories) > 0:
if "encoder" in sub_directories and "decoder" in sub_directories:
print("WARNING: there is an older version of encoder-decoder saved in" +\
" the output directory. The default behaviour is to overwrite them.")
# Empty the output directory
for directory_to_remove in sub_directories:
# Remove all files into the subdirectory
files_to_remove = os.listdir(os.path.join(save_directory, directory_to_remove))
for file_to_remove in files_to_remove:
os.remove(os.path.join(save_directory, directory_to_remove, file_to_remove))
# Remove the subdirectory itself
os.rmdir(os.path.join(save_directory, directory_to_remove))
assert(len(os.listdir(save_directory)) == 0) # sanity check
# Create the "encoder" directory inside the output directory and save the encoder into it
if not os.path.exists(os.path.join(save_directory, "encoder")):
os.mkdir(os.path.join(save_directory, "encoder"))
self.encoder.save_pretrained(os.path.join(save_directory, "encoder"))
# Create the "encoder" directory inside the output directory and save the decoder into it
if not os.path.exists(os.path.join(save_directory, "decoder")):
os.mkdir(os.path.join(save_directory, "decoder"))
self.decoder.save_pretrained(os.path.join(save_directory, "decoder"))
def forward(self, encoder_input_ids, decoder_input_ids, **kwargs):
@@ -186,51 +221,56 @@ class PreTrainedEncoderDecoder(nn.Module):
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_encoder, kwargs_decoder = self.prepare_model_kwargs(**kwargs)
# 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]
else:
encoder_outputs = ()
kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
decoder_outputs = self.decoder(decoder_input_ids, encoder_hidden_states, **kwargs_decoder)
return decoder_outputs + encoder_outputs
@staticmethod
def prepare_model_kwargs(**kwargs):
""" Prepare the encoder and decoder's keyword arguments.
Keyword arguments come in 3 flavors:
- encoder-specific (prefixed by `encoder_`)
- decoder-specific (prefixed by `decoder_`)
- 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_")
if not argument.startswith("encoder_") and not argument.startswith("decoder_")
}
kwargs_decoder = kwargs_common.copy()
kwargs_encoder = kwargs_common.copy()
kwargs_encoder.update(
decoder_kwargs = kwargs_common.copy()
encoder_kwargs = kwargs_common.copy()
encoder_kwargs.update(
{
argument[len("encoder_") :]: value
for argument, value in kwargs.items()
if argument.startswith("encoder_")
}
)
kwargs_decoder.update(
decoder_kwargs.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
decoder_kwargs["encoder_attention_mask"] = encoder_kwargs.get("attention_mask", None)
return encoder_kwargs, decoder_kwargs
class Model2Model(PreTrainedEncoderDecoder):

5
transformers/modeling_gpt2.py
View File

@@ -345,7 +345,7 @@ class GPT2Model(GPT2PreTrainedModel):
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
model = GPT2Model.from_pretrained('gpt2')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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
@@ -523,7 +523,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
model = GPT2LMHeadModel.from_pretrained('gpt2')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
outputs = model(input_ids, labels=input_ids)
loss, logits = outputs[:2]
@@ -634,6 +634,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
"""
def __init__(self, config):
super(GPT2DoubleHeadsModel, self).__init__(config)
config.num_labels = 1
self.transformer = GPT2Model(config)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
self.multiple_choice_head = SequenceSummary(config)

5
transformers/modeling_openai.py
View File

@@ -349,7 +349,7 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
model = OpenAIGPTModel.from_pretrained('openai-gpt')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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
@@ -491,7 +491,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
model = OpenAIGPTLMHeadModel.from_pretrained('openai-gpt')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
outputs = model(input_ids, labels=input_ids)
loss, logits = outputs[:2]
@@ -590,6 +590,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
def __init__(self, config):
super(OpenAIGPTDoubleHeadsModel, self).__init__(config)
config.num_labels = 1
self.transformer = OpenAIGPTModel(config)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
self.multiple_choice_head = SequenceSummary(config)

50
transformers/modeling_roberta.py
View File

@@ -51,24 +51,44 @@ class RobertaEmbeddings(BertEmbeddings):
padding_idx=self.padding_idx)
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=device)
position_ids = position_ids.unsqueeze(0).expand(input_shape)
if input_ids is not None:
# Create the position ids from the input token ids. Any padded tokens remain padded.
position_ids = self.create_position_ids_from_input_ids(input_ids).to(input_ids.device)
else:
position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
return super(RobertaEmbeddings, self).forward(input_ids,
token_type_ids=token_type_ids,
position_ids=position_ids,
inputs_embeds=inputs_embeds)
def create_position_ids_from_input_ids(self, x):
""" Replace non-padding symbols with their position numbers. Position numbers begin at
padding_idx+1. Padding symbols are ignored. This is modified from fairseq's
`utils.make_positions`.
:param torch.Tensor x:
:return torch.Tensor:
"""
mask = x.ne(self.padding_idx).long()
incremental_indicies = torch.cumsum(mask, dim=1) * mask
return incremental_indicies + self.padding_idx
def create_position_ids_from_inputs_embeds(self, inputs_embeds):
""" We are provided embeddings directly. We cannot infer which are padded so just generate
sequential position ids.
:param torch.Tensor inputs_embeds:
:return torch.Tensor:
"""
input_shape = inputs_embeds.size()[:-1]
sequence_length = input_shape[1]
position_ids = torch.arange(self.padding_idx+1, sequence_length+self.padding_idx+1, dtype=torch.long,
device=inputs_embeds.device)
return position_ids.unsqueeze(0).expand(input_shape)
ROBERTA_START_DOCSTRING = r""" The RoBERTa model was proposed in
`RoBERTa: A Robustly Optimized BERT Pretraining Approach`_
@@ -168,7 +188,7 @@ class RobertaModel(BertModel):
tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
model = RobertaModel.from_pretrained('roberta-base')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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
@@ -216,7 +236,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
model = RobertaForMaskedLM.from_pretrained('roberta-base')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
outputs = model(input_ids, masked_lm_labels=input_ids)
loss, prediction_scores = outputs[:2]
@@ -307,7 +327,7 @@ class RobertaForSequenceClassification(BertPreTrainedModel):
tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
model = RobertaForSequenceClassification.from_pretrained('roberta-base')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
labels = torch.tensor([1]).unsqueeze(0) # Batch size 1
outputs = model(input_ids, labels=labels)
loss, logits = outputs[:2]

886
transformers/modeling_t5.py Normal file
View File

@@ -0,0 +1,886 @@
# coding=utf-8
# Copyright 2018 Mesh TensorFlow authors, T5 Authors and 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 T5 model. """
from __future__ import absolute_import, division, print_function, unicode_literals
import json
import logging
import math
import os
import sys
import copy
import itertools
from io import open
import torch
from torch import nn
import torch.nn.functional as F
from torch.nn import CrossEntropyLoss, MSELoss
from .modeling_utils import PreTrainedModel, prune_linear_layer
from .configuration_t5 import T5Config
from .file_utils import add_start_docstrings, DUMMY_INPUTS, DUMMY_MASK
logger = logging.getLogger(__name__)
####################################################
# This dict contrains shortcut names and associated url
# for the pretrained weights provided with the models
####################################################
T5_PRETRAINED_MODEL_ARCHIVE_MAP = {
't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-small-pytorch_model.bin",
't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-base-pytorch_model.bin",
't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-large-pytorch_model.bin",
't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-3b-pytorch_model.bin",
't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-11b-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_t5(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 = []
tf_weights = {}
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)
tf_weights[name] = array
for txt_name in names:
name = txt_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)))
tf_weights.pop(txt_name, None)
continue
if '_slot_' in name[-1]:
logger.info("Skipping {}".format("/".join(name)))
tf_weights.pop(txt_name, None)
continue
pointer = model
array = tf_weights[txt_name]
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] in ['kernel', 'scale', 'embedding']:
pointer = getattr(pointer, 'weight')
# elif l[0] == 'scale':
# pointer = getattr(pointer, 'weight')
# elif l[0] == 'output_bias' or l[0] == 'beta':
# pointer = getattr(pointer, 'bias')
# 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 l[0] not in ['kernel', 'scale', 'embedding']:
pointer = getattr(pointer, 'weight')
if l[0] != 'embedding':
logger.info("Transposing numpy weight of shape {} for {}".format(array.shape, name))
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.astype(np.float32))
tf_weights.pop(txt_name, None)
logger.info("Weights not copied to PyTorch model: {}".format(', '.join(tf_weights.keys())))
# logger.info("Weights not copied to PyTorch model: {}".format(', '.join(tf_weights.keys())))
return model
####################################################
# PyTorch Models are constructed by sub-classing
# - torch.nn.Module for the layers and
# - PreTrainedModel for the models (it-self a sub-class of torch.nn.Module)
####################################################
class T5LayerNorm(nn.Module):
def __init__(self, hidden_size, eps=1e-6):
""" Construct a layernorm module in the T5 style
No bias and no substraction of mean.
"""
super(T5LayerNorm, self).__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.variance_epsilon = eps
def forward(self, x):
variance = x.pow(2).mean(-1, keepdim=True)
x = x / torch.sqrt(variance + self.variance_epsilon)
return self.weight * x
class T5DenseReluDense(nn.Module):
def __init__(self, config):
super(T5DenseReluDense, self).__init__()
self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)
self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
self.dropout = nn.Dropout(config.dropout_rate)
def forward(self, hidden_states):
h = self.wi(hidden_states)
h = F.relu(h)
h = self.dropout(h)
h = self.wo(h)
return h
class T5LayerFF(nn.Module):
def __init__(self, config):
super(T5LayerFF, self).__init__()
self.DenseReluDense = T5DenseReluDense(config)
self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
self.dropout = nn.Dropout(config.dropout_rate)
def forward(self, hidden_states):
norm_x = self.layer_norm(hidden_states)
y = self.DenseReluDense(norm_x)
layer_output = hidden_states + self.dropout(y)
return layer_output
class T5Attention(nn.Module):
NEW_ID = itertools.count()
def __init__(self, config, has_relative_attention_bias=False):
super(T5Attention, self).__init__()
self.layer_id = next(T5Attention.NEW_ID)
self.is_decoder = config.is_decoder
self.has_relative_attention_bias = has_relative_attention_bias
self.output_attentions = config.output_attentions
self.relative_attention_num_buckets = config.relative_attention_num_buckets
self.d_model = config.d_model
self.d_kv = config.d_kv
self.n_heads = config.num_heads
self.dropout = config.dropout_rate
self.inner_dim = self.n_heads * self.d_kv
# Mesh TensorFlow initialization to avoid scaling before softmax
self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)
if self.has_relative_attention_bias:
self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
self.pruned_heads = set()
def prune_heads(self, heads):
if len(heads) == 0:
return
mask = torch.ones(self.n_heads, self.d_kv)
heads = set(heads) - self.pruned_heads
for head in heads:
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.q = prune_linear_layer(self.q, index)
self.k = prune_linear_layer(self.k, index)
self.v = prune_linear_layer(self.v, index)
self.o = prune_linear_layer(self.o, index, dim=1)
# Update hyper params
self.n_heads = self.n_heads - len(heads)
self.inner_dim = self.d_kv * self.n_heads
self.pruned_heads = self.pruned_heads.union(heads)
@staticmethod
def _relative_position_bucket(relative_position,
bidirectional=True,
num_buckets=32,
max_distance=128):
"""
Adapted from Mesh Tensorflow:
https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
Translate relative position to a bucket number for relative attention.
The relative position is defined as memory_position - query_position, i.e.
the distance in tokens from the attending position to the attended-to
position. If bidirectional=False, then positive relative positions are
invalid.
We use smaller buckets for small absolute relative_position and larger buckets
for larger absolute relative_positions. All relative positions >=max_distance
map to the same bucket. All relative positions <=-max_distance map to the
same bucket. This should allow for more graceful generalization to longer
sequences than the model has been trained on.
Args:
relative_position: an int32 Tensor
bidirectional: a boolean - whether the attention is bidirectional
num_buckets: an integer
max_distance: an integer
Returns:
a Tensor with the same shape as relative_position, containing int32
values in the range [0, num_buckets)
"""
ret = 0
n = -relative_position
if bidirectional:
num_buckets //= 2
ret += (n < 0).to(torch.long) * num_buckets # mtf.to_int32(mtf.less(n, 0)) * num_buckets
n = torch.abs(n)
else:
n = torch.max(n, torch.zeros_like(n))
# now n is in the range [0, inf)
# half of the buckets are for exact increments in positions
max_exact = num_buckets // 2
is_small = (n < max_exact)
# The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
val_if_large = max_exact + (
torch.log(n.float() / max_exact)
/ math.log(max_distance / max_exact) * (num_buckets - max_exact)).to(torch.long)
val_if_large = torch.min(val_if_large, torch.full_like(val_if_large, num_buckets - 1))
ret += torch.where(is_small, n, val_if_large)
return ret
def compute_bias(self, qlen, klen):
""" Compute binned relative position bias """
context_position = torch.arange(qlen, dtype=torch.long)[:, None]
memory_position = torch.arange(klen, dtype=torch.long)[None, :]
relative_position = memory_position - context_position # shape (qlen, klen)
rp_bucket = self._relative_position_bucket(relative_position, # shape (qlen, klen)
bidirectional=not self.is_decoder,
num_buckets=self.relative_attention_num_buckets)
values = self.relative_attention_bias(rp_bucket) # shape (qlen, klen, num_heads)
values = values.permute([2, 0, 1]).unsqueeze(0) # shape (1, num_heads, qlen, klen)
return values
def forward(self, input, mask=None, kv=None, position_bias=None, cache=None, head_mask=None):
"""
Self-attention (if kv is None) or attention over source sentence (provided by kv).
"""
# Input is (bs, qlen, dim)
# Mask is (bs, klen) (non-causal) or (bs, klen, klen)
bs, qlen, dim = input.size()
if kv is None:
klen = qlen if cache is None else cache['slen'] + qlen
else:
klen = kv.size(1)
def shape(x):
""" projection """
return x.view(bs, -1, self.n_heads, self.d_kv).transpose(1, 2)
def unshape(x):
""" compute context """
return x.transpose(1, 2).contiguous().view(bs, -1, self.inner_dim)
q = shape(self.q(input)) # (bs, n_heads, qlen, dim_per_head)
if kv is None:
k = shape(self.k(input)) # (bs, n_heads, qlen, dim_per_head)
v = shape(self.v(input)) # (bs, n_heads, qlen, dim_per_head)
elif cache is None or self.layer_id not in cache:
k = v = kv
k = shape(self.k(k)) # (bs, n_heads, qlen, dim_per_head)
v = shape(self.v(v)) # (bs, n_heads, qlen, dim_per_head)
if cache is not None:
if self.layer_id in cache:
if kv is None:
k_, v_ = cache[self.layer_id]
k = torch.cat([k_, k], dim=2) # (bs, n_heads, klen, dim_per_head)
v = torch.cat([v_, v], dim=2) # (bs, n_heads, klen, dim_per_head)
else:
k, v = cache[self.layer_id]
cache[self.layer_id] = (k, v)
# q = q / math.sqrt(dim_per_head) # No scaling in T5
scores = torch.einsum('bnqd,bnkd->bnqk', q, k) # (bs, n_heads, qlen, klen)
if position_bias is None:
if not self.has_relative_attention_bias:
raise ValueError("No position_bias provided and no weights to compute position_bias")
position_bias = self.compute_bias(qlen, klen)
if mask is not None:
position_bias = position_bias + mask # (bs, n_heads, qlen, klen)
scores += position_bias
weights = F.softmax(scores.float(), dim=-1).type_as(scores) # (bs, n_heads, qlen, klen)
weights = F.dropout(weights, p=self.dropout, training=self.training) # (bs, n_heads, qlen, klen)
# Mask heads if we want to
if head_mask is not None:
weights = weights * head_mask
context = torch.matmul(weights, v) # (bs, n_heads, qlen, dim_per_head)
context = unshape(context) # (bs, qlen, dim)
context = self.o(context)
outputs = (context,)
if self.output_attentions:
outputs = outputs + (weights,)
if self.has_relative_attention_bias:
outputs = outputs + (position_bias,)
return outputs
class T5LayerSelfAttention(nn.Module):
def __init__(self, config, has_relative_attention_bias=False):
super(T5LayerSelfAttention, self).__init__()
self.SelfAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
self.dropout = nn.Dropout(config.dropout_rate)
def forward(self, hidden_states, attention_mask=None, position_bias=None, head_mask=None):
norm_x = self.layer_norm(hidden_states)
attention_output = self.SelfAttention(norm_x,
mask=attention_mask,
position_bias=position_bias,
head_mask=head_mask)
y = attention_output[0]
layer_output = hidden_states + self.dropout(y)
outputs = (layer_output,) + attention_output[1:] # add attentions if we output them
return outputs
class T5LayerCrossAttention(nn.Module):
def __init__(self, config, has_relative_attention_bias=False):
super(T5LayerCrossAttention, self).__init__()
self.EncDecAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
self.dropout = nn.Dropout(config.dropout_rate)
def forward(self, hidden_states, kv, attention_mask=None, position_bias=None, head_mask=None):
norm_x = self.layer_norm(hidden_states)
attention_output = self.EncDecAttention(norm_x,
mask=attention_mask,
kv=kv,
position_bias=position_bias,
head_mask=head_mask)
y = attention_output[0]
layer_output = hidden_states + self.dropout(y)
outputs = (layer_output,) + attention_output[1:] # add attentions if we output them
return outputs
class T5Block(nn.Module):
def __init__(self, config, has_relative_attention_bias=False):
super(T5Block, self).__init__()
self.is_decoder = config.is_decoder
self.layer = nn.ModuleList()
self.layer.append(T5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias))
if self.is_decoder:
self.layer.append(T5LayerCrossAttention(config, has_relative_attention_bias=has_relative_attention_bias))
self.layer.append(T5LayerFF(config))
else:
self.layer.append(T5LayerFF(config))
def forward(self, hidden_states, attention_mask=None, position_bias=None,
encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None,
head_mask=None):
self_attention_outputs = self.layer[0](hidden_states,
attention_mask=attention_mask,
position_bias=position_bias,
head_mask=head_mask)
hidden_states = self_attention_outputs[0]
outputs = self_attention_outputs[1:] # Keep self-attention outputs and relative position weights
if not self.is_decoder:
hidden_states = self.layer[1](hidden_states)
else:
cross_attention_outputs = self.layer[1](hidden_states,
kv=encoder_hidden_states,
attention_mask=encoder_attention_mask,
position_bias=encoder_decoder_position_bias,
head_mask=head_mask)
hidden_states = cross_attention_outputs[0]
outputs = outputs + cross_attention_outputs[1:] # Keep cross-attention outputs and relative position weights
hidden_states = self.layer[2](hidden_states)
outputs = (hidden_states,) + outputs # add attentions if we output them
return outputs # hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
class T5PreTrainedModel(PreTrainedModel):
""" An abstract class to handle weights initialization and
a simple interface for dowloading and loading pretrained models.
"""
config_class = T5Config
pretrained_model_archive_map = T5_PRETRAINED_MODEL_ARCHIVE_MAP
load_tf_weights = load_tf_weights_in_t5
base_model_prefix = "transformer"
@property
def dummy_inputs(self):
input_ids = torch.tensor(DUMMY_INPUTS)
input_mask = torch.tensor(DUMMY_MASK)
dummy_inputs = {'decoder_input_ids': input_ids,
'encoder_input_ids': input_ids,
'decoder_attention_mask': input_mask}
return dummy_inputs
def _init_weights(self, module):
""" Initialize the weights """
factor = self.config.initializer_factor # Used for testing weights initialization
if isinstance(module, T5LayerNorm):
module.weight.data.fill_(factor*1.0)
elif isinstance(module, (T5Model, T5WithLMHeadModel)):
# Mesh TensorFlow embeddings initialization
# See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624
module.shared.weight.data.normal_(mean=0.0, std=factor*1.0)
elif isinstance(module, T5DenseReluDense):
# Mesh TensorFlow FF initialization
# See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56
# and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89
module.wi.weight.data.normal_(mean=0.0, std=factor*((self.config.d_model) ** -0.5))
if hasattr(module.wi, 'bias') and module.wi.bias is not None:
module.wi.bias.data.zero_()
module.wo.weight.data.normal_(mean=0.0, std=factor*((self.config.d_ff) ** -0.5))
if hasattr(module.wo, 'bias') and module.wo.bias is not None:
module.wo.bias.data.zero_()
elif isinstance(module, T5Attention):
# Mesh TensorFlow attention initialization to avoid scaling before softmax
# See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136
d_model = self.config.d_model
d_kv = self.config.d_kv
n_heads = self.config.num_heads
module.q.weight.data.normal_(mean=0.0, std=factor*((d_model * d_kv) ** -0.5))
module.k.weight.data.normal_(mean=0.0, std=factor*(d_model ** -0.5))
module.v.weight.data.normal_(mean=0.0, std=factor*(d_model ** -0.5))
module.o.weight.data.normal_(mean=0.0, std=factor*((n_heads * d_kv) ** -0.5))
if module.has_relative_attention_bias:
module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor*((d_model) ** -0.5))
class T5Stack(T5PreTrainedModel):
def __init__(self, config):
super(T5Stack, self).__init__(config)
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.is_decoder = config.is_decoder
self.block = nn.ModuleList([T5Block(config, has_relative_attention_bias=bool(i == 0))
for i in range(config.num_layers)])
self.final_layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
self.dropout = nn.Dropout(config.dropout_rate)
self.init_weights()
def forward(self,
hidden_states,
attention_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
head_mask=None):
batch_size, seq_length = hidden_states.shape[0], hidden_states.shape[1]
if attention_mask is None:
attention_mask = torch.ones(batch_size, seq_length).to(hidden_states.device)
if self.is_decoder and encoder_attention_mask is None:
encoder_seq_length = encoder_hidden_states.shape[1]
encoder_attention_mask = torch.ones(batch_size, encoder_seq_length).to(hidden_states.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, :, :]
elif attention_mask.dim() == 2:
# 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 self.config.is_decoder:
seq_ids = torch.arange(seq_length, device=hidden_states.device)
causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
causal_mask = causal_mask.to(attention_mask)
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 -1e9 for masked positions.
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
# T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
# Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
# extended_attention_mask = (extended_attention_mask == extended_attention_mask.transpose(-1, -2))
extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
extended_attention_mask = (1.0 - extended_attention_mask) * -1e9
if self.is_decoder:
# 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, :]
# T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
# Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
# encoder_extended_attention_mask = (encoder_extended_attention_mask == encoder_extended_attention_mask.transpose(-1, -2))
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) * -1e9
else:
encoder_extended_attention_mask = None
# 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_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_layers
all_hidden_states = ()
all_attentions = ()
position_bias = None
encoder_decoder_position_bias = None
hidden_states = self.dropout(hidden_states)
for i, layer_module in enumerate(self.block):
if self.output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
layer_outputs = layer_module(hidden_states,
attention_mask=extended_attention_mask,
position_bias=position_bias,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_extended_attention_mask,
encoder_decoder_position_bias=encoder_decoder_position_bias,
head_mask=head_mask[i])
# layer_outputs is a tuple with:
# hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
hidden_states = layer_outputs[0]
if i == 0:
# We share the position biases between the layers - the first layer store them
# layer_outputs = hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
position_bias = layer_outputs[2 if self.output_attentions else 1]
if self.is_decoder:
encoder_decoder_position_bias = layer_outputs[4 if self.output_attentions else 2]
if self.output_attentions:
all_attentions = all_attentions + (layer_outputs[1],) # We keep only self-attention weights for now
hidden_states = self.final_layer_norm(hidden_states)
layer_output = self.dropout(hidden_states)
# Add last layer
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)
T5_START_DOCSTRING = r""" The T5 model was proposed in
`Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`_
by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.
It's an encoder decoder transformer pre-trained in a text-to-text denoising generative setting.
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.
.. _`Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`:
https://arxiv.org/abs/1910.10683
.. _`torch.nn.Module`:
https://pytorch.org/docs/stable/nn.html#module
Parameters:
config (:class:`~transformers.T5Config`): 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.
"""
T5_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, T5 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]``
(b) For single sequences:
``tokens: [CLS] the dog is hairy . [SEP]``
T5 is a model with relative position embeddings so you should be able to pad the inputs on
the right or the left.
Indices can be obtained using :class:`transformers.T5Tokenizer`.
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.
**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 T5 Model transformer outputting raw hidden-states"
"without any specific head on top.",
T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
class T5Model(T5PreTrainedModel):
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.
**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 = T5Tokenizer.from_pretrained('t5-small')
model = T5Model.from_pretrained('t5-small')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
outputs = model(input_ids=input_ids)
last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple
"""
def __init__(self, config):
super(T5Model, self).__init__(config)
self.shared = nn.Embedding(config.vocab_size, config.d_model)
encoder_config = copy.deepcopy(config)
self.encoder = T5Stack(encoder_config)
decoder_config = copy.deepcopy(config)
decoder_config.is_decoder = True
self.decoder = T5Stack(decoder_config)
self.init_weights()
def get_input_embeddings(self):
return self.shared
def set_input_embeddings(self, new_embeddings):
self.shared = 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, **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_common = dict((k, v) for k, v in kwargs.items()
if not k.startswith("encoder_") and not k.startswith("decoder_"))
kwargs_encoder = kwargs_common.copy()
kwargs_decoder = kwargs_common.copy()
kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
# Encode if needed (training, first prediction pass)
encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
encoder_attention_mask = kwargs_encoder.get("attention_mask", None)
if encoder_hidden_states is None:
# Convert encoder inputs in embeddings if needed
hidden_states = kwargs_encoder.pop("inputs_embeds", None)
if hidden_states is None:
encoder_inputs_ids = kwargs_encoder.pop("input_ids")
hidden_states = self.shared(encoder_inputs_ids) # Convert inputs in embeddings
if encoder_attention_mask is not None:
# Apply masking
encoder_attention_mask = (encoder_attention_mask != 0).to(hidden_states)
hidden_states = hidden_states * encoder_attention_mask.unsqueeze(-1)
encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
encoder_hidden_states = encoder_outputs[0]
else:
encoder_outputs = ()
# Decode
# Convert decoder inputs in embeddings if needed
hidden_states = kwargs_decoder.pop("inputs_embeds", None)
if hidden_states is None:
decoder_inputs_ids = kwargs_decoder.pop("input_ids")
hidden_states = self.shared(decoder_inputs_ids)
kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
kwargs_decoder["encoder_attention_mask"] = encoder_attention_mask
decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
return decoder_outputs + encoder_outputs
@add_start_docstrings("""T5 Model with a `language modeling` head on top. """,
T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
class T5WithLMHeadModel(T5PreTrainedModel):
r"""
**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 ``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 = T5Tokenizer.from_pretrained('t5-small')
model = T5WithLMHeadModel.from_pretrained('t5-small')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
outputs = model(input_ids=input_ids, lm_labels=input_ids)
loss, prediction_scores = outputs[:2]
"""
def __init__(self, config):
super(T5WithLMHeadModel, self).__init__(config)
self.model_dim = config.d_model
self.shared = nn.Embedding(config.vocab_size, config.d_model)
encoder_config = copy.deepcopy(config)
self.encoder = T5Stack(encoder_config)
decoder_config = copy.deepcopy(config)
decoder_config.is_decoder = True
self.decoder = T5Stack(decoder_config)
self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
self.init_weights()
def get_input_embeddings(self):
return self.shared
def set_input_embeddings(self, new_embeddings):
self.shared = new_embeddings
def get_output_embeddings(self):
return self.lm_head
def forward(self, **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.
lm_labels = kwargs.pop('decoder_lm_labels', None)
kwargs_common = dict((k, v) for k, v in kwargs.items()
if not k.startswith("encoder_") and not k.startswith("decoder_"))
kwargs_encoder = kwargs_common.copy()
kwargs_decoder = kwargs_common.copy()
kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
# Encode if needed (training, first prediction pass)
encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
if encoder_hidden_states is None:
# Convert encoder inputs in embeddings if needed
hidden_states = kwargs_encoder.pop("inputs_embeds", None)
if hidden_states is None:
encoder_inputs_ids = kwargs_encoder.pop("input_ids")
hidden_states = self.shared(encoder_inputs_ids) # Convert inputs in embeddings
encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
encoder_hidden_states = encoder_outputs[0]
else:
encoder_outputs = ()
# Decode
# Convert decoder inputs in embeddings if needed
hidden_states = kwargs_decoder.pop("inputs_embeds", None)
if hidden_states is None:
decoder_inputs_ids = kwargs_decoder.pop("input_ids")
hidden_states = self.shared(decoder_inputs_ids)
kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get("attention_mask", None)
decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
sequence_output = decoder_outputs[0]
# Rescale output before projecting on vocab
# See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
sequence_output = sequence_output * (self.model_dim ** -0.5)
lm_logits = self.lm_head(sequence_output)
decoder_outputs = (lm_logits,) + decoder_outputs[1:] # Add hidden states and attention if they are here
if lm_labels is not None:
shift_logits = lm_logits[..., :-1, :].contiguous()
shift_labels = lm_labels[..., 1:].contiguous()
loss_fct = CrossEntropyLoss(ignore_index=-1)
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1))
decoder_outputs = (loss,) + decoder_outputs # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666
return decoder_outputs + encoder_outputs

4
transformers/modeling_tf_albert.py
View File

@@ -587,8 +587,8 @@ class TFAlbertModel(TFAlbertPreTrainedModel):
import tensorflow as tf
from transformers import AlbertTokenizer, TFAlbertModel
tokenizer = AlbertTokenizer.from_pretrained('bert-base-uncased')
model = TFAlbertModel.from_pretrained('bert-base-uncased')
tokenizer = AlbertTokenizer.from_pretrained('albert-base-v1')
model = TFAlbertModel.from_pretrained('albert-base-v1')
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

349
transformers/modeling_tf_auto.py
View File

@@ -18,21 +18,48 @@ from __future__ import absolute_import, division, print_function, unicode_litera
import logging
from .modeling_tf_bert import TFBertModel, TFBertForMaskedLM, TFBertForSequenceClassification, TFBertForQuestionAnswering
from .modeling_tf_openai import TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel
from .modeling_tf_gpt2 import TFGPT2Model, TFGPT2LMHeadModel
from .modeling_tf_transfo_xl import TFTransfoXLModel, TFTransfoXLLMHeadModel
from .modeling_tf_xlnet import TFXLNetModel, TFXLNetLMHeadModel, TFXLNetForSequenceClassification, TFXLNetForQuestionAnsweringSimple
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 .configuration_auto import (BertConfig, CTRLConfig, DistilBertConfig,
GPT2Config, OpenAIGPTConfig, RobertaConfig,
TransfoXLConfig, XLMConfig, XLNetConfig)
from .modeling_tf_bert import TFBertModel, TFBertForMaskedLM, TFBertForSequenceClassification, \
TFBertForQuestionAnswering, TFBertForTokenClassification, TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_tf_openai import TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel, TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_tf_gpt2 import TFGPT2Model, TFGPT2LMHeadModel, TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_tf_transfo_xl import TFTransfoXLModel, TFTransfoXLLMHeadModel, TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_tf_xlnet import TFXLNetModel, TFXLNetLMHeadModel, TFXLNetForSequenceClassification, \
TFXLNetForQuestionAnsweringSimple, TFXLNetForTokenClassification, TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_tf_xlm import TFXLMModel, TFXLMWithLMHeadModel, TFXLMForSequenceClassification, \
TFXLMForQuestionAnsweringSimple, TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_tf_roberta import TFRobertaModel, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, \
TFRobertaForTokenClassification, TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_tf_distilbert import TFDistilBertModel, TFDistilBertForQuestionAnswering, TFDistilBertForMaskedLM, TFDistilBertForSequenceClassification, TFDistilBertForTokenClassification, TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_tf_ctrl import TFCTRLModel, TFCTRLLMHeadModel, TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_tf_albert import TFAlbertModel, TFAlbertForMaskedLM, TFAlbertForSequenceClassification, TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_tf_t5 import TFT5Model, TFT5WithLMHeadModel, TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP
from .file_utils import add_start_docstrings
logger = logging.getLogger(__name__)
TF_ALL_PRETRAINED_MODEL_ARCHIVE_MAP = dict((key, value)
for pretrained_map in [
TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP,
]
for key, value, in pretrained_map.items())
class TFAutoModel(object):
r"""
:class:`~transformers.TFAutoModel` is a generic model class
@@ -45,6 +72,7 @@ class TFAutoModel(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 `t5`: TFT5Model (T5 model)
- contains `distilbert`: TFDistilBertModel (DistilBERT model)
- contains `roberta`: TFRobertaModel (RoBERTa model)
- contains `bert`: TFBertModel (Bert model)
@@ -59,7 +87,50 @@ class TFAutoModel(object):
"""
def __init__(self):
raise EnvironmentError("TFAutoModel is designed to be instantiated "
"using the `TFAutoModel.from_pretrained(pretrained_model_name_or_path)` method.")
"using the `TFAutoModel.from_pretrained(pretrained_model_name_or_path)` or "
"`TFAutoModel.from_config(config)` methods.")
@classmethod
def from_config(cls, config):
r""" Instantiates one of the base model classes of the library
from a configuration.
config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
The model class to instantiate is selected based on the configuration class:
- isInstance of `distilbert` configuration class: TFDistilBertModel (DistilBERT model)
- isInstance of `roberta` configuration class: TFRobertaModel (RoBERTa model)
- isInstance of `bert` configuration class: TFBertModel (Bert model)
- isInstance of `openai-gpt` configuration class: TFOpenAIGPTModel (OpenAI GPT model)
- isInstance of `gpt2` configuration class: TFGPT2Model (OpenAI GPT-2 model)
- isInstance of `ctrl` configuration class: TFCTRLModel (Salesforce CTRL model)
- isInstance of `transfo-xl` configuration class: TFTransfoXLModel (Transformer-XL model)
- isInstance of `xlnet` configuration class: TFXLNetModel (XLNet model)
- isInstance of `xlm` configuration class: TFXLMModel (XLM model)
Examples::
config = BertConfig.from_pretrained('bert-base-uncased') # Download configuration from S3 and cache.
model = TFAutoModel.from_config(config) # E.g. model was saved using `save_pretrained('./test/saved_model/')`
"""
if isinstance(config, DistilBertConfig):
return TFDistilBertModel(config)
elif isinstance(config, RobertaConfig):
return TFRobertaModel(config)
elif isinstance(config, BertConfig):
return TFBertModel(config)
elif isinstance(config, OpenAIGPTConfig):
return TFOpenAIGPTModel(config)
elif isinstance(config, GPT2Config):
return TFGPT2Model(config)
elif isinstance(config, TransfoXLConfig):
return TFTransfoXLModel(config)
elif isinstance(config, XLNetConfig):
return TFXLNetModel(config)
elif isinstance(config, XLMConfig):
return TFXLMModel(config)
elif isinstance(config, CTRLConfig):
return TFCTRLModel(config)
raise ValueError("Unrecognized configuration class {}".format(config))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -68,6 +139,7 @@ class TFAutoModel(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 `t5`: TFT5Model (T5 model)
- contains `distilbert`: TFDistilBertModel (DistilBERT model)
- contains `roberta`: TFRobertaModel (RoBERTa model)
- contains `bert`: TFTFBertModel (Bert model)
@@ -81,6 +153,7 @@ class TFAutoModel(object):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
- a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
@@ -136,8 +209,12 @@ class TFAutoModel(object):
model = TFAutoModel.from_pretrained('./pt_model/bert_pytorch_model.bin', from_pt=True, config=config)
"""
if 'distilbert' in pretrained_model_name_or_path:
if 't5' in pretrained_model_name_or_path:
return TFT5Model.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'distilbert' in pretrained_model_name_or_path:
return TFDistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'albert' in pretrained_model_name_or_path:
return TFAlbertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'roberta' in pretrained_model_name_or_path:
return TFRobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'bert' in pretrained_model_name_or_path:
@@ -156,7 +233,7 @@ class TFAutoModel(object):
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', "
"'distilbert', 'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
"'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))
@@ -172,6 +249,7 @@ class TFAutoModelWithLMHead(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 `t5`: TFT5WithLMHeadModel (T5 model)
- contains `distilbert`: TFDistilBertForMaskedLM (DistilBERT model)
- contains `roberta`: TFRobertaForMaskedLM (RoBERTa model)
- contains `bert`: TFBertForMaskedLM (Bert model)
@@ -186,7 +264,50 @@ class TFAutoModelWithLMHead(object):
"""
def __init__(self):
raise EnvironmentError("TFAutoModelWithLMHead is designed to be instantiated "
"using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
"using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` or "
"`TFAutoModelWithLMHead.from_config(config)` methods.")
@classmethod
def from_config(cls, config):
r""" Instantiates one of the base model classes of the library
from a configuration.
config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
The model class to instantiate is selected based on the configuration class:
- isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
- isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
- isInstance of `bert` configuration class: BertModel (Bert model)
- isInstance of `openai-gpt` configuration class: OpenAIGPTModel (OpenAI GPT model)
- isInstance of `gpt2` configuration class: GPT2Model (OpenAI GPT-2 model)
- isInstance of `ctrl` configuration class: CTRLModel (Salesforce CTRL model)
- isInstance of `transfo-xl` configuration class: TransfoXLModel (Transformer-XL model)
- isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
- isInstance of `xlm` configuration class: XLMModel (XLM model)
Examples::
config = BertConfig.from_pretrained('bert-base-uncased') # Download configuration from S3 and cache.
model = AutoModelWithLMHead.from_config(config) # E.g. model was saved using `save_pretrained('./test/saved_model/')`
"""
if isinstance(config, DistilBertConfig):
return TFDistilBertForMaskedLM(config)
elif isinstance(config, RobertaConfig):
return TFRobertaForMaskedLM(config)
elif isinstance(config, BertConfig):
return TFBertForMaskedLM(config)
elif isinstance(config, OpenAIGPTConfig):
return TFOpenAIGPTLMHeadModel(config)
elif isinstance(config, GPT2Config):
return TFGPT2LMHeadModel(config)
elif isinstance(config, TransfoXLConfig):
return TFTransfoXLLMHeadModel(config)
elif isinstance(config, XLNetConfig):
return TFXLNetLMHeadModel(config)
elif isinstance(config, XLMConfig):
return TFXLMWithLMHeadModel(config)
elif isinstance(config, CTRLConfig):
return TFCTRLLMHeadModel(config)
raise ValueError("Unrecognized configuration class {}".format(config))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -198,6 +319,7 @@ class TFAutoModelWithLMHead(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 `t5`: TFT5WithLMHeadModel (T5 model)
- contains `distilbert`: TFDistilBertForMaskedLM (DistilBERT model)
- contains `roberta`: TFRobertaForMaskedLM (RoBERTa model)
- contains `bert`: TFBertForMaskedLM (Bert model)
@@ -212,6 +334,7 @@ class TFAutoModelWithLMHead(object):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
- a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
@@ -267,8 +390,12 @@ class TFAutoModelWithLMHead(object):
model = TFAutoModelWithLMHead.from_pretrained('./pt_model/bert_pytorch_model.bin', from_pt=True, config=config)
"""
if 'distilbert' in pretrained_model_name_or_path:
if 't5' in pretrained_model_name_or_path:
return TFT5WithLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'distilbert' in pretrained_model_name_or_path:
return TFDistilBertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'albert' in pretrained_model_name_or_path:
return TFAlbertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'roberta' in pretrained_model_name_or_path:
return TFRobertaForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'bert' in pretrained_model_name_or_path:
@@ -287,7 +414,7 @@ class TFAutoModelWithLMHead(object):
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', "
"'distilbert', 'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
"'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))
@@ -312,8 +439,39 @@ class TFAutoModelForSequenceClassification(object):
This class cannot be instantiated using `__init__()` (throws an error).
"""
def __init__(self):
raise EnvironmentError("TFAutoModelWithLMHead is designed to be instantiated "
"using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
raise EnvironmentError("TFAutoModelForSequenceClassification is designed to be instantiated "
"using the `TFAutoModelForSequenceClassification.from_pretrained(pretrained_model_name_or_path)` or "
"`TFAutoModelForSequenceClassification.from_config(config)` methods.")
@classmethod
def from_config(cls, config):
r""" Instantiates one of the base model classes of the library
from a configuration.
config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
The model class to instantiate is selected based on the configuration class:
- isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
- isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
- isInstance of `bert` configuration class: BertModel (Bert model)
- isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
- isInstance of `xlm` configuration class: XLMModel (XLM model)
Examples::
config = BertConfig.from_pretrained('bert-base-uncased') # Download configuration from S3 and cache.
model = AutoModelForSequenceClassification.from_config(config) # E.g. model was saved using `save_pretrained('./test/saved_model/')`
"""
if isinstance(config, DistilBertConfig):
return TFDistilBertForSequenceClassification(config)
elif isinstance(config, RobertaConfig):
return TFRobertaForSequenceClassification(config)
elif isinstance(config, BertConfig):
return TFBertForSequenceClassification(config)
elif isinstance(config, XLNetConfig):
return TFXLNetForSequenceClassification(config)
elif isinstance(config, XLMConfig):
return TFXLMForSequenceClassification(config)
raise ValueError("Unrecognized configuration class {}".format(config))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -338,6 +496,7 @@ class TFAutoModelForSequenceClassification(object):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
- a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
@@ -395,6 +554,8 @@ class TFAutoModelForSequenceClassification(object):
"""
if 'distilbert' in pretrained_model_name_or_path:
return TFDistilBertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'albert' in pretrained_model_name_or_path:
return TFAlbertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'roberta' in pretrained_model_name_or_path:
return TFRobertaForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'bert' in pretrained_model_name_or_path:
@@ -405,7 +566,7 @@ class TFAutoModelForSequenceClassification(object):
return TFXLMForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
raise ValueError("Unrecognized model identifier in {}. Should contains one of "
"'bert', 'xlnet', 'xlm', 'roberta'".format(pretrained_model_name_or_path))
"'distilbert', 'bert', 'xlnet', 'xlm', 'roberta'".format(pretrained_model_name_or_path))
class TFAutoModelForQuestionAnswering(object):
@@ -428,8 +589,36 @@ class TFAutoModelForQuestionAnswering(object):
This class cannot be instantiated using `__init__()` (throws an error).
"""
def __init__(self):
raise EnvironmentError("TFAutoModelWithLMHead is designed to be instantiated "
"using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
raise EnvironmentError("TFAutoModelForQuestionAnswering is designed to be instantiated "
"using the `TFAutoModelForQuestionAnswering.from_pretrained(pretrained_model_name_or_path)` or "
"`TFAutoModelForQuestionAnswering.from_config(config)` methods.")
@classmethod
def from_config(cls, config):
r""" Instantiates one of the base model classes of the library
from a configuration.
config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
The model class to instantiate is selected based on the configuration class:
- isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
- isInstance of `bert` configuration class: BertModel (Bert model)
- isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
- isInstance of `xlm` configuration class: XLMModel (XLM model)
Examples::
config = BertConfig.from_pretrained('bert-base-uncased') # Download configuration from S3 and cache.
model = AutoModelForSequenceClassification.from_config(config) # E.g. model was saved using `save_pretrained('./test/saved_model/')`
"""
if isinstance(config, DistilBertConfig):
return TFDistilBertForQuestionAnswering(config)
elif isinstance(config, BertConfig):
return TFBertForQuestionAnswering(config)
elif isinstance(config, XLNetConfig):
return TFXLNetForQuestionAnswering(config)
elif isinstance(config, XLMConfig):
return TFXLMForQuestionAnswering(config)
raise ValueError("Unrecognized configuration class {}".format(config))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -453,6 +642,7 @@ class TFAutoModelForQuestionAnswering(object):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
- a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
@@ -518,4 +708,121 @@ class TFAutoModelForQuestionAnswering(object):
return TFXLMForQuestionAnsweringSimple.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
raise ValueError("Unrecognized model identifier in {}. Should contains one of "
"'bert', 'xlnet', 'xlm'".format(pretrained_model_name_or_path))
"'distilbert', 'bert', 'xlnet', 'xlm'".format(pretrained_model_name_or_path))
class TFAutoModelForTokenClassification:
def __init__(self):
raise EnvironmentError("TFAutoModelForTokenClassification is designed to be instantiated "
"using the `TFAutoModelForTokenClassification.from_pretrained(pretrained_model_name_or_path)` or "
"`AutoModelForTokenClassification.from_config(config)` methods.")
@classmethod
def from_config(cls, config):
r""" Instantiates one of the base model classes of the library
from a configuration.
config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
The model class to instantiate is selected based on the configuration class:
- isInstance of `bert` configuration class: BertModel (Bert model)
- isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
- isInstance of `distilbert` configuration class: DistilBertModel (DistilBert model)
- isInstance of `roberta` configuration class: RobteraModel (Roberta model)
Examples::
config = BertConfig.from_pretrained('bert-base-uncased') # Download configuration from S3 and cache.
model = TFAutoModelForTokenClassification.from_config(config) # E.g. model was saved using `save_pretrained('./test/saved_model/')`
"""
if isinstance(config, BertConfig):
return TFBertForTokenClassification(config)
elif isinstance(config, XLNetConfig):
return TFXLNetForTokenClassification(config)
elif isinstance(config, DistilBertConfig):
return TFDistilBertForTokenClassification(config)
elif isinstance(config, RobertaConfig):
return TFRobertaForTokenClassification(config)
raise ValueError("Unrecognized configuration class {}".format(config))
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
r""" Instantiates one of the question answering model classes of the library
from a pre-trained model configuration.
The `from_pretrained()` method takes care of returning the correct model class instance
using pattern matching on the `pretrained_model_name_or_path` string.
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 `bert`: BertForTokenClassification (Bert model)
- contains `xlnet`: XLNetForTokenClassification (XLNet model)
- contains `distilbert`: DistilBertForTokenClassification (DistilBert model)
- contains `roberta`: RobertaForTokenClassification (Roberta model)
The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
To train the model, you should first set it back in training mode with `model.train()`
Params:
pretrained_model_name_or_path: 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/``.
- 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.
Examples::
model = TFAutoModelForTokenClassification.from_pretrained('bert-base-uncased') # Download model and configuration from S3 and cache.
model = TFAutoModelForTokenClassification.from_pretrained('./test/bert_model/') # E.g. model was saved using `save_pretrained('./test/saved_model/')`
model = TFAutoModelForTokenClassification.from_pretrained('bert-base-uncased', output_attention=True) # Update configuration during loading
assert model.config.output_attention == True
# Loading from a TF checkpoint file instead of a PyTorch model (slower)
config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
model = TFAutoModelForTokenClassification.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
"""
if 'bert' in pretrained_model_name_or_path:
return TFBertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'xlnet' in pretrained_model_name_or_path:
return TFXLNetForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'distilbert' in pretrained_model_name_or_path:
return TFDistilBertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
elif 'roberta' in pretrained_model_name_or_path:
return TFRobertaForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
raise ValueError("Unrecognized model identifier in {}. Should contains one of "
"'bert', 'xlnet', 'distilbert', 'roberta'".format(pretrained_model_name_or_path))

20
transformers/modeling_tf_bert.py
View File

@@ -48,6 +48,12 @@ TF_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-tf_model.h5",
'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-tf_model.h5",
'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-tf_model.h5",
'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-tf_model.h5",
'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-tf_model.h5",
'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-tf_model.h5",
'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-tf_model.h5",
'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/tf_model.h5",
'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/tf_model.h5",
}
@@ -671,7 +677,7 @@ class TFBertModel(TFBertPreTrainedModel):
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = TFBertModel.from_pretrained('bert-base-uncased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[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
@@ -710,7 +716,7 @@ class TFBertForPreTraining(TFBertPreTrainedModel):
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = TFBertForPreTraining.from_pretrained('bert-base-uncased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
prediction_scores, seq_relationship_scores = outputs[:2]
@@ -759,7 +765,7 @@ class TFBertForMaskedLM(TFBertPreTrainedModel):
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = TFBertForMaskedLM.from_pretrained('bert-base-uncased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
prediction_scores = outputs[0]
@@ -806,7 +812,7 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel):
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = TFBertForNextSentencePrediction.from_pretrained('bert-base-uncased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
seq_relationship_scores = outputs[0]
@@ -851,7 +857,7 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel):
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = TFBertForSequenceClassification.from_pretrained('bert-base-uncased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
logits = outputs[0]
@@ -988,7 +994,7 @@ class TFBertForTokenClassification(TFBertPreTrainedModel):
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = TFBertForTokenClassification.from_pretrained('bert-base-uncased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
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]
@@ -1041,7 +1047,7 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel):
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = TFBertForQuestionAnswering.from_pretrained('bert-base-uncased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
start_scores, end_scores = outputs[:2]

4
transformers/modeling_tf_ctrl.py
View File

@@ -418,7 +418,7 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
tokenizer = CTRLTokenizer.from_pretrained('ctrl')
model = TFCTRLModel.from_pretrained('ctrl')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[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
@@ -481,7 +481,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel):
tokenizer = CTRLTokenizer.from_pretrained('ctrl')
model = TFCTRLLMHeadModel.from_pretrained('ctrl')
input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
outputs = model(input_ids, labels=input_ids)
loss, logits = outputs[:2]

5
transformers/modeling_tf_gpt2.py
View File

@@ -454,7 +454,7 @@ class TFGPT2Model(TFGPT2PreTrainedModel):
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
model = TFGPT2Model.from_pretrained('gpt2')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[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
@@ -495,7 +495,7 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel):
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
model = TFGPT2LMHeadModel.from_pretrained('gpt2')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
logits = outputs[0]
@@ -574,6 +574,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
"""
def __init__(self, config, *inputs, **kwargs):
super(TFGPT2DoubleHeadsModel, self).__init__(config, *inputs, **kwargs)
config.num_labels = 1
self.transformer = TFGPT2MainLayer(config, name='transformer')
self.multiple_choice_head = TFSequenceSummary(config, initializer_range=config.initializer_range, name='multiple_choice_head')

5
transformers/modeling_tf_openai.py
View File

@@ -431,7 +431,7 @@ class TFOpenAIGPTModel(TFOpenAIGPTPreTrainedModel):
tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
model = TFOpenAIGPTModel.from_pretrained('openai-gpt')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[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
@@ -467,7 +467,7 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel):
tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
model = TFOpenAIGPTLMHeadModel.from_pretrained('openai-gpt')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
logits = outputs[0]
@@ -538,6 +538,7 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
"""
def __init__(self, config, *inputs, **kwargs):
super(TFOpenAIGPTDoubleHeadsModel, self).__init__(config, *inputs, **kwargs)
config.num_labels = 1
self.transformer = TFOpenAIGPTMainLayer(config, name='transformer')
self.multiple_choice_head = TFSequenceSummary(config, initializer_range=config.initializer_range, name='multiple_choice_head')

26
transformers/modeling_tf_pytorch_utils.py
View File

@@ -78,6 +78,7 @@ def load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_i
logger.info("Loading PyTorch weights from {}".format(pt_path))
pt_state_dict = torch.load(pt_path, map_location='cpu')
logger.info("PyTorch checkpoint contains {:,} parameters".format(sum(t.numel() for t in pt_state_dict.values())))
return load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=tf_inputs, allow_missing_keys=allow_missing_keys)
@@ -118,9 +119,6 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
new_key = key.replace('gamma', 'weight')
if 'beta' in key:
new_key = key.replace('beta', 'bias')
# DialoGPT format
if key == 'lm_head.decoder.weight':
new_key = 'lm_head.weight'
if new_key:
old_keys.append(key)
new_keys.append(new_key)
@@ -134,7 +132,7 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
start_prefix_to_remove = tf_model.base_model_prefix + '.'
symbolic_weights = tf_model.trainable_weights + tf_model.non_trainable_weights
tf_loaded_numel = 0
weight_value_tuples = []
all_pytorch_weights = set(list(pt_state_dict.keys()))
for symbolic_weight in symbolic_weights:
@@ -142,7 +140,11 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
name, transpose = convert_tf_weight_name_to_pt_weight_name(sw_name, start_prefix_to_remove=start_prefix_to_remove)
# Find associated numpy array in pytorch model state dict
assert name in pt_state_dict, "{} not found in PyTorch model".format(name)
if name not in pt_state_dict:
if allow_missing_keys:
continue
raise AttributeError("{} not found in PyTorch model".format(name))
array = pt_state_dict[name].numpy()
if transpose:
@@ -159,7 +161,8 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
e.args += (symbolic_weight.shape, array.shape)
raise e
logger.info("Initialize TF weight {}".format(symbolic_weight.name))
tf_loaded_numel += array.size
# logger.warning("Initialize TF weight {}".format(symbolic_weight.name))
weight_value_tuples.append((symbolic_weight, array))
all_pytorch_weights.discard(name)
@@ -169,6 +172,8 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
if tf_inputs is not None:
tfo = tf_model(tf_inputs, training=False) # Make sure restore ops are run
logger.info("Loaded {:,} parameters in the TF 2.0 model.".format(tf_loaded_numel))
logger.info("Weights or buffers not loaded from PyTorch model: {}".format(all_pytorch_weights))
return tf_model
@@ -246,6 +251,7 @@ def load_tf2_weights_in_pytorch_model(pt_model, tf_weights, allow_missing_keys=F
all_tf_weights = set(list(tf_weights_map.keys()))
loaded_pt_weights_data_ptr = {}
missing_keys_pt = []
for pt_weight_name, pt_weight in current_pt_params_dict.items():
# Handle PyTorch shared weight ()not duplicated in TF 2.0
if pt_weight.data_ptr() in loaded_pt_weights_data_ptr:
@@ -254,7 +260,10 @@ def load_tf2_weights_in_pytorch_model(pt_model, tf_weights, allow_missing_keys=F
# Find associated numpy array in pytorch model state dict
if pt_weight_name not in tf_weights_map:
raise ValueError("{} not found in TF 2.0 model".format(pt_weight_name))
if allow_missing_keys:
missing_keys_pt.append(pt_weight_name)
continue
raise AttributeError("{} not found in TF 2.0 model".format(pt_weight_name))
array, transpose = tf_weights_map[pt_weight_name]
@@ -272,13 +281,14 @@ def load_tf2_weights_in_pytorch_model(pt_model, tf_weights, allow_missing_keys=F
e.args += (pt_weight.shape, array.shape)
raise e
logger.info("Initialize PyTorch weight {}".format(pt_weight_name))
# logger.warning("Initialize PyTorch weight {}".format(pt_weight_name))
new_pt_params_dict[pt_weight_name] = torch.from_numpy(array)
loaded_pt_weights_data_ptr[pt_weight.data_ptr()] = torch.from_numpy(array)
all_tf_weights.discard(pt_weight_name)
missing_keys, unexpected_keys = pt_model.load_state_dict(new_pt_params_dict, strict=False)
missing_keys += missing_keys_pt
if len(missing_keys) > 0:
logger.info("Weights of {} not initialized from TF 2.0 model: {}".format(

6
transformers/modeling_tf_roberta.py
View File

@@ -199,7 +199,7 @@ class TFRobertaModel(TFRobertaPreTrainedModel):
tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
model = TFRobertaModel.from_pretrained('roberta-base')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[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
@@ -276,7 +276,7 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel):
tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
model = TFRobertaForMaskedLM.from_pretrained('roberta-base')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids, masked_lm_labels=input_ids)
prediction_scores = outputs[0]
@@ -347,7 +347,7 @@ class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel):
tokenizer = RoertaTokenizer.from_pretrained('roberta-base')
model = TFRobertaForSequenceClassification.from_pretrained('roberta-base')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
labels = tf.constant([1])[None, :] # Batch size 1
outputs = model(input_ids)
logits = outputs[0]

775
transformers/modeling_tf_t5.py Normal file
View File

@@ -0,0 +1,775 @@
# coding=utf-8
# Copyright 2018 T5 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 T5 model. """
from __future__ import absolute_import, division, print_function, unicode_literals
import logging
import math
import copy
import itertools
import tensorflow as tf
from .configuration_t5 import T5Config
from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, shape_list
from .file_utils import add_start_docstrings, DUMMY_INPUTS, DUMMY_MASK
logger = logging.getLogger(__name__)
TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP = {
't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-small-tf_model.h5",
't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-base-tf_model.h5",
't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-large-tf_model.h5",
't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-3b-tf_model.h5",
't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-11b-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 (it-self a sub-class of tf.keras.Model)
####################################################
class TFT5LayerNorm(tf.keras.layers.Layer):
def __init__(self, epsilon=1e-6, **kwargs):
""" Construct a layernorm module in the T5 style
No bias and no substraction of mean.
"""
super(TFT5LayerNorm, self).__init__(**kwargs)
self.variance_epsilon = epsilon
def build(self, input_shape):
"""Build shared word embedding layer """
self.weight = self.add_weight(
"weight",
shape=(input_shape[-1],),
initializer='ones')
super(TFT5LayerNorm, self).build(input_shape)
def call(self, x):
variance = tf.math.reduce_mean(tf.math.square(x), axis=-1, keepdims=True)
x = x * tf.math.rsqrt(variance + self.variance_epsilon)
return self.weight * x
class TFT5DenseReluDense(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super(TFT5DenseReluDense, self).__init__(**kwargs)
self.wi = tf.keras.layers.Dense(config.d_ff, use_bias=False, name='wi')
self.wo = tf.keras.layers.Dense(config.d_model, use_bias=False, name='wo')
self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
self.act = tf.keras.activations.relu
def call(self, hidden_states, training=False):
h = self.wi(hidden_states)
h = self.act(h)
h = self.dropout(h, training=training)
h = self.wo(h)
return h
class TFT5LayerFF(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super(TFT5LayerFF, self).__init__(**kwargs)
self.DenseReluDense = TFT5DenseReluDense(config, name='DenseReluDense')
self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
name='layer_norm')
self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
def call(self, hidden_states, training=False):
norm_x = self.layer_norm(hidden_states)
y = self.DenseReluDense(norm_x, training=training)
layer_output = hidden_states + self.dropout(y, training=training)
return layer_output
class TFT5Attention(tf.keras.layers.Layer):
NEW_ID = itertools.count()
def __init__(self, config, has_relative_attention_bias=False, **kwargs):
super(TFT5Attention, self).__init__(**kwargs)
self.layer_id = next(TFT5Attention.NEW_ID)
self.is_decoder = config.is_decoder
self.has_relative_attention_bias = has_relative_attention_bias
self.output_attentions = config.output_attentions
self.relative_attention_num_buckets = config.relative_attention_num_buckets
self.d_model = config.d_model
self.d_kv = config.d_kv
self.n_heads = config.num_heads
self.inner_dim = self.n_heads * self.d_kv
# Mesh TensorFlow initialization to avoid scaling before softmax
self.q = tf.keras.layers.Dense(self.inner_dim, use_bias=False, name='q')
self.k = tf.keras.layers.Dense(self.inner_dim, use_bias=False, name='k')
self.v = tf.keras.layers.Dense(self.inner_dim, use_bias=False, name='v')
self.o = tf.keras.layers.Dense(self.d_model, use_bias=False, name='o')
self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
if self.has_relative_attention_bias:
self.relative_attention_bias = tf.keras.layers.Embedding(self.relative_attention_num_buckets,
self.n_heads,
name='relative_attention_bias')
self.pruned_heads = set()
def prune_heads(self, heads):
raise NotImplementedError
@staticmethod
def _relative_position_bucket(relative_position,
bidirectional=True,
num_buckets=32,
max_distance=128):
"""
Adapted from Mesh Tensorflow:
https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
Translate relative position to a bucket number for relative attention.
The relative position is defined as memory_position - query_position, i.e.
the distance in tokens from the attending position to the attended-to
position. If bidirectional=False, then positive relative positions are
invalid.
We use smaller buckets for small absolute relative_position and larger buckets
for larger absolute relative_positions. All relative positions >=max_distance
map to the same bucket. All relative positions <=-max_distance map to the
same bucket. This should allow for more graceful generalization to longer
sequences than the model has been trained on.
Args:
relative_position: an int32 Tensor
bidirectional: a boolean - whether the attention is bidirectional
num_buckets: an integer
max_distance: an integer
Returns:
a Tensor with the same shape as relative_position, containing int32
values in the range [0, num_buckets)
"""
ret = 0
n = -relative_position
if bidirectional:
num_buckets //= 2
ret += tf.dtypes.cast(tf.math.less(n, 0), tf.int32) * num_buckets
n = tf.math.abs(n)
else:
n = tf.math.maximum(n, 0)
# now n is in the range [0, inf)
max_exact = num_buckets // 2
is_small = tf.math.less(n, max_exact)
val_if_large = max_exact + tf.dtypes.cast(
tf.math.log(tf.dtypes.cast(n, tf.float32) / max_exact)
/ math.log(max_distance / max_exact) * (num_buckets - max_exact), tf.int32)
val_if_large = tf.math.minimum(val_if_large, num_buckets - 1)
ret += tf.where(is_small, n, val_if_large)
return ret
def compute_bias(self, qlen, klen):
""" Compute binned relative position bias """
context_position = tf.range(qlen)[:, None]
memory_position = tf.range(klen)[None, :]
relative_position = memory_position - context_position # shape (qlen, klen)
rp_bucket = self._relative_position_bucket(relative_position,
bidirectional=not self.is_decoder,
num_buckets=self.relative_attention_num_buckets)
values = self.relative_attention_bias(rp_bucket) # shape (qlen, klen, num_heads)
values = tf.expand_dims(tf.transpose(values, [2, 0, 1]), axis=0) # shape (1, num_heads, qlen, klen)
return values
def call(self, input, mask=None, kv=None, position_bias=None, cache=None, head_mask=None, training=False):
"""
Self-attention (if kv is None) or attention over source sentence (provided by kv).
"""
# Input is (bs, qlen, dim)
# Mask is (bs, klen) (non-causal) or (bs, klen, klen)
bs, qlen, dim = shape_list(input)
if kv is None:
klen = qlen if cache is None else cache['slen'] + qlen
else:
klen = shape_list(kv)[1]
def shape(x):
""" projection """
return tf.transpose(tf.reshape(x, (bs, -1, self.n_heads, self.d_kv)), perm=(0, 2, 1, 3))
def unshape(x):
""" compute context """
return tf.reshape(tf.transpose(x, perm=(0, 2, 1, 3)), (bs, -1, self.inner_dim))
q = shape(self.q(input)) # (bs, n_heads, qlen, dim_per_head)
if kv is None:
k = shape(self.k(input)) # (bs, n_heads, qlen, dim_per_head)
v = shape(self.v(input)) # (bs, n_heads, qlen, dim_per_head)
elif cache is None or self.layer_id not in cache:
k = v = kv
k = shape(self.k(k)) # (bs, n_heads, qlen, dim_per_head)
v = shape(self.v(v)) # (bs, n_heads, qlen, dim_per_head)
if cache is not None:
if self.layer_id in cache:
if kv is None:
k_, v_ = cache[self.layer_id]
k = tf.concat([k_, k], axis=2) # (bs, n_heads, klen, dim_per_head)
v = tf.concat([v_, v], axis=2) # (bs, n_heads, klen, dim_per_head)
else:
k, v = cache[self.layer_id]
cache[self.layer_id] = (k, v)
# q = q / math.sqrt(dim_per_head) # No scaling in T5
# scores = tf.matmul(q, k, transpose_b=True) # (bs, n_heads, qlen, klen)
scores = tf.einsum('bnqd,bnkd->bnqk', q, k) # (bs, n_heads, qlen, klen)
if position_bias is None:
if not self.has_relative_attention_bias:
raise ValueError("No position_bias provided and no weights to compute position_bias")
position_bias = self.compute_bias(qlen, klen)
if mask is not None:
position_bias = position_bias + mask
# mask = (mask == 0).expand_as(scores) # (bs, n_heads, qlen, klen)
# scores.masked_fill_(mask, -float('inf')) # (bs, n_heads, qlen, klen)
scores += position_bias
weights = tf.nn.softmax(scores, axis=-1) # (bs, n_heads, qlen, klen)
weights = self.dropout(weights, training=training) # (bs, n_heads, qlen, klen)
# Mask heads if we want to
if head_mask is not None:
weights = weights * head_mask
context = tf.matmul(weights, v) # (bs, n_heads, qlen, dim_per_head)
context = unshape(context) # (bs, qlen, dim)
context = self.o(context)
outputs = (context,)
if self.output_attentions:
outputs = outputs + (weights,)
if self.has_relative_attention_bias:
outputs = outputs + (position_bias,)
return outputs
class TFT5LayerSelfAttention(tf.keras.layers.Layer):
def __init__(self, config, has_relative_attention_bias=False, **kwargs):
super(TFT5LayerSelfAttention, self).__init__(**kwargs)
self.SelfAttention = TFT5Attention(config,
has_relative_attention_bias=has_relative_attention_bias,
name='SelfAttention')
self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
name='layer_norm')
self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
def call(self, hidden_states, attention_mask=None, position_bias=None,
head_mask=None, training=False):
norm_x = self.layer_norm(hidden_states)
attention_output = self.SelfAttention(norm_x,
mask=attention_mask,
position_bias=position_bias,
head_mask=head_mask,
training=training)
y = attention_output[0]
layer_output = hidden_states + self.dropout(y, training=training)
outputs = (layer_output,) + attention_output[1:] # add attentions if we output them
return outputs
class TFT5LayerCrossAttention(tf.keras.layers.Layer):
def __init__(self, config, has_relative_attention_bias=False, **kwargs):
super(TFT5LayerCrossAttention, self).__init__(**kwargs)
self.EncDecAttention = TFT5Attention(config,
has_relative_attention_bias=has_relative_attention_bias,
name='EncDecAttention')
self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
name='layer_norm')
self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
def call(self, hidden_states, kv, attention_mask=None, position_bias=None,
head_mask=None, training=False):
norm_x = self.layer_norm(hidden_states)
attention_output = self.EncDecAttention(norm_x,
mask=attention_mask,
kv=kv,
position_bias=position_bias,
head_mask=head_mask,
training=training)
y = attention_output[0]
layer_output = hidden_states + self.dropout(y, training=training)
outputs = (layer_output,) + attention_output[1:] # add attentions if we output them
return outputs
class TFT5Block(tf.keras.layers.Layer):
def __init__(self, config, has_relative_attention_bias=False, **kwargs):
super(TFT5Block, self).__init__(**kwargs)
self.is_decoder = config.is_decoder
self.layer = []
self.layer.append(TFT5LayerSelfAttention(config,
has_relative_attention_bias=has_relative_attention_bias,
name='layer_._0'))
if self.is_decoder:
self.layer.append(TFT5LayerCrossAttention(config,
has_relative_attention_bias=has_relative_attention_bias,
name='layer_._1'))
self.layer.append(TFT5LayerFF(config, name='layer_._2'))
else:
self.layer.append(TFT5LayerFF(config, name='layer_._1'))
def call(self, hidden_states, attention_mask=None, position_bias=None,
encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None,
head_mask=None, training=False):
self_attention_outputs = self.layer[0](hidden_states,
attention_mask=attention_mask,
position_bias=position_bias,
head_mask=head_mask,
training=training)
hidden_states = self_attention_outputs[0]
outputs = self_attention_outputs[1:]
if not self.is_decoder:
hidden_states = self.layer[1](hidden_states, training=training)
else:
cross_attention_outputs = self.layer[1](hidden_states,
kv=encoder_hidden_states,
attention_mask=encoder_attention_mask,
position_bias=encoder_decoder_position_bias,
head_mask=head_mask,
training=training)
hidden_states = cross_attention_outputs[0]
outputs = outputs + cross_attention_outputs[1:]
hidden_states = self.layer[2](hidden_states, training=training)
outputs = (hidden_states,) + outputs # add attentions if we output them
return outputs # hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
####################################################
# The full model without a specific pretrained or finetuning head is
# provided as a tf.keras.layers.Layer usually called "TFT5MainLayer"
####################################################
class TFT5MainLayer(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super(TFT5MainLayer, self).__init__(**kwargs)
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.is_decoder = config.is_decoder
self.config = config
self.num_hidden_layers = config.num_layers
self.block = [TFT5Block(config,
has_relative_attention_bias=bool(i == 0),
name='block_._{}'.format(i))
for i in range(config.num_layers)]
self.final_layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
name='final_layer_norm')
self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
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, hidden_states, attention_mask=None, encoder_hidden_states=None,
encoder_attention_mask=None, head_mask=None, training=False):
batch_size, seq_length = shape_list(hidden_states)[:2]
if attention_mask is None:
attention_mask = tf.fill((batch_size, seq_length), 1)
if self.is_decoder and encoder_attention_mask is None:
encoder_seq_length = encoder_hidden_states.shape[1]
encoder_attention_mask = tf.fill((batch_size, encoder_seq_length), 1)
# 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.
attention_mask = tf.cast(attention_mask, dtype=tf.float32)
num_dims_attention_mask = len(shape_list(attention_mask))
if num_dims_attention_mask == 3:
extended_attention_mask = attention_mask[:, None, :, :]
elif num_dims_attention_mask == 2:
# 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 self.config.is_decoder:
seq_ids = tf.range(seq_length)
causal_mask = tf.less_equal(tf.tile(seq_ids[None, None, :], (batch_size, seq_length, 1)),
seq_ids[None, :, None])
causal_mask = tf.cast(causal_mask, dtype=tf.float32)
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.
# T5 has a mask that can compare sequence ids, we can simulate this here with this transposistion
# Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
# extended_attention_mask = tf.math.equal(extended_attention_mask,
# tf.transpose(extended_attention_mask, perm=(-1, -2)))
extended_attention_mask = (1.0 - extended_attention_mask) * -1e9
if self.is_decoder:
# 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]
encoder_attention_mask = tf.cast(encoder_attention_mask, dtype=tf.float32)
num_dims_encoder_attention_mask = len(shape_list(encoder_attention_mask))
if num_dims_encoder_attention_mask == 3:
encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
if num_dims_encoder_attention_mask == 2:
encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
# T5 has a mask that can compare sequence ids, we can simulate this here with this transposistion
# Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
# encoder_extended_attention_mask = tf.math.equal(encoder_extended_attention_mask,
# tf.transpose(encoder_extended_attention_mask, perm=(-1, -2)))
encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e9
else:
encoder_extended_attention_mask = None
# 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)
all_hidden_states = ()
all_attentions = ()
position_bias = None
encoder_decoder_position_bias = None
for i, layer_module in enumerate(self.block):
if self.output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
layer_outputs = layer_module(hidden_states,
attention_mask=extended_attention_mask,
position_bias=position_bias,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_extended_attention_mask,
encoder_decoder_position_bias=encoder_decoder_position_bias,
head_mask=head_mask[i],
training=training)
hidden_states = layer_outputs[0]
if i == 0:
# We share the position biases between the layers - the first layer store them
# layer_outputs = hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
position_bias = layer_outputs[2 if self.output_attentions else 1]
if self.is_decoder:
encoder_decoder_position_bias = layer_outputs[4 if self.output_attentions else 2]
if self.output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
hidden_states = self.final_layer_norm(hidden_states)
layer_output = self.dropout(hidden_states, training=training)
# Add last layer
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)
####################################################
# TFT5PreTrainedModel 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 TFT5PreTrainedModel(TFPreTrainedModel):
""" An abstract class to handle weights initialization and
a simple interface for dowloading and loading pretrained models.
"""
config_class = T5Config
pretrained_model_archive_map = TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP
base_model_prefix = "transformer"
@property
def dummy_inputs(self):
input_ids = tf.constant(DUMMY_INPUTS)
input_mask = tf.constant(DUMMY_MASK)
dummy_inputs = {'decoder_input_ids': input_ids,
'encoder_input_ids': input_ids,
'decoder_attention_mask': input_mask}
return dummy_inputs
T5_START_DOCSTRING = r""" The T5 model was proposed in
`Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`_
by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.
It's an encoder decoder transformer pre-trained in a text-to-text denoising generative setting.
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.
.. _`Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`:
https://arxiv.org/abs/1910.10683
.. _`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.T5Config`): 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.
"""
T5_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, T5 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]``
(b) For single sequences:
``tokens: [CLS] the dog is hairy . [SEP]``
T5 is a model with relative position embeddings so you should be able to pad the inputs on
the right or the left.
Indices can be obtained using :class:`transformers.T5Tokenizer`.
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.
**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 T5 Model transformer outputting raw hidden-states"
"without any specific head on top.",
T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
class TFT5Model(TFT5PreTrainedModel):
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.
**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 T5Tokenizer, TFT5Model
tokenizer = T5Tokenizer.from_pretrained('t5-small')
model = TFT5Model.from_pretrained('t5-small')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
outputs = model(input_ids=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(TFT5Model, self).__init__(config, *inputs, **kwargs)
self.shared = TFSharedEmbeddings(config.vocab_size, config.d_model,
name='shared')
encoder_config = copy.deepcopy(config)
self.encoder = TFT5MainLayer(encoder_config, name='encoder')
decoder_config = copy.deepcopy(config)
decoder_config.is_decoder = True
self.decoder = TFT5MainLayer(decoder_config, name='decoder')
def get_input_embeddings(self):
return self.shared
def get_output_embeddings(self):
return self.shared
def call(self, decoder_input_ids, **kwargs):
# We allow two 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
# The last option is useful to use the tf.keras fit() method.
if isinstance(decoder_input_ids, dict):
kwargs.update(decoder_input_ids)
else:
kwargs['decoder_input_ids'] = decoder_input_ids
kwargs_common = dict((k, v) for k, v in kwargs.items()
if not k.startswith("encoder_") and not k.startswith("decoder_"))
kwargs_encoder = kwargs_common.copy()
kwargs_decoder = kwargs_common.copy()
kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
# Encode if needed (training, first prediction pass)
encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
if encoder_hidden_states is None:
# Convert encoder inputs in embeddings if needed
hidden_states = kwargs_encoder.pop("inputs_embeds", None)
if hidden_states is None:
encoder_inputs_ids = kwargs_encoder.pop("input_ids")
hidden_states = self.shared(encoder_inputs_ids) # Convert inputs in embeddings
encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
encoder_hidden_states = encoder_outputs[0]
else:
encoder_outputs = ()
# Decode
# Convert decoder inputs in embeddings if needed
hidden_states = kwargs_decoder.pop("inputs_embeds", None)
if hidden_states is None:
decoder_inputs_ids = kwargs_decoder.pop("input_ids")
hidden_states = self.shared(decoder_inputs_ids)
kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get("attention_mask", None)
decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
return decoder_outputs + encoder_outputs
@add_start_docstrings("""T5 Model with a `language modeling` head on top. """,
T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
class TFT5WithLMHeadModel(TFT5PreTrainedModel):
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 T5Tokenizer, TFT5WithLMHeadModel
tokenizer = T5Tokenizer.from_pretrained('t5-small')
model = TFT5WithLMHeadModel.from_pretrained('t5-small')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
outputs = model(input_ids=input_ids)
prediction_scores = outputs[0]
"""
def __init__(self, config, *inputs, **kwargs):
super(TFT5WithLMHeadModel, self).__init__(config, *inputs, **kwargs)
self.model_dim = config.d_model
self.shared = TFSharedEmbeddings(config.vocab_size, config.d_model,
name='shared')
encoder_config = copy.deepcopy(config)
self.encoder = TFT5MainLayer(encoder_config, name='encoder')
decoder_config = copy.deepcopy(config)
decoder_config.is_decoder = True
self.decoder = TFT5MainLayer(decoder_config, name='decoder')
def get_input_embeddings(self):
return self.shared
def get_output_embeddings(self):
return self.shared
def call(self, decoder_input_ids, **kwargs):
# We allow two 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
# The last option is useful to use the tf.keras fit() method.
if isinstance(decoder_input_ids, dict):
kwargs.update(decoder_input_ids)
else:
kwargs['decoder_input_ids'] = decoder_input_ids
kwargs_common = dict((k, v) for k, v in kwargs.items()
if not k.startswith("encoder_") and not k.startswith("decoder_"))
kwargs_encoder = kwargs_common.copy()
kwargs_decoder = kwargs_common.copy()
kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
# Encode if needed (training, first prediction pass)
encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
if encoder_hidden_states is None:
# Convert encoder inputs in embeddings if needed
hidden_states = kwargs_encoder.pop("inputs_embeds", None)
if hidden_states is None:
encoder_inputs_ids = kwargs_encoder.pop("input_ids")
hidden_states = self.shared(encoder_inputs_ids) # Convert inputs in embeddings
encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
encoder_hidden_states = encoder_outputs[0]
else:
encoder_outputs = ()
# Decode
# Convert decoder inputs in embeddings if needed
hidden_states = kwargs_decoder.pop("inputs_embeds", None)
if hidden_states is None:
decoder_inputs_ids = kwargs_decoder.pop("input_ids")
hidden_states = self.shared(decoder_inputs_ids)
kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get("attention_mask", None)
decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
sequence_output = decoder_outputs[0] * (self.model_dim ** -0.5)
lm_logits = self.shared(sequence_output, mode="linear")
decoder_outputs = (lm_logits,) + decoder_outputs[1:]
return decoder_outputs + encoder_outputs

10
transformers/modeling_tf_transfo_xl.py
View File

@@ -353,7 +353,7 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.n_token = config.n_token
self.n_token = config.vocab_size
self.d_embed = config.d_embed
self.d_model = config.d_model
@@ -361,7 +361,7 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
self.d_head = config.d_head
self.untie_r = config.untie_r
self.word_emb = TFAdaptiveEmbedding(config.n_token, config.d_embed, config.d_model, config.cutoffs,
self.word_emb = TFAdaptiveEmbedding(config.vocab_size, config.d_embed, config.d_model, config.cutoffs,
div_val=config.div_val, init_std=config.init_std, name='word_emb')
self.drop = tf.keras.layers.Dropout(config.dropout)
@@ -673,7 +673,7 @@ class TFTransfoXLModel(TFTransfoXLPreTrainedModel):
tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
model = TFTransfoXLModel.from_pretrained('transfo-xl-wt103')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
last_hidden_states, mems = outputs[:2]
@@ -715,7 +715,7 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
model = TFTransfoXLLMHeadModel.from_pretrained('transfo-xl-wt103')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
prediction_scores, mems = outputs[:2]
@@ -729,7 +729,7 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
raise NotImplementedError
# use adaptive softmax (including standard softmax)
else:
self.crit = TFAdaptiveSoftmaxMask(config.n_token, config.d_embed, config.d_model,
self.crit = TFAdaptiveSoftmaxMask(config.vocab_size, config.d_embed, config.d_model,
config.cutoffs, div_val=config.div_val, name='crit')
def reset_length(self, tgt_len, ext_len, mem_len):

12
transformers/modeling_tf_transfo_xl_utilities.py
View File

@@ -25,15 +25,15 @@ import tensorflow as tf
from .modeling_tf_utils import shape_list
class TFAdaptiveSoftmaxMask(tf.keras.layers.Layer):
def __init__(self, n_token, d_embed, d_proj, cutoffs, div_val=1,
def __init__(self, vocab_size, d_embed, d_proj, cutoffs, div_val=1,
keep_order=False, **kwargs):
super(TFAdaptiveSoftmaxMask, self).__init__(**kwargs)
self.n_token = n_token
self.vocab_size = vocab_size
self.d_embed = d_embed
self.d_proj = d_proj
self.cutoffs = cutoffs + [n_token]
self.cutoffs = cutoffs + [vocab_size]
self.cutoff_ends = [0] + self.cutoffs
self.div_val = div_val
@@ -66,11 +66,11 @@ class TFAdaptiveSoftmaxMask(tf.keras.layers.Layer):
self.out_projs.append(weight)
else:
self.out_projs.append(None)
weight = self.add_weight(shape=(self.n_token, self.d_embed,),
weight = self.add_weight(shape=(self.vocab_size, self.d_embed,),
initializer='zeros',
trainable=True,
name='out_layers_._{}_._weight'.format(i))
bias = self.add_weight(shape=(self.n_token,),
bias = self.add_weight(shape=(self.vocab_size,),
initializer='zeros',
trainable=True,
name='out_layers_._{}_._bias'.format(i))
@@ -114,7 +114,7 @@ class TFAdaptiveSoftmaxMask(tf.keras.layers.Layer):
hidden, target = inputs
head_logprob = 0
if self.n_clusters == 0:
softmax_b = tf.get_variable('bias', [n_token], initializer=tf.zeros_initializer())
softmax_b = tf.get_variable('bias', [self.config.vocab_size], initializer=tf.zeros_initializer())
output = self._logit(hidden, self.out_layers[0][0], self.out_layers[0][1], self.out_projs[0])
if target is not None:
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=target, logits=output)

64
transformers/modeling_tf_utils.py
View File

@@ -22,15 +22,16 @@ import logging
import os
import tensorflow as tf
from tensorflow.python.keras.saving import hdf5_format
import h5py
from .configuration_utils import PretrainedConfig
from .file_utils import cached_path, WEIGHTS_NAME, TF_WEIGHTS_NAME, TF2_WEIGHTS_NAME
from .file_utils import (TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, WEIGHTS_NAME, DUMMY_INPUTS,
cached_path, hf_bucket_url, is_remote_url)
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.
@@ -59,7 +60,7 @@ class TFPreTrainedModel(tf.keras.Model):
Returns:
tf.Tensor with dummy inputs
"""
return tf.constant(DUMMY_INPUTS)
return {'input_ids': tf.constant(DUMMY_INPUTS)}
def __init__(self, config, *inputs, **kwargs):
super(TFPreTrainedModel, self).__init__(*inputs, **kwargs)
@@ -176,13 +177,16 @@ class TFPreTrainedModel(tf.keras.Model):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
- a path or url to a `PyTorch state_dict save file` (e.g. `./pt_model/pytorch_model.bin`). In this case, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the PyTorch checkpoint in a TensorFlow model using the provided conversion scripts and loading the TensorFlow 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`:
config: (`optional`) one of:
- an instance of a class derived from :class:`~transformers.PretrainedConfig`, or
- a string valid as input to :func:`~transformers.PretrainedConfig.from_pretrained()`
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
@@ -206,6 +210,9 @@ class TFPreTrainedModel(tf.keras.Model):
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:
@@ -229,11 +236,13 @@ class TFPreTrainedModel(tf.keras.Model):
force_download = kwargs.pop('force_download', False)
resume_download = kwargs.pop('resume_download', False)
proxies = kwargs.pop('proxies', None)
output_loading_info = kwargs.pop('output_loading_info', False)
# Load config
if config is None:
# Load config if we don't provide a configuration
if not isinstance(config, PretrainedConfig):
config_path = config if config is not None else pretrained_model_name_or_path
config, model_kwargs = cls.config_class.from_pretrained(
pretrained_model_name_or_path, *model_args,
config_path, *model_args,
cache_dir=cache_dir, return_unused_kwargs=True,
force_download=force_download,
resume_download=resume_download,
@@ -257,10 +266,14 @@ class TFPreTrainedModel(tf.keras.Model):
raise EnvironmentError("Error no file named {} found in directory {} or `from_pt` set to False".format(
[WEIGHTS_NAME, TF2_WEIGHTS_NAME],
pretrained_model_name_or_path))
elif os.path.isfile(pretrained_model_name_or_path):
elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
archive_file = pretrained_model_name_or_path
elif os.path.isfile(pretrained_model_name_or_path + ".index"):
archive_file = pretrained_model_name_or_path + ".index"
else:
raise EnvironmentError("Error file {} not found".format(pretrained_model_name_or_path))
archive_file = hf_bucket_url(pretrained_model_name_or_path, postfix=TF2_WEIGHTS_NAME)
if from_pt:
raise EnvironmentError("Loading a TF model from a PyTorch checkpoint is not supported when using a model identifier name.")
# redirect to the cache, if necessary
try:
@@ -293,17 +306,46 @@ class TFPreTrainedModel(tf.keras.Model):
if from_pt:
# Load from a PyTorch checkpoint
return load_pytorch_checkpoint_in_tf2_model(model, resolved_archive_file)
return load_pytorch_checkpoint_in_tf2_model(model, resolved_archive_file, allow_missing_keys=True)
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
try:
model.load_weights(resolved_archive_file, by_name=True)
except OSError:
raise OSError("Unable to load weights from h5 file. "
"If you tried to load a TF 2.0 model from a PyTorch checkpoint, please set from_pt=True. ")
ret = model(model.dummy_inputs, training=False) # Make sure restore ops are run
# Check if the models are the same to output loading informations
with h5py.File(resolved_archive_file, 'r') as f:
if 'layer_names' not in f.attrs and 'model_weights' in f:
f = f['model_weights']
hdf5_layer_names = set(hdf5_format.load_attributes_from_hdf5_group(f, 'layer_names'))
model_layer_names = set(layer.name for layer in model.layers)
missing_keys = list(model_layer_names - hdf5_layer_names)
unexpected_keys = list(hdf5_layer_names - model_layer_names)
error_msgs = []
if len(missing_keys) > 0:
logger.info("Layers of {} not initialized from pretrained model: {}".format(
model.__class__.__name__, missing_keys))
if len(unexpected_keys) > 0:
logger.info("Layers from pretrained model not used in {}: {}".format(
model.__class__.__name__, unexpected_keys))
if len(error_msgs) > 0:
raise RuntimeError('Error(s) in loading weights for {}:\n\t{}'.format(
model.__class__.__name__, "\n\t".join(error_msgs)))
if output_loading_info:
loading_info = {"missing_keys": missing_keys,
"unexpected_keys": unexpected_keys,
"error_msgs": error_msgs}
return model, loading_info
return model
class TFConv1D(tf.keras.layers.Layer):

10
transformers/modeling_tf_xlm.py
View File

@@ -460,7 +460,7 @@ class TFXLMPreTrainedModel(TFPreTrainedModel):
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]
return {'input_ids': inputs_list, 'attention_mask': attns_list, 'langs': langs_list}
XLM_START_DOCSTRING = r""" The XLM model was proposed in
@@ -576,7 +576,7 @@ class TFXLMModel(TFXLMPreTrainedModel):
tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
model = TFXLMModel.from_pretrained('xlm-mlm-en-2048')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[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
@@ -649,7 +649,7 @@ class TFXLMWithLMHeadModel(TFXLMPreTrainedModel):
tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
model = TFXLMWithLMHeadModel.from_pretrained('xlm-mlm-en-2048')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[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
@@ -695,7 +695,7 @@ class TFXLMForSequenceClassification(TFXLMPreTrainedModel):
tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
model = TFXLMForSequenceClassification.from_pretrained('xlm-mlm-en-2048')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
labels = tf.constant([1])[None, :] # Batch size 1
outputs = model(input_ids)
logits = outputs[0]
@@ -743,7 +743,7 @@ class TFXLMForQuestionAnsweringSimple(TFXLMPreTrainedModel):
tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
model = TFXLMForQuestionAnsweringSimple.from_pretrained('xlm-mlm-en-2048')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
start_scores, end_scores = outputs[:2]

14
transformers/modeling_tf_xlnet.py
View File

@@ -366,7 +366,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
self.use_bfloat16 = config.use_bfloat16
self.initializer_range = config.initializer_range
self.word_embedding = TFSharedEmbeddings(config.n_token, config.d_model, initializer_range=config.initializer_range, name='word_embedding')
self.word_embedding = TFSharedEmbeddings(config.vocab_size, config.d_model, initializer_range=config.initializer_range, name='word_embedding')
self.layer = [TFXLNetLayer(config, name='layer_._{}'.format(i)) for i in range(config.n_layer)]
self.dropout = tf.keras.layers.Dropout(config.dropout)
@@ -552,7 +552,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
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
input_mask = 1.0 - tf.cast(attention_mask, dtype=dtype_float)
if input_mask is not None and perm_mask is not None:
data_mask = input_mask[None] + perm_mask
elif input_mask is not None and perm_mask is None:
@@ -811,7 +811,7 @@ class TFXLNetModel(TFXLNetPreTrainedModel):
tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
model = TFXLNetModel.from_pretrained('xlnet-large-cased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[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
@@ -855,7 +855,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
model = TFXLNetLMHeadModel.from_pretrained('xlnet-large-cased')
# We show how to setup inputs to predict a next token using a bi-directional context.
input_ids = tf.constant(tokenizer.encode("Hello, my dog is very <mask>"))[None, :] # We will predict the masked token
input_ids = tf.constant(tokenizer.encode("Hello, my dog is very <mask>", add_special_tokens=True))[None, :] # We will predict the masked token
perm_mask = tf.zeros((1, input_ids.shape[1], input_ids.shape[1]))
perm_mask[:, :, -1] = 1.0 # Previous tokens don't see last token
target_mapping = tf.zeros((1, 1, input_ids.shape[1])) # Shape [1, 1, seq_length] => let's predict one token
@@ -911,7 +911,7 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel):
tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
model = TFXLNetForSequenceClassification.from_pretrained('xlnet-large-cased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
logits = outputs[0]
@@ -1022,7 +1022,7 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
tokenizer = XLNetTokenizer.from_pretrained('xlnet-base-cased')
model = TFXLNetForQuestionAnsweringSimple.from_pretrained('xlnet-base-cased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
outputs = model(input_ids)
start_scores, end_scores = outputs[:2]
@@ -1086,7 +1086,7 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
# tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
# model = XLMForQuestionAnswering.from_pretrained('xlnet-large-cased')
# input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
# input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
# start_positions = tf.constant([1])
# end_positions = tf.constant([3])
# outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)

24
transformers/modeling_transfo_xl.py
View File

@@ -36,7 +36,7 @@ from torch.nn.parameter import Parameter
from .modeling_utils import PreTrainedModel, Conv1D, prune_conv1d_layer, SequenceSummary
from .configuration_transfo_xl import TransfoXLConfig
from .modeling_transfo_xl_utilities import ProjectedAdaptiveLogSoftmax, sample_logits
from .modeling_transfo_xl_utilities import ProjectedAdaptiveLogSoftmax, sample_logits, LogUniformSampler
from .file_utils import add_start_docstrings
logger = logging.getLogger(__name__)
@@ -582,7 +582,7 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
model = TransfoXLModel.from_pretrained('transfo-xl-wt103')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
outputs = model(input_ids)
last_hidden_states, mems = outputs[:2]
@@ -592,14 +592,14 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.n_token = config.n_token
self.n_token = config.vocab_size
self.d_embed = config.d_embed
self.d_model = config.d_model
self.n_head = config.n_head
self.d_head = config.d_head
self.word_emb = AdaptiveEmbedding(config.n_token, config.d_embed, config.d_model, config.cutoffs,
self.word_emb = AdaptiveEmbedding(config.vocab_size, config.d_embed, config.d_model, config.cutoffs,
div_val=config.div_val)
self.drop = nn.Dropout(config.dropout)
@@ -825,7 +825,7 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
model = TransfoXLLMHeadModel.from_pretrained('transfo-xl-wt103')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
outputs = model(input_ids)
prediction_scores, mems = outputs[:2]
@@ -836,11 +836,11 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
self.sample_softmax = config.sample_softmax
# use sampled softmax
if config.sample_softmax > 0:
self.out_layer = nn.Linear(config.d_model, config.n_token)
self.sampler = LogUniformSampler(config.n_token, config.sample_softmax)
self.out_layer = nn.Linear(config.d_model, config.vocab_size)
self.sampler = LogUniformSampler(config.vocab_size, config.sample_softmax)
# use adaptive softmax (including standard softmax)
else:
self.crit = ProjectedAdaptiveLogSoftmax(config.n_token, config.d_embed, config.d_model,
self.crit = ProjectedAdaptiveLogSoftmax(config.vocab_size, config.d_embed, config.d_model,
config.cutoffs, div_val=config.div_val)
self.init_weights()
@@ -908,3 +908,11 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
outputs = [softmax_output, None] + outputs
return outputs # (loss), logits or None if labels is not None (speed up adaptive softmax), new_mems, (all hidden states), (all attentions)
def get_output_embeddings(self):
""" Double-check if you are using adaptive softmax.
"""
if self.sample_softmax > 0:
return self.out_layer
else:
return self.crit.out_layers[-1]

450
transformers/modeling_utils.py
View File

@@ -1,5 +1,5 @@
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright 2018 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
@@ -31,11 +31,11 @@ from torch.nn import CrossEntropyLoss
from torch.nn import functional as F
from .configuration_utils import PretrainedConfig
from .file_utils import cached_path, WEIGHTS_NAME, TF_WEIGHTS_NAME, TF2_WEIGHTS_NAME
from .file_utils import (TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, WEIGHTS_NAME, DUMMY_INPUTS,
cached_path, hf_bucket_url, is_remote_url)
logger = logging.getLogger(__name__)
try:
from torch.nn import Identity
except ImportError:
@@ -71,6 +71,15 @@ class PreTrainedModel(nn.Module):
load_tf_weights = lambda model, config, path: None
base_model_prefix = ""
@property
def dummy_inputs(self):
""" Dummy inputs to do a forward pass in the network.
Returns:
torch.Tensor with dummy inputs
"""
return {'input_ids': torch.tensor(DUMMY_INPUTS)}
def __init__(self, config, *inputs, **kwargs):
super(PreTrainedModel, self).__init__()
if not isinstance(config, PretrainedConfig):
@@ -160,7 +169,6 @@ class PreTrainedModel(nn.Module):
base_model.vocab_size = new_num_tokens
# Tie weights again if needed
if hasattr(self, 'tie_weights'):
self.tie_weights()
return model_embeds
@@ -265,6 +273,7 @@ class PreTrainedModel(nn.Module):
pretrained_model_name_or_path: either:
- a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
- a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
- a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
- 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.
- None if you are both providing the configuration and state dictionary (resp. with keyword arguments ``config`` and ``state_dict``)
@@ -272,7 +281,9 @@ class PreTrainedModel(nn.Module):
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`:
config: (`optional`) one of:
- an instance of a class derived from :class:`~transformers.PretrainedConfig`, or
- a string valid as input to :func:`~transformers.PretrainedConfig.from_pretrained()`
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
@@ -318,11 +329,6 @@ class PreTrainedModel(nn.Module):
model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_tf=True, config=config)
"""
if pretrained_model_name_or_path is not None and (
"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)
@@ -332,10 +338,11 @@ class PreTrainedModel(nn.Module):
proxies = kwargs.pop('proxies', None)
output_loading_info = kwargs.pop('output_loading_info', False)
# Load config
if config is None:
# Load config if we don't provide a configuration
if not isinstance(config, PretrainedConfig):
config_path = config if config is not None else pretrained_model_name_or_path
config, model_kwargs = cls.config_class.from_pretrained(
pretrained_model_name_or_path, *model_args,
config_path, *model_args,
cache_dir=cache_dir, return_unused_kwargs=True,
force_download=force_download,
resume_download=resume_download,
@@ -363,11 +370,16 @@ class PreTrainedModel(nn.Module):
raise EnvironmentError("Error no file named {} found in directory {} or `from_tf` set to False".format(
[WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME + ".index"],
pretrained_model_name_or_path))
elif os.path.isfile(pretrained_model_name_or_path):
elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
archive_file = pretrained_model_name_or_path
else:
assert from_tf, "Error finding file {}, no file or TF 1.X checkpoint found".format(pretrained_model_name_or_path)
elif os.path.isfile(pretrained_model_name_or_path + ".index"):
assert from_tf, "We found a TensorFlow checkpoint at {}, please set from_tf to True to load from this checkpoint".format(
pretrained_model_name_or_path + ".index")
archive_file = pretrained_model_name_or_path + ".index"
else:
archive_file = hf_bucket_url(pretrained_model_name_or_path, postfix=WEIGHTS_NAME)
if from_tf:
raise EnvironmentError("Loading a PyTorch model from a TF checkpoint is not supported when using a model identifier name.")
# redirect to the cache, if necessary
try:
@@ -399,7 +411,11 @@ class PreTrainedModel(nn.Module):
model = cls(config, *model_args, **model_kwargs)
if state_dict is None and not from_tf:
try:
state_dict = torch.load(resolved_archive_file, map_location='cpu')
except:
raise OSError("Unable to load weights from pytorch checkpoint file. "
"If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True. ")
missing_keys = []
unexpected_keys = []
@@ -428,8 +444,6 @@ class PreTrainedModel(nn.Module):
new_key = key.replace('gamma', 'weight')
if 'beta' in key:
new_key = key.replace('beta', 'bias')
if key == 'lm_head.decoder.weight':
new_key = 'lm_head.weight'
if new_key:
old_keys.append(key)
new_keys.append(new_key)
@@ -471,8 +485,7 @@ class PreTrainedModel(nn.Module):
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
model.__class__.__name__, "\n\t".join(error_msgs)))
if hasattr(model, 'tie_weights'):
model.tie_weights() # make sure word embedding weights are still tied
model.tie_weights() # make sure word embedding weights are still tied if needed
# Set model in evaluation mode to desactivate DropOut modules by default
model.eval()
@@ -483,6 +496,403 @@ class PreTrainedModel(nn.Module):
return model
def prepare_inputs_for_generation(self, input_ids, **kwargs):
return {"input_ids": input_ids}
@torch.no_grad()
def generate(self, input_ids=None, max_length=None, do_sample=None, num_beams=None,
temperature=None, top_k=None, top_p=None, repetition_penalty=None,
bos_token_id=None, pad_token_id=None, eos_token_ids=None,
length_penalty=None, num_return_sequences=None):
""" Sequence generator for models with a LM head.
The method currently supports greedy or penalized greedy decoding, sampling with top-k or nucleus sampling
and beam-search.
Adapted in part from Facebook's XLM beam search code: https://github.com/facebookresearch/XLM
Params:
**input_ids**: (`optional`) `torch.LongTensor` of shape (1, sequence_length)
The sequence used as a prompt for the generation. If `None` the method initializes
it as an empty `torch.LongTensor` of shape (1,)
**max_length**: (`optional`) int
The max length of the sequence to be generated. Between 1 and infinity. Default to 20.
**do_sample**: (`optional`) bool
If set to `False` we use greedy decoding; otherwise sampling. Default to greedy sampling.
**num_beams**: (`optional`) int
Number of beams for beam search. 1 means no beam serach. Default to 1.
**temperature**: (`optional`) float
The value used to module the next token probabilities.
**top_k**: (`optional`) int
The number of highest probability vocabulary tokens to keep for top-k-filtering. Between 1 and infinity. Default to 50.
**top_p**: (`optional`) float
The cumulative probability of parameter highest probability vocabulary tokens to keep for nucleus sampling. Must be between 0 and 1. Default to 1.
**repetition_penalty**: (`optional`) float
The parameter for repetition penalty. Between 1.0 and + infinity. 1.0 means no penalty. Default to 1.
**bos_token_id**: (`optional`) int
Beginning of sentence token if no prompt is provided. Default to 0.
**eos_token_ids**: (`optional`) int or list of int
End of sequence token or list of tokens to stop the generation. Default to 0.
**length_penalty**: (`optional`) int
Exponential penalty to the length. Default to 0.
**length_penalty**: (`optional`) float
Exponential penalty to the length. Default to 1.
**num_return_sequences**: (`optional`) int
The number of independantly computed returned sequences for each element in the batch. Default to 1.
"""
# We cannot generate if the model does not have a LM head
if self.get_output_embeddings() is None:
raise AttributeError("You tried to generate sequences with a model that does not have a LM Head."
"Please use another model class (e.g. `OpenAIGPTLMHeadModel`)")
max_length = max_length if max_length is not None else self.config.max_length
do_sample = do_sample if do_sample is not None else self.config.do_sample
num_beams = num_beams if num_beams is not None else self.config.num_beams
temperature = temperature if temperature is not None else self.config.temperature
top_k = top_k if top_k is not None else self.config.top_k
top_p = top_p if top_p is not None else self.config.top_p
repetition_penalty = repetition_penalty if repetition_penalty is not None else self.config.repetition_penalty
bos_token_id = bos_token_id if bos_token_id is not None else self.config.bos_token_id
pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id
eos_token_ids = eos_token_ids if eos_token_ids is not None else self.config.eos_token_ids
length_penalty = length_penalty if length_penalty is not None else self.config.length_penalty
num_return_sequences = num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
if input_ids is not None:
batch_size = input_ids.shape[0] # overriden by the input batch_size
else:
batch_size = 1
if isinstance(eos_token_ids, int):
eos_token_ids = [eos_token_ids]
assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictely positive integer."
assert isinstance(do_sample, bool), "`do_sample` should be a boolean."
assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictely positive integer."
# assert temperature >= 0, "`temperature` should be positive."
assert isinstance(top_k, int) and top_k >= 0, "`top_k` should be a positive integer."
assert 0 <= top_p <= 1, "`top_p` should be between 0 and 1."
assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
assert isinstance(bos_token_id, int) and bos_token_id >= 0, "`bos_token_id` should be a positive integer."
assert isinstance(pad_token_id, int) and pad_token_id >= 0, "`pad_token_id` should be a positive integer."
assert isinstance(eos_token_ids, (list, tuple)) and (e >= 0 for e in eos_token_ids), \
"`eos_token_ids` should be a positive integer or a list/tuple of positive integers."
assert length_penalty > 0, "`length_penalty` should be strictely positive."
assert isinstance(num_return_sequences, int) and num_return_sequences > 0, "`num_return_sequences` should be a strictely positive integer."
if input_ids is None:
input_ids = torch.full((batch_size, 1), bos_token_id, dtype=torch.long, device=next(self.parameters()).device)
else:
assert input_ids.dim() == 2, "Input prompt should be of shape (batch_size, sequence length)."
# current position and vocab size
cur_len = input_ids.shape[1]
vocab_size = self.config.vocab_size
if num_return_sequences != 1:
# Expand input to num return sequences
input_ids = input_ids.unsqueeze(1).expand(batch_size, num_return_sequences, cur_len)
input_ids = input_ids.contiguous().view(batch_size * num_return_sequences, cur_len) # (batch_size * num_return_sequences, cur_len)
effective_batch_size = batch_size * num_return_sequences
else:
effective_batch_size = batch_size
if num_beams > 1:
output = self._generate_beam_search(input_ids, cur_len, max_length, do_sample,
temperature, top_k, top_p, repetition_penalty,
pad_token_id, eos_token_ids, effective_batch_size,
length_penalty, num_beams, vocab_size)
else:
output = self._generate_no_beam_search(input_ids, cur_len, max_length, do_sample,
temperature, top_k, top_p, repetition_penalty,
pad_token_id, eos_token_ids, effective_batch_size)
if num_return_sequences != 1:
output = output.view(batch_size, num_return_sequences, -1)
return output
def _generate_no_beam_search(self, input_ids, cur_len, max_length, do_sample,
temperature, top_k, top_p, repetition_penalty,
pad_token_id, eos_token_ids, batch_size):
""" Generate sequences for each example without beam search (num_beams == 1).
All returned sequence are generated independantly.
"""
# current position / max lengths / length of generated sentences / unfinished sentences
unfinished_sents = input_ids.new(batch_size).fill_(1)
# TODO: add cached compute states
pasts = None
while cur_len < max_length:
model_inputs = self.prepare_inputs_for_generation(input_ids, pasts=pasts)
outputs = self(**model_inputs)
next_token_logits = outputs[0][:, -1, :]
# repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858)
if repetition_penalty != 1.0:
for i in range(batch_size):
for previous_tokens in set(input_ids[i].tolist()):
next_token_logits[i, previous_tokens] /= repetition_penalty
if do_sample:
# Temperature (higher temperature => more likely to sample low probability tokens)
if temperature > 0 and temperature != 1.0:
next_token_logits = next_token_logits / temperature
# Top-p/top-k filtering
next_token_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
# Sample
next_token = torch.multinomial(F.softmax(next_token_logits, dim=-1), num_samples=1).squeeze(1)
else:
# Greedy decoding
next_token = torch.argmax(next_token_logits, dim=-1)
# update generations and finished sentences
tokens_to_add = next_token * unfinished_sents + pad_token_id * (1 - unfinished_sents)
input_ids = torch.cat([input_ids, tokens_to_add.unsqueeze(-1)], dim=-1)
for eos_token_id in eos_token_ids:
unfinished_sents.mul_(tokens_to_add.ne(eos_token_id).long())
cur_len = cur_len + 1
# stop when there is a </s> in each sentence, or if we exceed the maximul length
if unfinished_sents.max() == 0:
break
# add eos_token_ids to unfinished sentences
if cur_len == max_length:
input_ids[:, -1].masked_fill_(unfinished_sents.to(dtype=torch.bool), eos_token_ids[0])
return input_ids
def _generate_beam_search(self, input_ids, cur_len, max_length, do_sample,
temperature, top_k, top_p, repetition_penalty,
pad_token_id, eos_token_ids, batch_size,
length_penalty, num_beams, vocab_size):
""" Generate sequences for each example with beam search.
"""
# Expand input to num beams
input_ids = input_ids.unsqueeze(1).expand(batch_size, num_beams, cur_len)
input_ids = input_ids.contiguous().view(batch_size * num_beams, cur_len) # (batch_size * num_beams, cur_len)
# generated hypotheses
generated_hyps = [BeamHypotheses(num_beams, max_length, length_penalty, early_stopping=False) for _ in range(batch_size)]
# scores for each sentence in the beam
beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
beam_scores[:, 1:] = -1e9
beam_scores = beam_scores.view(-1) # shape (batch_size * num_beams,)
# cache compute states
pasts = None # self.prepare_pasts()
# done sentences
done = [False for _ in range(batch_size)]
while cur_len < max_length:
model_inputs = self.prepare_inputs_for_generation(input_ids, pasts=pasts)
scores = self(**model_inputs)[0] # (batch_size * num_beams, cur_len, vocab_size)
scores = scores[:, -1, :] # (batch_size * num_beams, vocab_size)
# repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858)
if repetition_penalty != 1.0:
for i in range(batch_size * num_beams):
for previous_tokens in set(input_ids[i].tolist()):
scores[i, previous_tokens] /= repetition_penalty
if do_sample:
# Temperature (higher temperature => more likely to sample low probability tokens)
if temperature > 0 and temperature != 1.0:
scores = scores / temperature
# Top-p/top-k filtering
scores = top_k_top_p_filtering(scores, top_k=top_k, top_p=top_p, min_tokens_to_keep=2) # (batch_size * num_beams, vocab_size)
# Sample 2 next words for each beam (so we have some spare tokens and match output of greedy beam search)
next_words = torch.multinomial(F.softmax(scores, dim=-1), num_samples=2) # (batch_size * num_beams, 2)
# Compute next scores
_scores = F.log_softmax(scores, dim=-1) # (batch_size * num_beams, vocab_size)
_scores = torch.gather(_scores, -1, next_words) # (batch_size * num_beams, 2)
next_scores = _scores + beam_scores[:, None].expand_as(_scores) # (batch_size * num_beams, 2)
# Match shape of greedy beam search
next_words = next_words.view(batch_size, 2 * num_beams) # (batch_size, 2 * num_beams)
next_scores = next_scores.view(batch_size, 2 * num_beams) # (batch_size, 2 * num_beams)
else:
# do greedy beam search
scores = F.log_softmax(scores, dim=-1) # (batch_size * num_beams, vocab_size)
assert scores.size() == (batch_size * num_beams, vocab_size)
# Add the log prob of the new beams to the log prob of the beginning of the sequence (sum of logs == log of the product)
_scores = scores + beam_scores[:, None].expand_as(scores) # (batch_size * num_beams, vocab_size)
# re-organize to group the beam together (we are keeping top hypothesis accross beams)
_scores = _scores.view(batch_size, num_beams * vocab_size) # (batch_size, num_beams * vocab_size)
next_scores, next_words = torch.topk(_scores, 2*num_beams, dim=1, largest=True, sorted=True)
assert next_scores.size() == next_words.size() == (batch_size, 2 * num_beams)
# next batch beam content
# list of (batch_size * num_beams) tuple(next hypothesis score, next word, current position in the batch)
next_batch_beam = []
# for each sentence
for batch_ex in range(batch_size):
# if we are done with this sentence
done[batch_ex] = done[batch_ex] or generated_hyps[batch_ex].is_done(next_scores[batch_ex].max().item())
if done[batch_ex]:
next_batch_beam.extend([(0, pad_token_id, 0)] * num_beams) # pad the batch
continue
# next sentence beam content
next_sent_beam = []
# next words for this sentence
for idx, score in zip(next_words[batch_ex], next_scores[batch_ex]):
# get beam and word IDs
beam_id = idx // vocab_size
word_id = idx % vocab_size
# end of sentence, or next word
if word_id.item() in eos_token_ids or cur_len + 1 == max_length:
generated_hyps[batch_ex].add(input_ids[batch_ex * num_beams + beam_id, :cur_len].clone(), score.item())
else:
next_sent_beam.append((score, word_id, batch_ex * num_beams + beam_id))
# the beam for next step is full
if len(next_sent_beam) == num_beams:
break
# update next beam content
assert len(next_sent_beam) == 0 if cur_len + 1 == max_length else num_beams
if len(next_sent_beam) == 0:
next_sent_beam = [(0, pad_token_id, 0)] * num_beams # pad the batch
next_batch_beam.extend(next_sent_beam)
assert len(next_batch_beam) == num_beams * (batch_ex + 1)
# sanity check / prepare next batch
assert len(next_batch_beam) == batch_size * num_beams
beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
beam_words = input_ids.new([x[1] for x in next_batch_beam])
beam_idx = input_ids.new([x[2] for x in next_batch_beam])
# re-order batch and internal states
input_ids = input_ids[beam_idx, :]
input_ids = torch.cat([input_ids, beam_words.unsqueeze(1)], dim=-1)
# TODO: Activate cache
# for k in cache.keys():
# if k != 'slen':
# cache[k] = (cache[k][0][beam_idx], cache[k][1][beam_idx])
# update current length
cur_len = cur_len + 1
# stop when we are done with each sentence
if all(done):
break
# visualize hypotheses
# print([len(x) for x in generated_hyps], cur_len)
# globals().update( locals() );
# !import code; code.interact(local=vars())
# for ii in range(batch_size):
# for ss, ww in sorted(generated_hyps[ii].hyp, key=lambda x: x[0], reverse=True):
# print("%.3f " % ss + " ".join(self.dico[x] for x in ww.tolist()))
# print("")
# select the best hypotheses
tgt_len = input_ids.new(batch_size)
best = []
for i, hypotheses in enumerate(generated_hyps):
best_hyp = max(hypotheses.hyp, key=lambda x: x[0])[1]
tgt_len[i] = len(best_hyp) + 1 # +1 for the <EOS> symbol
best.append(best_hyp)
# generate target batch
decoded = input_ids.new(batch_size, tgt_len.max().item()).fill_(pad_token_id)
for i, hypo in enumerate(best):
decoded[i, :tgt_len[i] - 1] = hypo
decoded[i, tgt_len[i] - 1] = eos_token_ids[0]
return decoded
def top_k_top_p_filtering(logits, top_k=0, top_p=1.0, filter_value=-float('Inf'), min_tokens_to_keep=1):
""" Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
Args:
logits: logits distribution shape (batch size, vocabulary size)
if top_k > 0: keep only top k tokens with highest probability (top-k filtering).
if top_p < 1.0: 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)
Make sure we keep at least min_tokens_to_keep per batch example in the output
From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
"""
if top_k > 0:
top_k = min(max(top_k, min_tokens_to_keep), logits.size(-1)) # Safety check
# Remove all tokens with a probability less than the last token of the top-k
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
logits[indices_to_remove] = filter_value
if top_p < 1.0:
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
# Remove tokens with cumulative probability above the threshold (token with 0 are kept)
sorted_indices_to_remove = cumulative_probs > top_p
if min_tokens_to_keep > 1:
# Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
sorted_indices_to_remove[..., :min_tokens_to_keep] = 0
# Shift the indices to the right to keep also the first token above the threshold
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
# 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
class BeamHypotheses(object):
def __init__(self, n_hyp, max_length, length_penalty, early_stopping):
"""
Initialize n-best list of hypotheses.
"""
self.max_length = max_length - 1 # ignoring bos_token
self.length_penalty = length_penalty
self.early_stopping = early_stopping
self.n_hyp = n_hyp
self.hyp = []
self.worst_score = 1e9
def __len__(self):
"""
Number of hypotheses in the list.
"""
return len(self.hyp)
def add(self, hyp, sum_logprobs):
"""
Add a new hypothesis to the list.
"""
score = sum_logprobs / len(hyp) ** self.length_penalty
if len(self) < self.n_hyp or score > self.worst_score:
self.hyp.append((score, hyp))
if len(self) > self.n_hyp:
sorted_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.hyp)])
del self.hyp[sorted_scores[0][1]]
self.worst_score = sorted_scores[1][0]
else:
self.worst_score = min(score, self.worst_score)
def is_done(self, best_sum_logprobs):
"""
If there are enough hypotheses and that none of the hypotheses being generated
can become better than the worst one in the heap, then we are done with this sentence.
"""
if len(self) < self.n_hyp:
return False
elif self.early_stopping:
return True
else:
return self.worst_score >= best_sum_logprobs / self.max_length ** self.length_penalty
class Conv1D(nn.Module):
def __init__(self, nf, nx):

32
transformers/modeling_xlm.py
View File

@@ -227,6 +227,16 @@ class XLMPreTrainedModel(PreTrainedModel):
def __init__(self, *inputs, **kwargs):
super(XLMPreTrainedModel, self).__init__(*inputs, **kwargs)
@property
def dummy_inputs(self):
inputs_list = torch.tensor([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
attns_list = torch.tensor([[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 = torch.tensor([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
else:
langs_list = None
return {'input_ids': inputs_list, 'attention_mask': attns_list, 'langs': langs_list}
def _init_weights(self, module):
""" Initialize the weights. """
if isinstance(module, nn.Embedding):
@@ -336,7 +346,7 @@ class XLMModel(XLMPreTrainedModel):
tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
model = XLMModel.from_pretrained('xlm-mlm-en-2048')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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
@@ -624,7 +634,7 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
model = XLMWithLMHeadModel.from_pretrained('xlm-mlm-en-2048')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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
@@ -639,6 +649,18 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
def get_output_embeddings(self):
return self.pred_layer.proj
def prepare_inputs_for_generation(self, input_ids, **kwargs):
mask_token_id = self.config.mask_token_id
lang_id = self.config.lang_id
mask_token = torch.full((1, 1), mask_token_id, dtype=torch.long, device=input_ids.device)
input_ids = torch.cat([input_ids, mask_token], dim=1)
if lang_id is not None:
langs = torch.full_like(input_ids, lang_id)
else:
langs = None
return {"input_ids": input_ids, "langs": langs}
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,
@@ -686,7 +708,7 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
model = XLMForSequenceClassification.from_pretrained('xlm-mlm-en-2048')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
labels = torch.tensor([1]).unsqueeze(0) # Batch size 1
outputs = model(input_ids, labels=labels)
loss, logits = outputs[:2]
@@ -770,7 +792,7 @@ class XLMForQuestionAnsweringSimple(XLMPreTrainedModel):
tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
model = XLMForQuestionAnsweringSimple.from_pretrained('xlm-mlm-en-2048')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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)
@@ -866,7 +888,7 @@ class XLMForQuestionAnswering(XLMPreTrainedModel):
tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
model = XLMForQuestionAnswering.from_pretrained('xlm-mlm-en-2048')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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)

298
transformers/modeling_xlm_roberta.py Normal file
View File

@@ -0,0 +1,298 @@
# coding=utf-8
# Copyright 2019 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 XLM-RoBERTa model. """
from __future__ import (absolute_import, division, print_function,
unicode_literals)
import logging
from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification, RobertaForMultipleChoice, RobertaForTokenClassification
from .configuration_xlm_roberta import XLMRobertaConfig
from .file_utils import add_start_docstrings
logger = logging.getLogger(__name__)
XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP = {
'xlm-roberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-base-pytorch_model.bin",
'xlm-roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-pytorch_model.bin",
'xlm-roberta-large-finetuned-conll02-dutch': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll02-dutch-pytorch_model.bin",
'xlm-roberta-large-finetuned-conll02-spanish': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll02-spanish-pytorch_model.bin",
'xlm-roberta-large-finetuned-conll03-english': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll03-english-pytorch_model.bin",
'xlm-roberta-large-finetuned-conll03-german': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll03-german-pytorch_model.bin",
}
XLM_ROBERTA_START_DOCSTRING = r""" The XLM-RoBERTa model was proposed in
`Unsupervised Cross-lingual Representation Learning at Scale`_
by Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov. It is based on Facebook's RoBERTa model released in 2019.
It is a large multi-lingual language model, trained on 2.5TB of filtered CommonCrawl data.
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.
.. _`Unsupervised Cross-lingual Representation Learning at Scale`:
https://arxiv.org/abs/1911.02116
.. _`torch.nn.Module`:
https://pytorch.org/docs/stable/nn.html#module
Parameters:
config (:class:`~transformers.XLMRobertaConfig`): 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.
"""
XLM_ROBERTA_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, XLM-RoBERTa 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 XLMRobertaTokenizer.encode function with
the ``add_special_tokens`` parameter set to ``True``.
XLM-RoBERTa 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 XLM-RoBERTa 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 XLM-RoBERTa Model transformer outputting raw hidden-states without any specific head on top.",
XLM_ROBERTA_START_DOCSTRING, XLM_ROBERTA_INPUTS_DOCSTRING)
class XLMRobertaModel(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, XLM-RoBERTa input sequence should be formatted with <s> and </s> tokens as follows:
(a) For sequence pairs:
``tokens: <s> is this jack ##son ##ville ? </s> </s> no it is not . </s>``
``token_type_ids: 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1``
(b) For single sequences:
``tokens: <s> the dog is hairy . </s>``
``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 = XLMRobertaTokenizer.from_pretrained('xlm-roberta-large')
model = XLMRobertaModel.from_pretrained('xlm-roberta-large')
input_ids = torch.tensor(tokenizer.encode("Schloß Nymphenburg ist sehr schön .")).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 = XLMRobertaConfig
pretrained_model_archive_map = XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
@add_start_docstrings("""XLM-RoBERTa Model with a `language modeling` head on top. """,
XLM_ROBERTA_START_DOCSTRING, XLM_ROBERTA_INPUTS_DOCSTRING)
class XLMRobertaForMaskedLM(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 = XLMRobertaTokenizer.from_pretrained('xlm-roberta-large')
model = XLMRobertaForMaskedLM.from_pretrained('xlm-roberta-large')
input_ids = torch.tensor(tokenizer.encode("Schloß Nymphenburg ist sehr schön .")).unsqueeze(0) # Batch size 1
outputs = model(input_ids, masked_lm_labels=input_ids)
loss, prediction_scores = outputs[:2]
"""
config_class = XLMRobertaConfig
pretrained_model_archive_map = XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
@add_start_docstrings("""XLM-RoBERTa Model transformer with a sequence classification/regression head on top (a linear layer
on top of the pooled output) e.g. for GLUE tasks. """,
XLM_ROBERTA_START_DOCSTRING, XLM_ROBERTA_INPUTS_DOCSTRING)
class XLMRobertaForSequenceClassification(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 = XLMRobertaTokenizer.from_pretrained('xlm-roberta-large')
model = XLMRobertaForSequenceClassification.from_pretrained('xlm-roberta-large')
input_ids = torch.tensor(tokenizer.encode("Schloß Nymphenburg ist sehr schön .")).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 = XLMRobertaConfig
pretrained_model_archive_map = XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
@add_start_docstrings("""XLM-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. """,
XLM_ROBERTA_START_DOCSTRING, XLM_ROBERTA_INPUTS_DOCSTRING)
class XLMRobertaForMultipleChoice(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 = XLMRobertaTokenizer.from_pretrained('xlm-roberta-large')
model = XLMRobertaForMultipleChoice.from_pretrained('xlm-roberta-large')
choices = ["Schloß Nymphenburg ist sehr schön .", "Der Schloßkanal auch !"]
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 = XLMRobertaConfig
pretrained_model_archive_map = XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
@add_start_docstrings("""XLM-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. """,
XLM_ROBERTA_START_DOCSTRING, XLM_ROBERTA_INPUTS_DOCSTRING)
class XLMRobertaForTokenClassification(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 = XLMRobertaTokenizer.from_pretrained('xlm-roberta-large')
model = XLMRobertaForTokenClassification.from_pretrained('xlm-roberta-large')
input_ids = torch.tensor(tokenizer.encode("Schloß Nymphenburg ist sehr schön .", 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 = XLMRobertaConfig
pretrained_model_archive_map = XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP

38
transformers/modeling_xlnet.py
View File

@@ -589,7 +589,7 @@ class XLNetModel(XLNetPreTrainedModel):
tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
model = XLNetModel.from_pretrained('xlnet-large-cased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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
@@ -609,7 +609,7 @@ class XLNetModel(XLNetPreTrainedModel):
self.clamp_len = config.clamp_len
self.n_layer = config.n_layer
self.word_embedding = nn.Embedding(config.n_token, config.d_model)
self.word_embedding = nn.Embedding(config.vocab_size, config.d_model)
self.mask_emb = nn.Parameter(torch.FloatTensor(1, 1, config.d_model))
self.layer = nn.ModuleList([XLNetLayer(config) for _ in range(config.n_layer)])
self.dropout = nn.Dropout(config.dropout)
@@ -925,7 +925,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
model = XLNetLMHeadModel.from_pretrained('xlnet-large-cased')
# We show how to setup inputs to predict a next token using a bi-directional context.
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is very <mask>")).unsqueeze(0) # We will predict the masked token
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is very <mask>", add_special_tokens=True)).unsqueeze(0) # We will predict the masked token
perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float)
perm_mask[:, :, -1] = 1.0 # Previous tokens don't see last token
target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float) # Shape [1, 1, seq_length] => let's predict one token
@@ -940,13 +940,37 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
self.same_length = config.same_length
self.transformer = XLNetModel(config)
self.lm_loss = nn.Linear(config.d_model, config.n_token, bias=True)
self.lm_loss = nn.Linear(config.d_model, config.vocab_size, bias=True)
self.init_weights()
def get_output_embeddings(self):
return self.lm_loss
def prepare_inputs_for_generation(self, input_ids, **model_kwargs):
# Add dummy token at the end (no attention on this one)
dummy_token = torch.zeros((1, 1), dtype=torch.long, device=input_ids.device)
input_ids = torch.cat([input_ids, dummy_token], dim=1)
# Build permutation mask so that previous tokens don't see last token
perm_mask = torch.zeros(
(input_ids.shape[0], input_ids.shape[1], input_ids.shape[1]),
dtype=torch.float, device=input_ids.device
)
perm_mask[:, :, -1] = 1.0
# We'll only predict the last token
target_mapping = torch.zeros(
(input_ids.shape[0], 1, input_ids.shape[1]),
dtype=torch.float, device=input_ids.device
)
target_mapping[0, 0, -1] = 1.0
return {"input_ids": input_ids,
"perm_mask": perm_mask,
"target_mapping": target_mapping
}
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,
@@ -1007,7 +1031,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
model = XLNetForSequenceClassification.from_pretrained('xlnet-large-cased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
labels = torch.tensor([1]).unsqueeze(0) # Batch size 1
outputs = model(input_ids, labels=labels)
loss, logits = outputs[:2]
@@ -1294,7 +1318,7 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
model = XLMForQuestionAnswering.from_pretrained('xlnet-large-cased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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)
@@ -1409,7 +1433,7 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
model = XLMForQuestionAnswering.from_pretrained('xlnet-large-cased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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)

907
transformers/pipelines.py Executable file
View File

@@ -0,0 +1,907 @@
# 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.
from __future__ import absolute_import, division, print_function, unicode_literals
import sys
import csv
import json
import os
import pickle
import logging
import six
from abc import ABC, abstractmethod
from contextlib import contextmanager
from itertools import groupby
from os.path import abspath, exists
from typing import Union, Optional, Tuple, List, Dict
import numpy as np
from transformers import (AutoConfig, AutoTokenizer, PreTrainedTokenizer,
PretrainedConfig, ModelCard, SquadExample,
squad_convert_examples_to_features, is_tf_available,
is_torch_available, BasicTokenizer,
ALL_PRETRAINED_CONFIG_ARCHIVE_MAP)
if is_tf_available():
import tensorflow as tf
from transformers import TFAutoModel, TFAutoModelForSequenceClassification, \
TFAutoModelForQuestionAnswering, TFAutoModelForTokenClassification
if is_torch_available():
import torch
from transformers import AutoModel, AutoModelForSequenceClassification, \
AutoModelForQuestionAnswering, AutoModelForTokenClassification
logger = logging.getLogger(__name__)
def get_framework(model=None):
""" Select framework (TensorFlow/PyTorch) to use.
If both frameworks are installed and no specific model is provided, defaults to using PyTorch.
"""
if is_tf_available() and is_torch_available() and model is not None and not isinstance(model, str):
# Both framework are available but the use supplied a model class instance.
# Try to guess which framework to use from the model classname
framework = 'tf' if model.__class__.__name__.startswith('TF') else 'pt'
elif not is_tf_available() and not is_torch_available():
raise ImportError("At least one of TensorFlow 2.0 or PyTorch should be installed. "
"To install TensorFlow 2.0, read the instructions at https://www.tensorflow.org/install/ "
"To install PyTorch, read the instructions at https://pytorch.org/.")
else:
# framework = 'tf' if is_tf_available() else 'pt'
framework = 'pt' if is_torch_available() else 'tf'
return framework
class ArgumentHandler(ABC):
"""
Base interface for handling varargs for each Pipeline
"""
@abstractmethod
def __call__(self, *args, **kwargs):
raise NotImplementedError()
class DefaultArgumentHandler(ArgumentHandler):
"""
Default varargs argument parser handling parameters for each Pipeline
"""
def __call__(self, *args, **kwargs):
if 'X' in kwargs:
return kwargs['X']
elif 'data' in kwargs:
return kwargs['data']
elif len(args) == 1:
if isinstance(args[0], list):
return args[0]
else:
return [args[0]]
elif len(args) > 1:
return list(args)
raise ValueError('Unable to infer the format of the provided data (X=, data=, ...)')
class PipelineDataFormat:
"""
Base class for all the pipeline supported data format both for reading and writing.
Supported data formats currently includes:
- JSON
- CSV
- stdin/stdout (pipe)
PipelineDataFormat also includes some utilities to work with multi-columns like mapping from datasets columns
to pipelines keyword arguments through the `dataset_kwarg_1=dataset_column_1` format.
"""
SUPPORTED_FORMATS = ['json', 'csv', 'pipe']
def __init__(self, output_path: Optional[str], input_path: Optional[str], column: Optional[str], overwrite=False):
self.output_path = output_path
self.input_path = input_path
self.column = column.split(',') if column is not None else ['']
self.is_multi_columns = len(self.column) > 1
if self.is_multi_columns:
self.column = [tuple(c.split('=')) if '=' in c else (c, c) for c in self.column]
if output_path is not None and not overwrite:
if exists(abspath(self.output_path)):
raise OSError('{} already exists on disk'.format(self.output_path))
if input_path is not None:
if not exists(abspath(self.input_path)):
raise OSError('{} doesnt exist on disk'.format(self.input_path))
@abstractmethod
def __iter__(self):
raise NotImplementedError()
@abstractmethod
def save(self, data: dict):
"""
Save the provided data object with the representation for the current `DataFormat`.
:param data: data to store
:return:
"""
raise NotImplementedError()
def save_binary(self, data: Union[dict, List[dict]]) -> str:
"""
Save the provided data object as a pickle-formatted binary data on the disk.
:param data: data to store
:return: (str) Path where the data has been saved
"""
path, _ = os.path.splitext(self.output_path)
binary_path = os.path.extsep.join((path, 'pickle'))
with open(binary_path, 'wb+') as f_output:
pickle.dump(data, f_output)
return binary_path
@staticmethod
def from_str(format: str, output_path: Optional[str], input_path: Optional[str], column: Optional[str], overwrite=False):
if format == 'json':
return JsonPipelineDataFormat(output_path, input_path, column, overwrite=overwrite)
elif format == 'csv':
return CsvPipelineDataFormat(output_path, input_path, column, overwrite=overwrite)
elif format == 'pipe':
return PipedPipelineDataFormat(output_path, input_path, column, overwrite=overwrite)
else:
raise KeyError('Unknown reader {} (Available reader are json/csv/pipe)'.format(format))
class CsvPipelineDataFormat(PipelineDataFormat):
def __init__(self, output_path: Optional[str], input_path: Optional[str], column: Optional[str], overwrite=False):
super().__init__(output_path, input_path, column, overwrite=overwrite)
def __iter__(self):
with open(self.input_path, 'r') as f:
reader = csv.DictReader(f)
for row in reader:
if self.is_multi_columns:
yield {k: row[c] for k, c in self.column}
else:
yield row[self.column[0]]
def save(self, data: List[dict]):
with open(self.output_path, 'w') as f:
if len(data) > 0:
writer = csv.DictWriter(f, list(data[0].keys()))
writer.writeheader()
writer.writerows(data)
class JsonPipelineDataFormat(PipelineDataFormat):
def __init__(self, output_path: Optional[str], input_path: Optional[str], column: Optional[str], overwrite=False):
super().__init__(output_path, input_path, column, overwrite=overwrite)
with open(input_path, 'r') as f:
self._entries = json.load(f)
def __iter__(self):
for entry in self._entries:
if self.is_multi_columns:
yield {k: entry[c] for k, c in self.column}
else:
yield entry[self.column[0]]
def save(self, data: dict):
with open(self.output_path, 'w') as f:
json.dump(data, f)
class PipedPipelineDataFormat(PipelineDataFormat):
"""
Read data from piped input to the python process.
For multi columns data, columns should separated by \t
If columns are provided, then the output will be a dictionary with {column_x: value_x}
"""
def __iter__(self):
for line in sys.stdin:
# Split for multi-columns
if '\t' in line:
line = line.split('\t')
if self.column:
# Dictionary to map arguments
yield {kwargs: l for (kwargs, _), l in zip(self.column, line)}
else:
yield tuple(line)
# No dictionary to map arguments
else:
yield line
def save(self, data: dict):
print(data)
def save_binary(self, data: Union[dict, List[dict]]) -> str:
if self.output_path is None:
raise KeyError(
'When using piped input on pipeline outputting large object requires an output file path. '
'Please provide such output path through --output argument.'
)
return super().save_binary(data)
class _ScikitCompat(ABC):
"""
Interface layer for the Scikit and Keras compatibility.
"""
@abstractmethod
def transform(self, X):
raise NotImplementedError()
@abstractmethod
def predict(self, X):
raise NotImplementedError()
class Pipeline(_ScikitCompat):
"""
Base class implementing pipelined operations.
Pipeline workflow is defined as a sequence of the following operations:
Input -> Tokenization -> Model Inference -> Post-Processing (Task dependent) -> Output
Pipeline supports running on CPU or GPU through the device argument. Users can specify
device argument as an integer, -1 meaning "CPU", >= 0 referring the CUDA device ordinal.
Some pipeline, like for instance FeatureExtractionPipeline ('feature-extraction') outputs large
tensor object as nested-lists. In order to avoid dumping such large structure as textual data we
provide the binary_output constructor argument. If set to True, the output will be stored in the
pickle format.
Arguments:
**model**: ``(str, PretrainedModel, TFPretrainedModel)``:
Reference to the model to use through this pipeline.
**tokenizer**: ``(str, PreTrainedTokenizer)``:
Reference to the tokenizer to use through this pipeline.
**args_parser**: ``ArgumentHandler``:
Reference to the object in charge of parsing supplied pipeline parameters.
**device**: ``int``:
Device ordinal for CPU/GPU supports. Setting this to -1 will leverage CPU, >=0 will run the model
on the associated CUDA device id.
**binary_output** ``bool`` (default: False):
Flag indicating if the output the pipeline should happen in a binary format (i.e. pickle) or as raw text.
Return:
Pipeline returns list or dictionary depending on:
- Does the user provided multiple sample
- The pipeline expose multiple fields in the output object
Examples:
nlp = pipeline('ner')
nlp = pipeline('ner', model='...', config='...', tokenizer='...')
nlp = NerPipeline(model='...', config='...', tokenizer='...')
nlp = QuestionAnsweringPipeline(model=AutoModel.from_pretrained('...'), tokenizer='...')
"""
default_input_names = None
def __init__(self, model, tokenizer: PreTrainedTokenizer = None,
modelcard: ModelCard = None, framework: Optional[str] = None,
args_parser: ArgumentHandler = None, device: int = -1,
binary_output: bool = False):
if framework is None:
framework = get_framework()
self.model = model
self.tokenizer = tokenizer
self.modelcard = modelcard
self.framework = framework
self.device = device
self.binary_output = binary_output
self._args_parser = args_parser or DefaultArgumentHandler()
# Special handling
if self.device >= 0 and self.framework == 'pt':
self.model = self.model.to('cuda:{}'.format(self.device))
def save_pretrained(self, save_directory):
"""
Save the pipeline's model and tokenizer to the specified save_directory
"""
if not os.path.isdir(save_directory):
logger.error("Provided path ({}) should be a directory".format(save_directory))
return
self.model.save_pretrained(save_directory)
self.tokenizer.save_pretrained(save_directory)
self.modelcard.save_pretrained(save_directory)
def transform(self, X):
"""
Scikit / Keras interface to transformers' pipelines. This method will forward to __call__().
"""
return self(X=X)
def predict(self, X):
"""
Scikit / Keras interface to transformers' pipelines. This method will forward to __call__().
Se
"""
return self(X=X)
@contextmanager
def device_placement(self):
"""
Context Manager allowing tensor allocation on the user-specified device in framework agnostic way.
example:
# Explicitly ask for tensor allocation on CUDA device :0
nlp = pipeline(..., device=0)
with nlp.device_placement():
# Every framework specific tensor allocation will be done on the request device
output = nlp(...)
Returns:
Context manager
"""
if self.framework == 'tf':
with tf.device('/CPU:0' if self.device == -1 else '/device:GPU:{}'.format(self.device)):
yield
else:
if self.device >= 0:
torch.cuda.set_device(self.device)
yield
def inputs_for_model(self, features: Union[dict, List[dict]]) -> Dict:
"""
Generates the input dictionary with model-specific parameters.
Returns:
dict holding all the required parameters for model's forward
"""
args = ['input_ids', 'attention_mask']
model_type = type(self.model).__name__.lower()
if 'distilbert' not in model_type and 'xlm' not in model_type:
args += ['token_type_ids']
# PR #1548 (CLI) There is an issue with attention_mask
# if 'xlnet' in model_type or 'xlm' in model_type:
# args += ['cls_index', 'p_mask']
if isinstance(features, dict):
return {k: features[k] for k in args}
else:
return {k: [feature[k] for feature in features] for k in args}
def __call__(self, *texts, **kwargs):
# Parse arguments
inputs = self._args_parser(*texts, **kwargs)
# Encode for forward
with self.device_placement():
inputs = self.tokenizer.batch_encode_plus(
inputs, add_special_tokens=True,
return_tensors=self.framework,
max_length=self.tokenizer.max_len
)
# Filter out features not available on specific models
inputs = self.inputs_for_model(inputs)
return self._forward(inputs)
def _forward(self, inputs):
"""
Internal framework specific forward dispatching.
Args:
inputs: dict holding all the keyworded arguments for required by the model forward method.
Returns:
Numpy array
"""
if self.framework == 'tf':
# TODO trace model
predictions = self.model(inputs, training=False)[0]
else:
with torch.no_grad():
predictions = self.model(**inputs)[0].cpu()
return predictions.numpy()
class FeatureExtractionPipeline(Pipeline):
"""
Feature extraction pipeline using Model head.
"""
def __init__(self, model,
tokenizer: PreTrainedTokenizer = None,
modelcard: ModelCard = None,
framework: Optional[str] = None,
args_parser: ArgumentHandler = None,
device: int = -1):
super().__init__(model=model,
tokenizer=tokenizer,
modelcard=modelcard,
framework=framework,
args_parser=args_parser,
device=device,
binary_output=True)
def __call__(self, *args, **kwargs):
return super().__call__(*args, **kwargs).tolist()
class TextClassificationPipeline(Pipeline):
"""
Text classification pipeline using ModelForTextClassification head.
"""
def __call__(self, *args, **kwargs):
outputs = super().__call__(*args, **kwargs)
scores = np.exp(outputs) / np.exp(outputs).sum(-1)
return [{'label': self.model.config.id2label[item.argmax()], 'score': item.max()} for item in scores]
class NerPipeline(Pipeline):
"""
Named Entity Recognition pipeline using ModelForTokenClassification head.
"""
default_input_names = 'sequences'
def __init__(self, model, tokenizer: PreTrainedTokenizer = None,
modelcard: ModelCard = None, framework: Optional[str] = None,
args_parser: ArgumentHandler = None, device: int = -1,
binary_output: bool = False, ignore_labels=['O']):
super().__init__(model=model,
tokenizer=tokenizer,
modelcard=modelcard,
framework=framework,
args_parser=args_parser,
device=device,
binary_output=binary_output)
self._basic_tokenizer = BasicTokenizer(do_lower_case=False)
self.ignore_labels = ignore_labels
def __call__(self, *texts, **kwargs):
inputs, answers = self._args_parser(*texts, **kwargs), []
for sentence in inputs:
# Manage correct placement of the tensors
with self.device_placement():
tokens = self.tokenizer.encode_plus(
sentence, return_attention_mask=False,
return_tensors=self.framework,
max_length=self.tokenizer.max_len
)
# Forward
if self.framework == 'tf':
entities = self.model(tokens)[0][0].numpy()
input_ids = tokens['input_ids'].numpy()[0]
else:
with torch.no_grad():
entities = self.model(**tokens)[0][0].cpu().numpy()
input_ids = tokens['input_ids'].cpu().numpy()[0]
score = np.exp(entities) / np.exp(entities).sum(-1, keepdims=True)
labels_idx = score.argmax(axis=-1)
answer = []
for idx, label_idx in enumerate(labels_idx):
if self.model.config.id2label[label_idx] not in self.ignore_labels:
answer += [{
'word': self.tokenizer.decode([int(input_ids[idx])]),
'score': score[idx][label_idx].item(),
'entity': self.model.config.id2label[label_idx]
}]
# Append
answers += [answer]
if len(answers) == 1:
return answers[0]
return answers
class QuestionAnsweringArgumentHandler(ArgumentHandler):
"""
QuestionAnsweringPipeline requires the user to provide multiple arguments (i.e. question & context) to be mapped
to internal SquadExample / SquadFeature structures.
QuestionAnsweringArgumentHandler manages all the possible to create SquadExample from the command-line supplied
arguments.
"""
def __call__(self, *args, **kwargs):
# Position args, handling is sensibly the same as X and data, so forwarding to avoid duplicating
if args is not None and len(args) > 0:
if len(args) == 1:
kwargs['X'] = args[0]
else:
kwargs['X'] = list(args)
# Generic compatibility with sklearn and Keras
# Batched data
if 'X' in kwargs or 'data' in kwargs:
inputs = kwargs['X'] if 'X' in kwargs else kwargs['data']
if isinstance(inputs, dict):
inputs = [inputs]
else:
# Copy to avoid overriding arguments
inputs = [i for i in inputs]
for i, item in enumerate(inputs):
if isinstance(item, dict):
if any(k not in item for k in ['question', 'context']):
raise KeyError('You need to provide a dictionary with keys {question:..., context:...}')
inputs[i] = QuestionAnsweringPipeline.create_sample(**item)
elif not isinstance(item, SquadExample):
raise ValueError(
'{} argument needs to be of type (list[SquadExample | dict], SquadExample, dict)'
.format('X' if 'X' in kwargs else 'data')
)
# Tabular input
elif 'question' in kwargs and 'context' in kwargs:
if isinstance(kwargs['question'], str):
kwargs['question'] = [kwargs['question']]
if isinstance(kwargs['context'], str):
kwargs['context'] = [kwargs['context']]
inputs = [QuestionAnsweringPipeline.create_sample(q, c) for q, c in zip(kwargs['question'], kwargs['context'])]
else:
raise ValueError('Unknown arguments {}'.format(kwargs))
if not isinstance(inputs, list):
inputs = [inputs]
return inputs
class QuestionAnsweringPipeline(Pipeline):
"""
Question Answering pipeline using ModelForQuestionAnswering head.
"""
default_input_names = 'question,context'
def __init__(self, model,
tokenizer: Optional[PreTrainedTokenizer],
modelcard: Optional[ModelCard],
framework: Optional[str] = None,
device: int = -1, **kwargs):
super().__init__(model=model,
tokenizer=tokenizer,
modelcard=modelcard,
framework=framework,
args_parser=QuestionAnsweringArgumentHandler(),
device=device, **kwargs)
@staticmethod
def create_sample(question: Union[str, List[str]], context: Union[str, List[str]]) -> Union[SquadExample, List[SquadExample]]:
"""
QuestionAnsweringPipeline leverages the SquadExample/SquadFeatures internally.
This helper method encapsulate all the logic for converting question(s) and context(s) to SquadExample(s).
We currently support extractive question answering.
Arguments:
question: (str, List[str]) The question to be ask for the associated context
context: (str, List[str]) The context in which we will look for the answer.
Returns:
SquadExample initialized with the corresponding question and context.
"""
if isinstance(question, list):
return [SquadExample(None, q, c, None, None, None) for q, c in zip(question, context)]
else:
return SquadExample(None, question, context, None, None, None)
def __call__(self, *texts, **kwargs):
"""
Args:
We support multiple use-cases, the following are exclusive:
X: sequence of SquadExample
data: sequence of SquadExample
question: (str, List[str]), batch of question(s) to map along with context
context: (str, List[str]), batch of context(s) associated with the provided question keyword argument
Returns:
dict: {'answer': str, 'score": float, 'start": int, "end": int}
answer: the textual answer in the intial context
score: the score the current answer scored for the model
start: the character index in the original string corresponding to the beginning of the answer' span
end: the character index in the original string corresponding to the ending of the answer' span
"""
# Set defaults values
kwargs.setdefault('topk', 1)
kwargs.setdefault('doc_stride', 128)
kwargs.setdefault('max_answer_len', 15)
kwargs.setdefault('max_seq_len', 384)
kwargs.setdefault('max_question_len', 64)
if kwargs['topk'] < 1:
raise ValueError('topk parameter should be >= 1 (got {})'.format(kwargs['topk']))
if kwargs['max_answer_len'] < 1:
raise ValueError('max_answer_len parameter should be >= 1 (got {})'.format(kwargs['max_answer_len']))
# Convert inputs to features
examples = self._args_parser(*texts, **kwargs)
features = squad_convert_examples_to_features(examples, self.tokenizer, kwargs['max_seq_len'], kwargs['doc_stride'], kwargs['max_question_len'], False)
fw_args = self.inputs_for_model([f.__dict__ for f in features])
# Manage tensor allocation on correct device
with self.device_placement():
if self.framework == 'tf':
fw_args = {k: tf.constant(v) for (k, v) in fw_args.items()}
start, end = self.model(fw_args)
start, end = start.numpy(), end.numpy()
else:
with torch.no_grad():
# Retrieve the score for the context tokens only (removing question tokens)
fw_args = {k: torch.tensor(v) for (k, v) in fw_args.items()}
start, end = self.model(**fw_args)
start, end = start.cpu().numpy(), end.cpu().numpy()
answers = []
for (example, feature, start_, end_) in zip(examples, features, start, end):
# Normalize logits and spans to retrieve the answer
start_ = np.exp(start_) / np.sum(np.exp(start_))
end_ = np.exp(end_) / np.sum(np.exp(end_))
# Mask padding and question
start_, end_ = start_ * np.abs(np.array(feature.p_mask) - 1), end_ * np.abs(np.array(feature.p_mask) - 1)
# TODO : What happens if not possible
# Mask CLS
start_[0] = end_[0] = 0
starts, ends, scores = self.decode(start_, end_, kwargs['topk'], kwargs['max_answer_len'])
char_to_word = np.array(example.char_to_word_offset)
# Convert the answer (tokens) back to the original text
answers += [
{
'score': score.item(),
'start': np.where(char_to_word == feature.token_to_orig_map[s])[0][0].item(),
'end': np.where(char_to_word == feature.token_to_orig_map[e])[0][-1].item(),
'answer': ' '.join(example.doc_tokens[feature.token_to_orig_map[s]:feature.token_to_orig_map[e] + 1])
}
for s, e, score in zip(starts, ends, scores)
]
if len(answers) == 1:
return answers[0]
return answers
def decode(self, start: np.ndarray, end: np.ndarray, topk: int, max_answer_len: int) -> Tuple:
"""
Take the output of any QuestionAnswering head and will generate probalities for each span to be
the actual answer.
In addition, it filters out some unwanted/impossible cases like answer len being greater than
max_answer_len or answer end position being before the starting position.
The method supports output the k-best answer through the topk argument.
Args:
start: numpy array, holding individual start probabilities for each token
end: numpy array, holding individual end probabilities for each token
topk: int, indicates how many possible answer span(s) to extract from the model's output
max_answer_len: int, maximum size of the answer to extract from the model's output
"""
# Ensure we have batch axis
if start.ndim == 1:
start = start[None]
if end.ndim == 1:
end = end[None]
# Compute the score of each tuple(start, end) to be the real answer
outer = np.matmul(np.expand_dims(start, -1), np.expand_dims(end, 1))
# Remove candidate with end < start and end - start > max_answer_len
candidates = np.tril(np.triu(outer), max_answer_len - 1)
# Inspired by Chen & al. (https://github.com/facebookresearch/DrQA)
scores_flat = candidates.flatten()
if topk == 1:
idx_sort = [np.argmax(scores_flat)]
elif len(scores_flat) < topk:
idx_sort = np.argsort(-scores_flat)
else:
idx = np.argpartition(-scores_flat, topk)[0:topk]
idx_sort = idx[np.argsort(-scores_flat[idx])]
start, end = np.unravel_index(idx_sort, candidates.shape)[1:]
return start, end, candidates[0, start, end]
def span_to_answer(self, text: str, start: int, end: int):
"""
When decoding from token probalities, this method maps token indexes to actual word in
the initial context.
Args:
text: str, the actual context to extract the answer from
start: int, starting answer token index
end: int, ending answer token index
Returns:
dict: {'answer': str, 'start': int, 'end': int}
"""
words = []
token_idx = char_start_idx = char_end_idx = chars_idx = 0
for i, word in enumerate(text.split(" ")):
token = self.tokenizer.tokenize(word)
# Append words if they are in the span
if start <= token_idx <= end:
if token_idx == start:
char_start_idx = chars_idx
if token_idx == end:
char_end_idx = chars_idx + len(word)
words += [word]
# Stop if we went over the end of the answer
if token_idx > end:
break
# Append the subtokenization length to the running index
token_idx += len(token)
chars_idx += len(word) + 1
# Join text with spaces
return {'answer': ' '.join(words), 'start': max(0, char_start_idx), 'end': min(len(text), char_end_idx)}
# Register all the supported task here
SUPPORTED_TASKS = {
'feature-extraction': {
'impl': FeatureExtractionPipeline,
'tf': TFAutoModel if is_tf_available() else None,
'pt': AutoModel if is_torch_available() else None,
'default': {
'model': {
'pt': 'distilbert-base-uncased',
'tf': 'distilbert-base-uncased',
},
'config': None,
'tokenizer': 'distilbert-base-uncased'
}
},
'sentiment-analysis': {
'impl': TextClassificationPipeline,
'tf': TFAutoModelForSequenceClassification if is_tf_available() else None,
'pt': AutoModelForSequenceClassification if is_torch_available() else None,
'default': {
'model': {
'pt': 'https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-finetuned-sst-2-english-pytorch_model.bin',
'tf': 'https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-finetuned-sst-2-english-tf_model.h5',
},
'config': 'https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-finetuned-sst-2-english-config.json',
'tokenizer': 'distilbert-base-uncased'
}
},
'ner': {
'impl': NerPipeline,
'tf': TFAutoModelForTokenClassification if is_tf_available() else None,
'pt': AutoModelForTokenClassification if is_torch_available() else None,
'default': {
'model': {
'pt':'https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-finetuned-conll03-english-pytorch_model.bin',
'tf': 'https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-finetuned-conll03-english-tf_model.h5',
},
'config': 'https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-finetuned-conll03-english-config.json',
'tokenizer': 'bert-large-cased'
}
},
'question-answering': {
'impl': QuestionAnsweringPipeline,
'tf': TFAutoModelForQuestionAnswering if is_tf_available() else None,
'pt': AutoModelForQuestionAnswering if is_torch_available() else None,
'default': {
'model': {
'pt': 'distilbert-base-uncased-distilled-squad',
'tf': 'distilbert-base-uncased-distilled-squad',
},
'config': None,
'tokenizer': 'distilbert-base-uncased'
}
}
}
def pipeline(task: str, model: Optional = None,
config: Optional[Union[str, PretrainedConfig]] = None,
tokenizer: Optional[Union[str, PreTrainedTokenizer]] = None,
modelcard: Optional[Union[str, ModelCard]] = None,
**kwargs) -> Pipeline:
"""
Utility factory method to build a pipeline.
Pipeline are made of:
A Tokenizer instance in charge of mapping raw textual input to token
A Model instance
Some (optional) post processing for enhancing model's output
Examples:
pipeline('sentiment-analysis')
pipeline('question-answering', model='distilbert-base-uncased-distilled-squad', tokenizer='bert-base-cased')
pipeline('ner', model=AutoModel.from_pretrained(...), tokenizer=AutoTokenizer.from_pretrained(...)
pipeline('ner', model='https://...pytorch-model.bin', config='https://...config.json', tokenizer='bert-base-cased')
"""
# Retrieve the task
if task not in SUPPORTED_TASKS:
raise KeyError("Unknown task {}, available tasks are {}".format(task, list(SUPPORTED_TASKS.keys())))
framework = get_framework(model)
targeted_task = SUPPORTED_TASKS[task]
task, model_class = targeted_task['impl'], targeted_task[framework]
# Use default model/config/tokenizer for the task if no model is provided
if model is None:
models, config, tokenizer = tuple(targeted_task['default'].values())
model = models[framework]
# Try to infer tokenizer from model or config name (if provided as str)
if tokenizer is None:
if isinstance(model, str) and model in ALL_PRETRAINED_CONFIG_ARCHIVE_MAP:
tokenizer = model
elif isinstance(config, str) and config in ALL_PRETRAINED_CONFIG_ARCHIVE_MAP:
tokenizer = config
else:
# Impossible to guest what is the right tokenizer here
raise Exception("Impossible to guess which tokenizer to use. "
"Please provided a PretrainedTokenizer class or a path/url/shortcut name to a pretrained tokenizer.")
# Try to infer modelcard from model or config name (if provided as str)
if modelcard is None:
# Try to fallback on one of the provided string for model or config (will replace the suffix)
if isinstance(model, str):
modelcard = model
elif isinstance(config, str):
modelcard = config
# Instantiate tokenizer if needed
if isinstance(tokenizer, six.string_types):
tokenizer = AutoTokenizer.from_pretrained(tokenizer)
# Instantiate config if needed
if isinstance(config, str):
config = AutoConfig.from_pretrained(config)
# Instantiate modelcard if needed
if isinstance(modelcard, str):
modelcard = ModelCard.from_pretrained(modelcard)
# Instantiate model if needed
if isinstance(model, str):
# Handle transparent TF/PT model conversion
model_kwargs = {}
if framework == 'pt' and model.endswith('.h5'):
model_kwargs['from_tf'] = True
logger.warning('Model might be a TensorFlow model (ending with `.h5`) but TensorFlow is not available. '
'Trying to load the model with PyTorch.')
elif framework == 'tf' and model.endswith('.bin'):
model_kwargs['from_pt'] = True
logger.warning('Model might be a PyTorch model (ending with `.bin`) but PyTorch is not available. '
'Trying to load the model with Tensorflow.')
model = model_class.from_pretrained(model, config=config, **model_kwargs)
return task(model=model, tokenizer=tokenizer, modelcard=modelcard, framework=framework, **kwargs)

23
transformers/tests/configuration_common_test.py
View File

@@ -16,15 +16,12 @@ from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import copy
import os
import shutil
import json
import random
import uuid
import tempfile
import unittest
import logging
from .tokenization_tests_commons import TemporaryDirectory
class ConfigTester(object):
@@ -48,16 +45,28 @@ class ConfigTester(object):
def create_and_test_config_to_json_file(self):
config_first = self.config_class(**self.inputs_dict)
json_file_path = os.path.join(os.getcwd(), "config_" + str(uuid.uuid4()) + ".json")
with TemporaryDirectory() as tmpdirname:
json_file_path = os.path.join(tmpdirname, "config.json")
config_first.to_json_file(json_file_path)
config_second = self.config_class.from_json_file(json_file_path)
os.remove(json_file_path)
self.parent.assertEqual(config_second.to_dict(), config_first.to_dict())
def create_and_test_config_from_and_save_pretrained(self):
config_first = self.config_class(**self.inputs_dict)
with TemporaryDirectory() as tmpdirname:
config_first.save_pretrained(tmpdirname)
config_second = self.config_class.from_pretrained(tmpdirname)
self.parent.assertEqual(config_second.to_dict(), config_first.to_dict())
def run_common_tests(self):
self.create_and_test_config_common_properties()
self.create_and_test_config_to_json_string()
self.create_and_test_config_to_json_file()
self.create_and_test_config_from_and_save_pretrained()
if __name__ == "__main__":
unittest.main()

0
transformers/tests/fixtures/empty.txt vendored Normal file
View File

26
transformers/tests/hf_api_test.py
View File

@@ -15,18 +15,30 @@
from __future__ import absolute_import, division, print_function
import os
import six
import time
import unittest
from transformers.hf_api import HfApi, S3Obj, PresignedUrl, HfFolder, HTTPError
import requests
import six
from transformers.hf_api import HfApi, HfFolder, HTTPError, PresignedUrl, S3Obj
USER = "__DUMMY_TRANSFORMERS_USER__"
PASS = "__DUMMY_TRANSFORMERS_PASS__"
FILE_KEY = "Test-{}.txt".format(int(time.time()))
FILE_PATH = os.path.join(
FILES = [
(
"Test-{}.txt".format(int(time.time())),
os.path.join(
os.path.dirname(os.path.abspath(__file__)), "fixtures/input.txt"
)
),
(
"yoyo {}.txt".format(int(time.time())), # space is intentional
os.path.join(
os.path.dirname(os.path.abspath(__file__)), "fixtures/empty.txt"
)
),
]
@@ -57,15 +69,21 @@ class HfApiEndpointsTest(HfApiCommonTest):
self.assertEqual(user, USER)
def test_presign(self):
for FILE_KEY, FILE_PATH in FILES:
urls = self._api.presign(token=self._token, filename=FILE_KEY)
self.assertIsInstance(urls, PresignedUrl)
self.assertEqual(urls.type, "text/plain")
def test_presign_and_upload(self):
for FILE_KEY, FILE_PATH in FILES:
access_url = self._api.presign_and_upload(
token=self._token, filename=FILE_KEY, filepath=FILE_PATH
)
self.assertIsInstance(access_url, six.string_types)
with open(FILE_PATH, 'r') as f:
body = f.read()
r = requests.get(access_url)
self.assertEqual(r.text, body)
def test_list_objs(self):
objs = self._api.list_objs(token=self._token)

89
transformers/tests/model_card_test.py Normal file
View File

@@ -0,0 +1,89 @@
# 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.
from __future__ import absolute_import, division, print_function, unicode_literals
import os
import json
import unittest
from transformers.modelcard import ModelCard
from .tokenization_tests_commons import TemporaryDirectory
class ModelCardTester(unittest.TestCase):
def setUp(self):
self.inputs_dict = {'model_details': {
'Organization': 'testing',
'Model date': 'today',
'Model version': 'v2.1, Developed by Test Corp in 2019.',
'Architecture': 'Convolutional Neural Network.',
},
'metrics': 'BLEU and ROUGE-1',
'evaluation_data':{
'Datasets':{
'BLEU': 'My-great-dataset-v1',
'ROUGE-1': 'My-short-dataset-v2.1',
},
'Preprocessing': 'See details on https://arxiv.org/pdf/1810.03993.pdf'
},
'training_data':{
'Dataset': 'English Wikipedia dump dated 2018-12-01',
'Preprocessing': 'Using SentencePiece vocabulary of size 52k tokens. See details on https://arxiv.org/pdf/1810.03993.pdf'
},
'quantitative_analyses': {
'BLEU': 55.1,
'ROUGE-1': 76,
},
}
def test_model_card_common_properties(self):
modelcard = ModelCard.from_dict(self.inputs_dict)
self.assertTrue(hasattr(modelcard, 'model_details'))
self.assertTrue(hasattr(modelcard, 'intended_use'))
self.assertTrue(hasattr(modelcard, 'factors'))
self.assertTrue(hasattr(modelcard, 'metrics'))
self.assertTrue(hasattr(modelcard, 'evaluation_data'))
self.assertTrue(hasattr(modelcard, 'training_data'))
self.assertTrue(hasattr(modelcard, 'quantitative_analyses'))
self.assertTrue(hasattr(modelcard, 'ethical_considerations'))
self.assertTrue(hasattr(modelcard, 'caveats_and_recommendations'))
def test_model_card_to_json_string(self):
modelcard = ModelCard.from_dict(self.inputs_dict)
obj = json.loads(modelcard.to_json_string())
for key, value in self.inputs_dict.items():
self.assertEqual(obj[key], value)
def test_model_card_to_json_file(self):
model_card_first = ModelCard.from_dict(self.inputs_dict)
with TemporaryDirectory() as tmpdirname:
filename = os.path.join(tmpdirname, u"modelcard.json")
model_card_first.to_json_file(filename)
model_card_second = ModelCard.from_json_file(filename)
self.assertEqual(model_card_second.to_dict(), model_card_first.to_dict())
def test_model_card_from_and_save_pretrained(self):
model_card_first = ModelCard.from_dict(self.inputs_dict)
with TemporaryDirectory() as tmpdirname:
model_card_first.save_pretrained(tmpdirname)
model_card_second = ModelCard.from_pretrained(tmpdirname)
self.assertEqual(model_card_second.to_dict(), model_card_first.to_dict())
if __name__ == "__main__":
unittest.main()

9
transformers/tests/modeling_albert_test.py
View File

@@ -17,13 +17,12 @@ from __future__ import division
from __future__ import print_function
import unittest
import shutil
from transformers import is_torch_available
from .modeling_common_test import (CommonTestCases, ids_tensor)
from .configuration_common_test import ConfigTester
from .utils import require_torch, slow, torch_device
from .utils import CACHE_DIR, require_torch, slow, torch_device
if is_torch_available():
from transformers import (AlbertConfig, AlbertModel, AlbertForMaskedLM,
@@ -110,7 +109,7 @@ class AlbertModelTest(CommonTestCases.CommonModelTester):
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = AlbertConfig(
vocab_size_or_config_json_file=self.vocab_size,
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
@@ -230,10 +229,8 @@ class AlbertModelTest(CommonTestCases.CommonModelTester):
@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)
model = AlbertModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model)
if __name__ == "__main__":

7
transformers/tests/modeling_auto_test.py
View File

@@ -22,7 +22,7 @@ import logging
from transformers import is_torch_available
from .utils import require_torch, slow
from .utils import require_torch, slow, SMALL_MODEL_IDENTIFIER
if is_torch_available():
from transformers import (AutoConfig, BertConfig,
@@ -92,6 +92,11 @@ class AutoModelTest(unittest.TestCase):
self.assertIsNotNone(model)
self.assertIsInstance(model, BertForQuestionAnswering)
def test_from_pretrained_identifier(self):
logging.basicConfig(level=logging.INFO)
model = AutoModelWithLMHead.from_pretrained(SMALL_MODEL_IDENTIFIER)
self.assertIsInstance(model, BertForMaskedLM)
if __name__ == "__main__":
unittest.main()

9
transformers/tests/modeling_bert_test.py
View File

@@ -17,13 +17,12 @@ from __future__ import division
from __future__ import print_function
import unittest
import shutil
from transformers import is_torch_available
from .modeling_common_test import (CommonTestCases, ids_tensor, floats_tensor)
from .configuration_common_test import ConfigTester
from .utils import require_torch, slow, torch_device
from .utils import CACHE_DIR, require_torch, slow, torch_device
if is_torch_available():
from transformers import (BertConfig, BertModel, BertForMaskedLM,
@@ -109,7 +108,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = BertConfig(
vocab_size_or_config_json_file=self.vocab_size,
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
@@ -360,10 +359,8 @@ class BertModelTest(CommonTestCases.CommonModelTester):
@slow
def test_model_from_pretrained(self):
cache_dir = "/tmp/transformers_test/"
for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
model = BertModel.from_pretrained(model_name, cache_dir=cache_dir)
shutil.rmtree(cache_dir)
model = BertModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model)

132
transformers/tests/modeling_common_test.py
View File

@@ -18,7 +18,7 @@ from __future__ import print_function
import copy
import sys
import os
import os.path
import shutil
import tempfile
import json
@@ -30,7 +30,7 @@ import logging
from transformers import is_torch_available
from .utils import require_torch, slow, torch_device
from .utils import CACHE_DIR, require_torch, slow, torch_device
if is_torch_available():
import torch
@@ -58,7 +58,7 @@ else:
def _config_zero_init(config):
configs_no_init = copy.deepcopy(config)
for key in configs_no_init.__dict__.keys():
if '_range' in key or '_std' in key:
if '_range' in key or '_std' in key or 'initializer_factor' in key:
setattr(configs_no_init, key, 0.0)
return configs_no_init
@@ -73,6 +73,7 @@ class CommonTestCases:
test_pruning = True
test_resize_embeddings = True
test_head_masking = True
is_encoder_decoder = False
def test_save_load(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -83,6 +84,8 @@ class CommonTestCases:
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
out_2 = outputs[0].numpy()
out_2[np.isnan(out_2)] = 0
with TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
@@ -93,9 +96,7 @@ class CommonTestCases:
# Make sure we don't have nans
out_1 = after_outputs[0].cpu().numpy()
out_2 = outputs[0].cpu().numpy()
out_1 = out_1[~np.isnan(out_1)]
out_2 = out_2[~np.isnan(out_2)]
out_1[np.isnan(out_1)] = 0
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
@@ -117,19 +118,31 @@ class CommonTestCases:
model = model_class(config)
model.to(torch_device)
model.eval()
first, second = model(inputs_dict["input_ids"])[0], model(inputs_dict["input_ids"])[0]
self.assertEqual(first.ne(second).sum().item(), 0)
with torch.no_grad():
first = model(**inputs_dict)[0]
second = model(**inputs_dict)[0]
out_1 = first.cpu().numpy()
out_2 = second.cpu().numpy()
out_1 = out_1[~np.isnan(out_1)]
out_2 = out_2[~np.isnan(out_2)]
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
decoder_seq_length = self.model_tester.decoder_seq_length if hasattr(self.model_tester, 'decoder_seq_length') else self.model_tester.seq_length
encoder_seq_length = self.model_tester.encoder_seq_length if hasattr(self.model_tester, 'encoder_seq_length') else self.model_tester.seq_length
decoder_key_length = self.model_tester.key_length if hasattr(self.model_tester, 'key_length') else decoder_seq_length
encoder_key_length = self.model_tester.key_length if hasattr(self.model_tester, 'key_length') else encoder_seq_length
for model_class in self.all_model_classes:
config.output_attentions = True
config.output_hidden_states = False
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
self.assertEqual(model.config.output_attentions, True)
@@ -138,28 +151,42 @@ class CommonTestCases:
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads,
self.model_tester.seq_length,
self.model_tester.key_len if hasattr(self.model_tester, 'key_len') else self.model_tester.seq_length])
encoder_seq_length ,
encoder_key_length])
out_len = len(outputs)
if self.is_encoder_decoder:
self.assertEqual(out_len % 2, 0)
decoder_attentions = outputs[(out_len // 2)-1]
self.assertEqual(model.config.output_attentions, True)
self.assertEqual(model.config.output_hidden_states, False)
self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(decoder_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads,
decoder_seq_length,
decoder_key_length
])
# Check attention is always last and order is fine
config.output_attentions = True
config.output_hidden_states = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
self.assertEqual(out_len+1, len(outputs))
self.assertEqual(out_len + (2 if self.is_encoder_decoder else 1), len(outputs))
self.assertEqual(model.config.output_attentions, True)
self.assertEqual(model.config.output_hidden_states, True)
attentions = outputs[-1]
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self_attentions = outputs[-1]
self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads,
self.model_tester.seq_length,
self.model_tester.key_len if hasattr(self.model_tester, 'key_len') else self.model_tester.seq_length])
encoder_seq_length,
encoder_key_length])
def test_torchscript(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -191,20 +218,21 @@ class CommonTestCases:
inputs = inputs_dict['input_ids'] # Let's keep only input_ids
try:
torch.jit.trace(model, inputs)
traced_gpt2 = torch.jit.trace(model, inputs)
except RuntimeError:
self.fail("Couldn't trace module.")
with TemporaryDirectory() as tmp_dir_name:
pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
try:
traced_gpt2 = torch.jit.trace(model, inputs)
torch.jit.save(traced_gpt2, "traced_model.pt")
except RuntimeError:
torch.jit.save(traced_gpt2, pt_file_name)
except Exception:
self.fail("Couldn't save module.")
try:
loaded_model = torch.jit.load("traced_model.pt")
os.remove("traced_model.pt")
except ValueError:
loaded_model = torch.jit.load(pt_file_name)
except Exception:
self.fail("Couldn't load module.")
model.to(torch_device)
@@ -223,7 +251,6 @@ class CommonTestCases:
self.assertTrue(models_equal)
def test_headmasking(self):
if not self.test_head_masking:
return
@@ -278,7 +305,6 @@ class CommonTestCases:
self.assertNotEqual(
attentions[-1][..., -1, :, :].flatten().sum().item(), 0.0)
def test_head_pruning(self):
if not self.test_pruning:
return
@@ -297,6 +323,7 @@ class CommonTestCases:
heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)),
-1: [0]}
model.prune_heads(heads_to_prune)
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
@@ -326,20 +353,19 @@ class CommonTestCases:
heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)),
-1: [0]}
model.prune_heads(heads_to_prune)
directory = "pruned_model"
if not os.path.exists(directory):
os.makedirs(directory)
model.save_pretrained(directory)
model = model_class.from_pretrained(directory)
with TemporaryDirectory() as temp_dir_name:
model.save_pretrained(temp_dir_name)
model = model_class.from_pretrained(temp_dir_name)
model.to(torch_device)
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
self.assertEqual(attentions[0].shape[-3], 1)
self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
shutil.rmtree(directory)
def test_head_pruning_save_load_from_config_init(self):
if not self.test_pruning:
@@ -362,6 +388,7 @@ class CommonTestCases:
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
@@ -389,6 +416,7 @@ class CommonTestCases:
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
@@ -397,15 +425,12 @@ class CommonTestCases:
self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
directory = "pruned_model"
if not os.path.exists(directory):
os.makedirs(directory)
model.save_pretrained(directory)
model = model_class.from_pretrained(directory)
with TemporaryDirectory() as temp_dir_name:
model.save_pretrained(temp_dir_name)
model = model_class.from_pretrained(temp_dir_name)
model.to(torch_device)
shutil.rmtree(directory)
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
@@ -417,6 +442,7 @@ class CommonTestCases:
heads_to_prune = {0: [0], 2: [1, 2]}
model.prune_heads(heads_to_prune)
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
@@ -427,7 +453,6 @@ class CommonTestCases:
self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2], 2: [1, 2]})
def test_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -437,6 +462,7 @@ class CommonTestCases:
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
hidden_states = outputs[-1]
self.assertEqual(model.config.output_attentions, False)
@@ -444,7 +470,8 @@ class CommonTestCases:
self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[self.model_tester.seq_length, self.model_tester.hidden_size])
[self.model_tester.encoder_seq_length if hasattr(self.model_tester, 'encoder_seq_length') else self.model_tester.seq_length,
self.model_tester.hidden_size])
def test_resize_tokens_embeddings(self):
original_config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -550,8 +577,14 @@ class CommonTestCases:
def test_inputs_embeds(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if not self.is_encoder_decoder:
input_ids = inputs_dict["input_ids"]
del inputs_dict["input_ids"]
else:
encoder_input_ids = inputs_dict["encoder_input_ids"]
decoder_input_ids = inputs_dict["decoder_input_ids"]
del inputs_dict["encoder_input_ids"]
del inputs_dict["decoder_input_ids"]
for model_class in self.all_model_classes:
model = model_class(config)
@@ -559,9 +592,14 @@ class CommonTestCases:
model.eval()
wte = model.get_input_embeddings()
if not self.is_encoder_decoder:
inputs_dict["inputs_embeds"] = wte(input_ids)
outputs = model(**inputs_dict)
else:
inputs_dict["encoder_inputs_embeds"] = wte(encoder_input_ids)
inputs_dict["decoder_inputs_embeds"] = wte(decoder_input_ids)
with torch.no_grad():
outputs = model(**inputs_dict)
class GPTModelTester(CommonModelTester):
@@ -633,7 +671,7 @@ class CommonTestCases:
mc_token_ids = ids_tensor([self.batch_size, self.n_choices], self.seq_length)
config = self.config_class(
vocab_size_or_config_json_file=self.vocab_size,
vocab_size=self.vocab_size,
n_positions=self.n_positions,
n_embd=self.hidden_size,
n_layer=self.num_hidden_layers,
@@ -649,6 +687,7 @@ class CommonTestCases:
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(input_ids, position_ids, token_type_ids)
outputs = model(input_ids, position_ids)
outputs = model(input_ids)
@@ -664,6 +703,7 @@ class CommonTestCases:
model = self.lm_head_model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(input_ids, position_ids, token_type_ids, lm_labels)
loss, lm_logits = outputs[:2]
@@ -681,6 +721,7 @@ class CommonTestCases:
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(input_ids)
presents = outputs[-1]
self.parent.assertEqual(self.num_hidden_layers, len(presents))
@@ -694,6 +735,7 @@ class CommonTestCases:
model = self.double_head_model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(input_ids, mc_token_ids, lm_labels=lm_labels, mc_labels=mc_labels,
token_type_ids=token_type_ids, position_ids=position_ids)
lm_loss, mc_loss, lm_logits, mc_logits = outputs[:4]
@@ -711,10 +753,8 @@ class CommonTestCases:
[[], []])
def create_and_check_model_from_pretrained(self):
cache_dir = "/tmp/transformers_test/"
for model_name in list(self.base_model_class.pretrained_model_archive_map.keys())[:1]:
model = self.base_model_class.from_pretrained(model_name, cache_dir=cache_dir)
shutil.rmtree(cache_dir)
model = self.base_model_class.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.parent.assertIsNotNone(model)
def prepare_config_and_inputs_for_common(self):

9
transformers/tests/modeling_ctrl_test.py
View File

@@ -16,7 +16,6 @@ from __future__ import division
from __future__ import print_function
import unittest
import shutil
import pdb
from transformers import is_torch_available
@@ -27,7 +26,7 @@ if is_torch_available():
from .modeling_common_test import (CommonTestCases, ids_tensor)
from .configuration_common_test import ConfigTester
from .utils import require_torch, slow, torch_device
from .utils import CACHE_DIR, require_torch, slow, torch_device
@require_torch
@@ -114,7 +113,7 @@ class CTRLModelTest(CommonTestCases.CommonModelTester):
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = CTRLConfig(
vocab_size_or_config_json_file=self.vocab_size,
vocab_size=self.vocab_size,
n_embd=self.hidden_size,
n_layer=self.num_hidden_layers,
n_head=self.num_attention_heads,
@@ -205,10 +204,8 @@ class CTRLModelTest(CommonTestCases.CommonModelTester):
@slow
def test_model_from_pretrained(self):
cache_dir = "/tmp/transformers_test/"
for model_name in list(CTRL_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
model = CTRLModel.from_pretrained(model_name, cache_dir=cache_dir)
shutil.rmtree(cache_dir)
model = CTRLModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model)

8
transformers/tests/modeling_distilbert_test.py
View File

@@ -27,7 +27,7 @@ if is_torch_available():
from .modeling_common_test import (CommonTestCases, ids_tensor)
from .configuration_common_test import ConfigTester
from .utils import require_torch, slow, torch_device
from .utils import CACHE_DIR, require_torch, slow, torch_device
@require_torch
@@ -105,7 +105,7 @@ class DistilBertModelTest(CommonTestCases.CommonModelTester):
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = DistilBertConfig(
vocab_size_or_config_json_file=self.vocab_size,
vocab_size=self.vocab_size,
dim=self.hidden_size,
n_layers=self.num_hidden_layers,
n_heads=self.num_attention_heads,
@@ -235,10 +235,8 @@ class DistilBertModelTest(CommonTestCases.CommonModelTester):
# @slow
# def test_model_from_pretrained(self):
# cache_dir = "/tmp/transformers_test/"
# for model_name in list(DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
# model = DistilBertModel.from_pretrained(model_name, cache_dir=cache_dir)
# shutil.rmtree(cache_dir)
# model = DistilBertModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
# self.assertIsNotNone(model)
if __name__ == "__main__":

9
transformers/tests/modeling_gpt2_test.py
View File

@@ -17,7 +17,6 @@ from __future__ import division
from __future__ import print_function
import unittest
import shutil
from transformers import is_torch_available
@@ -27,7 +26,7 @@ if is_torch_available():
from .modeling_common_test import (CommonTestCases, ids_tensor)
from .configuration_common_test import ConfigTester
from .utils import require_torch, slow, torch_device
from .utils import CACHE_DIR, require_torch, slow, torch_device
@require_torch
@@ -110,7 +109,7 @@ class GPT2ModelTest(CommonTestCases.CommonModelTester):
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = GPT2Config(
vocab_size_or_config_json_file=self.vocab_size,
vocab_size=self.vocab_size,
n_embd=self.hidden_size,
n_layer=self.num_hidden_layers,
n_head=self.num_attention_heads,
@@ -239,10 +238,8 @@ class GPT2ModelTest(CommonTestCases.CommonModelTester):
@slow
def test_model_from_pretrained(self):
cache_dir = "/tmp/transformers_test/"
for model_name in list(GPT2_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
model = GPT2Model.from_pretrained(model_name, cache_dir=cache_dir)
shutil.rmtree(cache_dir)
model = GPT2Model.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model)

9
transformers/tests/modeling_openai_test.py
View File

@@ -17,7 +17,6 @@ from __future__ import division
from __future__ import print_function
import unittest
import shutil
from transformers import is_torch_available
@@ -27,7 +26,7 @@ if is_torch_available():
from .modeling_common_test import (CommonTestCases, ids_tensor)
from .configuration_common_test import ConfigTester
from .utils import require_torch, slow, torch_device
from .utils import CACHE_DIR, require_torch, slow, torch_device
@require_torch
@@ -98,7 +97,7 @@ class OpenAIGPTModelTest(CommonTestCases.CommonModelTester):
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = OpenAIGPTConfig(
vocab_size_or_config_json_file=self.vocab_size,
vocab_size=self.vocab_size,
n_embd=self.hidden_size,
n_layer=self.num_hidden_layers,
n_head=self.num_attention_heads,
@@ -207,10 +206,8 @@ class OpenAIGPTModelTest(CommonTestCases.CommonModelTester):
@slow
def test_model_from_pretrained(self):
cache_dir = "/tmp/transformers_test/"
for model_name in list(OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
model = OpenAIGPTModel.from_pretrained(model_name, cache_dir=cache_dir)
shutil.rmtree(cache_dir)
model = OpenAIGPTModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model)

61
transformers/tests/modeling_roberta_test.py
View File

@@ -17,7 +17,6 @@ from __future__ import division
from __future__ import print_function
import unittest
import shutil
from transformers import is_torch_available
@@ -25,11 +24,12 @@ if is_torch_available():
import torch
from transformers import (RobertaConfig, RobertaModel, RobertaForMaskedLM,
RobertaForSequenceClassification, RobertaForTokenClassification)
from transformers.modeling_roberta import RobertaEmbeddings
from transformers.modeling_roberta import ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
from .modeling_common_test import (CommonTestCases, ids_tensor)
from .configuration_common_test import ConfigTester
from .utils import require_torch, slow, torch_device
from .utils import CACHE_DIR, require_torch, slow, torch_device
@require_torch
@@ -106,7 +106,7 @@ class RobertaModelTest(CommonTestCases.CommonModelTester):
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = RobertaConfig(
vocab_size_or_config_json_file=self.vocab_size,
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
@@ -199,12 +199,61 @@ class RobertaModelTest(CommonTestCases.CommonModelTester):
@slow
def test_model_from_pretrained(self):
cache_dir = "/tmp/transformers_test/"
for model_name in list(ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
model = RobertaModel.from_pretrained(model_name, cache_dir=cache_dir)
shutil.rmtree(cache_dir)
model = RobertaModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model)
def test_create_position_ids_respects_padding_index(self):
""" Ensure that the default position ids only assign a sequential . This is a regression
test for https://github.com/huggingface/transformers/issues/1761
The position ids should be masked with the embedding object's padding index. Therefore, the
first available non-padding position index is RobertaEmbeddings.padding_idx + 1
"""
config = self.model_tester.prepare_config_and_inputs()[0]
model = RobertaEmbeddings(config=config)
input_ids = torch.as_tensor([[12, 31, 13, model.padding_idx]])
expected_positions = torch.as_tensor([[
0 + model.padding_idx + 1,
1 + model.padding_idx + 1,
2 + model.padding_idx + 1,
model.padding_idx
]])
position_ids = model.create_position_ids_from_input_ids(input_ids)
self.assertEqual(
position_ids.shape,
expected_positions.shape
)
self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
def test_create_position_ids_from_inputs_embeds(self):
""" Ensure that the default position ids only assign a sequential . This is a regression
test for https://github.com/huggingface/transformers/issues/1761
The position ids should be masked with the embedding object's padding index. Therefore, the
first available non-padding position index is RobertaEmbeddings.padding_idx + 1
"""
config = self.model_tester.prepare_config_and_inputs()[0]
embeddings = RobertaEmbeddings(config=config)
inputs_embeds = torch.Tensor(2, 4, 30)
expected_single_positions = [
0 + embeddings.padding_idx + 1,
1 + embeddings.padding_idx + 1,
2 + embeddings.padding_idx + 1,
3 + embeddings.padding_idx + 1,
]
expected_positions = torch.as_tensor([expected_single_positions, expected_single_positions])
position_ids = embeddings.create_position_ids_from_inputs_embeds(inputs_embeds)
self.assertEqual(
position_ids.shape,
expected_positions.shape
)
self.assertTrue(
torch.all(torch.eq(position_ids, expected_positions))
)
class RobertaModelIntegrationTest(unittest.TestCase):

182
transformers/tests/modeling_t5_test.py Normal file
View File

@@ -0,0 +1,182 @@
# coding=utf-8
# Copyright 2018 Google T5 Authors and 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.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import unittest
from transformers import is_torch_available
from .modeling_common_test import (CommonTestCases, ids_tensor, floats_tensor)
from .configuration_common_test import ConfigTester
from .utils import CACHE_DIR, require_torch, slow, torch_device
if is_torch_available():
from transformers import (T5Config, T5Model, T5WithLMHeadModel)
from transformers.modeling_t5 import T5_PRETRAINED_MODEL_ARCHIVE_MAP
@require_torch
class T5ModelTest(CommonTestCases.CommonModelTester):
all_model_classes = (T5Model, T5WithLMHeadModel) if is_torch_available() else ()
test_pruning = False
test_torchscript = False
test_resize_embeddings = False
is_encoder_decoder = True
class T5ModelTester(object):
def __init__(self,
parent,
batch_size=13,
encoder_seq_length=7,
decoder_seq_length=9,
is_training=True,
use_attention_mask=True,
use_labels=True,
vocab_size=99,
n_positions=14,
hidden_size=32,
num_hidden_layers=5,
num_attention_heads=4,
d_ff=37,
relative_attention_num_buckets=8,
dropout_rate=0.1,
initializer_factor=0.002,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.encoder_seq_length = encoder_seq_length
self.decoder_seq_length = decoder_seq_length
self.is_training = is_training
self.use_attention_mask = use_attention_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.n_positions = n_positions
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.d_ff = d_ff
self.relative_attention_num_buckets = relative_attention_num_buckets
self.dropout_rate = dropout_rate
self.initializer_factor = initializer_factor
self.scope = scope
def prepare_config_and_inputs(self):
encoder_input_ids = ids_tensor([self.batch_size, self.encoder_seq_length], self.vocab_size)
decoder_input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
encoder_attention_mask = None
decoder_attention_mask = None
if self.use_attention_mask:
encoder_attention_mask = ids_tensor([self.batch_size, self.encoder_seq_length], vocab_size=2)
decoder_attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
decoder_lm_labels = None
if self.use_labels:
decoder_lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
config = T5Config(
vocab_size=self.vocab_size,
n_positions=self.n_positions,
d_model=self.hidden_size,
d_ff=self.d_ff,
d_kv=self.hidden_size // self.num_attention_heads,
num_layers=self.num_hidden_layers,
num_heads=self.num_attention_heads,
relative_attention_num_buckets=self.relative_attention_num_buckets,
dropout_rate=self.dropout_rate,
initializer_factor=self.initializer_factor)
return (config, encoder_input_ids, decoder_input_ids, encoder_attention_mask, decoder_attention_mask, decoder_lm_labels)
def check_loss_output(self, result):
self.parent.assertListEqual(
list(result["loss"].size()),
[])
def create_and_check_t5_model(self, config, encoder_input_ids, decoder_input_ids, encoder_attention_mask, decoder_attention_mask, decoder_lm_labels):
model = T5Model(config=config)
model.eval()
decoder_output, encoder_output = model(encoder_input_ids=encoder_input_ids,
decoder_input_ids=decoder_input_ids,
encoder_attention_mask=encoder_attention_mask,
decoder_attention_mask=decoder_attention_mask)
decoder_output, encoder_output = model(encoder_input_ids=encoder_input_ids,
decoder_input_ids=decoder_input_ids)
result = {
"encoder_output": encoder_output,
"decoder_output": decoder_output,
}
self.parent.assertListEqual(
list(result["encoder_output"].size()),
[self.batch_size, self.encoder_seq_length, self.hidden_size])
self.parent.assertListEqual(
list(result["decoder_output"].size()),
[self.batch_size, self.decoder_seq_length, self.hidden_size])
def create_and_check_t5_with_lm_head(self, config, encoder_input_ids, decoder_input_ids, encoder_attention_mask, decoder_attention_mask, decoder_lm_labels):
model = T5WithLMHeadModel(config=config)
model.eval()
outputs = model(encoder_input_ids=encoder_input_ids, decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask, decoder_lm_labels=decoder_lm_labels)
loss, prediction_scores = outputs[0], outputs[1]
result = {
"loss": loss,
"prediction_scores": prediction_scores,
}
self.parent.assertListEqual(
list(result["prediction_scores"].size()),
[self.batch_size, self.decoder_seq_length, self.vocab_size])
self.check_loss_output(result)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(config, encoder_input_ids, decoder_input_ids, encoder_attention_mask,
decoder_attention_mask, decoder_lm_labels) = config_and_inputs
inputs_dict = {'encoder_input_ids': encoder_input_ids,
'decoder_input_ids': decoder_input_ids,
'decoder_attention_mask': decoder_attention_mask,
'encoder_attention_mask': encoder_attention_mask}
return config, inputs_dict
def setUp(self):
self.model_tester = T5ModelTest.T5ModelTester(self)
self.config_tester = ConfigTester(self, config_class=T5Config, d_model=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_t5_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_t5_model(*config_and_inputs)
def test_with_lm_head(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_t5_with_lm_head(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
for model_name in list(T5_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
model = T5Model.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model)
if __name__ == "__main__":
unittest.main()

Some files were not shown because too many files have changed in this diff Show More