unpin isort for pypi

* split beam search and no beam search test

* fix test

* clean generate tests

.circleci/config.yml

@@ -1,43 +1,106 @@
 version: 2
 jobs:
-    build_py3:
-        working_directory: ~/pytorch-transformers
+    run_tests_torch_and_tf:
+        working_directory: ~/transformers
         docker:
-            - image: circleci/python:3.5
+            - image: circleci/python:3.6
+        environment:
+            OMP_NUM_THREADS: 1
         resource_class: xlarge
         parallelism: 1
         steps:
             - checkout
-            - run: sudo pip install --progress-bar off .
-            - run: sudo pip install pytest codecov pytest-cov
-            - run: sudo pip install tensorboardX scikit-learn
-            - run: python -m pytest -sv ./pytorch_transformers/tests/ --cov
-            - run: python -m pytest -sv ./examples/
+            - run: sudo pip install .[sklearn,tf-cpu,torch,testing]
+            - run: sudo pip install codecov pytest-cov
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
             - run: codecov
-    build_py2:
-        working_directory: ~/pytorch-transformers
-        resource_class: large
-        parallelism: 1
+
+    run_tests_torch:
+        working_directory: ~/transformers
         docker:
-            - image: circleci/python:2.7
+            - image: circleci/python:3.7
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
         steps:
             - checkout
-            - run: sudo pip install --progress-bar off .
-            - run: sudo pip install pytest codecov pytest-cov
-            - run: python -m pytest -sv ./pytorch_transformers/tests/ --cov
+            - run: sudo pip install .[sklearn,torch,testing]
+            - run: sudo pip install codecov pytest-cov
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
             - run: codecov
-    deploy_doc:
-        working_directory: ~/pytorch-transformers
+    run_tests_tf:
+        working_directory: ~/transformers
         docker:
-            - image: circleci/python:3.5
+            - image: circleci/python:3.7
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - run: sudo pip install .[sklearn,tf-cpu,testing]
+            - run: sudo pip install codecov pytest-cov
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
+            - run: codecov
+    run_tests_custom_tokenizers:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.6
+        environment:
+            RUN_CUSTOM_TOKENIZERS: yes
+        steps:
+            - checkout
+            - run: sudo pip install .[mecab,testing]
+            - run: python -m pytest -sv ./tests/test_tokenization_bert_japanese.py
+    run_examples_torch:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.6
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - run: sudo pip install .[sklearn,torch,testing]
+            - run: sudo pip install -r examples/requirements.txt
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./examples/
+    deploy_doc:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.6
         steps:
             - add_ssh_keys:
-                  fingerprints:
-                      - "5b:7a:95:18:07:8c:aa:76:4c:60:35:88:ad:60:56:71"
+                fingerprints:
+                    - "5b:7a:95:18:07:8c:aa:76:4c:60:35:88:ad:60:56:71"
             - checkout
-            - run: sudo pip install --progress-bar off -r docs/requirements.txt
-            - run: sudo pip install --progress-bar off -r requirements.txt
-            - run: cd docs && make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+            - run: sudo pip install .[tf,torch,docs]
+            - run: ./.circleci/deploy.sh
+    check_code_quality:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.6
+        resource_class: medium
+        parallelism: 1
+        steps:
+            - checkout
+            # we need a version of isort with https://github.com/timothycrosley/isort/pull/1000
+            - run: sudo pip install git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort
+            - run: sudo pip install .[tf,torch,quality]
+            - run: black --check --line-length 119 --target-version py35 examples templates tests src utils
+            - run: isort --check-only --recursive examples templates tests src utils
+            - run: flake8 examples templates tests src utils
+    check_repository_consistency:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.6
+        resource_class: small
+        parallelism: 1
+        steps:
+            - checkout
+            - run: sudo pip install requests
+            - run: python ./utils/link_tester.py
 workflow_filters: &workflow_filters
     filters:
         branches:
@@ -47,6 +110,11 @@ workflows:
     version: 2
     build_and_test:
         jobs:
-            - build_py3
-            - build_py2
-            - deploy_doc: *workflow_filters
+            - check_code_quality
+            - check_repository_consistency
+            - run_examples_torch
+            - run_tests_custom_tokenizers
+            - run_tests_torch_and_tf
+            - run_tests_torch
+            - run_tests_tf
+            - deploy_doc: *workflow_filters

.circleci/deploy.sh

@@ -0,0 +1,29 @@
+cd docs
+
+function deploy_doc(){
+	echo "Creating doc at commit $1 and pushing to folder $2"
+	git checkout $1
+	if [ ! -z "$2" ]
+	then
+		if [ -d "$dir/$2" ]; then
+			echo "Directory" $2 "already exists"
+		else
+			echo "Pushing version" $2
+			make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
+		fi
+	else
+		echo "Pushing master"
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+	fi
+}
+
+deploy_doc "master"
+deploy_doc "b33a385" v1.0.0
+deploy_doc "fe02e45" v1.1.0
+deploy_doc "89fd345" v1.2.0
+deploy_doc "fc9faa8" v2.0.0
+deploy_doc "3ddce1d" v2.1.1
+deploy_doc "3616209" v2.2.0
+deploy_doc "d0f8b9a" v2.3.0
+deploy_doc "6664ea9" v2.4.0
+deploy_doc "fb560dc" v2.5.0

.coveragerc

@@ -1,5 +1,5 @@
 [run]
-source=pytorch_transformers
+source=transformers
 omit =
     # skip convertion scripts from testing for now
     */convert_*

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

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

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

@@ -0,0 +1,20 @@
+---
+name: "\U0001F31F New model addition"
+about: Submit a proposal/request to implement a new Transformer-based model
+title: ''
+labels: ''
+assignees: ''
+
+---
+
+# 🌟 New model addition
+
+## Model description
+
+<!-- Important information -->
+
+## Open source status
+
+* [ ] the model implementation is available: (give details)
+* [ ] the model weights are available: (give details)
+* [ ] who are the authors: (mention them, if possible by @gh-username)

.github/ISSUE_TEMPLATE/bug-report.md

@@ -1,25 +1,29 @@
 ---
 name: "\U0001F41B Bug Report"
-about: Submit a bug report to help us improve PyTorch Transformers
+about: Submit a bug report to help us improve transformers
+title: ''
+labels: ''
+assignees: ''
+
 ---
 
-## 🐛 Bug
+# 🐛 Bug
 
-<!-- Important information -->
+## Information
 
-Model I am using (Bert, XLNet....):
+Model I am using (Bert, XLNet ...):
 
-Language I am using the model on (English, Chinese....):
+Language I am using the model on (English, Chinese ...):
 
-The problem arise when using:
-* [ ] the official example scripts: (give details)
-* [ ] my own modified scripts: (give details)
+The problem arises when using:
+* [ ] the official example scripts: (give details below)
+* [ ] my own modified scripts: (give details below)
 
 The tasks I am working on is:
 * [ ] an official GLUE/SQUaD task: (give the name)
-* [ ] my own task or dataset: (give details)
+* [ ] my own task or dataset: (give details below)
 
-## To Reproduce
+## To reproduce
 
 Steps to reproduce the behavior:
 
@@ -27,22 +31,22 @@ Steps to reproduce the behavior:
 2.
 3.
 
-<!-- If you have a code sample, error messages, stack traces, please provide it here as well. -->
+<!-- If you have code snippets, error messages, stack traces please provide them here as well.
+     Important! Use code tags to correctly format your code. See https://help.github.com/en/github/writing-on-github/creating-and-highlighting-code-blocks#syntax-highlighting
+     Do not use screenshots, as they are hard to read and (more importantly) don't allow others to copy-and-paste your code.-->
 
 ## Expected behavior
 
-<!-- A clear and concise description of what you expected to happen. -->
+<!-- A clear and concise description of what you would expect to happen. -->
 
-## Environment
-
-* OS:
-* Python version:
-* PyTorch version:
-* PyTorch Transformers version (or branch):
-* Using GPU ?
-* Distributed of parallel setup ?
-* Any other relevant information:
-
-## Additional context
-
-<!-- Add any other context about the problem here. -->
+## Environment info
+<!-- You can run the command `python transformers-cli env` and copy-and-paste its output below.
+     Don't forget to fill out the missing fields in that output! -->
+     
+- `transformers` version:
+- Platform:
+- Python version:
+- PyTorch version (GPU?):
+- Tensorflow version (GPU?):
+- Using GPU in script?:
+- Using distributed or parallel set-up in script?:

.github/ISSUE_TEMPLATE/feature-request.md

@@ -1,16 +1,25 @@
 ---
-name: "\U0001F680 Feature Request"
-about: Submit a proposal/request for a new PyTorch Transformers feature
+name: "\U0001F680 Feature request"
+about: Submit a proposal/request for a new transformers feature
+title: ''
+labels: ''
+assignees: ''
+
 ---
 
-## 🚀 Feature
+# 🚀 Feature request
 
-<!-- A clear and concise description of the feature proposal. Please provide a link to the paper and code in case they exist. -->
+<!-- A clear and concise description of the feature proposal.
+     Please provide a link to the paper and code in case they exist. -->
 
 ## Motivation
 
-<!-- Please outline the motivation for the proposal. Is your feature request related to a problem? e.g., I'm always frustrated when [...]. If this is related to another GitHub issue, please link here too. -->
+<!-- Please outline the motivation for the proposal. Is your feature request
+     related to a problem? e.g., I'm always frustrated when [...]. If this is related
+     to another GitHub issue, please link here too. -->
 
-## Additional context
+## Your contribution
 
-<!-- Add any other context or screenshots about the feature request here. -->
+<!-- Is there any way that you could help, e.g. by submitting a PR?
+     Make sure to read the CONTRIBUTING.MD readme:
+     https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md -->

.github/ISSUE_TEMPLATE/migration.md

@@ -1,43 +1,57 @@
 ---
-name: "\U0001F4DA Migration from PyTorch-pretrained-Bert"
-about: Report a problem when migrating from PyTorch-pretrained-Bert to PyTorch-Transformers
+name: "\U0001F4DA Migration from pytorch-pretrained-bert or pytorch-transformers"
+about: Report a problem when migrating from pytorch-pretrained-bert or pytorch-transformers to transformers
+title: ''
+labels: ''
+assignees: ''
+
 ---
 
-## 📚 Migration
+# 📚 Migration
+
+## Information
 
 <!-- Important information -->
 
-Model I am using (Bert, XLNet....):
+Model I am using (Bert, XLNet ...):
 
-Language I am using the model on (English, Chinese....):
+Language I am using the model on (English, Chinese ...):
 
-The problem arise when using:
-* [ ] the official example scripts: (give details)
-* [ ] my own modified scripts: (give details)
+The problem arises when using:
+* [ ] the official example scripts: (give details below)
+* [ ] my own modified scripts: (give details below)
 
 The tasks I am working on is:
 * [ ] an official GLUE/SQUaD task: (give the name)
-* [ ] my own task or dataset: (give details)
+* [ ] my own task or dataset: (give details below)
 
-Details of the issue:
+## Details
 
-<!-- A clear and concise description of the migration issue. If you have code snippets, please provide it here as well. -->
+<!-- A clear and concise description of the migration issue.
+    If you have code snippets, please provide it here as well.
+    Important! Use code tags to correctly format your code. See https://help.github.com/en/github/writing-on-github/creating-and-highlighting-code-blocks#syntax-highlighting
+    Do not use screenshots, as they are hard to read and (more importantly) don't allow others to copy-and-paste your code.
+    -->
 
-## Environment
+## Environment info
+<!-- You can run the command `python transformers-cli env` and copy-and-paste its output below.
+     Don't forget to fill out the missing fields in that output! -->
+ 
+- `transformers` version:
+- Platform:
+- Python version:
+- PyTorch version (GPU?):
+- Tensorflow version (GPU?):
+- Using GPU in script?:
+- Using distributed or parallel set-up in script?:
+
+<!-- IMPORTANT: which version of the former library do you use? -->
+* `pytorch-transformers` or `pytorch-pretrained-bert` version (or branch):
 
-* OS:
-* Python version:
-* PyTorch version:
-* PyTorch Transformers version (or branch):
-* Using GPU ?
-* Distributed of parallel setup ?
-* Any other relevant information:
 
 ## Checklist
 
 - [ ] I have read the migration guide in the readme.
+ ([pytorch-transformers](https://github.com/huggingface/transformers#migrating-from-pytorch-transformers-to-transformers);
+  [pytorch-pretrained-bert](https://github.com/huggingface/transformers#migrating-from-pytorch-pretrained-bert-to-transformers))
 - [ ] I checked if a related official extension example runs on my machine.
-
-## Additional context
-
-<!-- Add any other context about the problem here. -->

.github/ISSUE_TEMPLATE/question-help.md

@@ -1,8 +1,29 @@
 ---
-name: "❓Questions & Help"
-about: Start a general discussion related to PyTorch Transformers
+name: "❓ Questions & Help"
+about: Post your general questions on Stack Overflow tagged huggingface-transformers
+title: ''
+labels: ''
+assignees: ''
+
 ---
 
-## ❓ Questions & Help
+# ❓ Questions & Help
 
-<!-- A clear and concise description of the question. -->
+<!-- The GitHub issue tracker is primarly intended for bugs, feature requests,
+     new models and benchmarks, and migration questions. For all other questions,
+     we direct you to Stack Overflow (SO) where a whole community of PyTorch and
+     Tensorflow enthusiast can help you out. Make sure to tag your question with the
+     right deep learning framework as well as the huggingface-transformers tag: 
+     https://stackoverflow.com/questions/tagged/huggingface-transformers 
+     
+     If your question wasn't answered after a period of time on Stack Overflow, you
+     can always open a question on GitHub. You should then link to the SO question 
+     that you posted.
+     -->
+
+## Details
+<!-- Description of your issue -->
+
+<!-- You should first ask your question on SO, and only if
+     you didn't get an answer ask it here on GitHub. -->
+**A link to original question on Stack Overflow**: 

.github/workflows/github-push.yml

@@ -0,0 +1,19 @@
+name: GitHub-hosted runner
+
+on: push
+
+jobs:
+  check_code_quality:
+    runs-on: ubuntu-18.04
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python
+      uses: actions/setup-python@v1
+      with:
+        python-version: 3.7
+    - name: Install dependencies
+      run: |
+        pip install .[tf,torch,quality]
+
+
+

.github/workflows/self-push.yml

@@ -0,0 +1,50 @@
+name: Self-hosted runner (push)
+
+on: 
+  # push:
+  #   branches:
+  #     - master
+  # pull_request:
+  repository_dispatch:
+
+
+jobs:
+  run_tests_torch_and_tf_gpu:
+    runs-on: self-hosted
+    steps:
+    - uses: actions/checkout@v2
+    - name: Python version
+      run: |
+        which python
+        python --version
+        pip --version
+    - name: Current dir
+      run: pwd
+    - run: nvidia-smi
+    - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+      run: |
+        python -m venv .env
+        source .env/bin/activate
+        which python
+        python --version
+        pip --version
+    - name: Install dependencies
+      run: |
+        source .env/bin/activate
+        pip install .[sklearn,tf,torch,testing]
+        pip uninstall -y tensorflow
+
+    - name: Are GPUs recognized by our DL frameworks
+      run: |
+        source .env/bin/activate
+        python -c "import torch; print(torch.cuda.is_available())"
+
+    - name: Run all non-slow tests on GPU
+      env:
+        TF_FORCE_GPU_ALLOW_GROWTH: "true"
+        # TF_GPU_MEMORY_LIMIT: 4096
+        OMP_NUM_THREADS: 1
+        USE_CUDA: yes
+      run: |
+        source .env/bin/activate
+        python -m pytest -n 2 --dist=loadfile -s -v ./tests/

.github/workflows/self-scheduled.yml

@@ -0,0 +1,51 @@
+name: Self-hosted runner (scheduled)
+
+on:
+  push:
+    branches:
+      - ci_*
+  repository_dispatch:
+  schedule:
+    - cron: "0 0 * * *"
+
+jobs:
+  run_all_tests_torch_and_tf_gpu:
+    runs-on: self-hosted
+    steps:
+    - uses: actions/checkout@v2
+    - name: Python version
+      run: |
+        which python
+        python --version
+        pip --version
+    - name: Current dir
+      run: pwd
+    - run: nvidia-smi
+    - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+      run: |
+        python -m venv .env
+        source .env/bin/activate
+        which python
+        python --version
+        pip --version
+    - name: Install dependencies
+      run: |
+        source .env/bin/activate
+        pip install .[sklearn,tf,torch,testing]
+
+    - name: Are GPUs recognized by our DL frameworks
+      run: |
+        source .env/bin/activate
+        python -c "import torch; print(torch.cuda.is_available())"
+        python -c "import tensorflow as tf; print(tf.test.is_built_with_cuda(), tf.config.list_physical_devices('GPU'))"
+
+    - name: Run all tests on GPU
+      env:
+        TF_FORCE_GPU_ALLOW_GROWTH: "true"
+        OMP_NUM_THREADS: 1
+        RUN_SLOW: yes
+        USE_CUDA: yes
+      run: |
+        source .env/bin/activate
+        python -m pytest -n 1 --dist=loadfile -s -v ./tests/
+        

.gitignore

@@ -118,6 +118,9 @@ dmypy.json
 # vscode
 .vscode
 
+# Pycharm
+.idea
+
 # TF code
 tensorflow_code
 
@@ -130,4 +133,15 @@ runs
 examples/runs
 
 # data
-data
+/data
+serialization_dir
+
+# emacs
+*.*~
+debug.env
+
+# vim
+.*.swp
+
+#ctags
+tags

CONTRIBUTING.md

@@ -0,0 +1,260 @@
+# How to contribute to transformers?
+
+Everyone is welcome to contribute, and we value everybody's contribution. Code
+is thus not the only way to help the community. Answering questions, helping
+others, reaching out and improving the documentations are immensely valuable to
+the community.
+
+It also helps us if you spread the word: reference the library from blog posts
+on the awesome projects it made possible, shout out on Twitter every time it has
+helped you, or simply star the repo to say "thank you".
+
+## You can contribute in so many ways!
+
+There are 4 ways you can contribute to transformers:
+* Fixing outstanding issues with the existing code;
+* Implementing new models;
+* Contributing to the examples or to the documentation;
+* Submitting issues related to bugs or desired new features.
+
+*All are equally valuable to the community.*
+
+## Submitting a new issue or feature request
+
+Do your best to follow these guidelines when submitting an issue or a feature
+request. It will make it easier for us to come back to you quickly and with good
+feedback.
+
+### Did you find a bug?
+
+The transformers are robust and reliable thanks to the users who notify us of
+the problems they encounter. So thank you for reporting an issue.
+
+First, we would really appreciate it if you could **make sure the bug was not
+already reported** (use the search bar on Github under Issues).
+
+Did not find it? :( So we can act quickly on it, please follow these steps:
+
+* Include your **OS type and version**, the versions of **Python**, **PyTorch** and
+  **Tensorflow** when applicable;
+* A short, self-contained, code snippet that allows us to reproduce the bug in
+  less than 30s;
+* Provide the *full* traceback if an exception is raised.
+
+To get the OS and software versions automatically, you can run the following command:
+
+```bash
+python transformers-cli env
+```
+
+### Do you want to implement a new model?
+
+Awesome! Please provide the following information:
+
+* Short description of the model and link to the paper;
+* Link to the implementation if it is open-source;
+* Link to the model weights if they are available.
+
+If you are willing to contribute the model yourself, let us know so we can best
+guide you.
+
+We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder.
+
+### Do you want a new feature (that is not a model)?
+
+A world-class feature request addresses the following points:
+
+1. Motivation first:
+  * Is it related to a problem/frustration with the library? If so, please explain
+    why. Providing a code snippet that demonstrates the problem is best.
+  * Is it related to something you would need for a project? We'd love to hear
+    about it!
+  * Is it something you worked on and think could benefit the community?
+    Awesome! Tell us what problem it solved for you.
+2. Write a *full paragraph* describing the feature;
+3. Provide a **code snippet** that demonstrates its future use;
+4. In case this is related to a paper, please attach a link;
+5. Attach any additional information (drawings, screenshots, etc.) you think may help.
+
+If your issue is well written we're already 80% of the way there by the time you
+post it.
+
+We have added **templates** to guide you in the process of adding a new example script for training or testing the models in the library. You can find them in the [`templates`](./templates) folder.
+
+## Start contributing! (Pull Requests)
+
+Before writing code, we strongly advise you to search through the exising PRs or
+issues to make sure that nobody is already working on the same thing. If you are
+unsure, it is always a good idea to open an issue to get some feedback.
+
+You will need basic `git` proficiency to be able to contribute to
+`transformers`. `git` is not the easiest tool to use but it has the greatest
+manual. Type `git --help` in a shell and enjoy. If you prefer books, [Pro
+Git](https://git-scm.com/book/en/v2) is a very good reference.
+
+Follow these steps to start contributing:
+
+1. Fork the [repository](https://github.com/huggingface/transformers) by
+   clicking on the 'Fork' button on the repository's page. This creates a copy of the code
+   under your GitHub user account.
+
+2. Clone your fork to your local disk, and add the base repository as a remote:
+
+   ```bash
+   $ git clone git@github.com:<your Github handle>/transformers.git
+   $ cd transformers
+   $ git remote add upstream https://github.com/huggingface/transformers.git
+   ```
+
+3. Create a new branch to hold your development changes:
+
+   ```bash
+   $ git checkout -b a-descriptive-name-for-my-changes
+   ```
+
+   **do not** work on the `master` branch.
+
+4. Set up a development environment by running the following command in a virtual environment:
+
+   ```bash
+   $ pip install -e ".[dev]"
+   ```
+
+   (If transformers was already installed in the virtual environment, remove
+   it with `pip uninstall transformers` before reinstalling it in editable
+   mode with the `-e` flag.)
+
+   Right now, we need an unreleased version of `isort` to avoid a
+   [bug](https://github.com/timothycrosley/isort/pull/1000):
+
+   ```bash
+   $ pip install -U git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort
+   ```
+
+5. Develop the features on your branch.
+
+   As you work on the features, you should make sure that the test suite
+   passes:
+
+   ```bash
+   $ make test
+   ```
+
+   `transformers` relies on `black` and `isort` to format its source code
+   consistently. After you make changes, format them with:
+
+   ```bash
+   $ make style
+   ```
+
+   `transformers` also uses `flake8` to check for coding mistakes. Quality
+   control runs in CI, however you can also run the same checks with:
+
+   ```bash
+   $ make quality
+   ```
+
+   Once you're happy with your changes, add changed files using `git add` and
+   make a commit with `git commit` to record your changes locally:
+
+   ```bash
+   $ git add modified_file.py
+   $ git commit
+   ```
+
+   Please write [good commit
+   messages](https://chris.beams.io/posts/git-commit/).
+
+   It is a good idea to sync your copy of the code with the original
+   repository regularly. This way you can quickly account for changes:
+
+   ```bash
+   $ git fetch upstream
+   $ git rebase upstream/master
+   ```
+
+   Push the changes to your account using:
+
+   ```bash
+   $ git push -u origin a-descriptive-name-for-my-changes
+   ```
+
+6. Once you are satisfied (**and the checklist below is happy too**), go to the
+   webpage of your fork on GitHub. Click on 'Pull request' to send your changes
+   to the project maintainers for review.
+
+7. It's ok if maintainers ask you for changes. It happens to core contributors
+   too! So everyone can see the changes in the Pull request, work in your local
+   branch and push the changes to your fork. They will automatically appear in
+   the pull request.
+
+
+### Checklist
+
+1. The title of your pull request should be a summary of its contribution;
+2. If your pull request adresses an issue, please mention the issue number in
+   the pull request description to make sure they are linked (and people
+   consulting the issue know you are working on it);
+3. To indicate a work in progress please prefix the title with `[WIP]`. These
+   are useful to avoid duplicated work, and to differentiate it from PRs ready
+   to be merged;
+4. Make sure existing tests pass;
+5. Add high-coverage tests. No quality test, no merge. 
+ - If you are adding a new model, make sure that you use `ModelTester.all_model_classes = (MyModel, MyModelWithLMHead,...)`, which triggers the common tests.
+ - If you are adding new `@slow` tests, make sure they pass using `RUN_SLOW=1 python -m pytest tests/test_my_new_model.py`. 
+CircleCI does not run them. 
+6. All public methods must have informative docstrings;
+
+### Tests
+
+You can run 🤗 Transformers tests with `unittest` or `pytest`.
+
+We like `pytest` and `pytest-xdist` because it's faster. From the root of the
+repository, here's how to run tests with `pytest` for the library:
+
+```bash
+$ python -m pytest -n auto --dist=loadfile -s -v ./tests/
+```
+
+and for the examples:
+
+```bash
+$ pip install -r examples/requirements.txt  # only needed the first time
+$ python -m pytest -n auto --dist=loadfile -s -v ./examples/
+```
+
+In fact, that's how `make test` and `make test-examples` are implemented!
+
+You can specify a smaller set of tests in order to test only the feature
+you're working on.
+
+By default, slow tests are skipped. Set the `RUN_SLOW` environment variable to
+`yes` to run them. This will download many gigabytes of models — make sure you
+have enough disk space and a good Internet connection, or a lot of patience!
+
+```bash
+$ RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./tests/
+$ RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./examples/
+```
+
+Likewise, set the `RUN_CUSTOM_TOKENIZERS` environment variable to `yes` to run
+tests for custom tokenizers, which don't run by default either.
+
+🤗 Transformers uses `pytest` as a test runner only. It doesn't use any
+`pytest`-specific features in the test suite itself.
+
+This means `unittest` is fully supported. Here's how to run tests with
+`unittest`:
+
+```bash
+$ python -m unittest discover -s tests -t . -v
+$ python -m unittest discover -s examples -t examples -v
+```
+
+
+### Style guide
+
+For documentation strings, `transformers` follows the [google
+style](https://google.github.io/styleguide/pyguide.html).
+
+#### This guide was heavily inspired by the awesome [scikit-learn guide to contributing](https://github.com/scikit-learn/scikit-learn/blob/master/CONTRIBUTING.md)

Makefile

@@ -0,0 +1,24 @@
+.PHONY: quality style test test-examples
+
+# Check that source code meets quality standards
+
+quality:
+	black --check --line-length 119 --target-version py35 examples templates tests src utils
+	isort --check-only --recursive examples templates tests src utils
+	flake8 examples templates tests src utils
+
+# Format source code automatically
+
+style:
+	black --line-length 119 --target-version py35 examples templates tests src utils
+	isort --recursive examples templates tests src utils
+
+# Run tests for the library
+
+test:
+	python -m pytest -n auto --dist=loadfile -s -v ./tests/
+
+# Run tests for examples
+
+test-examples:
+	python -m pytest -n auto --dist=loadfile -s -v ./examples/

README.md

@@ -1,10 +1,152 @@
-# 👾 PyTorch-Transformers
+<p align="center">
+    <br>
+    <img src="https://raw.githubusercontent.com/huggingface/transformers/master/docs/source/imgs/transformers_logo_name.png" width="400"/>
+    <br>
+<p>
+<p align="center">
+    <a href="https://circleci.com/gh/huggingface/transformers">
+        <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/master">
+    </a>
+    <a href="https://github.com/huggingface/transformers/blob/master/LICENSE">
+        <img alt="GitHub" src="https://img.shields.io/github/license/huggingface/transformers.svg?color=blue">
+    </a>
+    <a href="https://huggingface.co/transformers/index.html">
+        <img alt="Documentation" src="https://img.shields.io/website/http/huggingface.co/transformers/index.html.svg?down_color=red&down_message=offline&up_message=online">
+    </a>
+    <a href="https://github.com/huggingface/transformers/releases">
+        <img alt="GitHub release" src="https://img.shields.io/github/release/huggingface/transformers.svg">
+    </a>
+</p>
 
-[![CircleCI](https://circleci.com/gh/huggingface/pytorch-transformers.svg?style=svg)](https://circleci.com/gh/huggingface/pytorch-transformers)
+<h3 align="center">
+<p>State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch
+</h3>
 
-PyTorch-Transformers (formerly known as `pytorch-pretrained-bert`) is a library of state-of-the-art pre-trained models for Natural Language Processing (NLP).
+🤗 Transformers (formerly known as `pytorch-transformers` and `pytorch-pretrained-bert`) provides state-of-the-art general-purpose architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet, CTRL...) for Natural Language Understanding (NLU) and Natural Language Generation (NLG) with over 32+ pretrained models in 100+ languages and deep interoperability between TensorFlow 2.0 and PyTorch.
 
-The library currently contains PyTorch implementations, pre-trained model weights, usage scripts and conversion utilities for the following models:
+[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/0)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/0)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/1)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/1)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/2)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/2)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/3)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/3)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/4)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/4)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/5)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/5)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/6)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/6)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/7)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/7)
+
+### Features
+
+- As easy to use as pytorch-transformers
+- As powerful and concise as Keras
+- High performance on NLU and NLG tasks
+- Low barrier to entry for educators and practitioners
+
+State-of-the-art NLP for everyone
+- Deep learning researchers
+- Hands-on practitioners
+- AI/ML/NLP teachers and educators
+
+Lower compute costs, smaller carbon footprint
+- Researchers can share trained models instead of always retraining
+- Practitioners can reduce compute time and production costs
+- 10 architectures with over 30 pretrained models, some in more than 100 languages
+
+Choose the right framework for every part of a model's lifetime
+- Train state-of-the-art models in 3 lines of code
+- Deep interoperability between TensorFlow 2.0 and PyTorch models
+- Move a single model between TF2.0/PyTorch frameworks at will
+- Seamlessly pick the right framework for training, evaluation, production
+
+
+| Section | Description |
+|-|-|
+| [Installation](#installation) | How to install the package |
+| [Model architectures](#model-architectures) | Architectures (with pretrained weights) |
+| [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][(v2.5.0)](https://huggingface.co/transformers/v2.5.0)[(v2.4.0/v2.4.1)](https://huggingface.co/transformers/v2.4.0)[(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
+
+This repo is tested on Python 3.6+, PyTorch 1.0.0+ and TensorFlow 2.0.0-rc1
+
+You should install 🤗 Transformers in a [virtual environment](https://docs.python.org/3/library/venv.html). If you're unfamiliar with Python virtual environments, check out the [user guide](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/).
+
+Create a virtual environment with the version of Python you're going to use and activate it.
+
+Now, if you want to use 🤗 Transformers, you can install it with pip. If you'd like to play with the examples, you must install it from source.
+
+### With pip
+
+First you need to install one of, or both, TensorFlow 2.0 and PyTorch.
+Please refer to [TensorFlow installation page](https://www.tensorflow.org/install/pip#tensorflow-2.0-rc-is-available) and/or [PyTorch installation page](https://pytorch.org/get-started/locally/#start-locally) regarding the specific install command for your platform.
+
+When TensorFlow 2.0 and/or PyTorch has been installed, 🤗 Transformers can be installed using pip as follows:
+
+```bash
+pip install transformers
+```
+
+### From source
+
+Here also, you first need to install one of, or both, TensorFlow 2.0 and PyTorch.
+Please refer to [TensorFlow installation page](https://www.tensorflow.org/install/pip#tensorflow-2.0-rc-is-available) and/or [PyTorch installation page](https://pytorch.org/get-started/locally/#start-locally) regarding the specific install command for your platform.
+
+When TensorFlow 2.0 and/or PyTorch has been installed, you can install from source by cloning the repository and running:
+
+```bash
+git clone https://github.com/huggingface/transformers
+cd transformers
+pip install .
+```
+
+When you update the repository, you should upgrade the transformers installation and its dependencies as follows:
+
+```bash
+git pull
+pip install --upgrade .
+```
+
+### Run the examples
+
+Examples are included in the repository but are not shipped with the library.
+
+Therefore, in order to run the latest versions of the examples, you need to install from source, as described above.
+
+Look at the [README](https://github.com/huggingface/transformers/blob/master/examples/README.md) for how to run examples.
+
+### Tests
+
+A series of tests are included for the library and for some example scripts. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples).
+
+Depending on which framework is installed (TensorFlow 2.0 and/or PyTorch), the irrelevant tests will be skipped. Ensure that both frameworks are installed if you want to execute all tests.
+
+Here's the easiest way to run tests for the library:
+
+```bash
+pip install -e ".[testing]"
+make test
+```
+
+and for the examples:
+
+```bash
+pip install -e ".[testing]"
+pip install -r examples/requirements.txt
+make test-examples
+```
+
+For details, refer to the [contributing guide](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md#tests).
+
+### Do you want to run a Transformer model on a mobile device?
+
+You should check out our [`swift-coreml-transformers`](https://github.com/huggingface/swift-coreml-transformers) repo.
+
+It contains a set of tools to convert PyTorch or TensorFlow 2.0 trained Transformer models (currently contains `GPT-2`, `DistilGPT-2`, `BERT`, and `DistilBERT`) to CoreML models that run on iOS devices.
+
+At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models to productizing them in CoreML, or prototype a model or an app in CoreML then research its hyperparameters or architecture from TensorFlow 2.0 and/or PyTorch. Super exciting!
+
+## Model architectures
+
+🤗 Transformers currently provides the following NLU/NLG architectures:
 
 1. **[BERT](https://github.com/google-research/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
 2. **[GPT](https://github.com/openai/finetune-transformer-lm)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
@@ -13,80 +155,54 @@ The library currently contains PyTorch implementations, pre-trained model weight
 5. **[XLNet](https://github.com/zihangdai/xlnet/)** (from Google/CMU) released with the paper [​XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
 6. **[XLM](https://github.com/facebookresearch/XLM/)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 7. **[RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/roberta)** (from Facebook), released together with the paper a [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
-8. **[DistilBERT](https://github.com/huggingface/pytorch-transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the blogpost [Smaller, faster, cheaper, lighter: Introducing DistilBERT, a distilled version of BERT](https://medium.com/huggingface/distilbert-8cf3380435b5
-) by Victor Sanh, Lysandre Debut and Thomas Wolf.
+8. **[DistilBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
+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.
+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. **[MMBT](https://github.com/facebookresearch/mmbt/)** (from Facebook), released together with the paper a [Supervised Multimodal Bitransformers for Classifying Images and Text](https://arxiv.org/pdf/1909.02950.pdf) by Douwe Kiela, Suvrat Bhooshan, Hamed Firooz, Davide Testuggine.
+15. **[FlauBERT](https://github.com/getalp/Flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
+16. **[BART](https://github.com/pytorch/fairseq/tree/master/examples/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
+17. **[ELECTRA](https://github.com/google-research/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+18. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
+19. 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/pytorch-transformers/examples.html).
+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).
 
-| Section | Description |
-|-|-|
-| [Installation](#installation) | How to install the package |
-| [Quick tour: Usage](#quick-tour) | Tokenizers & models usage: Bert and GPT-2 |
-| [Quick tour: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
-| [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-pytorch-transformers) | Migrating your code from pytorch-pretrained-bert to pytorch-transformers |
-| [Documentation](https://huggingface.co/pytorch-transformers/) | Full API documentation and more |
+## Online demo
 
-## Installation
+**[Write With Transformer](https://transformer.huggingface.co)**, built by the Hugging Face team at transformer.huggingface.co, is the official demo of this repo’s text generation capabilities.
+You can use it to experiment with completions generated by `GPT2Model`, `TransfoXLModel`, and `XLNetModel`.
 
-This repo is tested on Python 2.7 and 3.5+ (examples are tested only on python 3.5+) and PyTorch 1.0.0+
-
-### With pip
-
-PyTorch-Transformers can be installed by pip as follows:
-
-```bash
-pip install pytorch-transformers
-```
-
-### From source
-
-Clone the repository and run:
-
-```bash
-pip install [--editable] .
-```
-
-### Tests
-
-A series of tests is included for the library and the example scripts. Library tests can be found in the [tests folder](https://github.com/huggingface/pytorch-transformers/tree/master/pytorch_transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/pytorch-transformers/tree/master/examples).
-
-These tests can be run using `pytest` (install pytest if needed with `pip install pytest`).
-
-You can run the tests from the root of the cloned repository with the commands:
-
-```bash
-python -m pytest -sv ./pytorch_transformers/tests/
-python -m pytest -sv ./examples/
-```
-
-### Do you want to run a Transformer model on a mobile device?
-
-You should check out our [`swift-coreml-transformers`](https://github.com/huggingface/swift-coreml-transformers) repo.
-
-It contains an example of a conversion script from a Pytorch trained Transformer model (here, `GPT-2`) to a CoreML model that runs on iOS devices.
-
-At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch to productizing them in CoreML,
-or prototype a model or an app in CoreML then research its hyperparameters or architecture from PyTorch. Super exciting!
+> “🦄 Write with transformer is to writing what calculators are to calculus.”
 
+![write_with_transformer](https://transformer.huggingface.co/front/assets/thumbnail-large.png)
 
 ## Quick tour
 
-Let's do a very quick overview of PyTorch-Transformers. Detailed examples for each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the [full documentation](https://huggingface.co/pytorch-transformers/).
+Let's do a very quick overview of the model architectures in 🤗 Transformers. Detailed examples for each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the [full documentation](https://huggingface.co/transformers/).
 
 ```python
 import torch
-from pytorch_transformers import *
+from transformers import *
 
-# PyTorch-Transformers has a unified API
-# for 7 transformer architectures and 30 pretrained weights.
+# Transformers has a unified API
+# for 10 transformer architectures and 30 pretrained weights.
 #          Model          | Tokenizer          | Pretrained weights shortcut
-MODELS = [(BertModel,       BertTokenizer,      'bert-base-uncased'),
-          (OpenAIGPTModel,  OpenAIGPTTokenizer, 'openai-gpt'),
-          (GPT2Model,       GPT2Tokenizer,      'gpt2'),
-          (TransfoXLModel,  TransfoXLTokenizer, 'transfo-xl-wt103'),
-          (XLNetModel,      XLNetTokenizer,     'xlnet-base-cased'),
-          (XLMModel,        XLMTokenizer,       'xlm-mlm-enfr-1024'),
-          (RobertaModel,    RobertaTokenizer,   'roberta-base')]
+MODELS = [(BertModel,       BertTokenizer,       'bert-base-uncased'),
+          (OpenAIGPTModel,  OpenAIGPTTokenizer,  'openai-gpt'),
+          (GPT2Model,       GPT2Tokenizer,       'gpt2'),
+          (CTRLModel,       CTRLTokenizer,       'ctrl'),
+          (TransfoXLModel,  TransfoXLTokenizer,  'transfo-xl-wt103'),
+          (XLNetModel,      XLNetTokenizer,      'xlnet-base-cased'),
+          (XLMModel,        XLMTokenizer,        'xlm-mlm-enfr-1024'),
+          (DistilBertModel, DistilBertTokenizer, 'distilbert-base-cased'),
+          (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`
 
 # Let's encode some text in a sequence of hidden-states using each model:
 for model_class, tokenizer_class, pretrained_weights in MODELS:
@@ -101,44 +217,97 @@ for model_class, tokenizer_class, pretrained_weights in MODELS:
 
 # Each architecture is provided with several class for fine-tuning on down-stream tasks, e.g.
 BERT_MODEL_CLASSES = [BertModel, BertForPreTraining, BertForMaskedLM, BertForNextSentencePrediction,
-                      BertForSequenceClassification, BertForMultipleChoice, BertForTokenClassification,
-                      BertForQuestionAnswering]
+                      BertForSequenceClassification, BertForTokenClassification, BertForQuestionAnswering]
 
 # All the classes for an architecture can be initiated from pretrained weights for this architecture
 # Note that additional weights added for fine-tuning are only initialized
 # and need to be trained on the down-stream task
-tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+pretrained_weights = 'bert-base-uncased'
+tokenizer = BertTokenizer.from_pretrained(pretrained_weights)
 for model_class in BERT_MODEL_CLASSES:
     # Load pretrained model/tokenizer
-    model = model_class.from_pretrained('bert-base-uncased')
+    model = model_class.from_pretrained(pretrained_weights)
 
-# Models can return full list of hidden-states & attentions weights at each layer
-model = model_class.from_pretrained(pretrained_weights,
-                                    output_hidden_states=True,
-                                    output_attentions=True)
-input_ids = torch.tensor([tokenizer.encode("Let's see all hidden-states and attentions on this text")])
-all_hidden_states, all_attentions = model(input_ids)[-2:]
+    # Models can return full list of hidden-states & attentions weights at each layer
+    model = model_class.from_pretrained(pretrained_weights,
+                                        output_hidden_states=True,
+                                        output_attentions=True)
+    input_ids = torch.tensor([tokenizer.encode("Let's see all hidden-states and attentions on this text")])
+    all_hidden_states, all_attentions = model(input_ids)[-2:]
 
-# Models are compatible with Torchscript
-model = model_class.from_pretrained(pretrained_weights, torchscript=True)
-traced_model = torch.jit.trace(model, (input_ids,))
+    # Models are compatible with Torchscript
+    model = model_class.from_pretrained(pretrained_weights, torchscript=True)
+    traced_model = torch.jit.trace(model, (input_ids,))
 
-# Simple serialization for models and tokenizers
-model.save_pretrained('./directory/to/save/')  # save
-model = model_class.from_pretrained('./directory/to/save/')  # re-load
-tokenizer.save_pretrained('./directory/to/save/')  # save
-tokenizer = tokenizer_class.from_pretrained('./directory/to/save/')  # re-load
+    # Simple serialization for models and tokenizers
+    model.save_pretrained('./directory/to/save/')  # save
+    model = model_class.from_pretrained('./directory/to/save/')  # re-load
+    tokenizer.save_pretrained('./directory/to/save/')  # save
+    tokenizer = BertTokenizer.from_pretrained('./directory/to/save/')  # re-load
 
-# SOTA examples for GLUE, SQUAD, text generation...
+    # SOTA examples for GLUE, SQUAD, text generation...
+```
+
+## Quick tour TF 2.0 training and PyTorch interoperability
+
+Let's do a quick example of how a TensorFlow 2.0 model can be trained in 12 lines of code with 🤗 Transformers and then loaded in PyTorch for fast inspection/tests.
+
+```python
+import tensorflow as tf
+import tensorflow_datasets
+from transformers import *
+
+# Load dataset, tokenizer, model from pretrained model/vocabulary
+tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
+model = TFBertForSequenceClassification.from_pretrained('bert-base-cased')
+data = tensorflow_datasets.load('glue/mrpc')
+
+# Prepare dataset for GLUE as a tf.data.Dataset instance
+train_dataset = glue_convert_examples_to_features(data['train'], tokenizer, max_length=128, task='mrpc')
+valid_dataset = glue_convert_examples_to_features(data['validation'], tokenizer, max_length=128, task='mrpc')
+train_dataset = train_dataset.shuffle(100).batch(32).repeat(2)
+valid_dataset = valid_dataset.batch(64)
+
+# Prepare training: Compile tf.keras model with optimizer, loss and learning rate schedule
+optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
+loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
+metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')
+model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
+
+# Train and evaluate using tf.keras.Model.fit()
+history = model.fit(train_dataset, epochs=2, steps_per_epoch=115,
+                    validation_data=valid_dataset, validation_steps=7)
+
+# Load the TensorFlow model in PyTorch for inspection
+model.save_pretrained('./save/')
+pytorch_model = BertForSequenceClassification.from_pretrained('./save/', from_tf=True)
+
+# Quickly test a few predictions - MRPC is a paraphrasing task, let's see if our model learned the task
+sentence_0 = "This research was consistent with his findings."
+sentence_1 = "His findings were compatible with this research."
+sentence_2 = "His findings were not compatible with this research."
+inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
+inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
+
+pred_1 = pytorch_model(inputs_1['input_ids'], token_type_ids=inputs_1['token_type_ids'])[0].argmax().item()
+pred_2 = pytorch_model(inputs_2['input_ids'], token_type_ids=inputs_2['token_type_ids'])[0].argmax().item()
+
+print("sentence_1 is", "a paraphrase" if pred_1 else "not a paraphrase", "of sentence_0")
+print("sentence_2 is", "a paraphrase" if pred_2 else "not a paraphrase", "of sentence_0")
 ```
 
 ## Quick tour of the fine-tuning/usage scripts
 
+**Important**
+Before running the fine-tuning scripts, please read the
+[instructions](#run-the-examples) on how to
+setup your environment to run the examples.
+
 The library comprises several example scripts with SOTA performances for NLU and NLG tasks:
 
 - `run_glue.py`: an example fine-tuning Bert, XLNet and XLM on nine different GLUE tasks (*sequence-level classification*)
 - `run_squad.py`: an example fine-tuning Bert, XLNet and XLM on the question answering dataset SQuAD 2.0 (*token-level classification*)
-- `run_generation.py`: an example using GPT, GPT-2, Transformer-XL and XLNet for conditional language generation
+- `run_generation.py`: an example using GPT, GPT-2, CTRL, Transformer-XL and XLNet for conditional language generation
 - other model-specific examples (see the documentation).
 
 Here are three quick usage examples for these scripts:
@@ -276,10 +445,10 @@ python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json ../models/wwm_uncase
 
 This is the model provided as `bert-large-uncased-whole-word-masking-finetuned-squad`.
 
-### `run_generation.py`: Text generation with GPT, GPT-2, Transformer-XL and XLNet
+### `run_generation.py`: Text generation with GPT, GPT-2, CTRL, Transformer-XL and XLNet
 
 A conditional generation script is also included to generate text from a prompt.
-The generation script includes the [tricks](https://github.com/rusiaaman/XLNet-gen#methodology) proposed by by Aman Rusia to get high quality generation with memory models like Transformer-XL and XLNet (include a predefined text to make short inputs longer).
+The generation script includes the [tricks](https://github.com/rusiaaman/XLNet-gen#methodology) proposed by Aman Rusia to get high-quality generation with memory models like Transformer-XL and XLNet (include a predefined text to make short inputs longer).
 
 Here is how to run the script with the small version of OpenAI GPT-2 model:
 
@@ -290,19 +459,129 @@ python ./examples/run_generation.py \
     --model_name_or_path=gpt2 \
 ```
 
-## Migrating from pytorch-pretrained-bert to pytorch-transformers
+and from the Salesforce CTRL model:
+```shell
+python ./examples/run_generation.py \
+    --model_type=ctrl \
+    --length=20 \
+    --model_name_or_path=ctrl \
+    --temperature=0 \
+    --repetition_penalty=1.2 \
+```
 
-Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to `pytorch-transformers`
+## Quick tour of model 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)**. Optionally, join an existing organization or create a new one. 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)
+```
+
+If you want your model to be namespaced by your organization name rather than your username, add the following flag to any command:
+```shell
+--organization organization_name
+```
+
+Your model will then be accessible through its identifier, a concatenation of your username (or organization name) and the folder name above:
+```python
+"username/pretrained_model"
+# or if an org:
+"organization_name/pretrained_model"
+```
+
+**Please add a README.md model card** to the repo under `model_cards/` with: model description, training params (dataset, preprocessing, hardware used, hyperparameters), evaluation results, intended uses & limitations, etc.
+
+Your model now has a page on huggingface.co/models 🔥
+
+Anyone can load it from code:
+```python
+tokenizer = AutoTokenizer.from_pretrained("namespace/pretrained_model")
+model = AutoModel.from_pretrained("namespace/pretrained_model")
+```
+
+List all your files on S3:
+```shell
+transformers-cli s3 ls
+```
+
+You can also delete unneeded files:
+
+```shell
+transformers-cli s3 rm …
+```
+
+## 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.
+ - `text-classification`: Initialize a `TextClassificationPipeline` directly, or see `sentiment-analysis` for an example.
+ - `question-answering`: Provided some context and a question refering to the context, it will extract the answer to the question in the context.
+ - `fill-mask`: Takes an input sequence containing a masked token (e.g. `<mask>`) and return list of most probable filled sequences, with their probabilities.
+
+```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`.
+
+### Positional order of some models' keywords inputs (`attention_mask`, `token_type_ids`...) changed
+
+To be able to use Torchscript (see #1010, #1204 and #1195) the specific order of some models **keywords inputs** (`attention_mask`, `token_type_ids`...) has been changed.
+
+If you used to call the models with keyword names for keyword arguments, e.g. `model(inputs_ids, attention_mask=attention_mask, token_type_ids=token_type_ids)`, this should not cause any change.
+
+If you used to call the models with positional inputs for keyword arguments, e.g. `model(inputs_ids, attention_mask, token_type_ids)`, you may have to double check the exact order of input arguments.
+
+
+## Migrating from pytorch-pretrained-bert to transformers
+
+Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to `transformers`.
 
 ### Models always output `tuples`
 
-The main breaking change when migrating from `pytorch-pretrained-bert` to `pytorch-transformers` is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
+The main breaking change when migrating from `pytorch-pretrained-bert` to `transformers` is that every model's forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
 
-The exact content of the tuples for each model are detailed in the models' docstrings and the [documentation](https://huggingface.co/pytorch-transformers/).
+The exact content of the tuples for each model is detailed in the models' docstrings and the [documentation](https://huggingface.co/transformers/).
 
 In pretty much every case, you will be fine by taking the first element of the output as the output you previously used in `pytorch-pretrained-bert`.
 
-Here is a `pytorch-pretrained-bert` to `pytorch-transformers` conversion example for a `BertForSequenceClassification` classification model:
+Here is a `pytorch-pretrained-bert` to `transformers` conversion example for a `BertForSequenceClassification` classification model:
 
 ```python
 # Let's load our model
@@ -311,11 +590,11 @@ model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
 # If you used to have this line in pytorch-pretrained-bert:
 loss = model(input_ids, labels=labels)
 
-# Now just use this line in pytorch-transformers to extract the loss from the output tuple:
+# Now just use this line in transformers to extract the loss from the output tuple:
 outputs = model(input_ids, labels=labels)
 loss = outputs[0]
 
-# In pytorch-transformers you can also have access to the logits:
+# In transformers you can also have access to the logits:
 loss, logits = outputs[:2]
 
 # And even the attention weights if you configure the model to output them (and other outputs too, see the docstrings and documentation)
@@ -324,13 +603,17 @@ outputs = model(input_ids, labels=labels)
 loss, logits, attentions = outputs
 ```
 
+### Using hidden states
+
+By enabling the configuration option `output_hidden_states`, it was possible to retrieve the last hidden states of the encoder. In `pytorch-transformers` as well as `transformers` the return value has changed slightly: `all_hidden_states` now also includes the hidden state of the embeddings in addition to those of the encoding layers. This allows users to easily access the embeddings final state.
+
 ### Serialization
 
-Breaking change in the `from_pretrained()`method:
+Breaking change in the `from_pretrained()` method:
 
-1. Models are now set in evaluation mode by default when instantiated with the `from_pretrained()` method. To train them don't forget to set them back in training mode (`model.train()`) to activate the dropout modules.
+1. Models are now set in evaluation mode by default when instantiated with the `from_pretrained()` method. To train them, don't forget to set them back in training mode (`model.train()`) to activate the dropout modules.
 
-2. The additional `*input` and `**kwargs` arguments supplied to the `from_pretrained()` method used to be directly passed to the underlying model's class `__init__()` method. They are now used to update the model configuration attribute instead which can break derived model classes build based on the previous `BertForSequenceClassification` examples. We are working on a way to mitigate this breaking change in [#866](https://github.com/huggingface/pytorch-transformers/pull/866) by forwarding the the model `__init__()` method (i) the provided positional arguments and (ii) the keyword arguments which do not match any configuration class attributes.
+2. The additional `*input` and `**kwargs` arguments supplied to the `from_pretrained()` method used to be directly passed to the underlying model's class `__init__()` method. They are now used to update the model configuration attribute instead, which can break derived model classes built based on the previous `BertForSequenceClassification` examples. We are working on a way to mitigate this breaking change in [#866](https://github.com/huggingface/transformers/pull/866) by forwarding the the model's `__init__()` method (i) the provided positional arguments and (ii) the keyword arguments which do not match any configuration class attributes.
 
 Also, while not a breaking change, the serialization methods have been standardized and you probably should switch to the new method `save_pretrained(save_directory)` if you were using any other serialization method before.
 
@@ -375,23 +658,24 @@ Here is a conversion examples from `BertAdam` with a linear warmup and decay sch
 # Parameters:
 lr = 1e-3
 max_grad_norm = 1.0
-num_total_steps = 1000
+num_training_steps = 1000
 num_warmup_steps = 100
-warmup_proportion = float(num_warmup_steps) / float(num_total_steps)  # 0.1
+warmup_proportion = float(num_warmup_steps) / float(num_training_steps)  # 0.1
 
 ### Previously BertAdam optimizer was instantiated like this:
-optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
+optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_training_steps)
 ### and used like this:
 for batch in train_data:
     loss = model(batch)
     loss.backward()
     optimizer.step()
 
-### In PyTorch-Transformers, optimizer and schedules are splitted and instantiated like this:
+### In Transformers, optimizer and schedules are splitted and instantiated like this:
 optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False)  # To reproduce BertAdam specific behavior set correct_bias=False
-scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_steps)  # PyTorch scheduler
+scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)  # PyTorch scheduler
 ### and used like this:
 for batch in train_data:
+    model.train()
     loss = model(batch)
     loss.backward()
     torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm)  # Gradient clipping is not in AdamW anymore (so you can use amp without issue)
@@ -402,4 +686,13 @@ for batch in train_data:
 
 ## Citation
 
-At the moment, there is no paper associated to PyTorch-Transformers but we are working on preparing one. In the meantime, please include a mention of the library and a link to the present repository if you use this work in a published or open-source project.
+We now have a paper you can cite for the 🤗 Transformers library:
+```bibtex
+@article{Wolf2019HuggingFacesTS,
+  title={HuggingFace's Transformers: State-of-the-art Natural Language Processing},
+  author={Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and R'emi Louf and Morgan Funtowicz and Jamie Brew},
+  journal={ArXiv},
+  year={2019},
+  volume={abs/1910.03771}
+}
+```

deploy_multi_version_doc.sh

@@ -0,0 +1,23 @@
+cd docs
+
+function deploy_doc(){
+	echo "Creating doc at commit $1 and pushing to folder $2"
+	git checkout $1
+	if [ ! -z "$2" ] 
+	then
+		echo "Pushing version" $2
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
+	else
+		echo "Pushing master"
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+	fi
+}
+
+deploy_doc "master" 
+deploy_doc "b33a385" v1.0.0
+deploy_doc "fe02e45" v1.1.0
+deploy_doc "89fd345" v1.2.0
+deploy_doc "fc9faa8" v2.0.0
+deploy_doc "3ddce1d" v2.1.1
+deploy_doc "f2f3294" v2.2.0
+deploy_doc "d0f8b9a" v2.3.0

docker/Dockerfile

@@ -1,7 +0,0 @@
-FROM pytorch/pytorch:latest
-
-RUN git clone https://github.com/NVIDIA/apex.git && cd apex && python setup.py install --cuda_ext --cpp_ext
-
-RUN pip install pytorch_transformers
-
-WORKDIR /workspace

docker/transformers-cpu/Dockerfile

@@ -0,0 +1,26 @@
+FROM ubuntu:18.04
+LABEL maintainer="Hugging Face"
+LABEL repository="transformers"
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   curl \
+                   ca-certificates \
+                   python3 \
+                   python3-pip && \
+    rm -rf /var/lib/apt/lists
+
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+    jupyter \
+    tensorflow-cpu \
+    torch
+
+WORKDIR /workspace
+COPY . transformers/
+RUN cd transformers/ && \
+    python3 -m pip install --no-cache-dir .
+
+CMD ["/bin/bash"]

docker/transformers-gpu/Dockerfile

@@ -0,0 +1,26 @@
+FROM nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04
+LABEL maintainer="Hugging Face"
+LABEL repository="transformers"
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   curl \
+                   ca-certificates \
+                   python3 \
+                   python3-pip && \
+    rm -rf /var/lib/apt/lists
+
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+    jupyter \
+    tensorflow \
+    torch
+
+WORKDIR /workspace
+COPY . transformers/
+RUN cd transformers/ && \
+    python3 -m pip install --no-cache-dir .
+
+CMD ["/bin/bash"]

docker/transformers-pytorch-cpu/Dockerfile

@@ -0,0 +1,25 @@
+FROM ubuntu:18.04
+LABEL maintainer="Hugging Face"
+LABEL repository="transformers"
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   curl \
+                   ca-certificates \
+                   python3 \
+                   python3-pip && \
+    rm -rf /var/lib/apt/lists
+
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+    jupyter \
+    torch
+
+WORKDIR /workspace
+COPY . transformers/
+RUN cd transformers/ && \
+    python3 -m pip install --no-cache-dir .
+
+CMD ["/bin/bash"]

docker/transformers-pytorch-gpu/Dockerfile

@@ -0,0 +1,25 @@
+FROM nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04
+LABEL maintainer="Hugging Face"
+LABEL repository="transformers"
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   curl \
+                   ca-certificates \
+                   python3 \
+                   python3-pip && \
+    rm -rf /var/lib/apt/lists
+
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+    mkl \
+    torch
+
+WORKDIR /workspace
+COPY . transformers/
+RUN cd transformers/ && \
+    python3 -m pip install --no-cache-dir .
+
+CMD ["/bin/bash"]

docker/transformers-tensorflow-cpu/Dockerfile

@@ -0,0 +1,25 @@
+FROM ubuntu:18.04
+LABEL maintainer="Hugging Face"
+LABEL repository="transformers"
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   curl \
+                   ca-certificates \
+                   python3 \
+                   python3-pip && \
+    rm -rf /var/lib/apt/lists
+
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+    mkl \
+    tensorflow-cpu
+
+WORKDIR /workspace
+COPY . transformers/
+RUN cd transformers/ && \
+    python3 -m pip install --no-cache-dir .
+
+CMD ["/bin/bash"]

docker/transformers-tensorflow-gpu/Dockerfile

@@ -0,0 +1,25 @@
+FROM nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04
+LABEL maintainer="Hugging Face"
+LABEL repository="transformers"
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   curl \
+                   ca-certificates \
+                   python3 \
+                   python3-pip && \
+    rm -rf /var/lib/apt/lists
+
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+    mkl \
+    tensorflow
+
+WORKDIR /workspace
+COPY . transformers/
+RUN cd transformers/ && \
+    python3 -m pip install --no-cache-dir .
+
+CMD ["/bin/bash"]

docs/README.md

@@ -1,25 +1,25 @@
 # Generating the documentation
 
 To generate the documentation, you first have to build it. Several packages are necessary to build the doc,
-you can install them using:
+you can install them with the following command, at the root of the code repository:
 
 ```bash
-pip install -r requirements.txt
+pip install -e ".[docs]"
 ```
- 
+
 ## Packages installed
 
-Here's an overview of all the packages installed. If you ran the previous command installing all packages from 
+Here's an overview of all the packages installed. If you ran the previous command installing all packages from
 `requirements.txt`, you do not need to run the following commands.
 
-Building it requires the package `sphinx` that you can 
+Building it requires the package `sphinx` that you can
 install using:
 
 ```bash
 pip install -U sphinx
 ```
 
-You would also need the custom installed [theme](https://github.com/readthedocs/sphinx_rtd_theme) by 
+You would also need the custom installed [theme](https://github.com/readthedocs/sphinx_rtd_theme) by
 [Read The Docs](https://readthedocs.org/). You can install it using the following command:
 
 ```bash
@@ -34,16 +34,25 @@ pip install recommonmark
 
 ## Building the documentation
 
+Make sure that there is a symlink from the `example` file (in /examples) inside the source folder. Run the following
+command to generate it:
+
+```bash
+ln -s ../../examples/README.md examples.md
+```
+
 Once you have setup `sphinx`, you can build the documentation by running the following command in the `/docs` folder:
 
 ```bash
 make html
 ```
 
+A folder called ``_build/html`` should have been created. You can now open the file ``_build/html/index.html`` in your browser. 
+
 ---
 **NOTE**
 
-If you are adding/removing elements from the toc-tree or from any strutural item, it is recommended to clean the build
+If you are adding/removing elements from the toc-tree or from any structural item, it is recommended to clean the build
 directory before rebuilding. Run the following command to clean and build:
 
 ```bash

docs/requirements.txt

@@ -1,28 +0,0 @@
-alabaster==0.7.12
-Babel==2.7.0
-certifi==2019.6.16
-chardet==3.0.4
-commonmark==0.9.0
-docutils==0.14
-future==0.17.1
-idna==2.8
-imagesize==1.1.0
-Jinja2==2.10.1
-MarkupSafe==1.1.1
-packaging==19.0
-Pygments==2.4.2
-pyparsing==2.4.0
-pytz==2019.1
-recommonmark==0.5.0
-requests==2.22.0
-six==1.12.0
-snowballstemmer==1.9.0
-Sphinx==2.1.2
-sphinx-rtd-theme==0.4.3
-sphinxcontrib-applehelp==1.0.1
-sphinxcontrib-devhelp==1.0.1
-sphinxcontrib-htmlhelp==1.0.2
-sphinxcontrib-jsmath==1.0.1
-sphinxcontrib-qthelp==1.0.2
-sphinxcontrib-serializinghtml==1.1.3
-urllib3==1.25.3

docs/source/_static/css/huggingface.css

@@ -1,4 +1,24 @@
-huggingface.css
+/* Our DOM objects */
+
+.framework-selector {
+    display: flex;
+    flex-direction: row;
+    justify-content: flex-end;
+}
+
+.framework-selector > button {
+    background-color: white;
+    color: #6670FF;
+    border: 1px solid #6670FF;
+    padding: 5px;
+}
+
+.framework-selector > button.selected{
+    background-color: #6670FF;
+    color: white;
+    border: 1px solid #6670FF;
+    padding: 5px;
+}
 
 /* The literal code blocks */
 .rst-content tt.literal, .rst-content tt.literal, .rst-content code.literal {
@@ -44,11 +64,11 @@ huggingface.css
 /* The text items on the toc tree */
 .wy-menu-vertical a {
     color: #FFFFDD;
-    font-family: Calibre-Light;
+    font-family: Calibre-Light, sans-serif;
 }
 .wy-menu-vertical header, .wy-menu-vertical p.caption{
     color: white;
-    font-family: Calibre-Light;
+    font-family: Calibre-Light, sans-serif;
 }
 
 /* The color inside the selected toc tree block */
@@ -85,7 +105,7 @@ a {
     border-right: solid 2px #FB8D68;
     border-left: solid 2px #FB8D68;
     color: #FB8D68;
-    font-family: Calibre-Light;
+    font-family: Calibre-Light, sans-serif;
     border-top: none;
     font-style: normal !important;
 }
@@ -136,14 +156,14 @@ a {
 
 /* class and method names in doc */
 .rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) tt.descclassname, .rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) code.descname, .rst-content dl:not(.docutils) tt.descclassname, .rst-content dl:not(.docutils) code.descclassname{
-    font-family: Calibre;
+    font-family: Calibre, sans-serif;
     font-size: 20px !important;
 }
 
 /* class name in doc*/
 .rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) code.descname{
     margin-right: 10px;
-    font-family: Calibre-Medium;
+    font-family: Calibre-Medium, sans-serif;
 }
 
 /* Method and class parameters */
@@ -160,17 +180,17 @@ a {
 
 /* FONTS */
 body{
-    font-family: Calibre;
+    font-family: Calibre, sans-serif;
     font-size: 16px;
 }
 
 h1 {
-    font-family: Calibre-Thin;
+    font-family: Calibre-Thin, sans-serif;
     font-size: 70px;
 }
 
 h2, .rst-content .toctree-wrapper p.caption, h3, h4, h5, h6, legend{
-    font-family: Calibre-Medium;
+    font-family: Calibre-Medium, sans-serif;
 }
 
 @font-face {
@@ -197,3 +217,40 @@ h2, .rst-content .toctree-wrapper p.caption, h3, h4, h5, h6, legend{
     font-weight:400;
 }
 
+
+/**
+ * Nav Links to other parts of huggingface.co
+ */
+ div.menu {
+    position: absolute;
+    top: 0;
+    right: 0;
+    padding-top: 20px;
+    padding-right: 20px;
+    z-index: 1000;
+}
+div.menu a {
+    font-size: 14px;
+    letter-spacing: 0.3px;
+    text-transform: uppercase;
+    color: white;
+    -webkit-font-smoothing: antialiased;
+    background: linear-gradient(0deg, #6671ffb8, #9a66ffb8 50%);
+    padding: 10px 16px 6px 16px;
+    border-radius: 3px;
+    margin-left: 12px;
+    position: relative;
+}
+div.menu a:active {
+    top: 1px;
+}
+@media (min-width: 768px) and (max-width: 1750px) {
+    .wy-breadcrumbs {
+        margin-top: 32px;
+    }
+}
+@media (max-width: 768px) {
+    div.menu {
+        display: none;
+    }
+}

docs/source/benchmarks.md

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

docs/source/bertology.rst

@@ -15,4 +15,4 @@ In order to help this new field develop, we have included a few additional featu
 * accessing all the attention weights for each head of BERT/GPT/GPT-2,
 * retrieving heads output values and gradients to be able to compute head importance score and prune head as explained in https://arxiv.org/abs/1905.10650.
 
-To help you understand and use these features, we have added a specific example script: `bertology.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_bertology.py>`_ while extract information and prune a model pre-trained on GLUE.
+To help you understand and use these features, we have added a specific example script: `bertology.py <https://github.com/huggingface/transformers/blob/master/examples/run_bertology.py>`_ while extract information and prune a model pre-trained on GLUE.

docs/source/conf.py

@@ -14,19 +14,19 @@
 #
 import os
 import sys
-sys.path.insert(0, os.path.abspath('../..'))
+sys.path.insert(0, os.path.abspath('../../src'))
 
 
 # -- Project information -----------------------------------------------------
 
-project = u'pytorch-transformers'
-copyright = u'2019, huggingface'
+project = u'transformers'
+copyright = u'2020, huggingface'
 author = u'huggingface'
 
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'1.0.0'
+release = u'2.8.0'
 
 
 # -- General configuration ---------------------------------------------------
@@ -43,7 +43,8 @@ extensions = [
     'sphinx.ext.coverage',
     'sphinx.ext.napoleon',
     'recommonmark',
-    'sphinx.ext.viewcode'
+    'sphinx.ext.viewcode',
+    'sphinx_markdown_tables'
 ]
 
 # Add any paths that contain templates here, relative to this directory.
@@ -104,11 +105,17 @@ html_static_path = ['_static']
 #
 # html_sidebars = {}
 
+# This must be the name of an image file (path relative to the configuration 
+# directory) that is the favicon of the docs. Modern browsers use this as 
+# the icon for tabs, windows and bookmarks. It should be a Windows-style 
+# icon file (.ico).
+html_favicon = 'favicon.ico'
+
 
 # -- Options for HTMLHelp output ---------------------------------------------
 
 # Output file base name for HTML help builder.
-htmlhelp_basename = 'pytorch-transformersdoc'
+htmlhelp_basename = 'transformersdoc'
 
 
 # -- Options for LaTeX output ------------------------------------------------
@@ -135,7 +142,7 @@ latex_elements = {
 # (source start file, target name, title,
 #  author, documentclass [howto, manual, or own class]).
 latex_documents = [
-    (master_doc, 'pytorch-transformers.tex', u'pytorch-transformers Documentation',
+    (master_doc, 'transformers.tex', u'transformers Documentation',
      u'huggingface', 'manual'),
 ]
 
@@ -145,7 +152,7 @@ latex_documents = [
 # One entry per manual page. List of tuples
 # (source start file, name, description, authors, manual section).
 man_pages = [
-    (master_doc, 'pytorch-transformers', u'pytorch-transformers Documentation',
+    (master_doc, 'transformers', u'transformers Documentation',
      [author], 1)
 ]
 
@@ -156,8 +163,8 @@ man_pages = [
 # (source start file, target name, title, author,
 #  dir menu entry, description, category)
 texinfo_documents = [
-    (master_doc, 'pytorch-transformers', u'pytorch-transformers Documentation',
-     author, 'pytorch-transformers', 'One line description of project.',
+    (master_doc, 'transformers', u'transformers Documentation',
+     author, 'transformers', 'One line description of project.',
      'Miscellaneous'),
 ]
 

docs/source/converting_tensorflow_models.rst

@@ -3,10 +3,16 @@ Converting Tensorflow Checkpoints
 
 A command-line interface is provided to convert original Bert/GPT/GPT-2/Transformer-XL/XLNet/XLM checkpoints in models than be loaded using the ``from_pretrained`` methods of the library.
 
+.. note::
+    Since 2.3.0 the conversion script is now part of the transformers CLI (**transformers-cli**)
+    available in any transformers >= 2.3.0 installation.
+
+    The documentation below reflects the **transformers-cli convert** command format.
+
 BERT
 ^^^^
 
-You can convert any TensorFlow checkpoint for BERT (in particular `the pre-trained models released by Google <https://github.com/google-research/bert#pre-trained-models>`_\ ) in a PyTorch save file by using the `convert_tf_checkpoint_to_pytorch.py <https://github.com/huggingface/pytorch-transformers/blob/master/pytorch_transformers/convert_tf_checkpoint_to_pytorch.py>`_ script.
+You can convert any TensorFlow checkpoint for BERT (in particular `the pre-trained models released by Google <https://github.com/google-research/bert#pre-trained-models>`_\ ) in a PyTorch save file by using the `convert_tf_checkpoint_to_pytorch.py <https://github.com/huggingface/transformers/blob/master/transformers/convert_tf_checkpoint_to_pytorch.py>`_ script.
 
 This CLI takes as input a TensorFlow checkpoint (three files starting with ``bert_model.ckpt``\ ) and the associated configuration file (\ ``bert_config.json``\ ), and creates a PyTorch model for this configuration, loads the weights from the TensorFlow checkpoint in the PyTorch model and saves the resulting model in a standard PyTorch save file that can be imported using ``torch.load()`` (see examples in `run_bert_extract_features.py <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/run_bert_extract_features.py>`_\ , `run_bert_classifier.py <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/run_bert_classifier.py>`_ and `run_bert_squad.py <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/run_bert_squad.py>`_\ ).
 
@@ -20,10 +26,10 @@ Here is an example of the conversion process for a pre-trained ``BERT-Base Uncas
 
    export BERT_BASE_DIR=/path/to/bert/uncased_L-12_H-768_A-12
 
-   pytorch_transformers bert \
-     $BERT_BASE_DIR/bert_model.ckpt \
-     $BERT_BASE_DIR/bert_config.json \
-     $BERT_BASE_DIR/pytorch_model.bin
+   transformers-cli convert --model_type bert \
+     --tf_checkpoint $BERT_BASE_DIR/bert_model.ckpt \
+     --config $BERT_BASE_DIR/bert_config.json \
+     --pytorch_dump_output $BERT_BASE_DIR/pytorch_model.bin
 
 You can download Google's pre-trained models for the conversion `here <https://github.com/google-research/bert#pre-trained-models>`__.
 
@@ -36,10 +42,12 @@ Here is an example of the conversion process for a pre-trained OpenAI GPT model,
 
    export OPENAI_GPT_CHECKPOINT_FOLDER_PATH=/path/to/openai/pretrained/numpy/weights
 
-   pytorch_transformers gpt \
-     $OPENAI_GPT_CHECKPOINT_FOLDER_PATH \
-     $PYTORCH_DUMP_OUTPUT \
-     [OPENAI_GPT_CONFIG]
+   transformers-cli convert --model_type gpt \
+     --tf_checkpoint $OPENAI_GPT_CHECKPOINT_FOLDER_PATH \
+     --pytorch_dump_output $PYTORCH_DUMP_OUTPUT \
+     [--config OPENAI_GPT_CONFIG] \
+     [--finetuning_task_name OPENAI_GPT_FINETUNED_TASK] \
+
 
 OpenAI GPT-2
 ^^^^^^^^^^^^
@@ -50,10 +58,11 @@ Here is an example of the conversion process for a pre-trained OpenAI GPT-2 mode
 
    export OPENAI_GPT2_CHECKPOINT_PATH=/path/to/gpt2/pretrained/weights
 
-   pytorch_transformers gpt2 \
-     $OPENAI_GPT2_CHECKPOINT_PATH \
-     $PYTORCH_DUMP_OUTPUT \
-     [OPENAI_GPT2_CONFIG]
+   transformers-cli convert --model_type gpt2 \
+     --tf_checkpoint $OPENAI_GPT2_CHECKPOINT_PATH \
+     --pytorch_dump_output $PYTORCH_DUMP_OUTPUT \
+     [--config OPENAI_GPT2_CONFIG] \
+     [--finetuning_task_name OPENAI_GPT2_FINETUNED_TASK]
 
 Transformer-XL
 ^^^^^^^^^^^^^^
@@ -64,27 +73,28 @@ Here is an example of the conversion process for a pre-trained Transformer-XL mo
 
    export TRANSFO_XL_CHECKPOINT_FOLDER_PATH=/path/to/transfo/xl/checkpoint
 
-   pytorch_transformers transfo_xl \
-     $TRANSFO_XL_CHECKPOINT_FOLDER_PATH \
-     $PYTORCH_DUMP_OUTPUT \
-     [TRANSFO_XL_CONFIG]
+   transformers-cli convert --model_type transfo_xl \
+     --tf_checkpoint $TRANSFO_XL_CHECKPOINT_FOLDER_PATH \
+     --pytorch_dump_output $PYTORCH_DUMP_OUTPUT \
+     [--config TRANSFO_XL_CONFIG] \
+     [--finetuning_task_name TRANSFO_XL_FINETUNED_TASK]
 
 
 XLNet
 ^^^^^
 
-Here is an example of the conversion process for a pre-trained XLNet model, fine-tuned on STS-B using the TensorFlow script:
+Here is an example of the conversion process for a pre-trained XLNet model:
 
 .. code-block:: shell
 
    export TRANSFO_XL_CHECKPOINT_PATH=/path/to/xlnet/checkpoint
    export TRANSFO_XL_CONFIG_PATH=/path/to/xlnet/config
 
-   pytorch_transformers xlnet \
-     $TRANSFO_XL_CHECKPOINT_PATH \
-     $TRANSFO_XL_CONFIG_PATH \
-     $PYTORCH_DUMP_OUTPUT \
-     STS-B \
+   transformers-cli convert --model_type xlnet \
+     --tf_checkpoint $TRANSFO_XL_CHECKPOINT_PATH \
+     --config $TRANSFO_XL_CONFIG_PATH \
+     --pytorch_dump_output $PYTORCH_DUMP_OUTPUT \
+     [--finetuning_task_name XLNET_FINETUNED_TASK] \
 
 
 XLM
@@ -96,6 +106,8 @@ Here is an example of the conversion process for a pre-trained XLM model:
 
    export XLM_CHECKPOINT_PATH=/path/to/xlm/checkpoint
 
-   pytorch_transformers xlm \
-     $XLM_CHECKPOINT_PATH \
-     $PYTORCH_DUMP_OUTPUT \
+   transformers-cli convert --model_type xlm \
+     --tf_checkpoint $XLM_CHECKPOINT_PATH \
+     --pytorch_dump_output $PYTORCH_DUMP_OUTPUT
+    [--config XML_CONFIG] \
+    [--finetuning_task_name XML_FINETUNED_TASK]

docs/source/examples.md

@@ -0,0 +1 @@
+../../examples/README.md

docs/source/examples.rst

@@ -1,686 +0,0 @@
-examples.rst
-
-Examples
-================================================
-
-.. list-table::
-   :header-rows: 1
-
-   * - Sub-section
-     - Description
-   * - `Training large models: introduction, tools and examples <#introduction>`_
-     - How to use gradient-accumulation, multi-gpu training, distributed training, optimize on CPU and 16-bits training to train Bert models
-   * - `Fine-tuning with BERT: running the examples <#fine-tuning-bert-examples>`_
-     - Running the examples in `examples <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples>`_\ : ``extract_classif.py``\ , ``run_bert_classifier.py``\ , ``run_bert_squad.py`` and ``run_lm_finetuning.py``
-   * - `Fine-tuning with OpenAI GPT, Transformer-XL, GPT-2 as well as BERT and RoBERTa <#fine-tuning>`_
-     - Running the examples in `examples <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples>`_\ : ``run_openai_gpt.py``\ , ``run_transfo_xl.py``, ``run_gpt2.py`` and ``run_lm_finetuning.py``
-   * - `Fine-tuning BERT-large on GPUs <#fine-tuning-bert-large>`_
-     - How to fine tune ``BERT large``
-
-
-.. _introduction:
-
-Training large models: introduction, tools and examples
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-BERT-base and BERT-large are respectively 110M and 340M parameters models and it can be difficult to fine-tune them on a single GPU with the recommended batch size for good performance (in most case a batch size of 32).
-
-To help with fine-tuning these models, we have included several techniques that you can activate in the fine-tuning scripts `run_bert_classifier.py <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/run_bert_classifier.py>`_ and `run_bert_squad.py <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/run_bert_squad.py>`_\ : gradient-accumulation, multi-gpu training, distributed training and 16-bits training . For more details on how to use these techniques you can read `the tips on training large batches in PyTorch <https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255>`_ that I published earlier this year.
-
-Here is how to use these techniques in our scripts:
-
-
-* **Gradient Accumulation**\ : Gradient accumulation can be used by supplying a integer greater than 1 to the ``--gradient_accumulation_steps`` argument. The batch at each step will be divided by this integer and gradient will be accumulated over ``gradient_accumulation_steps`` steps.
-* **Multi-GPU**\ : Multi-GPU is automatically activated when several GPUs are detected and the batches are splitted over the GPUs.
-* **Distributed training**\ : Distributed training can be activated by supplying an integer greater or equal to 0 to the ``--local_rank`` argument (see below).
-* **16-bits training**\ : 16-bits training, also called mixed-precision training, can reduce the memory requirement of your model on the GPU by using half-precision training, basically allowing to double the batch size. If you have a recent GPU (starting from NVIDIA Volta architecture) you should see no decrease in speed. A good introduction to Mixed precision training can be found `here <https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/>`__ and a full documentation is `here <https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html>`__. In our scripts, this option can be activated by setting the ``--fp16`` flag and you can play with loss scaling using the ``--loss_scale`` flag (see the previously linked documentation for details on loss scaling). The loss scale can be zero in which case the scale is dynamically adjusted or a positive power of two in which case the scaling is static.
-
-To use 16-bits training and distributed training, you need to install NVIDIA's apex extension `as detailed here <https://github.com/nvidia/apex>`__. You will find more information regarding the internals of ``apex`` and how to use ``apex`` in `the doc and the associated repository <https://github.com/nvidia/apex>`_. The results of the tests performed on pytorch-BERT by the NVIDIA team (and my trials at reproducing them) can be consulted in `the relevant PR of the present repository <https://github.com/huggingface/pytorch-pretrained-BERT/pull/116>`_.
-
-Note: To use *Distributed Training*\ , you will need to run one training script on each of your machines. This can be done for example by running the following command on each server (see `the above mentioned blog post <https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255>`_\ ) for more details):
-
-.. code-block:: bash
-
-    python -m torch.distributed.launch \
-        --nproc_per_node=4 \
-        --nnodes=2 \
-        --node_rank=$THIS_MACHINE_INDEX \
-        --master_addr="192.168.1.1" \
-        --master_port=1234 run_bert_classifier.py \
-        (--arg1 --arg2 --arg3 and all other arguments of the run_classifier script)
-
-Where ``$THIS_MACHINE_INDEX`` is an sequential index assigned to each of your machine (0, 1, 2...) and the machine with rank 0 has an IP address ``192.168.1.1`` and an open port ``1234``.
-
-.. _fine-tuning-bert-examples:
-
-Fine-tuning with BERT: running the examples
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-We showcase several fine-tuning examples based on (and extended from) `the original implementation <https://github.com/google-research/bert/>`_\ :
-
-
-* a *sequence-level classifier* on nine different GLUE tasks,
-* a *token-level classifier* on the question answering dataset SQuAD, and
-* a *sequence-level multiple-choice classifier* on the SWAG classification corpus.
-* a *BERT language model* on another target corpus
-
-GLUE results on dev set
-~~~~~~~~~~~~~~~~~~~~~~~
-
-We get the following results on the dev set of GLUE benchmark with an uncased BERT base
-model (`bert-base-uncased`). All experiments ran on 8 V100 GPUs with a total train batch size of 24. Some of 
-these tasks have a small dataset and training can lead to high variance in the results between different runs.
-We report the median on 5 runs (with different seeds) for each of the metrics.
-
-.. list-table::
-   :header-rows: 1
-
-   * - Task
-     - Metric
-     - Result
-   * - CoLA
-     - Matthew's corr.
-     - 55.75
-   * - SST-2
-     - accuracy
-     - 92.09
-   * - MRPC
-     - F1/accuracy
-     - 90.48/86.27
-   * - STS-B
-     - Pearson/Spearman corr.
-     - 89.03/88.64
-   * - QQP
-     - accuracy/F1
-     - 90.92/87.72
-   * - MNLI
-     - matched acc./mismatched acc.
-     - 83.74/84.06
-   * - QNLI
-     - accuracy
-     - 91.07
-   * - RTE
-     - accuracy
-     - 68.59
-   * - WNLI
-     - accuracy
-     - 43.66
-
-
-Some of these results are significantly different from the ones reported on the test set
-of GLUE benchmark on the website. For QQP and WNLI, please refer to `FAQ #12 <https://gluebenchmark.com/faq>`_ on the webite.
-
-Before running anyone of these GLUE tasks you should download the
-`GLUE data <https://gluebenchmark.com/tasks>`_ by running
-`this script <https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e>`_
-and unpack it to some directory ``$GLUE_DIR``.
-
-.. code-block:: shell
-
-   export GLUE_DIR=/path/to/glue
-   export TASK_NAME=MRPC
-
-   python run_bert_classifier.py \
-     --task_name $TASK_NAME \
-     --do_train \
-     --do_eval \
-     --do_lower_case \
-     --data_dir $GLUE_DIR/$TASK_NAME \
-     --bert_model bert-base-uncased \
-     --max_seq_length 128 \
-     --train_batch_size 32 \
-     --learning_rate 2e-5 \
-     --num_train_epochs 3.0 \
-     --output_dir /tmp/$TASK_NAME/
-
-where task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI.
-
-The dev set results will be present within the text file 'eval_results.txt' in the specified output_dir. In case of MNLI, since there are two separate dev sets, matched and mismatched, there will be a separate output folder called '/tmp/MNLI-MM/' in addition to '/tmp/MNLI/'.
-
-The code has not been tested with half-precision training with apex on any GLUE task apart from MRPC, MNLI, CoLA, SST-2. The following section provides details on how to run half-precision training with MRPC. With that being said, there shouldn't be any issues in running half-precision training with the remaining GLUE tasks as well, since the data processor for each task inherits from the base class DataProcessor.
-
-MRPC
-~~~~
-
-This example code fine-tunes BERT on the Microsoft Research Paraphrase
-Corpus (MRPC) corpus and runs in less than 10 minutes on a single K-80 and in 27 seconds (!) on single tesla V100 16GB with apex installed.
-
-Before running this example you should download the
-`GLUE data <https://gluebenchmark.com/tasks>`_ by running
-`this script <https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e>`_
-and unpack it to some directory ``$GLUE_DIR``.
-
-.. code-block:: shell
-
-   export GLUE_DIR=/path/to/glue
-
-   python run_bert_classifier.py \
-     --task_name MRPC \
-     --do_train \
-     --do_eval \
-     --do_lower_case \
-     --data_dir $GLUE_DIR/MRPC/ \
-     --bert_model bert-base-uncased \
-     --max_seq_length 128 \
-     --train_batch_size 32 \
-     --learning_rate 2e-5 \
-     --num_train_epochs 3.0 \
-     --output_dir /tmp/mrpc_output/
-
-Our test ran on a few seeds with `the original implementation hyper-parameters <https://github.com/google-research/bert#sentence-and-sentence-pair-classification-tasks>`__ gave evaluation results between 84% and 88%.
-
-**Fast run with apex and 16 bit precision: fine-tuning on MRPC in 27 seconds!**
-First install apex as indicated `here <https://github.com/NVIDIA/apex>`__.
-Then run
-
-.. code-block:: shell
-
-   export GLUE_DIR=/path/to/glue
-
-   python run_bert_classifier.py \
-     --task_name MRPC \
-     --do_train \
-     --do_eval \
-     --do_lower_case \
-     --data_dir $GLUE_DIR/MRPC/ \
-     --bert_model bert-base-uncased \
-     --max_seq_length 128 \
-     --train_batch_size 32 \
-     --learning_rate 2e-5 \
-     --num_train_epochs 3.0 \
-     --output_dir /tmp/mrpc_output/ \
-     --fp16
-
-**Distributed training**
-Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking model to reach a F1 > 92 on MRPC:
-
-.. code-block:: bash
-
-    python -m torch.distributed.launch \
-        --nproc_per_node 8 run_bert_classifier.py \
-        --bert_model bert-large-uncased-whole-word-masking \
-        --task_name MRPC \
-        --do_train \
-        --do_eval \
-        --do_lower_case \
-        --data_dir $GLUE_DIR/MRPC/ \
-        --max_seq_length 128 \
-        --train_batch_size 8 \
-        --learning_rate 2e-5 \
-        --num_train_epochs 3.0 \
-         --output_dir /tmp/mrpc_output/
-
-Training with these hyper-parameters gave us the following results:
-
-.. code-block:: bash
-
-     acc = 0.8823529411764706
-     acc_and_f1 = 0.901702786377709
-     eval_loss = 0.3418912578906332
-     f1 = 0.9210526315789473
-     global_step = 174
-     loss = 0.07231863956341798
-
-Here is an example on MNLI:
-
-.. code-block:: bash
-
-    python -m torch.distributed.launch \
-        --nproc_per_node 8 run_bert_classifier.py \
-        --bert_model bert-large-uncased-whole-word-masking \
-        --task_name mnli \
-        --do_train \
-        --do_eval \
-        --do_lower_case \
-        --data_dir /datadrive/bert_data/glue_data//MNLI/ \
-        --max_seq_length 128 \
-        --train_batch_size 8 \
-        --learning_rate 2e-5 \
-        --num_train_epochs 3.0 \
-        --output_dir ../models/wwm-uncased-finetuned-mnli/ \
-        --overwrite_output_dir
-
-.. code-block:: bash
-
-   ***** Eval results *****
-     acc = 0.8679706601466992
-     eval_loss = 0.4911287787382479
-     global_step = 18408
-     loss = 0.04755385363816904
-
-   ***** Eval results *****
-     acc = 0.8747965825874695
-     eval_loss = 0.45516540421714036
-     global_step = 18408
-     loss = 0.04755385363816904
-
-This is the example of the ``bert-large-uncased-whole-word-masking-finetuned-mnli`` model
-
-SQuAD
-~~~~~
-
-This example code fine-tunes BERT on the SQuAD dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) on a single tesla V100 16GB.
-
-The data for SQuAD can be downloaded with the following links and should be saved in a ``$SQUAD_DIR`` directory.
-
-
-* `train-v1.1.json <https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json>`_
-* `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>`_
-
-.. code-block:: shell
-
-   export SQUAD_DIR=/path/to/SQUAD
-
-   python run_bert_squad.py \
-     --bert_model bert-base-uncased \
-     --do_train \
-     --do_predict \
-     --do_lower_case \
-     --train_file $SQUAD_DIR/train-v1.1.json \
-     --predict_file $SQUAD_DIR/dev-v1.1.json \
-     --train_batch_size 12 \
-     --learning_rate 3e-5 \
-     --num_train_epochs 2.0 \
-     --max_seq_length 384 \
-     --doc_stride 128 \
-     --output_dir /tmp/debug_squad/
-
-Training with the previous hyper-parameters gave us the following results:
-
-.. code-block:: bash
-
-   python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json /tmp/debug_squad/predictions.json
-   {"f1": 88.52381567990474, "exact_match": 81.22043519394512}
-
-**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:
-
-.. code-block:: bash
-
-   python -m torch.distributed.launch --nproc_per_node=8 \
-    run_bert_squad.py \
-    --bert_model bert-large-uncased-whole-word-masking  \
-    --do_train \
-    --do_predict \
-    --do_lower_case \
-    --train_file $SQUAD_DIR/train-v1.1.json \
-    --predict_file $SQUAD_DIR/dev-v1.1.json \
-    --learning_rate 3e-5 \
-    --num_train_epochs 2 \
-    --max_seq_length 384 \
-    --doc_stride 128 \
-    --output_dir ../models/wwm_uncased_finetuned_squad/ \
-    --train_batch_size 24 \
-    --gradient_accumulation_steps 12
-
-Training with these hyper-parameters gave us the following results:
-
-.. code-block:: bash
-
-   python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json ../models/wwm_uncased_finetuned_squad/predictions.json
-   {"exact_match": 86.91579943235573, "f1": 93.1532499015869}
-
-This is the model provided as ``bert-large-uncased-whole-word-masking-finetuned-squad``.
-
-And here is the model provided as ``bert-large-cased-whole-word-masking-finetuned-squad``\ :
-
-.. code-block:: bash
-
-    python -m torch.distributed.launch --nproc_per_node=8  run_bert_squad.py \
-        --bert_model bert-large-cased-whole-word-masking \
-        --do_train \
-        --do_predict \
-        --do_lower_case \
-        --train_file $SQUAD_DIR/train-v1.1.json \
-        --predict_file $SQUAD_DIR/dev-v1.1.json \
-        --learning_rate 3e-5 \
-        --num_train_epochs 2 \
-        --max_seq_length 384 \
-        --doc_stride 128 \
-        --output_dir ../models/wwm_cased_finetuned_squad/ \
-        --train_batch_size 24 \
-        --gradient_accumulation_steps 12
-
-Training with these hyper-parameters gave us the following results:
-
-.. code-block:: bash
-
-   python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json ../models/wwm_uncased_finetuned_squad/predictions.json
-   {"exact_match": 84.18164616840113, "f1": 91.58645594850135}
-
-SWAG
-~~~~
-
-The data for SWAG can be downloaded by cloning the following `repository <https://github.com/rowanz/swagaf>`_
-
-.. code-block:: shell
-
-   export SWAG_DIR=/path/to/SWAG
-
-   python run_bert_swag.py \
-     --bert_model bert-base-uncased \
-     --do_train \
-     --do_lower_case \
-     --do_eval \
-     --data_dir $SWAG_DIR/data \
-     --train_batch_size 16 \
-     --learning_rate 2e-5 \
-     --num_train_epochs 3.0 \
-     --max_seq_length 80 \
-     --output_dir /tmp/swag_output/ \
-     --gradient_accumulation_steps 4
-
-Training with the previous hyper-parameters on a single GPU gave us the following results:
-
-.. code-block::
-
-   eval_accuracy = 0.8062081375587323
-   eval_loss = 0.5966546792367169
-   global_step = 13788
-   loss = 0.06423990014260186
-
-LM Fine-tuning
-~~~~~~~~~~~~~~
-
-The data should be a text file in the same format as `sample_text.txt <./pytorch_transformers/tests/fixtures/sample_text.txt/sample_text.txt>`_  (one sentence per line, docs separated by empty line).
-You can download an `exemplary training corpus <https://ext-bert-sample.obs.eu-de.otc.t-systems.com/small_wiki_sentence_corpus.txt>`_ generated from wikipedia articles and split into ~500k sentences with spaCy.
-Training one epoch on this corpus takes about 1:20h on 4 x NVIDIA Tesla P100 with ``train_batch_size=200`` and ``max_seq_length=128``\ :
-
-Thank to the work of @Rocketknight1 and @tholor there are now **several scripts** that can be used to fine-tune BERT using the pretraining objective (combination of masked-language modeling and next sentence prediction loss). These scripts are detailed in the `README <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/lm_finetuning/README.md>`_ of the `examples/lm_finetuning/ <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/lm_finetuning/>`_ folder.
-
-.. _fine-tuning:
-
-OpenAI GPT, Transformer-XL and GPT-2: running the examples
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-We provide three examples of scripts for OpenAI GPT, Transformer-XL, OpenAI GPT-2, BERT and RoBERTa based on (and extended from) the respective original implementations:
-
-
-* fine-tuning OpenAI GPT on the ROCStories dataset
-* evaluating Transformer-XL on Wikitext 103
-* unconditional and conditional generation from a pre-trained OpenAI GPT-2 model
-* fine-tuning GPT/GPT-2 on a causal language modeling task and BERT/RoBERTa on a masked language modeling task
-
-Fine-tuning OpenAI GPT on the RocStories dataset
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-This example code fine-tunes OpenAI GPT on the RocStories dataset.
-
-Before running this example you should download the
-`RocStories dataset <https://github.com/snigdhac/StoryComprehension_EMNLP/tree/master/Dataset/RoCStories>`_ and unpack it to some directory ``$ROC_STORIES_DIR``.
-
-.. code-block:: shell
-
-   export ROC_STORIES_DIR=/path/to/RocStories
-
-   python run_openai_gpt.py \
-     --model_name openai-gpt \
-     --do_train \
-     --do_eval \
-     --train_dataset $ROC_STORIES_DIR/cloze_test_val__spring2016\ -\ cloze_test_ALL_val.csv \
-     --eval_dataset $ROC_STORIES_DIR/cloze_test_test__spring2016\ -\ cloze_test_ALL_test.csv \
-     --output_dir ../log \
-     --train_batch_size 16 \
-
-This command runs in about 10 min on a single K-80 an gives an evaluation accuracy of about 87.7% (the authors report a median accuracy with the TensorFlow code of 85.8% and the OpenAI GPT paper reports a best single run accuracy of 86.5%).
-
-Evaluating the pre-trained Transformer-XL on the WikiText 103 dataset
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-This example code evaluate the pre-trained Transformer-XL on the WikiText 103 dataset.
-This command will download a pre-processed version of the WikiText 103 dataset in which the vocabulary has been computed.
-
-.. code-block:: shell
-
-   python run_transfo_xl.py --work_dir ../log
-
-This command runs in about 1 min on a V100 and gives an evaluation perplexity of 18.22 on WikiText-103 (the authors report a perplexity of about 18.3 on this dataset with the TensorFlow code).
-
-Unconditional and conditional generation from OpenAI's GPT-2 model
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-This example code is identical to the original unconditional and conditional generation codes.
-
-Conditional generation:
-
-.. code-block:: shell
-
-   python run_gpt2.py
-
-Unconditional generation:
-
-.. code-block:: shell
-
-   python run_gpt2.py --unconditional
-
-The same option as in the original scripts are provided, please refer to the code of the example and the original repository of OpenAI.
-
-
-Causal LM fine-tuning on GPT/GPT-2, Masked LM fine-tuning on BERT/RoBERTa
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-Before running the following examples you should download the `WikiText-2 dataset <https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/>`__ and unpack it to some directory `$WIKITEXT_2_DATASET`
-The following results were obtained using the `raw` WikiText-2 (no tokens were replaced before the tokenization).
-
-This example fine-tunes GPT-2 on the WikiText-2 dataset. The loss function is a causal language modeling loss (perplexity).
-
-.. code-block:: bash
-
-
-    export WIKITEXT_2_DATASET=/path/to/wikitext_dataset
-
-    python run_lm_finetuning.py
-        --output_dir=output
-        --model_type=gpt2
-        --model_name_or_path=gpt2
-        --do_train
-        --train_data_file=$WIKITEXT_2_DATASET/wiki.train.raw
-        --do_eval
-        --eval_data_file=$WIKITEXT_2_DATASET/wiki.test.raw
-
-This takes about half an hour to train on a single K80 GPU and about one minute for the evaluation to run.
-It reaches a score of about 20 perplexity once fine-tuned on the dataset.
-
-This example fine-tunes RoBERTa on the WikiText-2 dataset. The loss function is a masked language modeling loss (masked perplexity).
-The `--mlm` flag is necessary to fine-tune BERT/RoBERTa on masked language modeling.
-
-.. code-block:: bash
-
-
-    export WIKITEXT_2_DATASET=/path/to/wikitext_dataset
-
-    python run_lm_finetuning.py
-        --output_dir=output
-        --model_type=roberta
-        --model_name_or_path=roberta-base
-        --do_train
-        --train_data_file=$WIKITEXT_2_DATASET/wiki.train.raw
-        --do_eval
-        --eval_data_file=$WIKITEXT_2_DATASET/wiki.test.raw
-        --mlm
-
-.. _fine-tuning-BERT-large:
-
-Fine-tuning BERT-large on GPUs
-------------------------------
-
-The options we list above allow to fine-tune BERT-large rather easily on GPU(s) instead of the TPU used by the original implementation.
-
-For example, fine-tuning BERT-large on SQuAD can be done on a server with 4 k-80 (these are pretty old now) in 18 hours. Our results are similar to the TensorFlow implementation results (actually slightly higher):
-
-.. code-block:: bash
-
-   {"exact_match": 84.56953642384106, "f1": 91.04028647786927}
-
-To get these results we used a combination of:
-
-
-* multi-GPU training (automatically activated on a multi-GPU server),
-* 2 steps of gradient accumulation and
-* perform the optimization step on CPU to store Adam's averages in RAM.
-
-Here is the full list of hyper-parameters for this run:
-
-.. code-block:: bash
-
-   export SQUAD_DIR=/path/to/SQUAD
-
-   python ./run_bert_squad.py \
-     --bert_model bert-large-uncased \
-     --do_train \
-     --do_predict \
-     --do_lower_case \
-     --train_file $SQUAD_DIR/train-v1.1.json \
-     --predict_file $SQUAD_DIR/dev-v1.1.json \
-     --learning_rate 3e-5 \
-     --num_train_epochs 2 \
-     --max_seq_length 384 \
-     --doc_stride 128 \
-     --output_dir /tmp/debug_squad/ \
-     --train_batch_size 24 \
-     --gradient_accumulation_steps 2
-
-If you have a recent GPU (starting from NVIDIA Volta series), you should try **16-bit fine-tuning** (FP16).
-
-Here is an example of hyper-parameters for a FP16 run we tried:
-
-.. code-block:: bash
-
-   export SQUAD_DIR=/path/to/SQUAD
-
-   python ./run_bert_squad.py \
-     --bert_model bert-large-uncased \
-     --do_train \
-     --do_predict \
-     --do_lower_case \
-     --train_file $SQUAD_DIR/train-v1.1.json \
-     --predict_file $SQUAD_DIR/dev-v1.1.json \
-     --learning_rate 3e-5 \
-     --num_train_epochs 2 \
-     --max_seq_length 384 \
-     --doc_stride 128 \
-     --output_dir /tmp/debug_squad/ \
-     --train_batch_size 24 \
-     --fp16 \
-     --loss_scale 128
-
-The results were similar to the above FP32 results (actually slightly higher):
-
-.. code-block:: bash
-
-   {"exact_match": 84.65468306527909, "f1": 91.238669287002}
-
-Here is an example with the recent ``bert-large-uncased-whole-word-masking``\ :
-
-.. code-block:: bash
-
-   python -m torch.distributed.launch --nproc_per_node=8 \
-     run_bert_squad.py \
-     --bert_model bert-large-uncased-whole-word-masking \
-     --do_train \
-     --do_predict \
-     --do_lower_case \
-     --train_file $SQUAD_DIR/train-v1.1.json \
-     --predict_file $SQUAD_DIR/dev-v1.1.json \
-     --learning_rate 3e-5 \
-     --num_train_epochs 2 \
-     --max_seq_length 384 \
-     --doc_stride 128 \
-     --output_dir /tmp/debug_squad/ \
-     --train_batch_size 24 \
-     --gradient_accumulation_steps 2
-
-Fine-tuning XLNet
------------------
-
-STS-B
-~~~~~
-
-This example code fine-tunes XLNet on the STS-B corpus.
-
-Before running this example you should download the
-`GLUE data <https://gluebenchmark.com/tasks>`_ by running
-`this script <https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e>`_
-and unpack it to some directory ``$GLUE_DIR``.
-
-.. code-block:: shell
-
-   export GLUE_DIR=/path/to/glue
-
-   python run_xlnet_classifier.py \
-    --task_name STS-B \
-    --do_train \
-    --do_eval \
-    --data_dir $GLUE_DIR/STS-B/ \
-    --max_seq_length 128 \
-    --train_batch_size 8 \
-    --gradient_accumulation_steps 1 \
-    --learning_rate 5e-5 \
-    --num_train_epochs 3.0 \
-    --output_dir /tmp/mrpc_output/
-
-Our test ran on a few seeds with `the original implementation hyper-parameters <https://github.com/zihangdai/xlnet#1-sts-b-sentence-pair-relevance-regression-with-gpus>`__ gave evaluation results between 84% and 88%.
-
-**Distributed training**
-Here is an example using distributed training on 8 V100 GPUs to reach XXXX:
-
-.. code-block:: bash
-
-   python -m torch.distributed.launch --nproc_per_node 8 \
-    run_xlnet_classifier.py \
-    --task_name STS-B \
-    --do_train \
-    --do_eval \
-    --data_dir $GLUE_DIR/STS-B/ \
-    --max_seq_length 128 \
-    --train_batch_size 8 \
-    --gradient_accumulation_steps 1 \
-    --learning_rate 5e-5 \
-    --num_train_epochs 3.0 \
-    --output_dir /tmp/mrpc_output/
-
-Training with these hyper-parameters gave us the following results:
-
-.. code-block:: bash
-
-     acc = 0.8823529411764706
-     acc_and_f1 = 0.901702786377709
-     eval_loss = 0.3418912578906332
-     f1 = 0.9210526315789473
-     global_step = 174
-     loss = 0.07231863956341798
-
-Here is an example on MNLI:
-
-.. code-block:: bash
-
-    python -m torch.distributed.launch --nproc_per_node 8 run_bert_classifier.py \
-        --bert_model bert-large-uncased-whole-word-masking \
-        --task_name mnli \
-        --do_train \
-        --do_eval \
-        --data_dir /datadrive/bert_data/glue_data//MNLI/ \
-        --max_seq_length 128 \
-        --train_batch_size 8 \
-        --learning_rate 2e-5 \
-        --num_train_epochs 3.0 \
-        --output_dir ../models/wwm-uncased-finetuned-mnli/ \
-        --overwrite_output_dir
-
-.. code-block:: bash
-
-   ***** Eval results *****
-     acc = 0.8679706601466992
-     eval_loss = 0.4911287787382479
-     global_step = 18408
-     loss = 0.04755385363816904
-
-   ***** Eval results *****
-     acc = 0.8747965825874695
-     eval_loss = 0.45516540421714036
-     global_step = 18408
-     loss = 0.04755385363816904
-
-This is the example of the ``bert-large-uncased-whole-word-masking-finetuned-mnli`` model.

docs/source/glossary.rst

@@ -0,0 +1,145 @@
+Glossary
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Every model is different yet bears similarities with the others. Therefore most models use the same inputs, which are
+detailed here alongside usage examples.
+
+Input IDs
+--------------------------
+
+The input ids are often the only required parameters to be passed to the model as input. *They are token indices,
+numerical representations of tokens building the sequences that will be used as input by the model*.
+
+Each tokenizer works differently but the underlying mechanism remains the same. Here's an example using the BERT
+tokenizer, which is a `WordPiece <https://arxiv.org/pdf/1609.08144.pdf>`__ tokenizer:
+
+::
+
+    from transformers import BertTokenizer
+    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+
+    sequence = "A Titan RTX has 24GB of VRAM"
+
+The tokenizer takes care of splitting the sequence into tokens available in the tokenizer vocabulary.
+
+::
+
+    # Continuation of the previous script
+    tokenized_sequence = tokenizer.tokenize(sequence)
+    assert tokenized_sequence == ['A', 'Titan', 'R', '##T', '##X', 'has', '24', '##GB', 'of', 'V', '##RA', '##M']
+
+These tokens can then be converted into IDs which are understandable by the model. Several methods are available for
+this, the recommended being `encode` or `encode_plus`, which leverage the Rust implementation of
+`huggingface/tokenizers <https://github.com/huggingface/tokenizers>`__ for peak performance.
+
+::
+
+    # Continuation of the previous script
+    encoded_sequence = tokenizer.encode(sequence)
+    assert encoded_sequence == [101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102]
+
+The `encode` and `encode_plus` methods automatically add "special tokens" which are special IDs the model uses.
+
+Attention mask
+--------------------------
+
+The attention mask is an optional argument used when batching sequences together. This argument indicates to the
+model which tokens should be attended to, and which should not.
+
+For example, consider these two sequences:
+
+::
+
+    from transformers import BertTokenizer
+    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+
+    sequence_a = "This is a short sequence."
+    sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."
+
+    encoded_sequence_a = tokenizer.encode(sequence_a)
+    assert len(encoded_sequence_a) == 8
+
+    encoded_sequence_b = tokenizer.encode(sequence_b)
+    assert len(encoded_sequence_b) == 19
+
+These two sequences have different lengths and therefore can't be put together in a same tensor as-is. The first
+sequence needs to be padded up to the length of the second one, or the second one needs to be truncated down to
+the length of the first one.
+
+In the first case, the list of IDs will be extended by the padding indices:
+
+::
+
+    # Continuation of the previous script
+    padded_sequence_a = tokenizer.encode(sequence_a, max_length=19, pad_to_max_length=True)
+
+    assert padded_sequence_a == [101, 1188, 1110, 170, 1603, 4954,  119, 102,    0,    0,    0,    0,    0,    0,    0,    0,   0,   0,   0]
+    assert encoded_sequence_b == [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]
+
+These can then be converted into a tensor in PyTorch or TensorFlow. The attention mask is a binary tensor indicating
+the position of the padded indices so that the model does not attend to them. For the
+:class:`~transformers.BertTokenizer`, :obj:`1` indicate a value that should be attended to while :obj:`0` indicate
+a padded value.
+
+The method :func:`~transformers.PreTrainedTokenizer.encode_plus` may be used to obtain the attention mask directly:
+
+::
+
+    # Continuation of the previous script
+    sequence_a_dict = tokenizer.encode_plus(sequence_a, max_length=19, pad_to_max_length=True)
+
+    assert sequence_a_dict['input_ids'] == [101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
+    assert sequence_a_dict['attention_mask'] == [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
+
+
+Token Type IDs
+--------------------------
+
+Some models' purpose is to do sequence classification or question answering. These require two different sequences to
+be encoded in the same input IDs. They are usually separated by special tokens, such as the classifier and separator
+tokens. For example, the BERT model builds its two sequence input as such:
+
+::
+
+    from transformers import BertTokenizer
+    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+
+    # [CLS] SEQ_A [SEP] SEQ_B [SEP]
+
+    sequence_a = "HuggingFace is based in NYC"
+    sequence_b = "Where is HuggingFace based?"
+
+    encoded_sequence = tokenizer.encode(sequence_a, sequence_b)
+    assert tokenizer.decode(encoded_sequence) == "[CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]"
+
+This is enough for some models to understand where one sequence ends and where another begins. However, other models
+such as BERT have an additional mechanism, which are the segment IDs. The Token Type IDs are a binary mask identifying
+the different sequences in the model.
+
+We can leverage :func:`~transformers.PreTrainedTokenizer.encode_plus` to output the Token Type IDs for us:
+
+::
+
+    # Continuation of the previous script
+    encoded_dict = tokenizer.encode_plus(sequence_a, sequence_b)
+
+    assert encoded_dict['input_ids'] == [101, 20164, 10932, 2271, 7954, 1110, 1359, 1107, 17520, 102, 2777, 1110, 20164, 10932, 2271, 7954, 1359, 136, 102]
+    assert encoded_dict['token_type_ids'] == [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]
+
+The first sequence, the "context" used for the question, has all its tokens represented by :obj:`0`, whereas the
+question has all its tokens represented by :obj:`1`. Some models, like :class:`~transformers.XLNetModel` use an
+additional token represented by a :obj:`2`.
+
+
+Position IDs
+--------------------------
+
+The position IDs are used by the model to identify which token is at which position. Contrary to RNNs that have the
+position of each token embedded within them, transformers are unaware of the position of each token. The position
+IDs are created for this purpose.
+
+They are an optional parameter. If no position IDs are passed to the model, they are automatically created as absolute
+positional embeddings.
+
+Absolute positional embeddings are selected in the range ``[0, config.max_position_embeddings - 1]``. Some models
+use other types of positional embeddings, such as sinusoidal position embeddings or relative position embeddings.

docs/source/index.rst

@@ -1,9 +1,43 @@
-Pytorch-Transformers
+Transformers
 ================================================================================================================================================
 
-PyTorch-Transformers is a library of state-of-the-art pre-trained models for Natural Language Processing (NLP).
+🤗 Transformers (formerly known as `pytorch-transformers` and `pytorch-pretrained-bert`) provides general-purpose architectures
+(BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet...) for Natural Language Understanding (NLU) and Natural Language Generation
+(NLG) with over 32+ pretrained models in 100+ languages and deep interoperability between TensorFlow 2.0 and PyTorch.
 
-The library currently contains PyTorch implementations, pre-trained model weights, usage scripts and conversion utilities for the following models:
+This is the documentation of our repository `transformers <https://github.com/huggingface/transformers>`__.
+
+Features
+---------------------------------------------------
+
+- As easy to use as pytorch-transformers
+- As powerful and concise as Keras
+- High performance on NLU and NLG tasks
+- Low barrier to entry for educators and practitioners
+
+State-of-the-art NLP for everyone:
+
+- Deep learning researchers
+- Hands-on practitioners
+- AI/ML/NLP teachers and educators
+
+Lower compute costs, smaller carbon footprint:
+
+- Researchers can share trained models instead of always retraining
+- Practitioners can reduce compute time and production costs
+- 8 architectures with over 30 pretrained models, some in more than 100 languages
+
+Choose the right framework for every part of a model's lifetime:
+
+- Train state-of-the-art models in 3 lines of code
+- Deep interoperability between TensorFlow 2.0 and PyTorch models
+- Move a single model between TF2.0/PyTorch frameworks at will
+- Seamlessly pick the right framework for training, evaluation, production
+
+Contents
+---------------------------------
+
+The library currently contains PyTorch and Tensorflow implementations, pre-trained model weights, usage scripts and conversion utilities for the following models:
 
 1. `BERT <https://github.com/google-research/bert>`_ (from Google) released with the paper `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding <https://arxiv.org/abs/1810.04805>`_ by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
 2. `GPT <https://github.com/openai/finetune-transformer-lm>`_ (from OpenAI) released with the paper `Improving Language Understanding by Generative Pre-Training <https://blog.openai.com/language-unsupervised>`_ by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
@@ -12,7 +46,12 @@ The library currently contains PyTorch implementations, pre-trained model weight
 5. `XLNet <https://github.com/zihangdai/xlnet>`_ (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive Pretraining for Language Understanding <https://arxiv.org/abs/1906.08237>`_ by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
 6. `XLM <https://github.com/facebookresearch/XLM>`_ (from Facebook) released together with the paper `Cross-lingual Language Model Pretraining <https://arxiv.org/abs/1901.07291>`_ by Guillaume Lample and Alexis Conneau.
 7. `RoBERTa <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`_ (from Facebook), released together with the paper a `Robustly Optimized BERT Pretraining Approach <https://arxiv.org/abs/1907.11692>`_ by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
-8. `DistilBERT <https://huggingface.co/pytorch-transformers/model_doc/distilbert.html>`_ (from HuggingFace) released together with the blog post `Smaller, faster, cheaper, lighter: Introducing DistilBERT, a distilled version of BERT <https://medium.com/huggingface/distilbert-8cf3380435b5>`_ by Victor Sanh, Lysandre Debut and Thomas Wolf.
+8. `DistilBERT <https://huggingface.co/transformers/model_doc/distilbert.html>`_ (from HuggingFace) released together with the paper `DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter <https://arxiv.org/abs/1910.01108>`_ by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into `DistilGPT2 <https://github.com/huggingface/transformers/tree/master/examples/distillation>`_.
+9. `CTRL <https://github.com/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.
+13. `FlauBERT <https://github.com/getalp/Flaubert>`_ (from CNRS) released with the paper `FlauBERT: Unsupervised Language Model Pre-training for French <https://arxiv.org/abs/1912.05372>`_ by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
 
 .. toctree::
     :maxdepth: 2
@@ -20,7 +59,10 @@ The library currently contains PyTorch implementations, pre-trained model weight
 
     installation
     quickstart
+    glossary
     pretrained_models
+    usage
+    model_sharing
     examples
     notebooks
     serialization
@@ -28,6 +70,8 @@ The library currently contains PyTorch implementations, pre-trained model weight
     migration
     bertology
     torchscript
+    multilingual
+    benchmarks
 
 .. toctree::
     :maxdepth: 2
@@ -36,7 +80,9 @@ The library currently contains PyTorch implementations, pre-trained model weight
     main_classes/configuration
     main_classes/model
     main_classes/tokenizer
+    main_classes/pipelines
     main_classes/optimizer_schedules
+    main_classes/processors
 
 .. toctree::
     :maxdepth: 2
@@ -51,3 +97,11 @@ The library currently contains PyTorch implementations, pre-trained model weight
     model_doc/xlnet
     model_doc/roberta
     model_doc/distilbert
+    model_doc/ctrl
+    model_doc/camembert
+    model_doc/albert
+    model_doc/xlmroberta
+    model_doc/flaubert
+    model_doc/bart
+    model_doc/t5
+    model_doc/electra

docs/source/installation.md

@@ -0,0 +1,51 @@
+# Installation
+
+Transformers is tested on Python 3.6+ and PyTorch 1.1.0
+
+## With pip
+
+PyTorch Transformers can be installed using pip as follows:
+
+``` bash
+pip install transformers
+```
+
+## From source
+
+To install from source, clone the repository and install with:
+
+``` bash
+git clone https://github.com/huggingface/transformers.git
+cd transformers
+pip install .
+```
+
+## Tests
+
+An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples).
+
+Refer to the [contributing guide](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md#tests) for details about running tests.
+
+## OpenAI GPT original tokenization workflow
+
+If you want to reproduce the original tokenization process of the `OpenAI GPT` paper, you will need to install `ftfy` and `SpaCy`:
+
+``` bash
+pip install spacy ftfy==4.4.3
+python -m spacy download en
+```
+
+If you don't install `ftfy` and `SpaCy`, the `OpenAI GPT` tokenizer will default to tokenize using BERT's `BasicTokenizer` followed by Byte-Pair Encoding (which should be fine for most usage, don't worry).
+
+## Note on model downloads (Continuous Integration or large-scale deployments)
+
+If you expect to be downloading large volumes of models (more than 1,000) from our hosted bucket (for instance through your CI setup, or a large-scale production deployment), please cache the model files on your end. It will be way faster, and cheaper. Feel free to contact us privately if you need any help.
+
+## Do you want to run a Transformer model on a mobile device?
+
+You should check out our [swift-coreml-transformers](https://github.com/huggingface/swift-coreml-transformers) repo.
+
+It contains a set of tools to convert PyTorch or TensorFlow 2.0 trained Transformer models (currently contains `GPT-2`, `DistilGPT-2`, `BERT`, and `DistilBERT`) to CoreML models that run on iOS devices.
+
+At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch to productizing them in CoreML,
+or prototype a model or an app in CoreML then research its hyperparameters or architecture from PyTorch. Super exciting!

docs/source/installation.rst

@@ -1,71 +0,0 @@
-Installation
-================================================
-
-PyTorch-Transformers is tested on Python 2.7 and 3.5+ (examples are tested only on python 3.5+) and PyTorch 1.1.0
-
-With pip
-^^^^^^^^
-
-PyTorch Transformers can be installed using pip as follows:
-
-.. code-block:: bash
-
-   pip install pytorch-transformers
-
-From source
-^^^^^^^^^^^
-
-To install from source, clone the repository and install with:
-
-.. code-block:: bash
-
-    git clone https://github.com/huggingface/pytorch-transformers.git
-    cd pytorch-transformers
-    pip install [--editable] .
-
-
-Tests
-^^^^^
-
-An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the `tests folder <https://github.com/huggingface/pytorch-transformers/tree/master/pytorch_transformers/tests>`_ and examples tests in the `examples folder <https://github.com/huggingface/pytorch-transformers/tree/master/examples>`_.
-
-Tests can be run using `pytest` (install pytest if needed with `pip install pytest`).
-
-Run all the tests from the root of the cloned repository with the commands:
-
-.. code-block:: bash
-
-    python -m pytest -sv ./pytorch_transformers/tests/
-    python -m pytest -sv ./examples/
-
-
-OpenAI GPT original tokenization workflow
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-If you want to reproduce the original tokenization process of the ``OpenAI GPT`` paper, you will need to install ``ftfy`` (use version 4.4.3 if you are using Python 2) and ``SpaCy`` :
-
-.. code-block:: bash
-
-   pip install spacy ftfy==4.4.3
-   python -m spacy download en
-
-If you don't install ``ftfy`` and ``SpaCy``\ , the ``OpenAI GPT`` tokenizer will default to tokenize using BERT's ``BasicTokenizer`` followed by Byte-Pair Encoding (which should be fine for most usage, don't worry).
-
-
-Note on model downloads (Continuous Integration or large-scale deployments)
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-If you expect to be downloading large volumes of models (more than 1,000) from our hosted bucket (for instance through your CI setup, or a large-scale production deployment), please cache the model files on your end. It will be way faster, and cheaper. Feel free to contact us privately if you need any help.
-
-
-Do you want to run a Transformer model on a mobile device?
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-You should check out our `swift-coreml-transformers <https://github.com/huggingface/swift-coreml-transformers>`_ repo.
-
-It contains an example of a conversion script from a Pytorch trained Transformer model (here, ``GPT-2``) to a CoreML model that runs on iOS devices.
-
-It also contains an implementation of BERT for Question answering.
-
-At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch to productizing them in CoreML,
-or prototype a model or an app in CoreML then research its hyperparameters or architecture from PyTorch. Super exciting!

docs/source/main_classes/configuration.rst

@@ -6,5 +6,5 @@ The base class ``PretrainedConfig`` implements the common methods for loading/sa
 ``PretrainedConfig``
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.PretrainedConfig
+.. autoclass:: transformers.PretrainedConfig
     :members:

docs/source/main_classes/model.rst

@@ -11,5 +11,11 @@ The base class ``PreTrainedModel`` implements the common methods for loading/sav
 ``PreTrainedModel``
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.PreTrainedModel
+.. autoclass:: transformers.PreTrainedModel
+    :members:
+
+``TFPreTrainedModel``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFPreTrainedModel
     :members:

docs/source/main_classes/optimizer_schedules.rst

@@ -5,41 +5,45 @@ The ``.optimization`` module provides:
 
 - an optimizer with weight decay fixed that can be used to fine-tuned models, and
 - several schedules in the form of schedule objects that inherit from ``_LRSchedule``:
+- a gradient accumulation class to accumulate the gradients of multiple batches
 
 ``AdamW``
 ~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.AdamW
+.. autoclass:: transformers.AdamW
     :members:
 
+``AdamWeightDecay``
+~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AdamWeightDecay
+    :members:
+
+.. autofunction:: transformers.create_optimizer
+
 Schedules
 ----------------------------------------------------
 
 Learning Rate Schedules
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-.. autoclass:: pytorch_transformers.ConstantLRSchedule
-    :members:
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+.. autofunction:: transformers.get_constant_schedule
 
 
-.. autoclass:: pytorch_transformers.WarmupConstantSchedule
-    :members:
+.. autofunction:: transformers.get_constant_schedule_with_warmup
 
 .. image:: /imgs/warmup_constant_schedule.png
     :target: /imgs/warmup_constant_schedule.png
     :alt:
 
 
-.. autoclass:: pytorch_transformers.WarmupCosineSchedule
-    :members:
+.. autofunction:: transformers.get_cosine_schedule_with_warmup
 
 .. image:: /imgs/warmup_cosine_schedule.png
     :target: /imgs/warmup_cosine_schedule.png
     :alt:
 
 
-.. autoclass:: pytorch_transformers.WarmupCosineWithHardRestartsSchedule
-    :members:
+.. autofunction:: transformers.get_cosine_with_hard_restarts_schedule_with_warmup
 
 .. image:: /imgs/warmup_cosine_hard_restarts_schedule.png
     :target: /imgs/warmup_cosine_hard_restarts_schedule.png
@@ -47,9 +51,22 @@ Learning Rate Schedules
 
 
 
-.. autoclass:: pytorch_transformers.WarmupLinearSchedule
-    :members:
+.. autofunction:: transformers.get_linear_schedule_with_warmup
 
 .. image:: /imgs/warmup_linear_schedule.png
     :target: /imgs/warmup_linear_schedule.png
     :alt:
+
+``Warmup``
+~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.WarmUp
+    :members:
+
+Gradient Strategies
+----------------------------------------------------
+
+``GradientAccumulator``
+~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.GradientAccumulator

docs/source/main_classes/pipelines.rst

@@ -0,0 +1,68 @@
+Pipelines
+----------------------------------------------------
+
+The pipelines are a great and easy way to use models for inference. These pipelines are objects that abstract most
+of the complex code from the library, offering a simple API dedicated to several tasks, including Named Entity
+Recognition, Masked Language Modeling, Sentiment Analysis, Feature Extraction and Question Answering.
+
+There are two categories of pipeline abstractions to be aware about:
+
+- The :class:`~transformers.pipeline` which is the most powerful object encapsulating all other pipelines
+- The other task-specific pipelines, such as :class:`~transformers.NerPipeline`
+  or :class:`~transformers.QuestionAnsweringPipeline`
+
+The pipeline abstraction
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The `pipeline` abstraction is a wrapper around all the other available pipelines. It is instantiated as any
+other pipeline but requires an additional argument which is the `task`.
+
+.. autoclass:: transformers.pipeline
+    :members:
+
+
+The task specific pipelines
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Parent class: Pipeline
+=========================================
+
+.. autoclass:: transformers.Pipeline
+    :members: predict, transform, save_pretrained
+
+NerPipeline
+==========================================
+
+.. autoclass:: transformers.NerPipeline
+
+TokenClassificationPipeline
+==========================================
+
+This class is an alias of the :class:`~transformers.NerPipeline` defined above. Please refer to that pipeline for
+documentation and usage examples.
+
+FillMaskPipeline
+==========================================
+
+.. autoclass:: transformers.FillMaskPipeline
+
+FeatureExtractionPipeline
+==========================================
+
+.. autoclass:: transformers.FeatureExtractionPipeline
+
+TextClassificationPipeline
+==========================================
+
+.. autoclass:: transformers.TextClassificationPipeline
+
+QuestionAnsweringPipeline
+==========================================
+
+.. autoclass:: transformers.QuestionAnsweringPipeline
+
+
+SummarizationPipeline
+==========================================
+
+.. autoclass:: transformers.SummarizationPipeline

docs/source/main_classes/processors.rst

@@ -0,0 +1,153 @@
+Processors
+----------------------------------------------------
+
+This library includes processors for several traditional tasks. These processors can be used to process a dataset into
+examples that can be fed to a model.
+
+Processors
+~~~~~~~~~~~~~~~~~~~~~
+
+All processors follow the same architecture which is that of the
+:class:`~transformers.data.processors.utils.DataProcessor`. The processor returns a list
+of :class:`~transformers.data.processors.utils.InputExample`. These
+:class:`~transformers.data.processors.utils.InputExample` can be converted to
+:class:`~transformers.data.processors.utils.InputFeatures` in order to be fed to the model.
+
+.. autoclass:: transformers.data.processors.utils.DataProcessor
+    :members:
+
+
+.. autoclass:: transformers.data.processors.utils.InputExample
+    :members:
+
+
+.. autoclass:: transformers.data.processors.utils.InputFeatures
+    :members:
+
+
+GLUE
+~~~~~~~~~~~~~~~~~~~~~
+
+`General Language Understanding Evaluation (GLUE) <https://gluebenchmark.com/>`__ is a benchmark that evaluates
+the performance of models across a diverse set of existing NLU tasks. It was released together with the paper
+`GLUE: A multi-task benchmark and analysis platform for natural language understanding <https://openreview.net/pdf?id=rJ4km2R5t7>`__
+
+This library hosts a total of 10 processors for the following tasks: MRPC, MNLI, MNLI (mismatched),
+CoLA, SST2, STSB, QQP, QNLI, RTE and WNLI.
+
+Those processors are:
+    - :class:`~transformers.data.processors.utils.MrpcProcessor`
+    - :class:`~transformers.data.processors.utils.MnliProcessor`
+    - :class:`~transformers.data.processors.utils.MnliMismatchedProcessor`
+    - :class:`~transformers.data.processors.utils.Sst2Processor`
+    - :class:`~transformers.data.processors.utils.StsbProcessor`
+    - :class:`~transformers.data.processors.utils.QqpProcessor`
+    - :class:`~transformers.data.processors.utils.QnliProcessor`
+    - :class:`~transformers.data.processors.utils.RteProcessor`
+    - :class:`~transformers.data.processors.utils.WnliProcessor`
+
+Additionally, the following method  can be used to load values from a data file and convert them to a list of
+:class:`~transformers.data.processors.utils.InputExample`.
+
+.. automethod:: transformers.data.processors.glue.glue_convert_examples_to_features
+
+Example usage
+^^^^^^^^^^^^^^^^^^^^^^^^^
+
+An example using these processors is given in the `run_glue.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_glue.py>`__ script.
+
+
+XNLI
+~~~~~~~~~~~~~~~~~~~~~
+
+`The Cross-Lingual NLI Corpus (XNLI) <https://www.nyu.edu/projects/bowman/xnli/>`__ is a benchmark that evaluates
+the quality of cross-lingual text representations. 
+XNLI is crowd-sourced dataset based on `MultiNLI <http://www.nyu.edu/projects/bowman/multinli/>`: pairs of text are labeled with textual entailment 
+annotations for 15 different languages (including both high-resource language such as English and low-resource languages such as Swahili).
+
+It was released together with the paper
+`XNLI: Evaluating Cross-lingual Sentence Representations <https://arxiv.org/abs/1809.05053>`__
+
+This library hosts the processor to load the XNLI data:
+    - :class:`~transformers.data.processors.utils.XnliProcessor`
+
+Please note that since the gold labels are available on the test set, evaluation is performed on the test set.
+
+An example using these processors is given in the
+`run_xnli.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_xnli.py>`__ script.
+
+
+SQuAD
+~~~~~~~~~~~~~~~~~~~~~
+
+`The Stanford Question Answering Dataset (SQuAD) <https://rajpurkar.github.io/SQuAD-explorer//>`__ is a benchmark that evaluates
+the performance of models on question answering. Two versions are available, v1.1 and v2.0. The first version (v1.1) was released together with the paper
+`SQuAD: 100,000+ Questions for Machine Comprehension of Text <https://arxiv.org/abs/1606.05250>`__. The second version (v2.0) was released alongside 
+the paper `Know What You Don't Know: Unanswerable Questions for SQuAD <https://arxiv.org/abs/1806.03822>`__.
+
+This library hosts a processor for each of the two versions:
+
+Processors
+^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Those processors are:
+    - :class:`~transformers.data.processors.utils.SquadV1Processor`
+    - :class:`~transformers.data.processors.utils.SquadV2Processor`
+
+They both inherit from the abstract class :class:`~transformers.data.processors.utils.SquadProcessor`
+
+.. autoclass:: transformers.data.processors.squad.SquadProcessor
+    :members:
+
+Additionally, the following method can be used to convert SQuAD examples into :class:`~transformers.data.processors.utils.SquadFeatures`
+that can be used as model inputs.
+
+.. automethod:: transformers.data.processors.squad.squad_convert_examples_to_features
+
+These processors as well as the aforementionned method can be used with files containing the data as well as with the `tensorflow_datasets` package.
+Examples are given below.
+
+
+Example usage
+^^^^^^^^^^^^^^^^^^^^^^^^^
+Here is an example using the processors as well as the conversion method using data files:
+
+Example::
+
+    # Loading a V2 processor
+    processor = SquadV2Processor()
+    examples = processor.get_dev_examples(squad_v2_data_dir)
+
+    # Loading a V1 processor
+    processor = SquadV1Processor()
+    examples = processor.get_dev_examples(squad_v1_data_dir)
+
+    features = squad_convert_examples_to_features( 
+        examples=examples,
+        tokenizer=tokenizer,
+        max_seq_length=max_seq_length,
+        doc_stride=args.doc_stride,
+        max_query_length=max_query_length,
+        is_training=not evaluate,
+    )
+
+Using `tensorflow_datasets` is as easy as using a data file:
+
+Example::
+
+    # tensorflow_datasets only handle Squad V1.
+    tfds_examples = tfds.load("squad")
+    examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
+
+    features = squad_convert_examples_to_features( 
+        examples=examples,
+        tokenizer=tokenizer,
+        max_seq_length=max_seq_length,
+        doc_stride=args.doc_stride,
+        max_query_length=max_query_length,
+        is_training=not evaluate,
+    )
+
+
+Another example using these processors is given in the
+`run_squad.py <https://github.com/huggingface/transformers/blob/master/examples/run_squad.py>`__ script.

docs/source/main_classes/tokenizer.rst

@@ -12,5 +12,5 @@ The base class ``PreTrainedTokenizer`` implements the common methods for loading
 ``PreTrainedTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.PreTrainedTokenizer
+.. autoclass:: transformers.PreTrainedTokenizer
     :members:

docs/source/migration.md

@@ -1,17 +1,17 @@
 # Migrating from pytorch-pretrained-bert
 
 
-Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to `pytorch-transformers`
+Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to `transformers`
 
 ### Models always output `tuples`
 
-The main breaking change when migrating from `pytorch-pretrained-bert` to `pytorch-transformers` is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
+The main breaking change when migrating from `pytorch-pretrained-bert` to `transformers` is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
 
-The exact content of the tuples for each model are detailled in the models' docstrings and the [documentation](https://huggingface.co/pytorch-transformers/).
+The exact content of the tuples for each model are detailled in the models' docstrings and the [documentation](https://huggingface.co/transformers/).
 
 In pretty much every case, you will be fine by taking the first element of the output as the output you previously used in `pytorch-pretrained-bert`.
 
-Here is a `pytorch-pretrained-bert` to `pytorch-transformers` conversion example for a `BertForSequenceClassification` classification model:
+Here is a `pytorch-pretrained-bert` to `transformers` conversion example for a `BertForSequenceClassification` classification model:
 
 ```python
 # Let's load our model
@@ -20,14 +20,14 @@ model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
 # If you used to have this line in pytorch-pretrained-bert:
 loss = model(input_ids, labels=labels)
 
-# Now just use this line in pytorch-transformers to extract the loss from the output tuple:
+# Now just use this line in transformers to extract the loss from the output tuple:
 outputs = model(input_ids, labels=labels)
 loss = outputs[0]
 
-# In pytorch-transformers you can also have access to the logits:
+# In transformers you can also have access to the logits:
 loss, logits = outputs[:2]
 
-# And even the attention weigths if you configure the model to output them (and other outputs too, see the docstrings and documentation)
+# And even the attention weights if you configure the model to output them (and other outputs too, see the docstrings and documentation)
 model = BertForSequenceClassification.from_pretrained('bert-base-uncased', output_attentions=True)
 outputs = model(input_ids, labels=labels)
 loss, logits, attentions = outputs
@@ -84,26 +84,26 @@ Here is a conversion examples from `BertAdam` with a linear warmup and decay sch
 # Parameters:
 lr = 1e-3
 max_grad_norm = 1.0
-num_total_steps = 1000
+num_training_steps = 1000
 num_warmup_steps = 100
-warmup_proportion = float(num_warmup_steps) / float(num_total_steps)  # 0.1
+warmup_proportion = float(num_warmup_steps) / float(num_training_steps)  # 0.1
 
 ### Previously BertAdam optimizer was instantiated like this:
-optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
+optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, num_training_steps=num_training_steps)
 ### and used like this:
 for batch in train_data:
     loss = model(batch)
     loss.backward()
     optimizer.step()
 
-### In PyTorch-Transformers, optimizer and schedules are splitted and instantiated like this:
+### In Transformers, optimizer and schedules are splitted and instantiated like this:
 optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False)  # To reproduce BertAdam specific behavior set correct_bias=False
-scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_steps)  # PyTorch scheduler
+scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)  # PyTorch scheduler
 ### and used like this:
 for batch in train_data:
     loss = model(batch)
     loss.backward()
     torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm)  # Gradient clipping is not in AdamW anymore (so you can use amp without issue)
-    scheduler.step()
     optimizer.step()
+    scheduler.step()
 ```

docs/source/model_doc/albert.rst

@@ -0,0 +1,94 @@
+ALBERT
+----------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~
+
+The ALBERT model was proposed in `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. It presents
+two parameter-reduction techniques to lower memory consumption and increase the trainig speed of BERT:
+
+- Splitting the embedding matrix into two smaller matrices
+- Using repeating layers split among groups
+
+The abstract from the paper is the following:
+
+*Increasing model size when pretraining natural language representations often results in improved performance on
+downstream tasks. However, at some point further model increases become harder due to GPU/TPU memory limitations,
+longer training times, and unexpected model degradation. To address these problems, we present two parameter-reduction
+techniques to lower memory consumption and increase the training speed of BERT. Comprehensive empirical evidence shows
+that our proposed methods lead to models that scale much better compared to the original BERT. We also use a
+self-supervised loss that focuses on modeling inter-sentence coherence, and show it consistently helps downstream
+tasks with multi-sentence inputs. As a result, our best model establishes new state-of-the-art results on the GLUE,
+RACE, and SQuAD benchmarks while having fewer parameters compared to BERT-large.*
+
+Tips:
+
+- ALBERT is a model with absolute position embeddings so it's usually advised to pad the inputs on
+  the right rather than the left.
+- ALBERT uses repeating layers which results in a small memory footprint, however the computational cost remains
+  similar to a BERT-like architecture with the same number of hidden layers as it has to iterate through the same
+  number of (repeating) layers.
+
+AlbertConfig
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AlbertConfig
+    :members:
+
+
+AlbertTokenizer
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AlbertTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+AlbertModel
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AlbertModel
+    :members:
+
+
+AlbertForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AlbertForMaskedLM
+    :members:
+
+
+AlbertForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AlbertForSequenceClassification
+    :members:
+
+
+AlbertForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AlbertForQuestionAnswering
+    :members:
+
+
+TFAlbertModel
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFAlbertModel
+    :members:
+
+
+TFAlbertForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFAlbertForMaskedLM
+    :members:
+
+
+TFAlbertForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFAlbertForSequenceClassification
+    :members:

docs/source/model_doc/auto.rst

@@ -3,7 +3,7 @@ AutoModels
 
 In many cases, the architecture you want to use can be guessed from the name or the path of the pretrained model you are supplying to the ``from_pretrained`` method.
 
-AutoClasses are here to do this job for you so that you automatically retreive the relevant model given the name/path to the pretrained weights/config/vocabulary:
+AutoClasses are here to do this job for you so that you automatically retrieve the relevant model given the name/path to the pretrained weights/config/vocabulary:
 
 Instantiating one of ``AutoModel``, ``AutoConfig`` and ``AutoTokenizer`` will directly create a class of the relevant architecture (ex: ``model = AutoModel.from_pretrained('bert-base-cased')`` will create a instance of ``BertModel``).
 
@@ -11,19 +11,55 @@ Instantiating one of ``AutoModel``, ``AutoConfig`` and ``AutoTokenizer`` will di
 ``AutoConfig``
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.AutoConfig
-    :members:
-
-
-``AutoModel``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: pytorch_transformers.AutoModel
+.. autoclass:: transformers.AutoConfig
     :members:
 
 
 ``AutoTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.AutoTokenizer
+.. autoclass:: transformers.AutoTokenizer
     :members:
+
+
+``AutoModel``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AutoModel
+    :members:
+
+
+``AutoModelForPreTraining``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AutoModelForPreTraining
+    :members:
+
+
+``AutoModelWithLMHead``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AutoModelWithLMHead
+    :members:
+
+
+``AutoModelForSequenceClassification``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AutoModelForSequenceClassification
+    :members:
+
+
+``AutoModelForQuestionAnswering``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AutoModelForQuestionAnswering
+    :members:
+
+
+``AutoModelForTokenClassification``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AutoModelForTokenClassification
+    :members:
+

docs/source/model_doc/bart.rst

@@ -0,0 +1,56 @@
+Bart
+----------------------------------------------------
+**DISCLAIMER:** This model is still a work in progress, if you see something strange,
+file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
+@sshleifer
+
+Paper
+~~~~~
+The Bart model was `proposed <https://arxiv.org/abs/1910.13461>`_ by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer on 29 Oct, 2019.
+According to the abstract,
+
+- Bart uses a standard seq2seq/machine translation architecture with a bidirectional encoder (like BERT) and a left-to-right decoder (like GPT).
+- The pretraining task involves randomly shuffling the order of the original sentences and a novel in-filling scheme, where spans of text are replaced with a single mask token.
+- BART is particularly effective when fine tuned for text generation but also works well for comprehension tasks. It matches the performance of RoBERTa with comparable training resources on GLUE and SQuAD, achieves new state-of-the-art results on a range of abstractive dialogue, question answering, and summarization tasks, with gains of up to 6 ROUGE.
+
+The Authors' code can be found `here <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_
+
+
+Implementation Notes
+~~~~~~~~~~~~~~~~~~~~
+- Bart doesn't use :obj:`token_type_ids` for sequence classification. Use BartTokenizer.encode to get the proper splitting.
+- The forward pass of ``BartModel`` will create decoder inputs (using the helper function ``transformers.modeling_bart._prepare_bart_decoder_inputs``)  if they are not passed. This is different than some other modeling APIs.
+- Model predictions are intended to be identical to the original implementation. This only works, however, if the string you pass to ``fairseq.encode`` starts with a space.
+- ``BartForConditionalGeneration.generate`` should be used for conditional generation tasks like summarization, see the example in that docstrings
+- Models that load the ``"bart-large-cnn"`` weights will not have a ``mask_token_id``, or be able to perform mask filling tasks.
+
+
+
+BartModel
+~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BartModel
+    :members: forward
+
+.. autofunction:: transformers.modeling_bart._prepare_bart_decoder_inputs
+
+
+BartForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BartForConditionalGeneration
+    :members: generate, forward
+
+
+BartForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BartForSequenceClassification
+    :members: forward
+
+BartConfig
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BartConfig
+    :members:
+

docs/source/model_doc/bert.rst

@@ -1,72 +1,163 @@
 BERT
 ----------------------------------------------------
 
-``BertConfig``
+Overview
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.BertConfig
+The BERT model was proposed in `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding <https://arxiv.org/abs/1810.04805>`__
+by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova. It's a bidirectional transformer
+pre-trained using a combination of masked language modeling objective and next sentence prediction
+on a large corpus comprising the Toronto Book Corpus and Wikipedia.
+
+The abstract from the paper is the following:
+
+*We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations
+from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional
+representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result,
+the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models
+for a wide range of tasks, such as question answering and language inference, without substantial task-specific
+architecture modifications.*
+
+*BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural
+language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI
+accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute
+improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement).*
+
+Tips:
+
+- BERT is a model with absolute position embeddings so it's usually advised to pad the inputs on
+  the right rather than the left.
+- BERT was trained with a masked language modeling (MLM) objective. It is therefore efficient at predicting masked
+  tokens and at NLU in general, but is not optimal for text generation. Models trained with a causal language
+  modeling (CLM) objective are better in that regard.
+- Alongside MLM, BERT was trained using a next sentence prediction (NSP) objective using the [CLS] token as a sequence
+  approximate. The user may use this token (the first token in a sequence built with special tokens) to get a sequence
+  prediction rather than a token prediction. However, averaging over the sequence may yield better results than using
+  the [CLS] token.
+
+BertConfig
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BertConfig
     :members:
 
 
-``BertTokenizer``
+BertTokenizer
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.BertTokenizer
-    :members:
+.. autoclass:: transformers.BertTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
 
 
-``BertModel``
+BertModel
 ~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.BertModel
+.. autoclass:: transformers.BertModel
     :members:
 
 
-``BertForPreTraining``
+BertForPreTraining
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.BertForPreTraining
+.. autoclass:: transformers.BertForPreTraining
     :members:
 
 
-``BertForMaskedLM``
+BertForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.BertForMaskedLM
+.. autoclass:: transformers.BertForMaskedLM
     :members:
 
 
-``BertForNextSentencePrediction``
+BertForNextSentencePrediction
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.BertForNextSentencePrediction
+.. autoclass:: transformers.BertForNextSentencePrediction
     :members:
 
 
-``BertForSequenceClassification``
+BertForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.BertForSequenceClassification
+.. autoclass:: transformers.BertForSequenceClassification
     :members:
 
 
-``BertForMultipleChoice``
+BertForMultipleChoice
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.BertForMultipleChoice
+.. autoclass:: transformers.BertForMultipleChoice
     :members:
 
 
-``BertForTokenClassification``
+BertForTokenClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.BertForTokenClassification
+.. autoclass:: transformers.BertForTokenClassification
     :members:
 
 
-``BertForQuestionAnswering``
+BertForQuestionAnswering
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.BertForQuestionAnswering
+.. autoclass:: transformers.BertForQuestionAnswering
+    :members:
+
+
+TFBertModel
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFBertModel
+    :members:
+
+
+TFBertForPreTraining
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFBertForPreTraining
+    :members:
+
+
+TFBertForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFBertForMaskedLM
+    :members:
+
+
+TFBertForNextSentencePrediction
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFBertForNextSentencePrediction
+    :members:
+
+
+TFBertForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFBertForSequenceClassification
+    :members:
+
+
+TFBertForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFBertForMultipleChoice
+    :members:
+
+
+TFBertForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFBertForTokenClassification
+    :members:
+
+
+TFBertForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFBertForQuestionAnswering
     :members:
 

docs/source/model_doc/camembert.rst

@@ -0,0 +1,100 @@
+CamemBERT
+----------------------------------------------------
+
+The CamemBERT model was proposed in `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. It is based on Facebook's RoBERTa model released in 2019. It is a model
+trained on 138GB of French text.
+
+The abstract from the paper is the following:
+
+*Pretrained language models are now ubiquitous in Natural Language Processing. Despite their success,
+most available models have either been trained on English data or on the concatenation of data in multiple
+languages. This makes practical use of such models --in all languages except English-- very limited. Aiming
+to address this issue for French, we release CamemBERT, a French version of the Bi-directional Encoders for
+Transformers (BERT). We measure the performance of CamemBERT compared to multilingual models in multiple
+downstream tasks, namely part-of-speech tagging, dependency parsing, named-entity recognition, and natural
+language inference. CamemBERT improves the state of the art for most of the tasks considered. We release the
+pretrained model for CamemBERT hoping to foster research and downstream applications for French NLP.*
+
+Tips:
+
+- This implementation is the same as RoBERTa. Refer to the `documentation of RoBERTa <./roberta.html>`__ for usage
+  examples as well as the information relative to the inputs and outputs.
+
+CamembertConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CamembertConfig
+    :members:
+
+
+CamembertTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CamembertTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+CamembertModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CamembertModel
+    :members:
+
+
+CamembertForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CamembertForMaskedLM
+    :members:
+
+
+CamembertForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CamembertForSequenceClassification
+    :members:
+
+
+CamembertForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CamembertForMultipleChoice
+    :members:
+
+
+CamembertForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CamembertForTokenClassification
+    :members:
+
+
+TFCamembertModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFCamembertModel
+    :members:
+
+
+TFCamembertForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFCamembertForMaskedLM
+    :members:
+
+
+TFCamembertForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFCamembertForSequenceClassification
+    :members:
+
+
+TFCamembertForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFCamembertForTokenClassification
+    :members:

docs/source/model_doc/ctrl.rst

@@ -0,0 +1,75 @@
+CTRL
+----------------------------------------------------
+
+CTRL model was proposed in `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.
+It's a causal (unidirectional) transformer pre-trained using language modeling on a very large
+corpus of ~140 GB of text data with the first token reserved as a control code (such as Links, Books, Wikipedia etc.).
+
+The abstract from the paper is the following:
+
+*Large-scale language models show promising text generation capabilities, but users cannot easily control particular
+aspects of the generated text. We release CTRL, a 1.63 billion-parameter conditional transformer language model,
+trained to condition on control codes that govern style, content, and task-specific behavior. Control codes were
+derived from structure that naturally co-occurs with raw text, preserving the advantages of unsupervised learning
+while providing more explicit control over text generation. These codes also allow CTRL to predict which parts of
+the training data are most likely given a sequence. This provides a potential method for analyzing large amounts
+of data via model-based source attribution.*
+
+Tips:
+
+- CTRL makes use of control codes to generate text: it requires generations to be started by certain words, sentences
+  or links to generate coherent text. Refer to the `original implementation <https://github.com/salesforce/ctrl>`__
+  for more information.
+- CTRL is a model with absolute position embeddings so it's usually advised to pad the inputs on
+  the right rather than the left.
+- CTRL was trained with a causal language modeling (CLM) objective and is therefore powerful at predicting the next
+  token in a sequence. Leveraging this feature allows CTRL to generate syntactically coherent text as
+  it can be observed in the `run_generation.py` example script.
+- The PyTorch models can take the `past` as input, which is the previously computed key/value attention pairs. Using
+  this `past` value prevents the model from re-computing pre-computed values in the context of text generation.
+  See `reusing the past in generative models <../quickstart.html#using-the-past>`_ for more information on the usage
+  of this argument.
+
+
+CTRLConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CTRLConfig
+    :members:
+
+
+CTRLTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CTRLTokenizer
+    :members: save_vocabulary
+
+
+CTRLModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CTRLModel
+    :members:
+
+
+CTRLLMHeadModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CTRLLMHeadModel
+    :members:
+
+
+TFCTRLModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFCTRLModel
+    :members:
+
+
+TFCTRLLMHeadModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFCTRLLMHeadModel
+    :members:
+

docs/source/model_doc/distilbert.rst

@@ -1,43 +1,97 @@
 DistilBERT
 ----------------------------------------------------
 
-``DistilBertConfig``
+The DistilBERT model was proposed in the blog post
+`Smaller, faster, cheaper, lighter: Introducing DistilBERT, a distilled version of BERT <https://medium.com/huggingface/distilbert-8cf3380435b5>`__,
+and the paper `DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter <https://arxiv.org/abs/1910.01108>`__.
+DistilBERT is a small, fast, cheap and light Transformer model trained by distilling Bert base. It has 40% less
+parameters than `bert-base-uncased`, runs 60% faster while preserving over 95% of Bert's performances as measured on
+the GLUE language understanding benchmark.
+
+The abstract from the paper is the following:
+
+*As Transfer Learning from large-scale pre-trained models becomes more prevalent in Natural Language Processing (NLP),
+operating these large models in on-the-edge and/or under constrained computational training or inference budgets
+remains challenging. In this work, we propose a method to pre-train a smaller general-purpose language representation
+model, called DistilBERT, which can then be fine-tuned with good performances on a wide range of tasks like its larger
+counterparts. While most prior work investigated the use of distillation for building task-specific models, we
+leverage knowledge distillation during the pre-training phase and show that it is possible to reduce the size of a
+BERT model by 40%, while retaining 97% of its language understanding capabilities and being 60% faster. To leverage
+the inductive biases learned by larger models during pre-training, we introduce a triple loss combining language
+modeling, distillation and cosine-distance losses. Our smaller, faster and lighter model is cheaper to pre-train
+and we demonstrate its capabilities for on-device computations in a proof-of-concept experiment and a comparative
+on-device study.*
+
+Tips:
+
+- DistilBert doesn't have `token_type_ids`, you don't need to indicate which token belongs to which segment. Just separate your segments with the separation token `tokenizer.sep_token` (or `[SEP]`)
+- DistilBert doesn't have options to select the input positions (`position_ids` input). This could be added if necessary though, just let's us know if you need this option.
+
+
+DistilBertConfig
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.DistilBertConfig
+.. autoclass:: transformers.DistilBertConfig
     :members:
 
 
-``DistilBertTokenizer``
+DistilBertTokenizer
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.DistilBertTokenizer
+.. autoclass:: transformers.DistilBertTokenizer
     :members:
 
 
-``DistilBertModel``
+DistilBertModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.DistilBertModel
+.. autoclass:: transformers.DistilBertModel
     :members:
 
 
-``DistilBertForMaskedLM``
+DistilBertForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.DistilBertForMaskedLM
+.. autoclass:: transformers.DistilBertForMaskedLM
     :members:
 
 
-``DistilBertForSequenceClassification``
+DistilBertForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.DistilBertForSequenceClassification
+.. autoclass:: transformers.DistilBertForSequenceClassification
     :members:
 
 
-``DistilBertForQuestionAnswering``
+DistilBertForQuestionAnswering
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.DistilBertForQuestionAnswering
+.. autoclass:: transformers.DistilBertForQuestionAnswering
+    :members:
+
+TFDistilBertModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFDistilBertModel
+    :members:
+
+
+TFDistilBertForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFDistilBertForMaskedLM
+    :members:
+
+
+TFDistilBertForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFDistilBertForSequenceClassification
+    :members:
+
+
+TFDistilBertForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFDistilBertForQuestionAnswering
     :members:

docs/source/model_doc/electra.rst

@@ -0,0 +1,115 @@
+ELECTRA
+----------------------------------------------------
+
+The ELECTRA model was proposed in the paper.
+`ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators <https://openreview.net/pdf?id=r1xMH1BtvB>`__.
+ELECTRA is a new pre-training approach which trains two transformer models: the generator and the discriminator. The
+generator's role is to replace tokens in a sequence, and is therefore trained as a masked language model. The discriminator,
+which is the model we're interested in, tries to identify which tokens were replaced by the generator in the sequence.
+
+The abstract from the paper is the following:
+
+*Masked language modeling (MLM) pre-training methods such as BERT corrupt
+the input by replacing some tokens with [MASK] and then train a model to
+reconstruct the original tokens. While they produce good results when transferred
+to downstream NLP tasks, they generally require large amounts of compute to be
+effective. As an alternative, we propose a more sample-efficient pre-training task
+called replaced token detection. Instead of masking the input, our approach
+corrupts it by replacing some tokens with plausible alternatives sampled from a small
+generator network. Then, instead of training a model that predicts the original
+identities of the corrupted tokens, we train a discriminative model that predicts
+whether each token in the corrupted input was replaced by a generator sample
+or not. Thorough experiments demonstrate this new pre-training task is more
+efficient than MLM because the task is defined over all input tokens rather than
+just the small subset that was masked out. As a result, the contextual representations
+learned by our approach substantially outperform the ones learned by BERT
+given the same model size, data, and compute. The gains are particularly strong
+for small models; for example, we train a model on one GPU for 4 days that
+outperforms GPT (trained using 30x more compute) on the GLUE natural language
+understanding benchmark. Our approach also works well at scale, where it
+performs comparably to RoBERTa and XLNet while using less than 1/4 of their
+compute and outperforms them when using the same amount of compute.*
+
+Tips:
+
+- ELECTRA is the pre-training approach, therefore there is nearly no changes done to the underlying model: BERT. The
+  only change is the separation of the embedding size and the hidden size -> The embedding size is generally smaller,
+  while the hidden size is larger. An additional projection layer (linear) is used to project the embeddings from
+  their embedding size to the hidden size. In the case where the embedding size is the same as the hidden size, no
+  projection layer is used.
+- The ELECTRA checkpoints saved using `Google Research's implementation <https://github.com/google-research/electra>`__
+  contain both the generator and discriminator. The conversion script requires the user to name which model to export
+  into the correct architecture. Once converted to the HuggingFace format, these checkpoints may be loaded into all
+  available ELECTRA models, however. This means that the discriminator may be loaded in the `ElectraForMaskedLM` model,
+  and the generator may be loaded in the `ElectraForPreTraining` model (the classification head will be randomly
+  initialized as it doesn't exist in the generator).
+
+
+ElectraConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ElectraConfig
+    :members:
+
+
+ElectraTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ElectraTokenizer
+    :members:
+
+
+ElectraModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ElectraModel
+    :members:
+
+
+ElectraForPreTraining
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ElectraForPreTraining
+    :members:
+
+
+ElectraForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ElectraForMaskedLM
+    :members:
+
+
+ElectraForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ElectraForTokenClassification
+    :members:
+
+
+TFElectraModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFElectraModel
+    :members:
+
+
+TFElectraForPreTraining
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFElectraForPreTraining
+    :members:
+
+
+TFElectraForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFElectraForMaskedLM
+    :members:
+
+
+TFElectraForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFElectraForTokenClassification
+    :members:

docs/source/model_doc/flaubert.rst

@@ -0,0 +1,72 @@
+FlauBERT
+----------------------------------------------------
+
+The FlauBERT model was proposed in the paper
+`FlauBERT: Unsupervised Language Model Pre-training for French <https://arxiv.org/abs/1912.05372>`__ by Hang Le et al.
+It's a transformer pre-trained using a masked language modeling (MLM) objective (BERT-like).
+
+The abstract from the paper is the following:
+
+*Language models have become a key step to achieve state-of-the art results in many different Natural Language
+Processing (NLP) tasks. Leveraging the huge amount of unlabeled texts nowadays available, they provide an efficient
+way to pre-train continuous word representations that can be fine-tuned for a downstream task, along with their
+contextualization at the sentence level. This has been widely demonstrated for English using contextualized
+representations (Dai and Le, 2015; Peters et al., 2018; Howard and Ruder, 2018; Radford et al., 2018; Devlin et
+al., 2019; Yang et al., 2019b). In this paper, we introduce and share FlauBERT, a model learned on a very large
+and heterogeneous French corpus. Models of different sizes are trained using the new CNRS (French National Centre
+for Scientific Research) Jean Zay supercomputer. We apply our French language models to diverse NLP tasks (text
+classification, paraphrasing, natural language inference, parsing, word sense disambiguation) and show that most
+of the time they outperform other pre-training approaches. Different versions of FlauBERT as well as a unified
+evaluation protocol for the downstream tasks, called FLUE (French Language Understanding Evaluation), are shared
+to the research community for further reproducible experiments in French NLP.*
+
+
+FlaubertConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaubertConfig
+    :members:
+
+
+FlaubertTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaubertTokenizer
+    :members:
+
+
+FlaubertModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaubertModel
+    :members:
+
+
+FlaubertWithLMHeadModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaubertWithLMHeadModel
+    :members:
+
+
+FlaubertForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaubertForSequenceClassification
+    :members:
+
+
+FlaubertForQuestionAnsweringSimple
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaubertForQuestionAnsweringSimple
+    :members:
+
+
+FlaubertForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaubertForQuestionAnswering
+    :members:
+
+

docs/source/model_doc/gpt.rst

@@ -1,36 +1,92 @@
 OpenAI GPT
 ----------------------------------------------------
 
-``OpenAIGPTConfig``
+Overview
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.OpenAIGPTConfig
+OpenAI GPT model was proposed in `Improving Language Understanding by Generative Pre-Training <https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/language_understanding_paper.pdf>`__
+by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever. It's a causal (unidirectional)
+transformer pre-trained using language modeling on a large corpus will long range dependencies, the Toronto Book Corpus.
+
+The abstract from the paper is the following:
+
+*Natural language understanding comprises a wide range of diverse tasks such
+as textual entailment, question answering, semantic similarity assessment, and
+document classification. Although large unlabeled text corpora are abundant,
+labeled data for learning these specific tasks is scarce, making it challenging for
+discriminatively trained models to perform adequately. We demonstrate that large
+gains on these tasks can be realized by generative pre-training of a language model
+on a diverse corpus of unlabeled text, followed by discriminative fine-tuning on each
+specific task. In contrast to previous approaches, we make use of task-aware input
+transformations during fine-tuning to achieve effective transfer while requiring
+minimal changes to the model architecture. We demonstrate the effectiveness of
+our approach on a wide range of benchmarks for natural language understanding.
+Our general task-agnostic model outperforms discriminatively trained models that
+use architectures specifically crafted for each task, significantly improving upon the
+state of the art in 9 out of the 12 tasks studied.*
+
+Tips:
+
+- GPT is a model with absolute position embeddings so it's usually advised to pad the inputs on
+  the right rather than the left.
+- GPT was trained with a causal language modeling (CLM) objective and is therefore powerful at predicting the next
+  token in a sequence. Leveraging this feature allows GPT-2 to generate syntactically coherent text as
+  it can be observed in the `run_generation.py` example script.
+
+`Write With Transformer <https://transformer.huggingface.co/doc/gpt>`__ is a webapp created and hosted by
+Hugging Face showcasing the generative capabilities of several models. GPT is one of them.
+
+OpenAIGPTConfig
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.OpenAIGPTConfig
     :members:
 
 
-``OpenAIGPTTokenizer``
+OpenAIGPTTokenizer
 ~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.OpenAIGPTTokenizer
-    :members:
+.. autoclass:: transformers.OpenAIGPTTokenizer
+    :members: save_vocabulary
 
 
-``OpenAIGPTModel``
+OpenAIGPTModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.OpenAIGPTModel
+.. autoclass:: transformers.OpenAIGPTModel
     :members:
 
 
-``OpenAIGPTLMHeadModel``
+OpenAIGPTLMHeadModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.OpenAIGPTLMHeadModel
+.. autoclass:: transformers.OpenAIGPTLMHeadModel
     :members:
 
 
-``OpenAIGPTDoubleHeadsModel``
+OpenAIGPTDoubleHeadsModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.OpenAIGPTDoubleHeadsModel
+.. autoclass:: transformers.OpenAIGPTDoubleHeadsModel
+    :members:
+
+
+TFOpenAIGPTModel
+~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFOpenAIGPTModel
+    :members:
+
+
+TFOpenAIGPTLMHeadModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFOpenAIGPTLMHeadModel
+    :members:
+
+
+TFOpenAIGPTDoubleHeadsModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFOpenAIGPTDoubleHeadsModel
     :members:

docs/source/model_doc/gpt2.rst

@@ -1,36 +1,91 @@
 OpenAI GPT2
 ----------------------------------------------------
 
-``GPT2Config``
+Overview
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.GPT2Config
+OpenAI GPT-2 model was proposed in
+`Language Models are Unsupervised Multitask Learners <https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf>`_
+by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
+It's a causal (unidirectional) transformer pre-trained using  language modeling on a very large
+corpus of ~40 GB of text data.
+
+The abstract from the paper is the following:
+
+*GPT-2 is a large transformer-based language model with 1.5 billion parameters, trained on a dataset[1]
+of 8 million web pages. GPT-2 is trained with a simple objective: predict the next word, given all of the previous
+words within some text. The diversity of the dataset causes this simple goal to contain naturally occurring
+demonstrations of many tasks across diverse domains. GPT-2 is a direct scale-up of GPT, with more than 10X
+the parameters and trained on more than 10X the amount of data.*
+
+Tips:
+
+- GPT-2 is a model with absolute position embeddings so it's usually advised to pad the inputs on
+  the right rather than the left.
+- GPT-2 was trained with a causal language modeling (CLM) objective and is therefore powerful at predicting the next
+  token in a sequence. Leveraging this feature allows GPT-2 to generate syntactically coherent text as
+  it can be observed in the `run_generation.py` example script.
+- The PyTorch models can take the `past` as input, which is the previously computed key/value attention pairs. Using
+  this `past` value prevents the model from re-computing pre-computed values in the context of text generation.
+  See `reusing the past in generative models <../quickstart.html#using-the-past>`_ for more information on the usage
+  of this argument.
+
+`Write With Transformer <https://transformer.huggingface.co/doc/gpt2-large>`__ is a webapp created and hosted by
+Hugging Face showcasing the generative capabilities of several models. GPT-2 is one of them and is available in five
+different sizes: small, medium, large, xl and a distilled version of the small checkpoint: distilgpt-2.
+
+
+GPT2Config
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.GPT2Config
     :members:
 
 
-``GPT2Tokenizer``
+GPT2Tokenizer
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.GPT2Tokenizer
+.. autoclass:: transformers.GPT2Tokenizer
+    :members: save_vocabulary
+
+
+GPT2Model
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.GPT2Model
     :members:
 
 
-``GPT2Model``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: pytorch_transformers.GPT2Model
-    :members:
-
-
-``GPT2LMHeadModel``
+GPT2LMHeadModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.GPT2LMHeadModel
+.. autoclass:: transformers.GPT2LMHeadModel
     :members:
 
 
-``GPT2DoubleHeadsModel``
+GPT2DoubleHeadsModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.GPT2DoubleHeadsModel
+.. autoclass:: transformers.GPT2DoubleHeadsModel
+    :members:
+
+
+TFGPT2Model
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFGPT2Model
+    :members:
+
+
+TFGPT2LMHeadModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFGPT2LMHeadModel
+    :members:
+
+
+TFGPT2DoubleHeadsModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFGPT2DoubleHeadsModel
     :members:

docs/source/model_doc/roberta.rst

@@ -1,36 +1,98 @@
 RoBERTa
 ----------------------------------------------------
 
-``RobertaConfig``
+The RoBERTa model was proposed in `RoBERTa: A Robustly Optimized BERT Pretraining Approach <https://arxiv.org/abs/1907.11692>`_
+by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer,
+Veselin Stoyanov. It is based on Google's BERT model released in 2018.
+
+It builds on BERT and modifies key hyperparameters, removing the next-sentence pretraining
+objective and training with much larger mini-batches and learning rates.
+
+The abstract from the paper is the following:
+
+*Language model pretraining has led to significant performance gains but careful comparison between different
+approaches is challenging. Training is computationally expensive, often done on private datasets of different sizes,
+and, as we will show, hyperparameter choices have significant impact on the final results. We present a replication
+study of BERT pretraining (Devlin et al., 2019) that carefully measures the impact of many key hyperparameters and
+training data size. We find that BERT was significantly undertrained, and can match or exceed the performance of
+every model published after it. Our best model achieves state-of-the-art results on GLUE, RACE and SQuAD. These
+results highlight the importance of previously overlooked design choices, and raise questions about the source
+of recently reported improvements. We release our models and code.*
+
+Tips:
+
+- This implementation is the same as :class:`~transformers.BertModel` with a tiny embeddings tweak as well as a
+  setup for Roberta pretrained models.
+- RoBERTa has the same architecture as BERT, but uses a byte-level BPE as a tokenizer (same as GPT-2) and uses a
+  different pre-training scheme.
+- RoBERTa doesn't have `token_type_ids`, you don't need to indicate which token belongs to which segment. Just separate your segments with the separation token `tokenizer.sep_token` (or `</s>`)
+- `Camembert <./camembert.html>`__ is a wrapper around RoBERTa. Refer to this page for usage examples.
+
+RobertaConfig
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.RobertaConfig
+.. autoclass:: transformers.RobertaConfig
     :members:
 
 
-``RobertaTokenizer``
+RobertaTokenizer
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.RobertaTokenizer
-    :members:
+.. autoclass:: transformers.RobertaTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
 
 
-``RobertaModel``
+RobertaModel
 ~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.RobertaModel
+.. autoclass:: transformers.RobertaModel
     :members:
 
 
-``RobertaForMaskedLM``
+RobertaForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.RobertaForMaskedLM
+.. autoclass:: transformers.RobertaForMaskedLM
     :members:
 
 
-``RobertaForSequenceClassification``
+RobertaForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.RobertaForSequenceClassification
+.. autoclass:: transformers.RobertaForSequenceClassification
+    :members:
+
+
+RobertaForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RobertaForTokenClassification
+    :members:
+
+TFRobertaModel
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRobertaModel
+    :members:
+
+
+TFRobertaForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRobertaForMaskedLM
+    :members:
+
+
+TFRobertaForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRobertaForSequenceClassification
+    :members:
+
+
+TFRobertaForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRobertaForTokenClassification
     :members:

docs/source/model_doc/t5.rst

@@ -0,0 +1,101 @@
+T5
+----------------------------------------------------
+**DISCLAIMER:** This model is still a work in progress, if you see something strange,
+file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`_
+
+Overview
+~~~~~
+The T5 model was presented in `Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer <https://arxiv.org/pdf/1910.10683.pdf>`_ by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu in 
+Here the abstract: 
+
+*Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. 
+In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. 
+Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. 
+By combining the insights from our exploration with scale and our new "Colossal Clean Crawled Corpus", we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. 
+To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.*
+
+The Authors' code can be found `here <https://github.com/google-research/text-to-text-transfer-transformer>`_ .
+
+Training
+~~~~~~~~~~~~~~~~~~~~
+T5 is an encoder-decoder model and converts all NLP problems into a text-to-text format. It is trained using teacher forcing.
+This means that for training we always need an input sequence and a target sequence. 
+The input sequence is fed to the model using ``input_ids``. The target sequence is shifted to the right, *i.e.* perprended by a start-sequence token and fed to the decoder using the `decoder_input_ids`. In teacher-forcing style, the target sequence is then appended by the EOS token and corresponds to the ``lm_labels``. The PAD token is hereby used as the start-sequence token.
+T5 can be trained / fine-tuned both in a supervised and unsupervised fashion.
+
+- Unsupervised denoising training
+  In this setup spans of the input sequence are masked by so-called sentinel tokens (*a.k.a* unique mask tokens) 
+  and the output sequence is formed as a concatenation of the same sentinel tokens and the *real* masked tokens. 
+  Each sentinel tokens represents a unique mask token for this sentence and should start with ``<extra_id_1>``, ``<extrac_id_2>``, ... up to ``<extra_id_100>``. As a default 100 sentinel tokens are available in ``T5Tokenizer``.
+  *E.g.* the sentence "The cute dog walks in the park" with the masks put on "cute dog" and "the" should be processed as follows: 
+
+::
+
+  input_ids = tokenizer.encode('The <extra_id_1> walks in <extra_id_2> park', return_tensors='pt')
+  lm_labels = tokenizer.encode('<extra_id_1> cute dog <extra_id_2> the <extra_id_3> </s>', return_tensors='pt')
+  # the forward function automatically creates the correct decoder_input_ids
+  model(input_ids=input_ids, lm_labels=lm_labels)
+
+- Supervised training
+  In this setup the input sequence and output sequence are standard sequence to sequence input output mapping.
+  In translation, *e.g.* the input sequence "The house is wonderful." and output sequence "Das Haus ist wunderbar." should 
+  be processed as follows:
+  
+::
+
+  input_ids = tokenizer.encode('translate English to German: The house is wonderful. </s>', return_tensors='pt')
+  lm_labels = tokenizer.encode('Das Haus ist wunderbar. </s>', return_tensors='pt')
+  # the forward function automatically creates the correct decoder_input_ids
+  model(input_ids=input_ids, lm_labels=lm_labels)
+
+Tips
+~~~~~~~~~~~~~~~~~~~~
+- T5 is an encoder-decoder model pre-trained on a multi-task mixture of unsupervised 
+  and supervised tasks and for which each task is converted into a text-to-text format.
+  T5 works well on a variety of tasks out-of-the-box by prepending a different prefix to the input corresponding to each task, e.g.: for translation: *translate English to German: ..., summarize: ...*.
+  For more information about which prefix to use, it is easiest to look into Appendix D of the `paper <https://arxiv.org/pdf/1910.10683.pdf>`_ .
+- For sequence to sequence generation, it is recommended to use ``T5ForConditionalGeneration.generate()``. The method takes care of feeding the encoded input via cross-attention layers to the decoder and auto-regressively generates the decoder output.
+- T5 uses relative scalar embeddings. Encoder input padding can be done on the left and on the right.
+
+
+T5Config
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.T5Config
+    :members:
+
+
+T5Tokenizer
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.T5Tokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+T5Model
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.T5Model
+    :members:
+
+
+T5ForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.T5ForConditionalGeneration
+    :members:
+
+
+TFT5Model
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFT5Model
+    :members:
+
+
+TFT5ForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFT5ForConditionalGeneration
+    :members:

docs/source/model_doc/transformerxl.rst

@@ -1,30 +1,73 @@
 Transformer XL
 ----------------------------------------------------
 
-
-``TransfoXLConfig``
+Overview
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.TransfoXLConfig
+The Transformer-XL model was proposed in
+`Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context <https://arxiv.org/abs/1901.02860>`__
+by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
+It's a causal (uni-directional) transformer with relative positioning (sinusoïdal) embeddings which can reuse
+previously computed hidden-states to attend to longer context (memory).
+This model also uses adaptive softmax inputs and outputs (tied).
+
+The abstract from the paper is the following:
+
+*Transformers have a potential of learning longer-term dependency, but are limited by a fixed-length context in the
+setting of language modeling. We propose a novel neural architecture Transformer-XL that enables learning dependency
+beyond a fixed length without disrupting temporal coherence. It consists of a segment-level recurrence mechanism and
+a novel positional encoding scheme. Our method not only enables capturing longer-term dependency, but also resolves
+the context fragmentation problem. As a result, Transformer-XL learns dependency that is 80% longer than RNNs and
+450% longer than vanilla Transformers, achieves better performance on both short and long sequences, and is up
+to 1,800+ times faster than vanilla Transformers during evaluation. Notably, we improve the state-of-the-art results
+of bpc/perplexity to 0.99 on enwiki8, 1.08 on text8, 18.3 on WikiText-103, 21.8 on One Billion Word, and 54.5 on
+Penn Treebank (without finetuning). When trained only on WikiText-103, Transformer-XL manages to generate reasonably
+coherent, novel text articles with thousands of tokens.*
+
+Tips:
+
+- Transformer-XL uses relative sinusoidal positional embeddings. Padding can be done on the left or on the right.
+  The original implementation trains on SQuAD with padding on the left, therefore the padding defaults are set to left.
+- Transformer-XL is one of the few models that has no sequence length limit.
+
+
+TransfoXLConfig
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TransfoXLConfig
     :members:
 
 
-``TransfoXLTokenizer``
+TransfoXLTokenizer
 ~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.TransfoXLTokenizer
-    :members:
+.. autoclass:: transformers.TransfoXLTokenizer
+    :members: save_vocabulary
 
 
-``TransfoXLModel``
+TransfoXLModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.TransfoXLModel
+.. autoclass:: transformers.TransfoXLModel
     :members:
 
 
-``TransfoXLLMHeadModel``
+TransfoXLLMHeadModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.TransfoXLLMHeadModel
+.. autoclass:: transformers.TransfoXLLMHeadModel
+    :members:
+
+
+TFTransfoXLModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFTransfoXLModel
+    :members:
+
+
+TFTransfoXLLMHeadModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFTransfoXLLMHeadModel
     :members:

docs/source/model_doc/xlm.rst

@@ -1,41 +1,107 @@
 XLM
 ----------------------------------------------------
 
-``XLMConfig``
+Overview
+~~~~~~~~~~~~~~~~~~~~~
+
+The XLM model was proposed in `Cross-lingual Language Model Pretraining <https://arxiv.org/abs/1901.07291>`_
+by Guillaume Lample*, Alexis Conneau*. It's a transformer pre-trained using one of the following objectives:
+
+- a causal language modeling (CLM) objective (next token prediction),
+- a masked language modeling (MLM) objective (Bert-like), or
+- a Translation Language Modeling (TLM) object (extension of Bert's MLM to multiple language inputs)
+
+The abstract from the paper is the following:
+
+*Recent studies have demonstrated the efficiency of generative pretraining for English natural language understanding.
+In this work, we extend this approach to multiple languages and show the effectiveness of cross-lingual pretraining.
+We propose two methods to learn cross-lingual language models (XLMs): one unsupervised that only relies on monolingual
+data, and one supervised that leverages parallel data with a new cross-lingual language model objective. We obtain
+state-of-the-art results on cross-lingual classification, unsupervised and supervised machine translation. On XNLI,
+our approach pushes the state of the art by an absolute gain of 4.9% accuracy. On unsupervised machine translation,
+we obtain 34.3 BLEU on WMT'16 German-English, improving the previous state of the art by more than 9 BLEU. On
+supervised machine translation, we obtain a new state of the art of 38.5 BLEU on WMT'16 Romanian-English, outperforming
+the previous best approach by more than 4 BLEU. Our code and pretrained models will be made publicly available.*
+
+Tips:
+
+- XLM has many different checkpoints, which were trained using different objectives: CLM, MLM or TLM. Make sure to
+  select the correct objective for your task (e.g. MLM checkpoints are not suitable for generation).
+- XLM has multilingual checkpoints which leverage a specific `lang` parameter. Check out the
+  `multi-lingual <../multilingual.html>`__ page for more information.
+
+
+XLMConfig
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLMConfig
+.. autoclass:: transformers.XLMConfig
     :members:
 
-``XLMTokenizer``
+XLMTokenizer
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLMTokenizer
-    :members:
+.. autoclass:: transformers.XLMTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
 
-``XLMModel``
+XLMModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLMModel
+.. autoclass:: transformers.XLMModel
     :members:
 
 
-``XLMWithLMHeadModel``
+XLMWithLMHeadModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLMWithLMHeadModel
+.. autoclass:: transformers.XLMWithLMHeadModel
     :members:
 
 
-``XLMForSequenceClassification``
+XLMForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLMForSequenceClassification
+.. autoclass:: transformers.XLMForSequenceClassification
     :members:
 
 
-``XLMForQuestionAnswering``
+XLMForQuestionAnsweringSimple
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLMForQuestionAnswering
+.. autoclass:: transformers.XLMForQuestionAnsweringSimple
+    :members:
+
+
+XLMForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMForQuestionAnswering
+    :members:
+
+
+TFXLMModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLMModel
+    :members:
+
+
+TFXLMWithLMHeadModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLMWithLMHeadModel
+    :members:
+
+
+TFXLMForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLMForSequenceClassification
+    :members:
+
+
+TFXLMForQuestionAnsweringSimple
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLMForQuestionAnsweringSimple
     :members:

docs/source/model_doc/xlmroberta.rst

@@ -0,0 +1,106 @@
+XLM-RoBERTa
+------------------------------------------
+
+The XLM-RoBERTa model was proposed in `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. 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.
+
+The abstract from the paper is the following:
+
+*This paper shows that pretraining multilingual language models at scale leads to significant performance gains for
+a wide range of cross-lingual transfer tasks. We train a Transformer-based masked language model on one hundred
+languages, using more than two terabytes of filtered CommonCrawl data. Our model, dubbed XLM-R, significantly
+outperforms multilingual BERT (mBERT) on a variety of cross-lingual benchmarks, including +13.8% average accuracy
+on XNLI, +12.3% average F1 score on MLQA, and +2.1% average F1 score on NER. XLM-R performs particularly well on
+low-resource languages, improving 11.8% in XNLI accuracy for Swahili and 9.2% for Urdu over the previous XLM model.
+We also present a detailed empirical evaluation of the key factors that are required to achieve these gains,
+including the trade-offs between (1) positive transfer and capacity dilution and (2) the performance of high and
+low resource languages at scale. Finally, we show, for the first time, the possibility of multilingual modeling
+without sacrificing per-language performance; XLM-Ris very competitive with strong monolingual models on the GLUE
+and XNLI benchmarks. We will make XLM-R code, data, and models publicly available.*
+
+Tips:
+
+- XLM-R is a multilingual model trained on 100 different languages. Unlike some XLM multilingual models, it does
+  not require `lang` tensors to understand which language is used, and should be able to determine the correct
+  language from the input ids.
+- This implementation is the same as RoBERTa. Refer to the `documentation of RoBERTa <./roberta.html>`__ for usage
+  examples as well as the information relative to the inputs and outputs.
+
+XLMRobertaConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMRobertaConfig
+    :members:
+
+
+XLMRobertaTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMRobertaTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+XLMRobertaModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMRobertaModel
+    :members:
+
+
+XLMRobertaForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMRobertaForMaskedLM
+    :members:
+
+
+XLMRobertaForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMRobertaForSequenceClassification
+    :members:
+
+
+XLMRobertaForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMRobertaForMultipleChoice
+    :members:
+
+
+XLMRobertaForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMRobertaForTokenClassification
+    :members:
+
+
+TFXLMRobertaModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLMRobertaModel
+    :members:
+
+
+TFXLMRobertaForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLMRobertaForMaskedLM
+    :members:
+
+
+TFXLMRobertaForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLMRobertaForSequenceClassification
+    :members:
+
+
+TFXLMRobertaForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLMRobertaForTokenClassification
+    :members:

docs/source/model_doc/xlnet.rst

@@ -1,43 +1,125 @@
 XLNet
 ----------------------------------------------------
 
-``XLNetConfig``
+Overview
+~~~~~~~~~~~~~~~~~~~~~
+
+The XLNet model was proposed in `XLNet: Generalized Autoregressive Pretraining for Language Understanding <https://arxiv.org/abs/1906.08237>`_
+by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
+XLnet is an extension of the Transformer-XL model pre-trained using an autoregressive method
+to learn bidirectional contexts by maximizing the expected likelihood over all permutations
+of the input sequence factorization order.
+
+The abstract from the paper is the following:
+
+*With the capability of modeling bidirectional contexts, denoising autoencoding based pretraining like BERT achieves
+better performance than pretraining approaches based on autoregressive language modeling. However, relying on
+corrupting the input with masks, BERT neglects dependency between the masked positions and suffers from a
+pretrain-finetune discrepancy. In light of these pros and cons, we propose XLNet, a generalized autoregressive
+pretraining method that (1) enables learning bidirectional contexts by maximizing the expected likelihood over
+all permutations of the factorization order and (2) overcomes the limitations of BERT thanks to its autoregressive
+formulation. Furthermore, XLNet integrates ideas from Transformer-XL, the state-of-the-art autoregressive model,
+into pretraining. Empirically, under comparable experiment settings, XLNet outperforms BERT on 20 tasks, often by
+a large margin, including question answering, natural language inference, sentiment analysis, and document ranking.*
+
+Tips:
+
+- The specific attention pattern can be controlled at training and test time using the `perm_mask` input.
+- Due to the difficulty of training a fully auto-regressive model over various factorization order,
+  XLNet is pretrained using only a sub-set of the output tokens as target which are selected
+  with the `target_mapping` input.
+- To use XLNet for sequential decoding (i.e. not in fully bi-directional setting), use the `perm_mask` and
+  `target_mapping` inputs to control the attention span and outputs (see examples in `examples/run_generation.py`)
+- XLNet is one of the few models that has no sequence length limit.
+
+
+XLNetConfig
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLNetConfig
+.. autoclass:: transformers.XLNetConfig
     :members:
 
 
-``XLNetTokenizer``
+XLNetTokenizer
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLNetTokenizer
-    :members:
+.. autoclass:: transformers.XLNetTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
 
 
-``XLNetModel``
+XLNetModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLNetModel
+.. autoclass:: transformers.XLNetModel
     :members:
 
 
-``XLNetLMHeadModel``
+XLNetLMHeadModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLNetLMHeadModel
+.. autoclass:: transformers.XLNetLMHeadModel
     :members:
 
 
-``XLNetForSequenceClassification``
+XLNetForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLNetForSequenceClassification
+.. autoclass:: transformers.XLNetForSequenceClassification
     :members:
 
 
-``XLNetForQuestionAnswering``
+XLNetForTokenClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: pytorch_transformers.XLNetForQuestionAnswering
+.. autoclass:: transformers.XLNetForTokenClassification
+    :members:
+
+
+XLNetForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLNetForMultipleChoice
+    :members:
+
+
+XLNetForQuestionAnsweringSimple
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLNetForQuestionAnsweringSimple
+    :members:
+
+
+XLNetForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLNetForQuestionAnswering
+    :members:
+
+
+TFXLNetModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLNetModel
+    :members:
+
+
+TFXLNetLMHeadModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLNetLMHeadModel
+    :members:
+
+
+TFXLNetForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLNetForSequenceClassification
+    :members:
+
+
+TFXLNetForQuestionAnsweringSimple
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFXLNetForQuestionAnsweringSimple
     :members:

docs/source/model_sharing.md

@@ -0,0 +1,55 @@
+# 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)**. Optionally, join an existing organization or create a new one. 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)
+```
+
+If you want your model to be namespaced by your organization name rather than your username, add the following flag to any command:
+```shell
+--organization organization_name
+```
+
+Your model will then be accessible through its identifier, a concatenation of your username (or organization name) and the folder name above:
+```python
+"username/pretrained_model"
+# or if an org:
+"organization_name/pretrained_model"
+```
+
+**Please add a README.md model card** to the repo under `model_cards/` with: model description, training params (dataset, preprocessing, hardware used, hyperparameters), evaluation results, intended uses & limitations, etc.
+
+Your model now has a page on huggingface.co/models 🔥
+
+Anyone can load it from code:
+```python
+tokenizer = AutoTokenizer.from_pretrained("namespace/pretrained_model")
+model = AutoModel.from_pretrained("namespace/pretrained_model")
+```
+
+List all your files on S3:
+```shell
+transformers-cli s3 ls
+```
+
+You can also delete unneeded files:
+
+```shell
+transformers-cli s3 rm …
+```

docs/source/multilingual.rst

@@ -0,0 +1,107 @@
+Multi-lingual models
+================================================
+
+Most of the models available in this library are mono-lingual models (English, Chinese and German). A few
+multi-lingual models are available and have a different mechanisms than mono-lingual models.
+This page details the usage of these models.
+
+The two models that currently support multiple languages are BERT and XLM.
+
+XLM
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+XLM has a total of 10 different checkpoints, only one of which is mono-lingual. The 9 remaining model checkpoints can
+be split in two categories: the checkpoints that make use of language embeddings, and those that don't
+
+XLM & Language Embeddings
+------------------------------------------------
+
+This section concerns the following checkpoints:
+
+- ``xlm-mlm-ende-1024`` (Masked language modeling, English-German)
+- ``xlm-mlm-enfr-1024`` (Masked language modeling, English-French)
+- ``xlm-mlm-enro-1024`` (Masked language modeling, English-Romanian)
+- ``xlm-mlm-xnli15-1024`` (Masked language modeling, XNLI languages)
+- ``xlm-mlm-tlm-xnli15-1024`` (Masked language modeling + Translation, XNLI languages)
+- ``xlm-clm-enfr-1024`` (Causal language modeling, English-French)
+- ``xlm-clm-ende-1024`` (Causal language modeling, English-German)
+
+These checkpoints require language embeddings that will specify the language used at inference time. These language
+embeddings are represented as a tensor that is of the same shape as the input ids passed to the model. The values in
+these tensors depend on the language used and are identifiable using the ``lang2id`` and ``id2lang`` attributes
+from the tokenizer.
+
+Here is an example using the ``xlm-clm-enfr-1024`` checkpoint (Causal language modeling, English-French):
+
+
+.. code-block::
+
+    import torch
+    from transformers import XLMTokenizer, XLMWithLMHeadModel
+
+    tokenizer = XLMTokenizer.from_pretrained("xlm-clm-1024-enfr")
+
+
+The different languages this model/tokenizer handles, as well as the ids of these languages are visible using the
+``lang2id`` attribute:
+
+.. code-block::
+
+    # Continuation of the previous script
+    print(tokenizer.lang2id)  # {'en': 0, 'fr': 1}
+
+
+These ids should be used when passing a language parameter during a model pass. Let's define our inputs:
+
+.. code-block::
+
+    # Continuation of the previous script
+    input_ids = torch.tensor([tokenizer.encode("Wikipedia was used to")]) # batch size of 1
+
+
+We should now define the language embedding by using the previously defined language id. We want to create a tensor
+filled with the appropriate language ids, of the same size as input_ids. For english, the id is 0:
+
+.. code-block::
+
+    # Continuation of the previous script
+    language_id = tokenizer.lang2id['en']  # 0
+    langs = torch.tensor([language_id] * input_ids.shape[1])  # torch.tensor([0, 0, 0, ..., 0])
+
+    # We reshape it to be of size (batch_size, sequence_length)
+    langs = langs.view(1, -1) # is now of shape [1, sequence_length] (we have a batch size of 1)
+
+
+You can then feed it all as input to your model:
+
+.. code-block::
+
+    # Continuation of the previous script
+    outputs = model(input_ids, langs=langs)
+
+
+The example `run_generation.py <https://github.com/huggingface/transformers/blob/master/examples/run_generation.py>`__
+can generate text using the CLM checkpoints from XLM, using the language embeddings.
+
+XLM without Language Embeddings
+------------------------------------------------
+
+This section concerns the following checkpoints:
+
+- ``xlm-mlm-17-1280`` (Masked language modeling, 17 languages)
+- ``xlm-mlm-100-1280`` (Masked language modeling, 100 languages)
+
+These checkpoints do not require language embeddings at inference time. These models are used to have generic
+sentence representations, differently from previously-mentioned XLM checkpoints.
+
+
+BERT
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+BERT has two checkpoints that can be used for multi-lingual tasks:
+
+- ``bert-base-multilingual-uncased`` (Masked language modeling + Next sentence prediction, 102 languages)
+- ``bert-base-multilingual-cased`` (Masked language modeling + Next sentence prediction, 104 languages)
+
+These checkpoints do not require language embeddings at inference time. They should identify the language
+used in the context and infer accordingly.

docs/source/notebooks.rst

@@ -1,16 +1,16 @@
 Notebooks
 ================================================
 
-We include `three Jupyter Notebooks <https://github.com/huggingface/pytorch-transformers/tree/master/notebooks>`_ that can be used to check that the predictions of the PyTorch model are identical to the predictions of the original TensorFlow model.
+We include `three Jupyter Notebooks <https://github.com/huggingface/transformers/tree/master/notebooks>`_ that can be used to check that the predictions of the PyTorch model are identical to the predictions of the original TensorFlow model.
 
 
 *
-  The first NoteBook (\ `Comparing-TF-and-PT-models.ipynb <https://github.com/huggingface/pytorch-transformers/blob/master/notebooks/Comparing-TF-and-PT-models.ipynb>`_\ ) extracts the hidden states of a full sequence on each layers of the TensorFlow and the PyTorch models and computes the standard deviation between them. In the given example, we get a standard deviation of 1.5e-7 to 9e-7 on the various hidden state of the models.
+  The first NoteBook (\ `Comparing-TF-and-PT-models.ipynb <https://github.com/huggingface/transformers/blob/master/notebooks/Comparing-TF-and-PT-models.ipynb>`_\ ) extracts the hidden states of a full sequence on each layers of the TensorFlow and the PyTorch models and computes the standard deviation between them. In the given example, we get a standard deviation of 1.5e-7 to 9e-7 on the various hidden state of the models.
 
 *
-  The second NoteBook (\ `Comparing-TF-and-PT-models-SQuAD.ipynb <https://github.com/huggingface/pytorch-transformers/blob/master/notebooks/Comparing-TF-and-PT-models-SQuAD.ipynb>`_\ ) compares the loss computed by the TensorFlow and the PyTorch models for identical initialization of the fine-tuning layer of the ``BertForQuestionAnswering`` and computes the standard deviation between them. In the given example, we get a standard deviation of 2.5e-7 between the models.
+  The second NoteBook (\ `Comparing-TF-and-PT-models-SQuAD.ipynb <https://github.com/huggingface/transformers/blob/master/notebooks/Comparing-TF-and-PT-models-SQuAD.ipynb>`_\ ) compares the loss computed by the TensorFlow and the PyTorch models for identical initialization of the fine-tuning layer of the ``BertForQuestionAnswering`` and computes the standard deviation between them. In the given example, we get a standard deviation of 2.5e-7 between the models.
 
 *
-  The third NoteBook (\ `Comparing-TF-and-PT-models-MLM-NSP.ipynb <https://github.com/huggingface/pytorch-transformers/blob/master/notebooks/Comparing-TF-and-PT-models-MLM-NSP.ipynb>`_\ ) compares the predictions computed by the TensorFlow and the PyTorch models for masked token language modeling using the pre-trained masked language modeling model.
+  The third NoteBook (\ `Comparing-TF-and-PT-models-MLM-NSP.ipynb <https://github.com/huggingface/transformers/blob/master/notebooks/Comparing-TF-and-PT-models-MLM-NSP.ipynb>`_\ ) compares the predictions computed by the TensorFlow and the PyTorch models for masked token language modeling using the pre-trained masked language modeling model.
 
 Please follow the instructions given in the notebooks to run and modify them.

docs/source/pretrained_models.rst

@@ -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                                                                                                                  |
@@ -44,15 +45,53 @@ Here is the full list of the currently provided pretrained models together with
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``bert-large-uncased-whole-word-masking-finetuned-squad``  | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
 |                   |                                                            | | The ``bert-large-uncased-whole-word-masking`` model fine-tuned on SQuAD                                                             |
-|                   |                                                            | (see details of fine-tuning in the `example section <https://github.com/huggingface/pytorch-transformers/tree/master/examples>`__).   |
+|                   |                                                            | (see details of fine-tuning in the `example section <https://github.com/huggingface/transformers/tree/master/examples>`__).           |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``bert-large-cased-whole-word-masking-finetuned-squad``    | | 24-layer, 1024-hidden, 16-heads, 340M parameters                                                                                    |
 |                   |                                                            | | The ``bert-large-cased-whole-word-masking`` model fine-tuned on SQuAD                                                               |
-|                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/pytorch-transformers/examples.html>`__)                   |
+|                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/transformers/examples.html>`__)                           |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``bert-base-cased-finetuned-mrpc``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | The ``bert-base-cased`` model fine-tuned on MRPC                                                                                    |
-|                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/pytorch-transformers/examples.html>`__)                   |
+|                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/transformers/examples.html>`__)                           |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-german-dbmdz-cased``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on cased German text by DBMDZ                                                                                               |
+|                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-german-dbmdz-uncased``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on uncased German text by DBMDZ                                                                                             |
+|                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``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/>`__).                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-dutch-cased``                                  | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on cased Dutch text.                                                                                                        |
+|                   |                                                            | (see `details on wietsedv repository <https://github.com/wietsedv/bertje/>`__).                                                       |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | GPT               | ``openai-gpt``                                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | OpenAI GPT English model                                                                                                            |
@@ -65,6 +104,9 @@ Here is the full list of the currently provided pretrained models together with
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``gpt2-large``                                             | | 36-layer, 1280-hidden, 20-heads, 774M parameters.                                                                                   |
 |                   |                                                            | | OpenAI's Large-sized GPT-2 English model                                                                                            |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``gpt2-xl``                                                | | 48-layer, 1600-hidden, 25-heads, 1558M parameters.                                                                                  |
+|                   |                                                            | | OpenAI's XL-sized GPT-2 English model                                                                                               |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | Transformer-XL    | ``transfo-xl-wt103``                                       | | 18-layer, 1024-hidden, 16-heads, 257M parameters.                                                                                   |
 |                   |                                                            | | English model trained on wikitext-103                                                                                               |
@@ -79,10 +121,10 @@ Here is the full list of the currently provided pretrained models together with
 |                   |                                                            | | XLM English model                                                                                                                   |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``xlm-mlm-ende-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
-|                   |                                                            | | XLM English-German Multi-language model                                                                                             |
+|                   |                                                            | | XLM English-German model trained on the concatenation of English and German wikipedia                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``xlm-mlm-enfr-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
-|                   |                                                            | | XLM English-French Multi-language model                                                                                             |
+|                   |                                                            | | XLM English-French model trained on the concatenation of English and French wikipedia                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``xlm-mlm-enro-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
 |                   |                                                            | | XLM English-Romanian Multi-language model                                                                                           |
@@ -93,11 +135,17 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``xlm-mlm-tlm-xnli15-1024``                                | | 12-layer, 1024-hidden, 8-heads                                                                                                      |
 |                   |                                                            | | XLM Model pre-trained with MLM + TLM on the `15 XNLI languages <https://github.com/facebookresearch/XNLI>`__.                       |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``xlm-clm-enfr-1024``                                      | | 12-layer, 1024-hidden, 8-heads                                                                                                      |
-|                   |                                                            | | XLM English model trained with CLM (Causal Language Modeling)                                                                       |
+|                   | ``xlm-clm-enfr-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
+|                   |                                                            | | XLM English-French model trained with CLM (Causal Language Modeling) on the concatenation of English and French wikipedia           |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``xlm-clm-ende-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
-|                   |                                                            | | XLM English-German Multi-language model trained with CLM (Causal Language Modeling)                                                 |
+|                   |                                                            | | XLM English-German model trained with CLM (Causal Language Modeling) on the concatenation of English and German wikipedia           |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlm-mlm-17-1280``                                        | | 16-layer, 1280-hidden, 16-heads                                                                                                     |
+|                   |                                                            | | XLM model trained with MLM (Masked Language Modeling) on 17 languages.                                                              |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlm-mlm-100-1280``                                       | | 16-layer, 1280-hidden, 16-heads                                                                                                     |
+|                   |                                                            | | XLM model trained with MLM (Masked Language Modeling) on 100 languages.                                                             |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | RoBERTa           | ``roberta-base``                                           | | 12-layer, 768-hidden, 12-heads, 125M parameters                                                                                     |
 |                   |                                                            | | RoBERTa using the BERT-base architecture                                                                                            |
@@ -110,14 +158,129 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``roberta-large-mnli``                                     | | 24-layer, 1024-hidden, 16-heads, 355M parameters                                                                                    |
 |                   |                                                            | | ``roberta-large`` fine-tuned on `MNLI <http://www.nyu.edu/projects/bowman/multinli/>`__.                                            |
 |                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__)                                                   |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``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>`__)                                               |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``roberta-large-openai-detector``                          | | 24-layer, 1024-hidden, 16-heads, 355M parameters                                                                                    |
+|                   |                                                            | | ``roberta-large`` fine-tuned by OpenAI on the outputs of the 1.5B-parameter GPT-2 model.                                            |
+|                   |                                                            | (see `details <https://github.com/openai/gpt-2-output-dataset/tree/master/detector>`__)                                               |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | DistilBERT        | ``distilbert-base-uncased``                                | | 6-layer, 768-hidden, 12-heads, 66M parameters                                                                                       |
 |                   |                                                            | | The DistilBERT model distilled from the BERT model `bert-base-uncased` checkpoint                                                   |
-|                   |                                                            | (see `details <https://medium.com/huggingface/distilbert-8cf3380435b5>`__)                                                            |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``distilbert-base-uncased-distilled-squad``                | | 6-layer, 768-hidden, 12-heads, 66M parameters                                                                                       |
 |                   |                                                            | | The DistilBERT model distilled from the BERT model `bert-base-uncased` checkpoint, with an additional linear layer.                 |
-|                   |                                                            | (see `details <https://medium.com/huggingface/distilbert-8cf3380435b5>`__)                                                            |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilbert-base-cased``                                  | | 6-layer, 768-hidden, 12-heads, 65M parameters                                                                                       |
+|                   |                                                            | | The DistilBERT model distilled from the BERT model `bert-base-cased` checkpoint                                                     |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilbert-base-cased-distilled-squad``                  | | 6-layer, 768-hidden, 12-heads, 65M parameters                                                                                       |
+|                   |                                                            | | The DistilBERT model distilled from the BERT model `bert-base-cased` checkpoint, with an additional question answering layer.       |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilgpt2``                                             | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
+|                   |                                                            | | The DistilGPT2 model distilled from the GPT2 model `gpt2` checkpoint.                                                               |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``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>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilbert-base-multilingual-cased``                     | | 6-layer, 768-hidden, 12-heads, 134M parameters                                                                                      |
+|                   |                                                            | | The multilingual DistilBERT model distilled from the Multilingual BERT model `bert-base-multilingual-cased` checkpoint.             |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| CTRL              | ``ctrl``                                                   | | 48-layer, 1280-hidden, 16-heads, 1.6B parameters                                                                                    |
+|                   |                                                            | | Salesforce's Large-sized CTRL English model                                                                                         |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| CamemBERT         | ``camembert-base``                                         | | 12-layer, 768-hidden, 12-heads, 110M parameters                                                                                     |
+|                   |                                                            | | CamemBERT using the BERT-base architecture                                                                                          |
+|                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/camembert>`__)                                                 |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| ALBERT            | ``albert-base-v1``                                         | | 12 repeating layers, 128 embedding, 768-hidden, 12-heads, 11M parameters                                                            |
+|                   |                                                            | | ALBERT base model                                                                                                                   |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``albert-large-v1``                                        | | 24 repeating layers, 128 embedding, 1024-hidden, 16-heads, 17M parameters                                                           |
+|                   |                                                            | | ALBERT large model                                                                                                                  |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``albert-xlarge-v1``                                       | | 24 repeating layers, 128 embedding, 2048-hidden, 16-heads, 58M parameters                                                           |
+|                   |                                                            | | ALBERT xlarge model                                                                                                                 |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``albert-xxlarge-v1``                                      | | 12 repeating layer, 128 embedding, 4096-hidden, 64-heads, 223M parameters                                                           |
+|                   |                                                            | | ALBERT xxlarge model                                                                                                                |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``albert-base-v2``                                         | | 12 repeating layers, 128 embedding, 768-hidden, 12-heads, 11M parameters                                                            |
+|                   |                                                            | | ALBERT base model with no dropout, additional training data and longer training                                                     |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``albert-large-v2``                                        | | 24 repeating layers, 128 embedding, 1024-hidden, 16-heads, 17M parameters                                                           |
+|                   |                                                            | | ALBERT large model with no dropout, additional training data and longer training                                                    |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``albert-xlarge-v2``                                       | | 24 repeating layers, 128 embedding, 2048-hidden, 16-heads, 58M parameters                                                           |
+|                   |                                                            | | ALBERT xlarge model with no dropout, additional training data and longer training                                                   |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``albert-xxlarge-v2``                                      | | 12 repeating layer, 128 embedding, 4096-hidden, 64-heads, 223M parameters                                                           |
+|                   |                                                            | | 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                                                          |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| FlauBERT          | ``flaubert-small-cased``                                   | | 6-layer, 512-hidden, 8-heads, 54M parameters                                                                                        |
+|                   |                                                            | | FlauBERT small architecture                                                                                                         |
+|                   |                                                            | (see `details <https://github.com/getalp/Flaubert>`__)                                                                                |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``flaubert-base-uncased``                                  | | 12-layer, 768-hidden, 12-heads, 137M parameters                                                                                     |
+|                   |                                                            | | FlauBERT base architecture with uncased vocabulary                                                                                  |
+|                   |                                                            | (see `details <https://github.com/getalp/Flaubert>`__)                                                                                |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``flaubert-base-cased``                                    | | 12-layer, 768-hidden, 12-heads, 138M parameters                                                                                     |
+|                   |                                                            | | FlauBERT base architecture with cased vocabulary                                                                                    |
+|                   |                                                            | (see `details <https://github.com/getalp/Flaubert>`__)                                                                                |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``flaubert-large-cased``                                   | | 24-layer, 1024-hidden, 16-heads, 373M parameters                                                                                    |
+|                   |                                                            | | FlauBERT large architecture                                                                                                         |
+|                   |                                                            | (see `details <https://github.com/getalp/Flaubert>`__)                                                                                |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| Bart              | ``bart-large``                                             | | 12-layer, 1024-hidden, 16-heads, 406M parameters                                                                                    |
+|                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_)                                                       |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bart-large-mnli``                                        | | Adds a 2 layer classification head with 1 million parameters                                                                        |
+|                   |                                                            | | bart-large base architecture with a classification head, finetuned on MNLI                                                          |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bart-large-cnn``                                         | | 12-layer, 1024-hidden, 16-heads, 406M parameters       (same as base)                                                               |
+|                   |                                                            | | bart-large base architecture finetuned on cnn summarization task                                                                    |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-
-.. <https://huggingface.co/pytorch-transformers/examples.html>`__

docs/source/quickstart.md

@@ -2,7 +2,7 @@
 
 ## Philosophy
 
-PyTorch-Transformers is an opinionated library built for NLP researchers seeking to use/study/extend large-scale transformers models.
+Transformers is an opinionated library built for NLP researchers seeking to use/study/extend large-scale transformers models.
 
 The library was designed with two strong goals in mind:
 
@@ -19,12 +19,12 @@ The library was designed with two strong goals in mind:
 
 A few other goals:
 
-- expose the models internals as consistently as possible:
+- expose the models' internals as consistently as possible:
 
   - we give access, using a single API to the full hidden-states and attention weights,
   - tokenizer and base model's API are standardized to easily switch between models.
 
-- incorporate a subjective selection of promising tools for fine-tuning/investiguating these models:
+- incorporate a subjective selection of promising tools for fine-tuning/investigating these models:
 
   - a simple/consistent way to add new tokens to the vocabulary and embeddings for fine-tuning,
   - simple ways to mask and prune transformer heads.
@@ -33,13 +33,13 @@ A few other goals:
 
 The library is build around three type of classes for each models:
 
-- **model classes** which are PyTorch models (`torch.nn.Modules`) of the 6 models architectures currently provided in the library, e.g. `BertModel`
+- **model classes** which are PyTorch models (`torch.nn.Modules`) of the 8 models architectures currently provided in the library, e.g. `BertModel`
 - **configuration classes** which store all the parameters required to build a model, e.g. `BertConfig`. You don't always need to instantiate these your-self, in particular if you are using a pretrained model without any modification, creating the model will automatically take care of instantiating the configuration (which is part of the model)
 - **tokenizer classes** which store the vocabulary for each model and provide methods for encoding/decoding strings in list of token embeddings indices to be fed to a model, e.g. `BertTokenizer`
 
 All these classes can be instantiated from pretrained instances and saved locally using two methods:
 
-- `from_pretrained()` let you instantiate a model/configuration/tokenizer from a pretrained version either provided by the library itself (currently 27 models are provided as listed [here](https://huggingface.co/pytorch-transformers/pretrained_models.html)) or stored locally (or on a server) by the user,
+- `from_pretrained()` let you instantiate a model/configuration/tokenizer from a pretrained version either provided by the library itself (currently 27 models are provided as listed [here](https://huggingface.co/transformers/pretrained_models.html)) or stored locally (or on a server) by the user,
 - `save_pretrained()` let you save a model/configuration/tokenizer locally so that it can be reloaded using `from_pretrained()`.
 
 We'll finish this quickstart tour by going through a few simple quick-start examples to see how we can instantiate and use these classes. The rest of the documentation is organized in two parts:
@@ -51,7 +51,7 @@ We'll finish this quickstart tour by going through a few simple quick-start exam
 
 Here are two examples showcasing a few `Bert` and `GPT2` classes and pre-trained models.
 
-See full API reference for examples for each model classe.
+See full API reference for examples for each model class.
 
 ### BERT example
 
@@ -59,7 +59,7 @@ Let's start by preparing a tokenized input (a list of token embeddings indices t
 
 ```python
 import torch
-from pytorch_transformers import BertTokenizer, BertModel, BertForMaskedLM
+from transformers import BertTokenizer, BertModel, BertForMaskedLM
 
 # OPTIONAL: if you want to have more information on what's happening under the hood, activate the logger as follows
 import logging
@@ -93,8 +93,8 @@ Let's see how we can use `BertModel` to encode our inputs in hidden-states:
 # Load pre-trained model (weights)
 model = BertModel.from_pretrained('bert-base-uncased')
 
-# Set the model in evaluation mode to desactivate the DropOut modules
-# This is IMPORTANT to have reproductible results during evaluation!
+# 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
@@ -106,7 +106,7 @@ model.to('cuda')
 with torch.no_grad():
     # See the models docstrings for the detail of the inputs
     outputs = model(tokens_tensor, token_type_ids=segments_tensors)
-    # PyTorch-Transformers models always output tuples.
+    # Transformers models always output tuples.
     # See the models docstrings for the detail of all the outputs
     # In our case, the first element is the hidden state of the last layer of the Bert model
     encoded_layers = outputs[0]
@@ -145,7 +145,7 @@ First let's prepare a tokenized input from our text string using `GPT2Tokenizer`
 
 ```python
 import torch
-from pytorch_transformers import GPT2Tokenizer, GPT2LMHeadModel
+from transformers import GPT2Tokenizer, GPT2LMHeadModel
 
 # OPTIONAL: if you want to have more information on what's happening, activate the logger as follows
 import logging
@@ -168,8 +168,8 @@ Let's see how to use `GPT2LMHeadModel` to generate the next token following our
 # Load pre-trained model (weights)
 model = GPT2LMHeadModel.from_pretrained('gpt2')
 
-# Set the model in evaluation mode to desactivate the DropOut modules
-# This is IMPORTANT to have reproductible results during evaluation!
+# 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
@@ -188,3 +188,35 @@ assert predicted_text == 'Who was Jim Henson? Jim Henson was a man'
 ```
 
 Examples for each model class of each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the [documentation](#documentation).
+
+#### Using the past
+
+GPT-2 as well as some other models (GPT, XLNet, Transfo-XL, CTRL) make use of a `past` or `mems` attribute which can be used to prevent re-computing the key/value pairs when using sequential decoding. It is useful when generating sequences as a big part of the attention mechanism benefits from previous computations.
+
+Here is a fully-working example using the `past` with `GPT2LMHeadModel` and argmax decoding (which should only be used as an example, as argmax decoding introduces a lot of repetition):
+
+```python
+from transformers import GPT2LMHeadModel, GPT2Tokenizer
+import torch
+
+tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+model = GPT2LMHeadModel.from_pretrained('gpt2')
+
+generated = tokenizer.encode("The Manhattan bridge")
+context = torch.tensor([generated])
+past = None
+
+for i in range(100):
+    print(i)
+    output, past = model(context, past=past)
+    token = torch.argmax(output[..., -1, :])
+
+    generated += [token.tolist()]
+    context = token.unsqueeze(0)
+
+sequence = tokenizer.decode(generated)
+
+print(sequence)
+```
+
+The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.

docs/source/serialization.rst

@@ -33,6 +33,8 @@ where
     * ``bert-large-uncased-whole-word-masking``: 24-layer, 1024-hidden, 16-heads, 340M parameters - Trained with Whole Word Masking (mask all of the the tokens corresponding to a word at once)
     * ``bert-large-cased-whole-word-masking``: 24-layer, 1024-hidden, 16-heads, 340M parameters - Trained with Whole Word Masking (mask all of the the tokens corresponding to a word at once)
     * ``bert-large-uncased-whole-word-masking-finetuned-squad``: The ``bert-large-uncased-whole-word-masking`` model finetuned on SQuAD (using the ``run_bert_squad.py`` examples). Results: *exact_match: 86.91579943235573, f1: 93.1532499015869*
+    * ``bert-base-german-dbmdz-cased``: Trained on German data only, 12-layer, 768-hidden, 12-heads, 110M parameters `Performance Evaluation <https://github.com/dbmdz/german-bert>`__
+    * ``bert-base-german-dbmdz-uncased``: Trained on (uncased) German data only, 12-layer, 768-hidden, 12-heads, 110M parameters `Performance Evaluation <https://github.com/dbmdz/german-bert>`__
     * ``openai-gpt``: OpenAI GPT English model, 12-layer, 768-hidden, 12-heads, 110M parameters
     * ``gpt2``: OpenAI GPT-2 English model, 12-layer, 768-hidden, 12-heads, 117M parameters
     * ``gpt2-medium``: OpenAI GPT-2 English model, 24-layer, 1024-hidden, 16-heads, 345M parameters
@@ -45,7 +47,7 @@ where
     * ``bert_config.json`` or ``openai_gpt_config.json`` a configuration file for the model, and
     * ``pytorch_model.bin`` a PyTorch dump of a pre-trained instance of ``BertForPreTraining``\ , ``OpenAIGPTModel``\ , ``TransfoXLModel``\ , ``GPT2LMHeadModel`` (saved with the usual ``torch.save()``\ )
 
-  If ``PRE_TRAINED_MODEL_NAME_OR_PATH`` is a shortcut name, the pre-trained weights will be downloaded from AWS S3 (see the links `here <https://github.com/huggingface/pytorch-transformers/blob/master/pytorch_transformers/modeling_bert.py>`__\ ) and stored in a cache folder to avoid future download (the cache folder can be found at ``~/.pytorch_pretrained_bert/``\ ).
+  If ``PRE_TRAINED_MODEL_NAME_OR_PATH`` is a shortcut name, the pre-trained weights will be downloaded from AWS S3 (see the links `here <https://github.com/huggingface/transformers/blob/master/transformers/modeling_bert.py>`__\ ) and stored in a cache folder to avoid future download (the cache folder can be found at ``~/.pytorch_pretrained_bert/``\ ).
 
 *
   ``cache_dir`` can be an optional path to a specific directory to download and cache the pre-trained model weights. This option is useful in particular when you are using distributed training: to avoid concurrent access to the same weights you can set for example ``cache_dir='./pretrained_model_{}'.format(args.local_rank)`` (see the section on distributed training for more information).
@@ -104,7 +106,7 @@ This section explain how you can save and re-load a fine-tuned model (BERT, GPT,
 There are three types of files you need to save to be able to reload a fine-tuned model:
 
 
-* the model it-self which should be saved following PyTorch serialization `best practices <https://pytorch.org/docs/stable/notes/serialization.html#best-practices>`__\ ,
+* the model itself which should be saved following PyTorch serialization `best practices <https://pytorch.org/docs/stable/notes/serialization.html#best-practices>`__\ ,
 * the configuration file of the model which is saved as a JSON file, and
 * the vocabulary (and the merges for the BPE-based models GPT and GPT-2).
 
@@ -122,7 +124,7 @@ Here is the recommended way of saving the model, configuration and vocabulary to
 
 .. code-block:: python
 
-   from pytorch_transformers import WEIGHTS_NAME, CONFIG_NAME
+   from transformers import WEIGHTS_NAME, CONFIG_NAME
 
    output_dir = "./models/"
 

docs/source/torchscript.rst

@@ -12,7 +12,7 @@ According to Pytorch's documentation: "TorchScript is a way to create serializab
 Pytorch's two modules `JIT and TRACE <https://pytorch.org/docs/stable/jit.html>`_ allow the developer to export
 their model to be re-used in other programs, such as efficiency-oriented C++ programs.
 
-We have provided an interface that allows the export of `pytorch-transformers` models to TorchScript so that they can
+We have provided an interface that allows the export of `transformers` models to TorchScript so that they can
 be reused in a different environment than a Pytorch-based python program. Here we explain how to use our models so that
 they can be exported, and what to be mindful of when using these models with TorchScript.
 
@@ -74,7 +74,7 @@ according to a ``BertConfig`` class and then saved to disk under the filename ``
 
 .. code-block:: python
 
-    from pytorch_transformers import BertModel, BertTokenizer, BertConfig
+    from transformers import BertModel, BertTokenizer, BertConfig
     import torch
 
     enc = BertTokenizer.from_pretrained("bert-base-uncased")

docs/source/usage.rst

@@ -0,0 +1,727 @@
+Usage
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+This page shows the most frequent use-cases when using the library. The models available allow for many different
+configurations and a great versatility in use-cases. The most simple ones are presented here, showcasing usage
+for tasks such as question answering, sequence classification, named entity recognition and others.
+
+These examples leverage auto-models, which are classes that will instantiate a model according to a given checkpoint,
+automatically selecting the correct model architecture. Please check the :class:`~transformers.AutoModel` documentation
+for more information.
+Feel free to modify the code to be more specific and adapt it to your specific use-case.
+
+In order for a model to perform well on a task, it must be loaded from a checkpoint corresponding to that task. These
+checkpoints are usually pre-trained on a large corpus of data and fine-tuned on a specific task. This means the
+following:
+
+- Not all models were fine-tuned on all tasks. If you want to fine-tune a model on a specific task, you can leverage
+  one of the `run_$TASK.py` script in the
+  `examples <https://github.com/huggingface/transformers/tree/master/examples>`_ directory.
+- Fine-tuned models were fine-tuned on a specific dataset. This dataset may or may not overlap with your use-case
+  and domain. As mentioned previously, you may leverage the
+  `examples <https://github.com/huggingface/transformers/tree/master/examples>`_ scripts to fine-tune your model, or you
+  may create your own training script.
+
+In order to do an inference on a task, several mechanisms are made available by the library:
+
+- Pipelines: very easy-to-use abstractions, which require as little as two lines of code.
+- Using a model directly with a tokenizer (PyTorch/TensorFlow): the full inference using the model. Less abstraction,
+  but much more powerful.
+
+Both approaches are showcased here.
+
+.. note::
+
+    All tasks presented here leverage pre-trained checkpoints that were fine-tuned on specific tasks. Loading a
+    checkpoint that was not fine-tuned on a specific task would load only the base transformer layers and not the
+    additional head that is used for the task, initializing the weights of that head randomly.
+
+    This would produce random output.
+
+Sequence Classification
+--------------------------
+
+Sequence classification is the task of classifying sequences according to a given number of classes. An example
+of sequence classification is the GLUE dataset, which is entirely based on that task. If you would like to fine-tune
+a model on a GLUE sequence classification task, you may leverage the
+`run_glue.py <https://github.com/huggingface/transformers/tree/master/examples/run_glue.py>`_ or
+`run_tf_glue.py <https://github.com/huggingface/transformers/tree/master/examples/run_tf_glue.py>`_ scripts.
+
+Here is an example using the pipelines do to sentiment analysis: identifying if a sequence is positive or negative.
+It leverages a fine-tuned model on sst2, which is a GLUE task.
+
+::
+
+    from transformers import pipeline
+
+    nlp = pipeline("sentiment-analysis")
+
+    print(nlp("I hate you"))
+    print(nlp("I love you"))
+
+This returns a label ("POSITIVE" or "NEGATIVE") alongside a score, as follows:
+
+::
+
+    [{'label': 'NEGATIVE', 'score': 0.9991129}]
+    [{'label': 'POSITIVE', 'score': 0.99986565}]
+
+
+Here is an example of doing a sequence classification using a model to determine if two sequences are paraphrases
+of each other. The process is the following:
+
+- Instantiate a tokenizer and a model from the checkpoint name. The model is identified as a BERT model and loads it
+  with the weights stored in the checkpoint.
+- Build a sequence from the two sentences, with the correct model-specific separators token type ids
+  and attention masks (:func:`~transformers.PreTrainedTokenizer.encode` and
+  :func:`~transformers.PreTrainedTokenizer.encode_plus` take care of this)
+- Pass this sequence through the model so that it is classified in one of the two available classes: 0
+  (not a paraphrase) and 1 (is a paraphrase)
+- Compute the softmax of the result to get probabilities over the classes
+- Print the results
+
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoTokenizer, AutoModelForSequenceClassification
+    import torch
+
+    tokenizer = AutoTokenizer.from_pretrained("bert-base-cased-finetuned-mrpc")
+    model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased-finetuned-mrpc")
+
+    classes = ["not paraphrase", "is paraphrase"]
+
+    sequence_0 = "The company HuggingFace is based in New York City"
+    sequence_1 = "Apples are especially bad for your health"
+    sequence_2 = "HuggingFace's headquarters are situated in Manhattan"
+
+    paraphrase = tokenizer.encode_plus(sequence_0, sequence_2, return_tensors="pt")
+    not_paraphrase = tokenizer.encode_plus(sequence_0, sequence_1, return_tensors="pt")
+
+    paraphrase_classification_logits = model(**paraphrase)[0]
+    not_paraphrase_classification_logits = model(**not_paraphrase)[0]
+
+    paraphrase_results = torch.softmax(paraphrase_classification_logits, dim=1).tolist()[0]
+    not_paraphrase_results = torch.softmax(not_paraphrase_classification_logits, dim=1).tolist()[0]
+
+    print("Should be paraphrase")
+    for i in range(len(classes)):
+        print(f"{classes[i]}: {round(paraphrase_results[i] * 100)}%")
+
+    print("\nShould not be paraphrase")
+    for i in range(len(classes)):
+        print(f"{classes[i]}: {round(not_paraphrase_results[i] * 100)}%")
+    ## TENSORFLOW CODE
+    from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
+    import tensorflow as tf
+
+    tokenizer = AutoTokenizer.from_pretrained("bert-base-cased-finetuned-mrpc")
+    model = TFAutoModelForSequenceClassification.from_pretrained("bert-base-cased-finetuned-mrpc")
+
+    classes = ["not paraphrase", "is paraphrase"]
+
+    sequence_0 = "The company HuggingFace is based in New York City"
+    sequence_1 = "Apples are especially bad for your health"
+    sequence_2 = "HuggingFace's headquarters are situated in Manhattan"
+
+    paraphrase = tokenizer.encode_plus(sequence_0, sequence_2, return_tensors="tf")
+    not_paraphrase = tokenizer.encode_plus(sequence_0, sequence_1, return_tensors="tf")
+
+    paraphrase_classification_logits = model(paraphrase)[0]
+    not_paraphrase_classification_logits = model(not_paraphrase)[0]
+
+    paraphrase_results = tf.nn.softmax(paraphrase_classification_logits, axis=1).numpy()[0]
+    not_paraphrase_results = tf.nn.softmax(not_paraphrase_classification_logits, axis=1).numpy()[0]
+
+    print("Should be paraphrase")
+    for i in range(len(classes)):
+        print(f"{classes[i]}: {round(paraphrase_results[i] * 100)}%")
+
+    print("\nShould not be paraphrase")
+    for i in range(len(classes)):
+        print(f"{classes[i]}: {round(not_paraphrase_results[i] * 100)}%")
+
+This outputs the following results:
+
+::
+
+    Should be paraphrase
+    not paraphrase: 10%
+    is paraphrase: 90%
+
+    Should not be paraphrase
+    not paraphrase: 94%
+    is paraphrase: 6%
+
+Extractive Question Answering
+----------------------------------------------------
+
+Extractive Question Answering is the task of extracting an answer from a text given a question. An example of a
+question answering dataset is the SQuAD dataset, which is entirely based on that task. If you would like to fine-tune
+a model on a SQuAD task, you may leverage the `run_squad.py`.
+
+Here is an example using the pipelines do to question answering: extracting an answer from a text given a question.
+It leverages a fine-tuned model on SQuAD.
+
+::
+
+    from transformers import pipeline
+
+    nlp = pipeline("question-answering")
+
+    context = r"""
+    Extractive Question Answering is the task of extracting an answer from a text given a question. An example of a
+    question answering dataset is the SQuAD dataset, which is entirely based on that task. If you would like to fine-tune
+    a model on a SQuAD task, you may leverage the `run_squad.py`.
+    """
+
+    print(nlp(question="What is extractive question answering?", context=context))
+    print(nlp(question="What is a good example of a question answering dataset?", context=context))
+
+This returns an answer extracted from the text, a confidence score, alongside "start" and "end" values which
+are the positions of the extracted answer in the text.
+
+::
+
+    {'score': 0.622232091629833, 'start': 34, 'end': 96, 'answer': 'the task of extracting an answer from a text given a question.'}
+    {'score': 0.5115299158662765, 'start': 147, 'end': 161, 'answer': 'SQuAD dataset,'}
+
+
+Here is an example of question answering using a model and a tokenizer. The process is the following:
+
+- Instantiate a tokenizer and a model from the checkpoint name. The model is identified as a BERT model and loads it
+  with the weights stored in the checkpoint.
+- Define a text and a few questions.
+- Iterate over the questions and build a sequence from the text and the current question, with the correct
+  model-specific separators token type ids and attention masks
+- Pass this sequence through the model. This outputs a range of scores across the entire sequence tokens (question and
+  text), for both the start and end positions.
+- Compute the softmax of the result to get probabilities over the tokens
+- Fetch the tokens from the identified start and stop values, convert those tokens to a string.
+- Print the results
+
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoTokenizer, AutoModelForQuestionAnswering
+    import torch
+
+    tokenizer = AutoTokenizer.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
+    model = AutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
+
+    text = r"""
+    🤗 Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides general-purpose
+    architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet…) for Natural Language Understanding (NLU) and Natural
+    Language Generation (NLG) with over 32+ pretrained models in 100+ languages and deep interoperability between
+    TensorFlow 2.0 and PyTorch.
+    """
+
+    questions = [
+        "How many pretrained models are available in Transformers?",
+        "What does Transformers provide?",
+        "Transformers provides interoperability between which frameworks?",
+    ]
+
+    for question in questions:
+        inputs = tokenizer.encode_plus(question, text, add_special_tokens=True, return_tensors="pt")
+        input_ids = inputs["input_ids"].tolist()[0]
+
+        text_tokens = tokenizer.convert_ids_to_tokens(input_ids)
+        answer_start_scores, answer_end_scores = model(**inputs)
+
+        answer_start = torch.argmax(
+            answer_start_scores
+        )  # Get the most likely beginning of answer with the argmax of the score
+        answer_end = torch.argmax(answer_end_scores) + 1  # Get the most likely end of answer with the argmax of the score
+
+        answer = tokenizer.convert_tokens_to_string(tokenizer.convert_ids_to_tokens(input_ids[answer_start:answer_end]))
+
+        print(f"Question: {question}")
+        print(f"Answer: {answer}\n")
+    ## TENSORFLOW CODE
+    from transformers import AutoTokenizer, TFAutoModelForQuestionAnswering
+    import tensorflow as tf
+
+    tokenizer = AutoTokenizer.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
+    model = TFAutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
+
+    text = r"""
+    🤗 Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides general-purpose
+    architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet…) for Natural Language Understanding (NLU) and Natural
+    Language Generation (NLG) with over 32+ pretrained models in 100+ languages and deep interoperability between
+    TensorFlow 2.0 and PyTorch.
+    """
+
+    questions = [
+        "How many pretrained models are available in Transformers?",
+        "What does Transformers provide?",
+        "Transformers provides interoperability between which frameworks?",
+    ]
+
+    for question in questions:
+        inputs = tokenizer.encode_plus(question, text, add_special_tokens=True, return_tensors="tf")
+        input_ids = inputs["input_ids"].numpy()[0]
+
+        text_tokens = tokenizer.convert_ids_to_tokens(input_ids)
+        answer_start_scores, answer_end_scores = model(inputs)
+
+        answer_start = tf.argmax(
+            answer_start_scores, axis=1
+        ).numpy()[0]  # Get the most likely beginning of answer with the argmax of the score
+        answer_end = (
+            tf.argmax(answer_end_scores, axis=1) + 1
+        ).numpy()[0]  # Get the most likely end of answer with the argmax of the score
+        answer = tokenizer.convert_tokens_to_string(tokenizer.convert_ids_to_tokens(input_ids[answer_start:answer_end]))
+
+        print(f"Question: {question}")
+        print(f"Answer: {answer}\n")
+
+This outputs the questions followed by the predicted answers:
+
+::
+
+    Question: How many pretrained models are available in Transformers?
+    Answer: over 32 +
+
+    Question: What does Transformers provide?
+    Answer: general - purpose architectures
+
+    Question: Transformers provides interoperability between which frameworks?
+    Answer: tensorflow 2 . 0 and pytorch
+
+
+
+Language Modeling
+----------------------------------------------------
+
+Language modeling is the task of fitting a model to a corpus, which can be domain specific. All popular transformer
+based models are trained using a variant of language modeling, e.g. BERT with masked language modeling, GPT-2 with
+causal language modeling.
+
+Language modeling can be useful outside of pre-training as well, for example to shift the model distribution to be
+domain-specific: using a language model trained over a very large corpus, and then fine-tuning it to a news dataset
+or on scientific papers e.g. `LysandreJik/arxiv-nlp <https://huggingface.co/lysandre/arxiv-nlp>`__.
+
+Masked Language Modeling
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Masked language modeling is the task of masking tokens in a sequence with a masking token, and prompting the model to
+fill that mask with an appropriate token. This allows the model to attend to both the right context (tokens on the
+right of the mask) and the left context (tokens on the left of the mask). Such a training creates a strong basis
+for downstream tasks requiring bi-directional context such as SQuAD (question answering,
+see `Lewis, Lui, Goyal et al. <https://arxiv.org/abs/1910.13461>`__, part 4.2).
+
+Here is an example of using pipelines to replace a mask from a sequence:
+
+::
+
+    from transformers import pipeline
+
+    nlp = pipeline("fill-mask")
+    print(nlp(f"HuggingFace is creating a {nlp.tokenizer.mask_token} that the community uses to solve NLP tasks."))
+
+This outputs the sequences with the mask filled, the confidence score as well as the token id in the tokenizer
+vocabulary:
+
+::
+
+    [
+        {'sequence': '<s> HuggingFace is creating a tool that the community uses to solve NLP tasks.</s>', 'score': 0.15627853572368622, 'token': 3944},
+        {'sequence': '<s> HuggingFace is creating a framework that the community uses to solve NLP tasks.</s>', 'score': 0.11690319329500198, 'token': 7208},
+        {'sequence': '<s> HuggingFace is creating a library that the community uses to solve NLP tasks.</s>', 'score': 0.058063216507434845, 'token': 5560},
+        {'sequence': '<s> HuggingFace is creating a database that the community uses to solve NLP tasks.</s>', 'score': 0.04211743175983429, 'token': 8503},
+        {'sequence': '<s> HuggingFace is creating a prototype that the community uses to solve NLP tasks.</s>', 'score': 0.024718601256608963, 'token': 17715}
+    ]
+
+Here is an example doing masked language modeling using a model and a tokenizer. The process is the following:
+
+- Instantiate a tokenizer and a model from the checkpoint name. The model is identified as a DistilBERT model and
+  loads it with the weights stored in the checkpoint.
+- Define a sequence with a masked token, placing the :obj:`tokenizer.mask_token` instead of a word.
+- Encode that sequence into IDs and find the position of the masked token in that list of IDs.
+- Retrieve the predictions at the index of the mask token: this tensor has the same size as the vocabulary, and the
+  values are the scores attributed to each token. The model gives higher score to tokens he deems probable in that
+  context.
+- Retrieve the top 5 tokens using the PyTorch :obj:`topk` or TensorFlow :obj:`top_k` methods.
+- Replace the mask token by the tokens and print the results
+
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoModelWithLMHead, AutoTokenizer
+    import torch
+
+    tokenizer = AutoTokenizer.from_pretrained("distilbert-base-cased")
+    model = AutoModelWithLMHead.from_pretrained("distilbert-base-cased")
+
+    sequence = f"Distilled models are smaller than the models they mimic. Using them instead of the large versions would help {tokenizer.mask_token} our carbon footprint."
+
+    input = tokenizer.encode(sequence, return_tensors="pt")
+    mask_token_index = torch.where(input == tokenizer.mask_token_id)[1]
+
+    token_logits = model(input)[0]
+    mask_token_logits = token_logits[0, mask_token_index, :]
+
+    top_5_tokens = torch.topk(mask_token_logits, 5, dim=1).indices[0].tolist()
+
+    for token in top_5_tokens:
+        print(sequence.replace(tokenizer.mask_token, tokenizer.decode([token])))
+    ## TENSORFLOW CODE
+    from transformers import TFAutoModelWithLMHead, AutoTokenizer
+    import tensorflow as tf
+
+    tokenizer = AutoTokenizer.from_pretrained("distilbert-base-cased")
+    model = TFAutoModelWithLMHead.from_pretrained("distilbert-base-cased")
+
+    sequence = f"Distilled models are smaller than the models they mimic. Using them instead of the large versions would help {tokenizer.mask_token} our carbon footprint."
+
+    input = tokenizer.encode(sequence, return_tensors="tf")
+    mask_token_index = tf.where(input == tokenizer.mask_token_id)[0, 1]
+
+    token_logits = model(input)[0]
+    mask_token_logits = token_logits[0, mask_token_index, :]
+
+    top_5_tokens = tf.math.top_k(mask_token_logits, 5).indices.numpy()
+
+    for token in top_5_tokens:
+        print(sequence.replace(tokenizer.mask_token, tokenizer.decode([token])))
+
+This prints five sequences, with the top 5 tokens predicted by the model:
+
+::
+
+    Distilled models are smaller than the models they mimic. Using them instead of the large versions would help reduce our carbon footprint.
+    Distilled models are smaller than the models they mimic. Using them instead of the large versions would help increase our carbon footprint.
+    Distilled models are smaller than the models they mimic. Using them instead of the large versions would help decrease our carbon footprint.
+    Distilled models are smaller than the models they mimic. Using them instead of the large versions would help offset our carbon footprint.
+    Distilled models are smaller than the models they mimic. Using them instead of the large versions would help improve our carbon footprint.
+
+
+Causal Language Modeling
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Causal language modeling is the task of predicting the token following a sequence of tokens. In this situation, the
+model only attends to the left context (tokens on the left of the mask). Such a training is particularly interesting
+for generation tasks.
+
+There is currently no pipeline to do causal language modeling/generation.
+
+Here is an example using the tokenizer and model. leveraging the :func:`~transformers.PreTrainedModel.generate` method
+to generate the tokens following the initial sequence in PyTorch, and creating a simple loop in TensorFlow.
+
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoModelWithLMHead, AutoTokenizer
+
+    tokenizer = AutoTokenizer.from_pretrained("gpt2")
+    model = AutoModelWithLMHead.from_pretrained("gpt2")
+
+    sequence = f"Hugging Face is based in DUMBO, New York City, and is"
+
+    input = tokenizer.encode(sequence, return_tensors="pt")
+    generated = model.generate(input, max_length=50, do_sample=True)
+
+    resulting_string = tokenizer.decode(generated.tolist()[0])
+    print(resulting_string)
+    ## TENSORFLOW CODE
+    from transformers import TFAutoModelWithLMHead, AutoTokenizer
+    import tensorflow as tf
+
+    tokenizer = AutoTokenizer.from_pretrained("gpt2")
+    model = TFAutoModelWithLMHead.from_pretrained("gpt2")
+
+    sequence = f"Hugging Face is based in DUMBO, New York City, and is"
+    input = tokenizer.encode(sequence, return_tensors="tf")
+    generated = model.generate(input, max_length=50, do_sample=True)
+
+    resulting_string = tokenizer.decode(generated.tolist()[0])
+    print(resulting_string)
+
+
+This outputs a (hopefully) coherent string from the original sequence, as the
+:func:`~transformers.PreTrainedModel.generate` samples from a top_p/tok_k distribution:
+
+::
+
+    Hugging Face is based in DUMBO, New York City, and is a live-action TV series based on the novel by John
+    Carpenter, and its producers, David Kustlin and Steve Pichar. The film is directed by!
+
+
+Named Entity Recognition
+----------------------------------------------------
+
+Named Entity Recognition (NER) is the task of classifying tokens according to a class, for example identifying a
+token as a person, an organisation or a location.
+An example of a named entity recognition dataset is the CoNLL-2003 dataset, which is entirely based on that task.
+If you would like to fine-tune a model on an NER task, you may leverage the `ner/run_ner.py` (PyTorch),
+`ner/run_pl_ner.py` (leveraging pytorch-lightning) or the `ner/run_tf_ner.py` (TensorFlow) scripts.
+
+Here is an example using the pipelines do to named entity recognition, trying to identify tokens as belonging to one
+of 9 classes:
+
+- O, Outside of a named entity
+- B-MIS, Beginning of a miscellaneous entity right after another miscellaneous entity
+- I-MIS, Miscellaneous entity
+- B-PER, Beginning of a person's name right after another person's name
+- I-PER, Person's name
+- B-ORG, Beginning of an organisation right after another organisation
+- I-ORG, Organisation
+- B-LOC, Beginning of a location right after another location
+- I-LOC, Location
+
+It leverages a fine-tuned model on CoNLL-2003, fine-tuned by `@stefan-it <https://github.com/stefan-it>`__ from
+`dbmdz <https://github.com/dbmdz>`__.
+
+::
+
+    from transformers import pipeline
+
+    nlp = pipeline("ner")
+
+    sequence = "Hugging Face Inc. is a company based in New York City. Its headquarters are in DUMBO, therefore very" \
+               "close to the Manhattan Bridge which is visible from the window."
+
+    print(nlp(sequence))
+
+This outputs a list of all words that have been identified as an entity from the 9 classes defined above. Here is the
+expected results:
+
+::
+
+    [
+        {'word': 'Hu', 'score': 0.9995632767677307, 'entity': 'I-ORG'},
+        {'word': '##gging', 'score': 0.9915938973426819, 'entity': 'I-ORG'},
+        {'word': 'Face', 'score': 0.9982671737670898, 'entity': 'I-ORG'},
+        {'word': 'Inc', 'score': 0.9994403719902039, 'entity': 'I-ORG'},
+        {'word': 'New', 'score': 0.9994346499443054, 'entity': 'I-LOC'},
+        {'word': 'York', 'score': 0.9993270635604858, 'entity': 'I-LOC'},
+        {'word': 'City', 'score': 0.9993864893913269, 'entity': 'I-LOC'},
+        {'word': 'D', 'score': 0.9825621843338013, 'entity': 'I-LOC'},
+        {'word': '##UM', 'score': 0.936983048915863, 'entity': 'I-LOC'},
+        {'word': '##BO', 'score': 0.8987102508544922, 'entity': 'I-LOC'},
+        {'word': 'Manhattan', 'score': 0.9758241176605225, 'entity': 'I-LOC'},
+        {'word': 'Bridge', 'score': 0.990249514579773, 'entity': 'I-LOC'}
+    ]
+
+Note how the words "Hugging Face" have been identified as an organisation, and "New York City", "DUMBO" and
+"Manhattan Bridge" have been identified as locations.
+
+Here is an example doing named entity recognition using a model and a tokenizer. The process is the following:
+
+- Instantiate a tokenizer and a model from the checkpoint name. The model is identified as a BERT model and
+  loads it with the weights stored in the checkpoint.
+- Define the label list with which the model was trained on.
+- Define a sequence with known entities, such as "Hugging Face" as an organisation and "New York City" as a location.
+- Split words into tokens so that they can be mapped to the predictions. We use a small hack by firstly completely
+  encoding and decoding the sequence, so that we're left with a string that contains the special tokens.
+- Encode that sequence into IDs (special tokens are added automatically).
+- Retrieve the predictions by passing the input to the model and getting the first output. This results in a
+  distribution over the 9 possible classes for each token. We take the argmax to retrieve the most likely class
+  for each token.
+- Zip together each token with its prediction and print it.
+
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoModelForTokenClassification, AutoTokenizer
+    import torch
+
+    model = AutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english")
+    tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
+
+    label_list = [
+        "O",       # Outside of a named entity
+        "B-MISC",  # Beginning of a miscellaneous entity right after another miscellaneous entity
+        "I-MISC",  # Miscellaneous entity
+        "B-PER",   # Beginning of a person's name right after another person's name
+        "I-PER",   # Person's name
+        "B-ORG",   # Beginning of an organisation right after another organisation
+        "I-ORG",   # Organisation
+        "B-LOC",   # Beginning of a location right after another location
+        "I-LOC"    # Location
+    ]
+
+    sequence = "Hugging Face Inc. is a company based in New York City. Its headquarters are in DUMBO, therefore very" \
+               "close to the Manhattan Bridge."
+
+    # Bit of a hack to get the tokens with the special tokens
+    tokens = tokenizer.tokenize(tokenizer.decode(tokenizer.encode(sequence)))
+    inputs = tokenizer.encode(sequence, return_tensors="pt")
+
+    outputs = model(inputs)[0]
+    predictions = torch.argmax(outputs, dim=2)
+
+    print([(token, label_list[prediction]) for token, prediction in zip(tokens, predictions[0].tolist())])
+    ## TENSORFLOW CODE
+    from transformers import TFAutoModelForTokenClassification, AutoTokenizer
+    import tensorflow as tf
+
+    model = TFAutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english")
+    tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
+
+    label_list = [
+        "O",       # Outside of a named entity
+        "B-MISC",  # Beginning of a miscellaneous entity right after another miscellaneous entity
+        "I-MISC",  # Miscellaneous entity
+        "B-PER",   # Beginning of a person's name right after another person's name
+        "I-PER",   # Person's name
+        "B-ORG",   # Beginning of an organisation right after another organisation
+        "I-ORG",   # Organisation
+        "B-LOC",   # Beginning of a location right after another location
+        "I-LOC"    # Location
+    ]
+
+    sequence = "Hugging Face Inc. is a company based in New York City. Its headquarters are in DUMBO, therefore very" \
+               "close to the Manhattan Bridge."
+
+    # Bit of a hack to get the tokens with the special tokens
+    tokens = tokenizer.tokenize(tokenizer.decode(tokenizer.encode(sequence)))
+    inputs = tokenizer.encode(sequence, return_tensors="tf")
+
+    outputs = model(inputs)[0]
+    predictions = tf.argmax(outputs, axis=2)
+
+    print([(token, label_list[prediction]) for token, prediction in zip(tokens, predictions[0].numpy())])
+
+This outputs a list of each token mapped to their prediction. Differently from the pipeline, here every token has
+a prediction as we didn't remove the "0" class which means that no particular entity was found on that token. The
+following array should be the output:
+
+::
+
+    [('[CLS]', 'O'), ('Hu', 'I-ORG'), ('##gging', 'I-ORG'), ('Face', 'I-ORG'), ('Inc', 'I-ORG'), ('.', 'O'), ('is', 'O'), ('a', 'O'), ('company', 'O'), ('based', 'O'), ('in', 'O'), ('New', 'I-LOC'), ('York', 'I-LOC'), ('City', 'I-LOC'), ('.', 'O'), ('Its', 'O'), ('headquarters', 'O'), ('are', 'O'), ('in', 'O'), ('D', 'I-LOC'), ('##UM', 'I-LOC'), ('##BO', 'I-LOC'), (',', 'O'), ('therefore', 'O'), ('very', 'O'), ('##c', 'O'), ('##lose', 'O'), ('to', 'O'), ('the', 'O'), ('Manhattan', 'I-LOC'), ('Bridge', 'I-LOC'), ('.', 'O'), ('[SEP]', 'O')]   
+Summarization
+----------------------------------------------------
+
+Summarization is the task of summarizing a text / an article into a shorter text.
+
+An example of a summarization dataset is the CNN / Daily Mail dataset, which consists of long news articles and was created for the task of summarization.
+If you would like to fine-tune a model on a summarization task, you may leverage the ``examples/summarization/bart/run_train.sh`` (leveraging pytorch-lightning) script.
+
+Here is an example using the pipelines do to summarization. 
+It leverages a Bart model that was fine-tuned on the CNN / Daily Mail data set.
+
+::
+
+    from transformers import pipeline
+
+    summarizer = pipeline("summarization")
+
+    ARTICLE = """ New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County, New York. 
+    A year later, she got married again in Westchester County, but to a different man and without divorcing her first husband. 
+    Only 18 days after that marriage, she got hitched yet again. Then, Barrientos declared "I do" five more times, sometimes only within two weeks of each other. 
+    In 2010, she married once more, this time in the Bronx. In an application for a marriage license, she stated it was her "first and only" marriage. 
+    Barrientos, now 39, is facing two criminal counts of "offering a false instrument for filing in the first degree," referring to her false statements on the 
+    2010 marriage license application, according to court documents. 
+    Prosecutors said the marriages were part of an immigration scam. 
+    On Friday, she pleaded not guilty at State Supreme Court in the Bronx, according to her attorney, Christopher Wright, who declined to comment further. 
+    After leaving court, Barrientos was arrested and charged with theft of service and criminal trespass for allegedly sneaking into the New York subway through an emergency exit, said Detective 
+    Annette Markowski, a police spokeswoman. In total, Barrientos has been married 10 times, with nine of her marriages occurring between 1999 and 2002. 
+    All occurred either in Westchester County, Long Island, New Jersey or the Bronx. She is believed to still be married to four men, and at one time, she was married to eight men at once, prosecutors say. 
+    Prosecutors said the immigration scam involved some of her husbands, who filed for permanent residence status shortly after the marriages. 
+    Any divorces happened only after such filings were approved. It was unclear whether any of the men will be prosecuted. 
+    The case was referred to the Bronx District Attorney\'s Office by Immigration and Customs Enforcement and the Department of Homeland Security\'s 
+    Investigation Division. Seven of the men are from so-called "red-flagged" countries, including Egypt, Turkey, Georgia, Pakistan and Mali. 
+    Her eighth husband, Rashid Rajput, was deported in 2006 to his native Pakistan after an investigation by the Joint Terrorism Task Force. 
+    If convicted, Barrientos faces up to four years in prison.  Her next court appearance is scheduled for May 18.
+    """
+    
+    print(summarizer(ARTICLE, max_length=130, min_length=30))
+
+Because the summarization pipeline depends on the ``PretrainedModel.generate()`` method, we can override the default arguments 
+of ``PretrainedModel.generate()`` directly in the pipeline as is shown for ``max_length`` and ``min_length`` above.
+This outputs the following summary:
+
+::
+
+  Liana Barrientos has been married 10 times, sometimes within two weeks of each other. Prosecutors say the marriages were part of an immigration scam. She pleaded not guilty at State Supreme Court in the Bronx on Friday.
+  
+Here is an example doing summarization using a model and a tokenizer. The process is the following:
+
+- Instantiate a tokenizer and a model from the checkpoint name. Summarization is usually done using an encoder-decoder model, such as ``Bart`` or ``T5``.
+- Define the article that should be summarizaed.
+- Leverage the ``PretrainedModel.generate()`` method.
+- Add the T5 specific prefix "summarize: ".
+
+Here Google`s T5 model is used that was only pre-trained on a multi-task mixed data set (including CNN / Daily Mail), but nevertheless yields very good results.
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoModelWithLMHead, AutoTokenizer
+
+    model = AutoModelWithLMHead.from_pretrained("t5-base")
+    tokenizer = AutoTokenizer.from_pretrained("t5-base")
+
+    # T5 uses a max_length of 512 so we cut the article to 512 tokens.
+    inputs = tokenizer.encode("summarize: " + ARTICLE, return_tensors="pt", max_length=512)
+    outputs = model.generate(inputs, max_length=150, min_length=40, length_penalty=2.0, num_beams=4, early_stopping=True)
+    print(outputs)
+    
+    ## TENSORFLOW CODE
+    from transformers import TFAutoModelWithLMHead, AutoTokenizer
+
+    model = TFAutoModelWithLMHead.from_pretrained("t5-base")
+    tokenizer = AutoTokenizer.from_pretrained("t5-base")
+
+    # T5 uses a max_length of 512 so we cut the article to 512 tokens.
+    inputs = tokenizer.encode("summarize: " + ARTICLE, return_tensors="tf", max_length=512)
+    outputs = model.generate(inputs, max_length=150, min_length=40, length_penalty=2.0, num_beams=4, early_stopping=True)
+    print(outputs)  
+Translation
+----------------------------------------------------
+
+Translation is the task of translating a text from one language to another.
+
+An example of a translation dataset is the WMT English to German dataset, which has English sentences as the input data 
+and German sentences as the target data.
+
+Here is an example using the pipelines do to translation. 
+It leverages a T5 model that was only pre-trained on a multi-task mixture dataset (including WMT), but yields impressive 
+translation results nevertheless.
+
+::
+
+    from transformers import pipeline
+
+    translator = pipeline("translation_en_to_de")
+    print(translator("Hugging Face is a technology company based in New York and Paris", max_length=40))
+
+Because the translation pipeline depends on the ``PretrainedModel.generate()`` method, we can override the default arguments 
+of ``PretrainedModel.generate()`` directly in the pipeline as is shown for ``max_length`` above.
+This outputs the following translation into German:
+
+::
+
+  Hugging Face ist ein Technologieunternehmen mit Sitz in New York und Paris.
+  
+Here is an example doing translation using a model and a tokenizer. The process is the following:
+
+- Instantiate a tokenizer and a model from the checkpoint name. Summarization is usually done using an encoder-decoder model, such as ``Bart`` or ``T5``.
+- Define the article that should be summarizaed.
+- Leverage the ``PretrainedModel.generate()`` method.
+- Add the T5 specific prefix "translate English to German: "
+
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoModelWithLMHead, AutoTokenizer
+
+    model = AutoModelWithLMHead.from_pretrained("t5-base")
+    tokenizer = AutoTokenizer.from_pretrained("t5-base")
+
+    inputs = tokenizer.encode("translate English to German: Hugging Face is a technology company based in New York and Paris", return_tensors="pt")
+    outputs = model.generate(inputs, max_length=40, num_beams=4, early_stopping=True)
+
+    print(outputs)
+    
+    ## TENSORFLOW CODE
+    from transformers import TFAutoModelWithLMHead, AutoTokenizer
+
+    model = TFAutoModelWithLMHead.from_pretrained("t5-base")
+    tokenizer = AutoTokenizer.from_pretrained("t5-base")
+
+    inputs = tokenizer.encode("translate English to German: Hugging Face is a technology company based in New York and Paris", return_tensors="tf")
+    outputs = model.generate(inputs, max_length=40, num_beams=4, early_stopping=True)
+
+    print(outputs)

examples/README.md

@@ -0,0 +1,621 @@
+# Examples
+
+In this section a few examples are put together. All of these examples work for several models, making use of the very
+similar API between the different models.
+
+**Important**
+To run the latest versions of the examples, you have to install from source and install some specific requirements for the examples.
+Execute the following steps in a new virtual environment:
+
+```bash
+git clone https://github.com/huggingface/transformers
+cd transformers
+pip install .
+pip install -r ./examples/requirements.txt
+```
+
+| Section                    | Description                                                                                                                                                |
+|----------------------------|------------------------------------------------------------------------------------------------------------------------------------------
+| [TensorFlow 2.0 models on GLUE](#TensorFlow-2.0-Bert-models-on-GLUE) | Examples running BERT TensorFlow 2.0 model on the GLUE tasks. |
+| [Language Model training](#language-model-training) | Fine-tuning (or training from scratch) the library models for language modeling on a text dataset. Causal language modeling for GPT/GPT-2, masked language modeling for BERT/RoBERTa. |
+| [Language Generation](#language-generation) | Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet. |
+| [GLUE](#glue) | Examples running BERT/XLM/XLNet/RoBERTa on the 9 GLUE tasks. Examples feature distributed training as well as half-precision. |
+| [SQuAD](#squad) | Using BERT/RoBERTa/XLNet/XLM for question answering, examples with distributed training. |
+| [Multiple Choice](#multiple-choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks. |
+| [Named Entity Recognition](https://github.com/huggingface/transformers/tree/master/examples/ner) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training. |
+| [XNLI](#xnli) | Examples running BERT/XLM on the XNLI benchmark. |
+| [Adversarial evaluation of model performances](#adversarial-evaluation-of-model-performances) | Testing a model with adversarial evaluation of natural language inference on the Heuristic Analysis for NLI Systems (HANS) dataset (McCoy et al., 2019.) |
+
+## TensorFlow 2.0 Bert models on GLUE
+
+Based on the script [`run_tf_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/run_tf_glue.py).
+
+Fine-tuning the library TensorFlow 2.0 Bert model for sequence classification on the  MRPC task of the GLUE benchmark: [General Language Understanding Evaluation](https://gluebenchmark.com/).
+
+This script has an option for mixed precision (Automatic Mixed Precision / AMP) to run models on Tensor Cores (NVIDIA Volta/Turing GPUs) and future hardware and an option for XLA, which uses the XLA compiler to reduce model runtime.
+Options are toggled using `USE_XLA` or `USE_AMP` variables in the script.
+These options and the below benchmark are provided by @tlkh.
+
+Quick benchmarks from the script (no other modifications):
+
+| GPU    | Mode | Time (2nd epoch) | Val Acc (3 runs) |
+| --------- | -------- | ----------------------- | ----------------------|
+| Titan V | FP32 | 41s | 0.8438/0.8281/0.8333 |
+| Titan V | AMP | 26s | 0.8281/0.8568/0.8411 |
+| V100    | FP32 | 35s | 0.8646/0.8359/0.8464 |
+| V100    | AMP | 22s | 0.8646/0.8385/0.8411 |
+| 1080 Ti | FP32 | 55s | - |
+
+Mixed precision (AMP) reduces the training time considerably for the same hardware and hyper-parameters (same batch size was used).
+
+## Language model training
+
+Based on the script [`run_language_modeling.py`](https://github.com/huggingface/transformers/blob/master/examples/run_language_modeling.py).
+
+Fine-tuning (or training from scratch) the library models for language modeling on a text dataset for GPT, GPT-2, BERT and RoBERTa (DistilBERT 
+to be added soon). GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa 
+are fine-tuned using a masked language modeling (MLM) loss.
+
+Before running the following example, you should get a file that contains text on which the language model will be
+trained or fine-tuned. A good example of such text is the [WikiText-2 dataset](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/).
+
+We will refer to two different files: `$TRAIN_FILE`, which contains text for training, and `$TEST_FILE`, which contains
+text that will be used for evaluation.
+
+### GPT-2/GPT and causal language modeling
+
+The following example fine-tunes GPT-2 on WikiText-2. We're using the raw WikiText-2 (no tokens were replaced before
+the tokenization). The loss here is that of causal language modeling.
+
+```bash
+export TRAIN_FILE=/path/to/dataset/wiki.train.raw
+export TEST_FILE=/path/to/dataset/wiki.test.raw
+
+python run_language_modeling.py \
+    --output_dir=output \
+    --model_type=gpt2 \
+    --model_name_or_path=gpt2 \
+    --do_train \
+    --train_data_file=$TRAIN_FILE \
+    --do_eval \
+    --eval_data_file=$TEST_FILE
+```
+
+This takes about half an hour to train on a single K80 GPU and about one minute for the evaluation to run. It reaches
+a score of ~20 perplexity once fine-tuned on the dataset.
+
+### RoBERTa/BERT and masked language modeling
+
+The following example fine-tunes RoBERTa on WikiText-2. Here too, we're using the raw WikiText-2. The loss is different
+as BERT/RoBERTa have a bidirectional mechanism; we're therefore using the same loss that was used during their
+pre-training: masked language modeling.
+
+In accordance to the RoBERTa paper, we use dynamic masking rather than static masking. The model may, therefore, converge
+slightly slower (over-fitting takes more epochs).
+
+We use the `--mlm` flag so that the script may change its loss function.
+
+```bash
+export TRAIN_FILE=/path/to/dataset/wiki.train.raw
+export TEST_FILE=/path/to/dataset/wiki.test.raw
+
+python run_language_modeling.py \
+    --output_dir=output \
+    --model_type=roberta \
+    --model_name_or_path=roberta-base \
+    --do_train \
+    --train_data_file=$TRAIN_FILE \
+    --do_eval \
+    --eval_data_file=$TEST_FILE \
+    --mlm
+```
+
+## Language generation
+
+Based on the script [`run_generation.py`](https://github.com/huggingface/transformers/blob/master/examples/run_generation.py).
+
+Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL, XLNet, CTRL.
+A similar script is used for our official demo [Write With Transfomer](https://transformer.huggingface.co), where you
+can try out the different models available in the library.
+
+Example usage:
+
+```bash
+python run_generation.py \
+    --model_type=gpt2 \
+    --model_name_or_path=gpt2
+```
+
+## GLUE
+
+Based on the script [`run_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/run_glue.py).
+
+Fine-tuning the library models for sequence classification on the GLUE benchmark: [General Language Understanding
+Evaluation](https://gluebenchmark.com/). This script can fine-tune the following models: BERT, XLM, XLNet and RoBERTa.
+
+GLUE is made up of a total of 9 different tasks. We get the following results on the dev set of the benchmark with an
+uncased  BERT base model (the checkpoint `bert-base-uncased`). All experiments ran single V100 GPUs with a total train
+batch sizes between 16 and 64. Some of these tasks have a small dataset and training can lead to high variance in the results
+between different runs. We report the median on 5 runs (with different seeds) for each of the metrics.
+
+| Task  | Metric                       | Result      |
+|-------|------------------------------|-------------|
+| CoLA  | Matthew's corr               | 49.23       |
+| SST-2 | Accuracy                     | 91.97       |
+| MRPC  | F1/Accuracy                  | 89.47/85.29 |
+| STS-B | Person/Spearman corr.        | 83.95/83.70 |
+| QQP   | Accuracy/F1                  | 88.40/84.31 |
+| MNLI  | Matched acc./Mismatched acc. | 80.61/81.08 |
+| QNLI  | Accuracy                     | 87.46       |
+| RTE   | Accuracy                     | 61.73       |
+| WNLI  | Accuracy                     | 45.07       |
+
+Some of these results are significantly different from the ones reported on the test set
+of GLUE benchmark on the website. For QQP and WNLI, please refer to [FAQ #12](https://gluebenchmark.com/faq) on the webite.
+
+Before running any one of these GLUE tasks you should download the
+[GLUE data](https://gluebenchmark.com/tasks) by running
+[this script](https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e)
+and unpack it to some directory `$GLUE_DIR`.
+
+```bash
+export GLUE_DIR=/path/to/glue
+export TASK_NAME=MRPC
+
+python run_glue.py \
+  --model_type bert \
+  --model_name_or_path bert-base-cased \
+  --task_name $TASK_NAME \
+  --do_train \
+  --do_eval \
+  --do_lower_case \
+  --data_dir $GLUE_DIR/$TASK_NAME \
+  --max_seq_length 128 \
+  --per_gpu_train_batch_size 32 \
+  --learning_rate 2e-5 \
+  --num_train_epochs 3.0 \
+  --output_dir /tmp/$TASK_NAME/
+```
+
+where task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI.
+
+The dev set results will be present within the text file `eval_results.txt` in the specified output_dir.
+In case of MNLI, since there are two separate dev sets (matched and mismatched), there will be a separate
+output folder called `/tmp/MNLI-MM/` in addition to `/tmp/MNLI/`.
+
+The code has not been tested with half-precision training with apex on any GLUE task apart from MRPC, MNLI,
+CoLA, SST-2. The following section provides details on how to run half-precision training with MRPC. With that being
+said, there shouldn’t be any issues in running half-precision training with the remaining GLUE tasks as well,
+since the data processor for each task inherits from the base class DataProcessor.
+
+### MRPC
+
+#### Fine-tuning example
+
+The following examples fine-tune BERT on the Microsoft Research Paraphrase Corpus (MRPC) corpus and runs in less
+than 10 minutes on a single K-80 and in 27 seconds (!) on single tesla V100 16GB with apex installed.
+
+Before running any one of these GLUE tasks you should download the
+[GLUE data](https://gluebenchmark.com/tasks) by running
+[this script](https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e)
+and unpack it to some directory `$GLUE_DIR`.
+
+```bash
+export GLUE_DIR=/path/to/glue
+
+python run_glue.py \
+  --model_type bert \
+  --model_name_or_path bert-base-cased \
+  --task_name MRPC \
+  --do_train \
+  --do_eval \
+  --do_lower_case \
+  --data_dir $GLUE_DIR/MRPC/ \
+  --max_seq_length 128 \
+  --per_gpu_train_batch_size 32 \
+  --learning_rate 2e-5 \
+  --num_train_epochs 3.0 \
+  --output_dir /tmp/mrpc_output/
+```
+
+Our test ran on a few seeds with [the original implementation hyper-
+parameters](https://github.com/google-research/bert#sentence-and-sentence-pair-classification-tasks) gave evaluation
+results between 84% and 88%.
+
+#### Using Apex and mixed-precision
+
+Using Apex and 16 bit precision, the fine-tuning on MRPC only takes 27 seconds. First install
+[apex](https://github.com/NVIDIA/apex), then run the following example:
+
+```bash
+export GLUE_DIR=/path/to/glue
+
+python run_glue.py \
+  --model_type bert \
+  --model_name_or_path bert-base-cased \
+  --task_name MRPC \
+  --do_train \
+  --do_eval \
+  --do_lower_case \
+  --data_dir $GLUE_DIR/MRPC/ \
+  --max_seq_length 128 \
+  --per_gpu_train_batch_size 32 \
+  --learning_rate 2e-5 \
+  --num_train_epochs 3.0 \
+  --output_dir /tmp/mrpc_output/ \
+  --fp16
+```
+
+#### Distributed training
+
+Here is an example using distributed training on 8 V100 GPUs. The model used is the BERT whole-word-masking and it
+reaches F1 > 92 on MRPC.
+
+```bash
+export GLUE_DIR=/path/to/glue
+
+python -m torch.distributed.launch \
+    --nproc_per_node 8 run_glue.py \
+    --model_type bert \
+    --model_name_or_path bert-base-cased \
+    --task_name MRPC \
+    --do_train \
+    --do_eval \
+    --do_lower_case \
+    --data_dir $GLUE_DIR/MRPC/ \
+    --max_seq_length 128 \
+    --per_gpu_train_batch_size 8 \
+    --learning_rate 2e-5 \
+    --num_train_epochs 3.0 \
+    --output_dir /tmp/mrpc_output/
+```
+
+Training with these hyper-parameters gave us the following results:
+
+```bash
+acc = 0.8823529411764706
+acc_and_f1 = 0.901702786377709
+eval_loss = 0.3418912578906332
+f1 = 0.9210526315789473
+global_step = 174
+loss = 0.07231863956341798
+```
+
+### MNLI
+
+The following example uses the BERT-large, uncased, whole-word-masking model and fine-tunes it on the MNLI task.
+
+```bash
+export GLUE_DIR=/path/to/glue
+
+python -m torch.distributed.launch \
+    --nproc_per_node 8 run_glue.py \
+    --model_type bert \
+    --model_name_or_path bert-base-cased \
+    --task_name mnli \
+    --do_train \
+    --do_eval \
+    --do_lower_case \
+    --data_dir $GLUE_DIR/MNLI/ \
+    --max_seq_length 128 \
+    --per_gpu_train_batch_size 8 \
+    --learning_rate 2e-5 \
+    --num_train_epochs 3.0 \
+    --output_dir output_dir \
+```
+
+The results  are the following:
+
+```bash
+***** Eval results *****
+  acc = 0.8679706601466992
+  eval_loss = 0.4911287787382479
+  global_step = 18408
+  loss = 0.04755385363816904
+
+***** Eval results *****
+  acc = 0.8747965825874695
+  eval_loss = 0.45516540421714036
+  global_step = 18408
+  loss = 0.04755385363816904
+```
+
+## Multiple Choice
+
+Based on the script [`run_multiple_choice.py`]().
+
+#### Fine-tuning on SWAG
+Download [swag](https://github.com/rowanz/swagaf/tree/master/data) data
+
+```bash
+#training on 4 tesla V100(16GB) GPUS
+export SWAG_DIR=/path/to/swag_data_dir
+python ./examples/run_multiple_choice.py \
+--model_type roberta \
+--task_name swag \
+--model_name_or_path roberta-base \
+--do_train \
+--do_eval \
+--do_lower_case \
+--data_dir $SWAG_DIR \
+--learning_rate 5e-5 \
+--num_train_epochs 3 \
+--max_seq_length 80 \
+--output_dir models_bert/swag_base \
+--per_gpu_eval_batch_size=16 \
+--per_gpu_train_batch_size=16 \
+--gradient_accumulation_steps 2 \
+--overwrite_output
+```
+Training with the defined hyper-parameters yields the following results:
+```
+***** Eval results *****
+eval_acc = 0.8338998300509847
+eval_loss = 0.44457291918821606
+```
+
+## SQuAD
+
+Based on the script [`run_squad.py`](https://github.com/huggingface/transformers/blob/master/examples/run_squad.py).
+
+#### Fine-tuning BERT on SQuAD1.0
+
+This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large)
+on a single tesla V100 16GB. The data for SQuAD can be downloaded with the following links and should be saved in a
+$SQUAD_DIR directory.
+
+* [train-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json)
+* [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
+
+python run_squad.py \
+  --model_type bert \
+  --model_name_or_path bert-base-uncased \
+  --do_train \
+  --do_eval \
+  --do_lower_case \
+  --train_file $SQUAD_DIR/train-v1.1.json \
+  --predict_file $SQUAD_DIR/dev-v1.1.json \
+  --per_gpu_train_batch_size 12 \
+  --learning_rate 3e-5 \
+  --num_train_epochs 2.0 \
+  --max_seq_length 384 \
+  --doc_stride 128 \
+  --output_dir /tmp/debug_squad/
+```
+
+Training with the previously defined hyper-parameters yields the following results:
+
+```bash
+f1 = 88.52
+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 SQuAD1.1:
+
+```bash
+python -m torch.distributed.launch --nproc_per_node=8 ./examples/run_squad.py \
+    --model_type bert \
+    --model_name_or_path bert-large-uncased-whole-word-masking \
+    --do_train \
+    --do_eval \
+    --do_lower_case \
+    --train_file $SQUAD_DIR/train-v1.1.json \
+    --predict_file $SQUAD_DIR/dev-v1.1.json \
+    --learning_rate 3e-5 \
+    --num_train_epochs 2 \
+    --max_seq_length 384 \
+    --doc_stride 128 \
+    --output_dir ./examples/models/wwm_uncased_finetuned_squad/ \
+    --per_gpu_eval_batch_size=3   \
+    --per_gpu_train_batch_size=3   \
+```
+
+Training with the previously defined hyper-parameters yields the following results:
+
+```bash
+f1 = 93.15
+exact_match = 86.91
+```
+
+This fine-tuned model is available as a checkpoint under the reference
+`bert-large-uncased-whole-word-masking-finetuned-squad`.
+
+#### Fine-tuning XLNet on SQuAD
+
+This example code fine-tunes XLNet on 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
+
+python run_squad.py \
+    --model_type xlnet \
+    --model_name_or_path xlnet-large-cased \
+    --do_train \
+    --do_eval \
+    --do_lower_case \
+    --train_file $SQUAD_DIR/train-v1.1.json \
+    --predict_file $SQUAD_DIR/dev-v1.1.json \
+    --learning_rate 3e-5 \
+    --num_train_epochs 2 \
+    --max_seq_length 384 \
+    --doc_stride 128 \
+    --output_dir ./wwm_cased_finetuned_squad/ \
+    --per_gpu_eval_batch_size=4  \
+    --per_gpu_train_batch_size=4   \
+    --save_steps 5000
+```
+
+##### 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 $SQUAD_DIR/dev-v2.0.json \
+    --learning_rate 3e-5 \
+    --num_train_epochs 4 \
+    --max_seq_length 384 \
+    --doc_stride 128 \
+    --output_dir ./wwm_cased_finetuned_squad/ \
+    --per_gpu_eval_batch_size=2  \
+    --per_gpu_train_batch_size=2   \
+    --save_steps 5000
+```
+
+Larger batch size may improve the performance while costing more memory.
+
+##### Results for SQuAD1.0 with the previously defined hyper-parameters:
+
+```python
+{
+"exact": 85.45884578997162,
+"f1": 92.5974600601065,
+"total": 10570,
+"HasAns_exact": 85.45884578997162,
+"HasAns_f1": 92.59746006010651,
+"HasAns_total": 10570
+}
+```
+
+##### Results for SQuAD2.0 with the previously defined hyper-parameters:
+
+```python
+{
+"exact": 80.4177545691906,
+"f1": 84.07154997729623,
+"total": 11873,
+"HasAns_exact": 76.73751686909581,
+"HasAns_f1": 84.05558584352873,
+"HasAns_total": 5928,
+"NoAns_exact": 84.0874684608915,
+"NoAns_f1": 84.0874684608915,
+"NoAns_total": 5945
+}
+```
+
+
+
+
+## XNLI
+
+Based on the script [`run_xnli.py`](https://github.com/huggingface/transformers/blob/master/examples/run_xnli.py).
+
+[XNLI](https://www.nyu.edu/projects/bowman/xnli/) is crowd-sourced dataset based on [MultiNLI](http://www.nyu.edu/projects/bowman/multinli/). It is an evaluation benchmark for cross-lingual text representations. Pairs of text are labeled with textual entailment annotations for 15 different languages (including both high-resource language such as English and low-resource languages such as Swahili).
+
+#### Fine-tuning on XNLI
+
+This example code fine-tunes mBERT (multi-lingual BERT) on the XNLI dataset. It runs in 106 mins
+on a single tesla V100 16GB. The data for XNLI can be downloaded with the following links and should be both saved (and un-zipped) in a
+`$XNLI_DIR` directory.
+
+* [XNLI 1.0](https://www.nyu.edu/projects/bowman/xnli/XNLI-1.0.zip)
+* [XNLI-MT 1.0](https://www.nyu.edu/projects/bowman/xnli/XNLI-MT-1.0.zip)
+
+```bash
+export XNLI_DIR=/path/to/XNLI
+
+python run_xnli.py \
+  --model_type bert \
+  --model_name_or_path bert-base-multilingual-cased \
+  --language de \
+  --train_language en \
+  --do_train \
+  --do_eval \
+  --data_dir $XNLI_DIR \
+  --per_gpu_train_batch_size 32 \
+  --learning_rate 5e-5 \
+  --num_train_epochs 2.0 \
+  --max_seq_length 128 \
+  --output_dir /tmp/debug_xnli/ \
+  --save_steps -1
+```
+
+Training with the previously defined hyper-parameters yields the following results on the **test** set:
+
+```bash
+acc = 0.7093812375249501
+```
+
+## MM-IMDb
+
+Based on the script [`run_mmimdb.py`](https://github.com/huggingface/transformers/blob/master/examples/mm-imdb/run_mmimdb.py).
+
+[MM-IMDb](http://lisi1.unal.edu.co/mmimdb/) is a Multimodal dataset with around 26,000 movies including images, plots and other metadata.
+
+### Training on MM-IMDb
+
+```
+python run_mmimdb.py \
+    --data_dir /path/to/mmimdb/dataset/ \
+    --model_type bert \
+    --model_name_or_path bert-base-uncased \
+    --output_dir /path/to/save/dir/ \
+    --do_train \
+    --do_eval \
+    --max_seq_len 512 \
+    --gradient_accumulation_steps 20 \
+    --num_image_embeds 3 \
+    --num_train_epochs 100 \
+    --patience 5
+```
+
+## Adversarial evaluation of model performances
+
+Here is an example on evaluating a model using adversarial evaluation of natural language inference with the Heuristic Analysis for NLI Systems (HANS) dataset [McCoy et al., 2019](https://arxiv.org/abs/1902.01007). The example was gracefully provided by [Nafise Sadat Moosavi](https://github.com/ns-moosavi).
+
+The HANS dataset can be downloaded from [this location](https://github.com/tommccoy1/hans).
+
+This is an example of using test_hans.py:
+
+```bash
+export HANS_DIR=path-to-hans
+export MODEL_TYPE=type-of-the-model-e.g.-bert-roberta-xlnet-etc
+export MODEL_PATH=path-to-the-model-directory-that-is-trained-on-NLI-e.g.-by-using-run_glue.py
+
+python examples/hans/test_hans.py \
+        --task_name hans \
+        --model_type $MODEL_TYPE \
+        --do_eval \
+        --do_lower_case \
+        --data_dir $HANS_DIR \
+        --model_name_or_path $MODEL_PATH \
+        --max_seq_length 128 \
+        --output_dir $MODEL_PATH \
+```
+
+This will create the hans_predictions.txt file in MODEL_PATH, which can then be evaluated using hans/evaluate_heur_output.py from the HANS dataset.
+
+The results of the BERT-base model that is trained on MNLI using batch size 8 and the random seed 42 on the HANS dataset is as follows:
+
+```bash
+Heuristic entailed results:
+lexical_overlap: 0.9702
+subsequence: 0.9942
+constituent: 0.9962
+
+Heuristic non-entailed results:
+lexical_overlap: 0.199
+subsequence: 0.0396
+constituent: 0.118
+```

examples/benchmarks.py

@@ -0,0 +1,664 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Benchmarking the library on inference and training """
+
+# If checking the tensors placement
+# tf.debugging.set_log_device_placement(True)
+
+import argparse
+import csv
+import timeit
+from time import time
+from typing import List
+
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    MemorySummary,
+    is_tf_available,
+    is_torch_available,
+    start_memory_tracing,
+    stop_memory_tracing,
+)
+
+
+if is_tf_available():
+    import tensorflow as tf
+    from transformers import TFAutoModel
+
+if is_torch_available():
+    import torch
+    from transformers import AutoModel
+
+
+input_text = """Bent over their instruments, three hundred Fertilizers were plunged, as
+the Director of Hatcheries and Conditioning entered the room, in the
+
+
+
+scarcely breathing silence, the absent-minded, soliloquizing hum or
+whistle, of absorbed concentration. A troop of newly arrived students,
+very young, pink and callow, followed nervously, rather abjectly, at the
+Director's heels. Each of them carried a notebook, in which, whenever
+the great man spoke, he desperately scribbled. Straight from the
+horse's mouth. It was a rare privilege. The D. H. C. for Central London
+always made a point of personally conducting his new students round
+the various departments.
+
+"Just to give you a general idea," he would explain to them. For of
+course some sort of general idea they must have, if they were to do
+their work intelligently-though as little of one, if they were to be good
+and happy members of society, as possible. For particulars, as every
+one knows, make for virtue and happiness; generalities are intellectu-
+ally necessary evils. Not philosophers but fret-sawyers and stamp col-
+lectors compose the backbone of society.
+
+"To-morrow," he would add, smiling at them with a slightly menacing
+geniality, "you'll be settling down to serious work. You won't have time
+for generalities. Meanwhile ..."
+
+Meanwhile, it was a privilege. Straight from the horse's mouth into the
+notebook. The boys scribbled like mad.
+
+Tall and rather thin but upright, the Director advanced into the room.
+He had a long chin and big rather prominent teeth, just covered, when
+he was not talking, by his full, floridly curved lips. Old, young? Thirty?
+Fifty? Fifty-five? It was hard to say. And anyhow the question didn't
+arise; in this year of stability, A. F. 632, it didn't occur to you to ask it.
+
+"I shall begin at the beginning," said the D.H.C. and the more zealous
+students recorded his intention in their notebooks: Begin at the begin-
+ning. "These," he waved his hand, "are the incubators." And opening
+an insulated door he showed them racks upon racks of numbered test-
+tubes. "The week's supply of ova. Kept," he explained, "at blood heat;
+whereas the male gametes," and here he opened another door, "they
+have to be kept at thirty-five instead of thirty-seven. Full blood heat
+sterilizes." Rams wrapped in theremogene beget no lambs.
+
+Still leaning against the incubators he gave them, while the pencils
+scurried illegibly across the pages, a brief description of the modern
+
+
+
+fertilizing process; spoke first, of course, of its surgical introduc-
+tion-"the operation undergone voluntarily for the good of Society, not
+to mention the fact that it carries a bonus amounting to six months'
+salary"; continued with some account of the technique for preserving
+the excised ovary alive and actively developing; passed on to a consid-
+eration of optimum temperature, salinity, viscosity; referred to the liq-
+uor in which the detached and ripened eggs were kept; and, leading
+his charges to the work tables, actually showed them how this liquor
+was drawn off from the test-tubes; how it was let out drop by drop
+onto the specially warmed slides of the microscopes; how the eggs
+which it contained were inspected for abnormalities, counted and
+transferred to a porous receptacle; how (and he now took them to
+watch the operation) this receptacle was immersed in a warm bouillon
+containing free-swimming spermatozoa-at a minimum concentration
+of one hundred thousand per cubic centimetre, he insisted; and how,
+after ten minutes, the container was lifted out of the liquor and its
+contents re-examined; how, if any of the eggs remained unfertilized, it
+was again immersed, and, if necessary, yet again; how the fertilized
+ova went back to the incubators; where the Alphas and Betas re-
+mained until definitely bottled; while the Gammas, Deltas and Epsilons
+were brought out again, after only thirty-six hours, to undergo Bo-
+kanovsky's Process.
+
+"Bokanovsky's Process," repeated the Director, and the students un-
+derlined the words in their little notebooks.
+
+One egg, one embryo, one adult-normality. But a bokanovskified egg
+will bud, will proliferate, will divide. From eight to ninety-six buds, and
+every bud will grow into a perfectly formed embryo, and every embryo
+into a full-sized adult. Making ninety-six human beings grow where
+only one grew before. Progress.
+
+"Essentially," the D.H.C. concluded, "bokanovskification consists of a
+series of arrests of development. We check the normal growth and,
+paradoxically enough, the egg responds by budding."
+
+Responds by budding. The pencils were busy.
+
+He pointed. On a very slowly moving band a rack-full of test-tubes was
+entering a large metal box, another, rack-full was emerging. Machinery
+faintly purred. It took eight minutes for the tubes to go through, he
+
+
+
+told them. Eight minutes of hard X-rays being about as much as an
+egg can stand. A few died; of the rest, the least susceptible divided
+into two; most put out four buds; some eight; all were returned to the
+incubators, where the buds began to develop; then, after two days,
+were suddenly chilled, chilled and checked. Two, four, eight, the buds
+in their turn budded; and having budded were dosed almost to death
+with alcohol; consequently burgeoned again and having budded-bud
+out of bud out of bud-were thereafter-further arrest being generally
+fatal-left to develop in peace. By which time the original egg was in a
+fair way to becoming anything from eight to ninety-six embryos- a
+prodigious improvement, you will agree, on nature. Identical twins-but
+not in piddling twos and threes as in the old viviparous days, when an
+egg would sometimes accidentally divide; actually by dozens, by
+scores at a time.
+
+"Scores," the Director repeated and flung out his arms, as though he
+were distributing largesse. "Scores."
+
+But one of the students was fool enough to ask where the advantage
+lay.
+
+"My good boy!" The Director wheeled sharply round on him. "Can't you
+see? Can't you see?" He raised a hand; his expression was solemn.
+"Bokanovsky's Process is one of the major instruments of social stabil-
+ity!"
+
+Major instruments of social stability.
+
+Standard men and women; in uniform batches. The whole of a small
+factory staffed with the products of a single bokanovskified egg.
+
+"Ninety-six identical twins working ninety-six identical machines!" The
+voice was almost tremulous with enthusiasm. "You really know where
+you are. For the first time in history." He quoted the planetary motto.
+"Community, Identity, Stability." Grand words. "If we could bo-
+kanovskify indefinitely the whole problem would be solved."
+
+Solved by standard Gammas, unvarying Deltas, uniform Epsilons. Mil-
+lions of identical twins. The principle of mass production at last applied
+to biology.
+
+
+
+"But, alas," the Director shook his head, "we can't bokanovskify indefi-
+nitely."
+
+Ninety-six seemed to be the limit; seventy-two a good average. From
+the same ovary and with gametes of the same male to manufacture as
+many batches of identical twins as possible-that was the best (sadly a
+second best) that they could do. And even that was difficult.
+
+"For in nature it takes thirty years for two hundred eggs to reach ma-
+turity. But our business is to stabilize the population at this moment,
+here and now. Dribbling out twins over a quarter of a century-what
+would be the use of that?"
+
+Obviously, no use at all. But Podsnap's Technique had immensely ac-
+celerated the process of ripening. They could make sure of at least a
+hundred and fifty mature eggs within two years. Fertilize and bo-
+kanovskify-in other words, multiply by seventy-two-and you get an
+average of nearly eleven thousand brothers and sisters in a hundred
+and fifty batches of identical twins, all within two years of the same
+age.
+
+"And in exceptional cases we can make one ovary yield us over fifteen
+thousand adult individuals."
+
+Beckoning to a fair-haired, ruddy young man who happened to be
+passing at the moment. "Mr. Foster," he called. The ruddy young man
+approached. "Can you tell us the record for a single ovary, Mr. Foster?"
+
+"Sixteen thousand and twelve in this Centre," Mr. Foster replied with-
+out hesitation. He spoke very quickly, had a vivacious blue eye, and
+took an evident pleasure in quoting figures. "Sixteen thousand and
+twelve; in one hundred and eighty-nine batches of identicals. But of
+course they've done much better," he rattled on, "in some of the tropi-
+cal Centres. Singapore has often produced over sixteen thousand five
+hundred; and Mombasa has actually touched the seventeen thousand
+mark. But then they have unfair advantages. You should see the way a
+negro ovary responds to pituitary! It's quite astonishing, when you're
+used to working with European material. Still," he added, with a laugh
+(but the light of combat was in his eyes and the lift of his chin was
+challenging), "still, we mean to beat them if we can. I'm working on a
+wonderful Delta-Minus ovary at this moment. Only just eighteen
+
+
+
+months old. Over twelve thousand seven hundred children already, ei-
+ther decanted or in embryo. And still going strong. We'll beat them
+yet."
+
+"That's the spirit I like!" cried the Director, and clapped Mr. Foster on
+the shoulder. "Come along with us, and give these boys the benefit of
+your expert knowledge."
+
+Mr. Foster smiled modestly. "With pleasure." They went.
+In the Bottling Room all was harmonious bustle and ordered activity.
+Flaps of fresh sow's peritoneum ready cut to the proper size came
+shooting up in little lifts from the Organ Store in the sub-basement.
+Whizz and then, click! the lift-hatches hew open; the bottle-liner had
+only to reach out a hand, take the flap, insert, smooth-down, and be-
+fore the lined bottle had had time to travel out of reach along the end-
+less band, whizz, click! another flap of peritoneum had shot up from
+the depths, ready to be slipped into yet another bottle, the next of that
+slow interminable procession on the band.
+
+Next to the Liners stood the Matriculators. The procession advanced;
+one by one the eggs were transferred from their test-tubes to the
+larger containers; deftly the peritoneal lining was slit, the morula
+dropped into place, the saline solution poured in ... and already the
+bottle had passed, and it was the turn of the labellers. Heredity, date
+of fertilization, membership of Bokanovsky Group-details were trans-
+ferred from test-tube to bottle. No longer anonymous, but named,
+identified, the procession marched slowly on; on through an opening in
+the wall, slowly on into the Social Predestination Room.
+"Eighty-eight cubic metres of card-index," said Mr. Foster with relish,
+as they entered."""
+
+
+def create_setup_and_compute(
+    model_names: List[str],
+    batch_sizes: List[int],
+    slice_sizes: List[int],
+    gpu: bool = True,
+    tensorflow: bool = False,
+    average_over: int = 3,
+    no_speed: bool = False,
+    no_memory: bool = False,
+    verbose: bool = False,
+    torchscript: bool = False,
+    xla: bool = False,
+    amp: bool = False,
+    fp16: bool = False,
+    save_to_csv: bool = False,
+    csv_filename: str = f"results_{round(time())}.csv",
+    csv_memory_filename: str = f"memory_{round(time())}.csv",
+):
+    if xla:
+        tf.config.optimizer.set_jit(True)
+    if amp:
+        tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})
+
+    if tensorflow:
+        dictionary = {model_name: {} for model_name in model_names}
+        results = _compute_tensorflow(
+            model_names, batch_sizes, slice_sizes, dictionary, average_over, amp, no_speed, no_memory, verbose
+        )
+    else:
+        device = "cuda" if (gpu and torch.cuda.is_available()) else "cpu"
+        dictionary = {model_name: {} for model_name in model_names}
+        results = _compute_pytorch(
+            model_names,
+            batch_sizes,
+            slice_sizes,
+            dictionary,
+            average_over,
+            device,
+            torchscript,
+            fp16,
+            no_speed,
+            no_memory,
+            verbose,
+        )
+
+    print("=========== RESULTS ===========")
+    for model_name in model_names:
+        print("\t" + f"======= MODEL CHECKPOINT: {model_name} =======")
+        for batch_size in results[model_name]["bs"]:
+            print("\t\t" + f"===== BATCH SIZE: {batch_size} =====")
+            for slice_size in results[model_name]["ss"]:
+                result = results[model_name]["results"][batch_size][slice_size]
+                memory = results[model_name]["memory"][batch_size][slice_size]
+                if isinstance(result, str):
+                    print(f"\t\t{model_name}/{batch_size}/{slice_size}: " f"{result} " f"{memory}")
+                else:
+                    print(
+                        f"\t\t{model_name}/{batch_size}/{slice_size}: "
+                        f"{(round(1000 * result) / 1000)}"
+                        f"s "
+                        f"{memory}"
+                    )
+
+    if save_to_csv:
+        with open(csv_filename, mode="w") as csv_file, open(csv_memory_filename, mode="w") as csv_memory_file:
+            fieldnames = [
+                "model",
+                "1x8",
+                "1x64",
+                "1x128",
+                "1x256",
+                "1x512",
+                "1x1024",
+                "2x8",
+                "2x64",
+                "2x128",
+                "2x256",
+                "2x512",
+                "2x1024",
+                "4x8",
+                "4x64",
+                "4x128",
+                "4x256",
+                "4x512",
+                "4x1024",
+                "8x8",
+                "8x64",
+                "8x128",
+                "8x256",
+                "8x512",
+                "8x1024",
+            ]
+
+            writer = csv.DictWriter(csv_file, fieldnames=fieldnames)
+            writer.writeheader()
+            memory_writer = csv.DictWriter(csv_memory_file, fieldnames=fieldnames)
+            memory_writer.writeheader()
+
+            for model_name in model_names:
+                model_results = {
+                    f"{bs}x{ss}": results[model_name]["results"][bs][ss]
+                    for bs in results[model_name]["results"]
+                    for ss in results[model_name]["results"][bs]
+                }
+                writer.writerow({"model": model_name, **model_results})
+
+                model_memory_results = {
+                    f"{bs}x{ss}": results[model_name]["memory"][bs][ss]
+                    for bs in results[model_name]["memory"]
+                    for ss in results[model_name]["memory"][bs]
+                }
+                memory_writer.writerow({"model": model_name, **model_memory_results})
+
+
+def print_summary_statistics(summary: MemorySummary):
+    print(
+        "\nLines by line memory consumption:\n"
+        + "\n".join(
+            f"{state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
+            for state in summary.sequential
+        )
+    )
+    print(
+        "\nLines with top memory consumption:\n"
+        + "\n".join(
+            f"=> {state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
+            for state in summary.cumulative[:6]
+        )
+    )
+    print(
+        "\nLines with lowest memory consumption:\n"
+        + "\n".join(
+            f"=> {state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
+            for state in summary.cumulative[-6:]
+        )
+    )
+    print(f"\nTotal memory increase: {summary.total}")
+
+
+def _compute_pytorch(
+    model_names,
+    batch_sizes,
+    slice_sizes,
+    dictionary,
+    average_over,
+    device,
+    torchscript,
+    fp16,
+    no_speed,
+    no_memory,
+    verbose,
+):
+    for c, model_name in enumerate(model_names):
+        print(f"{c + 1} / {len(model_names)}")
+        config = AutoConfig.from_pretrained(model_name, torchscript=torchscript)
+        model = AutoModel.from_pretrained(model_name, config=config)
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+        tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False)
+
+        max_input_size = tokenizer.max_model_input_sizes[model_name]
+
+        dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}, "memory": {}}
+        dictionary[model_name]["results"] = {i: {} for i in batch_sizes}
+        dictionary[model_name]["memory"] = {i: {} for i in batch_sizes}
+
+        for batch_size in batch_sizes:
+            if fp16:
+                model.half()
+            model.to(device)
+            model.eval()
+
+            for slice_size in slice_sizes:
+                if max_input_size is not None and slice_size > max_input_size:
+                    dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+                else:
+                    sequence = torch.tensor(tokenized_sequence[:slice_size], device=device).repeat(batch_size, 1)
+                    try:
+                        if torchscript:
+                            print("Tracing model with sequence size", sequence.shape)
+                            inference = torch.jit.trace(model, sequence)
+                            inference(sequence)
+                        else:
+                            inference = model
+                            inference(sequence)
+
+                        if not no_memory:
+                            # model.add_memory_hooks()  # Forward method tracing (only for PyTorch models)
+
+                            # Line by line memory tracing (all code in the module `transformers`) works for all models/arbitrary code
+                            trace = start_memory_tracing("transformers")
+                            inference(sequence)
+                            summary = stop_memory_tracing(trace)
+
+                            if verbose:
+                                print_summary_statistics(summary)
+
+                            dictionary[model_name]["memory"][batch_size][slice_size] = str(summary.total)
+                        else:
+                            dictionary[model_name]["memory"][batch_size][slice_size] = "N/A"
+
+                        if not no_speed:
+                            print("Going through model with sequence of shape", sequence.shape)
+                            runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
+                            average_time = sum(runtimes) / float(len(runtimes)) / 3.0
+                            dictionary[model_name]["results"][batch_size][slice_size] = average_time
+                        else:
+                            dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+
+                    except RuntimeError as e:
+                        print("Doesn't fit on GPU.", e)
+                        torch.cuda.empty_cache()
+                        dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+                        dictionary[model_name]["memory"][batch_size][slice_size] = "N/A"
+    return dictionary
+
+
+def _compute_tensorflow(
+    model_names, batch_sizes, slice_sizes, dictionary, average_over, amp, no_speed, no_memory, verbose
+):
+    for c, model_name in enumerate(model_names):
+        print(f"{c + 1} / {len(model_names)}")
+        config = AutoConfig.from_pretrained(model_name)
+        model = TFAutoModel.from_pretrained(model_name, config=config)
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+        tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False)
+
+        max_input_size = tokenizer.max_model_input_sizes[model_name]
+
+        dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}, "memory": {}}
+        dictionary[model_name]["results"] = {i: {} for i in batch_sizes}
+        dictionary[model_name]["memory"] = {i: {} for i in batch_sizes}
+
+        print("Using model", model)
+
+        @tf.function
+        def inference(inputs):
+            return model(inputs)
+
+        for batch_size in batch_sizes:
+            for slice_size in slice_sizes:
+                if max_input_size is not None and slice_size > max_input_size:
+                    dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+                else:
+                    sequence = tf.stack(
+                        [tf.squeeze(tf.constant(tokenized_sequence[:slice_size])[None, :])] * batch_size
+                    )
+
+                    try:
+                        print("Going through model with sequence of shape", sequence.shape)
+                        # To make sure that the model is traced + that the tensors are on the appropriate device
+                        inference(sequence)
+
+                        if not no_memory:
+                            # Line by line memory tracing (all code in the module `transformers`) works for all models/arbitrary code
+                            trace = start_memory_tracing("transformers")
+                            inference(sequence)
+                            summary = stop_memory_tracing(trace)
+
+                            if verbose:
+                                print_summary_statistics(summary)
+
+                            dictionary[model_name]["memory"][batch_size][slice_size] = str(summary.total)
+                        else:
+                            dictionary[model_name]["memory"][batch_size][slice_size] = "N/A"
+
+                        if not no_speed:
+                            runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
+                            average_time = sum(runtimes) / float(len(runtimes)) / 3.0
+                            dictionary[model_name]["results"][batch_size][slice_size] = average_time
+                        else:
+                            dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+
+                    except tf.errors.ResourceExhaustedError as e:
+                        print("Doesn't fit on GPU.", e)
+                        dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+                        dictionary[model_name]["memory"][batch_size][slice_size] = "N/A"
+    return dictionary
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--models",
+        required=False,
+        type=str,
+        default="all",
+        help="Model checkpoints to be provided "
+        "to the AutoModel classes. Leave "
+        "blank to benchmark the base version "
+        "of all available model "
+        "architectures.",
+    )
+    parser.add_argument("--verbose", required=False, action="store_true", help="Verbose memory tracing")
+    parser.add_argument("--no_speed", required=False, action="store_true", help="Don't perform speed measurments")
+    parser.add_argument("--no_memory", required=False, action="store_true", help="Don't perform memory measurments")
+    parser.add_argument(
+        "--torch", required=False, action="store_true", help="Benchmark the Pytorch version of the " "models"
+    )
+    parser.add_argument(
+        "--torch_cuda", required=False, action="store_true", help="Pytorch only: run on available " "cuda devices"
+    )
+    parser.add_argument(
+        "--torchscript",
+        required=False,
+        action="store_true",
+        help="Pytorch only: trace the models " "using torchscript",
+    )
+    parser.add_argument(
+        "--tensorflow",
+        required=False,
+        action="store_true",
+        help="Benchmark the TensorFlow version "
+        "of the models. Will run on GPU if "
+        "the correct dependencies are "
+        "installed",
+    )
+    parser.add_argument("--xla", required=False, action="store_true", help="TensorFlow only: use XLA acceleration.")
+    parser.add_argument(
+        "--amp",
+        required=False,
+        action="store_true",
+        help="TensorFlow only: use automatic mixed precision acceleration.",
+    )
+    parser.add_argument(
+        "--fp16", required=False, action="store_true", help="PyTorch only: use FP16 to accelerate inference."
+    )
+    parser.add_argument(
+        "--keras_predict",
+        required=False,
+        action="store_true",
+        help="Whether to use model.predict " "instead of model() to do a " "forward pass.",
+    )
+    parser.add_argument("--save_to_csv", required=False, action="store_true", help="Save to a CSV file.")
+    parser.add_argument(
+        "--csv_filename", required=False, default=None, help="CSV filename used if saving results to csv."
+    )
+    parser.add_argument(
+        "--average_over", required=False, default=30, type=int, help="Times an experiment will be run."
+    )
+    parser.add_argument("--batch_sizes", nargs="+", type=int, default=[1, 2, 4, 8])
+    parser.add_argument("--slice_sizes", nargs="+", type=int, default=[8, 64, 128, 256, 512, 1024])
+
+    args = parser.parse_args()
+    if args.models == "all":
+        args.models = [
+            "gpt2",
+            "bert-base-cased",
+            "xlnet-base-cased",
+            "xlm-mlm-en-2048",
+            "transfo-xl-wt103",
+            "openai-gpt",
+            "distilbert-base-uncased",
+            "distilgpt2",
+            "roberta-base",
+            "ctrl",
+        ]
+    else:
+        args.models = args.models.split()
+
+    print("Running with arguments", args)
+
+    if args.torch:
+        if is_torch_available():
+            create_setup_and_compute(
+                model_names=args.models,
+                batch_sizes=args.batch_sizes,
+                slice_sizes=args.slice_sizes,
+                tensorflow=False,
+                gpu=args.torch_cuda,
+                torchscript=args.torchscript,
+                fp16=args.fp16,
+                save_to_csv=args.save_to_csv,
+                csv_filename=args.csv_filename,
+                average_over=args.average_over,
+                no_speed=args.no_speed,
+                no_memory=args.no_memory,
+                verbose=args.verbose,
+            )
+        else:
+            raise ImportError("Trying to run a PyTorch benchmark but PyTorch was not found in the environment.")
+
+    if args.tensorflow:
+        if is_tf_available():
+            create_setup_and_compute(
+                model_names=args.models,
+                batch_sizes=args.batch_sizes,
+                slice_sizes=args.slice_sizes,
+                tensorflow=True,
+                xla=args.xla,
+                amp=args.amp,
+                save_to_csv=args.save_to_csv,
+                csv_filename=args.csv_filename,
+                average_over=args.average_over,
+                no_speed=args.no_speed,
+                no_memory=args.no_memory,
+                verbose=args.verbose,
+            )
+        else:
+            raise ImportError("Trying to run a TensorFlow benchmark but TensorFlow was not found in the environment.")
+
+
+if __name__ == "__main__":
+    main()

examples/contrib/README.md

@@ -0,0 +1,5 @@
+# Community contributed examples
+
+This folder contains examples which are not actively maintained (mostly contributed by the community).
+
+Using these examples together with a recent version of the library usually requires to make small (sometimes big) adaptations to get the scripts working.

examples/contrib/run_camembert.py

@@ -0,0 +1,43 @@
+import torch
+
+from transformers.modeling_camembert import CamembertForMaskedLM
+from transformers.tokenization_camembert import CamembertTokenizer
+
+
+def fill_mask(masked_input, model, tokenizer, topk=5):
+    # Adapted from https://github.com/pytorch/fairseq/blob/master/fairseq/models/roberta/hub_interface.py
+    assert masked_input.count("<mask>") == 1
+    input_ids = torch.tensor(tokenizer.encode(masked_input, add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+    logits = model(input_ids)[0]  # The last hidden-state is the first element of the output tuple
+    masked_index = (input_ids.squeeze() == tokenizer.mask_token_id).nonzero().item()
+    logits = logits[0, masked_index, :]
+    prob = logits.softmax(dim=0)
+    values, indices = prob.topk(k=topk, dim=0)
+    topk_predicted_token_bpe = " ".join(
+        [tokenizer.convert_ids_to_tokens(indices[i].item()) for i in range(len(indices))]
+    )
+    masked_token = tokenizer.mask_token
+    topk_filled_outputs = []
+    for index, predicted_token_bpe in enumerate(topk_predicted_token_bpe.split(" ")):
+        predicted_token = predicted_token_bpe.replace("\u2581", " ")
+        if " {0}".format(masked_token) in masked_input:
+            topk_filled_outputs.append(
+                (
+                    masked_input.replace(" {0}".format(masked_token), predicted_token),
+                    values[index].item(),
+                    predicted_token,
+                )
+            )
+        else:
+            topk_filled_outputs.append(
+                (masked_input.replace(masked_token, predicted_token), values[index].item(), predicted_token,)
+            )
+    return topk_filled_outputs
+
+
+tokenizer = CamembertTokenizer.from_pretrained("camembert-base")
+model = CamembertForMaskedLM.from_pretrained("camembert-base")
+model.eval()
+
+masked_input = "Le camembert est <mask> :)"
+print(fill_mask(masked_input, model, tokenizer, topk=3))

examples/contrib/run_openai_gpt.py

@@ -22,48 +22,54 @@
           --model_name openai-gpt \
           --do_train \
           --do_eval \
-          --train_dataset $ROC_STORIES_DIR/cloze_test_val__spring2016\ -\ cloze_test_ALL_val.csv \
-          --eval_dataset $ROC_STORIES_DIR/cloze_test_test__spring2016\ -\ cloze_test_ALL_test.csv \
+          --train_dataset "$ROC_STORIES_DIR/cloze_test_val__spring2016 - cloze_test_ALL_val.csv" \
+          --eval_dataset "$ROC_STORIES_DIR/cloze_test_test__spring2016 - cloze_test_ALL_test.csv" \
           --output_dir ../log \
           --train_batch_size 16 \
 """
 import argparse
-import os
 import csv
-import random
 import logging
-from tqdm import tqdm, trange
+import os
+import random
 
 import numpy as np
 import torch
-from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
-                              TensorDataset)
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
+from tqdm import tqdm, trange
 
-from pytorch_transformers import (OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer,
-                                     AdamW, cached_path, WEIGHTS_NAME, CONFIG_NAME,
-                                     WarmupLinearSchedule)
+from transformers import (
+    CONFIG_NAME,
+    WEIGHTS_NAME,
+    AdamW,
+    OpenAIGPTDoubleHeadsModel,
+    OpenAIGPTTokenizer,
+    get_linear_schedule_with_warmup,
+)
 
-ROCSTORIES_URL = "https://s3.amazonaws.com/datasets.huggingface.co/ROCStories.tar.gz"
 
-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__)
 
+
 def accuracy(out, labels):
     outputs = np.argmax(out, axis=1)
     return np.sum(outputs == labels)
 
+
 def load_rocstories_dataset(dataset_path):
     """ Output a list of tuples(story, 1st continuation, 2nd continuation, label) """
-    with open(dataset_path, encoding='utf_8') as f:
+    with open(dataset_path, encoding="utf_8") as f:
         f = csv.reader(f)
         output = []
-        next(f) # skip the first line
+        next(f)  # skip the first line
         for line in tqdm(f):
-            output.append((' '.join(line[1:5]), line[5], line[6], int(line[-1])-1))
+            output.append((" ".join(line[1:5]), line[5], line[6], int(line[-1]) - 1))
     return output
 
+
 def pre_process_datasets(encoded_datasets, input_len, cap_length, start_token, delimiter_token, clf_token):
     """ Pre-process datasets containing lists of tuples(story, 1st continuation, 2nd continuation, label)
 
@@ -75,61 +81,73 @@ def pre_process_datasets(encoded_datasets, input_len, cap_length, start_token, d
         n_batch = len(dataset)
         input_ids = np.zeros((n_batch, 2, input_len), dtype=np.int64)
         mc_token_ids = np.zeros((n_batch, 2), dtype=np.int64)
-        lm_labels = np.full((n_batch, 2, input_len), fill_value=-1, dtype=np.int64)
+        lm_labels = np.full((n_batch, 2, input_len), fill_value=-100, dtype=np.int64)
         mc_labels = np.zeros((n_batch,), dtype=np.int64)
         for i, (story, cont1, cont2, mc_label), in enumerate(dataset):
             with_cont1 = [start_token] + story[:cap_length] + [delimiter_token] + cont1[:cap_length] + [clf_token]
             with_cont2 = [start_token] + story[:cap_length] + [delimiter_token] + cont2[:cap_length] + [clf_token]
-            input_ids[i, 0, :len(with_cont1)] = with_cont1
-            input_ids[i, 1, :len(with_cont2)] = with_cont2
+            input_ids[i, 0, : len(with_cont1)] = with_cont1
+            input_ids[i, 1, : len(with_cont2)] = with_cont2
             mc_token_ids[i, 0] = len(with_cont1) - 1
             mc_token_ids[i, 1] = len(with_cont2) - 1
-            lm_labels[i, 0, :len(with_cont1)] = with_cont1
-            lm_labels[i, 1, :len(with_cont2)] = with_cont2
+            lm_labels[i, 0, : len(with_cont1)] = with_cont1
+            lm_labels[i, 1, : len(with_cont2)] = with_cont2
             mc_labels[i] = mc_label
         all_inputs = (input_ids, mc_token_ids, lm_labels, mc_labels)
         tensor_datasets.append(tuple(torch.tensor(t) for t in all_inputs))
     return tensor_datasets
 
+
 def main():
     parser = argparse.ArgumentParser()
-    parser.add_argument('--model_name', type=str, default='openai-gpt',
-                        help='pretrained model name')
-    parser.add_argument("--do_train", action='store_true', help="Whether to run training.")
-    parser.add_argument("--do_eval", action='store_true', help="Whether to run eval on the dev set.")
-    parser.add_argument("--output_dir", default=None, type=str, required=True,
-                        help="The output directory where the model predictions and checkpoints will be written.")
-    parser.add_argument('--train_dataset', type=str, default='')
-    parser.add_argument('--eval_dataset', type=str, default='')
-    parser.add_argument('--seed', type=int, default=42)
-    parser.add_argument('--num_train_epochs', type=int, default=3)
-    parser.add_argument('--train_batch_size', type=int, default=8)
-    parser.add_argument('--eval_batch_size', type=int, default=16)
-    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
-                        help="Epsilon for Adam optimizer.")
-    parser.add_argument('--max_grad_norm', type=int, default=1)
-    parser.add_argument("--max_steps", default=-1, type=int,
-                        help="If > 0: set total number of training \
-                        steps to perform. Override num_train_epochs.")
-    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
-                        help="Number of updates steps to accumulate before\
-                        performing a backward/update pass.")
-    parser.add_argument('--learning_rate', type=float, default=6.25e-5)
-    parser.add_argument("--warmup_steps", default=0, type=int,
-                        help="Linear warmup over warmup_steps.")
-    parser.add_argument('--lr_schedule', type=str, default='warmup_linear')
-    parser.add_argument('--weight_decay', type=float, default=0.01)
-    parser.add_argument('--lm_coef', type=float, default=0.9)
-    parser.add_argument('--n_valid', type=int, default=374)
+    parser.add_argument("--model_name", type=str, default="openai-gpt", help="pretrained model name")
+    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
+    parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
+    parser.add_argument(
+        "--output_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument("--train_dataset", type=str, default="")
+    parser.add_argument("--eval_dataset", type=str, default="")
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--num_train_epochs", type=int, default=3)
+    parser.add_argument("--train_batch_size", type=int, default=8)
+    parser.add_argument("--eval_batch_size", type=int, default=16)
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", type=int, default=1)
+    parser.add_argument(
+        "--max_steps",
+        default=-1,
+        type=int,
+        help="If > 0: set total number of training \
+                        steps to perform. Override num_train_epochs.",
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before\
+                        performing a backward/update pass.",
+    )
+    parser.add_argument("--learning_rate", type=float, default=6.25e-5)
+    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
+    parser.add_argument("--lr_schedule", type=str, default="warmup_linear")
+    parser.add_argument("--weight_decay", type=float, default=0.01)
+    parser.add_argument("--lm_coef", type=float, default=0.9)
+    parser.add_argument("--n_valid", type=int, default=374)
 
-    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
-    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
+    parser.add_argument("--server_ip", type=str, default="", help="Can be used for distant debugging.")
+    parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")
     args = parser.parse_args()
     print(args)
 
     if args.server_ip and args.server_port:
         # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
         import ptvsd
+
         print("Waiting for debugger attach")
         ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
         ptvsd.wait_for_attach()
@@ -152,15 +170,15 @@ def main():
     # Load tokenizer and model
     # This loading functions also add new tokens and embeddings called `special tokens`
     # These new embeddings will be fine-tuned on the RocStories dataset
-    special_tokens = ['_start_', '_delimiter_', '_classify_']
-    tokenizer = OpenAIGPTTokenizer.from_pretrained(args.model_name, special_tokens=special_tokens)
-    special_tokens_ids = list(tokenizer.convert_tokens_to_ids(token) for token in special_tokens)
-    model = OpenAIGPTDoubleHeadsModel.from_pretrained(args.model_name, num_special_tokens=len(special_tokens))
+    special_tokens = ["_start_", "_delimiter_", "_classify_"]
+    tokenizer = OpenAIGPTTokenizer.from_pretrained(args.model_name)
+    tokenizer.add_tokens(special_tokens)
+    special_tokens_ids = tokenizer.convert_tokens_to_ids(special_tokens)
+    model = OpenAIGPTDoubleHeadsModel.from_pretrained(args.model_name)
+    model.resize_token_embeddings(len(tokenizer))
     model.to(device)
 
     # Load and encode the datasets
-    if not args.train_dataset and not args.eval_dataset:
-        roc_stories = cached_path(ROCSTORIES_URL)
     def tokenize_and_encode(obj):
         """ Tokenize and encode a nested object """
         if isinstance(obj, str):
@@ -168,6 +186,7 @@ def main():
         elif isinstance(obj, int):
             return obj
         return list(tokenize_and_encode(o) for o in obj)
+
     logger.info("Encoding dataset...")
     train_dataset = load_rocstories_dataset(args.train_dataset)
     eval_dataset = load_rocstories_dataset(args.eval_dataset)
@@ -176,8 +195,11 @@ def main():
 
     # Compute the max input length for the Transformer
     max_length = model.config.n_positions // 2 - 2
-    input_length = max(len(story[:max_length]) + max(len(cont1[:max_length]), len(cont2[:max_length])) + 3  \
-                           for dataset in encoded_datasets for story, cont1, cont2, _ in dataset)
+    input_length = max(
+        len(story[:max_length]) + max(len(cont1[:max_length]), len(cont2[:max_length])) + 3
+        for dataset in encoded_datasets
+        for story, cont1, cont2, _ in dataset
+    )
     input_length = min(input_length, model.config.n_positions)  # Max size of input for the pre-trained model
 
     # Prepare inputs tensors and dataloaders
@@ -196,20 +218,23 @@ def main():
     if args.do_train:
         if args.max_steps > 0:
             t_total = args.max_steps
-            args.num_train_epochs = args.max_steps //\
-                (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+            args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
         else:
-            t_total = len(train_dataloader)\
-                // args.gradient_accumulation_steps * args.num_train_epochs
+            t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
 
         param_optimizer = list(model.named_parameters())
-        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+        no_decay = ["bias", "LayerNorm.bias", "LayerNorm.weight"]
         optimizer_grouped_parameters = [
-            {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
-            {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
-            ]
+            {
+                "params": [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
+                "weight_decay": args.weight_decay,
+            },
+            {"params": [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
+        ]
         optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-        scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+        scheduler = get_linear_schedule_with_warmup(
+            optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
+        )
 
     if args.do_train:
         nb_tr_steps, tr_loss, exp_average_loss = 0, 0, None
@@ -221,21 +246,23 @@ def main():
             for step, batch in enumerate(tqdm_bar):
                 batch = tuple(t.to(device) for t in batch)
                 input_ids, mc_token_ids, lm_labels, mc_labels = batch
-                losses = model(input_ids, mc_token_ids, lm_labels, mc_labels)
+                losses = model(input_ids, mc_token_ids=mc_token_ids, lm_labels=lm_labels, mc_labels=mc_labels)
                 loss = args.lm_coef * losses[0] + losses[1]
                 loss.backward()
-                scheduler.step()
                 optimizer.step()
+                scheduler.step()
                 optimizer.zero_grad()
                 tr_loss += loss.item()
-                exp_average_loss = loss.item() if exp_average_loss is None else 0.7*exp_average_loss+0.3*loss.item()
+                exp_average_loss = (
+                    loss.item() if exp_average_loss is None else 0.7 * exp_average_loss + 0.3 * loss.item()
+                )
                 nb_tr_steps += 1
                 tqdm_bar.desc = "Training loss: {:.2e} lr: {:.2e}".format(exp_average_loss, scheduler.get_lr()[0])
 
     # Save a trained model
     if args.do_train:
         # Save a trained model, configuration and tokenizer
-        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
+        model_to_save = model.module if hasattr(model, "module") else model  # Only save the model itself
 
         # If we save using the predefined names, we can load using `from_pretrained`
         output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
@@ -258,10 +285,12 @@ def main():
             batch = tuple(t.to(device) for t in batch)
             input_ids, mc_token_ids, lm_labels, mc_labels = batch
             with torch.no_grad():
-               _, mc_loss, _, mc_logits = model(input_ids, mc_token_ids, lm_labels, mc_labels)
+                _, mc_loss, _, mc_logits = model(
+                    input_ids, mc_token_ids=mc_token_ids, lm_labels=lm_labels, mc_labels=mc_labels
+                )
 
             mc_logits = mc_logits.detach().cpu().numpy()
-            mc_labels = mc_labels.to('cpu').numpy()
+            mc_labels = mc_labels.to("cpu").numpy()
             tmp_eval_accuracy = accuracy(mc_logits, mc_labels)
 
             eval_loss += mc_loss.mean().item()
@@ -272,10 +301,8 @@ def main():
 
         eval_loss = eval_loss / nb_eval_steps
         eval_accuracy = eval_accuracy / nb_eval_examples
-        train_loss = tr_loss/nb_tr_steps if args.do_train else None
-        result = {'eval_loss': eval_loss,
-                  'eval_accuracy': eval_accuracy,
-                  'train_loss': train_loss}
+        train_loss = tr_loss / nb_tr_steps if args.do_train else None
+        result = {"eval_loss": eval_loss, "eval_accuracy": eval_accuracy, "train_loss": train_loss}
 
         output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
         with open(output_eval_file, "w") as writer:
@@ -284,5 +311,6 @@ def main():
                 logger.info("  %s = %s", key, str(result[key]))
                 writer.write("%s = %s\n" % (key, str(result[key])))
 
-if __name__ == '__main__':
+
+if __name__ == "__main__":
     main()

examples/contrib/run_swag.py

@@ -0,0 +1,737 @@
+# 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.
+"""BERT finetuning runner.
+   Finetuning the library models for multiple choice on SWAG (Bert).
+"""
+
+
+import argparse
+import csv
+import glob
+import logging
+import os
+import random
+
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm, trange
+
+from transformers import (
+    WEIGHTS_NAME,
+    AdamW,
+    BertConfig,
+    BertForMultipleChoice,
+    BertTokenizer,
+    get_linear_schedule_with_warmup,
+)
+
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except ImportError:
+    from tensorboardX import SummaryWriter
+
+
+logger = logging.getLogger(__name__)
+
+ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in [BertConfig]), ())
+
+MODEL_CLASSES = {
+    "bert": (BertConfig, BertForMultipleChoice, BertTokenizer),
+}
+
+
+class SwagExample(object):
+    """A single training/test example for the SWAG dataset."""
+
+    def __init__(self, swag_id, context_sentence, start_ending, ending_0, ending_1, ending_2, ending_3, label=None):
+        self.swag_id = swag_id
+        self.context_sentence = context_sentence
+        self.start_ending = start_ending
+        self.endings = [
+            ending_0,
+            ending_1,
+            ending_2,
+            ending_3,
+        ]
+        self.label = label
+
+    def __str__(self):
+        return self.__repr__()
+
+    def __repr__(self):
+        attributes = [
+            "swag_id: {}".format(self.swag_id),
+            "context_sentence: {}".format(self.context_sentence),
+            "start_ending: {}".format(self.start_ending),
+            "ending_0: {}".format(self.endings[0]),
+            "ending_1: {}".format(self.endings[1]),
+            "ending_2: {}".format(self.endings[2]),
+            "ending_3: {}".format(self.endings[3]),
+        ]
+
+        if self.label is not None:
+            attributes.append("label: {}".format(self.label))
+
+        return ", ".join(attributes)
+
+
+class InputFeatures(object):
+    def __init__(self, example_id, choices_features, label):
+        self.example_id = example_id
+        self.choices_features = [
+            {"input_ids": input_ids, "input_mask": input_mask, "segment_ids": segment_ids}
+            for _, input_ids, input_mask, segment_ids in choices_features
+        ]
+        self.label = label
+
+
+def read_swag_examples(input_file, is_training=True):
+    with open(input_file, "r", encoding="utf-8") as f:
+        lines = list(csv.reader(f))
+
+    if is_training and lines[0][-1] != "label":
+        raise ValueError("For training, the input file must contain a label column.")
+
+    examples = [
+        SwagExample(
+            swag_id=line[2],
+            context_sentence=line[4],
+            start_ending=line[5],  # in the swag dataset, the
+            # common beginning of each
+            # choice is stored in "sent2".
+            ending_0=line[7],
+            ending_1=line[8],
+            ending_2=line[9],
+            ending_3=line[10],
+            label=int(line[11]) if is_training else None,
+        )
+        for line in lines[1:]  # we skip the line with the column names
+    ]
+
+    return examples
+
+
+def convert_examples_to_features(examples, tokenizer, max_seq_length, is_training):
+    """Loads a data file into a list of `InputBatch`s."""
+
+    # Swag is a multiple choice task. To perform this task using Bert,
+    # we will use the formatting proposed in "Improving Language
+    # Understanding by Generative Pre-Training" and suggested by
+    # @jacobdevlin-google in this issue
+    # https://github.com/google-research/bert/issues/38.
+    #
+    # Each choice will correspond to a sample on which we run the
+    # inference. For a given Swag example, we will create the 4
+    # following inputs:
+    # - [CLS] context [SEP] choice_1 [SEP]
+    # - [CLS] context [SEP] choice_2 [SEP]
+    # - [CLS] context [SEP] choice_3 [SEP]
+    # - [CLS] context [SEP] choice_4 [SEP]
+    # The model will output a single value for each input. To get the
+    # final decision of the model, we will run a softmax over these 4
+    # outputs.
+    features = []
+    for example_index, example in tqdm(enumerate(examples)):
+        context_tokens = tokenizer.tokenize(example.context_sentence)
+        start_ending_tokens = tokenizer.tokenize(example.start_ending)
+
+        choices_features = []
+        for ending_index, ending in enumerate(example.endings):
+            # We create a copy of the context tokens in order to be
+            # able to shrink it according to ending_tokens
+            context_tokens_choice = context_tokens[:]
+            ending_tokens = start_ending_tokens + tokenizer.tokenize(ending)
+            # Modifies `context_tokens_choice` and `ending_tokens` in
+            # place so that the total length is less than the
+            # specified length.  Account for [CLS], [SEP], [SEP] with
+            # "- 3"
+            _truncate_seq_pair(context_tokens_choice, ending_tokens, max_seq_length - 3)
+
+            tokens = ["[CLS]"] + context_tokens_choice + ["[SEP]"] + ending_tokens + ["[SEP]"]
+            segment_ids = [0] * (len(context_tokens_choice) + 2) + [1] * (len(ending_tokens) + 1)
+
+            input_ids = tokenizer.convert_tokens_to_ids(tokens)
+            input_mask = [1] * len(input_ids)
+
+            # Zero-pad up to the sequence length.
+            padding = [0] * (max_seq_length - len(input_ids))
+            input_ids += padding
+            input_mask += padding
+            segment_ids += padding
+
+            assert len(input_ids) == max_seq_length
+            assert len(input_mask) == max_seq_length
+            assert len(segment_ids) == max_seq_length
+
+            choices_features.append((tokens, input_ids, input_mask, segment_ids))
+
+        label = example.label
+        if example_index < 5:
+            logger.info("*** Example ***")
+            logger.info("swag_id: {}".format(example.swag_id))
+            for choice_idx, (tokens, input_ids, input_mask, segment_ids) in enumerate(choices_features):
+                logger.info("choice: {}".format(choice_idx))
+                logger.info("tokens: {}".format(" ".join(tokens)))
+                logger.info("input_ids: {}".format(" ".join(map(str, input_ids))))
+                logger.info("input_mask: {}".format(" ".join(map(str, input_mask))))
+                logger.info("segment_ids: {}".format(" ".join(map(str, segment_ids))))
+            if is_training:
+                logger.info("label: {}".format(label))
+
+        features.append(InputFeatures(example_id=example.swag_id, choices_features=choices_features, label=label))
+
+    return features
+
+
+def _truncate_seq_pair(tokens_a, tokens_b, max_length):
+    """Truncates a sequence pair in place to the maximum length."""
+
+    # This is a simple heuristic which will always truncate the longer sequence
+    # one token at a time. This makes more sense than truncating an equal percent
+    # of tokens from each, since if one sequence is very short then each token
+    # that's truncated likely contains more information than a longer sequence.
+    while True:
+        total_length = len(tokens_a) + len(tokens_b)
+        if total_length <= max_length:
+            break
+        if len(tokens_a) > len(tokens_b):
+            tokens_a.pop()
+        else:
+            tokens_b.pop()
+
+
+def accuracy(out, labels):
+    outputs = np.argmax(out, axis=1)
+    return np.sum(outputs == labels)
+
+
+def select_field(features, field):
+    return [[choice[field] for choice in feature.choices_features] for feature in features]
+
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+
+def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Load data features from cache or dataset file
+    input_file = args.predict_file if evaluate else args.train_file
+    cached_features_file = os.path.join(
+        os.path.dirname(input_file),
+        "cached_{}_{}_{}".format(
+            "dev" if evaluate else "train",
+            list(filter(None, args.model_name_or_path.split("/"))).pop(),
+            str(args.max_seq_length),
+        ),
+    )
+    if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features = torch.load(cached_features_file)
+    else:
+        logger.info("Creating features from dataset file at %s", input_file)
+        examples = read_swag_examples(input_file)
+        features = convert_examples_to_features(examples, tokenizer, args.max_seq_length, not evaluate)
+
+        if args.local_rank in [-1, 0]:
+            logger.info("Saving features into cached file %s", cached_features_file)
+            torch.save(features, cached_features_file)
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Convert to Tensors and build dataset
+    all_input_ids = torch.tensor(select_field(features, "input_ids"), dtype=torch.long)
+    all_input_mask = torch.tensor(select_field(features, "input_mask"), dtype=torch.long)
+    all_segment_ids = torch.tensor(select_field(features, "segment_ids"), dtype=torch.long)
+    all_label = torch.tensor([f.label for f in features], dtype=torch.long)
+
+    if evaluate:
+        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label)
+    else:
+        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label)
+
+    if output_examples:
+        return dataset, examples, features
+    return dataset
+
+
+def train(args, train_dataset, model, tokenizer):
+    """ Train the model """
+    if args.local_rank in [-1, 0]:
+        tb_writer = SummaryWriter()
+
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+    else:
+        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+
+    # Prepare optimizer and schedule (linear warmup and decay)
+    no_decay = ["bias", "LayerNorm.weight"]
+    optimizer_grouped_parameters = [
+        {
+            "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
+            "weight_decay": args.weight_decay,
+        },
+        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
+    ]
+    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = get_linear_schedule_with_warmup(
+        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
+    )
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(
+            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
+        )
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info(
+        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
+        args.train_batch_size
+        * args.gradient_accumulation_steps
+        * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
+    )
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+
+    global_step = 0
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    set_seed(args)  # Added here for reproductibility
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+            model.train()
+            batch = tuple(t.to(args.device) for t in batch)
+            inputs = {
+                "input_ids": batch[0],
+                "attention_mask": batch[1],
+                # 'token_type_ids':  None if args.model_type == 'xlm' else batch[2],
+                "token_type_ids": batch[2],
+                "labels": batch[3],
+            }
+            # if args.model_type in ['xlnet', 'xlm']:
+            #     inputs.update({'cls_index': batch[5],
+            #                    'p_mask':       batch[6]})
+            outputs = model(**inputs)
+            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
+
+            if args.n_gpu > 1:
+                loss = loss.mean()  # mean() to average on multi-gpu parallel (not distributed) training
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+
+            if args.fp16:
+                with amp.scale_loss(loss, optimizer) as scaled_loss:
+                    scaled_loss.backward()
+                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+            else:
+                loss.backward()
+                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                optimizer.step()
+                scheduler.step()  # Update learning rate schedule
+                model.zero_grad()
+                global_step += 1
+
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    # Log metrics
+                    if (
+                        args.local_rank == -1 and args.evaluate_during_training
+                    ):  # Only evaluate when single GPU otherwise metrics may not average well
+                        results = evaluate(args, model, tokenizer)
+                        for key, value in results.items():
+                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
+                    tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
+                    logging_loss = tr_loss
+
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    # Save model checkpoint
+                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
+                    model_to_save = (
+                        model.module if hasattr(model, "module") else model
+                    )  # Take care of distributed/parallel training
+                    model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_vocabulary(output_dir)
+                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+
+    return global_step, tr_loss / global_step
+
+
+def evaluate(args, model, tokenizer, prefix=""):
+    dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
+
+    if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+        os.makedirs(args.output_dir)
+
+    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+    # Note that DistributedSampler samples randomly
+    eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
+    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+    # Eval!
+    logger.info("***** Running evaluation {} *****".format(prefix))
+    logger.info("  Num examples = %d", len(dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+
+    eval_loss, eval_accuracy = 0, 0
+    nb_eval_steps, nb_eval_examples = 0, 0
+
+    for batch in tqdm(eval_dataloader, desc="Evaluating"):
+        model.eval()
+        batch = tuple(t.to(args.device) for t in batch)
+        with torch.no_grad():
+            inputs = {
+                "input_ids": batch[0],
+                "attention_mask": batch[1],
+                # 'token_type_ids': None if args.model_type == 'xlm' else batch[2]  # XLM don't use segment_ids
+                "token_type_ids": batch[2],
+                "labels": batch[3],
+            }
+
+            # if args.model_type in ['xlnet', 'xlm']:
+            #     inputs.update({'cls_index': batch[4],
+            #                    'p_mask':    batch[5]})
+            outputs = model(**inputs)
+            tmp_eval_loss, logits = outputs[:2]
+            eval_loss += tmp_eval_loss.mean().item()
+
+        logits = logits.detach().cpu().numpy()
+        label_ids = inputs["labels"].to("cpu").numpy()
+        tmp_eval_accuracy = accuracy(logits, label_ids)
+        eval_accuracy += tmp_eval_accuracy
+
+        nb_eval_steps += 1
+        nb_eval_examples += inputs["input_ids"].size(0)
+
+    eval_loss = eval_loss / nb_eval_steps
+    eval_accuracy = eval_accuracy / nb_eval_examples
+    result = {"eval_loss": eval_loss, "eval_accuracy": eval_accuracy}
+
+    output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
+    with open(output_eval_file, "w") as writer:
+        logger.info("***** Eval results *****")
+        for key in sorted(result.keys()):
+            logger.info("%s = %s", key, str(result[key]))
+            writer.write("%s = %s\n" % (key, str(result[key])))
+
+    return result
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    # Required parameters
+    parser.add_argument(
+        "--train_file", default=None, type=str, required=True, help="SWAG csv for training. E.g., train.csv"
+    )
+    parser.add_argument(
+        "--predict_file",
+        default=None,
+        type=str,
+        required=True,
+        help="SWAG csv for predictions. E.g., val.csv or test.csv",
+    )
+    parser.add_argument(
+        "--model_type",
+        default=None,
+        type=str,
+        required=True,
+        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
+    )
+    parser.add_argument(
+        "--model_name_or_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
+    )
+    parser.add_argument(
+        "--output_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The output directory where the model checkpoints and predictions will be written.",
+    )
+
+    # Other parameters
+    parser.add_argument(
+        "--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name"
+    )
+    parser.add_argument(
+        "--tokenizer_name",
+        default="",
+        type=str,
+        help="Pretrained tokenizer name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--max_seq_length",
+        default=384,
+        type=int,
+        help="The maximum total input sequence length after tokenization. Sequences "
+        "longer than this will be truncated, and sequences shorter than this will be padded.",
+    )
+    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
+    parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
+    parser.add_argument(
+        "--evaluate_during_training", action="store_true", help="Rul evaluation during training at each logging step."
+    )
+    parser.add_argument(
+        "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
+    )
+
+    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.")
+    parser.add_argument(
+        "--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation."
+    )
+    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight deay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument(
+        "--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform."
+    )
+    parser.add_argument(
+        "--max_steps",
+        default=-1,
+        type=int,
+        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
+    )
+    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
+
+    parser.add_argument("--logging_steps", type=int, default=50, help="Log every X updates steps.")
+    parser.add_argument("--save_steps", type=int, default=50, help="Save checkpoint every X updates steps.")
+    parser.add_argument(
+        "--eval_all_checkpoints",
+        action="store_true",
+        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
+    )
+    parser.add_argument("--no_cuda", action="store_true", help="Whether not to use CUDA when available")
+    parser.add_argument(
+        "--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory"
+    )
+    parser.add_argument(
+        "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
+    )
+    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
+
+    parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
+    parser.add_argument(
+        "--fp16",
+        action="store_true",
+        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
+    )
+    parser.add_argument(
+        "--fp16_opt_level",
+        type=str,
+        default="O1",
+        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+        "See details at https://nvidia.github.io/apex/amp.html",
+    )
+    parser.add_argument("--server_ip", type=str, default="", help="Can be used for distant debugging.")
+    parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")
+    args = parser.parse_args()
+
+    if (
+        os.path.exists(args.output_dir)
+        and os.listdir(args.output_dir)
+        and args.do_train
+        and not args.overwrite_output_dir
+    ):
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
+                args.output_dir
+            )
+        )
+
+    # Setup distant debugging if needed
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    # Setup CUDA, GPU & distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend="nccl")
+        args.n_gpu = 1
+    args.device = device
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
+    )
+    logger.warning(
+        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+        args.local_rank,
+        device,
+        args.n_gpu,
+        bool(args.local_rank != -1),
+        args.fp16,
+    )
+
+    # Set seed
+    set_seed(args)
+
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    args.model_type = args.model_type.lower()
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
+    tokenizer = tokenizer_class.from_pretrained(
+        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case
+    )
+    model = model_class.from_pretrained(
+        args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path), config=config
+    )
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    model.to(args.device)
+
+    logger.info("Training/evaluation parameters %s", args)
+
+    # Training
+    if args.do_train:
+        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+
+    # Save the trained model and the tokenizer
+    if args.local_rank == -1 or torch.distributed.get_rank() == 0:
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = (
+            model.module if hasattr(model, "module") else model
+        )  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+
+        # Good practice: save your training arguments together with the trained model
+        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
+
+        # Load a trained model and vocabulary that you have fine-tuned
+        model = model_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
+        model.to(args.device)
+
+    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        if args.do_train:
+            checkpoints = [args.output_dir]
+        else:
+            # if do_train is False and do_eval is true, load model directly from pretrained.
+            checkpoints = [args.model_name_or_path]
+
+        if args.eval_all_checkpoints:
+            checkpoints = list(
+                os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
+            )
+            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs
+
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+
+        for checkpoint in checkpoints:
+            # Reload the model
+            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
+            model = model_class.from_pretrained(checkpoint)
+            tokenizer = tokenizer_class.from_pretrained(checkpoint)
+            model.to(args.device)
+
+            # Evaluate
+            result = evaluate(args, model, tokenizer, prefix=global_step)
+
+            result = dict((k + ("_{}".format(global_step) if global_step else ""), v) for k, v in result.items())
+            results.update(result)
+
+    logger.info("Results: {}".format(results))
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

examples/contrib/run_transfo_xl.py

@@ -19,55 +19,48 @@
 
     This script with default values evaluates a pretrained Transformer-XL on WikiText 103
 """
-from __future__ import absolute_import, division, print_function, unicode_literals
+
 
 import argparse
 import logging
-import time
 import math
+import time
 
 import torch
 
-from pytorch_transformers import TransfoXLLMHeadModel, TransfoXLCorpus, TransfoXLTokenizer
+from transformers import TransfoXLCorpus, TransfoXLLMHeadModel
 
-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__)
 
+
 def main():
-    parser = argparse.ArgumentParser(description='PyTorch Transformer Language Model')
-    parser.add_argument('--model_name', type=str, default='transfo-xl-wt103',
-                        help='pretrained model name')
-    parser.add_argument('--split', type=str, default='test',
-                        choices=['all', 'valid', 'test'],
-                        help='which split to evaluate')
-    parser.add_argument('--batch_size', type=int, default=10,
-                        help='batch size')
-    parser.add_argument('--tgt_len', type=int, default=128,
-                        help='number of tokens to predict')
-    parser.add_argument('--ext_len', type=int, default=0,
-                        help='length of the extended context')
-    parser.add_argument('--mem_len', type=int, default=1600,
-                        help='length of the retained previous heads')
-    parser.add_argument('--clamp_len', type=int, default=1000,
-                        help='max positional embedding index')
-    parser.add_argument('--no_cuda', action='store_true',
-                        help='Do not use CUDA even though CUA is available')
-    parser.add_argument('--work_dir', type=str, required=True,
-                        help='path to the work_dir')
-    parser.add_argument('--no_log', action='store_true',
-                        help='do not log the eval result')
-    parser.add_argument('--same_length', action='store_true',
-                        help='set same length attention with masking')
-    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
-    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
+    parser = argparse.ArgumentParser(description="PyTorch Transformer Language Model")
+    parser.add_argument("--model_name", type=str, default="transfo-xl-wt103", help="pretrained model name")
+    parser.add_argument(
+        "--split", type=str, default="test", choices=["all", "valid", "test"], help="which split to evaluate"
+    )
+    parser.add_argument("--batch_size", type=int, default=10, help="batch size")
+    parser.add_argument("--tgt_len", type=int, default=128, help="number of tokens to predict")
+    parser.add_argument("--ext_len", type=int, default=0, help="length of the extended context")
+    parser.add_argument("--mem_len", type=int, default=1600, help="length of the retained previous heads")
+    parser.add_argument("--clamp_len", type=int, default=1000, help="max positional embedding index")
+    parser.add_argument("--no_cuda", action="store_true", help="Do not use CUDA even though CUA is available")
+    parser.add_argument("--work_dir", type=str, required=True, help="path to the work_dir")
+    parser.add_argument("--no_log", action="store_true", help="do not log the eval result")
+    parser.add_argument("--same_length", action="store_true", help="set same length attention with masking")
+    parser.add_argument("--server_ip", type=str, default="", help="Can be used for distant debugging.")
+    parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")
     args = parser.parse_args()
-    assert args.ext_len >= 0, 'extended context length must be non-negative'
+    assert args.ext_len >= 0, "extended context length must be non-negative"
 
     if args.server_ip and args.server_port:
         # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
         import ptvsd
+
         print("Waiting for debugger attach")
         ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
         ptvsd.wait_for_attach()
@@ -80,21 +73,20 @@ def main():
     # The pre-processing involve computing word frequencies to prepare the Adaptive input and SoftMax
     # and tokenizing the dataset
     # The pre-processed corpus is a convertion (using the conversion script )
-    tokenizer = TransfoXLTokenizer.from_pretrained(args.model_name)
     corpus = TransfoXLCorpus.from_pretrained(args.model_name)
-    ntokens = len(corpus.vocab)
 
-    va_iter = corpus.get_iterator('valid', args.batch_size, args.tgt_len,
-        device=device, ext_len=args.ext_len)
-    te_iter = corpus.get_iterator('test', args.batch_size, args.tgt_len,
-        device=device, ext_len=args.ext_len)
+    va_iter = corpus.get_iterator("valid", args.batch_size, args.tgt_len, device=device, ext_len=args.ext_len)
+    te_iter = corpus.get_iterator("test", args.batch_size, args.tgt_len, device=device, ext_len=args.ext_len)
 
     # Load a pre-trained model
     model = TransfoXLLMHeadModel.from_pretrained(args.model_name)
     model = model.to(device)
 
-    logger.info('Evaluating with bsz {} tgt_len {} ext_len {} mem_len {} clamp_len {}'.format(
-        args.batch_size, args.tgt_len, args.ext_len, args.mem_len, args.clamp_len))
+    logger.info(
+        "Evaluating with bsz {} tgt_len {} ext_len {} mem_len {} clamp_len {}".format(
+            args.batch_size, args.tgt_len, args.ext_len, args.mem_len, args.clamp_len
+        )
+    )
 
     model.reset_length(args.tgt_len, args.ext_len, args.mem_len)
     if args.clamp_len > 0:
@@ -108,46 +100,45 @@ def main():
     def evaluate(eval_iter):
         # Turn on evaluation mode which disables dropout.
         model.eval()
-        total_len, total_loss = 0, 0.
+        total_len, total_loss = 0, 0.0
         start_time = time.time()
         with torch.no_grad():
             mems = None
             for idx, (data, target, seq_len) in enumerate(eval_iter):
-                ret = model(data, target, mems)
+                ret = model(data, lm_labels=target, mems=mems)
                 loss, _, mems = ret
                 loss = loss.mean()
                 total_loss += seq_len * loss.item()
                 total_len += seq_len
             total_time = time.time() - start_time
-        logger.info('Time : {:.2f}s, {:.2f}ms/segment'.format(
-                total_time, 1000 * total_time / (idx+1)))
+        logger.info("Time : {:.2f}s, {:.2f}ms/segment".format(total_time, 1000 * total_time / (idx + 1)))
         return total_loss / total_len
 
     # Run on test data.
-    if args.split == 'all':
+    if args.split == "all":
         test_loss = evaluate(te_iter)
         valid_loss = evaluate(va_iter)
-    elif args.split == 'valid':
+    elif args.split == "valid":
         valid_loss = evaluate(va_iter)
         test_loss = None
-    elif args.split == 'test':
+    elif args.split == "test":
         test_loss = evaluate(te_iter)
         valid_loss = None
 
     def format_log(loss, split):
-        log_str = '| {0} loss {1:5.2f} | {0} ppl {2:9.3f} '.format(
-            split, loss, math.exp(loss))
+        log_str = "| {0} loss {1:5.2f} | {0} ppl {2:9.3f} ".format(split, loss, math.exp(loss))
         return log_str
 
-    log_str = ''
+    log_str = ""
     if valid_loss is not None:
-        log_str += format_log(valid_loss, 'valid')
+        log_str += format_log(valid_loss, "valid")
     if test_loss is not None:
-        log_str += format_log(test_loss, 'test')
+        log_str += format_log(test_loss, "test")
 
-    logger.info('=' * 100)
+    logger.info("=" * 100)
     logger.info(log_str)
-    logger.info('=' * 100)
+    logger.info("=" * 100)
 
-if __name__ == '__main__':
+
+if __name__ == "__main__":
     main()

examples/distillation/README.md

@@ -1,35 +1,99 @@
-# DistilBERT
+# Distil*
 
-This folder contains the original code used to train DistilBERT as well as examples showcasing how to use DistilBERT.
+This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT, DistilRoBERTa and DistilGPT2.
 
-## What is DistilBERT
+**January 20, 2020 - Bug fixing** We have recently discovered and fixed [a bug](https://github.com/huggingface/transformers/commit/48cbf267c988b56c71a2380f748a3e6092ccaed3) in the evaluation of our `run_*.py` scripts that caused the reported metrics to be over-estimated on average. We have updated all the metrics with the latest runs.
 
-DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving over 95% of Bert's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
+**December 6, 2019 - Update** We release **DistilmBERT**: 92% of `bert-base-multilingual-cased` on XNLI. The model supports 104 different languages listed [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages).
+
+**November 19, 2019 - Update** We release German **DistilBERT**: 98.8% of `bert-base-german-dbmdz-cased` on NER tasks.
+
+**October 23, 2019 - Update** We release **DistilRoBERTa**: 95% of `RoBERTa-base`'s performance on GLUE, twice as fast as RoBERTa while being 35% smaller.
+
+**October 3, 2019 - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2. **The paper supersedes our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances. Please use the paper as a reference when comparing/reporting results on DistilBERT.**
+
+**September 19, 2019 - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 99% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
+
+
+## What is Distil*
+
+Distil* is a class of compressed models that started with DistilBERT. DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
+
+We have applied the same method to other Transformer architectures and released the weights:
+- GPT2: on the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 16.3 compared to 21.1 for **DistilGPT2** (after fine-tuning on the train set).
+- RoBERTa: **DistilRoBERTa** reaches 95% of `RoBERTa-base`'s performance on GLUE while being twice faster and 35% smaller.
+- German BERT: **German DistilBERT** reaches 99% of `bert-base-german-dbmdz-cased`'s performance on German NER (CoNLL-2003).
+- Multilingual BERT: **DistilmBERT** reaches 92% of Multilingual BERT's performance on XNLI while being twice faster and 25% smaller. The model supports 104 languages listed [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages).
+
+For more information on DistilBERT, please refer to our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108).
+
+Here are the results on the dev sets of GLUE:
+
+| Model                     | Macro-score                    | CoLA | MNLI | MRPC | QNLI | QQP  | RTE  | SST-2| STS-B| WNLI              |
+| :---:                     |    :---:                       | :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:             |
+| BERT-base-uncased         |  **79.5**                      | 56.3 | 84.7 | 88.6 | 91.8 | 89.6 | 69.3 | 92.7 | 89.0 | 53.5              |
+| DistilBERT-base-uncased   |  **77.0**                      | 51.3 | 82.1 | 87.5 | 89.2 | 88.5 | 59.9 | 91.3 | 86.9 | 56.3              |
+| BERT-base-cased           |  **78.2**                      | 58.2 | 83.9 | 87.8 | 91.0 | 89.2 | 66.1 | 91.7 | 89.2 | 46.5              |
+| DistilBERT-base-cased     |  **75.9**                      | 47.2 | 81.5 | 85.6 | 88.2 | 87.8 | 60.6 | 90.4 | 85.5 | 56.3              |
+| ---                       |    ---                         |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  ---              |
+| RoBERTa-base (reported)   |  **83.2**/**86.4**<sup>2</sup> | 63.6 | 87.6 | 90.2 | 92.8 | 91.9 | 78.7 | 94.8 | 91.2 | 57.7<sup>3</sup>  |
+| DistilRoBERTa<sup>1</sup> |  **79.0**/**82.3**<sup>2</sup> | 59.3 | 84.0 | 86.6 | 90.8 | 89.4 | 67.9 | 92.5 | 88.3 | 52.1              |
+
+<sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directly perform transfer learning on the pre-trained DistilRoBERTa.
+
+<sup>2</sup> Macro-score computed without WNLI.
+
+<sup>3</sup> We compute this score ourselves for completeness.
+
+Here are the results on the *test* sets for 6 of the languages available in XNLI. The results are computed in the zero shot setting (trained on the English portion and evaluated on the target language portion):
+
+| Model                        | English | Spanish | Chinese | German | Arabic  | Urdu |
+| :---:                        | :---:   | :---:   | :---:   | :---:  | :---:   | :---:|
+| mBERT base cased (computed)  | 82.1    | 74.6    | 69.1    | 72.3   | 66.4    | 58.5 |
+| mBERT base uncased (reported)| 81.4    | 74.3    | 63.8    | 70.5   | 62.1    | 58.3 |
+| DistilmBERT                  | 78.2    | 69.1    | 64.0    | 66.3   | 59.1    | 54.7 |
+
+## Setup
+
+This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`. 
+
+**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0).
 
-For more information on DistilBERT, please refer to our [detailed blog post](https://medium.com/huggingface/smaller-faster-cheaper-lighter-introducing-distilbert-a-distilled-version-of-bert-8cf3380435b5
-).
 
 ## How to use DistilBERT
 
-PyTorch-Transformers includes two pre-trained DistilBERT models, currently only provided for English (we are investigating the possibility to train and release a multilingual version of DistilBERT):
+Transformers includes five pre-trained Distil* models, currently only provided for English and German (we are investigating the possibility to train and release a multilingual version of DistilBERT):
 
 - `distilbert-base-uncased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-uncased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 66M parameters.
-- `distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.2 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
+- `distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knowledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.9 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
+- `distilbert-base-cased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-cased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 65M parameters.
+- `distilbert-base-cased-distilled-squad`: A finetuned version of `distilbert-base-cased` finetuned using (a second step of) knowledge distillation on SQuAD 1.0. This model reaches a F1 score of 87.1 on the dev set (for comparison, Bert `bert-base-cased` version reaches a 88.7 F1 score).
+- `distilbert-base-german-cased`: DistilBERT German language model pretrained on 1/2 of the data used to pretrain Bert using distillation with the supervision of the `bert-base-german-dbmdz-cased` version of German DBMDZ Bert. For NER tasks the model reaches a F1 score of 83.49 on the CoNLL-2003 test set (for comparison, `bert-base-german-dbmdz-cased` reaches a 84.52 F1 score), and a F1 score of 85.23 on the GermEval 2014 test set (`bert-base-german-dbmdz-cased` reaches a 86.89 F1 score).
+- `distilgpt2`: DistilGPT2 English language model pretrained with the supervision of `gpt2` (the smallest version of GPT2) on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset. The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
+- `distilroberta-base`: DistilRoBERTa English language model pretrained with the supervision of `roberta-base` solely on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset (it is ~4 times less training data than the teacher RoBERTa). The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 125M parameters for RoBERTa-base). On average DistilRoBERTa is twice as fast as Roberta-base.
+- `distilbert-base-multilingual-cased`: DistilmBERT multilingual model pretrained with the supervision of `bert-base-multilingual-cased` on the concatenation of Wikipedia in 104 different languages. The model supports the 104 languages listed [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages). The model has 6 layers, 768 dimension and 12 heads, totalizing 134M parameters (compared to 177M parameters for mBERT-base). On average DistilmBERT is twice as fast as mBERT-base.
 
 Using DistilBERT is very similar to using BERT. DistilBERT share the same tokenizer as BERT's `bert-base-uncased` even though we provide a link to this tokenizer under the `DistilBertTokenizer` name to have a consistent naming between the library models.
 
 ```python
-tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
-model = DistilBertModel.from_pretrained('distilbert-base-uncased')
+tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-cased')
+model = DistilBertModel.from_pretrained('distilbert-base-cased')
 
 input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)
 outputs = model(input_ids)
 last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 ```
 
-## How to train DistilBERT
+Similarly, using the other Distil* models simply consists in calling the base classes with a different pretrained checkpoint:
+- DistilBERT uncased: `model = DistilBertModel.from_pretrained('distilbert-base-uncased')`
+- DistilGPT2: `model = GPT2Model.from_pretrained('distilgpt2')`
+- DistilRoBERTa: `model = RobertaModel.from_pretrained('distilroberta-base')`
+- DistilmBERT: `model = DistilBertModel.from_pretrained('distilbert-base-multilingual-cased')`
 
-In the following, we will explain how you can train your own compressed model.
+
+## How to train Distil*
+
+In the following, we will explain how you can train DistilBERT.
 
 ### A. Preparing the data
 
@@ -42,16 +106,18 @@ First, we will binarize the data, i.e. tokenize the data and convert each token
 ```bash
 python scripts/binarized_data.py \
     --file_path data/dump.txt \
-    --bert_tokenizer bert-base-uncased \
+    --tokenizer_type bert \
+    --tokenizer_name bert-base-uncased \
     --dump_file data/binarized_text
 ```
 
-Our implementation of masked language modeling loss follows [XLM](https://github.com/facebookresearch/XLM)'s one and smoothes the probability of masking with a factor that put more emphasis on rare words. Thus we count the occurences of each tokens in the data:
+Our implementation of masked language modeling loss follows [XLM](https://github.com/facebookresearch/XLM)'s one and smoothes the probability of masking with a factor that put more emphasis on rare words. Thus we count the occurrences of each tokens in the data:
 
 ```bash
 python scripts/token_counts.py \
     --data_file data/binarized_text.bert-base-uncased.pickle \
-    --token_counts_dump data/token_counts.bert-base-uncased.pickle
+    --token_counts_dump data/token_counts.bert-base-uncased.pickle \
+    --vocab_size 30522
 ```
 
 ### B. Training
@@ -60,6 +126,12 @@ Training with distillation is really simple once you have pre-processed the data
 
 ```bash
 python train.py \
+    --student_type distilbert \
+    --student_config training_configs/distilbert-base-uncased.json \
+    --teacher_type bert \
+    --teacher_name bert-base-uncased \
+    --alpha_ce 5.0 --alpha_mlm 2.0 --alpha_cos 1.0 --alpha_clm 0.0 --mlm \
+    --freeze_pos_embs \
     --dump_path serialization_dir/my_first_training \
     --data_file data/binarized_text.bert-base-uncased.pickle \
     --token_counts data/token_counts.bert-base-uncased.pickle \
@@ -68,7 +140,7 @@ python train.py \
 
 By default, this will launch a training on a single GPU (even if more are available on the cluster). Other parameters are available in the command line, please look in `train.py` or run `python train.py --help` to list them.
 
-We highly encourage you to use distributed training for training DistilBert as the training corpus is quite large. Here's an example that runs a distributed training on a single node having 4 GPUs:
+We highly encourage you to use distributed training for training DistilBERT as the training corpus is quite large. Here's an example that runs a distributed training on a single node having 4 GPUs:
 
 ```bash
 export NODE_RANK=0
@@ -90,11 +162,30 @@ python -m torch.distributed.launch \
     train.py \
         --force \
         --n_gpu $WORLD_SIZE \
+        --student_type distilbert \
+        --student_config training_configs/distilbert-base-uncased.json \
+        --teacher_type bert \
+        --teacher_name bert-base-uncased \
+        --alpha_ce 0.33 --alpha_mlm 0.33 --alpha_cos 0.33 --alpha_clm 0.0 --mlm \
+        --freeze_pos_embs \
+        --dump_path serialization_dir/my_first_training \
         --data_file data/binarized_text.bert-base-uncased.pickle \
-        --token_counts data/token_counts.bert-base-uncased.pickle \
-        --dump_path serialization_dir/my_first_distillation
+        --token_counts data/token_counts.bert-base-uncased.pickle
 ```
 
-**Tips:** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract_for_distil.py` to create a valid initialization checkpoint and use `--from_pretrained_weights` and `--from_pretrained_config` arguments to use this initialization for the distilled training!
+**Tips:** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract.py` and `scripts/extract_distilbert.py` to create a valid initialization checkpoint and use `--student_pretrained_weights` argument to use this initialization for the distilled training!
 
 Happy distillation!
+
+## Citation
+
+If you find the resource useful, you should cite the following paper:
+
+```
+@inproceedings{sanh2019distilbert,
+  title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
+  author={Sanh, Victor and Debut, Lysandre and Chaumond, Julien and Wolf, Thomas},
+  booktitle={NeurIPS EMC^2 Workshop},
+  year={2019}
+}
+```

examples/distillation/dataset.py

@@ -1,201 +0,0 @@
-# coding=utf-8
-# Copyright 2019-present, the HuggingFace Inc. team and Facebook, 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.
-""" Dataloaders to train DistilBERT
-    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
-"""
-from typing import List
-import math
-from itertools import chain
-from collections import Counter
-import numpy as np
-import torch
-
-from utils import logger
-
-class Dataset:
-    def __init__(self,
-                 params,
-                 data):
-        self.params = params
-        self.tokens_per_batch = params.tokens_per_batch
-        self.batch_size = params.batch_size
-        self.shuffle = params.shuffle
-        self.group_by_size = params.group_by_size
-
-        self.token_ids = np.array(data)
-        self.lengths = np.uint16([len(t) for t in data])
-
-        self.check()
-        self.remove_long_sequences()
-        self.remove_empty_sequences()
-        self.check()
-        self.print_statistics()
-
-    def __len__(self):
-        return len(self.lengths)
-
-    def check(self):
-        """
-        Some sanity checks
-        """
-        assert len(self.token_ids) == len(self.lengths)
-
-    def remove_long_sequences(self):
-        """
-        Sequences that are too long are splitted by chunk of max_position_embeddings.
-        """
-        indices = self.lengths >= self.params.max_position_embeddings
-        logger.info(f'Splitting {sum(indices)} too long sequences.')
-
-        def divide_chunks(l, n):
-            return [l[i:i + n] for i in range(0, len(l), n)]
-
-        new_tok_ids = []
-        new_lengths = []
-        cls_id, sep_id = self.params.special_tok_ids['cls_token'], self.params.special_tok_ids['sep_token']
-        max_len = self.params.max_position_embeddings
-
-        for seq_, len_ in zip(self.token_ids, self.lengths):
-            if len_ <= max_len:
-                new_tok_ids.append(seq_)
-                new_lengths.append(len_)
-            else:
-                sub_seqs = []
-                for sub_s in divide_chunks(seq_, max_len-2):
-                    if sub_s[0] != cls_id:
-                        sub_s = np.insert(sub_s, 0, cls_id)
-                    if sub_s[-1] != sep_id:
-                        sub_s = np.insert(sub_s, len(sub_s), cls_id)
-                    assert len(sub_s) <= max_len
-                    sub_seqs.append(sub_s)
-
-                new_tok_ids.extend(sub_seqs)
-                new_lengths.extend([len(l) for l in sub_seqs])
-
-        self.token_ids = np.array(new_tok_ids)
-        self.lengths = np.array(new_lengths)
-
-    def remove_empty_sequences(self):
-        """
-        Too short sequences are simply removed. This could be tunedd.
-        """
-        init_size = len(self)
-        indices = self.lengths > 5
-        self.token_ids = self.token_ids[indices]
-        self.lengths = self.lengths[indices]
-        new_size = len(self)
-        logger.info(f'Remove {init_size - new_size} too short (<=5 tokens) sequences.')
-
-    def print_statistics(self):
-        """
-        Print some statistics on the corpus. Only the master process.
-        """
-        if not self.params.is_master:
-            return
-        logger.info(f'{len(self)} sequences')
-        # data_len = sum(self.lengths)
-        # nb_unique_tokens = len(Counter(list(chain(*self.token_ids))))
-        # logger.info(f'{data_len} tokens ({nb_unique_tokens} unique)')
-
-        # unk_idx = self.params.special_tok_ids['unk_token']
-        # nb_unkown = sum([(t==unk_idx).sum() for t in self.token_ids])
-        # logger.info(f'{nb_unkown} unknown tokens (covering {100*nb_unkown/data_len:.2f}% of the data)')
-
-    def select_data(self, a: int, b: int):
-        """
-        Select a subportion of the data.
-        """
-        n_sequences = len(self)
-        assert 0 <= a < b <= n_sequences, ValueError(f'`0 <= a < b <= n_sequences` is not met with a={a} and b={b}')
-
-        logger.info(f'Selecting sequences from {a} to {b} (excluded).')
-        self.token_ids = self.token_ids[a:b]
-        self.lengths = self.lengths[a:b]
-
-        self.check()
-
-    def split(self):
-        """
-        Distributed training: split the data accross the processes.
-        """
-        assert self.params.n_gpu > 1
-        logger.info('Splitting the data accross the processuses.')
-        n_seq = len(self)
-        n_seq_per_procesus = n_seq // self.params.world_size
-        a = n_seq_per_procesus * self.params.global_rank
-        b = a + n_seq_per_procesus
-        self.select_data(a=a, b=b)
-
-    def batch_sequences(self,
-                        token_ids: List[List[int]],
-                        lengths: List[int]):
-        """
-        Do the padding and transform into torch.tensor.
-        """
-        assert len(token_ids) == len(lengths)
-
-        # Max for paddings
-        max_seq_len_ = max(lengths)
-
-        # Pad token ids
-        pad_idx = self.params.special_tok_ids['pad_token']
-        tk_ = [list(t.astype(int)) + [pad_idx]*(max_seq_len_-len(t)) for t in token_ids]
-        assert len(tk_) == len(token_ids)
-        assert all(len(t) == max_seq_len_ for t in tk_)
-
-        tk_t = torch.tensor(tk_)                  # (bs, max_seq_len_)
-        lg_t = torch.tensor(lengths.astype(int))  # (bs)
-        return tk_t, lg_t
-
-    def get_batches_iterator(self,
-                             batches):
-        """
-        Return an iterator over batches.
-        """
-        for sequences_ids in batches:
-            token_ids, lengths = self.batch_sequences(self.token_ids[sequences_ids],
-                                                    self.lengths[sequences_ids])
-            yield (token_ids, lengths)
-
-    def get_iterator(self,
-                     seed: int = None):
-        """
-        Return a data iterator.
-        """
-        rng = np.random.RandomState(seed)
-
-        n_sequences = len(self)
-        indices = np.arange(n_sequences)
-
-        if self.group_by_size:
-            indices = indices[np.argsort(self.lengths[indices], kind='mergesort')]
-
-        if self.tokens_per_batch == -1:
-            batches = np.array_split(indices, math.ceil(len(indices) * 1. / self.batch_size))
-        else:
-            assert self.tokens_per_batch > 0
-            batch_ids = np.cumsum(self.lengths[indices]) // self.tokens_per_batch
-            _, bounds = np.unique(batch_ids, return_index=True)
-            batches = [indices[bounds[i]:bounds[i + 1]] for i in range(len(bounds) - 1)]
-            if bounds[-1] < len(indices):
-                batches.append(indices[bounds[-1]:])
-
-        if self.shuffle:
-            rng.shuffle(batches)
-
-        assert n_sequences == sum([len(x) for x in batches])
-        assert self.lengths[indices].sum() == sum([self.lengths[x].sum() for x in batches])
-
-        return self.get_batches_iterator(batches=batches)

examples/distillation/distiller.py

@@ -12,32 +12,39 @@
 # 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.
-""" The distiller to distil DistilBERT
-    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
+""" The distiller to distil the student.
+    Adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
 """
-import os
 import math
-from tensorboardX import SummaryWriter
-from tqdm import trange, tqdm
-import numpy as np
+import os
+import time
 
+import psutil
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from torch.optim import AdamW
+from torch.utils.data import BatchSampler, DataLoader, RandomSampler
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm
 
-from pytorch_transformers import AdamW, WarmupLinearSchedule
-
+from grouped_batch_sampler import GroupedBatchSampler, create_lengths_groups
+from lm_seqs_dataset import LmSeqsDataset
+from transformers import get_linear_schedule_with_warmup
 from utils import logger
-from dataset import Dataset
+
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except ImportError:
+    from tensorboardX import SummaryWriter
+
 
 class Distiller:
-    def __init__(self,
-                 params: dict,
-                 dataloader: Dataset,
-                 token_probs: torch.tensor,
-                 student: nn.Module,
-                 teacher: nn.Module):
-        logger.info('Initializing Distiller')
+    def __init__(
+        self, params: dict, dataset: LmSeqsDataset, token_probs: torch.tensor, student: nn.Module, teacher: nn.Module
+    ):
+        logger.info("Initializing Distiller")
         self.params = params
         self.dump_path = params.dump_path
         self.multi_gpu = params.multi_gpu
@@ -46,31 +53,45 @@ class Distiller:
         self.student = student
         self.teacher = teacher
 
-        self.dataloader = dataloader
-        if self.params.n_gpu > 1:
-            self.dataloader.split()
-        self.get_iterator(seed=params.seed)
+        self.student_config = student.config
+        self.vocab_size = student.config.vocab_size
+
+        if params.n_gpu <= 1:
+            sampler = RandomSampler(dataset)
+        else:
+            sampler = DistributedSampler(dataset)
+
+        if params.group_by_size:
+            groups = create_lengths_groups(lengths=dataset.lengths, k=params.max_model_input_size)
+            sampler = GroupedBatchSampler(sampler=sampler, group_ids=groups, batch_size=params.batch_size)
+        else:
+            sampler = BatchSampler(sampler=sampler, batch_size=params.batch_size, drop_last=False)
+
+        self.dataloader = DataLoader(dataset=dataset, batch_sampler=sampler, collate_fn=dataset.batch_sequences)
 
         self.temperature = params.temperature
-        assert self.temperature > 0.
+        assert self.temperature > 0.0
 
         self.alpha_ce = params.alpha_ce
         self.alpha_mlm = params.alpha_mlm
+        self.alpha_clm = params.alpha_clm
         self.alpha_mse = params.alpha_mse
-        assert self.alpha_ce >= 0.
-        assert self.alpha_mlm >= 0.
-        assert self.alpha_mse >= 0.
-        assert self.alpha_ce + self.alpha_mlm + self.alpha_mse > 0.
+        self.alpha_cos = params.alpha_cos
 
-        self.mlm_mask_prop = params.mlm_mask_prop
-        assert 0.0 <= self.mlm_mask_prop <= 1.0
-        assert params.word_mask + params.word_keep + params.word_rand == 1.0
-        self.pred_probs = torch.FloatTensor([params.word_mask, params.word_keep, params.word_rand])
-        self.pred_probs = self.pred_probs.to(f'cuda:{params.local_rank}') if params.n_gpu > 0 else self.pred_probs
-        self.token_probs = token_probs.to(f'cuda:{params.local_rank}') if params.n_gpu > 0 else token_probs
-        if self.fp16:
-            self.pred_probs = self.pred_probs.half()
-            self.token_probs = self.token_probs.half()
+        self.mlm = params.mlm
+        if self.mlm:
+            logger.info(f"Using MLM loss for LM step.")
+            self.mlm_mask_prop = params.mlm_mask_prop
+            assert 0.0 <= self.mlm_mask_prop <= 1.0
+            assert params.word_mask + params.word_keep + params.word_rand == 1.0
+            self.pred_probs = torch.FloatTensor([params.word_mask, params.word_keep, params.word_rand])
+            self.pred_probs = self.pred_probs.to(f"cuda:{params.local_rank}") if params.n_gpu > 0 else self.pred_probs
+            self.token_probs = token_probs.to(f"cuda:{params.local_rank}") if params.n_gpu > 0 else token_probs
+            if self.fp16:
+                self.pred_probs = self.pred_probs.half()
+                self.token_probs = self.token_probs.half()
+        else:
+            logger.info(f"Using CLM loss for LM step.")
 
         self.epoch = 0
         self.n_iter = 0
@@ -80,32 +101,55 @@ class Distiller:
         self.last_loss = 0
         self.last_loss_ce = 0
         self.last_loss_mlm = 0
-        self.last_loss_mse = 0
+        self.last_loss_clm = 0
+        if self.alpha_mse > 0.0:
+            self.last_loss_mse = 0
+        if self.alpha_cos > 0.0:
+            self.last_loss_cos = 0
+        self.last_log = 0
 
-        self.ce_loss_fct = nn.KLDivLoss(reduction='batchmean')
-        self.mlm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
-        self.mse_loss_fct = nn.MSELoss(reduction='sum')
+        self.ce_loss_fct = nn.KLDivLoss(reduction="batchmean")
+        self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
+        if self.alpha_mse > 0.0:
+            self.mse_loss_fct = nn.MSELoss(reduction="sum")
+        if self.alpha_cos > 0.0:
+            self.cosine_loss_fct = nn.CosineEmbeddingLoss(reduction="mean")
 
-        logger.info('--- Initializing model optimizer')
+        logger.info("--- Initializing model optimizer")
         assert params.gradient_accumulation_steps >= 1
-        self.num_steps_epoch = int(len(self.dataloader) / params.batch_size) + 1
-        num_train_optimization_steps = int(self.num_steps_epoch / params.gradient_accumulation_steps * params.n_epoch) + 1
-        warmup_steps = math.ceil(num_train_optimization_steps * params.warmup_prop)
+        self.num_steps_epoch = len(self.dataloader)
+        num_train_optimization_steps = (
+            int(self.num_steps_epoch / params.gradient_accumulation_steps * params.n_epoch) + 1
+        )
 
-        no_decay = ['bias', 'LayerNorm.weight']
+        no_decay = ["bias", "LayerNorm.weight"]
         optimizer_grouped_parameters = [
-            {'params': [p for n, p in student.named_parameters() if not any(nd in n for nd in no_decay) and p.requires_grad], 'weight_decay': params.weight_decay},
-            {'params': [p for n, p in student.named_parameters() if any(nd in n for nd in no_decay) and p.requires_grad], 'weight_decay': 0.0}
+            {
+                "params": [
+                    p for n, p in student.named_parameters() if not any(nd in n for nd in no_decay) and p.requires_grad
+                ],
+                "weight_decay": params.weight_decay,
+            },
+            {
+                "params": [
+                    p for n, p in student.named_parameters() if any(nd in n for nd in no_decay) and p.requires_grad
+                ],
+                "weight_decay": 0.0,
+            },
         ]
-        logger.info("------ Number of trainable parameters (student): %i" % sum([p.numel() for p in self.student.parameters() if p.requires_grad]))
+        logger.info(
+            "------ Number of trainable parameters (student): %i"
+            % sum([p.numel() for p in self.student.parameters() if p.requires_grad])
+        )
         logger.info("------ Number of parameters (student): %i" % sum([p.numel() for p in self.student.parameters()]))
-        self.optimizer = AdamW(optimizer_grouped_parameters,
-                               lr=params.learning_rate,
-                               eps=params.adam_epsilon,
-                               betas=(0.9, 0.98))
-        self.scheduler = WarmupLinearSchedule(self.optimizer,
-                                              warmup_steps=warmup_steps,
-                                              t_total=num_train_optimization_steps)
+        self.optimizer = AdamW(
+            optimizer_grouped_parameters, lr=params.learning_rate, eps=params.adam_epsilon, betas=(0.9, 0.98)
+        )
+
+        warmup_steps = math.ceil(num_train_optimization_steps * params.warmup_prop)
+        self.scheduler = get_linear_schedule_with_warmup(
+            self.optimizer, num_warmup_steps=warmup_steps, num_training_steps=num_train_optimization_steps
+        )
 
         if self.fp16:
             try:
@@ -113,58 +157,36 @@ class Distiller:
             except ImportError:
                 raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
             logger.info(f"Using fp16 training: {self.params.fp16_opt_level} level")
-            self.student, self.optimizer = amp.initialize(self.student,
-                                                          self.optimizer,
-                                                          opt_level=self.params.fp16_opt_level)
+            self.student, self.optimizer = amp.initialize(
+                self.student, self.optimizer, opt_level=self.params.fp16_opt_level
+            )
             self.teacher = self.teacher.half()
 
         if self.multi_gpu:
             if self.fp16:
                 from apex.parallel import DistributedDataParallel
+
                 logger.info("Using apex.parallel.DistributedDataParallel for distributed training.")
                 self.student = DistributedDataParallel(self.student)
             else:
                 from torch.nn.parallel import DistributedDataParallel
+
                 logger.info("Using nn.parallel.DistributedDataParallel for distributed training.")
-                self.student = DistributedDataParallel(self.student,
-                                                       device_ids=[params.local_rank],
-                                                       output_device=params.local_rank)
+                self.student = DistributedDataParallel(
+                    self.student,
+                    device_ids=[params.local_rank],
+                    output_device=params.local_rank,
+                    find_unused_parameters=True,
+                )
 
         self.is_master = params.is_master
         if self.is_master:
-            logger.info('--- Initializing Tensorboard')
-            self.tensorboard = SummaryWriter(log_dir=os.path.join(self.dump_path, 'log', 'train'))
-            self.tensorboard.add_text(tag='config', text_string=str(self.params), global_step=0)
+            logger.info("--- Initializing Tensorboard")
+            self.tensorboard = SummaryWriter(log_dir=os.path.join(self.dump_path, "log", "train"))
+            self.tensorboard.add_text(tag="config/training", text_string=str(self.params), global_step=0)
+            self.tensorboard.add_text(tag="config/student", text_string=str(self.student_config), global_step=0)
 
-    def get_iterator(self,
-                     seed: int = None):
-        """
-        Initialize the data iterator.
-        Each process has its own data iterator (iterating on his own random portion of the dataset).
-
-        Input:
-        ------
-            seed: `int` - The random seed.
-        """
-        logger.info('--- Initializing Data Iterator')
-        self.data_iterator = self.dataloader.get_iterator(seed=seed)
-
-    def get_batch(self):
-        """
-        Call the data iterator to output a new batch.
-        If the data iterator went through the whole dataset, create a new iterator.
-        """
-        assert hasattr(self, 'data_iterator')
-        try:
-            x = next(self.data_iterator)
-        except StopIteration:
-            logger.warning('--- Went through the whole dataset. Creating new data iterator.')
-            self.data_iterator = self.dataloader.get_iterator()
-            x = next(self.data_iterator)
-        return x
-
-    def prepare_batch(self,
-                      batch):
+    def prepare_batch_mlm(self, batch):
         """
         Prepare the batch: from the token_ids and the lenghts, compute the attention mask and the masked label for MLM.
 
@@ -178,13 +200,13 @@ class Distiller:
         -------
             token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
             attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
-            mlm_labels: `torch.tensor(bs, seq_length)` - The masked languge modeling labels. There is a -1 where there is nothing to predict.
+            mlm_labels: `torch.tensor(bs, seq_length)` - The masked languge modeling labels. There is a -100 where there is nothing to predict.
         """
         token_ids, lengths = batch
         token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
         assert token_ids.size(0) == lengths.size(0)
 
-        attn_mask = (torch.arange(token_ids.size(1), dtype=torch.long, device=lengths.device) < lengths[:, None])
+        attn_mask = torch.arange(token_ids.size(1), dtype=torch.long, device=lengths.device) < lengths[:, None]
 
         bs, max_seq_len = token_ids.size()
         mlm_labels = token_ids.new(token_ids.size()).copy_(token_ids)
@@ -192,11 +214,13 @@ class Distiller:
         x_prob = self.token_probs[token_ids.flatten()]
         n_tgt = math.ceil(self.mlm_mask_prop * lengths.sum().item())
         tgt_ids = torch.multinomial(x_prob / x_prob.sum(), n_tgt, replacement=False)
-        pred_mask = torch.zeros(bs * max_seq_len, dtype=torch.uint8, device=token_ids.device)
+        pred_mask = torch.zeros(
+            bs * max_seq_len, dtype=torch.bool, device=token_ids.device
+        )  # previously `dtype=torch.uint8`, cf pytorch 1.2.0 compatibility
         pred_mask[tgt_ids] = 1
         pred_mask = pred_mask.view(bs, max_seq_len)
 
-        pred_mask[token_ids == self.params.special_tok_ids['pad_token']] = 0
+        pred_mask[token_ids == self.params.special_tok_ids["pad_token"]] = 0
 
         # mask a number of words == 0 [8] (faster with fp16)
         if self.fp16:
@@ -205,24 +229,58 @@ class Distiller:
                 pred_mask = pred_mask.view(-1)
                 n2 = max(n1 % 8, 8 * (n1 // 8))
                 if n2 != n1:
-                    pred_mask[torch.nonzero(pred_mask).view(-1)[:n1-n2]] = 0
+                    pred_mask[torch.nonzero(pred_mask).view(-1)[: n1 - n2]] = 0
                 pred_mask = pred_mask.view(bs, max_seq_len)
                 assert pred_mask.sum().item() % 8 == 0, pred_mask.sum().item()
 
         _token_ids_real = token_ids[pred_mask]
-        _token_ids_rand = _token_ids_real.clone().random_(self.params.vocab_size)
-        _token_ids_mask = _token_ids_real.clone().fill_(self.params.special_tok_ids['mask_token'])
+        _token_ids_rand = _token_ids_real.clone().random_(self.vocab_size)
+        _token_ids_mask = _token_ids_real.clone().fill_(self.params.special_tok_ids["mask_token"])
         probs = torch.multinomial(self.pred_probs, len(_token_ids_real), replacement=True)
-        _token_ids = _token_ids_mask * (probs == 0).long() + _token_ids_real * (probs == 1).long() + _token_ids_rand * (probs == 2).long()
+        _token_ids = (
+            _token_ids_mask * (probs == 0).long()
+            + _token_ids_real * (probs == 1).long()
+            + _token_ids_rand * (probs == 2).long()
+        )
         token_ids = token_ids.masked_scatter(pred_mask, _token_ids)
 
-        mlm_labels[1-pred_mask] = -1
+        mlm_labels[~pred_mask] = -100  # previously `mlm_labels[1-pred_mask] = -1`, cf pytorch 1.2.0 compatibility
+
+        # sanity checks
+        assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size
 
         return token_ids, attn_mask, mlm_labels
 
-    def round_batch(self,
-                    x: torch.tensor,
-                    lengths: torch.tensor):
+    def prepare_batch_clm(self, batch):
+        """
+        Prepare the batch: from the token_ids and the lenghts, compute the attention mask and the labels for CLM.
+
+        Input:
+        ------
+            batch: `Tuple`
+                token_ids: `torch.tensor(bs, seq_length)` - The token ids for each of the sequence. It is padded.
+                lengths: `torch.tensor(bs)` - The lengths of each of the sequences in the batch.
+
+        Output:
+        -------
+            token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
+            attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
+            clm_labels: `torch.tensor(bs, seq_length)` - The causal languge modeling labels. There is a -100 where there is nothing to predict.
+        """
+        token_ids, lengths = batch
+        token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
+        assert token_ids.size(0) == lengths.size(0)
+
+        attn_mask = torch.arange(token_ids.size(1), dtype=torch.long, device=lengths.device) < lengths[:, None]
+        clm_labels = token_ids.new(token_ids.size()).copy_(token_ids)
+        clm_labels[~attn_mask] = -100  # previously `clm_labels[1-attn_mask] = -1`, cf pytorch 1.2.0 compatibility
+
+        # sanity checks
+        assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size
+
+        return token_ids, attn_mask, clm_labels
+
+    def round_batch(self, x: torch.tensor, lengths: torch.tensor):
         """
         For float16 only.
         Sub-sample sentences in a batch, and add padding, so that each dimension is a multiple of 8.
@@ -257,7 +315,10 @@ class Distiller:
         if ml1 % 8 != 0:
             pad = 8 - (ml1 % 8)
             ml2 = ml1 + pad
-            pad_id = self.params.special_tok_ids['pad_token']
+            if self.mlm:
+                pad_id = self.params.special_tok_ids["pad_token"]
+            else:
+                pad_id = self.params.special_tok_ids["unk_token"]
             padding_tensor = torch.zeros(bs2, pad, dtype=torch.long, device=x.device).fill_(pad_id)
             x = torch.cat([x, padding_tensor], 1)
             assert x.size() == (bs2, ml2)
@@ -270,36 +331,45 @@ class Distiller:
         """
         The real training loop.
         """
-        if self.is_master: logger.info('Starting training')
+        if self.is_master:
+            logger.info("Starting training")
+        self.last_log = time.time()
         self.student.train()
         self.teacher.eval()
 
         for _ in range(self.params.n_epoch):
-            if self.is_master: logger.info(f'--- Starting epoch {self.epoch}/{self.params.n_epoch-1}')
+            if self.is_master:
+                logger.info(f"--- Starting epoch {self.epoch}/{self.params.n_epoch-1}")
+            if self.multi_gpu:
+                torch.distributed.barrier()
 
-            iter_bar = trange(self.num_steps_epoch, desc="-Iter", disable=self.params.local_rank not in [-1, 0])
-            for __ in range(self.num_steps_epoch):
-                batch = self.get_batch()
+            iter_bar = tqdm(self.dataloader, desc="-Iter", disable=self.params.local_rank not in [-1, 0])
+            for batch in iter_bar:
                 if self.params.n_gpu > 0:
-                    batch = tuple(t.to(f'cuda:{self.params.local_rank}') for t in batch)
-                token_ids, attn_mask, mlm_labels = self.prepare_batch(batch=batch)
+                    batch = tuple(t.to(f"cuda:{self.params.local_rank}") for t in batch)
 
-                self.step(input_ids=token_ids, attention_mask=attn_mask, mlm_labels=mlm_labels)
+                if self.mlm:
+                    token_ids, attn_mask, lm_labels = self.prepare_batch_mlm(batch=batch)
+                else:
+                    token_ids, attn_mask, lm_labels = self.prepare_batch_clm(batch=batch)
+                self.step(input_ids=token_ids, attention_mask=attn_mask, lm_labels=lm_labels)
 
                 iter_bar.update()
-                iter_bar.set_postfix({'Last_loss': f'{self.last_loss:.2f}',
-                                      'Avg_cum_loss': f'{self.total_loss_epoch/self.n_iter:.2f}'})
+                iter_bar.set_postfix(
+                    {"Last_loss": f"{self.last_loss:.2f}", "Avg_cum_loss": f"{self.total_loss_epoch/self.n_iter:.2f}"}
+                )
             iter_bar.close()
 
-            if self.is_master: logger.info(f'--- Ending epoch {self.epoch}/{self.params.n_epoch-1}')
+            if self.is_master:
+                logger.info(f"--- Ending epoch {self.epoch}/{self.params.n_epoch-1}")
             self.end_epoch()
 
-        if self.is_master: logger.info('Training is finished')
+        if self.is_master:
+            logger.info(f"Save very last checkpoint as `pytorch_model.bin`.")
+            self.save_checkpoint(checkpoint_name=f"pytorch_model.bin")
+            logger.info("Training is finished")
 
-    def step(self,
-             input_ids: torch.tensor,
-             attention_mask: torch.tensor,
-             mlm_labels: torch.tensor):
+    def step(self, input_ids: torch.tensor, attention_mask: torch.tensor, lm_labels: torch.tensor):
         """
         One optimization step: forward of student AND teacher, backward on the loss (for gradient accumulation),
         and possibly a parameter update (depending on the gradient accumulation).
@@ -308,49 +378,94 @@ class Distiller:
         ------
         input_ids: `torch.tensor(bs, seq_length)` - The token ids.
         attention_mask: `torch.tensor(bs, seq_length)` - The attention mask for self attention.
-        mlm_labels: `torch.tensor(bs, seq_length)` - The masked language modeling labels.
+        lm_labels: `torch.tensor(bs, seq_length)` - The language modeling labels (mlm labels for MLM and clm labels for CLM).
         """
-        s_logits = self.student(input_ids=input_ids, attention_mask=attention_mask)[0]     # (bs, seq_length, voc_size)
-        with torch.no_grad():
-            t_logits = self.teacher(input_ids=input_ids, attention_mask=attention_mask)[0] # (bs, seq_length, voc_size)
+        if self.mlm:
+            s_logits, s_hidden_states = self.student(
+                input_ids=input_ids, attention_mask=attention_mask
+            )  # (bs, seq_length, voc_size)
+            with torch.no_grad():
+                t_logits, t_hidden_states = self.teacher(
+                    input_ids=input_ids, attention_mask=attention_mask
+                )  # (bs, seq_length, voc_size)
+        else:
+            s_logits, _, s_hidden_states = self.student(
+                input_ids=input_ids, attention_mask=None
+            )  # (bs, seq_length, voc_size)
+            with torch.no_grad():
+                t_logits, _, t_hidden_states = self.teacher(
+                    input_ids=input_ids, attention_mask=None
+                )  # (bs, seq_length, voc_size)
         assert s_logits.size() == t_logits.size()
 
-        #https://github.com/peterliht/knowledge-distillation-pytorch/blob/master/model/net.py#L100
-        #https://github.com/peterliht/knowledge-distillation-pytorch/issues/2
+        # https://github.com/peterliht/knowledge-distillation-pytorch/blob/master/model/net.py#L100
+        # https://github.com/peterliht/knowledge-distillation-pytorch/issues/2
         if self.params.restrict_ce_to_mask:
-            mask = (mlm_labels>-1).unsqueeze(-1).expand_as(s_logits)   # (bs, seq_lenth, voc_size)
+            mask = (lm_labels > -1).unsqueeze(-1).expand_as(s_logits)  # (bs, seq_lenth, voc_size)
         else:
-            mask = attention_mask.unsqueeze(-1).expand_as(s_logits)    # (bs, seq_lenth, voc_size)
-        s_logits_slct = torch.masked_select(s_logits, mask)            # (bs * seq_length * voc_size) modulo the 1s in mask
-        s_logits_slct = s_logits_slct.view(-1, s_logits.size(-1))      # (bs * seq_length, voc_size) modulo the 1s in mask
-        t_logits_slct = torch.masked_select(t_logits, mask)            # (bs * seq_length * voc_size) modulo the 1s in mask
-        t_logits_slct = t_logits_slct.view(-1, s_logits.size(-1))      # (bs * seq_length, voc_size) modulo the 1s in mask
+            mask = attention_mask.unsqueeze(-1).expand_as(s_logits)  # (bs, seq_lenth, voc_size)
+        s_logits_slct = torch.masked_select(s_logits, mask)  # (bs * seq_length * voc_size) modulo the 1s in mask
+        s_logits_slct = s_logits_slct.view(-1, s_logits.size(-1))  # (bs * seq_length, voc_size) modulo the 1s in mask
+        t_logits_slct = torch.masked_select(t_logits, mask)  # (bs * seq_length * voc_size) modulo the 1s in mask
+        t_logits_slct = t_logits_slct.view(-1, s_logits.size(-1))  # (bs * seq_length, voc_size) modulo the 1s in mask
         assert t_logits_slct.size() == s_logits_slct.size()
 
-        loss_ce = self.ce_loss_fct(F.log_softmax(s_logits_slct/self.temperature, dim=-1),
-                                   F.softmax(t_logits_slct/self.temperature, dim=-1)) * (self.temperature)**2
-        loss = self.alpha_ce*loss_ce
-        if self.alpha_mlm > 0.:
-            loss_mlm = self.mlm_loss_fct(s_logits.view(-1, s_logits.size(-1)), mlm_labels.view(-1))
+        loss_ce = (
+            self.ce_loss_fct(
+                F.log_softmax(s_logits_slct / self.temperature, dim=-1),
+                F.softmax(t_logits_slct / self.temperature, dim=-1),
+            )
+            * (self.temperature) ** 2
+        )
+        loss = self.alpha_ce * loss_ce
+
+        if self.alpha_mlm > 0.0:
+            loss_mlm = self.lm_loss_fct(s_logits.view(-1, s_logits.size(-1)), lm_labels.view(-1))
             loss += self.alpha_mlm * loss_mlm
-        if self.alpha_mse > 0.:
-            loss_mse = self.mse_loss_fct(s_logits_slct, t_logits_slct)/s_logits_slct.size(0) # Reproducing batchmean reduction
+        if self.alpha_clm > 0.0:
+            shift_logits = s_logits[..., :-1, :].contiguous()
+            shift_labels = lm_labels[..., 1:].contiguous()
+            loss_clm = self.lm_loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+            loss += self.alpha_clm * loss_clm
+
+        if self.alpha_mse > 0.0:
+            loss_mse = self.mse_loss_fct(s_logits_slct, t_logits_slct) / s_logits_slct.size(
+                0
+            )  # Reproducing batchmean reduction
             loss += self.alpha_mse * loss_mse
+        if self.alpha_cos > 0.0:
+            s_hidden_states = s_hidden_states[-1]  # (bs, seq_length, dim)
+            t_hidden_states = t_hidden_states[-1]  # (bs, seq_length, dim)
+            mask = attention_mask.unsqueeze(-1).expand_as(s_hidden_states)  # (bs, seq_length, dim)
+            assert s_hidden_states.size() == t_hidden_states.size()
+            dim = s_hidden_states.size(-1)
+
+            s_hidden_states_slct = torch.masked_select(s_hidden_states, mask)  # (bs * seq_length * dim)
+            s_hidden_states_slct = s_hidden_states_slct.view(-1, dim)  # (bs * seq_length, dim)
+            t_hidden_states_slct = torch.masked_select(t_hidden_states, mask)  # (bs * seq_length * dim)
+            t_hidden_states_slct = t_hidden_states_slct.view(-1, dim)  # (bs * seq_length, dim)
+
+            target = s_hidden_states_slct.new(s_hidden_states_slct.size(0)).fill_(1)  # (bs * seq_length,)
+            loss_cos = self.cosine_loss_fct(s_hidden_states_slct, t_hidden_states_slct, target)
+            loss += self.alpha_cos * loss_cos
 
         self.total_loss_epoch += loss.item()
         self.last_loss = loss.item()
         self.last_loss_ce = loss_ce.item()
-        if self.alpha_mlm > 0.:
+        if self.alpha_mlm > 0.0:
             self.last_loss_mlm = loss_mlm.item()
-        if self.alpha_mse > 0.:
+        if self.alpha_clm > 0.0:
+            self.last_loss_clm = loss_clm.item()
+        if self.alpha_mse > 0.0:
             self.last_loss_mse = loss_mse.item()
+        if self.alpha_cos > 0.0:
+            self.last_loss_cos = loss_cos.item()
 
         self.optimize(loss)
 
         self.n_sequences_epoch += input_ids.size(0)
 
-    def optimize(self,
-                 loss):
+    def optimize(self, loss):
         """
         Normalization on the loss (gradient accumulation or distributed training), followed by
         backward pass on the loss, possibly followed by a parameter update (depending on the gradient accumulation).
@@ -358,7 +473,7 @@ class Distiller:
         """
         # Check for NaN
         if (loss != loss).data.any():
-            logger.error('NaN detected')
+            logger.error("NaN detected")
             exit()
 
         if self.multi_gpu:
@@ -368,6 +483,7 @@ class Distiller:
 
         if self.fp16:
             from apex import amp
+
             with amp.scale_loss(loss, self.optimizer) as scaled_loss:
                 scaled_loss.backward()
         else:
@@ -379,9 +495,9 @@ class Distiller:
                 torch.nn.utils.clip_grad_norm_(amp.master_params(self.optimizer), self.params.max_grad_norm)
             else:
                 torch.nn.utils.clip_grad_norm_(self.student.parameters(), self.params.max_grad_norm)
-            self.scheduler.step()
             self.optimizer.step()
             self.optimizer.zero_grad()
+            self.scheduler.step()
 
     def iter(self):
         """
@@ -392,6 +508,7 @@ class Distiller:
 
         if self.n_total_iter % self.params.log_interval == 0:
             self.log_tensorboard()
+            self.last_log = time.time()
         if self.n_total_iter % self.params.checkpoint_interval == 0:
             self.save_checkpoint()
 
@@ -403,46 +520,84 @@ class Distiller:
             return
 
         for param_name, param in self.student.named_parameters():
-            self.tensorboard.add_scalar(tag='parameter_mean/' + param_name, scalar_value=param.data.mean(), global_step=self.n_total_iter)
-            self.tensorboard.add_scalar(tag='parameter_std/' + param_name, scalar_value=param.data.std(), global_step=self.n_total_iter)
+            self.tensorboard.add_scalar(
+                tag="parameter_mean/" + param_name, scalar_value=param.data.mean(), global_step=self.n_total_iter
+            )
+            self.tensorboard.add_scalar(
+                tag="parameter_std/" + param_name, scalar_value=param.data.std(), global_step=self.n_total_iter
+            )
             if param.grad is None:
                 continue
-            self.tensorboard.add_scalar(tag="grad_mean/" + param_name, scalar_value=param.grad.data.mean(),global_step=self.n_total_iter)
-            self.tensorboard.add_scalar(tag="grad_std/" + param_name, scalar_value=param.grad.data.std(), global_step=self.n_total_iter)
+            self.tensorboard.add_scalar(
+                tag="grad_mean/" + param_name, scalar_value=param.grad.data.mean(), global_step=self.n_total_iter
+            )
+            self.tensorboard.add_scalar(
+                tag="grad_std/" + param_name, scalar_value=param.grad.data.std(), global_step=self.n_total_iter
+            )
 
-        self.tensorboard.add_scalar(tag="losses/cum_avg_loss_epoch", scalar_value=self.total_loss_epoch/self.n_iter, global_step=self.n_total_iter)
+        self.tensorboard.add_scalar(
+            tag="losses/cum_avg_loss_epoch",
+            scalar_value=self.total_loss_epoch / self.n_iter,
+            global_step=self.n_total_iter,
+        )
         self.tensorboard.add_scalar(tag="losses/loss", scalar_value=self.last_loss, global_step=self.n_total_iter)
-        self.tensorboard.add_scalar(tag="losses/loss_ce", scalar_value=self.last_loss_ce, global_step=self.n_total_iter)
-        if self.alpha_mlm > 0.:
-            self.tensorboard.add_scalar(tag="losses/loss_mlm", scalar_value=self.last_loss_mlm, global_step=self.n_total_iter)
-        if self.alpha_mse > 0.:
-            self.tensorboard.add_scalar(tag="losses/loss_mse", scalar_value=self.last_loss_mse, global_step=self.n_total_iter)
-        self.tensorboard.add_scalar(tag="learning_rate/lr", scalar_value=self.scheduler.get_lr()[0], global_step=self.n_total_iter)
+        self.tensorboard.add_scalar(
+            tag="losses/loss_ce", scalar_value=self.last_loss_ce, global_step=self.n_total_iter
+        )
+        if self.alpha_mlm > 0.0:
+            self.tensorboard.add_scalar(
+                tag="losses/loss_mlm", scalar_value=self.last_loss_mlm, global_step=self.n_total_iter
+            )
+        if self.alpha_clm > 0.0:
+            self.tensorboard.add_scalar(
+                tag="losses/loss_clm", scalar_value=self.last_loss_clm, global_step=self.n_total_iter
+            )
+        if self.alpha_mse > 0.0:
+            self.tensorboard.add_scalar(
+                tag="losses/loss_mse", scalar_value=self.last_loss_mse, global_step=self.n_total_iter
+            )
+        if self.alpha_cos > 0.0:
+            self.tensorboard.add_scalar(
+                tag="losses/loss_cos", scalar_value=self.last_loss_cos, global_step=self.n_total_iter
+            )
+        self.tensorboard.add_scalar(
+            tag="learning_rate/lr", scalar_value=self.scheduler.get_lr()[0], global_step=self.n_total_iter
+        )
+
+        self.tensorboard.add_scalar(
+            tag="global/memory_usage",
+            scalar_value=psutil.virtual_memory()._asdict()["used"] / 1_000_000,
+            global_step=self.n_total_iter,
+        )
+        self.tensorboard.add_scalar(
+            tag="global/speed", scalar_value=time.time() - self.last_log, global_step=self.n_total_iter
+        )
 
     def end_epoch(self):
         """
         Finally arrived at the end of epoch (full pass on dataset).
         Do some tensorboard logging and checkpoint saving.
         """
-        logger.info(f'{self.n_sequences_epoch} sequences have been trained during this epoch.')
+        logger.info(f"{self.n_sequences_epoch} sequences have been trained during this epoch.")
 
         if self.is_master:
-            self.save_checkpoint(checkpoint_name=f'model_epoch_{self.epoch}.pth')
-            self.tensorboard.add_scalar(tag='epoch/loss', scalar_value=self.total_loss_epoch/self.n_iter, global_step=self.epoch)
+            self.save_checkpoint(checkpoint_name=f"model_epoch_{self.epoch}.pth")
+            self.tensorboard.add_scalar(
+                tag="epoch/loss", scalar_value=self.total_loss_epoch / self.n_iter, global_step=self.epoch
+            )
 
         self.epoch += 1
         self.n_sequences_epoch = 0
         self.n_iter = 0
         self.total_loss_epoch = 0
 
-    def save_checkpoint(self,
-                        checkpoint_name: str = 'checkpoint.pth'):
+    def save_checkpoint(self, checkpoint_name: str = "checkpoint.pth"):
         """
         Save the current state. Only by the master process.
         """
         if not self.is_master:
             return
-        mdl_to_save = self.student.module if hasattr(self.student, 'module') else self.student
+        mdl_to_save = self.student.module if hasattr(self.student, "module") else self.student
         mdl_to_save.config.save_pretrained(self.dump_path)
         state_dict = mdl_to_save.state_dict()
         torch.save(state_dict, os.path.join(self.dump_path, checkpoint_name))

examples/distillation/grouped_batch_sampler.py

@@ -0,0 +1,108 @@
+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team and Facebook, 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.
+""" Adapted from PyTorch Vision (https://github.com/pytorch/vision/blob/master/references/detection/group_by_aspect_ratio.py)
+"""
+import bisect
+import copy
+from collections import defaultdict
+
+import numpy as np
+from torch.utils.data.sampler import BatchSampler, Sampler
+
+from utils import logger
+
+
+def _quantize(x, bins):
+    bins = copy.deepcopy(bins)
+    bins = sorted(bins)
+    quantized = list(map(lambda y: bisect.bisect_right(bins, y), x))
+    return quantized
+
+
+def create_lengths_groups(lengths, k=0):
+    bins = np.arange(start=3, stop=k, step=4).tolist() if k > 0 else [10]
+    groups = _quantize(lengths, bins)
+    # count number of elements per group
+    counts = np.unique(groups, return_counts=True)[1]
+    fbins = [0] + bins + [np.inf]
+    logger.info("Using {} as bins for aspect lengths quantization".format(fbins))
+    logger.info("Count of instances per bin: {}".format(counts))
+    return groups
+
+
+class GroupedBatchSampler(BatchSampler):
+    """
+    Wraps another sampler to yield a mini-batch of indices.
+    It enforces that the batch only contain elements from the same group.
+    It also tries to provide mini-batches which follows an ordering which is
+    as close as possible to the ordering from the original sampler.
+    Arguments:
+        sampler (Sampler): Base sampler.
+        group_ids (list[int]): If the sampler produces indices in range [0, N),
+            `group_ids` must be a list of `N` ints which contains the group id of each sample.
+            The group ids must be a continuous set of integers starting from
+            0, i.e. they must be in the range [0, num_groups).
+        batch_size (int): Size of mini-batch.
+    """
+
+    def __init__(self, sampler, group_ids, batch_size):
+        if not isinstance(sampler, Sampler):
+            raise ValueError(
+                "sampler should be an instance of " "torch.utils.data.Sampler, but got sampler={}".format(sampler)
+            )
+        self.sampler = sampler
+        self.group_ids = group_ids
+        self.batch_size = batch_size
+
+    def __iter__(self):
+        buffer_per_group = defaultdict(list)
+        samples_per_group = defaultdict(list)
+
+        num_batches = 0
+        for idx in self.sampler:
+            group_id = self.group_ids[idx]
+            buffer_per_group[group_id].append(idx)
+            samples_per_group[group_id].append(idx)
+            if len(buffer_per_group[group_id]) == self.batch_size:
+                yield buffer_per_group[group_id]  # TODO
+                num_batches += 1
+                del buffer_per_group[group_id]
+            assert len(buffer_per_group[group_id]) < self.batch_size
+
+        # now we have run out of elements that satisfy
+        # the group criteria, let's return the remaining
+        # elements so that the size of the sampler is
+        # deterministic
+        expected_num_batches = len(self)
+        num_remaining = expected_num_batches - num_batches
+        if num_remaining > 0:
+            # for the remaining batches, group the batches by similar lengths
+            batch_idx = []
+            for group_id, idxs in sorted(buffer_per_group.items(), key=lambda x: x[0]):
+                batch_idx.extend(idxs)
+                if len(batch_idx) >= self.batch_size:
+                    yield batch_idx[: self.batch_size]
+                    batch_idx = batch_idx[self.batch_size :]
+                    num_remaining -= 1
+            if len(batch_idx) > 0:
+                yield batch_idx
+                num_remaining -= 1
+        assert num_remaining == 0
+
+    def __len__(self):
+        """
+        Return the number of mini-batches rather than the number of samples.
+        """
+        return (len(self.sampler) + self.batch_size - 1) // self.batch_size

examples/distillation/lm_seqs_dataset.py

@@ -0,0 +1,166 @@
+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team and Facebook, 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.
+""" Dataset to distilled models
+    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
+"""
+import numpy as np
+import torch
+from torch.utils.data import Dataset
+
+from utils import logger
+
+
+class LmSeqsDataset(Dataset):
+    """Custom Dataset wrapping language modeling sequences.
+
+    Each sample will be retrieved by indexing the list of token_ids and their corresponding lengths.
+
+    Input:
+    ------
+        params: `NameSpace` parameters
+        data: `List[np.array[int]]
+    """
+
+    def __init__(self, params, data):
+        self.params = params
+
+        self.token_ids = np.array(data)
+        self.lengths = np.array([len(t) for t in data])
+
+        self.check()
+        self.remove_long_sequences()
+        self.remove_empty_sequences()
+        self.remove_unknown_sequences()
+        self.check()
+        self.print_statistics()
+
+    def __getitem__(self, index):
+        return (self.token_ids[index], self.lengths[index])
+
+    def __len__(self):
+        return len(self.lengths)
+
+    def check(self):
+        """
+        Some sanity checks
+        """
+        assert len(self.token_ids) == len(self.lengths)
+        assert all(self.lengths[i] == len(self.token_ids[i]) for i in range(len(self.lengths)))
+
+    def remove_long_sequences(self):
+        """
+        Sequences that are too long are splitted by chunk of max_model_input_size.
+        """
+        max_len = self.params.max_model_input_size
+        indices = self.lengths > max_len
+        logger.info(f"Splitting {sum(indices)} too long sequences.")
+
+        def divide_chunks(l, n):
+            return [l[i : i + n] for i in range(0, len(l), n)]
+
+        new_tok_ids = []
+        new_lengths = []
+        if self.params.mlm:
+            cls_id, sep_id = self.params.special_tok_ids["cls_token"], self.params.special_tok_ids["sep_token"]
+        else:
+            cls_id, sep_id = self.params.special_tok_ids["bos_token"], self.params.special_tok_ids["eos_token"]
+
+        for seq_, len_ in zip(self.token_ids, self.lengths):
+            assert (seq_[0] == cls_id) and (seq_[-1] == sep_id), seq_
+            if len_ <= max_len:
+                new_tok_ids.append(seq_)
+                new_lengths.append(len_)
+            else:
+                sub_seqs = []
+                for sub_s in divide_chunks(seq_, max_len - 2):
+                    if sub_s[0] != cls_id:
+                        sub_s = np.insert(sub_s, 0, cls_id)
+                    if sub_s[-1] != sep_id:
+                        sub_s = np.insert(sub_s, len(sub_s), sep_id)
+                    assert len(sub_s) <= max_len
+                    assert (sub_s[0] == cls_id) and (sub_s[-1] == sep_id), sub_s
+                    sub_seqs.append(sub_s)
+
+                new_tok_ids.extend(sub_seqs)
+                new_lengths.extend([len(l) for l in sub_seqs])
+
+        self.token_ids = np.array(new_tok_ids)
+        self.lengths = np.array(new_lengths)
+
+    def remove_empty_sequences(self):
+        """
+        Too short sequences are simply removed. This could be tunedd.
+        """
+        init_size = len(self)
+        indices = self.lengths > 11
+        self.token_ids = self.token_ids[indices]
+        self.lengths = self.lengths[indices]
+        new_size = len(self)
+        logger.info(f"Remove {init_size - new_size} too short (<=11 tokens) sequences.")
+
+    def remove_unknown_sequences(self):
+        """
+        Remove sequences with a (too) high level of unknown tokens.
+        """
+        if "unk_token" not in self.params.special_tok_ids:
+            return
+        else:
+            unk_token_id = self.params.special_tok_ids["unk_token"]
+        init_size = len(self)
+        unk_occs = np.array([np.count_nonzero(a == unk_token_id) for a in self.token_ids])
+        indices = (unk_occs / self.lengths) < 0.5
+        self.token_ids = self.token_ids[indices]
+        self.lengths = self.lengths[indices]
+        new_size = len(self)
+        logger.info(f"Remove {init_size - new_size} sequences with a high level of unknown tokens (50%).")
+
+    def print_statistics(self):
+        """
+        Print some statistics on the corpus. Only the master process.
+        """
+        if not self.params.is_master:
+            return
+        logger.info(f"{len(self)} sequences")
+        # data_len = sum(self.lengths)
+        # nb_unique_tokens = len(Counter(list(chain(*self.token_ids))))
+        # logger.info(f'{data_len} tokens ({nb_unique_tokens} unique)')
+
+        # unk_idx = self.params.special_tok_ids['unk_token']
+        # nb_unkown = sum([(t==unk_idx).sum() for t in self.token_ids])
+        # logger.info(f'{nb_unkown} unknown tokens (covering {100*nb_unkown/data_len:.2f}% of the data)')
+
+    def batch_sequences(self, batch):
+        """
+        Do the padding and transform into torch.tensor.
+        """
+        token_ids = [t[0] for t in batch]
+        lengths = [t[1] for t in batch]
+        assert len(token_ids) == len(lengths)
+
+        # Max for paddings
+        max_seq_len_ = max(lengths)
+
+        # Pad token ids
+        if self.params.mlm:
+            pad_idx = self.params.special_tok_ids["pad_token"]
+        else:
+            pad_idx = self.params.special_tok_ids["unk_token"]
+        tk_ = [list(t.astype(int)) + [pad_idx] * (max_seq_len_ - len(t)) for t in token_ids]
+        assert len(tk_) == len(token_ids)
+        assert all(len(t) == max_seq_len_ for t in tk_)
+
+        tk_t = torch.tensor(tk_)  # (bs, max_seq_len_)
+        lg_t = torch.tensor(lengths)  # (bs)
+        return tk_t, lg_t

examples/distillation/requirements.txt

@@ -1 +1,7 @@
+transformers
+
 gitpython==3.0.2
+tensorboard>=1.14.0
+tensorboardX==1.8
+psutil==5.6.6
+scipy==1.3.1

examples/distillation/run_squad_w_distillation.py

@@ -0,0 +1,868 @@
+# 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.
+""" This is the exact same script as `examples/run_squad.py` (as of 2020, January 8th) with an additional and optional step of distillation."""
+
+import argparse
+import glob
+import logging
+import os
+import random
+import timeit
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm, trange
+
+from transformers import (
+    WEIGHTS_NAME,
+    AdamW,
+    BertConfig,
+    BertForQuestionAnswering,
+    BertTokenizer,
+    DistilBertConfig,
+    DistilBertForQuestionAnswering,
+    DistilBertTokenizer,
+    RobertaConfig,
+    RobertaForQuestionAnswering,
+    RobertaTokenizer,
+    XLMConfig,
+    XLMForQuestionAnswering,
+    XLMTokenizer,
+    XLNetConfig,
+    XLNetForQuestionAnswering,
+    XLNetTokenizer,
+    get_linear_schedule_with_warmup,
+    squad_convert_examples_to_features,
+)
+from transformers.data.metrics.squad_metrics import (
+    compute_predictions_log_probs,
+    compute_predictions_logits,
+    squad_evaluate,
+)
+from transformers.data.processors.squad import SquadResult, SquadV1Processor, SquadV2Processor
+
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except ImportError:
+    from tensorboardX import SummaryWriter
+
+
+logger = logging.getLogger(__name__)
+
+ALL_MODELS = sum(
+    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig)), ()
+)
+
+MODEL_CLASSES = {
+    "bert": (BertConfig, BertForQuestionAnswering, BertTokenizer),
+    "xlnet": (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
+    "xlm": (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
+    "distilbert": (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
+    "roberta": (RobertaConfig, RobertaForQuestionAnswering, RobertaTokenizer),
+}
+
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+
+def to_list(tensor):
+    return tensor.detach().cpu().tolist()
+
+
+def train(args, train_dataset, model, tokenizer, teacher=None):
+    """ Train the model """
+    if args.local_rank in [-1, 0]:
+        tb_writer = SummaryWriter()
+
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+    else:
+        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+
+    # Prepare optimizer and schedule (linear warmup and decay)
+    no_decay = ["bias", "LayerNorm.weight"]
+    optimizer_grouped_parameters = [
+        {
+            "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
+            "weight_decay": args.weight_decay,
+        },
+        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
+    ]
+    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = get_linear_schedule_with_warmup(
+        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
+    )
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
+        os.path.join(args.model_name_or_path, "scheduler.pt")
+    ):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
+
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(
+            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
+        )
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info(
+        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
+        args.train_batch_size
+        * args.gradient_accumulation_steps
+        * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
+    )
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+
+    global_step = 1
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        try:
+            # set global_step to gobal_step of last saved checkpoint from model path
+            checkpoint_suffix = args.model_name_or_path.split("-")[-1].split("/")[0]
+            global_step = int(checkpoint_suffix)
+            epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+            steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
+
+            logger.info("  Continuing training from checkpoint, will skip to saved global_step")
+            logger.info("  Continuing training from epoch %d", epochs_trained)
+            logger.info("  Continuing training from global step %d", global_step)
+            logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
+        except ValueError:
+            logger.info("  Starting fine-tuning.")
+
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(
+        epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]
+    )
+    # Added here for reproductibility
+    set_seed(args)
+
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
+            model.train()
+            if teacher is not None:
+                teacher.eval()
+            batch = tuple(t.to(args.device) for t in batch)
+
+            inputs = {
+                "input_ids": batch[0],
+                "attention_mask": batch[1],
+                "start_positions": batch[3],
+                "end_positions": batch[4],
+            }
+            if args.model_type != "distilbert":
+                inputs["token_type_ids"] = None if args.model_type == "xlm" else batch[2]
+            if args.model_type in ["xlnet", "xlm"]:
+                inputs.update({"cls_index": batch[5], "p_mask": batch[6]})
+                if args.version_2_with_negative:
+                    inputs.update({"is_impossible": batch[7]})
+            outputs = model(**inputs)
+            loss, start_logits_stu, end_logits_stu = outputs
+
+            # Distillation loss
+            if teacher is not None:
+                if "token_type_ids" not in inputs:
+                    inputs["token_type_ids"] = None if args.teacher_type == "xlm" else batch[2]
+                with torch.no_grad():
+                    start_logits_tea, end_logits_tea = teacher(
+                        input_ids=inputs["input_ids"],
+                        token_type_ids=inputs["token_type_ids"],
+                        attention_mask=inputs["attention_mask"],
+                    )
+                assert start_logits_tea.size() == start_logits_stu.size()
+                assert end_logits_tea.size() == end_logits_stu.size()
+
+                loss_fct = nn.KLDivLoss(reduction="batchmean")
+                loss_start = loss_fct(
+                    F.log_softmax(start_logits_stu / args.temperature, dim=-1),
+                    F.softmax(start_logits_tea / args.temperature, dim=-1),
+                ) * (args.temperature ** 2)
+                loss_end = loss_fct(
+                    F.log_softmax(end_logits_stu / args.temperature, dim=-1),
+                    F.softmax(end_logits_tea / args.temperature, dim=-1),
+                ) * (args.temperature ** 2)
+                loss_ce = (loss_start + loss_end) / 2.0
+
+                loss = args.alpha_ce * loss_ce + args.alpha_squad * loss
+
+            if args.n_gpu > 1:
+                loss = loss.mean()  # mean() to average on multi-gpu parallel (not distributed) training
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+
+            if args.fp16:
+                with amp.scale_loss(loss, optimizer) as scaled_loss:
+                    scaled_loss.backward()
+            else:
+                loss.backward()
+
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+
+                optimizer.step()
+                scheduler.step()  # Update learning rate schedule
+                model.zero_grad()
+                global_step += 1
+
+                # Log metrics
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    # Only evaluate when single GPU otherwise metrics may not average well
+                    if args.local_rank == -1 and args.evaluate_during_training:
+                        results = evaluate(args, model, tokenizer)
+                        for key, value in results.items():
+                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
+                    tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
+                    logging_loss = tr_loss
+
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    # Save model checkpoint
+                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
+                    model_to_save = (
+                        model.module if hasattr(model, "module") else model
+                    )  # Take care of distributed/parallel training
+                    model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+
+                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
+                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+
+    return global_step, tr_loss / global_step
+
+
+def evaluate(args, model, tokenizer, prefix=""):
+    dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
+
+    if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+        os.makedirs(args.output_dir)
+
+    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+
+    # Note that DistributedSampler samples randomly
+    eval_sampler = SequentialSampler(dataset)
+    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+    # multi-gpu evaluate
+    if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
+        model = torch.nn.DataParallel(model)
+
+    # Eval!
+    logger.info("***** Running evaluation {} *****".format(prefix))
+    logger.info("  Num examples = %d", len(dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+
+    all_results = []
+    start_time = timeit.default_timer()
+
+    for batch in tqdm(eval_dataloader, desc="Evaluating"):
+        model.eval()
+        batch = tuple(t.to(args.device) for t in batch)
+
+        with torch.no_grad():
+            inputs = {"input_ids": batch[0], "attention_mask": batch[1]}
+            if args.model_type != "distilbert":
+                inputs["token_type_ids"] = None if args.model_type == "xlm" else batch[2]  # XLM don't use segment_ids
+            example_indices = batch[3]
+            if args.model_type in ["xlnet", "xlm"]:
+                inputs.update({"cls_index": batch[4], "p_mask": batch[5]})
+
+            outputs = model(**inputs)
+
+        for i, example_index in enumerate(example_indices):
+            eval_feature = features[example_index.item()]
+            unique_id = int(eval_feature.unique_id)
+
+            output = [to_list(output[i]) for output in outputs]
+
+            # Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
+            # models only use two.
+            if len(output) >= 5:
+                start_logits = output[0]
+                start_top_index = output[1]
+                end_logits = output[2]
+                end_top_index = output[3]
+                cls_logits = output[4]
+
+                result = SquadResult(
+                    unique_id,
+                    start_logits,
+                    end_logits,
+                    start_top_index=start_top_index,
+                    end_top_index=end_top_index,
+                    cls_logits=cls_logits,
+                )
+
+            else:
+                start_logits, end_logits = output
+                result = SquadResult(unique_id, start_logits, end_logits)
+
+            all_results.append(result)
+
+    evalTime = timeit.default_timer() - start_time
+    logger.info("  Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))
+
+    # Compute predictions
+    output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
+    output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
+
+    if args.version_2_with_negative:
+        output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
+    else:
+        output_null_log_odds_file = None
+
+    if args.model_type in ["xlnet", "xlm"]:
+        # XLNet uses a more complex post-processing procedure
+        predictions = compute_predictions_log_probs(
+            examples,
+            features,
+            all_results,
+            args.n_best_size,
+            args.max_answer_length,
+            output_prediction_file,
+            output_nbest_file,
+            output_null_log_odds_file,
+            model.config.start_n_top,
+            model.config.end_n_top,
+            args.version_2_with_negative,
+            tokenizer,
+            args.verbose_logging,
+        )
+    else:
+        predictions = compute_predictions_logits(
+            examples,
+            features,
+            all_results,
+            args.n_best_size,
+            args.max_answer_length,
+            args.do_lower_case,
+            output_prediction_file,
+            output_nbest_file,
+            output_null_log_odds_file,
+            args.verbose_logging,
+            args.version_2_with_negative,
+            args.null_score_diff_threshold,
+            tokenizer,
+        )
+
+    # Compute the F1 and exact scores.
+    results = squad_evaluate(examples, predictions)
+    return results
+
+
+def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
+    if args.local_rank not in [-1, 0] and not evaluate:
+        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        torch.distributed.barrier()
+
+    # Load data features from cache or dataset file
+    input_file = args.predict_file if evaluate else args.train_file
+    cached_features_file = os.path.join(
+        os.path.dirname(input_file),
+        "cached_distillation_{}_{}_{}".format(
+            "dev" if evaluate else "train",
+            list(filter(None, args.model_name_or_path.split("/"))).pop(),
+            str(args.max_seq_length),
+        ),
+    )
+    if os.path.exists(cached_features_file) and not args.overwrite_cache:
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features_and_dataset = torch.load(cached_features_file)
+
+        try:
+            features, dataset, examples = (
+                features_and_dataset["features"],
+                features_and_dataset["dataset"],
+                features_and_dataset["examples"],
+            )
+        except KeyError:
+            raise DeprecationWarning(
+                "You seem to be loading features from an older version of this script please delete the "
+                "file %s in order for it to be created again" % cached_features_file
+            )
+    else:
+        logger.info("Creating features from dataset file at %s", input_file)
+        processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
+        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,
+            tokenizer=tokenizer,
+            max_seq_length=args.max_seq_length,
+            doc_stride=args.doc_stride,
+            max_query_length=args.max_query_length,
+            is_training=not evaluate,
+            return_dataset="pt",
+            threads=args.threads,
+        )
+
+        if args.local_rank in [-1, 0]:
+            logger.info("Saving features into cached file %s", cached_features_file)
+            torch.save({"features": features, "dataset": dataset, "examples": examples}, cached_features_file)
+
+    if args.local_rank == 0 and not evaluate:
+        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        torch.distributed.barrier()
+
+    if output_examples:
+        return dataset, examples, features
+    return dataset
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    # Required parameters
+    parser.add_argument(
+        "--model_type",
+        default=None,
+        type=str,
+        required=True,
+        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
+    )
+    parser.add_argument(
+        "--model_name_or_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
+    )
+    parser.add_argument(
+        "--output_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The output directory where the model checkpoints and predictions will be written.",
+    )
+
+    # Distillation parameters (optional)
+    parser.add_argument(
+        "--teacher_type",
+        default=None,
+        type=str,
+        help="Teacher type. Teacher tokenizer and student (model) tokenizer must output the same tokenization. Only for distillation.",
+    )
+    parser.add_argument(
+        "--teacher_name_or_path",
+        default=None,
+        type=str,
+        help="Path to the already SQuAD fine-tuned teacher model. Only for distillation.",
+    )
+    parser.add_argument(
+        "--alpha_ce", default=0.5, type=float, help="Distillation loss linear weight. Only for distillation."
+    )
+    parser.add_argument(
+        "--alpha_squad", default=0.5, type=float, help="True SQuAD loss linear weight. Only for distillation."
+    )
+    parser.add_argument(
+        "--temperature", default=2.0, type=float, help="Distillation temperature. Only for distillation."
+    )
+
+    # 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 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,
+        help="Pretrained tokenizer name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--cache_dir",
+        default="",
+        type=str,
+        help="Where do you want to store the pre-trained models downloaded from s3",
+    )
+
+    parser.add_argument(
+        "--version_2_with_negative",
+        action="store_true",
+        help="If true, the SQuAD examples contain some that do not have an answer.",
+    )
+    parser.add_argument(
+        "--null_score_diff_threshold",
+        type=float,
+        default=0.0,
+        help="If null_score - best_non_null is greater than the threshold predict null.",
+    )
+
+    parser.add_argument(
+        "--max_seq_length",
+        default=384,
+        type=int,
+        help="The maximum total input sequence length after WordPiece tokenization. Sequences "
+        "longer than this will be truncated, and sequences shorter than this will be padded.",
+    )
+    parser.add_argument(
+        "--doc_stride",
+        default=128,
+        type=int,
+        help="When splitting up a long document into chunks, how much stride to take between chunks.",
+    )
+    parser.add_argument(
+        "--max_query_length",
+        default=64,
+        type=int,
+        help="The maximum number of tokens for the question. Questions longer than this will "
+        "be truncated to this length.",
+    )
+    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
+    parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
+    parser.add_argument(
+        "--evaluate_during_training", action="store_true", help="Rul evaluation during training at each logging step."
+    )
+    parser.add_argument(
+        "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
+    )
+
+    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.")
+    parser.add_argument(
+        "--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation."
+    )
+    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument(
+        "--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform."
+    )
+    parser.add_argument(
+        "--max_steps",
+        default=-1,
+        type=int,
+        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
+    )
+    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
+    parser.add_argument(
+        "--n_best_size",
+        default=20,
+        type=int,
+        help="The total number of n-best predictions to generate in the nbest_predictions.json output file.",
+    )
+    parser.add_argument(
+        "--max_answer_length",
+        default=30,
+        type=int,
+        help="The maximum length of an answer that can be generated. This is needed because the start "
+        "and end predictions are not conditioned on one another.",
+    )
+    parser.add_argument(
+        "--verbose_logging",
+        action="store_true",
+        help="If true, all of the warnings related to data processing will be printed. "
+        "A number of warnings are expected for a normal SQuAD evaluation.",
+    )
+
+    parser.add_argument("--logging_steps", type=int, default=50, help="Log every X updates steps.")
+    parser.add_argument("--save_steps", type=int, default=50, help="Save checkpoint every X updates steps.")
+    parser.add_argument(
+        "--eval_all_checkpoints",
+        action="store_true",
+        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
+    )
+    parser.add_argument("--no_cuda", action="store_true", help="Whether not to use CUDA when available")
+    parser.add_argument(
+        "--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory"
+    )
+    parser.add_argument(
+        "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
+    )
+    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
+
+    parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
+    parser.add_argument(
+        "--fp16",
+        action="store_true",
+        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
+    )
+    parser.add_argument(
+        "--fp16_opt_level",
+        type=str,
+        default="O1",
+        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+        "See details at https://nvidia.github.io/apex/amp.html",
+    )
+    parser.add_argument("--server_ip", type=str, default="", help="Can be used for distant debugging.")
+    parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")
+
+    parser.add_argument("--threads", type=int, default=1, help="multiple threads for converting example to features")
+    args = parser.parse_args()
+
+    if (
+        os.path.exists(args.output_dir)
+        and os.listdir(args.output_dir)
+        and args.do_train
+        and not args.overwrite_output_dir
+    ):
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
+                args.output_dir
+            )
+        )
+
+    # Setup distant debugging if needed
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    # Setup CUDA, GPU & distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend="nccl")
+        args.n_gpu = 1
+    args.device = device
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
+    )
+    logger.warning(
+        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+        args.local_rank,
+        device,
+        args.n_gpu,
+        bool(args.local_rank != -1),
+        args.fp16,
+    )
+
+    # Set seed
+    set_seed(args)
+
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        # Make sure only the first process in distributed training will download model & vocab
+        torch.distributed.barrier()
+
+    args.model_type = args.model_type.lower()
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(
+        args.config_name if args.config_name else args.model_name_or_path,
+        cache_dir=args.cache_dir if args.cache_dir else None,
+    )
+    tokenizer = tokenizer_class.from_pretrained(
+        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+        do_lower_case=args.do_lower_case,
+        cache_dir=args.cache_dir if args.cache_dir else None,
+    )
+    model = model_class.from_pretrained(
+        args.model_name_or_path,
+        from_tf=bool(".ckpt" in args.model_name_or_path),
+        config=config,
+        cache_dir=args.cache_dir if args.cache_dir else None,
+    )
+
+    if args.teacher_type is not None:
+        assert args.teacher_name_or_path is not None
+        assert args.alpha_ce > 0.0
+        assert args.alpha_ce + args.alpha_squad > 0.0
+        assert args.teacher_type != "distilbert", "We constraint teachers not to be of type DistilBERT."
+        teacher_config_class, teacher_model_class, _ = MODEL_CLASSES[args.teacher_type]
+        teacher_config = teacher_config_class.from_pretrained(
+            args.teacher_name_or_path, cache_dir=args.cache_dir if args.cache_dir else None
+        )
+        teacher = teacher_model_class.from_pretrained(
+            args.teacher_name_or_path, config=teacher_config, cache_dir=args.cache_dir if args.cache_dir else None
+        )
+        teacher.to(args.device)
+    else:
+        teacher = None
+
+    if args.local_rank == 0:
+        # Make sure only the first process in distributed training will download model & vocab
+        torch.distributed.barrier()
+
+    model.to(args.device)
+
+    logger.info("Training/evaluation parameters %s", args)
+
+    # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
+    # Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
+    # remove the need for this code, but it is still valid.
+    if args.fp16:
+        try:
+            import apex
+
+            apex.amp.register_half_function(torch, "einsum")
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+
+    # Training
+    if args.do_train:
+        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer, teacher=teacher)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+
+    # Save the trained model and the tokenizer
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = (
+            model.module if hasattr(model, "module") else model
+        )  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+
+        # Good practice: save your training arguments together with the trained model
+        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
+
+        # Load a trained model and vocabulary that you have fine-tuned
+        model = model_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model.to(args.device)
+
+    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        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
+
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+
+        for checkpoint in checkpoints:
+            # Reload the model
+            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
+            model = model_class.from_pretrained(checkpoint)
+            model.to(args.device)
+
+            # Evaluate
+            result = evaluate(args, model, tokenizer, prefix=global_step)
+
+            result = dict((k + ("_{}".format(global_step) if global_step else ""), v) for k, v in result.items())
+            results.update(result)
+
+    logger.info("Results: {}".format(results))
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

examples/distillation/scripts/binarized_data.py

@@ -13,65 +13,84 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-Preprocessing script before training DistilBERT.
+Preprocessing script before distillation.
 """
 import argparse
+import logging
 import pickle
 import random
 import time
-import numpy as np
-from pytorch_transformers import BertTokenizer
 
-from examples.distillation.utils import logger
+import numpy as np
+
+from transformers import BertTokenizer, GPT2Tokenizer, RobertaTokenizer
+
+
+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__)
+
 
 def main():
-    parser = argparse.ArgumentParser(description="Preprocess the data to avoid re-doing it several times by (tokenization + token_to_ids).")
-    parser.add_argument('--file_path', type=str, default='data/dump.txt',
-                        help='The path to the data.')
-    parser.add_argument('--bert_tokenizer', type=str, default='bert-base-uncased',
-                        help="The tokenizer to use.")
-    parser.add_argument('--dump_file', type=str, default='data/dump',
-                        help='The dump file prefix.')
+    parser = argparse.ArgumentParser(
+        description="Preprocess the data to avoid re-doing it several times by (tokenization + token_to_ids)."
+    )
+    parser.add_argument("--file_path", type=str, default="data/dump.txt", help="The path to the data.")
+    parser.add_argument("--tokenizer_type", type=str, default="bert", choices=["bert", "roberta", "gpt2"])
+    parser.add_argument("--tokenizer_name", type=str, default="bert-base-uncased", help="The tokenizer to use.")
+    parser.add_argument("--dump_file", type=str, default="data/dump", help="The dump file prefix.")
     args = parser.parse_args()
 
+    logger.info(f"Loading Tokenizer ({args.tokenizer_name})")
+    if args.tokenizer_type == "bert":
+        tokenizer = BertTokenizer.from_pretrained(args.tokenizer_name)
+        bos = tokenizer.special_tokens_map["cls_token"]  # `[CLS]`
+        sep = tokenizer.special_tokens_map["sep_token"]  # `[SEP]`
+    elif args.tokenizer_type == "roberta":
+        tokenizer = RobertaTokenizer.from_pretrained(args.tokenizer_name)
+        bos = tokenizer.special_tokens_map["cls_token"]  # `<s>`
+        sep = tokenizer.special_tokens_map["sep_token"]  # `</s>`
+    elif args.tokenizer_type == "gpt2":
+        tokenizer = GPT2Tokenizer.from_pretrained(args.tokenizer_name)
+        bos = tokenizer.special_tokens_map["bos_token"]  # `<|endoftext|>`
+        sep = tokenizer.special_tokens_map["eos_token"]  # `<|endoftext|>`
 
-    logger.info(f'Loading Tokenizer ({args.bert_tokenizer})')
-    bert_tokenizer = BertTokenizer.from_pretrained(args.bert_tokenizer)
-
-
-    logger.info(f'Loading text from {args.file_path}')
-    with open(args.file_path, 'r', encoding='utf8') as fp:
+    logger.info(f"Loading text from {args.file_path}")
+    with open(args.file_path, "r", encoding="utf8") as fp:
         data = fp.readlines()
 
-
-    logger.info(f'Start encoding')
-    logger.info(f'{len(data)} examples to process.')
+    logger.info(f"Start encoding")
+    logger.info(f"{len(data)} examples to process.")
 
     rslt = []
     iter = 0
     interval = 10000
     start = time.time()
     for text in data:
-        text = f'[CLS] {text.strip()} [SEP]'
-        token_ids = bert_tokenizer.encode(text)
+        text = f"{bos} {text.strip()} {sep}"
+        token_ids = tokenizer.encode(text, add_special_tokens=False)
         rslt.append(token_ids)
 
         iter += 1
         if iter % interval == 0:
             end = time.time()
-            logger.info(f'{iter} examples processed. - {(end-start)/interval:.2f}s/expl')
+            logger.info(f"{iter} examples processed. - {(end-start):.2f}s/{interval}expl")
             start = time.time()
-    logger.info('Finished binarization')
-    logger.info(f'{len(data)} examples processed.')
+    logger.info("Finished binarization")
+    logger.info(f"{len(data)} examples processed.")
 
-
-    dp_file = f'{args.dump_file}.{args.bert_tokenizer}.pickle'
-    rslt_ = [np.uint16(d) for d in rslt]
+    dp_file = f"{args.dump_file}.{args.tokenizer_name}.pickle"
+    vocab_size = tokenizer.vocab_size
+    if vocab_size < (1 << 16):
+        rslt_ = [np.uint16(d) for d in rslt]
+    else:
+        rslt_ = [np.int32(d) for d in rslt]
     random.shuffle(rslt_)
-    logger.info(f'Dump to {dp_file}')
-    with open(dp_file, 'wb') as handle:
+    logger.info(f"Dump to {dp_file}")
+    with open(dp_file, "wb") as handle:
         pickle.dump(rslt_, handle, protocol=pickle.HIGHEST_PROTOCOL)
 
 
 if __name__ == "__main__":
-    main()
+    main()

examples/distillation/scripts/extract.py

@@ -0,0 +1,102 @@
+# coding=utf-8
+# Copyright 2019-present, 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.
+"""
+Preprocessing script before training the distilled model.
+Specific to RoBERTa -> DistilRoBERTa and GPT2 -> DistilGPT2.
+"""
+import argparse
+
+import torch
+
+from transformers import GPT2LMHeadModel, RobertaForMaskedLM
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Extraction some layers of the full RobertaForMaskedLM or GPT2LMHeadModel for Transfer Learned Distillation"
+    )
+    parser.add_argument("--model_type", default="roberta", choices=["roberta", "gpt2"])
+    parser.add_argument("--model_name", default="roberta-large", type=str)
+    parser.add_argument("--dump_checkpoint", default="serialization_dir/tf_roberta_048131723.pth", type=str)
+    parser.add_argument("--vocab_transform", action="store_true")
+    args = parser.parse_args()
+
+    if args.model_type == "roberta":
+        model = RobertaForMaskedLM.from_pretrained(args.model_name)
+        prefix = "roberta"
+    elif args.model_type == "gpt2":
+        model = GPT2LMHeadModel.from_pretrained(args.model_name)
+        prefix = "transformer"
+
+    state_dict = model.state_dict()
+    compressed_sd = {}
+
+    # Embeddings #
+    if args.model_type == "gpt2":
+        for param_name in ["wte.weight", "wpe.weight"]:
+            compressed_sd[f"{prefix}.{param_name}"] = state_dict[f"{prefix}.{param_name}"]
+    else:
+        for w in ["word_embeddings", "position_embeddings", "token_type_embeddings"]:
+            param_name = f"{prefix}.embeddings.{w}.weight"
+            compressed_sd[param_name] = state_dict[param_name]
+        for w in ["weight", "bias"]:
+            param_name = f"{prefix}.embeddings.LayerNorm.{w}"
+            compressed_sd[param_name] = state_dict[param_name]
+
+    # Transformer Blocks #
+    std_idx = 0
+    for teacher_idx in [0, 2, 4, 7, 9, 11]:
+        if args.model_type == "gpt2":
+            for layer in ["ln_1", "attn.c_attn", "attn.c_proj", "ln_2", "mlp.c_fc", "mlp.c_proj"]:
+                for w in ["weight", "bias"]:
+                    compressed_sd[f"{prefix}.h.{std_idx}.{layer}.{w}"] = state_dict[
+                        f"{prefix}.h.{teacher_idx}.{layer}.{w}"
+                    ]
+            compressed_sd[f"{prefix}.h.{std_idx}.attn.bias"] = state_dict[f"{prefix}.h.{teacher_idx}.attn.bias"]
+        else:
+            for layer in [
+                "attention.self.query",
+                "attention.self.key",
+                "attention.self.value",
+                "attention.output.dense",
+                "attention.output.LayerNorm",
+                "intermediate.dense",
+                "output.dense",
+                "output.LayerNorm",
+            ]:
+                for w in ["weight", "bias"]:
+                    compressed_sd[f"{prefix}.encoder.layer.{std_idx}.{layer}.{w}"] = state_dict[
+                        f"{prefix}.encoder.layer.{teacher_idx}.{layer}.{w}"
+                    ]
+        std_idx += 1
+
+    # Language Modeling Head ###s
+    if args.model_type == "roberta":
+        for layer in ["lm_head.decoder.weight", "lm_head.bias"]:
+            compressed_sd[f"{layer}"] = state_dict[f"{layer}"]
+        if args.vocab_transform:
+            for w in ["weight", "bias"]:
+                compressed_sd[f"lm_head.dense.{w}"] = state_dict[f"lm_head.dense.{w}"]
+                compressed_sd[f"lm_head.layer_norm.{w}"] = state_dict[f"lm_head.layer_norm.{w}"]
+    elif args.model_type == "gpt2":
+        for w in ["weight", "bias"]:
+            compressed_sd[f"{prefix}.ln_f.{w}"] = state_dict[f"{prefix}.ln_f.{w}"]
+        compressed_sd[f"lm_head.weight"] = state_dict[f"lm_head.weight"]
+
+    print(f"N layers selected for distillation: {std_idx}")
+    print(f"Number of params transfered for distillation: {len(compressed_sd.keys())}")
+
+    print(f"Save transfered checkpoint to {args.dump_checkpoint}.")
+    torch.save(compressed_sd, args.dump_checkpoint)

examples/distillation/scripts/extract_distilbert.py

@@ -0,0 +1,92 @@
+# coding=utf-8
+# Copyright 2019-present, 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.
+"""
+Preprocessing script before training DistilBERT.
+Specific to BERT -> DistilBERT.
+"""
+import argparse
+
+import torch
+
+from transformers import BertForMaskedLM
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Extraction some layers of the full BertForMaskedLM or RObertaForMaskedLM for Transfer Learned Distillation"
+    )
+    parser.add_argument("--model_type", default="bert", choices=["bert"])
+    parser.add_argument("--model_name", default="bert-base-uncased", type=str)
+    parser.add_argument("--dump_checkpoint", default="serialization_dir/tf_bert-base-uncased_0247911.pth", type=str)
+    parser.add_argument("--vocab_transform", action="store_true")
+    args = parser.parse_args()
+
+    if args.model_type == "bert":
+        model = BertForMaskedLM.from_pretrained(args.model_name)
+        prefix = "bert"
+    else:
+        raise ValueError(f'args.model_type should be "bert".')
+
+    state_dict = model.state_dict()
+    compressed_sd = {}
+
+    for w in ["word_embeddings", "position_embeddings"]:
+        compressed_sd[f"distilbert.embeddings.{w}.weight"] = state_dict[f"{prefix}.embeddings.{w}.weight"]
+    for w in ["weight", "bias"]:
+        compressed_sd[f"distilbert.embeddings.LayerNorm.{w}"] = state_dict[f"{prefix}.embeddings.LayerNorm.{w}"]
+
+    std_idx = 0
+    for teacher_idx in [0, 2, 4, 7, 9, 11]:
+        for w in ["weight", "bias"]:
+            compressed_sd[f"distilbert.transformer.layer.{std_idx}.attention.q_lin.{w}"] = state_dict[
+                f"{prefix}.encoder.layer.{teacher_idx}.attention.self.query.{w}"
+            ]
+            compressed_sd[f"distilbert.transformer.layer.{std_idx}.attention.k_lin.{w}"] = state_dict[
+                f"{prefix}.encoder.layer.{teacher_idx}.attention.self.key.{w}"
+            ]
+            compressed_sd[f"distilbert.transformer.layer.{std_idx}.attention.v_lin.{w}"] = state_dict[
+                f"{prefix}.encoder.layer.{teacher_idx}.attention.self.value.{w}"
+            ]
+
+            compressed_sd[f"distilbert.transformer.layer.{std_idx}.attention.out_lin.{w}"] = state_dict[
+                f"{prefix}.encoder.layer.{teacher_idx}.attention.output.dense.{w}"
+            ]
+            compressed_sd[f"distilbert.transformer.layer.{std_idx}.sa_layer_norm.{w}"] = state_dict[
+                f"{prefix}.encoder.layer.{teacher_idx}.attention.output.LayerNorm.{w}"
+            ]
+
+            compressed_sd[f"distilbert.transformer.layer.{std_idx}.ffn.lin1.{w}"] = state_dict[
+                f"{prefix}.encoder.layer.{teacher_idx}.intermediate.dense.{w}"
+            ]
+            compressed_sd[f"distilbert.transformer.layer.{std_idx}.ffn.lin2.{w}"] = state_dict[
+                f"{prefix}.encoder.layer.{teacher_idx}.output.dense.{w}"
+            ]
+            compressed_sd[f"distilbert.transformer.layer.{std_idx}.output_layer_norm.{w}"] = state_dict[
+                f"{prefix}.encoder.layer.{teacher_idx}.output.LayerNorm.{w}"
+            ]
+        std_idx += 1
+
+    compressed_sd[f"vocab_projector.weight"] = state_dict[f"cls.predictions.decoder.weight"]
+    compressed_sd[f"vocab_projector.bias"] = state_dict[f"cls.predictions.bias"]
+    if args.vocab_transform:
+        for w in ["weight", "bias"]:
+            compressed_sd[f"vocab_transform.{w}"] = state_dict[f"cls.predictions.transform.dense.{w}"]
+            compressed_sd[f"vocab_layer_norm.{w}"] = state_dict[f"cls.predictions.transform.LayerNorm.{w}"]
+
+    print(f"N layers selected for distillation: {std_idx}")
+    print(f"Number of params transfered for distillation: {len(compressed_sd.keys())}")
+
+    print(f"Save transfered checkpoint to {args.dump_checkpoint}.")
+    torch.save(compressed_sd, args.dump_checkpoint)

examples/distillation/scripts/extract_for_distil.py

@@ -1,76 +0,0 @@
-# coding=utf-8
-# Copyright 2019-present, 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.
-"""
-Preprocessing script before training DistilBERT.
-"""
-from pytorch_transformers import BertForPreTraining
-import torch
-import argparse
-
-if __name__ == '__main__':
-    parser = argparse.ArgumentParser(description="Extraction some layers of the full BertForPreTraining for Transfer Learned Distillation")
-    parser.add_argument("--bert_model", default='bert-base-uncased', type=str)
-    parser.add_argument("--dump_checkpoint", default='serialization_dir/transfer_learning_checkpoint_0247911.pth', type=str)
-    parser.add_argument("--vocab_transform", action='store_true')
-    args = parser.parse_args()
-
-
-    model = BertForPreTraining.from_pretrained(args.bert_model)
-
-    state_dict = model.state_dict()
-    compressed_sd = {}
-
-    for w in ['word_embeddings', 'position_embeddings']:
-        compressed_sd[f'distilbert.embeddings.{w}.weight'] = \
-            state_dict[f'bert.embeddings.{w}.weight']
-    for w in ['weight', 'bias']:
-        compressed_sd[f'distilbert.embeddings.LayerNorm.{w}'] = \
-            state_dict[f'bert.embeddings.LayerNorm.{w}']
-
-    std_idx = 0
-    for teacher_idx in [0, 2, 4, 7, 9, 11]:
-        for w in ['weight', 'bias']:
-            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.q_lin.{w}'] = \
-                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.self.query.{w}']
-            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.k_lin.{w}'] = \
-                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.self.key.{w}']
-            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.v_lin.{w}'] = \
-                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.self.value.{w}']
-
-            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.out_lin.{w}'] = \
-                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.output.dense.{w}']
-            compressed_sd[f'distilbert.transformer.layer.{std_idx}.sa_layer_norm.{w}'] = \
-                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.output.LayerNorm.{w}']
-
-            compressed_sd[f'distilbert.transformer.layer.{std_idx}.ffn.lin1.{w}'] = \
-                state_dict[f'bert.encoder.layer.{teacher_idx}.intermediate.dense.{w}']
-            compressed_sd[f'distilbert.transformer.layer.{std_idx}.ffn.lin2.{w}'] = \
-                state_dict[f'bert.encoder.layer.{teacher_idx}.output.dense.{w}']
-            compressed_sd[f'distilbert.transformer.layer.{std_idx}.output_layer_norm.{w}'] = \
-                state_dict[f'bert.encoder.layer.{teacher_idx}.output.LayerNorm.{w}']
-        std_idx += 1
-
-    compressed_sd[f'vocab_projector.weight'] = state_dict[f'cls.predictions.decoder.weight']
-    compressed_sd[f'vocab_projector.bias'] = state_dict[f'cls.predictions.bias']
-    if args.vocab_transform:
-        for w in ['weight', 'bias']:
-            compressed_sd[f'vocab_transform.{w}'] = state_dict[f'cls.predictions.transform.dense.{w}']
-            compressed_sd[f'vocab_layer_norm.{w}'] = state_dict[f'cls.predictions.transform.LayerNorm.{w}']
-
-    print(f'N layers selected for distillation: {std_idx}')
-    print(f'Number of params transfered for distillation: {len(compressed_sd.keys())}')
-
-    print(f'Save transfered checkpoint to {args.dump_checkpoint}.')
-    torch.save(compressed_sd, args.dump_checkpoint)

examples/distillation/scripts/token_counts.py

@@ -13,35 +13,44 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-Preprocessing script before training DistilBERT.
+Preprocessing script before training the distilled model.
 """
-from collections import Counter
 import argparse
+import logging
 import pickle
+from collections import Counter
 
-from examples.distillation.utils import logger
 
-if __name__ == '__main__':
-    parser = argparse.ArgumentParser(description="Token Counts for smoothing the masking probabilities in MLM (cf XLM/word2vec)")
-    parser.add_argument("--data_file", type=str, default="data/dump.bert-base-uncased.pickle",
-                        help="The binarized dataset.")
-    parser.add_argument("--token_counts_dump", type=str, default="data/token_counts.bert-base-uncased.pickle",
-                        help="The dump file.")
+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__)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Token Counts for smoothing the masking probabilities in MLM (cf XLM/word2vec)"
+    )
+    parser.add_argument(
+        "--data_file", type=str, default="data/dump.bert-base-uncased.pickle", help="The binarized dataset."
+    )
+    parser.add_argument(
+        "--token_counts_dump", type=str, default="data/token_counts.bert-base-uncased.pickle", help="The dump file."
+    )
     parser.add_argument("--vocab_size", default=30522, type=int)
     args = parser.parse_args()
 
-    logger.info(f'Loading data from {args.data_file}')
-    with open(args.data_file, 'rb') as fp:
+    logger.info(f"Loading data from {args.data_file}")
+    with open(args.data_file, "rb") as fp:
         data = pickle.load(fp)
 
-    logger.info('Counting occurences for MLM.')
+    logger.info("Counting occurences for MLM.")
     counter = Counter()
     for tk_ids in data:
         counter.update(tk_ids)
-    counts = [0]*args.vocab_size
+    counts = [0] * args.vocab_size
     for k, v in counter.items():
         counts[k] = v
 
-    logger.info(f'Dump to {args.token_counts_dump}')
-    with open(args.token_counts_dump, 'wb') as handle:
+    logger.info(f"Dump to {args.token_counts_dump}")
+    with open(args.token_counts_dump, "wb") as handle:
         pickle.dump(counts, handle, protocol=pickle.HIGHEST_PROTOCOL)

examples/distillation/train.py

@@ -13,221 +13,307 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-Training DistilBERT.
+Training the distilled model.
+Supported architectures include: BERT -> DistilBERT, RoBERTa -> DistilRoBERTa, GPT2 -> DistilGPT2.
 """
-import os
 import argparse
-import pickle
 import json
+import os
+import pickle
 import shutil
+
 import numpy as np
 import torch
 
-from pytorch_transformers import BertTokenizer, BertForMaskedLM
-from pytorch_transformers import DistilBertForMaskedLM, DistilBertConfig
-
 from distiller import Distiller
-from utils import git_log, logger, init_gpu_params, set_seed
-from dataset import Dataset
+from lm_seqs_dataset import LmSeqsDataset
+from transformers import (
+    BertConfig,
+    BertForMaskedLM,
+    BertTokenizer,
+    DistilBertConfig,
+    DistilBertForMaskedLM,
+    DistilBertTokenizer,
+    GPT2Config,
+    GPT2LMHeadModel,
+    GPT2Tokenizer,
+    RobertaConfig,
+    RobertaForMaskedLM,
+    RobertaTokenizer,
+)
+from utils import git_log, init_gpu_params, logger, set_seed
+
+
+MODEL_CLASSES = {
+    "distilbert": (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer),
+    "roberta": (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer),
+    "bert": (BertConfig, BertForMaskedLM, BertTokenizer),
+    "gpt2": (GPT2Config, GPT2LMHeadModel, GPT2Tokenizer),
+}
+
+
+def sanity_checks(args):
+    """
+    A bunch of args sanity checks to perform even starting...
+    """
+    assert (args.mlm and args.alpha_mlm > 0.0) or (not args.mlm and args.alpha_mlm == 0.0)
+    assert (args.alpha_mlm > 0.0 and args.alpha_clm == 0.0) or (args.alpha_mlm == 0.0 and args.alpha_clm > 0.0)
+    if args.mlm:
+        assert os.path.isfile(args.token_counts)
+        assert (args.student_type in ["roberta", "distilbert"]) and (args.teacher_type in ["roberta", "bert"])
+    else:
+        assert (args.student_type in ["gpt2"]) and (args.teacher_type in ["gpt2"])
+
+    assert args.teacher_type == args.student_type or (
+        args.student_type == "distilbert" and args.teacher_type == "bert"
+    )
+    assert os.path.isfile(args.student_config)
+    if args.student_pretrained_weights is not None:
+        assert os.path.isfile(args.student_pretrained_weights)
+
+    if args.freeze_token_type_embds:
+        assert args.student_type in ["roberta"]
+
+    assert args.alpha_ce >= 0.0
+    assert args.alpha_mlm >= 0.0
+    assert args.alpha_clm >= 0.0
+    assert args.alpha_mse >= 0.0
+    assert args.alpha_cos >= 0.0
+    assert args.alpha_ce + args.alpha_mlm + args.alpha_clm + args.alpha_mse + args.alpha_cos > 0.0
+
+
+def freeze_pos_embeddings(student, args):
+    if args.student_type == "roberta":
+        student.roberta.embeddings.position_embeddings.weight.requires_grad = False
+    elif args.student_type == "gpt2":
+        student.transformer.wpe.weight.requires_grad = False
+
+
+def freeze_token_type_embeddings(student, args):
+    if args.student_type == "roberta":
+        student.roberta.embeddings.token_type_embeddings.weight.requires_grad = False
 
 
 def main():
     parser = argparse.ArgumentParser(description="Training")
+    parser.add_argument("--force", action="store_true", help="Overwrite dump_path if it already exists.")
 
-    parser.add_argument("--dump_path", type=str, required=True,
-                        help="The output directory (log, checkpoints, parameters, etc.)")
-    parser.add_argument("--data_file", type=str, required=True,
-                        help="The binarized file (tokenized + tokens_to_ids) and grouped by sequence.")
-    parser.add_argument("--token_counts", type=str, required=True,
-                        help="The token counts in the data_file for MLM.")
-    parser.add_argument("--force", action='store_true',
-                        help="Overwrite dump_path if it already exists.")
+    parser.add_argument(
+        "--dump_path", type=str, required=True, help="The output directory (log, checkpoints, parameters, etc.)"
+    )
+    parser.add_argument(
+        "--data_file",
+        type=str,
+        required=True,
+        help="The binarized file (tokenized + tokens_to_ids) and grouped by sequence.",
+    )
 
-    parser.add_argument("--vocab_size", default=30522, type=int,
-                        help="The vocabulary size.")
-    parser.add_argument("--max_position_embeddings", default=512, type=int,
-                        help="Maximum sequence length we can model (including [CLS] and [SEP]).")
-    parser.add_argument("--sinusoidal_pos_embds", action='store_false',
-                        help="If true, the position embeddings are simply fixed with sinusoidal embeddings.")
-    parser.add_argument("--n_layers", default=6, type=int,
-                        help="Number of Transformer blocks.")
-    parser.add_argument("--n_heads", default=12, type=int,
-                        help="Number of heads in the self-attention module.")
-    parser.add_argument("--dim", default=768, type=int,
-                        help="Dimension through the network. Must be divisible by n_heads")
-    parser.add_argument("--hidden_dim", default=3072, type=int,
-                        help="Intermediate dimension in the FFN.")
-    parser.add_argument("--dropout", default=0.1, type=float,
-                        help="Dropout.")
-    parser.add_argument("--attention_dropout", default=0.1, type=float,
-                        help="Dropout in self-attention.")
-    parser.add_argument("--activation", default='gelu', type=str,
-                        help="Activation to use in self-attention")
-    parser.add_argument("--tie_weights_", action='store_false',
-                        help="If true, we tie the embeddings matrix with the projection over the vocabulary matrix. Default is true.")
+    parser.add_argument(
+        "--student_type",
+        type=str,
+        choices=["distilbert", "roberta", "gpt2"],
+        required=True,
+        help="The student type (DistilBERT, RoBERTa).",
+    )
+    parser.add_argument("--student_config", type=str, required=True, help="Path to the student configuration.")
+    parser.add_argument(
+        "--student_pretrained_weights", default=None, type=str, help="Load student initialization checkpoint."
+    )
 
-    parser.add_argument("--from_pretrained_weights", default=None, type=str,
-                        help="Load student initialization checkpoint.")
-    parser.add_argument("--from_pretrained_config", default=None, type=str,
-                        help="Load student initialization architecture config.")
-    parser.add_argument("--bert_model", default='bert-base-uncased', type=str,
-                        help="The teacher BERT model.")
+    parser.add_argument(
+        "--teacher_type", choices=["bert", "roberta", "gpt2"], required=True, help="Teacher type (BERT, RoBERTa)."
+    )
+    parser.add_argument("--teacher_name", type=str, required=True, help="The teacher model.")
 
-    parser.add_argument("--temperature", default=2., type=float,
-                        help="Temperature for the softmax temperature.")
-    parser.add_argument("--alpha_ce", default=0.5, type=float,
-                        help="Linear weight for the distillation loss. Must be >=0.")
-    parser.add_argument("--alpha_mlm", default=0.5, type=float,
-                        help="Linear weight for the MLM loss. Must be >=0.")
-    parser.add_argument("--alpha_mse", default=0.0, type=float,
-                        help="Linear weight of the MSE loss. Must be >=0.")
-    parser.add_argument("--mlm_mask_prop", default=0.15, type=float,
-                        help="Proportion of tokens for which we need to make a prediction.")
-    parser.add_argument("--word_mask", default=0.8, type=float,
-                        help="Proportion of tokens to mask out.")
-    parser.add_argument("--word_keep", default=0.1, type=float,
-                        help="Proportion of tokens to keep.")
-    parser.add_argument("--word_rand", default=0.1, type=float,
-                        help="Proportion of tokens to randomly replace.")
-    parser.add_argument("--mlm_smoothing", default=0.7, type=float,
-                        help="Smoothing parameter to emphasize more rare tokens (see XLM, similar to word2vec).")
-    parser.add_argument("--restrict_ce_to_mask", action='store_true',
-                        help="If true, compute the distilation loss only the [MLM] prediction distribution.")
+    parser.add_argument("--temperature", default=2.0, type=float, help="Temperature for the softmax temperature.")
+    parser.add_argument(
+        "--alpha_ce", default=0.5, type=float, help="Linear weight for the distillation loss. Must be >=0."
+    )
+    parser.add_argument(
+        "--alpha_mlm",
+        default=0.0,
+        type=float,
+        help="Linear weight for the MLM loss. Must be >=0. Should be used in coonjunction with `mlm` flag.",
+    )
+    parser.add_argument("--alpha_clm", default=0.5, type=float, help="Linear weight for the CLM loss. Must be >=0.")
+    parser.add_argument("--alpha_mse", default=0.0, type=float, help="Linear weight of the MSE loss. Must be >=0.")
+    parser.add_argument(
+        "--alpha_cos", default=0.0, type=float, help="Linear weight of the cosine embedding loss. Must be >=0."
+    )
 
-    parser.add_argument("--n_epoch", type=int, default=3,
-                        help="Number of pass on the whole dataset.")
-    parser.add_argument("--batch_size", type=int, default=5,
-                        help="Batch size (for each process).")
-    parser.add_argument("--tokens_per_batch", type=int, default=-1,
-                        help="If specified, modify the batches so that they have approximately this number of tokens.")
-    parser.add_argument("--shuffle", action='store_false',
-                        help="If true, shuffle the sequence order. Default is true.")
-    parser.add_argument("--group_by_size", action='store_false',
-                        help="If true, group sequences that have similar length into the same batch. Default is true.")
+    parser.add_argument(
+        "--mlm", action="store_true", help="The LM step: MLM or CLM. If `mlm` is True, the MLM is used over CLM."
+    )
+    parser.add_argument(
+        "--mlm_mask_prop",
+        default=0.15,
+        type=float,
+        help="Proportion of tokens for which we need to make a prediction.",
+    )
+    parser.add_argument("--word_mask", default=0.8, type=float, help="Proportion of tokens to mask out.")
+    parser.add_argument("--word_keep", default=0.1, type=float, help="Proportion of tokens to keep.")
+    parser.add_argument("--word_rand", default=0.1, type=float, help="Proportion of tokens to randomly replace.")
+    parser.add_argument(
+        "--mlm_smoothing",
+        default=0.7,
+        type=float,
+        help="Smoothing parameter to emphasize more rare tokens (see XLM, similar to word2vec).",
+    )
+    parser.add_argument("--token_counts", type=str, help="The token counts in the data_file for MLM.")
 
-    parser.add_argument("--gradient_accumulation_steps", type=int, default=50,
-                        help="Gradient accumulation for larger training batches.")
-    parser.add_argument("--warmup_prop", default=0.05, type=float,
-                        help="Linear warmup proportion.")
-    parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight deay if we apply some.")
-    parser.add_argument("--learning_rate", default=5e-4, type=float,
-                        help="The initial learning rate for Adam.")
-    parser.add_argument("--adam_epsilon", default=1e-6, type=float,
-                        help="Epsilon for Adam optimizer.")
-    parser.add_argument("--max_grad_norm", default=5.0, type=float,
-                        help="Max gradient norm.")
-    parser.add_argument("--initializer_range", default=0.02, type=float,
-                        help="Random initialization range.")
+    parser.add_argument(
+        "--restrict_ce_to_mask",
+        action="store_true",
+        help="If true, compute the distilation loss only the [MLM] prediction distribution.",
+    )
+    parser.add_argument(
+        "--freeze_pos_embs",
+        action="store_true",
+        help="Freeze positional embeddings during distillation. For student_type in ['roberta', 'gpt2'] only.",
+    )
+    parser.add_argument(
+        "--freeze_token_type_embds",
+        action="store_true",
+        help="Freeze token type embeddings during distillation if existent. For student_type in ['roberta'] only.",
+    )
 
-    parser.add_argument('--fp16', action='store_true',
-                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
-    parser.add_argument('--fp16_opt_level', type=str, default='O1',
-                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
-                             "See details at https://nvidia.github.io/apex/amp.html")
-    parser.add_argument("--n_gpu", type=int, default=1,
-                        help="Number of GPUs in the node.")
-    parser.add_argument("--local_rank", type=int, default=-1,
-                        help="Distributed training - Local rank")
-    parser.add_argument("--seed", type=int, default=56,
-                        help="Random seed")
+    parser.add_argument("--n_epoch", type=int, default=3, help="Number of pass on the whole dataset.")
+    parser.add_argument("--batch_size", type=int, default=5, help="Batch size (for each process).")
+    parser.add_argument(
+        "--group_by_size",
+        action="store_false",
+        help="If true, group sequences that have similar length into the same batch. Default is true.",
+    )
 
-    parser.add_argument("--log_interval", type=int, default=500,
-                        help="Tensorboard logging interval.")
-    parser.add_argument("--checkpoint_interval", type=int, default=4000,
-                        help="Checkpoint interval.")
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=50,
+        help="Gradient accumulation for larger training batches.",
+    )
+    parser.add_argument("--warmup_prop", default=0.05, type=float, help="Linear warmup proportion.")
+    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight deay if we apply some.")
+    parser.add_argument("--learning_rate", default=5e-4, type=float, help="The initial learning rate for Adam.")
+    parser.add_argument("--adam_epsilon", default=1e-6, type=float, help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=5.0, type=float, help="Max gradient norm.")
+    parser.add_argument("--initializer_range", default=0.02, type=float, help="Random initialization range.")
+
+    parser.add_argument(
+        "--fp16",
+        action="store_true",
+        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
+    )
+    parser.add_argument(
+        "--fp16_opt_level",
+        type=str,
+        default="O1",
+        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+        "See details at https://nvidia.github.io/apex/amp.html",
+    )
+    parser.add_argument("--n_gpu", type=int, default=1, help="Number of GPUs in the node.")
+    parser.add_argument("--local_rank", type=int, default=-1, help="Distributed training - Local rank")
+    parser.add_argument("--seed", type=int, default=56, help="Random seed")
+
+    parser.add_argument("--log_interval", type=int, default=500, help="Tensorboard logging interval.")
+    parser.add_argument("--checkpoint_interval", type=int, default=4000, help="Checkpoint interval.")
     args = parser.parse_args()
+    sanity_checks(args)
 
-
-    ## ARGS ##
+    # ARGS #
     init_gpu_params(args)
     set_seed(args)
     if args.is_master:
         if os.path.exists(args.dump_path):
             if not args.force:
-                raise ValueError(f'Serialization dir {args.dump_path} already exists, but you have not precised wheter to overwrite it'
-                                   'Use `--force` if you want to overwrite it')
+                raise ValueError(
+                    f"Serialization dir {args.dump_path} already exists, but you have not precised wheter to overwrite it"
+                    "Use `--force` if you want to overwrite it"
+                )
             else:
                 shutil.rmtree(args.dump_path)
 
         if not os.path.exists(args.dump_path):
             os.makedirs(args.dump_path)
-        logger.info(f'Experiment will be dumped and logged in {args.dump_path}')
+        logger.info(f"Experiment will be dumped and logged in {args.dump_path}")
 
-
-        ### SAVE PARAMS ###
-        logger.info(f'Param: {args}')
-        with open(os.path.join(args.dump_path, 'parameters.json'), 'w') as f:
+        # SAVE PARAMS #
+        logger.info(f"Param: {args}")
+        with open(os.path.join(args.dump_path, "parameters.json"), "w") as f:
             json.dump(vars(args), f, indent=4)
         git_log(args.dump_path)
-    assert (args.from_pretrained_weights is None and args.from_pretrained_config is None) or \
-           (args.from_pretrained_weights is not None and args.from_pretrained_config is not None)
 
+    student_config_class, student_model_class, _ = MODEL_CLASSES[args.student_type]
+    teacher_config_class, teacher_model_class, teacher_tokenizer_class = MODEL_CLASSES[args.teacher_type]
 
-    ### TOKENIZER ###
-    bert_tokenizer = BertTokenizer.from_pretrained(args.bert_model)
+    # TOKENIZER #
+    tokenizer = teacher_tokenizer_class.from_pretrained(args.teacher_name)
     special_tok_ids = {}
-    for tok_name, tok_symbol in bert_tokenizer.special_tokens_map.items():
-        idx = bert_tokenizer.all_special_tokens.index(tok_symbol)
-        special_tok_ids[tok_name] = bert_tokenizer.all_special_ids[idx]
-    logger.info(f'Special tokens {special_tok_ids}')
+    for tok_name, tok_symbol in tokenizer.special_tokens_map.items():
+        idx = tokenizer.all_special_tokens.index(tok_symbol)
+        special_tok_ids[tok_name] = tokenizer.all_special_ids[idx]
+    logger.info(f"Special tokens {special_tok_ids}")
     args.special_tok_ids = special_tok_ids
+    args.max_model_input_size = tokenizer.max_model_input_sizes[args.teacher_name]
 
-
-    ## DATA LOADER ##
-    logger.info(f'Loading data from {args.data_file}')
-    with open(args.data_file, 'rb') as fp:
+    # DATA LOADER #
+    logger.info(f"Loading data from {args.data_file}")
+    with open(args.data_file, "rb") as fp:
         data = pickle.load(fp)
 
+    if args.mlm:
+        logger.info(f"Loading token counts from {args.token_counts} (already pre-computed)")
+        with open(args.token_counts, "rb") as fp:
+            counts = pickle.load(fp)
 
-    assert os.path.isfile(args.token_counts)
-    logger.info(f'Loading token counts from {args.token_counts} (already pre-computed)')
-    with open(args.token_counts, 'rb') as fp:
-        counts = pickle.load(fp)
-        assert len(counts) == args.vocab_size
-    token_probs = np.maximum(counts, 1) ** -args.mlm_smoothing
-    for idx in special_tok_ids.values():
-        token_probs[idx] = 0.  # do not predict special tokens
-    token_probs = torch.from_numpy(token_probs)
-
-
-    train_dataloader = Dataset(params=args, data=data)
-    logger.info(f'Data loader created.')
-
-
-    ## STUDENT ##
-    if args.from_pretrained_weights is not None:
-        assert os.path.isfile(os.path.join(args.from_pretrained_weights))
-        assert os.path.isfile(os.path.join(args.from_pretrained_config))
-        logger.info(f'Loading pretrained weights from {args.from_pretrained_weights}')
-        logger.info(f'Loading pretrained config from {args.from_pretrained_config}')
-        stu_architecture_config = DistilBertConfig.from_json_file(args.from_pretrained_config)
-        student = DistilBertForMaskedLM.from_pretrained(args.from_pretrained_weights,
-                                                     config=stu_architecture_config)
+        token_probs = np.maximum(counts, 1) ** -args.mlm_smoothing
+        for idx in special_tok_ids.values():
+            token_probs[idx] = 0.0  # do not predict special tokens
+        token_probs = torch.from_numpy(token_probs)
     else:
-        args.vocab_size_or_config_json_file = args.vocab_size
-        stu_architecture_config = DistilBertConfig(**vars(args))
-        student = DistilBertForMaskedLM(stu_architecture_config)
+        token_probs = None
 
+    train_lm_seq_dataset = LmSeqsDataset(params=args, data=data)
+    logger.info(f"Data loader created.")
+
+    # STUDENT #
+    logger.info(f"Loading student config from {args.student_config}")
+    stu_architecture_config = student_config_class.from_pretrained(args.student_config)
+    stu_architecture_config.output_hidden_states = True
+
+    if args.student_pretrained_weights is not None:
+        logger.info(f"Loading pretrained weights from {args.student_pretrained_weights}")
+        student = student_model_class.from_pretrained(args.student_pretrained_weights, config=stu_architecture_config)
+    else:
+        student = student_model_class(stu_architecture_config)
 
     if args.n_gpu > 0:
-        student.to(f'cuda:{args.local_rank}')
-    logger.info(f'Student loaded.')
+        student.to(f"cuda:{args.local_rank}")
+    logger.info(f"Student loaded.")
 
-
-    ## TEACHER ##
-    teacher = BertForMaskedLM.from_pretrained(args.bert_model)
+    # TEACHER #
+    teacher = teacher_model_class.from_pretrained(args.teacher_name, output_hidden_states=True)
     if args.n_gpu > 0:
-        teacher.to(f'cuda:{args.local_rank}')
-    logger.info(f'Teacher loaded from {args.bert_model}.')
+        teacher.to(f"cuda:{args.local_rank}")
+    logger.info(f"Teacher loaded from {args.teacher_name}.")
 
-    ## DISTILLER ##
+    # FREEZING #
+    if args.freeze_pos_embs:
+        freeze_pos_embeddings(student, args)
+    if args.freeze_token_type_embds:
+        freeze_token_type_embeddings(student, args)
+
+    # SANITY CHECKS #
+    assert student.config.vocab_size == teacher.config.vocab_size
+    assert student.config.hidden_size == teacher.config.hidden_size
+    assert student.config.max_position_embeddings == teacher.config.max_position_embeddings
+    if args.mlm:
+        assert token_probs.size(0) == stu_architecture_config.vocab_size
+
+    # DISTILLER #
     torch.cuda.empty_cache()
-    distiller = Distiller(params=args,
-                          dataloader=train_dataloader,
-                          token_probs=token_probs,
-                          student=student,
-                          teacher=teacher)
+    distiller = Distiller(
+        params=args, dataset=train_lm_seq_dataset, token_probs=token_probs, student=student, teacher=teacher
+    )
     distiller.train()
     logger.info("Let's go get some drinks.")
 

examples/distillation/training_configs/distilbert-base-cased.json

@@ -0,0 +1,15 @@
+{
+	"activation": "gelu",
+	"attention_dropout": 0.1,
+	"dim": 768,
+	"dropout": 0.1,
+	"hidden_dim": 3072,
+	"initializer_range": 0.02,
+	"max_position_embeddings": 512,
+	"n_heads": 12,
+	"n_layers": 6,
+	"sinusoidal_pos_embds": true,
+	"tie_weights_": true,
+	"vocab_size": 28996
+  }
+  

examples/distillation/training_configs/distilbert-base-multilingual-cased.json

@@ -0,0 +1,15 @@
+{
+	"activation": "gelu",
+	"attention_dropout": 0.1,
+	"dim": 768,
+	"dropout": 0.1,
+	"hidden_dim": 3072,
+	"initializer_range": 0.02,
+	"max_position_embeddings": 512,
+	"n_heads": 12,
+	"n_layers": 6,
+	"sinusoidal_pos_embds": true,
+	"tie_weights_": true,
+	"vocab_size": 119547
+  }
+  

examples/distillation/training_configs/distilbert-base-uncased.json

@@ -0,0 +1,15 @@
+{
+	"activation": "gelu",
+	"attention_dropout": 0.1,
+	"dim": 768,
+	"dropout": 0.1,
+	"hidden_dim": 3072,
+	"initializer_range": 0.02,
+	"max_position_embeddings": 512,
+	"n_heads": 12,
+	"n_layers": 6,
+	"sinusoidal_pos_embds": true,
+	"tie_weights_": true,
+	"vocab_size": 30522
+  }
+  

examples/distillation/training_configs/distilgpt2.json

@@ -0,0 +1,10 @@
+{
+	"initializer_range": 0.02,
+	"layer_norm_epsilon": 0.00001,
+	"n_ctx": 1024,
+	"n_embd": 768,
+	"n_head": 12,
+	"n_layer": 6,
+	"n_positions": 1024,
+	"vocab_size": 50257
+}

examples/distillation/training_configs/distilroberta-base.json

@@ -0,0 +1,14 @@
+{
+    "vocab_size": 50265,
+    "hidden_size": 768,
+    "num_hidden_layers": 6,
+    "num_attention_heads": 12,
+    "intermediate_size": 3072,
+    "hidden_act": "gelu",
+    "hidden_dropout_prob": 0.1,
+    "attention_probs_dropout_prob": 0.1,
+    "max_position_embeddings": 514,
+    "type_vocab_size": 1,
+    "initializer_range": 0.02,
+    "layer_norm_eps": 0.00001
+}

examples/distillation/utils.py

@@ -15,17 +15,21 @@
 """ Utils to train DistilBERT
     adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
 """
-import git
 import json
+import logging
 import os
 import socket
-import torch
-import numpy as np
 
-import logging
-logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - PID: %(process)d -  %(message)s',
-                    datefmt = '%m/%d/%Y %H:%M:%S',
-                    level = logging.INFO)
+import git
+import numpy as np
+import torch
+
+
+logging.basicConfig(
+    format="%(asctime)s - %(levelname)s - %(name)s - PID: %(process)d -  %(message)s",
+    datefmt="%m/%d/%Y %H:%M:%S",
+    level=logging.INFO,
+)
 logger = logging.getLogger(__name__)
 
 
@@ -35,12 +39,12 @@ def git_log(folder_path: str):
     """
     repo = git.Repo(search_parent_directories=True)
     repo_infos = {
-        'repo_id': str(repo),
-        'repo_sha': str(repo.head.object.hexsha),
-        'repo_branch': str(repo.active_branch)
+        "repo_id": str(repo),
+        "repo_sha": str(repo.head.object.hexsha),
+        "repo_branch": str(repo.active_branch),
     }
 
-    with open(os.path.join(folder_path, 'git_log.json'), 'w') as f:
+    with open(os.path.join(folder_path, "git_log.json"), "w") as f:
         json.dump(repo_infos, f, indent=4)
 
 
@@ -57,21 +61,21 @@ def init_gpu_params(params):
 
     assert torch.cuda.is_available()
 
-    logger.info('Initializing GPUs')
+    logger.info("Initializing GPUs")
     if params.n_gpu > 1:
         assert params.local_rank != -1
 
-        params.world_size = int(os.environ['WORLD_SIZE'])
-        params.n_gpu_per_node = int(os.environ['N_GPU_NODE'])
-        params.global_rank = int(os.environ['RANK'])
+        params.world_size = int(os.environ["WORLD_SIZE"])
+        params.n_gpu_per_node = int(os.environ["N_GPU_NODE"])
+        params.global_rank = int(os.environ["RANK"])
 
         # number of nodes / node ID
         params.n_nodes = params.world_size // params.n_gpu_per_node
         params.node_id = params.global_rank // params.n_gpu_per_node
         params.multi_gpu = True
 
-        assert params.n_nodes == int(os.environ['N_NODES'])
-        assert params.node_id == int(os.environ['NODE_RANK'])
+        assert params.n_nodes == int(os.environ["N_NODES"])
+        assert params.node_id == int(os.environ["NODE_RANK"])
 
     # local job (single GPU)
     else:
@@ -114,8 +118,7 @@ def init_gpu_params(params):
     if params.multi_gpu:
         logger.info("Initializing PyTorch distributed")
         torch.distributed.init_process_group(
-            init_method='env://',
-            backend='nccl',
+            init_method="env://", backend="nccl",
         )
 
 

examples/glue/README.md

@@ -0,0 +1,9 @@
+# GLUE Benchmark
+
+Based on the script [`run_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/run_glue.py).
+
+#### Run PyTorch version using PyTorch-Lightning
+
+Run `bash run_pl.sh` from the `glue` directory. This will also install `pytorch-lightning` and the requirements in `examples/requirements.txt`. It is a shell pipeline that will automatically download, pre-process the data and run the specified models. Logs are saved in `lightning_logs` directory.
+
+Pass `--n_gpu` flag to change the number of GPUs. Default uses 1. At the end, the expected results are: `TEST RESULTS {'val_loss': tensor(0.0707), 'precision': 0.852427800698191, 'recall': 0.869537067011978, 'f1': 0.8608974358974358}`