Merge branch 'master' into conditional-generation

.circleci/config.yml

@@ -9,7 +9,7 @@ jobs:
         steps:
             - checkout
             - run: sudo pip install torch
-            - run: sudo pip install tensorflow==2.0.0-rc0
+            - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
             - run: sudo pip install pytest codecov pytest-cov
             - run: sudo pip install tensorboardX scikit-learn
@@ -38,7 +38,7 @@ jobs:
         parallelism: 1
         steps:
             - checkout
-            - run: sudo pip install tensorflow==2.0.0-rc0
+            - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
             - run: sudo pip install pytest codecov pytest-cov
             - run: sudo pip install tensorboardX scikit-learn
@@ -65,7 +65,7 @@ jobs:
             - image: circleci/python:2.7
         steps:
             - checkout
-            - run: sudo pip install tensorflow==2.0.0-rc0
+            - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
             - run: sudo pip install pytest codecov pytest-cov
             - run: python -m pytest -sv ./transformers/tests/ --cov

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

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

.gitignore

@@ -118,6 +118,9 @@ dmypy.json
 # vscode
 .vscode
 
+# Pycharm
+.idea
+
 # TF code
 tensorflow_code
 
@@ -132,3 +135,6 @@ examples/runs
 # data
 /data
 serialization_dir
+
+# emacs
+*.*~
\ No newline at end of file

CONTRIBUTING.md

+# 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, execute the following code and copy-paste
+the output:
+
+```
+import platform; print("Platform", platform.platform())
+import sys; print("Python", sys.version)
+import torch; print("PyTorch", torch.__version__)
+import tensorflow; print("Tensorflow", tensorflow.__version__)
+```
+
+### 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.
+
+### 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.
+
+## 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 git@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 -r requirements-dev.txt
+   ```
+
+5. Develop the features on your branch. Add changed files using `git add` and
+   then `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 pre-existing tests still pass;
+5. Add high-coverage tests. No quality test, no merge;
+6. All public methods must have informative doctrings;
+
+
+### 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)

README.md

@@ -22,7 +22,7 @@
 <p>State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch
 </h3>
 
-🤗 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...) 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.
+🤗 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.
 
 ### Features
 
@@ -39,7 +39,7 @@ State-of-the-art NLP for everyone
 Lower compute costs, smaller carbon footprint
 - Researchers can share trained models instead of always retraining
 - Practitioners can reduce compute time and production costs
-- 8 architectures with over 30 pretrained models, some in more than 100 languages
+- 10 architectures with over 30 pretrained models, some in more than 100 languages
 
 Choose the right framework for every part of a model's lifetime
 - Train state-of-the-art models in 3 lines of code
@@ -56,7 +56,7 @@ Choose the right framework for every part of a model's lifetime
 | [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: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
-| [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
+| [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-transformers to transformers |
 | [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
 | [Documentation](https://huggingface.co/transformers/) | Full API documentation and more |
 
@@ -105,13 +105,13 @@ python -m pytest -sv ./examples/
 
 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 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 8 NLU/NLG architectures:
+🤗 Transformers currently provides 10 NLU/NLG architectures:
 
 1. **[BERT](https://github.com/google-research/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
 2. **[GPT](https://github.com/openai/finetune-transformer-lm)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
@@ -121,6 +121,7 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 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/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation).
+9. **[CTRL](https://github.com/salesforce/ctrl/)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
 
 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).
 
@@ -147,6 +148,7 @@ from transformers import *
 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'),
@@ -174,10 +176,11 @@ BERT_MODEL_CLASSES = [BertModel, BertForPreTraining, BertForMaskedLM, BertForNex
 # 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,
@@ -240,8 +243,9 @@ sentence_2 = "His findings were not compatible with this research."
 inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
 inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
 
-pred_1 = pytorch_model(**inputs_1)[0].argmax().item()
-pred_2 = pytorch_model(**inputs_2)[0].argmax().item()
+pred_1 = pytorch_model(inputs_1['input_ids'], token_type_ids=inputs_1['token_type_ids'])[0].argmax().item()
+pred_2 = pytorch_model(inputs_2['input_ids'], token_type_ids=inputs_2['token_type_ids'])[0].argmax().item()
+
 print("sentence_1 is", "a paraphrase" if pred_1 else "not a paraphrase", "of sentence_0")
 print("sentence_2 is", "a paraphrase" if pred_2 else "not a paraphrase", "of sentence_0")
 ```
@@ -252,7 +256,7 @@ The library comprises several example scripts with SOTA performances for NLU and
 
 - `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:
@@ -390,7 +394,7 @@ 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 Aman Rusia to get high-quality generation with memory models like Transformer-XL and XLNet (include a predefined text to make short inputs longer).
@@ -404,6 +408,16 @@ python ./examples/run_generation.py \
     --model_name_or_path=gpt2 \
 ```
 
+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 \
+```
+
 ## 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`.
@@ -533,4 +547,13 @@ for batch in train_data:
 
 ## Citation
 
-At the moment, there is no paper associated with Transformers but we are working on preparing one. In the meantime, please include a mention of the library and a link to the present repository if you use this work in a published or open-source project.
+We now have a paper you can cite for the 🤗 Transformers library:
+```
+@article{Wolf2019HuggingFacesTS,
+  title={HuggingFace's Transformers: State-of-the-art Natural Language Processing},
+  author={Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and R'emi Louf and Morgan Funtowicz and Jamie Brew},
+  journal={ArXiv},
+  year={2019},
+  volume={abs/1910.03771}
+}
+```

docs/README.md

@@ -50,7 +50,7 @@ make html
 ---
 **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/source/benchmarks.md

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

docs/source/conf.py

@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'2.0.0'
+release = u'2.1.1'
 
 
 # -- General configuration ---------------------------------------------------

docs/source/index.rst

@@ -62,6 +62,8 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
     migration
     bertology
     torchscript
+    multilingual
+    benchmarks
 
 .. toctree::
     :maxdepth: 2
@@ -86,3 +88,4 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
     model_doc/xlnet
     model_doc/roberta
     model_doc/distilbert
+    model_doc/ctrl

docs/source/installation.rst → docs/source/installation.md

-Installation
-================================================
+# Installation
 
 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
-^^^^^^^^
+## With pip
 
 PyTorch Transformers can be installed using pip as follows:
 
-.. code-block:: bash
+``` bash
+pip install transformers
+```
 
-   pip install transformers
-
-From source
-^^^^^^^^^^^
+## From source
 
 To install from source, clone the repository and install with:
 
-.. code-block:: bash
-
-    git clone https://github.com/huggingface/transformers.git
-    cd transformers
-    pip install [--editable] .
-
+``` bash
+git clone https://github.com/huggingface/transformers.git
+cd transformers
+pip install [--editable] .
+```
 
-Tests
-^^^^^
+## 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/transformers/tests>`_ and examples tests in the `examples folder <https://github.com/huggingface/transformers/tree/master/examples>`_.
+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/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/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 ./transformers/tests/
-    python -m pytest -sv ./examples/
-
+``` bash
+python -m pytest -sv ./transformers/tests/
+python -m pytest -sv ./examples/
+```
 
-OpenAI GPT original tokenization workflow
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+## 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`` :
+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
+``` bash
+pip install spacy ftfy==4.4.3
+python -m spacy download en
+```
 
-   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).
 
-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)
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+## 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?
 
-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.
+You should check out our [swift-coreml-transformers](https://github.com/huggingface/swift-coreml-transformers) repo.
 
-It also contains an implementation of BERT for Question answering.
+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/model_doc/ctrl.rst

+CTRL
+----------------------------------------------------
+
+``CTRLConfig``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CTRLConfig
+    :members:
+
+
+``CTRLTokenizer``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CTRLTokenizer
+    :members:
+
+
+``CTRLModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CTRLModel
+    :members:
+
+
+``CTRLLMHeadModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CTRLLMHeadModel
+    :members:
+
+
+``TFCTRLModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFCTRLModel
+    :members:
+
+
+``TFCTRLLMHeadModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFCTRLLMHeadModel
+    :members:
+

docs/source/multilingual.rst

+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::
+
+    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::
+
+    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::
+
+    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::
+
+    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.
\ No newline at end of file

docs/source/pretrained_models.rst

@@ -53,6 +53,14 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``bert-base-cased-finetuned-mrpc``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | The ``bert-base-cased`` model fine-tuned on MRPC                                                                                    |
 |                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/transformers/examples.html>`__)                           |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-german-dbmdz-cased``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on cased German text by DBMDZ                                                                                               |
+|                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-german-dbmdz-uncased``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on uncased German text by DBMDZ                                                                                             |
+|                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | GPT               | ``openai-gpt``                                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | OpenAI GPT English model                                                                                                            |
@@ -128,5 +136,13 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``distilgpt2``                                             | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
 |                   |                                                            | | The DistilGPT2 model distilled from the GPT2 model `gpt2` checkpoint.                                                               |
 |                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilroberta-base``                                     | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
+|                   |                                                            | | The DistilRoBERTa model distilled from the RoBERTa model `roberta-base` checkpoint.                                                 |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| CTRL              | ``ctrl``                                                   | | 48-layer, 1280-hidden, 16-heads, 1.6B parameters                                                                                    |
+|                   |                                                            | | Salesforce's Large-sized CTRL English model                                                                                         |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+
 .. <https://huggingface.co/transformers/examples.html>`__
\ No newline at end of file

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

examples/README.md

@@ -5,13 +5,37 @@ similar API between the different models.
 
 | Section                    | Description                                                                                                                                                |
 |----------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| [TensorFlow 2.0 models on GLUE](#TensorFlow-2.0-Bert-models-on-GLUE) | Examples running BERT TensorFlow 2.0 model on the GLUE tasks. 
 | [Language Model fine-tuning](#language-model-fine-tuning) | Fine-tuning the library models for language modeling on a text dataset. Causal language modeling for GPT/GPT-2, masked language modeling for BERT/RoBERTa. |
 | [Language Generation](#language-generation) | Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet.                                         |
 | [GLUE](#glue) | Examples running BERT/XLM/XLNet/RoBERTa on the 9 GLUE tasks. Examples feature distributed training as well as half-precision.                              |
-| [SQuAD](#squad) | Using BERT for question answering, examples with distributed training.                                                                                  |
-| [Multiple Choice](#multiple choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks. 
+| [SQuAD](#squad) | Using BERT/XLM/XLNet/RoBERTa for question answering, examples with distributed training.                                                                                  |
+| [Multiple Choice](#multiple-choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks. 
+| [Named Entity Recognition](#named-entity-recognition) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training.                                                                                  |
 | [Abstractive summarization](#abstractive-summarization) | Fine-tuning the library models for abstractive summarization tasks on the CNN/Daily Mail dataset. |
 
+## TensorFlow 2.0 Bert models on GLUE
+
+Based on the script [`run_tf_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/run_tf_glue.py).
+
+Fine-tuning the library TensorFlow 2.0 Bert model for sequence classification on the  MRPC task of the GLUE benchmark: [General Language Understanding Evaluation](https://gluebenchmark.com/).
+
+This script has an option for mixed precision (Automatic Mixed Precision / AMP) to run models on Tensor Cores (NVIDIA Volta/Turing GPUs) and future hardware and an option for XLA, which uses the XLA compiler to reduce model runtime.
+Options are toggled using `USE_XLA` or `USE_AMP` variables in the script.
+These options and the below benchmark are provided by @tlkh.
+
+Quick benchmarks from the script (no other modifications):
+
+| GPU    | Mode | Time (2nd epoch) | Val Acc (3 runs) |
+| --------- | -------- | ----------------------- | ----------------------|
+| Titan V | FP32 | 41s | 0.8438/0.8281/0.8333 |
+| Titan V | AMP | 26s | 0.8281/0.8568/0.8411 |
+| V100    | FP32 | 35s | 0.8646/0.8359/0.8464 |
+| V100    | AMP | 22s | 0.8646/0.8385/0.8411 |
+| 1080 Ti | FP32 | 55s | - | 
+
+Mixed precision (AMP) reduces the training time considerably for the same hardware and hyper-parameters (same batch size was used).
+
 ## Language model fine-tuning
 
 Based on the script [`run_lm_finetuning.py`](https://github.com/huggingface/transformers/blob/master/examples/run_lm_finetuning.py).
@@ -78,7 +102,7 @@ python run_lm_finetuning.py \
 
 Based on the script [`run_generation.py`](https://github.com/huggingface/transformers/blob/master/examples/run_generation.py).
 
-Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet.
+Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL, XLNet, CTRL.
 A similar script is used for our official demo [Write With Transfomer](https://transformer.huggingface.co), where you
 can try out the different models available in the library.
 
@@ -284,17 +308,17 @@ The results  are the following:
   loss = 0.04755385363816904
 ```
 
-##Multiple Choice
+## 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/single_model_scripts/run_multiple_choice.py \
+python ./examples/run_multiple_choice.py \
 --model_type roberta \
 --task_name swag \
 --model_name_or_path roberta-base \
@@ -391,6 +415,107 @@ exact_match = 86.91
 This fine-tuned model is available as a checkpoint under the reference
 `bert-large-uncased-whole-word-masking-finetuned-squad`.
 
+## Named Entity Recognition
+
+Based on the script [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py).
+This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
+Details and results for the fine-tuning provided by @stefan-it.
+
+### Data (Download and pre-processing steps)
+
+Data can be obtained from the [GermEval 2014](https://sites.google.com/site/germeval2014ner/data) shared task page.
+
+Here are the commands for downloading and pre-processing train, dev and test datasets. The original data format has four (tab-separated) columns, in a pre-processing step only the two relevant columns (token and outer span NER annotation) are extracted:
+
+```bash
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-train.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > train.txt.tmp
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-dev.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > dev.txt.tmp
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-test.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > test.txt.tmp
+```
+
+The GermEval 2014 dataset contains some strange "control character" tokens like `'\x96', '\u200e', '\x95', '\xad' or '\x80'`. One problem with these tokens is, that `BertTokenizer` returns an empty token for them, resulting in misaligned `InputExample`s. I wrote a script that a) filters these tokens and b) splits longer sentences into smaller ones (once the max. subtoken length is reached).
+
+```bash
+wget "https://raw.githubusercontent.com/stefan-it/fine-tuned-berts-seq/master/scripts/preprocess.py"
+```
+Let's define some variables that we need for further pre-processing steps and training the model:
+
+```bash
+export MAX_LENGTH=128
+export BERT_MODEL=bert-base-multilingual-cased
+```
+
+Run the pre-processing script on training, dev and test datasets:
+
+```bash
+python3 preprocess.py train.txt.tmp $BERT_MODEL $MAX_LENGTH > train.txt
+python3 preprocess.py dev.txt.tmp $BERT_MODEL $MAX_LENGTH > dev.txt
+python3 preprocess.py test.txt.tmp $BERT_MODEL $MAX_LENGTH > test.txt
+```
+
+The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so an own set of labels must be used:
+
+```bash
+cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
+```
+
+### Training
+
+Additional environment variables must be set:
+
+```bash
+export OUTPUT_DIR=germeval-model
+export BATCH_SIZE=32
+export NUM_EPOCHS=3
+export SAVE_STEPS=750
+export SEED=1
+```
+
+To start training, just run:
+
+```bash
+python3 run_ner.py --data_dir ./ \
+--model_type bert \
+--labels ./labels.txt \
+--model_name_or_path $BERT_MODEL \
+--output_dir $OUTPUT_DIR \
+--max_seq_length  $MAX_LENGTH \
+--num_train_epochs $NUM_EPOCHS \
+--per_gpu_train_batch_size $BATCH_SIZE \
+--save_steps $SAVE_STEPS \
+--seed $SEED \
+--do_train \
+--do_eval \
+--do_predict
+```
+
+If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
+
+### Evaluation
+
+Evaluation on development dataset outputs the following for our example:
+
+```bash
+10/04/2019 00:42:06 - INFO - __main__ -   ***** Eval results  *****
+10/04/2019 00:42:06 - INFO - __main__ -     f1 = 0.8623348017621146
+10/04/2019 00:42:06 - INFO - __main__ -     loss = 0.07183869666975543
+10/04/2019 00:42:06 - INFO - __main__ -     precision = 0.8467916366258111
+10/04/2019 00:42:06 - INFO - __main__ -     recall = 0.8784592370979806
+```
+
+On the test dataset the following results could be achieved:
+
+```bash
+10/04/2019 00:42:42 - INFO - __main__ -   ***** Eval results  *****
+10/04/2019 00:42:42 - INFO - __main__ -     f1 = 0.8614389652384803
+10/04/2019 00:42:42 - INFO - __main__ -     loss = 0.07064602487454782
+10/04/2019 00:42:42 - INFO - __main__ -     precision = 0.8604651162790697
+10/04/2019 00:42:42 - INFO - __main__ -     recall = 0.8624150210424085
+```
+
 ## Abstractive summarization
 
 Based on the script

examples/contrib/run_swag.py

@@ -31,9 +31,13 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                               TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm, trange
 
-from tensorboardX import SummaryWriter
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
+from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
                                   BertForMultipleChoice, BertTokenizer)

examples/distillation/README.md

 # Distil*
 
-This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT and DistilGPT2.
+This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT, DistilRoBERTa and DistilGPT2.
 
-**2019, October 3rd - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2.
+**October 23rd, 2019 - Update** We release **DistilRoBERTa**: 95% of `RoBERTa-base`'s performance on GLUE, twice as fast as RoBERTa while being 35% smaller.
+
+**October 3rd, 2019 - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2. **The paper superseeds our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances. Please use the paper as a reference when comparing/reporting results on DistilBERT.**
+
+**September 19th, 2019 - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 97% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
 
-**2019, September 19th - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 97% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
 
 ## What is Distil*
 
 Distil* is a class of compressed models that started with DistilBERT. DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
 
-We have applied the same method to GPT2 and release the weights of the compressed model. On the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test of 15.8 compared to 19.3 for DistilGPT2 (after fine-tuning on the train set).
+We have applied the same method to other Transformer architectures and released the weights:
+- GPT2: on the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 15.0 compared to 18.5 for **DistilGPT2** (after fine-tuning on the train set).
+- RoBERTa: **DistilRoBERTa** reaches 95% of `RoBERTa-base` performance on GLUE while being twice faster and 35% smaller.
+- and more to come! 🤗🤗🤗
 
-For more information on DistilBERT, please refer to our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108). The paper superseeds our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances.
+For more information on DistilBERT, please refer to our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108).
 
 Here are the results on the dev sets of GLUE:
 
 | Model                     | Macro-score                    | CoLA | MNLI | MRPC | QNLI | QQP  | RTE  | SST-2| STS-B| WNLI              |
-| :---:      |    :---:    | :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:|
+| :---:                     |    :---:                       | :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:             |
 | BERT-base                 |  **77.6**                      | 48.9 | 84.3 | 88.6 | 89.3 | 89.5 | 71.3 | 91.7 | 91.2 | 43.7              |
 | DistilBERT                |  **76.8**                      | 49.1 | 81.8 | 90.2 | 90.2 | 89.2 | 62.9 | 92.7 | 90.7 | 44.4              |
+| ---                       |    ---                         |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  ---              |
+| RoBERTa-base (reported)   |  **83.2**/**86.4**<sup>2</sup> | 63.6 | 87.6 | 90.2 | 92.8 | 91.9 | 78.7 | 94.8 | 91.2 | 57.7<sup>3</sup> |
+| DistilRoBERTa<sup>1</sup> |  **79.0**/**82.3**<sup>2</sup> | 59.4 | 83.9 | 86.6 | 90.8 | 89.4 | 67.9 | 92.5 | 88.3 | 52.1              |
+
+<sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directy perform transfer learning on the pre-trained DistilRoBERTa.
+
+<sup>2</sup> Macro-score computed without WNLI.
+
+<sup>3</sup> We compute this score ourselves for completeness.
 
 ## Setup
 
 This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`. 
 
-**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0). It is important to note that there is a small internal bug in the current version of PyTorch available on pip that causes a memory leak in our training/distillation. It has been recently fixed and will likely be integrated into the next release. For the moment, we recommend to [compile PyTorch from source](https://github.com/pytorch/pytorch#from-source). Please refer to [issue 1179](https://github.com/huggingface/transformers/issues/1179) for more details.
+**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0).
+
 
 ## How to use DistilBERT
 
@@ -33,7 +49,8 @@ Transformers includes two pre-trained Distil* models, currently only provided fo
 
 - `distilbert-base-uncased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-uncased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 66M parameters.
 - `distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.9 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
-- `distilgpt2`: DistilGPT2 English language model pretrained with the supervision of `gpt2` (the smallest version of GPT2) on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset and . The model has 6 layers, 768 dimension and 12 heads, totalizing 82M (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
+- `distilgpt2`: DistilGPT2 English language model pretrained with the supervision of `gpt2` (the smallest version of GPT2) on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset. The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
+- `distilroberta-base`: DistilRoBERTa English language model pretrained with the supervision of `roberta-base` solely on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset (it is ~4 times less training data than the teacher RoBERTa). The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 125M parameters for RoBERTa-base). On average DistilRoBERTa is twice as fast as Roberta-base.
 - and more to come! 🤗🤗🤗
 
 Using DistilBERT is very similar to using BERT. DistilBERT share the same tokenizer as BERT's `bert-base-uncased` even though we provide a link to this tokenizer under the `DistilBertTokenizer` name to have a consistent naming between the library models.
@@ -47,7 +64,10 @@ outputs = model(input_ids)
 last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 ```
 
-Similarly, using DistilGPT2 simply consists in calling the GPT2 classes from a different pretrained checkpoint: `model = GPT2Model.from_pretrained('distilgpt2')`.
+Similarly, using the other Distil* models simply consists in calling the base classes with a different pretrained checkpoint:
+- DistilGPT2: `model = GPT2Model.from_pretrained('distilgpt2')`
+- DistilRoBERTa: `model = RobertaModel.from_pretrained('distilroberta-base')`
+
 
 ## How to train Distil*
 
@@ -134,3 +154,16 @@ python -m torch.distributed.launch \
 **Tips:** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract.py` and `scripts/extract_distilbert.py` to create a valid initialization checkpoint and use `--student_pretrained_weights` argument to use this initialization for the distilled training!
 
 Happy distillation!
+
+## Citation
+
+If you find the ressource useful, you should cite the following paper:
+
+```
+@inproceedings{sanh2019distilbert,
+  title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
+  author={Sanh, Victor and Debut, Lysandre and Chaumond, Julien and Wolf, Thomas},
+  booktitle={NeurIPS EMC^2 Workshop},
+  year={2019}
+}
+```
\ No newline at end of file

examples/distillation/distiller.py

@@ -19,7 +19,6 @@ import os
 import math
 import psutil
 import time
-from tensorboardX import SummaryWriter
 from tqdm import trange, tqdm
 import numpy as np
 import psutil
@@ -31,6 +30,11 @@ from torch.optim import AdamW
 from torch.utils.data.distributed import DistributedSampler
 from torch.utils.data import RandomSampler, BatchSampler, DataLoader
 
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
 from transformers import WarmupLinearSchedule
 
 from utils import logger

examples/distillation/run_squad_w_distillation.py

@@ -30,9 +30,13 @@ from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
 from torch.utils.data.distributed import DistributedSampler
 import torch.nn.functional as F
 import torch.nn as nn
-from tqdm import tqdm, trange
 
-from tensorboardX import SummaryWriter
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
+from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
                                   BertForQuestionAnswering, BertTokenizer,