switch to Thom script

Learning rate and perdiction
Beginning of GLUE script
2019-09-25 09:38:37 -04:00 · 2019-09-25 09:38:37 -04:00 · 2019-09-25 09:38:37 -04:00
223 changed files with 2536 additions and 19096 deletions
@@ -1,7 +1,7 @@
 version: 2
 jobs:
    build_py3_torch_and_tf:
-        working_directory: ~/transformers
+        working_directory: ~/pytorch-transformers
        docker:
            - image: circleci/python:3.5
        resource_class: xlarge
@@ -9,14 +9,14 @@ jobs:
        steps:
            - checkout
            - run: sudo pip install torch
-            - run: sudo pip install tensorflow
+            - run: sudo pip install tensorflow==2.0.0-rc0
            - 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 ./transformers/tests/ --cov
+            - run: python -m pytest -sv ./pytorch_transformers/tests/ --cov
            - run: codecov
    build_py3_torch:
-        working_directory: ~/transformers
+        working_directory: ~/pytorch-transformers
        docker:
            - image: circleci/python:3.5
        resource_class: xlarge
@@ -27,25 +27,25 @@ jobs:
            - 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 ./transformers/tests/ --cov
+            - run: python -m pytest -sv ./pytorch_transformers/tests/ --cov
            - run: python -m pytest -sv ./examples/
            - run: codecov
    build_py3_tf:
-        working_directory: ~/transformers
+        working_directory: ~/pytorch-transformers
        docker:
            - image: circleci/python:3.5
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install tensorflow
+            - run: sudo pip install tensorflow==2.0.0-rc0
            - 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 ./transformers/tests/ --cov
+            - run: python -m pytest -sv ./pytorch_transformers/tests/ --cov
            - run: codecov
    build_py2_torch:
-        working_directory: ~/transformers
+        working_directory: ~/pytorch-transformers
        resource_class: large
        parallelism: 1
        docker:
@@ -55,23 +55,23 @@ jobs:
            - run: sudo pip install torch
            - run: sudo pip install --progress-bar off .
            - run: sudo pip install pytest codecov pytest-cov
-            - run: python -m pytest -sv ./transformers/tests/ --cov
+            - run: python -m pytest -sv ./pytorch_transformers/tests/ --cov
            - run: codecov
    build_py2_tf:
-        working_directory: ~/transformers
+        working_directory: ~/pytorch-transformers
        resource_class: large
        parallelism: 1
        docker:
            - image: circleci/python:2.7
        steps:
            - checkout
-            - run: sudo pip install tensorflow
+            - run: sudo pip install tensorflow==2.0.0-rc0
            - run: sudo pip install --progress-bar off .
            - run: sudo pip install pytest codecov pytest-cov
-            - run: python -m pytest -sv ./transformers/tests/ --cov
+            - run: python -m pytest -sv ./pytorch_transformers/tests/ --cov
            - run: codecov
    deploy_doc:
-        working_directory: ~/transformers
+        working_directory: ~/pytorch-transformers
        docker:
            - image: circleci/python:3.5
        steps:
@@ -81,7 +81,8 @@ jobs:
            - checkout
            - run: sudo pip install --progress-bar off -r docs/requirements.txt
            - run: sudo pip install --progress-bar off -r requirements.txt
-            - run: ./.circleci/deploy.sh
+            - run: cd docs/source && ln -s ../../examples/README.md examples.md && cd -
+            - run: cd docs && make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
 workflow_filters: &workflow_filters
    filters:
        branches:
@@ -96,4 +97,4 @@ workflows:
            - build_py3_tf
            - build_py2_torch
            - build_py2_tf
-            - deploy_doc: *workflow_filters
+            - deploy_doc: *workflow_filters
@@ -1,26 +0,0 @@
-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
@@ -1,5 +1,5 @@
 [run]
-source=transformers
+source=pytorch_transformers
 omit =
    # skip convertion scripts from testing for now
    */convert_*
@@ -1,22 +0,0 @@
---
-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!
@@ -1,24 +0,0 @@
---
-name: "\U0001F31FNew 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)
-
-## Additional context
-
-<!-- Add any other context about the problem here. -->
@@ -1,10 +1,6 @@
 ---
 name: "\U0001F41B Bug Report"
 about: Submit a bug report to help us improve PyTorch Transformers
-title: ''
-labels: ''
-assignees: ''
-
 ---

 ## 🐛 Bug
@@ -49,4 +45,4 @@ Steps to reproduce the behavior:

 ## Additional context

-<!-- Add any other context about the problem here. -->
+<!-- Add any other context about the problem here. -->
@@ -1,10 +1,6 @@
 ---
 name: "\U0001F680 Feature Request"
 about: Submit a proposal/request for a new PyTorch Transformers feature
-title: ''
-labels: ''
-assignees: ''
-
 ---

 ## 🚀 Feature
@@ -17,4 +13,4 @@ assignees: ''

 ## Additional context

-<!-- Add any other context or screenshots about the feature request here. -->
+<!-- Add any other context or screenshots about the feature request here. -->
@@ -1,10 +1,6 @@
 ---
 name: "\U0001F4DA Migration from PyTorch-pretrained-Bert"
-about: Report a problem when migrating from PyTorch-pretrained-Bert to Transformers
-title: ''
-labels: ''
-assignees: ''
-
+about: Report a problem when migrating from PyTorch-pretrained-Bert to PyTorch-Transformers
 ---

 ## 📚 Migration
@@ -44,4 +40,4 @@ Details of the issue:

 ## Additional context

-<!-- Add any other context about the problem here. -->
+<!-- Add any other context about the problem here. -->
@@ -1,12 +1,8 @@
 ---
 name: "❓Questions & Help"
 about: Start a general discussion related to PyTorch Transformers
-title: ''
-labels: ''
-assignees: ''
-
 ---

 ## ❓ Questions & Help

-<!-- A clear and concise description of the question. -->
+<!-- A clear and concise description of the question. -->
@@ -118,9 +118,6 @@ dmypy.json
 # vscode
 .vscode

-# Pycharm
-.idea
-
 # TF code
 tensorflow_code

@@ -134,8 +131,4 @@ examples/runs

 # data
 /data
-serialization_dir
-
-# emacs
-*.*~
-debug.env
+serialization_dir
@@ -1,179 +0,0 @@
-# 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.
-
-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 -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)
@@ -1,128 +1,10 @@
-<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>
+# 👾 PyTorch-Transformers

-<h3 align="center">
-<p>State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch
-</h3>
+[![CircleCI](https://circleci.com/gh/huggingface/pytorch-transformers.svg?style=svg)](https://circleci.com/gh/huggingface/pytorch-transformers)

-🤗 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.
+PyTorch-Transformers (formerly known as `pytorch-pretrained-bert`) is a library of state-of-the-art pre-trained models for Natural Language Processing (NLP).

-### 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: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
-| [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-transformers to transformers |
-| [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
-| [Documentation][(v2.2.0)](https://huggingface.co/transformers/v2.2.0) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) [(master)](https://huggingface.co/transformers) | Full API documentation and more |
-
-## Installation
-
-This repo is tested on Python 2.7 and 3.5+ (examples are tested only on python 3.5+), PyTorch 1.0.0+ and TensorFlow 2.0.0-rc1
-
-### 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
-pip install [--editable] .
-```
-
-### Run the examples
-
-Examples are included in the repository but are not shipped with the library.
-Therefore, in order to run the latest versions of the examples you also need to install from source. To do so, create a new virtual environment and follow these steps:
-
-```bash
-git clone https://github.com/huggingface/transformers
-cd transformers
-pip install [--editable] .
-```
-
-### Tests
-
-A series of tests are included for the library and the example scripts. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples).
-
-These tests can be run using `pytest` (install pytest if needed with `pip install pytest`).
-
-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.
-
-You can run the tests from the root of the cloned repository with the commands:
-
-```bash
-python -m pytest -sv ./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 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 10 NLU/NLG architectures:
+The library currently contains PyTorch 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.
@@ -131,13 +13,61 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 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/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation).
-9. **[CTRL](https://github.com/salesforce/ctrl/)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
-10. **[CamemBERT](https://camembert-model.fr)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
-11. **[ALBERT](https://github.com/google-research/google-research/tree/master/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
-11. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
+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.

-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).
+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).
+
+| Section | Description |
+|-|-|
+| [Installation](#installation) | How to install the package |
+| [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: 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 |
+
+## Installation
+
+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!

 ## Online demo

@@ -150,26 +80,22 @@ You can use it to experiment with completions generated by `GPT2Model`, `Transfo

 ## Quick tour

-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/).
+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/).

 ```python
 import torch
-from transformers import *
+from pytorch_transformers import *

-# Transformers has a unified API
-# for 8 transformer architectures and 30 pretrained weights.
+# PyTorch-Transformers has a unified API
+# for 7 transformer architectures and 30 pretrained weights.
 #          Model          | Tokenizer          | Pretrained weights shortcut
-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-uncased'),
-          (RobertaModel,    RobertaTokenizer,    '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`
+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')]

 # Let's encode some text in a sequence of hidden-states using each model:
 for model_class, tokenizer_class, pretrained_weights in MODELS:
@@ -184,97 +110,44 @@ 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, BertForTokenClassification, BertForQuestionAnswering]
+                      BertForSequenceClassification, BertForMultipleChoice, 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
-pretrained_weights = 'bert-base-uncased'
-tokenizer = BertTokenizer.from_pretrained(pretrained_weights)
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
 for model_class in BERT_MODEL_CLASSES:
    # Load pretrained model/tokenizer
-    model = model_class.from_pretrained(pretrained_weights)
+    model = model_class.from_pretrained('bert-base-uncased')

-    # 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 = BertTokenizer.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 = tokenizer_class.from_pretrained('./directory/to/save/')  # re-load

-    # 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")
+# SOTA examples for GLUE, SQUAD, text generation...
 ```

 ## 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, CTRL, Transformer-XL and XLNet for conditional language generation
+- `run_generation.py`: an example using GPT, GPT-2, Transformer-XL and XLNet for conditional language generation
 - other model-specific examples (see the documentation).

 Here are three quick usage examples for these scripts:
@@ -412,10 +285,10 @@ 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, CTRL, Transformer-XL and XLNet
+### `run_generation.py`: Text generation with GPT, GPT-2, 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).
+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:

@@ -426,42 +299,19 @@ 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-pretrained-bert to pytorch-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`.
+Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to `pytorch-transformers`

 ### Models always output `tuples`

-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 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 exact content of the tuples for each model is detailed in the models' docstrings and the [documentation](https://huggingface.co/transformers/).
+The exact content of the tuples for each model are detailed in the models' docstrings and the [documentation](https://huggingface.co/pytorch-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 `transformers` conversion example for a `BertForSequenceClassification` classification model:
+Here is a `pytorch-pretrained-bert` to `pytorch-transformers` conversion example for a `BertForSequenceClassification` classification model:

 ```python
 # Let's load our model
@@ -470,11 +320,11 @@ 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 transformers to extract the loss from the output tuple:
+# Now just use this line in pytorch-transformers to extract the loss from the output tuple:
 outputs = model(input_ids, labels=labels)
 loss = outputs[0]

-# In transformers you can also have access to the logits:
+# In pytorch-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)
@@ -483,17 +333,13 @@ 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 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.
+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.

 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.

@@ -538,24 +384,23 @@ Here is a conversion examples from `BertAdam` with a linear warmup and decay sch
 # Parameters:
 lr = 1e-3
 max_grad_norm = 1.0
-num_training_steps = 1000
+num_total_steps = 1000
 num_warmup_steps = 100
-warmup_proportion = float(num_warmup_steps) / float(num_training_steps)  # 0.1
+warmup_proportion = float(num_warmup_steps) / float(num_total_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_training_steps)
+optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
 ### and used like this:
 for batch in train_data:
    loss = model(batch)
    loss.backward()
    optimizer.step()

-### In Transformers, optimizer and schedules are splitted and instantiated like this:
+### In PyTorch-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 = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)  # PyTorch scheduler
+scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_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)
@@ -566,13 +411,4 @@ for batch in train_data:

 ## Citation

-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}
-}
-```
+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.
@@ -1,22 +0,0 @@
-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
@@ -2,6 +2,6 @@ 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 transformers
+RUN pip install pytorch_transformers

 WORKDIR /workspace
@@ -34,11 +34,11 @@ 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 
+Make sure that there is a symlink from the `example` file (in /examples) inside the source folder. Run the followig 
 command to generate it:

 ```bash
-ln -s ../../examples/README.md examples.md
+ln -s ../../examples/README.md source/examples.md
 ```

 Once you have setup `sphinx`, you can build the documentation by running the following command in the `/docs` folder:
@@ -50,7 +50,7 @@ make html
 ---
 **NOTE**

-If you are adding/removing elements from the toc-tree or from any structural item, it is recommended to clean the build
+If you are adding/removing elements from the toc-tree or from any strutural item, it is recommended to clean the build
 directory before rebuilding. Run the following command to clean and build:

 ```bash
@@ -26,7 +26,4 @@ sphinxcontrib-jsmath==1.0.1
 sphinxcontrib-qthelp==1.0.2
 sphinxcontrib-serializinghtml==1.1.3
 urllib3==1.25.3
-sphinx-markdown-tables==0.0.9
-numpy==1.17.2
-tensorflow==2.0.0rc2
-torch==1.2.0
+sphinx-markdown-tables==0.0.9
@@ -1,3 +1,5 @@
+huggingface.css
+
 /* The literal code blocks */
 .rst-content tt.literal, .rst-content tt.literal, .rst-content code.literal {
    color: #6670FF;
@@ -42,11 +44,11 @@
 /* The text items on the toc tree */
 .wy-menu-vertical a {
    color: #FFFFDD;
-    font-family: Calibre-Light, sans-serif;
+    font-family: Calibre-Light;
 }
 .wy-menu-vertical header, .wy-menu-vertical p.caption{
    color: white;
-    font-family: Calibre-Light, sans-serif;
+    font-family: Calibre-Light;
 }

 /* The color inside the selected toc tree block */
@@ -83,7 +85,7 @@ a {
    border-right: solid 2px #FB8D68;
    border-left: solid 2px #FB8D68;
    color: #FB8D68;
-    font-family: Calibre-Light, sans-serif;
+    font-family: Calibre-Light;
    border-top: none;
    font-style: normal !important;
 }
@@ -134,14 +136,14 @@ 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, sans-serif;
+    font-family: Calibre;
    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, sans-serif;
+    font-family: Calibre-Medium;
 }

 /* Method and class parameters */
@@ -158,17 +160,17 @@ a {

 /* FONTS */
 body{
-    font-family: Calibre, sans-serif;
+    font-family: Calibre;
    font-size: 16px;
 }

 h1 {
-    font-family: Calibre-Thin, sans-serif;
+    font-family: Calibre-Thin;
    font-size: 70px;
 }

 h2, .rst-content .toctree-wrapper p.caption, h3, h4, h5, h6, legend{
-    font-family: Calibre-Medium, sans-serif;
+    font-family: Calibre-Medium;
 }

@font-face {
@@ -194,3 +196,4 @@ h2, .rst-content .toctree-wrapper p.caption, h3, h4, h5, h6, legend{
    src: url(./Calibre-Thin.otf);
    font-weight:400;
 }
+
@@ -1,54 +0,0 @@
-# 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).
@@ -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/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/pytorch-transformers/blob/master/examples/run_bertology.py>`_ while extract information and prune a model pre-trained on GLUE.
@@ -19,14 +19,14 @@ sys.path.insert(0, os.path.abspath('../..'))

 # -- Project information -----------------------------------------------------

-project = u'transformers'
+project = u'pytorch-transformers'
 copyright = u'2019, huggingface'
 author = u'huggingface'

 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'2.2.0'
+release = u'1.2.0'


 # -- General configuration ---------------------------------------------------
@@ -109,7 +109,7 @@ html_static_path = ['_static']
 # -- Options for HTMLHelp output ---------------------------------------------

 # Output file base name for HTML help builder.
-htmlhelp_basename = 'transformersdoc'
+htmlhelp_basename = 'pytorch-transformersdoc'


 # -- Options for LaTeX output ------------------------------------------------
@@ -136,7 +136,7 @@ latex_elements = {
 # (source start file, target name, title,
 #  author, documentclass [howto, manual, or own class]).
 latex_documents = [
-    (master_doc, 'transformers.tex', u'transformers Documentation',
+    (master_doc, 'pytorch-transformers.tex', u'pytorch-transformers Documentation',
     u'huggingface', 'manual'),
 ]

@@ -146,7 +146,7 @@ latex_documents = [
 # One entry per manual page. List of tuples
 # (source start file, name, description, authors, manual section).
 man_pages = [
-    (master_doc, 'transformers', u'transformers Documentation',
+    (master_doc, 'pytorch-transformers', u'pytorch-transformers Documentation',
     [author], 1)
 ]

@@ -157,8 +157,8 @@ man_pages = [
 # (source start file, target name, title, author,
 #  dir menu entry, description, category)
 texinfo_documents = [
-    (master_doc, 'transformers', u'transformers Documentation',
-     author, 'transformers', 'One line description of project.',
+    (master_doc, 'pytorch-transformers', u'pytorch-transformers Documentation',
+     author, 'pytorch-transformers', 'One line description of project.',
     'Miscellaneous'),
 ]

@@ -6,7 +6,7 @@ A command-line interface is provided to convert original Bert/GPT/GPT-2/Transfor
 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/transformers/blob/master/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/pytorch-transformers/blob/master/pytorch_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>`_\ ).

@@ -20,7 +20,7 @@ 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

-   transformers bert \
+   pytorch_transformers bert \
     $BERT_BASE_DIR/bert_model.ckpt \
     $BERT_BASE_DIR/bert_config.json \
     $BERT_BASE_DIR/pytorch_model.bin
@@ -36,7 +36,7 @@ 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

-   transformers gpt \
+   pytorch_transformers gpt \
     $OPENAI_GPT_CHECKPOINT_FOLDER_PATH \
     $PYTORCH_DUMP_OUTPUT \
     [OPENAI_GPT_CONFIG]
@@ -50,7 +50,7 @@ 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

-   transformers gpt2 \
+   pytorch_transformers gpt2 \
     $OPENAI_GPT2_CHECKPOINT_PATH \
     $PYTORCH_DUMP_OUTPUT \
     [OPENAI_GPT2_CONFIG]
@@ -64,7 +64,7 @@ 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

-   transformers transfo_xl \
+   pytorch_transformers transfo_xl \
     $TRANSFO_XL_CHECKPOINT_FOLDER_PATH \
     $PYTORCH_DUMP_OUTPUT \
     [TRANSFO_XL_CONFIG]
@@ -80,7 +80,7 @@ Here is an example of the conversion process for a pre-trained XLNet model, fine
   export TRANSFO_XL_CHECKPOINT_PATH=/path/to/xlnet/checkpoint
   export TRANSFO_XL_CONFIG_PATH=/path/to/xlnet/config

-   transformers xlnet \
+   pytorch_transformers xlnet \
     $TRANSFO_XL_CHECKPOINT_PATH \
     $TRANSFO_XL_CONFIG_PATH \
     $PYTORCH_DUMP_OUTPUT \
@@ -96,6 +96,6 @@ Here is an example of the conversion process for a pre-trained XLM model:

   export XLM_CHECKPOINT_PATH=/path/to/xlm/checkpoint

-   transformers xlm \
+   pytorch_transformers xlm \
     $XLM_CHECKPOINT_PATH \
     $PYTORCH_DUMP_OUTPUT \
@@ -1 +0,0 @@
-../../examples/README.md
@@ -1,43 +1,9 @@
-Transformers
+Pytorch-Transformers
 ================================================================================================================================================

-🤗 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.
+PyTorch-Transformers is a library of state-of-the-art pre-trained models for Natural Language Processing (NLP).

-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:
+The library currently contains PyTorch 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.
@@ -46,10 +12,7 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
 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/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/pytorch/fairseq/tree/master/examples/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.
+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.

 .. toctree::
    :maxdepth: 2
@@ -65,8 +28,6 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
    migration
    bertology
    torchscript
-    multilingual
-    benchmarks

 .. toctree::
    :maxdepth: 2
@@ -76,7 +37,6 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
    main_classes/model
    main_classes/tokenizer
    main_classes/optimizer_schedules
-    main_classes/processors

 .. toctree::
    :maxdepth: 2
@@ -91,6 +51,3 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
    model_doc/xlnet
    model_doc/roberta
    model_doc/distilbert
-    model_doc/ctrl
-    model_doc/camembert
-    model_doc/albert
@@ -1,58 +0,0 @@
-# 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
-
-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 [--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/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:
-
-``` bash
-python -m pytest -sv ./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`:
-
-``` 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!
@@ -0,0 +1,71 @@
+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!
@@ -6,5 +6,5 @@ The base class ``PretrainedConfig`` implements the common methods for loading/sa
 ``PretrainedConfig``
 ~~~~~~~~~~~~~~~~~~~~~

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

-.. autoclass:: transformers.PreTrainedModel
-    :members:
-
-``TFPreTrainedModel``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFPreTrainedModel
+.. autoclass:: pytorch_transformers.PreTrainedModel
    :members:
@@ -9,7 +9,7 @@ The ``.optimization`` module provides:
 ``AdamW``
 ~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.AdamW
+.. autoclass:: pytorch_transformers.AdamW
    :members:

 Schedules
@@ -18,17 +18,19 @@ Schedules
 Learning Rate Schedules
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

-.. autofunction:: transformers.get_constant_schedule
+.. autoclass:: pytorch_transformers.ConstantLRSchedule
+    :members:


-.. autofunction:: transformers.get_constant_schedule_with_warmup
+.. autoclass:: pytorch_transformers.WarmupConstantSchedule
+    :members:

 .. image:: /imgs/warmup_constant_schedule.png
    :target: /imgs/warmup_constant_schedule.png
    :alt:


-.. autofunction:: transformers.get_cosine_schedule_with_warmup
+.. autoclass:: pytorch_transformers.WarmupCosineSchedule
    :members:

 .. image:: /imgs/warmup_cosine_schedule.png
@@ -36,7 +38,8 @@ Learning Rate Schedules
    :alt:


-.. autofunction:: transformers.get_cosine_with_hard_restarts_schedule_with_warmup
+.. autoclass:: pytorch_transformers.WarmupCosineWithHardRestartsSchedule
+    :members:

 .. image:: /imgs/warmup_cosine_hard_restarts_schedule.png
    :target: /imgs/warmup_cosine_hard_restarts_schedule.png
@@ -44,7 +47,8 @@ Learning Rate Schedules



-.. autofunction:: transformers.get_linear_schedule_with_warmup
+.. autoclass:: pytorch_transformers.WarmupLinearSchedule
+    :members:

 .. image:: /imgs/warmup_linear_schedule.png
    :target: /imgs/warmup_linear_schedule.png
@@ -1,81 +0,0 @@
-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-ressource language such as English and low-ressource 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.
-
-Example usage
-^^^^^^^^^^^^^^^^^^^^^^^^^
-
-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.
@@ -12,5 +12,5 @@ The base class ``PreTrainedTokenizer`` implements the common methods for loading
 ``PreTrainedTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.PreTrainedTokenizer
+.. autoclass:: pytorch_transformers.PreTrainedTokenizer
    :members:
@@ -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 `transformers`
+Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to `pytorch-transformers`

 ### Models always output `tuples`

-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 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 exact content of the tuples for each model are detailled in the models' docstrings and the [documentation](https://huggingface.co/transformers/).
+The exact content of the tuples for each model are detailled in the models' docstrings and the [documentation](https://huggingface.co/pytorch-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 `transformers` conversion example for a `BertForSequenceClassification` classification model:
+Here is a `pytorch-pretrained-bert` to `pytorch-transformers` conversion example for a `BertForSequenceClassification` classification model:

 ```python
 # Let's load our model
@@ -20,11 +20,11 @@ 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 transformers to extract the loss from the output tuple:
+# Now just use this line in pytorch-transformers to extract the loss from the output tuple:
 outputs = model(input_ids, labels=labels)
 loss = outputs[0]

-# In transformers you can also have access to the logits:
+# In pytorch-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)
@@ -84,21 +84,21 @@ Here is a conversion examples from `BertAdam` with a linear warmup and decay sch
 # Parameters:
 lr = 1e-3
 max_grad_norm = 1.0
-num_training_steps = 1000
+num_total_steps = 1000
 num_warmup_steps = 100
-warmup_proportion = float(num_warmup_steps) / float(num_training_steps)  # 0.1
+warmup_proportion = float(num_warmup_steps) / float(num_total_steps)  # 0.1

 ### Previously BertAdam optimizer was instantiated like this:
-optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, num_training_steps=num_training_steps)
+optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
 ### and used like this:
 for batch in train_data:
    loss = model(batch)
    loss.backward()
    optimizer.step()

-### In Transformers, optimizer and schedules are splitted and instantiated like this:
+### In PyTorch-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 = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)  # PyTorch scheduler
+scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_steps)  # PyTorch scheduler
 ### and used like this:
 for batch in train_data:
    loss = model(batch)
@@ -1,64 +0,0 @@
-ALBERT
----------------------------------------------------
-
-``AlbrtConfig``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.AlbertConfig
-    :members:
-
-
-``AlbertTokenizer``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.AlbertTokenizer
-    :members:
-
-
-``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:
@@ -11,19 +11,19 @@ Instantiating one of ``AutoModel``, ``AutoConfig`` and ``AutoTokenizer`` will di
 ``AutoConfig``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.AutoConfig
+.. autoclass:: pytorch_transformers.AutoConfig
    :members:


 ``AutoModel``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.AutoModel
+.. autoclass:: pytorch_transformers.AutoModel
    :members:


 ``AutoTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.AutoTokenizer
+.. autoclass:: pytorch_transformers.AutoTokenizer
    :members:
@@ -4,125 +4,69 @@ BERT
 ``BertConfig``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.BertConfig
+.. autoclass:: pytorch_transformers.BertConfig
    :members:


 ``BertTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.BertTokenizer
+.. autoclass:: pytorch_transformers.BertTokenizer
    :members:


 ``BertModel``
 ~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.BertModel
+.. autoclass:: pytorch_transformers.BertModel
    :members:


 ``BertForPreTraining``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.BertForPreTraining
+.. autoclass:: pytorch_transformers.BertForPreTraining
    :members:


 ``BertForMaskedLM``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.BertForMaskedLM
+.. autoclass:: pytorch_transformers.BertForMaskedLM
    :members:


 ``BertForNextSentencePrediction``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.BertForNextSentencePrediction
+.. autoclass:: pytorch_transformers.BertForNextSentencePrediction
    :members:


 ``BertForSequenceClassification``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.BertForSequenceClassification
+.. autoclass:: pytorch_transformers.BertForSequenceClassification
    :members:


 ``BertForMultipleChoice``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.BertForMultipleChoice
+.. autoclass:: pytorch_transformers.BertForMultipleChoice
    :members:


 ``BertForTokenClassification``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.BertForTokenClassification
+.. autoclass:: pytorch_transformers.BertForTokenClassification
    :members:


 ``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
+.. autoclass:: pytorch_transformers.BertForQuestionAnswering
    :members:

@@ -1,50 +0,0 @@
-CamemBERT
----------------------------------------------------
-
-``CamembertConfig``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.CamembertConfig
-    :members:
-
-
-``CamembertTokenizer``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.CamembertTokenizer
-    :members:
-
-
-``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:
@@ -1,49 +0,0 @@
-CTRL
----------------------------------------------------
-
-Note: if you fine-tune a CTRL model using the Salesforce code (https://github.com/salesforce/ctrl),
-you'll be able to convert from TF to our HuggingFace/Transformers format using the 
-``convert_tf_to_huggingface_pytorch.py`` script (see `issue #1654 <https://github.com/huggingface/transformers/issues/1654>`_).
-
-
-``CTRLConfig``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. 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:
-
@@ -4,67 +4,40 @@ DistilBERT
 ``DistilBertConfig``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.DistilBertConfig
+.. autoclass:: pytorch_transformers.DistilBertConfig
    :members:


 ``DistilBertTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.DistilBertTokenizer
+.. autoclass:: pytorch_transformers.DistilBertTokenizer
    :members:


 ``DistilBertModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.DistilBertModel
+.. autoclass:: pytorch_transformers.DistilBertModel
    :members:


 ``DistilBertForMaskedLM``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.DistilBertForMaskedLM
+.. autoclass:: pytorch_transformers.DistilBertForMaskedLM
    :members:


 ``DistilBertForSequenceClassification``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.DistilBertForSequenceClassification
+.. autoclass:: pytorch_transformers.DistilBertForSequenceClassification
    :members:


 ``DistilBertForQuestionAnswering``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.DistilBertForQuestionAnswering
-    :members:
-
-``TFDistilBertModel``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFDistilBertModel
-    :members:
-
-
-``TFDistilBertForMaskedLM``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFDistilBertForMaskedLM
-    :members:
-
-
-``TFDistilBertForSequenceClassification``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFDistilBertForSequenceClassification
-    :members:
-
-
-``TFDistilBertForQuestionAnswering``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFDistilBertForQuestionAnswering
+.. autoclass:: pytorch_transformers.DistilBertForQuestionAnswering
    :members:
@@ -4,54 +4,33 @@ OpenAI GPT
 ``OpenAIGPTConfig``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.OpenAIGPTConfig
+.. autoclass:: pytorch_transformers.OpenAIGPTConfig
    :members:


 ``OpenAIGPTTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.OpenAIGPTTokenizer
+.. autoclass:: pytorch_transformers.OpenAIGPTTokenizer
    :members:


 ``OpenAIGPTModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.OpenAIGPTModel
+.. autoclass:: pytorch_transformers.OpenAIGPTModel
    :members:


 ``OpenAIGPTLMHeadModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.OpenAIGPTLMHeadModel
+.. autoclass:: pytorch_transformers.OpenAIGPTLMHeadModel
    :members:


 ``OpenAIGPTDoubleHeadsModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.OpenAIGPTDoubleHeadsModel
-    :members:
-
-
-``TFOpenAIGPTModel``
-~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFOpenAIGPTModel
-    :members:
-
-
-``TFOpenAIGPTLMHeadModel``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFOpenAIGPTLMHeadModel
-    :members:
-
-
-``TFOpenAIGPTDoubleHeadsModel``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFOpenAIGPTDoubleHeadsModel
+.. autoclass:: pytorch_transformers.OpenAIGPTDoubleHeadsModel
    :members:
@@ -4,54 +4,33 @@ OpenAI GPT2
 ``GPT2Config``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.GPT2Config
+.. autoclass:: pytorch_transformers.GPT2Config
    :members:


 ``GPT2Tokenizer``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.GPT2Tokenizer
+.. autoclass:: pytorch_transformers.GPT2Tokenizer
    :members:


 ``GPT2Model``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.GPT2Model
+.. autoclass:: pytorch_transformers.GPT2Model
    :members:


 ``GPT2LMHeadModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.GPT2LMHeadModel
+.. autoclass:: pytorch_transformers.GPT2LMHeadModel
    :members:


 ``GPT2DoubleHeadsModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.GPT2DoubleHeadsModel
-    :members:
-
-
-``TFGPT2Model``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFGPT2Model
-    :members:
-
-
-``TFGPT2LMHeadModel``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFGPT2LMHeadModel
-    :members:
-
-
-``TFGPT2DoubleHeadsModel``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFGPT2DoubleHeadsModel
+.. autoclass:: pytorch_transformers.GPT2DoubleHeadsModel
    :members:
@@ -4,54 +4,33 @@ RoBERTa
 ``RobertaConfig``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.RobertaConfig
+.. autoclass:: pytorch_transformers.RobertaConfig
    :members:


 ``RobertaTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.RobertaTokenizer
+.. autoclass:: pytorch_transformers.RobertaTokenizer
    :members:


 ``RobertaModel``
 ~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.RobertaModel
+.. autoclass:: pytorch_transformers.RobertaModel
    :members:


 ``RobertaForMaskedLM``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.RobertaForMaskedLM
+.. autoclass:: pytorch_transformers.RobertaForMaskedLM
    :members:


 ``RobertaForSequenceClassification``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.RobertaForSequenceClassification
-    :members:
-
-
-``TFRobertaModel``
-~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFRobertaModel
-    :members:
-
-
-``TFRobertaForMaskedLM``
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFRobertaForMaskedLM
-    :members:
-
-
-``TFRobertaForSequenceClassification``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFRobertaForSequenceClassification
+.. autoclass:: pytorch_transformers.RobertaForSequenceClassification
    :members:
@@ -5,40 +5,26 @@ Transformer XL
 ``TransfoXLConfig``
 ~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.TransfoXLConfig
+.. autoclass:: pytorch_transformers.TransfoXLConfig
    :members:


 ``TransfoXLTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.TransfoXLTokenizer
+.. autoclass:: pytorch_transformers.TransfoXLTokenizer
    :members:


 ``TransfoXLModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.TransfoXLModel
+.. autoclass:: pytorch_transformers.TransfoXLModel
    :members:


 ``TransfoXLLMHeadModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.TransfoXLLMHeadModel
-    :members:
-
-
-``TFTransfoXLModel``
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFTransfoXLModel
-    :members:
-
-
-``TFTransfoXLLMHeadModel``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFTransfoXLLMHeadModel
+.. autoclass:: pytorch_transformers.TransfoXLLMHeadModel
    :members:
@@ -4,66 +4,38 @@ XLM
 ``XLMConfig``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.XLMConfig
+.. autoclass:: pytorch_transformers.XLMConfig
    :members:

 ``XLMTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.XLMTokenizer
+.. autoclass:: pytorch_transformers.XLMTokenizer
    :members:

 ``XLMModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.XLMModel
+.. autoclass:: pytorch_transformers.XLMModel
    :members:


 ``XLMWithLMHeadModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.XLMWithLMHeadModel
+.. autoclass:: pytorch_transformers.XLMWithLMHeadModel
    :members:


 ``XLMForSequenceClassification``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.XLMForSequenceClassification
+.. autoclass:: pytorch_transformers.XLMForSequenceClassification
    :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
+.. autoclass:: pytorch_transformers.XLMForQuestionAnswering
    :members:
@@ -4,68 +4,40 @@ XLNet
 ``XLNetConfig``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.XLNetConfig
+.. autoclass:: pytorch_transformers.XLNetConfig
    :members:


 ``XLNetTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.XLNetTokenizer
+.. autoclass:: pytorch_transformers.XLNetTokenizer
    :members:


 ``XLNetModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.XLNetModel
+.. autoclass:: pytorch_transformers.XLNetModel
    :members:


 ``XLNetLMHeadModel``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.XLNetLMHeadModel
+.. autoclass:: pytorch_transformers.XLNetLMHeadModel
    :members:


 ``XLNetForSequenceClassification``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.XLNetForSequenceClassification
+.. autoclass:: pytorch_transformers.XLNetForSequenceClassification
    :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
+.. autoclass:: pytorch_transformers.XLNetForQuestionAnswering
    :members:
@@ -1,103 +0,0 @@
-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.
@@ -1,16 +1,16 @@
 Notebooks
 ================================================

-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.
+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.


 *
-  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 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 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 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 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.
+  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.

 Please follow the instructions given in the notebooks to run and modify them.
@@ -44,23 +44,15 @@ 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/transformers/tree/master/examples>`__).           |
+|                   |                                                            | (see details of fine-tuning in the `example section <https://github.com/huggingface/pytorch-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/transformers/examples.html>`__)                           |
+|                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/pytorch-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/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>`__).                                                         |
+|                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/pytorch-transformers/examples.html>`__)                   |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | GPT               | ``openai-gpt``                                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | OpenAI GPT English model                                                                                                            |
@@ -73,9 +65,6 @@ 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                                                                                               |
@@ -109,12 +98,6 @@ Here is the full list of the currently provided pretrained models together with
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``xlm-clm-ende-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
 |                   |                                                            | | 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                                                                                            |
@@ -127,70 +110,14 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``roberta-large-mnli``                                     | | 24-layer, 1024-hidden, 16-heads, 355M parameters                                                                                    |
 |                   |                                                            | | ``roberta-large`` fine-tuned on `MNLI <http://www.nyu.edu/projects/bowman/multinli/>`__.                                            |
 |                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__)                                                   |
-|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``roberta-base-openai-detector``                           | | 12-layer, 768-hidden, 12-heads, 125M parameters                                                                                     |
-|                   |                                                            | | ``roberta-base`` fine-tuned by OpenAI on the outputs of the 1.5B-parameter GPT-2 model.                                             |
-|                   |                                                            | (see `details <https://github.com/openai/gpt-2-output-dataset/tree/master/detector>`__)                                               |
-|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``roberta-large-openai-detector``                          | | 24-layer, 1024-hidden, 16-heads, 355M parameters                                                                                    |
-|                   |                                                            | | ``roberta-large`` fine-tuned by OpenAI on the outputs of the 1.5B-parameter GPT-2 model.                                            |
-|                   |                                                            | (see `details <https://github.com/openai/gpt-2-output-dataset/tree/master/detector>`__)                                               |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | DistilBERT        | ``distilbert-base-uncased``                                | | 6-layer, 768-hidden, 12-heads, 66M parameters                                                                                       |
 |                   |                                                            | | The DistilBERT model distilled from the BERT model `bert-base-uncased` checkpoint                                                   |
-|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   |                                                            | (see `details <https://medium.com/huggingface/distilbert-8cf3380435b5>`__)                                                            |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``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://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>`__)                                     |
-|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``distilroberta-base``                                     | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
-|                   |                                                            | | The DistilRoBERTa model distilled from the RoBERTa model `roberta-base` checkpoint.                                                 |
-|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
-+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-| CTRL              | ``ctrl``                                                   | | 48-layer, 1280-hidden, 16-heads, 1.6B parameters                                                                                    |
-|                   |                                                            | | Salesforce's Large-sized CTRL English model                                                                                         |
-+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-| CamemBERT         | ``camembert-base``                                         | | 12-layer, 768-hidden, 12-heads, 110M parameters                                                                                     |
-|                   |                                                            | | CamemBERT using the BERT-base architecture                                                                                          |
-|                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/camembert>`__)                                                 |
-+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-| 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/google-research/tree/master/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/google-research/tree/master/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/google-research/tree/master/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/google-research/tree/master/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/google-research/tree/master/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/google-research/tree/master/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/google-research/tree/master/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/google-research/tree/master/albert>`__)                                             |
+|                   |                                                            | (see `details <https://medium.com/huggingface/distilbert-8cf3380435b5>`__)                                                            |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+

-
-.. <https://huggingface.co/transformers/examples.html>`__
+.. <https://huggingface.co/pytorch-transformers/examples.html>`__
@@ -2,7 +2,7 @@

 ## Philosophy

-Transformers is an opinionated library built for NLP researchers seeking to use/study/extend large-scale transformers models.
+PyTorch-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/investigating these models:
+- incorporate a subjective selection of promising tools for fine-tuning/investiguating 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 8 models architectures currently provided in the library, e.g. `BertModel`
+- **model classes** which are PyTorch models (`torch.nn.Modules`) of the 6 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/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/pytorch-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 class.
+See full API reference for examples for each model classe.

 ### 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 transformers import BertTokenizer, BertModel, BertForMaskedLM
+from pytorch_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 deactivate the DropOut modules
-# This is IMPORTANT to have reproducible results during evaluation!
+# Set the model in evaluation mode to desactivate the DropOut modules
+# This is IMPORTANT to have reproductible 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)
-    # Transformers models always output tuples.
+    # PyTorch-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 transformers import GPT2Tokenizer, GPT2LMHeadModel
+from pytorch_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 deactivate the DropOut modules
-# This is IMPORTANT to have reproducible results during evaluation!
+# Set the model in evaluation mode to desactivate the DropOut modules
+# This is IMPORTANT to have reproductible results during evaluation!
 model.eval()

 # If you have a GPU, put everything on cuda
@@ -188,35 +188,3 @@ assert predicted_text == 'Who was Jim Henson? Jim Henson was a man'
 ```

 Examples for each model class of each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the [documentation](#documentation).
-
-#### Using the past
-
-GPT-2 as well as some other models (GPT, XLNet, Transfo-XL, CTRL) make use of a `past` or `mems` attribute which can be used to prevent re-computing the key/value pairs when using sequential decoding. It is useful when generating sequences as a big part of the attention mechanism benefits from previous computations.
-
-Here is a fully-working example using the `past` with `GPT2LMHeadModel` and argmax decoding (which should only be used as an example, as argmax decoding introduces a lot of repetition):
-
-```python
-from transformers import GPT2LMHeadModel, GPT2Tokenizer
-import torch
-
-tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
-model = GPT2LMHeadModel.from_pretrained('gpt2')
-
-generated = tokenizer.encode("The Manhattan bridge")
-context = torch.tensor([generated])
-past = None
-
-for i in range(100):
-    print(i)
-    output, past = model(context, past=past)
-    token = torch.argmax(output[0, :])
-
-    generated += [token.tolist()]
-    context = token.unsqueeze(0)
-
-sequence = tokenizer.decode(generated)
-
-print(sequence)
-```
-
-The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.
@@ -33,8 +33,6 @@ 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
@@ -47,7 +45,7 @@ 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/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/``\ ).
+  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/``\ ).

 *
  ``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).
@@ -106,7 +104,7 @@ 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 itself which should be saved following PyTorch serialization `best practices <https://pytorch.org/docs/stable/notes/serialization.html#best-practices>`__\ ,
+* the model it-self 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).

@@ -124,7 +122,7 @@ Here is the recommended way of saving the model, configuration and vocabulary to

 .. code-block:: python

-   from transformers import WEIGHTS_NAME, CONFIG_NAME
+   from pytorch_transformers import WEIGHTS_NAME, CONFIG_NAME

   output_dir = "./models/"

@@ -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 `transformers` models to TorchScript so that they can
+We have provided an interface that allows the export of `pytorch-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 transformers import BertModel, BertTokenizer, BertConfig
+    from pytorch_transformers import BertModel, BertTokenizer, BertConfig
    import torch

    enc = BertTokenizer.from_pretrained("bert-base-uncased")
@@ -3,52 +3,17 @@
 In this section a few examples are put together. All of these examples work for several models, making use of the very
 similar API between the different models.

-**Important**  
-To run the latest versions of the examples, you have to install from source. Execute the following steps in a new virtual environment:
-
-```bash
-git clone https://github.com/huggingface/transformers
-cd transformers
-pip install [--editable] .
-```
-
 | Section                    | Description                                                                                                                                                |
 |----------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| [TensorFlow 2.0 models on GLUE](#TensorFlow-2.0-Bert-models-on-GLUE) | Examples running BERT TensorFlow 2.0 model on the GLUE tasks. 
 | [Language Model fine-tuning](#language-model-fine-tuning) | Fine-tuning the library models for language modeling on a text dataset. Causal language modeling for GPT/GPT-2, masked language modeling for BERT/RoBERTa. |
 | [Language Generation](#language-generation) | Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet.                                         |
 | [GLUE](#glue) | Examples running BERT/XLM/XLNet/RoBERTa on the 9 GLUE tasks. Examples feature distributed training as well as half-precision.                              |
-| [SQuAD](#squad) | Using BERT/RoBERTa/XLNet/XLM for question answering, examples with distributed training.                                                                                  |
-| [Multiple Choice](#multiple-choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks. 
-| [Named Entity Recognition](#named-entity-recognition) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training.                                                                                  |
-| [XNLI](#xnli) | Examples running BERT/XLM on the XNLI benchmark. |
-| [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).
+| [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. 

 ## Language model fine-tuning

-Based on the script [`run_lm_finetuning.py`](https://github.com/huggingface/transformers/blob/master/examples/run_lm_finetuning.py).
+Based on the script [`run_lm_finetuning.py`](https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_lm_finetuning.py).

 Fine-tuning 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 
@@ -110,9 +75,9 @@ python run_lm_finetuning.py \

 ## Language generation

-Based on the script [`run_generation.py`](https://github.com/huggingface/transformers/blob/master/examples/run_generation.py).
+Based on the script [`run_generation.py`](https://github.com/huggingface/pytorch-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.
+Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet.
 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.

@@ -126,26 +91,26 @@ python run_generation.py \

 ## GLUE

-Based on the script [`run_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/run_glue.py).
+Based on the script [`run_glue.py`](https://github.com/huggingface/pytorch-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 on 8 V100 GPUs with a total train
+uncased  BERT base model (the checkpoint `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.

 | Task  | Metric                       | Result      |
 |-------|------------------------------|-------------|
-| CoLA  | Matthew's corr               | 48.87       |
-| SST-2 | Accuracy                     | 91.74       |
-| MRPC  | F1/Accuracy                  | 90.70/86.27 |
-| STS-B | Person/Spearman corr.        | 91.39/91.04 |
-| QQP   | Accuracy/F1                  | 90.79/87.66 |
-| MNLI  | Matched acc./Mismatched acc. | 83.70/84.83 |
-| QNLI  | Accuracy                     | 89.31       |
-| RTE   | Accuracy                     | 71.43       |
+| CoLA  | Matthew's corr               | 55.75       |
+| SST-2 | Accuracy                     | 92.09       |
+| MRPC  | F1/Accuracy                  | 90.48/86.27 |
+| STS-B | Person/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
@@ -318,17 +283,17 @@ 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/run_multiple_choice.py \
+python ./examples/single_model_scripts/run_multiple_choice.py \
 --model_type roberta \
 --task_name swag \
 --model_name_or_path roberta-base \
@@ -354,7 +319,7 @@ eval_loss = 0.44457291918821606

 ## SQuAD

-Based on the script [`run_squad.py`](https://github.com/huggingface/transformers/blob/master/examples/run_squad.py).
+Based on the script [`run_squad.py`](https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_squad.py).

 #### Fine-tuning on SQuAD

@@ -422,222 +387,6 @@ f1 = 93.15
 exact_match = 86.91
 ```

-This fine-tuned model is available as a checkpoint under the reference
+This fine-tuneds model is available as a checkpoint under the reference
 `bert-large-uncased-whole-word-masking-finetuned-squad`.

-#### Fine-tuning XLNet on SQuAD
-
-This example code fine-tunes XLNet on the SQuAD dataset. See above to download the data for SQuAD .
-
-```bash
-export SQUAD_DIR=/path/to/SQUAD
-
-python /data/home/hlu/transformers/examples/run_squad.py \
-    --model_type xlnet \
-    --model_name_or_path xlnet-large-cased \
-    --do_train \
-    --do_eval \
-    --do_lower_case \
-    --train_file /data/home/hlu/notebooks/NLP/examples/question_answering/train-v1.1.json \
-    --predict_file /data/home/hlu/notebooks/NLP/examples/question_answering/dev-v1.1.json \
-    --learning_rate 3e-5 \
-    --num_train_epochs 2 \
-    --max_seq_length 384 \
-    --doc_stride 128 \
-    --output_dir ./wwm_cased_finetuned_squad/ \
-    --per_gpu_eval_batch_size=4  \
-    --per_gpu_train_batch_size=4   \
-    --save_steps 5000
-```
-
-Training with the previously defined hyper-parameters yields the following results:
-
-```python
-{
-"exact": 85.45884578997162,
-"f1": 92.5974600601065,
-"total": 10570,
-"HasAns_exact": 85.45884578997162,
-"HasAns_f1": 92.59746006010651,
-"HasAns_total": 10570
-}
-```
-
-## Named Entity Recognition
-
-Based on the script [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py).
-This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
-Details and results for the fine-tuning provided by @stefan-it.
-
-### Data (Download and pre-processing steps)
-
-Data can be obtained from the [GermEval 2014](https://sites.google.com/site/germeval2014ner/data) shared task page.
-
-Here are the commands for downloading and pre-processing train, dev and test datasets. The original data format has four (tab-separated) columns, in a pre-processing step only the two relevant columns (token and outer span NER annotation) are extracted:
-
-```bash
-curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-train.tsv?attredirects=0&d=1' \
-| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > train.txt.tmp
-curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-dev.tsv?attredirects=0&d=1' \
-| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > dev.txt.tmp
-curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-test.tsv?attredirects=0&d=1' \
-| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > test.txt.tmp
-```
-
-The GermEval 2014 dataset contains some strange "control character" tokens like `'\x96', '\u200e', '\x95', '\xad' or '\x80'`. One problem with these tokens is, that `BertTokenizer` returns an empty token for them, resulting in misaligned `InputExample`s. I wrote a script that a) filters these tokens and b) splits longer sentences into smaller ones (once the max. subtoken length is reached).
-
-```bash
-wget "https://raw.githubusercontent.com/stefan-it/fine-tuned-berts-seq/master/scripts/preprocess.py"
-```
-Let's define some variables that we need for further pre-processing steps and training the model:
-
-```bash
-export MAX_LENGTH=128
-export BERT_MODEL=bert-base-multilingual-cased
-```
-
-Run the pre-processing script on training, dev and test datasets:
-
-```bash
-python3 preprocess.py train.txt.tmp $BERT_MODEL $MAX_LENGTH > train.txt
-python3 preprocess.py dev.txt.tmp $BERT_MODEL $MAX_LENGTH > dev.txt
-python3 preprocess.py test.txt.tmp $BERT_MODEL $MAX_LENGTH > test.txt
-```
-
-The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so an own set of labels must be used:
-
-```bash
-cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
-```
-
-### Training
-
-Additional environment variables must be set:
-
-```bash
-export OUTPUT_DIR=germeval-model
-export BATCH_SIZE=32
-export NUM_EPOCHS=3
-export SAVE_STEPS=750
-export SEED=1
-```
-
-To start training, just run:
-
-```bash
-python3 run_ner.py --data_dir ./ \
--model_type bert \
--labels ./labels.txt \
--model_name_or_path $BERT_MODEL \
--output_dir $OUTPUT_DIR \
--max_seq_length  $MAX_LENGTH \
--num_train_epochs $NUM_EPOCHS \
--per_gpu_train_batch_size $BATCH_SIZE \
--save_steps $SAVE_STEPS \
--seed $SEED \
--do_train \
--do_eval \
--do_predict
-```
-
-If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
-
-### Evaluation
-
-Evaluation on development dataset outputs the following for our example:
-
-```bash
-10/04/2019 00:42:06 - INFO - __main__ -   ***** Eval results  *****
-10/04/2019 00:42:06 - INFO - __main__ -     f1 = 0.8623348017621146
-10/04/2019 00:42:06 - INFO - __main__ -     loss = 0.07183869666975543
-10/04/2019 00:42:06 - INFO - __main__ -     precision = 0.8467916366258111
-10/04/2019 00:42:06 - INFO - __main__ -     recall = 0.8784592370979806
-```
-
-On the test dataset the following results could be achieved:
-
-```bash
-10/04/2019 00:42:42 - INFO - __main__ -   ***** Eval results  *****
-10/04/2019 00:42:42 - INFO - __main__ -     f1 = 0.8614389652384803
-10/04/2019 00:42:42 - INFO - __main__ -     loss = 0.07064602487454782
-10/04/2019 00:42:42 - INFO - __main__ -     precision = 0.8604651162790697
-10/04/2019 00:42:42 - INFO - __main__ -     recall = 0.8624150210424085
-```
-
-### Comparing BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased)
-
-Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased) with the same hyperparameters as specified in the [example documentation](https://huggingface.co/transformers/examples.html#named-entity-recognition) (one run):
-
-| Model | F-Score Dev | F-Score Test
-| --------------------------------- | ------- | --------
-| `bert-large-cased`            | 95.59 | 91.70
-| `roberta-large`                  | 95.96 | 91.87
-| `distilbert-base-uncased` | 94.34 | 90.32
-
-## Abstractive summarization
-
-Based on the script
-[`run_summarization_finetuning.py`](https://github.com/huggingface/transformers/blob/master/examples/run_summarization_finetuning.py).
-
-Before running this script you should download **both** CNN and Daily Mail
-datasets from [Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/)  (the
-links next to "Stories") in the same folder. Then uncompress the archives by running:
-
-```bash
-tar -xvf cnn_stories.tgz && tar -xvf dailymail_stories.tgz
-```
-
-note that the finetuning script **will not work** if you do not download both
-datasets. We will refer as `$DATA_PATH` the path to where you uncompressed both
-archive.
-
-```bash
-export DATA_PATH=/path/to/dataset/
-
-python run_summarization_finetuning.py \
-    --output_dir=output \
-    --model_type=bert2bert \
-    --model_name_or_path=bert2bert \
-    --do_train \
-    --data_path=$DATA_PATH \
-```
-
-## XNLI
-
-Based on the script [`run_xnli.py`](https://github.com/huggingface/transformers/blob/master/examples/run_xnli.py).
-
-[XNLI](https://www.nyu.edu/projects/bowman/xnli/) is crowd-sourced dataset based on [MultiNLI](http://www.nyu.edu/projects/bowman/multinli/). It is an evaluation benchmark for cross-lingual text representations. Pairs of text are labeled with textual entailment annotations for 15 different languages (including both high-ressource language such as English and low-ressource languages such as Swahili).
-
-#### Fine-tuning on XNLI
-
-This example code fine-tunes mBERT (multi-lingual BERT) on the XNLI dataset. It runs in 106 mins
-on a single tesla V100 16GB. The data for XNLI can be downloaded with the following links and should be both saved (and un-zipped) in a 
-`$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
-```
@@ -1,477 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The HuggingFace Inc. team.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" Benchmarking the library on inference and training """
-
-# If checking the tensors placement
-# tf.debugging.set_log_device_placement(True)
-
-from typing import List
-import timeit
-from transformers import is_tf_available, is_torch_available
-from time import time
-import argparse
-import csv
-
-if is_tf_available():
-    import tensorflow as tf
-    from transformers import TFAutoModel
-
-if is_torch_available():
-    import torch
-    from transformers import AutoModel
-
-from transformers import AutoConfig, AutoTokenizer
-
-input_text = """Bent over their instruments, three hundred Fertilizers were plunged, as 
-the Director of Hatcheries and Conditioning entered the room, in the 
-
-
-
-scarcely breathing silence, the absent-minded, soliloquizing hum or 
-whistle, of absorbed concentration. A troop of newly arrived students, 
-very young, pink and callow, followed nervously, rather abjectly, at the 
-Director's heels. Each of them carried a notebook, in which, whenever 
-the great man spoke, he desperately scribbled. Straight from the 
-horse's mouth. It was a rare privilege. The D. H. C. for Central London 
-always made a point of personally conducting his new students round 
-the various departments. 
-
-"Just to give you a general idea," he would explain to them. For of 
-course some sort of general idea they must have, if they were to do 
-their work intelligently-though as little of one, if they were to be good 
-and happy members of society, as possible. For particulars, as every 
-one knows, make for virtue and happiness; generalities are intellectu- 
-ally necessary evils. Not philosophers but fret-sawyers and stamp col- 
-lectors compose the backbone of society. 
-
-"To-morrow," he would add, smiling at them with a slightly menacing 
-geniality, "you'll be settling down to serious work. You won't have time 
-for generalities. Meanwhile ..." 
-
-Meanwhile, it was a privilege. Straight from the horse's mouth into the 
-notebook. The boys scribbled like mad. 
-
-Tall and rather thin but upright, the Director advanced into the room. 
-He had a long chin and big rather prominent teeth, just covered, when 
-he was not talking, by his full, floridly curved lips. Old, young? Thirty? 
-Fifty? Fifty-five? It was hard to say. And anyhow the question didn't 
-arise; in this year of stability, A. F. 632, it didn't occur to you to ask it. 
-
-"I shall begin at the beginning," said the D.H.C. and the more zealous 
-students recorded his intention in their notebooks: Begin at the begin- 
-ning. "These," he waved his hand, "are the incubators." And opening 
-an insulated door he showed them racks upon racks of numbered test- 
-tubes. "The week's supply of ova. Kept," he explained, "at blood heat; 
-whereas the male gametes," and here he opened another door, "they 
-have to be kept at thirty-five instead of thirty-seven. Full blood heat 
-sterilizes." Rams wrapped in theremogene beget no lambs. 
-
-Still leaning against the incubators he gave them, while the pencils 
-scurried illegibly across the pages, a brief description of the modern 
-
-
-
-fertilizing process; spoke first, of course, of its surgical introduc- 
-tion-"the operation undergone voluntarily for the good of Society, not 
-to mention the fact that it carries a bonus amounting to six months' 
-salary"; continued with some account of the technique for preserving 
-the excised ovary alive and actively developing; passed on to a consid- 
-eration of optimum temperature, salinity, viscosity; referred to the liq- 
-uor in which the detached and ripened eggs were kept; and, leading 
-his charges to the work tables, actually showed them how this liquor 
-was drawn off from the test-tubes; how it was let out drop by drop 
-onto the specially warmed slides of the microscopes; how the eggs 
-which it contained were inspected for abnormalities, counted and 
-transferred to a porous receptacle; how (and he now took them to 
-watch the operation) this receptacle was immersed in a warm bouillon 
-containing free-swimming spermatozoa-at a minimum concentration 
-of one hundred thousand per cubic centimetre, he insisted; and how, 
-after ten minutes, the container was lifted out of the liquor and its 
-contents re-examined; how, if any of the eggs remained unfertilized, it 
-was again immersed, and, if necessary, yet again; how the fertilized 
-ova went back to the incubators; where the Alphas and Betas re- 
-mained until definitely bottled; while the Gammas, Deltas and Epsilons 
-were brought out again, after only thirty-six hours, to undergo Bo- 
-kanovsky's Process. 
-
-"Bokanovsky's Process," repeated the Director, and the students un- 
-derlined the words in their little notebooks. 
-
-One egg, one embryo, one adult-normality. But a bokanovskified egg 
-will bud, will proliferate, will divide. From eight to ninety-six buds, and 
-every bud will grow into a perfectly formed embryo, and every embryo 
-into a full-sized adult. Making ninety-six human beings grow where 
-only one grew before. Progress. 
-
-"Essentially," the D.H.C. concluded, "bokanovskification consists of a 
-series of arrests of development. We check the normal growth and, 
-paradoxically enough, the egg responds by budding." 
-
-Responds by budding. The pencils were busy. 
-
-He pointed. On a very slowly moving band a rack-full of test-tubes was 
-entering a large metal box, another, rack-full was emerging. Machinery 
-faintly purred. It took eight minutes for the tubes to go through, he 
-
-
-
-told them. Eight minutes of hard X-rays being about as much as an 
-egg can stand. A few died; of the rest, the least susceptible divided 
-into two; most put out four buds; some eight; all were returned to the 
-incubators, where the buds began to develop; then, after two days, 
-were suddenly chilled, chilled and checked. Two, four, eight, the buds 
-in their turn budded; and having budded were dosed almost to death 
-with alcohol; consequently burgeoned again and having budded-bud 
-out of bud out of bud-were thereafter-further arrest being generally 
-fatal-left to develop in peace. By which time the original egg was in a 
-fair way to becoming anything from eight to ninety-six embryos- a 
-prodigious improvement, you will agree, on nature. Identical twins-but 
-not in piddling twos and threes as in the old viviparous days, when an 
-egg would sometimes accidentally divide; actually by dozens, by 
-scores at a time. 
-
-"Scores," the Director repeated and flung out his arms, as though he 
-were distributing largesse. "Scores." 
-
-But one of the students was fool enough to ask where the advantage 
-lay. 
-
-"My good boy!" The Director wheeled sharply round on him. "Can't you 
-see? Can't you see?" He raised a hand; his expression was solemn. 
-"Bokanovsky's Process is one of the major instruments of social stabil- 
-ity!" 
-
-Major instruments of social stability. 
-
-Standard men and women; in uniform batches. The whole of a small 
-factory staffed with the products of a single bokanovskified egg. 
-
-"Ninety-six identical twins working ninety-six identical machines!" The 
-voice was almost tremulous with enthusiasm. "You really know where 
-you are. For the first time in history." He quoted the planetary motto. 
-"Community, Identity, Stability." Grand words. "If we could bo- 
-kanovskify indefinitely the whole problem would be solved." 
-
-Solved by standard Gammas, unvarying Deltas, uniform Epsilons. Mil- 
-lions of identical twins. The principle of mass production at last applied 
-to biology. 
-
-
-
-"But, alas," the Director shook his head, "we can't bokanovskify indefi- 
-nitely." 
-
-Ninety-six seemed to be the limit; seventy-two a good average. From 
-the same ovary and with gametes of the same male to manufacture as 
-many batches of identical twins as possible-that was the best (sadly a 
-second best) that they could do. And even that was difficult. 
-
-"For in nature it takes thirty years for two hundred eggs to reach ma- 
-turity. But our business is to stabilize the population at this moment, 
-here and now. Dribbling out twins over a quarter of a century-what 
-would be the use of that?" 
-
-Obviously, no use at all. But Podsnap's Technique had immensely ac- 
-celerated the process of ripening. They could make sure of at least a 
-hundred and fifty mature eggs within two years. Fertilize and bo- 
-kanovskify-in other words, multiply by seventy-two-and you get an 
-average of nearly eleven thousand brothers and sisters in a hundred 
-and fifty batches of identical twins, all within two years of the same 
-age. 
-
-"And in exceptional cases we can make one ovary yield us over fifteen 
-thousand adult individuals." 
-
-Beckoning to a fair-haired, ruddy young man who happened to be 
-passing at the moment. "Mr. Foster," he called. The ruddy young man 
-approached. "Can you tell us the record for a single ovary, Mr. Foster?" 
-
-"Sixteen thousand and twelve in this Centre," Mr. Foster replied with- 
-out hesitation. He spoke very quickly, had a vivacious blue eye, and 
-took an evident pleasure in quoting figures. "Sixteen thousand and 
-twelve; in one hundred and eighty-nine batches of identicals. But of 
-course they've done much better," he rattled on, "in some of the tropi- 
-cal Centres. Singapore has often produced over sixteen thousand five 
-hundred; and Mombasa has actually touched the seventeen thousand 
-mark. But then they have unfair advantages. You should see the way a 
-negro ovary responds to pituitary! It's quite astonishing, when you're 
-used to working with European material. Still," he added, with a laugh 
-(but the light of combat was in his eyes and the lift of his chin was 
-challenging), "still, we mean to beat them if we can. I'm working on a 
-wonderful Delta-Minus ovary at this moment. Only just eighteen 
-
-
-
-months old. Over twelve thousand seven hundred children already, ei- 
-ther decanted or in embryo. And still going strong. We'll beat them 
-yet." 
-
-"That's the spirit I like!" cried the Director, and clapped Mr. Foster on 
-the shoulder. "Come along with us, and give these boys the benefit of 
-your expert knowledge." 
-
-Mr. Foster smiled modestly. "With pleasure." They went. 
-In the Bottling Room all was harmonious bustle and ordered activity. 
-Flaps of fresh sow's peritoneum ready cut to the proper size came 
-shooting up in little lifts from the Organ Store in the sub-basement. 
-Whizz and then, click! the lift-hatches hew open; the bottle-liner had 
-only to reach out a hand, take the flap, insert, smooth-down, and be- 
-fore the lined bottle had had time to travel out of reach along the end- 
-less band, whizz, click! another flap of peritoneum had shot up from 
-the depths, ready to be slipped into yet another bottle, the next of that 
-slow interminable procession on the band. 
-
-Next to the Liners stood the Matriculators. The procession advanced; 
-one by one the eggs were transferred from their test-tubes to the 
-larger containers; deftly the peritoneal lining was slit, the morula 
-dropped into place, the saline solution poured in ... and already the 
-bottle had passed, and it was the turn of the labellers. Heredity, date 
-of fertilization, membership of Bokanovsky Group-details were trans- 
-ferred from test-tube to bottle. No longer anonymous, but named, 
-identified, the procession marched slowly on; on through an opening in 
-the wall, slowly on into the Social Predestination Room. 
-"Eighty-eight cubic metres of card-index," said Mr. Foster with relish, 
-as they entered."""
-
-
-def create_setup_and_compute(model_names: List[str],
-                             gpu: bool = True,
-                             tensorflow: bool = False,
-                             average_over: int = 3,
-                             torchscript: bool = False,
-                             xla: bool = False,
-                             amp: bool = False,
-                             fp16: bool = False,
-                             save_to_csv: bool = False,
-                             csv_filename: str = f"results_{round(time())}.csv"):
-    if xla:
-        tf.config.optimizer.set_jit(True)
-    if amp:
-        tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})
-
-    if tensorflow:
-        dictionary = {model_name: {} for model_name in model_names}
-        results = _compute_tensorflow(model_names, dictionary, average_over, amp)
-    else:
-        device = 'cuda' if (gpu and torch.cuda.is_available()) else 'cpu'
-        dictionary = {model_name: {} for model_name in model_names}
-        results = _compute_pytorch(model_names, dictionary, average_over, device, torchscript, fp16)
-
-    print("=========== RESULTS ===========")
-    for model_name in model_names:
-        print("\t" + f"======= MODEL CHECKPOINT: {model_name} =======")
-        for batch_size in results[model_name]["bs"]:
-            print("\t\t" + f"===== BATCH SIZE: {batch_size} =====")
-            for slice_size in results[model_name]["ss"]:
-                result = results[model_name]['results'][batch_size][slice_size]
-                if isinstance(result, str):
-                    print(f"\t\t{model_name}/{batch_size}/{slice_size}: "
-                          f"{result}")
-                else:
-                    print(f"\t\t{model_name}/{batch_size}/{slice_size}: "
-                          f"{(round(1000 * result) / 1000)}"
-                          f"s")
-
-    if save_to_csv:
-        with open(csv_filename, mode='w') as csv_file:
-            fieldnames = ['model',
-                          '1x8', '1x64', '1x128', '1x256', '1x512', '1x1024',
-                          '2x8', '2x64', '2x128', '2x256', '2x512', '2x1024',
-                          '4x8', '4x64', '4x128', '4x256', '4x512', '4x1024',
-                          '8x8', '8x64', '8x128', '8x256', '8x512', '8x1024',
-                          ]
-
-            writer = csv.DictWriter(csv_file, fieldnames=fieldnames)
-            writer.writeheader()
-
-            for model_name in model_names:
-                model_results = {
-                    f'{bs}x{ss}': results[model_name]['results'][bs][ss]
-                    for bs in results[model_name]["results"]
-                    for ss in results[model_name]['results'][bs]
-                }
-                writer.writerow({'model': model_name, **model_results})
-
-
-def _compute_pytorch(model_names, dictionary, average_over, device, torchscript, fp16):
-    for c, model_name in enumerate(model_names):
-        print(f"{c + 1} / {len(model_names)}")
-        config = AutoConfig.from_pretrained(model_name, torchscript=torchscript)
-        model = AutoModel.from_pretrained(model_name, config=config)
-        tokenizer = AutoTokenizer.from_pretrained(model_name)
-
-        tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False)
-
-        max_input_size = tokenizer.max_model_input_sizes[model_name]
-        batch_sizes = [1, 2, 4, 8]
-        slice_sizes = [8, 64, 128, 256, 512, 1024]
-
-        dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}}
-        dictionary[model_name]["results"] = {i: {} for i in batch_sizes}
-
-        for batch_size in batch_sizes:
-            if fp16:
-                model.half()
-            model.to(device)
-            model.eval()
-            for slice_size in slice_sizes:
-                if max_input_size is not None and slice_size > max_input_size:
-                    dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
-                else:
-                    sequence = torch.tensor(tokenized_sequence[:slice_size], device=device).repeat(batch_size, 1)
-                    try:
-                        if torchscript:
-                            print("Tracing model with sequence size", sequence.shape)
-                            inference = torch.jit.trace(model, sequence)
-                            inference(sequence)
-                        else:
-                            inference = model
-                            inference(sequence)
-
-                        print("Going through model with sequence of shape", sequence.shape)
-                        runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
-                        average_time = sum(runtimes)/float(len(runtimes)) / 3.0
-                        dictionary[model_name]["results"][batch_size][slice_size] = average_time
-                    except RuntimeError as e:
-                        print("Doesn't fit on GPU.", e)
-                        torch.cuda.empty_cache()
-                        dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
-    return dictionary
-
-
-def _compute_tensorflow(model_names, dictionary, average_over, amp):
-    for c, model_name in enumerate(model_names):
-        print(f"{c + 1} / {len(model_names)}")
-        config = AutoConfig.from_pretrained(model_name)
-        model = TFAutoModel.from_pretrained(model_name, config=config)
-        tokenizer = AutoTokenizer.from_pretrained(model_name)
-
-        tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False)
-
-        max_input_size = tokenizer.max_model_input_sizes[model_name]
-        batch_sizes = [1, 2, 4, 8]
-        slice_sizes = [8, 64, 128, 256, 512, 1024]
-
-        dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}}
-        dictionary[model_name]["results"] = {i: {} for i in batch_sizes}
-
-        print("Using model", model)
-
-        @tf.function
-        def inference(inputs):
-            return model(inputs)
-
-        for batch_size in batch_sizes:
-            for slice_size in slice_sizes:
-                if max_input_size is not None and slice_size > max_input_size:
-                    dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
-                else:
-                    sequence = tf.stack([tf.squeeze(tf.constant(tokenized_sequence[:slice_size])[None, :])] * batch_size)
-
-                    try:
-                        print("Going through model with sequence of shape", sequence.shape)
-                        # To make sure that the model is traced + that the tensors are on the appropriate device
-                        inference(sequence)
-
-                        runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
-                        average_time = sum(runtimes)/float(len(runtimes)) / 3.0
-                        dictionary[model_name]["results"][batch_size][slice_size] = average_time
-                    except tf.errors.ResourceExhaustedError as e:
-                        print("Doesn't fit on GPU.", e)
-                        torch.cuda.empty_cache()
-                        dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
-    return dictionary
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument("--models", required=False, type=str, default='all', help="Model checkpoints to be provided "
-                                                                                  "to the AutoModel classes. Leave "
-                                                                                  "blank to benchmark the base version "
-                                                                                  "of all available model "
-                                                                                  "architectures.")
-    parser.add_argument("--torch", required=False, action="store_true", help="Benchmark the Pytorch version of the "
-                                                                             "models")
-    parser.add_argument("--torch_cuda", required=False, action="store_true", help="Pytorch only: run on available "
-                                                                                  "cuda devices")
-    parser.add_argument("--torchscript", required=False, action="store_true", help="Pytorch only: trace the models "
-                                                                                   "using torchscript")
-    parser.add_argument("--tensorflow", required=False, action="store_true", help="Benchmark the TensorFlow version "
-                                                                                  "of the models. Will run on GPU if "
-                                                                                  "the correct dependencies are "
-                                                                                  "installed")
-    parser.add_argument("--xla", required=False, action="store_true", help="TensorFlow only: use XLA acceleration.")
-    parser.add_argument("--amp", required=False, action="store_true", help="TensorFlow only: use automatic mixed precision acceleration.")
-    parser.add_argument("--fp16", required=False, action="store_true", help="PyTorch only: use FP16 to accelerate inference.")
-    parser.add_argument("--keras_predict", required=False, action="store_true", help="Whether to use model.predict "
-                                                                                     "instead of model() to do a "
-                                                                                     "forward pass.")
-    parser.add_argument("--save_to_csv", required=False, action="store_true", help="Save to a CSV file.")
-    parser.add_argument("--csv_filename", required=False, default=None, help="CSV filename used if saving results to csv.")
-    parser.add_argument("--average_over", required=False, default=30, type=int, help="Times an experiment will be run.")
-
-    args = parser.parse_args()
-    if args.models == 'all':
-        args.models = [
-            "gpt2",
-            "bert-base-cased",
-            "xlnet-base-cased",
-            "xlm-mlm-en-2048",
-            "transfo-xl-wt103",
-            "openai-gpt",
-            "distilbert-base-uncased",
-            "distilgpt2",
-            "roberta-base",
-            "ctrl"
-        ]
-    else:
-        args.models = args.models.split()
-
-    print("Running with arguments", args)
-
-    if args.torch:
-        if is_torch_available():
-            create_setup_and_compute(
-                model_names=args.models,
-                tensorflow=False,
-                gpu=args.torch_cuda,
-                torchscript=args.torchscript,
-                fp16=args.fp16,
-                save_to_csv=args.save_to_csv,
-                csv_filename=args.csv_filename,
-                average_over=args.average_over
-            )
-        else:
-            raise ImportError("Trying to run a PyTorch benchmark but PyTorch was not found in the environment.")
-
-    if args.tensorflow:
-        if is_tf_available():
-            create_setup_and_compute(
-                model_names=args.models,
-                tensorflow=True,
-                xla=args.xla,
-                amp=args.amp,
-                save_to_csv=args.save_to_csv,
-                csv_filename=args.csv_filename,
-                average_over=args.average_over
-            )
-        else:
-            raise ImportError("Trying to run a TensorFlow benchmark but TensorFlow was not found in the environment.")
-
-if __name__ == '__main__':
-    main()
-
@@ -1,48 +0,0 @@
-from pathlib import Path
-import tarfile
-import urllib.request
-
-import torch
-
-from transformers.tokenization_camembert import CamembertTokenizer
-from transformers.modeling_camembert import CamembertForMaskedLM
-
-
-def fill_mask(masked_input, model, tokenizer, topk=5):
-    # Adapted from https://github.com/pytorch/fairseq/blob/master/fairseq/models/roberta/hub_interface.py
-    assert masked_input.count('<mask>') == 1
-    input_ids = torch.tensor(tokenizer.encode(masked_input, add_special_tokens=True)).unsqueeze(0)  # Batch size 1
-    logits = model(input_ids)[0]  # The last hidden-state is the first element of the output tuple
-    masked_index = (input_ids.squeeze() == tokenizer.mask_token_id).nonzero().item()
-    logits = logits[0, masked_index, :]
-    prob = logits.softmax(dim=0)
-    values, indices = prob.topk(k=topk, dim=0)
-    topk_predicted_token_bpe = ' '.join([tokenizer.convert_ids_to_tokens(indices[i].item())
-                                         for i in range(len(indices))])
-    masked_token = tokenizer.mask_token
-    topk_filled_outputs = []
-    for index, predicted_token_bpe in enumerate(topk_predicted_token_bpe.split(' ')):
-        predicted_token = predicted_token_bpe.replace('\u2581', ' ')
-        if " {0}".format(masked_token) in masked_input:
-            topk_filled_outputs.append((
-                masked_input.replace(
-                    ' {0}'.format(masked_token), predicted_token
-                ),
-                values[index].item(),
-                predicted_token,
-            ))
-        else:
-            topk_filled_outputs.append((
-                masked_input.replace(masked_token, predicted_token),
-                values[index].item(),
-                predicted_token,
-            ))
-    return topk_filled_outputs
-
-
-tokenizer = CamembertTokenizer.from_pretrained('camembert-base')
-model = CamembertForMaskedLM.from_pretrained('camembert-base')
-model.eval()
-
-masked_input = "Le camembert est <mask> :)"
-print(fill_mask(masked_input, model, tokenizer, topk=3))
@@ -39,9 +39,9 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)

-from transformers import (OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer,
+from pytorch_transformers import (OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer,
                                     AdamW, cached_path, WEIGHTS_NAME, CONFIG_NAME,
-                                     get_linear_schedule_with_warmup)
+                                     WarmupLinearSchedule)

 ROCSTORIES_URL = "https://s3.amazonaws.com/datasets.huggingface.co/ROCStories.tar.gz"

@@ -211,7 +211,7 @@ def main():
            {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
            ]
        optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-        scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+        scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)

    if args.do_train:
        nb_tr_steps, tr_loss, exp_average_loss = 0, 0, None
@@ -237,7 +237,7 @@ def main():
    # Save a trained model
    if args.do_train:
        # Save a trained model, configuration and tokenizer
-        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model itself
+        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self

        # If we save using the predefined names, we can load using `from_pretrained`
        output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
@@ -31,18 +31,14 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-
-try:
-    from torch.utils.tensorboard import SummaryWriter
-except:
-    from tensorboardX import SummaryWriter
-
 from tqdm import tqdm, trange

-from transformers import (WEIGHTS_NAME, BertConfig,
+from tensorboardX import SummaryWriter
+
+from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
                                  BertForMultipleChoice, BertTokenizer)

-from transformers import AdamW, get_linear_schedule_with_warmup
+from pytorch_transformers import AdamW, WarmupLinearSchedule

 logger = logging.getLogger(__name__)

@@ -322,7 +318,7 @@ def train(args, train_dataset, model, tokenizer):
        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
    if args.fp16:
        try:
            from apex import amp
@@ -369,7 +365,7 @@ def train(args, train_dataset, model, tokenizer):
            #     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)
+            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)

            if args.n_gpu > 1:
                loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
@@ -651,7 +647,7 @@ def main():

        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
+            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs

        logger.info("Evaluate the following checkpoints: %s", checkpoints)

@@ -28,7 +28,7 @@ import math

 import torch

-from transformers import TransfoXLLMHeadModel, TransfoXLCorpus, TransfoXLTokenizer
+from pytorch_transformers import TransfoXLLMHeadModel, TransfoXLCorpus, TransfoXLTokenizer

 logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                    datefmt = '%m/%d/%Y %H:%M:%S',
@@ -1,57 +1,26 @@
-# Distil*
+# DistilBERT

-This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT, DistilRoBERTa and DistilGPT2.
+This folder contains the original code used to train DistilBERT as well as examples showcasing how to use DistilBERT.

-**October 23rd, 2019 - Update** We release **DistilRoBERTa**: 95% of `RoBERTa-base`'s performance on GLUE, twice as fast as RoBERTa while being 35% smaller.
+## What is DistilBERT

-**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.**
+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.

-**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!
-
-
-## 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 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).
-
-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.
+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
+).

 ## 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).
-
+**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/pytorch-transformers/issues/1179) for more details.

 ## How to use DistilBERT

-Transformers includes two pre-trained Distil* models, currently only provided for English (we are investigating the possibility to train and release a multilingual version of 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):

 - `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. 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! 🤗🤗🤗
+- `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).

 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.

@@ -64,14 +33,9 @@ outputs = model(input_ids)
 last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 ```

-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 DistilBERT

-
-## How to train Distil*
-
-In the following, we will explain how you can train DistilBERT.
+In the following, we will explain how you can train your own compressed model.

 ### A. Preparing the data

@@ -84,8 +48,7 @@ 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 \
-    --tokenizer_type bert \
-    --tokenizer_name bert-base-uncased \
+    --bert_tokenizer bert-base-uncased \
    --dump_file data/binarized_text
 ```

@@ -94,8 +57,7 @@ Our implementation of masked language modeling loss follows [XLM](https://github
 ```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 \
-    --vocab_size 30522
+    --token_counts_dump data/token_counts.bert-base-uncased.pickle
 ```

 ### B. Training
@@ -104,12 +66,6 @@ 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 \
@@ -118,7 +74,7 @@ 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
@@ -140,30 +96,11 @@ 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
+        --token_counts data/token_counts.bert-base-uncased.pickle \
+        --dump_path serialization_dir/my_first_distillation
 ```

-**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!
+**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!

 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}
-}
-```
@@ -12,33 +12,30 @@
 # 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
+""" Dataloaders to train DistilBERT
    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
 """
-import torch
-from torch.utils.data import Dataset
-
+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 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]]
-    """
-
+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.array([len(t) for t in data])
+        self.lengths = np.uint16([len(t) for t in data])

        self.check()
        self.remove_long_sequences()
@@ -46,9 +43,6 @@ class LmSeqsDataset(Dataset):
        self.check()
        self.print_statistics()

-    def __getitem__(self, index):
-        return (self.token_ids[index], self.lengths[index])
-
    def __len__(self):
        return len(self.lengths)

@@ -57,14 +51,12 @@ class LmSeqsDataset(Dataset):
        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.
+        Sequences that are too long are splitted by chunk of max_position_embeddings.
        """
-        max_len = self.params.max_model_input_size
-        indices = self.lengths > max_len
+        indices = self.lengths >= self.params.max_position_embeddings
        logger.info(f'Splitting {sum(indices)} too long sequences.')

        def divide_chunks(l, n):
@@ -72,13 +64,10 @@ class LmSeqsDataset(Dataset):

        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']
+        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):
-            assert (seq_[0] == cls_id) and (seq_[-1] == sep_id), seq_
            if len_ <= max_len:
                new_tok_ids.append(seq_)
                new_lengths.append(len_)
@@ -90,7 +79,6 @@ class LmSeqsDataset(Dataset):
                    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)
@@ -104,11 +92,11 @@ class LmSeqsDataset(Dataset):
        Too short sequences are simply removed. This could be tunedd.
        """
        init_size = len(self)
-        indices = self.lengths > 11
+        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 (<=11 tokens) sequences.')
+        logger.info(f'Remove {init_size - new_size} too short (<=5 tokens) sequences.')

    def print_statistics(self):
        """
@@ -125,27 +113,89 @@ class LmSeqsDataset(Dataset):
        # 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,
-                        batch):
+                        token_ids: List[List[int]],
+                        lengths: List[int]):
        """
        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']
+        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)  # (bs)
+        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)
@@ -12,39 +12,29 @@
 # 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 the student.
-    Adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
+""" The distiller to distil DistilBERT
+    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
 """
 import os
 import math
 import psutil
-import time
+from tensorboardX import SummaryWriter
 from tqdm import trange, tqdm
 import numpy as np
-import psutil

 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-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 get_linear_schedule_with_warmup
+from pytorch_transformers import AdamW, WarmupLinearSchedule

 from utils import logger
-from lm_seqs_dataset import LmSeqsDataset
-from grouped_batch_sampler import GroupedBatchSampler, create_lengths_groups
+from dataset import Dataset

 class Distiller:
    def __init__(self,
                 params: dict,
-                 dataset: LmSeqsDataset,
+                 dataloader: Dataset,
                 token_probs: torch.tensor,
                 student: nn.Module,
                 teacher: nn.Module):
@@ -57,47 +47,31 @@ class Distiller:
        self.student = student
        self.teacher = teacher

-        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.dataloader = dataloader
+        if self.params.n_gpu > 1:
+            self.dataloader.split()
+        self.get_iterator(seed=params.seed)

        self.temperature = params.temperature
        assert self.temperature > 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
-        self.alpha_cos = params.alpha_cos
+        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.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.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.epoch = 0
        self.n_iter = 0
@@ -107,22 +81,17 @@ class Distiller:
        self.last_loss = 0
        self.last_loss_ce = 0
        self.last_loss_mlm = 0
-        self.last_loss_clm = 0
-        if self.alpha_mse > 0.: self.last_loss_mse = 0
-        if self.alpha_cos > 0.: self.last_loss_cos = 0
-        self.last_log = 0
+        self.last_loss_mse = 0

        self.ce_loss_fct = nn.KLDivLoss(reduction='batchmean')
-        self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
-        if self.alpha_mse > 0.:
-            self.mse_loss_fct = nn.MSELoss(reduction='sum')
-        if self.alpha_cos > 0.:
-            self.cosine_loss_fct = nn.CosineEmbeddingLoss(reduction='mean')
+        self.mlm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
+        self.mse_loss_fct = nn.MSELoss(reduction='sum')

        logger.info('--- Initializing model optimizer')
        assert params.gradient_accumulation_steps >= 1
-        self.num_steps_epoch = len(self.dataloader)
+        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)

        no_decay = ['bias', 'LayerNorm.weight']
        optimizer_grouped_parameters = [
@@ -135,11 +104,9 @@ class Distiller:
                               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)
+        self.scheduler = WarmupLinearSchedule(self.optimizer,
+                                              warmup_steps=warmup_steps,
+                                              t_total=num_train_optimization_steps)

        if self.fp16:
            try:
@@ -162,18 +129,43 @@ class Distiller:
                logger.info("Using nn.parallel.DistributedDataParallel for distributed training.")
                self.student = DistributedDataParallel(self.student,
                                                       device_ids=[params.local_rank],
-                                                       output_device=params.local_rank,
-                                                       find_unused_parameters=True)
+                                                       output_device=params.local_rank)

        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/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)
+            self.tensorboard.add_text(tag='config', text_string=str(self.params), global_step=0)

-    def prepare_batch_mlm(self,
-                          batch):
+    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):
        """
        Prepare the batch: from the token_ids and the lenghts, compute the attention mask and the masked label for MLM.

@@ -219,7 +211,7 @@ class Distiller:
                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.vocab_size)
+        _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'])
        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()
@@ -227,41 +219,8 @@ class Distiller:

        mlm_labels[~pred_mask] = -1 # 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 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 -1 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] = -1 # 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):
@@ -299,10 +258,7 @@ class Distiller:
        if ml1 % 8 != 0:
            pad = 8 - (ml1 % 8)
            ml2 = ml1 + pad
-            if self.mlm:
-                pad_id = self.params.special_tok_ids['pad_token']
-            else:
-                pad_id = self.params.special_tok_ids['unk_token']
+            pad_id = self.params.special_tok_ids['pad_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)
@@ -316,25 +272,20 @@ class Distiller:
        The real training loop.
        """
        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.multi_gpu:
-                torch.distributed.barrier()

-            iter_bar = tqdm(self.dataloader, desc="-Iter", disable=self.params.local_rank not in [-1, 0])
-            for batch in iter_bar:
+            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()
                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)

-                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)
+                self.step(input_ids=token_ids, attention_mask=attn_mask, mlm_labels=mlm_labels)

                iter_bar.update()
                iter_bar.set_postfix({'Last_loss': f'{self.last_loss:.2f}',
@@ -352,7 +303,7 @@ class Distiller:
    def step(self,
             input_ids: torch.tensor,
             attention_mask: torch.tensor,
-             lm_labels: torch.tensor):
+             mlm_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).
@@ -361,22 +312,17 @@ 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.
-        lm_labels: `torch.tensor(bs, seq_length)` - The language modeling labels (mlm labels for MLM and clm labels for CLM).
+        mlm_labels: `torch.tensor(bs, seq_length)` - The masked language modeling labels.
        """
-        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)
+        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)
        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
        if self.params.restrict_ce_to_mask:
-            mask = (lm_labels>-1).unsqueeze(-1).expand_as(s_logits)    # (bs, seq_lenth, voc_size)
+            mask = (mlm_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
@@ -388,47 +334,20 @@ class Distiller:
        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.lm_loss_fct(s_logits.view(-1, s_logits.size(-1)), lm_labels.view(-1))
+            loss_mlm = self.mlm_loss_fct(s_logits.view(-1, s_logits.size(-1)), mlm_labels.view(-1))
            loss += self.alpha_mlm * loss_mlm
-        if self.alpha_clm > 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.:
            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.:
-            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.:
            self.last_loss_mlm = loss_mlm.item()
-        if self.alpha_clm > 0.:
-            self.last_loss_clm = loss_clm.item()
        if self.alpha_mse > 0.:
            self.last_loss_mse = loss_mse.item()
-        if self.alpha_cos > 0.:
-            self.last_loss_cos = loss_cos.item()

        self.optimize(loss)

@@ -477,7 +396,6 @@ 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()

@@ -501,16 +419,11 @@ class Distiller:
        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_clm > 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.:
            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.:
-            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):
        """
@@ -1,105 +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.
-""" 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
@@ -2,5 +2,3 @@ gitpython==3.0.2
 tensorboard>=1.14.0
 tensorboardX==1.8
 psutil==5.6.3
-scipy==1.3.1
-transformers==2.0.0
@@ -1,600 +0,0 @@
-# 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 2019, October 4th) with an additional and optional step of distillation."""
-
-from __future__ import absolute_import, division, print_function
-
-import argparse
-import logging
-import os
-import random
-import glob
-
-import numpy as np
-import torch
-from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
-                              TensorDataset)
-from torch.utils.data.distributed import DistributedSampler
-import torch.nn.functional as F
-import torch.nn as nn
-
-try:
-    from torch.utils.tensorboard import SummaryWriter
-except:
-    from tensorboardX import SummaryWriter
-
-from tqdm import tqdm, trange
-
-from transformers import (WEIGHTS_NAME, BertConfig,
-                                  BertForQuestionAnswering, BertTokenizer,
-                                  XLMConfig, XLMForQuestionAnswering,
-                                  XLMTokenizer, XLNetConfig,
-                                  XLNetForQuestionAnswering,
-                                  XLNetTokenizer,
-                                  DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
-
-from transformers import AdamW, get_linear_schedule_with_warmup
-
-from ..utils_squad import (read_squad_examples, convert_examples_to_features,
-                         RawResult, write_predictions,
-                         RawResultExtended, write_predictions_extended)
-
-# The follwing import is the official SQuAD evaluation script (2.0).
-# You can remove it from the dependencies if you are using this script outside of the library
-# We've added it here for automated tests (see examples/test_examples.py file)
-from ..utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad
-
-logger = logging.getLogger(__name__)
-
-ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
-                  for conf in (BertConfig, XLNetConfig, XLMConfig)), ())
-
-MODEL_CLASSES = {
-    'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
-    'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
-    'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
-    'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
-}
-
-def set_seed(args):
-    random.seed(args.seed)
-    np.random.seed(args.seed)
-    torch.manual_seed(args.seed)
-    if args.n_gpu > 0:
-        torch.cuda.manual_seed_all(args.seed)
-
-def to_list(tensor):
-    return tensor.detach().cpu().tolist()
-
-def train(args, train_dataset, model, tokenizer, 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)
-    if args.fp16:
-        try:
-            from apex import amp
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
-        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
-
-    # multi-gpu training (should be after apex fp16 initialization)
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    # Distributed training (should be after apex fp16 initialization)
-    if args.local_rank != -1:
-        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
-                                                          output_device=args.local_rank,
-                                                          find_unused_parameters=True)
-
-    # Train!
-    logger.info("***** Running training *****")
-    logger.info("  Num examples = %d", len(train_dataset))
-    logger.info("  Num Epochs = %d", args.num_train_epochs)
-    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
-    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
-                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
-    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
-    logger.info("  Total optimization steps = %d", t_total)
-
-    global_step = 0
-    tr_loss, logging_loss = 0.0, 0.0
-    model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
-    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
-    for _ in train_iterator:
-        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
-        for step, batch in enumerate(epoch_iterator):
-            model.train()
-            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]})
-            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.
-
-                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()
-                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)
-                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
-                    logger.info("Saving model checkpoint to %s", output_dir)
-
-            if args.max_steps > 0 and global_step > args.max_steps:
-                epoch_iterator.close()
-                break
-        if args.max_steps > 0 and global_step > args.max_steps:
-            train_iterator.close()
-            break
-
-    if args.local_rank in [-1, 0]:
-        tb_writer.close()
-
-    return global_step, tr_loss / global_step
-
-
-def evaluate(args, model, tokenizer, prefix=""):
-    dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
-
-    if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-        os.makedirs(args.output_dir)
-
-    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
-    # Note that DistributedSampler samples randomly
-    eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
-    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
-
-    # Eval!
-    logger.info("***** Running evaluation {} *****".format(prefix))
-    logger.info("  Num examples = %d", len(dataset))
-    logger.info("  Batch size = %d", args.eval_batch_size)
-    all_results = []
-    for batch in tqdm(eval_dataloader, desc="Evaluating"):
-        model.eval()
-        batch = tuple(t.to(args.device) for t in batch)
-        with torch.no_grad():
-            inputs = {'input_ids':      batch[0],
-                      'attention_mask': batch[1]
-                      }
-            if args.model_type != 'distilbert':
-                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]  # XLM don't use segment_ids
-            example_indices = batch[3]
-            if args.model_type in ['xlnet', 'xlm']:
-                inputs.update({'cls_index': batch[4],
-                               'p_mask':    batch[5]})
-            outputs = model(**inputs)
-
-        for i, example_index in enumerate(example_indices):
-            eval_feature = features[example_index.item()]
-            unique_id = int(eval_feature.unique_id)
-            if args.model_type in ['xlnet', 'xlm']:
-                # XLNet uses a more complex post-processing procedure
-                result = RawResultExtended(unique_id            = unique_id,
-                                           start_top_log_probs  = to_list(outputs[0][i]),
-                                           start_top_index      = to_list(outputs[1][i]),
-                                           end_top_log_probs    = to_list(outputs[2][i]),
-                                           end_top_index        = to_list(outputs[3][i]),
-                                           cls_logits           = to_list(outputs[4][i]))
-            else:
-                result = RawResult(unique_id    = unique_id,
-                                   start_logits = to_list(outputs[0][i]),
-                                   end_logits   = to_list(outputs[1][i]))
-            all_results.append(result)
-
-    # Compute predictions
-    output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
-    output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
-    if args.version_2_with_negative:
-        output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
-    else:
-        output_null_log_odds_file = None
-
-    if args.model_type in ['xlnet', 'xlm']:
-        # XLNet uses a more complex post-processing procedure
-        write_predictions_extended(examples, features, all_results, args.n_best_size,
-                        args.max_answer_length, output_prediction_file,
-                        output_nbest_file, output_null_log_odds_file, args.predict_file,
-                        model.config.start_n_top, model.config.end_n_top,
-                        args.version_2_with_negative, tokenizer, args.verbose_logging)
-    else:
-        write_predictions(examples, features, all_results, args.n_best_size,
-                        args.max_answer_length, args.do_lower_case, output_prediction_file,
-                        output_nbest_file, output_null_log_odds_file, args.verbose_logging,
-                        args.version_2_with_negative, args.null_score_diff_threshold)
-
-    # Evaluate with the official SQuAD script
-    evaluate_options = EVAL_OPTS(data_file=args.predict_file,
-                                 pred_file=output_prediction_file,
-                                 na_prob_file=output_null_log_odds_file)
-    results = evaluate_on_squad(evaluate_options)
-    return results
-
-
-def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
-    if args.local_rank not in [-1, 0] and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    # Load data features from cache or dataset file
-    input_file = args.predict_file if evaluate else args.train_file
-    cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
-        'dev' if evaluate else 'train',
-        list(filter(None, args.model_name_or_path.split('/'))).pop(),
-        str(args.max_seq_length)))
-    if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
-        logger.info("Loading features from cached file %s", cached_features_file)
-        features = torch.load(cached_features_file)
-    else:
-        logger.info("Creating features from dataset file at %s", input_file)
-        examples = read_squad_examples(input_file=input_file,
-                                                is_training=not evaluate,
-                                                version_2_with_negative=args.version_2_with_negative)
-        features = convert_examples_to_features(examples=examples,
-                                                tokenizer=tokenizer,
-                                                max_seq_length=args.max_seq_length,
-                                                doc_stride=args.doc_stride,
-                                                max_query_length=args.max_query_length,
-                                                is_training=not evaluate)
-        if args.local_rank in [-1, 0]:
-            logger.info("Saving features into cached file %s", cached_features_file)
-            torch.save(features, cached_features_file)
-
-    if args.local_rank == 0 and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    # Convert to Tensors and build dataset
-    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
-    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
-    all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
-    all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
-    all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
-    if evaluate:
-        all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
-        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
-                                all_example_index, all_cls_index, all_p_mask)
-    else:
-        all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
-        all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
-        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
-                                all_start_positions, all_end_positions,
-                                all_cls_index, all_p_mask)
-
-    if output_examples:
-        return dataset, examples, features
-    return dataset
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    ## Required parameters
-    parser.add_argument("--train_file", default=None, type=str, required=True,
-                        help="SQuAD json for training. E.g., train-v1.1.json")
-    parser.add_argument("--predict_file", default=None, type=str, required=True,
-                        help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
-    parser.add_argument("--model_type", default=None, type=str, required=True,
-                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
-    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
-                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
-    parser.add_argument("--output_dir", default=None, type=str, required=True,
-                        help="The output directory where the model checkpoints and predictions will be written.")
-
-    # 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("--config_name", default="", type=str,
-                        help="Pretrained config name or path if not the same as model_name")
-    parser.add_argument("--tokenizer_name", default="", type=str,
-                        help="Pretrained tokenizer name or path if not the same as model_name")
-    parser.add_argument("--cache_dir", default="", type=str,
-                        help="Where do you want to store the pre-trained models downloaded from s3")
-
-    parser.add_argument('--version_2_with_negative', action='store_true',
-                        help='If true, the SQuAD examples contain some that do not have an answer.')
-    parser.add_argument('--null_score_diff_threshold', type=float, default=0.0,
-                        help="If null_score - best_non_null is greater than the threshold predict null.")
-
-    parser.add_argument("--max_seq_length", default=384, type=int,
-                        help="The maximum total input sequence length after WordPiece tokenization. Sequences "
-                             "longer than this will be truncated, and sequences shorter than this will be padded.")
-    parser.add_argument("--doc_stride", default=128, type=int,
-                        help="When splitting up a long document into chunks, how much stride to take between chunks.")
-    parser.add_argument("--max_query_length", default=64, type=int,
-                        help="The maximum number of tokens for the question. Questions longer than this will "
-                             "be truncated to this length.")
-    parser.add_argument("--do_train", action='store_true',
-                        help="Whether to run training.")
-    parser.add_argument("--do_eval", action='store_true',
-                        help="Whether to run eval on the dev set.")
-    parser.add_argument("--evaluate_during_training", action='store_true',
-                        help="Rul evaluation during training at each logging step.")
-    parser.add_argument("--do_lower_case", action='store_true',
-                        help="Set this flag if you are using an uncased model.")
-
-    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
-                        help="Batch size per GPU/CPU for training.")
-    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
-                        help="Batch size per GPU/CPU for evaluation.")
-    parser.add_argument("--learning_rate", default=5e-5, type=float,
-                        help="The initial learning rate for Adam.")
-    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
-                        help="Number of updates steps to accumulate before performing a backward/update pass.")
-    parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight deay if we apply some.")
-    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
-                        help="Epsilon for Adam optimizer.")
-    parser.add_argument("--max_grad_norm", default=1.0, type=float,
-                        help="Max gradient norm.")
-    parser.add_argument("--num_train_epochs", default=3.0, type=float,
-                        help="Total number of training epochs to perform.")
-    parser.add_argument("--max_steps", default=-1, type=int,
-                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
-    parser.add_argument("--warmup_steps", default=0, type=int,
-                        help="Linear warmup over warmup_steps.")
-    parser.add_argument("--n_best_size", default=20, type=int,
-                        help="The total number of n-best predictions to generate in the nbest_predictions.json output file.")
-    parser.add_argument("--max_answer_length", default=30, type=int,
-                        help="The maximum length of an answer that can be generated. This is needed because the start "
-                             "and end predictions are not conditioned on one another.")
-    parser.add_argument("--verbose_logging", action='store_true',
-                        help="If true, all of the warnings related to data processing will be printed. "
-                             "A number of warnings are expected for a normal SQuAD evaluation.")
-
-    parser.add_argument('--logging_steps', type=int, default=50,
-                        help="Log every X updates steps.")
-    parser.add_argument('--save_steps', type=int, default=50,
-                        help="Save checkpoint every X updates steps.")
-    parser.add_argument("--eval_all_checkpoints", action='store_true',
-                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
-    parser.add_argument("--no_cuda", action='store_true',
-                        help="Whether not to use CUDA when available")
-    parser.add_argument('--overwrite_output_dir', action='store_true',
-                        help="Overwrite the content of the output directory")
-    parser.add_argument('--overwrite_cache', action='store_true',
-                        help="Overwrite the cached training and evaluation sets")
-    parser.add_argument('--seed', type=int, default=42,
-                        help="random seed for initialization")
-
-    parser.add_argument("--local_rank", type=int, default=-1,
-                        help="local_rank for distributed training on gpus")
-    parser.add_argument('--fp16', action='store_true',
-                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
-    parser.add_argument('--fp16_opt_level', type=str, default='O1',
-                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
-                             "See details at https://nvidia.github.io/apex/amp.html")
-    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
-    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
-    args = parser.parse_args()
-
-    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
-        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
-
-    # Setup distant debugging if needed
-    if args.server_ip and args.server_port:
-        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
-        import ptvsd
-        print("Waiting for debugger attach")
-        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
-        ptvsd.wait_for_attach()
-
-    # Setup CUDA, GPU & distributed training
-    if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
-    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
-        torch.cuda.set_device(args.local_rank)
-        device = torch.device("cuda", args.local_rank)
-        torch.distributed.init_process_group(backend='nccl')
-        args.n_gpu = 1
-    args.device = device
-
-    # Setup logging
-    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
-                        datefmt = '%m/%d/%Y %H:%M:%S',
-                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
-    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
-                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
-
-    # Set seed
-    set_seed(args)
-
-    # Load pretrained model and tokenizer
-    if args.local_rank not in [-1, 0]:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    args.model_type = args.model_type.lower()
-    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
-                                          cache_dir=args.cache_dir if args.cache_dir else None)
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
-                                                do_lower_case=args.do_lower_case,
-                                                cache_dir=args.cache_dir if args.cache_dir else None)
-    model = model_class.from_pretrained(args.model_name_or_path,
-                                        from_tf=bool('.ckpt' in args.model_name_or_path),
-                                        config=config,
-                                        cache_dir=args.cache_dir if args.cache_dir else None)
-
-    if args.teacher_type is not None:
-        assert args.teacher_name_or_path is not None
-        assert args.alpha_ce > 0.
-        assert args.alpha_ce + args.alpha_squad > 0.
-        assert args.teacher_type != 'distilbert', "We constraint teachers not to be of type DistilBERT."
-        teacher_config_class, teacher_model_class, _ = MODEL_CLASSES[args.teacher_type]
-        teacher_config = teacher_config_class.from_pretrained(args.teacher_name_or_path,
-                                                              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:
-        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, 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, cache_dir=args.cache_dir if args.cache_dir else None)
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir,
-                                                    do_lower_case=args.do_lower_case,
-                                                    cache_dir=args.cache_dir if args.cache_dir else None)
-        model.to(args.device)
-
-
-    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
-    results = {}
-    if args.do_eval and args.local_rank in [-1, 0]:
-        checkpoints = [args.output_dir]
-        if args.eval_all_checkpoints:
-            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
-            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs
-
-        logger.info("Evaluate the following checkpoints: %s", checkpoints)
-
-        for checkpoint in checkpoints:
-            # Reload the model
-            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
-            model = model_class.from_pretrained(checkpoint, cache_dir=args.cache_dir if args.cache_dir else None)
-            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()
@@ -13,14 +13,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-Preprocessing script before distillation.
+Preprocessing script before training DistilBERT.
 """
 import argparse
 import pickle
 import random
 import time
 import numpy as np
-from transformers import BertTokenizer, RobertaTokenizer, GPT2Tokenizer
+from pytorch_transformers import BertTokenizer
 import logging

 logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
@@ -32,27 +32,16 @@ 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('--tokenizer_type', type=str, default='bert', choices=['bert', 'roberta', 'gpt2'])
-    parser.add_argument('--tokenizer_name', type=str, default='bert-base-uncased',
+    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.')
    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:
@@ -67,8 +56,8 @@ def main():
    interval = 10000
    start = time.time()
    for text in data:
-        text = f'{bos} {text.strip()} {sep}'
-        token_ids = tokenizer.encode(text, add_special_tokens=False)
+        text = f'[CLS] {text.strip()} [SEP]'
+        token_ids = bert_tokenizer.encode(text)
        rslt.append(token_ids)

        iter += 1
@@ -80,7 +69,7 @@ def main():
    logger.info(f'{len(data)} examples processed.')


-    dp_file = f'{args.dump_file}.{args.tokenizer_name}.pickle'
+    dp_file = f'{args.dump_file}.{args.bert_tokenizer}.pickle'
    rslt_ = [np.uint16(d) for d in rslt]
    random.shuffle(rslt_)
    logger.info(f'Dump to {dp_file}')
@@ -1,89 +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 the distilled model.
-Specific to RoBERTa -> DistilRoBERTa and GPT2 -> DistilGPT2.
-"""
-from transformers import BertForMaskedLM, RobertaForMaskedLM, GPT2LMHeadModel
-import torch
-import argparse
-
-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)
@@ -14,58 +14,52 @@
 # limitations under the License.
 """
 Preprocessing script before training DistilBERT.
-Specific to BERT -> DistilBERT.
 """
-from transformers import BertForMaskedLM, RobertaForMaskedLM
+from pytorch_transformers import BertForPreTraining
 import torch
 import argparse

 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 = 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()


-    if args.model_type == 'bert':
-        model = BertForMaskedLM.from_pretrained(args.model_name)
-        prefix = 'bert'
-    else:
-        raise ValueError(f'args.model_type should be "bert".')
+    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'{prefix}.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'{prefix}.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'{prefix}.encoder.layer.{teacher_idx}.attention.self.query.{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'{prefix}.encoder.layer.{teacher_idx}.attention.self.key.{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'{prefix}.encoder.layer.{teacher_idx}.attention.self.value.{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'{prefix}.encoder.layer.{teacher_idx}.attention.output.dense.{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'{prefix}.encoder.layer.{teacher_idx}.attention.output.LayerNorm.{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'{prefix}.encoder.layer.{teacher_idx}.intermediate.dense.{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'{prefix}.encoder.layer.{teacher_idx}.output.dense.{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'{prefix}.encoder.layer.{teacher_idx}.output.LayerNorm.{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']
@@ -13,7 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-Preprocessing script before training the distilled model.
+Preprocessing script before training DistilBERT.
 """
 from collections import Counter
 import argparse
@@ -13,8 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-Training the distilled model.
-Supported architectures include: BERT -> DistilBERT, RoBERTa -> DistilRoBERTa, GPT2 -> DistilGPT2.
+Training DistilBERT.
 """
 import os
 import argparse
@@ -24,96 +23,64 @@ import shutil
 import numpy as np
 import torch

-from transformers import BertConfig, BertForMaskedLM, BertTokenizer
-from transformers import RobertaConfig, RobertaForMaskedLM, RobertaTokenizer
-from transformers import DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer
-from transformers import GPT2Config, GPT2LMHeadModel, GPT2Tokenizer
+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 lm_seqs_dataset import LmSeqsDataset
+from dataset import Dataset


-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.) or (not args.mlm and args.alpha_mlm == 0.)
-    assert (args.alpha_mlm > 0. and args.alpha_clm == 0.) or (args.alpha_mlm == 0. and args.alpha_clm > 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.
-    assert args.alpha_mlm >= 0.
-    assert args.alpha_clm >= 0.
-    assert args.alpha_mse >= 0.
-    assert args.alpha_cos >= 0.
-    assert args.alpha_ce + args.alpha_mlm + args.alpha_clm + args.alpha_mse + args.alpha_cos > 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("--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,
+    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("--from_pretrained_weights", default=None, type=str,
                        help="Load student initialization checkpoint.")
-
-    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("--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("--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=0.0, 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_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("--alpha_cos", default=0.0, type=float,
-                        help="Linear weight of the cosine embedding loss. Must be >=0.")
-
-    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,
@@ -124,20 +91,17 @@ def main():
                        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("--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("--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.")

@@ -173,7 +137,6 @@ def main():
    parser.add_argument("--checkpoint_interval", type=int, default=4000,
                        help="Checkpoint interval.")
    args = parser.parse_args()
-    sanity_checks(args)


    ## ARGS ##
@@ -197,19 +160,18 @@ def main():
        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 ###
-    tokenizer = teacher_tokenizer_class.from_pretrained(args.teacher_name)
+    bert_tokenizer = BertTokenizer.from_pretrained(args.bert_model)
    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]
+    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}')
    args.special_tok_ids = special_tok_ids
-    args.max_model_input_size = tokenizer.max_model_input_sizes[args.teacher_name]


    ## DATA LOADER ##
@@ -218,34 +180,34 @@ def main():
        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)
-        
-        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)
-    else:
-        token_probs = None
+    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_lm_seq_dataset = LmSeqsDataset(params=args, data=data)
+    train_dataloader = Dataset(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)
+    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)
    else:
-        student = student_model_class(stu_architecture_config)
+        args.vocab_size_or_config_json_file = args.vocab_size
+        stu_architecture_config = DistilBertConfig(**vars(args))
+        student = DistilBertForMaskedLM(stu_architecture_config)


    if args.n_gpu > 0:
@@ -254,31 +216,15 @@ def main():


    ## TEACHER ##
-    teacher = teacher_model_class.from_pretrained(args.teacher_name, output_hidden_states=True)
+    teacher = BertForMaskedLM.from_pretrained(args.bert_model)
    if args.n_gpu > 0:
        teacher.to(f'cuda:{args.local_rank}')
-    logger.info(f'Teacher loaded from {args.teacher_name}.')
-
-
-    ## 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
-
+    logger.info(f'Teacher loaded from {args.bert_model}.')

    ## DISTILLER ##
    torch.cuda.empty_cache()
    distiller = Distiller(params=args,
-                          dataset=train_lm_seq_dataset,
+                          dataloader=train_dataloader,
                          token_probs=token_probs,
                          student=student,
                          teacher=teacher)
@@ -1,15 +0,0 @@
-{
-	"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
-  }
-  
@@ -1,10 +0,0 @@
-{
-	"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
-}
@@ -1,4 +1,2 @@
 tensorboardX
-tensorboard
-scikit-learn
-seqeval
+scikit-learn
@@ -32,16 +32,15 @@ from torch.utils.data import DataLoader, SequentialSampler, TensorDataset, Subse
 from torch.utils.data.distributed import DistributedSampler
 from torch.nn import CrossEntropyLoss, MSELoss

-from transformers import (WEIGHTS_NAME,
+from pytorch_transformers import (WEIGHTS_NAME,
                                  BertConfig, BertForSequenceClassification, BertTokenizer,
                                  XLMConfig, XLMForSequenceClassification, XLMTokenizer,
                                  XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer)

 from run_glue import set_seed, load_and_cache_examples, ALL_MODELS, MODEL_CLASSES

-from transformers import glue_compute_metrics as compute_metrics
-from transformers import glue_output_modes as output_modes
-from transformers import glue_processors as processors
+from utils_glue import (compute_metrics, convert_examples_to_features,
+                        output_modes, processors)

 logger = logging.getLogger(__name__)

@@ -234,8 +233,6 @@ def main():
                        help="If > 0: limit the data to a subset of data_subset instances.")
    parser.add_argument("--overwrite_output_dir", action='store_true',
                        help="Whether to overwrite data in output directory")
-    parser.add_argument('--overwrite_cache', action='store_true',
-                        help="Overwrite the cached training and evaluation sets")

    parser.add_argument("--dont_normalize_importance_by_layer", action='store_true',
                        help="Don't normalize importance score by layers")
@@ -307,16 +304,10 @@ def main():
            break
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
-                                          num_labels=num_labels,
-                                          finetuning_task=args.task_name,
-                                          output_attentions=True,
-                                          cache_dir=args.cache_dir if args.cache_dir else None)
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
-                                                cache_dir=args.cache_dir if args.cache_dir else None)
-    model = model_class.from_pretrained(args.model_name_or_path,
-                                        from_tf=bool('.ckpt' in args.model_name_or_path),
-                                        config=config,
-                                        cache_dir=args.cache_dir if args.cache_dir else None)
+                                          num_labels=num_labels, finetuning_task=args.task_name,
+                                          output_attentions=True)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path)
+    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)

    if args.local_rank == 0:
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
@@ -14,7 +14,7 @@
 # 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.
-""" Conditional text generation with the auto-regressive models of the library (GPT/GPT-2/CTRL/Transformer-XL/XLNet)
+""" Conditional text generation with the auto-regressive models of the library (GPT/GPT-2/Transformer-XL/XLNet)
 """
 from __future__ import absolute_import, division, print_function, unicode_literals

@@ -26,14 +26,12 @@ import torch
 import torch.nn.functional as F
 import numpy as np

-from transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig
+from pytorch_transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig

-from transformers import GPT2LMHeadModel, GPT2Tokenizer
-from transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer
-from transformers import XLNetLMHeadModel, XLNetTokenizer
-from transformers import TransfoXLLMHeadModel, TransfoXLTokenizer
-from transformers import CTRLLMHeadModel, CTRLTokenizer
-from transformers import XLMWithLMHeadModel, XLMTokenizer
+from pytorch_transformers import GPT2LMHeadModel, GPT2Tokenizer
+from pytorch_transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer
+from pytorch_transformers import XLNetLMHeadModel, XLNetTokenizer
+from pytorch_transformers import TransfoXLLMHeadModel, TransfoXLTokenizer


 logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
@@ -43,15 +41,13 @@ logger = logging.getLogger(__name__)

 MAX_LENGTH = int(10000)  # Hardcoded max length to avoid infinite loop

-ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig)), ())
+ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig)), ())

 MODEL_CLASSES = {
    'gpt2': (GPT2LMHeadModel, GPT2Tokenizer),
-    'ctrl': (CTRLLMHeadModel, CTRLTokenizer),
    'openai-gpt': (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
    'xlnet': (XLNetLMHeadModel, XLNetTokenizer),
    'transfo-xl': (TransfoXLLMHeadModel, TransfoXLTokenizer),
-    'xlm': (XLMWithLMHeadModel, XLMTokenizer),
 }

 # Padding text to help Transformer-XL and XLNet with short prompts as proposed by Aman Rusia
@@ -79,12 +75,13 @@ def set_seed(args):
 def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
        Args:
-            logits: logits distribution shape (batch size x vocabulary size)
+            logits: logits distribution shape (vocabulary size)
            top_k > 0: keep only top k tokens with highest probability (top-k filtering).
            top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
    """
+    assert logits.dim() == 1  # batch size 1 for now - could be updated for more but the code would be less clear
    top_k = min(top_k, logits.size(-1))  # Safety check
    if top_k > 0:
        # Remove all tokens with a probability less than the last token of the top-k
@@ -101,14 +98,12 @@ def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')
        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
        sorted_indices_to_remove[..., 0] = 0

-        # scatter sorted tensors to original indexing
-        indices_to_remove = sorted_indices_to_remove.scatter(dim=1, index=sorted_indices, src=sorted_indices_to_remove)
+        indices_to_remove = sorted_indices[sorted_indices_to_remove]
        logits[indices_to_remove] = filter_value
    return logits


-def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0,
-                    is_xlnet=False, is_xlm_mlm=False, xlm_mask_token=None, xlm_lang=None, device='cpu'):
+def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, is_xlnet=False, device='cpu'):
    context = torch.tensor(context, dtype=torch.long, device=device)
    context = context.unsqueeze(0).repeat(num_samples, 1)
    generated = context
@@ -126,29 +121,11 @@ def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=
                target_mapping[0, 0, -1] = 1.0  # predict last token
                inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping}

-            if is_xlm_mlm and xlm_mask_token:
-                # XLM MLM models are direct models (predict same token, not next token)
-                # => need one additional dummy token in the input (will be masked and guessed)
-                input_ids = torch.cat((generated, torch.full((1, 1), xlm_mask_token, dtype=torch.long, device=device)), dim=1)
-                inputs = {'input_ids': input_ids}
-
-            if xlm_lang is not None:
-                inputs["langs"] = torch.tensor([xlm_lang] * inputs["input_ids"].shape[1], device=device).view(1, -1)
-
-            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states)
-            next_token_logits = outputs[0][:, -1, :] / (temperature if temperature > 0 else 1.)
-
-            # repetition penalty from CTRL (https://arxiv.org/abs/1909.05858)
-            for i in range(num_samples):
-                for _ in set(generated[i].tolist()):
-                    next_token_logits[i, _] /= repetition_penalty
-                
+            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet (cached hidden-states)
+            next_token_logits = outputs[0][0, -1, :] / temperature
            filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
-            if temperature == 0: # greedy sampling:
-                next_token = torch.argmax(filtered_logits, dim=-1).unsqueeze(-1)
-            else:
-                next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
-            generated = torch.cat((generated, next_token), dim=1)
+            next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
+            generated = torch.cat((generated, next_token.unsqueeze(0)), dim=1)
    return generated


@@ -160,21 +137,14 @@ def main():
                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
    parser.add_argument("--prompt", type=str, default="")
    parser.add_argument("--padding_text", type=str, default="")
-    parser.add_argument("--xlm_lang", type=str, default="", help="Optional language when used with the XLM model.")
    parser.add_argument("--length", type=int, default=20)
-    parser.add_argument("--num_samples", type=int, default=1)
-    parser.add_argument("--temperature", type=float, default=1.0,
-                        help="temperature of 0 implies greedy sampling")
-    parser.add_argument("--repetition_penalty", type=float, default=1.0,
-                        help="primarily useful for CTRL model; in that case, use 1.2")
+    parser.add_argument("--temperature", type=float, default=1.0)
    parser.add_argument("--top_k", type=int, default=0)
    parser.add_argument("--top_p", type=float, default=0.9)
    parser.add_argument("--no_cuda", action='store_true',
                        help="Avoid using CUDA when available")
    parser.add_argument('--seed', type=int, default=42,
                        help="random seed for initialization")
-    parser.add_argument('--stop_token', type=str, default=None,
-                        help="Token at which text generation is stopped")
    args = parser.parse_args()

    args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
@@ -196,61 +166,26 @@ def main():
    elif args.length < 0:
        args.length = MAX_LENGTH  # avoid infinite loop

-    logger.info(args)
-    if args.model_type in ["ctrl"]:
-        if args.temperature > 0.7:
-            logger.info('CTRL typically works better with lower temperatures (and lower top_k).')
-
+    print(args)
    while True:
-        xlm_lang = None
-        # XLM Language usage detailed in the issues #1414
-        if args.model_type in ["xlm"] and hasattr(tokenizer, 'lang2id') and hasattr(model.config, 'use_lang_emb') \
-                and model.config.use_lang_emb:
-            if args.xlm_lang:
-                language = args.xlm_lang
-            else:
-                language = None
-                while language not in tokenizer.lang2id.keys():
-                    language = input("Using XLM. Select language in " + str(list(tokenizer.lang2id.keys())) + " >>> ")
-            xlm_lang = tokenizer.lang2id[language]
-
-        # XLM masked-language modeling (MLM) models need masked token (see details in sample_sequence)
-        is_xlm_mlm = args.model_type in ["xlm"] and 'mlm' in args.model_name_or_path
-        if is_xlm_mlm:
-            xlm_mask_token = tokenizer.mask_token_id
-        else:
-            xlm_mask_token = None
-
        raw_text = args.prompt if args.prompt else input("Model prompt >>> ")
        if args.model_type in ["transfo-xl", "xlnet"]:
            # Models with memory likes to have a long prompt for short inputs.
            raw_text = (args.padding_text if args.padding_text else PADDING_TEXT) + raw_text
-        context_tokens = tokenizer.encode(raw_text, add_special_tokens=False)
-        if args.model_type == "ctrl":
-            if not any(context_tokens[0] == x for x in tokenizer.control_codes.values()):
-                logger.info("WARNING! You are not starting your generation from a control code so you won't get good results")
+        context_tokens = tokenizer.encode(raw_text)
        out = sample_sequence(
            model=model,
            context=context_tokens,
-            num_samples=args.num_samples,
            length=args.length,
            temperature=args.temperature,
            top_k=args.top_k,
            top_p=args.top_p,
-            repetition_penalty=args.repetition_penalty,
-            is_xlnet=bool(args.model_type == "xlnet"),
-            is_xlm_mlm=is_xlm_mlm,
-            xlm_mask_token=xlm_mask_token,
-            xlm_lang=xlm_lang,
            device=args.device,
+            is_xlnet=bool(args.model_type == "xlnet"),
        )
-        out = out[:, len(context_tokens):].tolist()
-        for o in out:
-            text = tokenizer.decode(o, clean_up_tokenization_spaces=True)
-            text = text[: text.find(args.stop_token) if args.stop_token else None]
-
-            print(text)
-
+        out = out[0, len(context_tokens):].tolist()
+        text = tokenizer.decode(out, clean_up_tokenization_spaces=True)
+        print(text)
        if args.prompt:
            break
    return text
@@ -28,15 +28,10 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-
-try:
-    from torch.utils.tensorboard import SummaryWriter
-except:
-    from tensorboardX import SummaryWriter
-
+from tensorboardX import SummaryWriter
 from tqdm import tqdm, trange

-from transformers import (WEIGHTS_NAME, BertConfig,
+from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
                                  BertForSequenceClassification, BertTokenizer,
                                  RobertaConfig,
                                  RobertaForSequenceClassification,
@@ -47,31 +42,25 @@ from transformers import (WEIGHTS_NAME, BertConfig,
                                  XLNetTokenizer,
                                  DistilBertConfig,
                                  DistilBertForSequenceClassification,
-                                  DistilBertTokenizer,
-                                  AlbertConfig,
-                                  AlbertForSequenceClassification, 
-                                  AlbertTokenizer,
-                                )
+                                  DistilBertTokenizer)

-from transformers import AdamW, get_linear_schedule_with_warmup
+from pytorch_transformers import AdamW, WarmupLinearSchedule

-from transformers import glue_compute_metrics as compute_metrics
-from transformers import glue_output_modes as output_modes
-from transformers import glue_processors as processors
-from transformers import glue_convert_examples_to_features as convert_examples_to_features
+from pytorch_transformers import glue_compute_metrics as compute_metrics
+from pytorch_transformers import glue_output_modes as output_modes
+from pytorch_transformers import glue_processors as processors
+from pytorch_transformers import glue_convert_examples_to_features as convert_examples_to_features

 logger = logging.getLogger(__name__)

-ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, 
-                                                                                RobertaConfig, DistilBertConfig)), ())
+ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, RobertaConfig)), ())

 MODEL_CLASSES = {
    'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
    'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer),
    'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
    'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
-    'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer),
-    'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer)
+    'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
 }


@@ -104,9 +93,8 @@ def train(args, train_dataset, model, tokenizer):
        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
-
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
    if args.fp16:
        try:
            from apex import amp
@@ -146,11 +134,10 @@ def train(args, train_dataset, model, tokenizer):
            batch = tuple(t.to(args.device) for t in batch)
            inputs = {'input_ids':      batch[0],
                      'attention_mask': batch[1],
+                      'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None,  # XLM, DistilBERT and RoBERTa don't use segment_ids
                      'labels':         batch[3]}
-            if args.model_type != 'distilbert':
-                inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet'] else None  # XLM, DistilBERT and RoBERTa don't use segment_ids
            outputs = model(**inputs)
-            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
+            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)

            if args.n_gpu > 1:
                loss = loss.mean() # mean() to average on multi-gpu parallel training
@@ -160,18 +147,15 @@ def train(args, train_dataset, model, tokenizer):
            if args.fp16:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
+                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
            else:
                loss.backward()
+                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
-                if args.fp16:
-                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
-                else:
-                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
-
-                optimizer.step()
                scheduler.step()  # Update learning rate schedule
+                optimizer.step()
                model.zero_grad()
                global_step += 1

@@ -225,10 +209,6 @@ def evaluate(args, model, tokenizer, prefix=""):
        eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

-        # multi-gpu eval
-        if args.n_gpu > 1:
-            model = torch.nn.DataParallel(model)
-
        # Eval!
        logger.info("***** Running evaluation {} *****".format(prefix))
        logger.info("  Num examples = %d", len(eval_dataset))
@@ -244,9 +224,8 @@ def evaluate(args, model, tokenizer, prefix=""):
            with torch.no_grad():
                inputs = {'input_ids':      batch[0],
                          'attention_mask': batch[1],
+                          'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None,  # XLM, DistilBERT and RoBERTa don't use segment_ids
                          'labels':         batch[3]}
-                if args.model_type != 'distilbert':
-                    inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet'] else None  # XLM, DistilBERT and RoBERTa don't use segment_ids
                outputs = model(**inputs)
                tmp_eval_loss, logits = outputs[:2]

@@ -267,7 +246,7 @@ def evaluate(args, model, tokenizer, prefix=""):
        result = compute_metrics(eval_task, preds, out_label_ids)
        results.update(result)

-        output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
+        output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
        with open(output_eval_file, "w") as writer:
            logger.info("***** Eval results {} *****".format(prefix))
            for key in sorted(result.keys()):
@@ -289,7 +268,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
        list(filter(None, args.model_name_or_path.split('/'))).pop(),
        str(args.max_seq_length),
        str(task)))
-    if os.path.exists(cached_features_file) and not args.overwrite_cache:
+    if os.path.exists(cached_features_file):
        logger.info("Loading features from cached file %s", cached_features_file)
        features = torch.load(cached_features_file)
    else:
@@ -323,7 +302,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
        all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
    elif output_mode == "regression":
        all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
- 
+
    dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
    return dataset

@@ -367,7 +346,7 @@ def main():
    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
                        help="Batch size per GPU/CPU for evaluation.")
    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
-                        help="Number of updates steps to accumulate before performing a backward/update pass.")     
+                        help="Number of updates steps to accumulate before performing a backward/update pass.")
    parser.add_argument("--learning_rate", default=5e-5, type=float,
                        help="The initial learning rate for Adam.")
    parser.add_argument("--weight_decay", default=0.0, type=float,
@@ -456,17 +435,9 @@ def main():

    args.model_type = args.model_type.lower()
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
-                                          num_labels=num_labels,
-                                          finetuning_task=args.task_name,
-                                          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)
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
+    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)

    if args.local_rank == 0:
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
@@ -501,7 +472,7 @@ def main():

        # Load a trained model and vocabulary that you have fine-tuned
        model = model_class.from_pretrained(args.output_dir)
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
        model.to(args.device)


@@ -512,15 +483,13 @@ def main():
        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 logging
+            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
-            prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
-            
            model = model_class.from_pretrained(checkpoint)
            model.to(args.device)
-            result = evaluate(args, model, tokenizer, prefix=prefix)
+            result = evaluate(args, model, tokenizer, prefix=global_step)
            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
            results.update(result)

@@ -27,22 +27,15 @@ import logging
 import os
 import pickle
 import random
-import re
-import shutil

 import numpy as np
 import torch
 from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler
 from torch.utils.data.distributed import DistributedSampler
-
-try:
-    from torch.utils.tensorboard import SummaryWriter
-except:
-    from tensorboardX import SummaryWriter
-
+from tensorboardX import SummaryWriter
 from tqdm import tqdm, trange

-from transformers import (WEIGHTS_NAME, AdamW, get_linear_schedule_with_warmup,
+from pytorch_transformers import (WEIGHTS_NAME, AdamW, WarmupLinearSchedule,
                                  BertConfig, BertForMaskedLM, BertTokenizer,
                                  GPT2Config, GPT2LMHeadModel, GPT2Tokenizer,
                                  OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer,
@@ -63,12 +56,12 @@ MODEL_CLASSES = {


 class TextDataset(Dataset):
-    def __init__(self, tokenizer, args, file_path='train', block_size=512):
+    def __init__(self, tokenizer, file_path='train', block_size=512):
        assert os.path.isfile(file_path)
        directory, filename = os.path.split(file_path)
-        cached_features_file = os.path.join(directory, args.model_name_or_path + '_cached_lm_' + str(block_size) + '_' + filename)
+        cached_features_file = os.path.join(directory, f'cached_lm_{block_size}_{filename}')

-        if os.path.exists(cached_features_file) and not args.overwrite_cache:
+        if os.path.exists(cached_features_file):
            logger.info("Loading features from cached file %s", cached_features_file)
            with open(cached_features_file, 'rb') as handle:
                self.examples = pickle.load(handle)
@@ -81,8 +74,9 @@ class TextDataset(Dataset):

            tokenized_text = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(text))

-            for i in range(0, len(tokenized_text)-block_size+1, block_size): # Truncate in block of block_size
-                self.examples.append(tokenizer.build_inputs_with_special_tokens(tokenized_text[i:i+block_size]))
+            while len(tokenized_text) >= block_size:  # Truncate in block of block_size
+                self.examples.append(tokenizer.add_special_tokens_single_sequence(tokenized_text[:block_size]))
+                tokenized_text = tokenized_text[block_size:]
            # Note that we are loosing the last truncated example here for the sake of simplicity (no padding)
            # If your dataset is small, first you should loook for a bigger one :-) and second you
            # can change this behavior by adding (model specific) padding.
@@ -99,7 +93,7 @@ class TextDataset(Dataset):


 def load_and_cache_examples(args, tokenizer, evaluate=False):
-    dataset = TextDataset(tokenizer, args, file_path=args.eval_data_file if evaluate else args.train_data_file, block_size=args.block_size)
+    dataset = TextDataset(tokenizer, file_path=args.eval_data_file if evaluate else args.train_data_file, block_size=args.block_size)
    return dataset


@@ -111,43 +105,11 @@ def set_seed(args):
        torch.cuda.manual_seed_all(args.seed)


-def _rotate_checkpoints(args, checkpoint_prefix, use_mtime=False):
-    if not args.save_total_limit:
-        return
-    if args.save_total_limit <= 0:
-        return
-
-    # Check if we should delete older checkpoint(s)
-    glob_checkpoints = glob.glob(os.path.join(args.output_dir, '{}-*'.format(checkpoint_prefix)))
-    if len(glob_checkpoints) <= args.save_total_limit:
-        return
-
-    ordering_and_checkpoint_path = []
-    for path in glob_checkpoints:
-        if use_mtime:
-            ordering_and_checkpoint_path.append((os.path.getmtime(path), path))
-        else:
-            regex_match = re.match('.*{}-([0-9]+)'.format(checkpoint_prefix), path)
-            if regex_match and regex_match.groups():
-                ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path))
-
-    checkpoints_sorted = sorted(ordering_and_checkpoint_path)
-    checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
-    number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - args.save_total_limit)
-    checkpoints_to_be_deleted = checkpoints_sorted[:number_of_checkpoints_to_delete]
-    for checkpoint in checkpoints_to_be_deleted:
-        logger.info("Deleting older checkpoint [{}] due to args.save_total_limit".format(checkpoint))
-        shutil.rmtree(checkpoint)
-
-
 def mask_tokens(inputs, tokenizer, args):
    """ Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. """
    labels = inputs.clone()
    # We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
-    probability_matrix = torch.full(labels.shape, args.mlm_probability)
-    special_tokens_mask = [tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()]
-    probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
-    masked_indices = torch.bernoulli(probability_matrix).bool()
+    masked_indices = torch.bernoulli(torch.full(labels.shape, args.mlm_probability)).bool()
    labels[~masked_indices] = -1  # We only compute loss on masked tokens

    # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
@@ -185,7 +147,7 @@ def train(args, train_dataset, model, tokenizer):
        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
    if args.fp16:
        try:
            from apex import amp
@@ -215,7 +177,6 @@ def train(args, train_dataset, model, tokenizer):

    global_step = 0
    tr_loss, logging_loss = 0.0, 0.0
-    model.resize_token_embeddings(len(tokenizer))
    model.zero_grad()
    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
    set_seed(args)  # Added here for reproducibility (even between python 2 and 3)
@@ -227,7 +188,7 @@ def train(args, train_dataset, model, tokenizer):
            labels = labels.to(args.device)
            model.train()
            outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
-            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
+            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)

            if args.n_gpu > 1:
                loss = loss.mean()  # mean() to average on multi-gpu parallel training
@@ -262,9 +223,8 @@ def train(args, train_dataset, model, tokenizer):
                    logging_loss = tr_loss

                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
-                    checkpoint_prefix = 'checkpoint'
                    # Save model checkpoint
-                    output_dir = os.path.join(args.output_dir, '{}-{}'.format(checkpoint_prefix, global_step))
+                    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
@@ -272,8 +232,6 @@ def train(args, train_dataset, model, tokenizer):
                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
                    logger.info("Saving model checkpoint to %s", output_dir)

-                    _rotate_checkpoints(args, checkpoint_prefix)
-
            if args.max_steps > 0 and global_step > args.max_steps:
                epoch_iterator.close()
                break
@@ -301,10 +259,6 @@ def evaluate(args, model, tokenizer, prefix=""):
    eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
    eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

-    # multi-gpu evaluate
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
    # Eval!
    logger.info("***** Running evaluation {} *****".format(prefix))
    logger.info("  Num examples = %d", len(eval_dataset))
@@ -314,12 +268,10 @@ def evaluate(args, model, tokenizer, prefix=""):
    model.eval()

    for batch in tqdm(eval_dataloader, desc="Evaluating"):
-        inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
-        inputs = inputs.to(args.device)
-        labels = labels.to(args.device)
+        batch = batch.to(args.device)

        with torch.no_grad():
-            outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
+            outputs = model(batch, masked_lm_labels=batch) if args.mlm else model(batch, labels=batch)
            lm_loss = outputs[0]
            eval_loss += lm_loss.mean().item()
        nb_eval_steps += 1
@@ -331,7 +283,7 @@ def evaluate(args, model, tokenizer, prefix=""):
        "perplexity": perplexity
    }

-    output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
+    output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
    with open(output_eval_file, "w") as writer:
        logger.info("***** Eval results {} *****".format(prefix))
        for key in sorted(result.keys()):
@@ -408,8 +360,6 @@ def main():
                        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('--save_total_limit', type=int, default=None,
-                        help='Limit the total amount of checkpoints, delete the older checkpoints in the output_dir, does not delete by default')
    parser.add_argument("--eval_all_checkpoints", action='store_true',
                        help="Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number")
    parser.add_argument("--no_cuda", action='store_true',
@@ -476,18 +426,12 @@ def main():
        torch.distributed.barrier()  # Barrier to make sure only the first process in distributed training download model & vocab

    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
-                                          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)
+    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)
    if args.block_size <= 0:
        args.block_size = tokenizer.max_len_single_sentence  # Our input block size will be the max possible for the model
    args.block_size = min(args.block_size, tokenizer.max_len_single_sentence)
-    model = model_class.from_pretrained(args.model_name_or_path,
-                                        from_tf=bool('.ckpt' in args.model_name_or_path),
-                                        config=config,
-                                        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)
    model.to(args.device)

    if args.local_rank == 0:
@@ -537,15 +481,13 @@ def main():
        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 logging
+            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
-            prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
-            
            model = model_class.from_pretrained(checkpoint)
            model.to(args.device)
-            result = evaluate(args, model, tokenizer, prefix=prefix)
+            result = evaluate(args, model, tokenizer, prefix=global_step)
            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
            results.update(result)

@@ -29,21 +29,16 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-
-try:
-    from torch.utils.tensorboard import SummaryWriter
-except:
-    from tensorboardX import SummaryWriter
-
+from tensorboardX import SummaryWriter
 from tqdm import tqdm, trange

-from transformers import (WEIGHTS_NAME, BertConfig,
+from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
                                  BertForMultipleChoice, BertTokenizer,
                                  XLNetConfig, XLNetForMultipleChoice,
                                  XLNetTokenizer, RobertaConfig,
                                  RobertaForMultipleChoice, RobertaTokenizer)

-from transformers import AdamW, get_linear_schedule_with_warmup
+from pytorch_transformers import AdamW, WarmupLinearSchedule

 from utils_multiple_choice import (convert_examples_to_features, processors)

@@ -101,7 +96,7 @@ def train(args, train_dataset, model, tokenizer):
        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
    if args.fp16:
        try:
            from apex import amp
@@ -146,7 +141,7 @@ def train(args, train_dataset, model, tokenizer):
                      'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None,  # XLM don't use segment_ids
                      'labels':         batch[3]}
            outputs = model(**inputs)
-            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
+            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)

            if args.n_gpu > 1:
                loss = loss.mean() # mean() to average on multi-gpu parallel training
@@ -229,10 +224,6 @@ def evaluate(args, model, tokenizer, prefix="", test=False):
        eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

-        # multi-gpu evaluate
-        if args.n_gpu > 1:
-            model = torch.nn.DataParallel(model)
-
        # Eval!
        logger.info("***** Running evaluation {} *****".format(prefix))
        logger.info("  Num examples = %d", len(eval_dataset))
@@ -302,7 +293,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False, test=False):
        list(filter(None, args.model_name_or_path.split('/'))).pop(),
        str(args.max_seq_length),
        str(task)))
-    if os.path.exists(cached_features_file) and not args.overwrite_cache:
+    if os.path.exists(cached_features_file):
        logger.info("Loading features from cached file %s", cached_features_file)
        features = torch.load(cached_features_file)
    else:
@@ -315,14 +306,14 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False, test=False):
        else:
            examples = processor.get_train_examples(args.data_dir)
        logger.info("Training number: %s", str(len(examples)))
-        features = convert_examples_to_features(
-            examples,
-            label_list,
-            args.max_seq_length,
-            tokenizer,
+        features = convert_examples_to_features(examples, label_list, args.max_seq_length, tokenizer,
+            cls_token_at_end=bool(args.model_type in ['xlnet']),            # xlnet has a cls token at the end
+            cls_token=tokenizer.cls_token,
+            sep_token=tokenizer.sep_token,
+            sep_token_extra=bool(args.model_type in ['roberta']),
+            cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
            pad_on_left=bool(args.model_type in ['xlnet']),                 # pad on the left for xlnet
-            pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0
-        )
+            pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0)
        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
            torch.save(features, cached_features_file)
@@ -371,7 +362,7 @@ def main():
                        help="Whether to run eval on the dev set.")
    parser.add_argument("--do_test", action='store_true', help='Whether to run test on the test set')
    parser.add_argument("--evaluate_during_training", action='store_true',
-                        help="Run evaluation during training at each logging step.")
+                        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.")

@@ -468,17 +459,9 @@ def main():

    args.model_type = args.model_type.lower()
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
-                                          num_labels=num_labels,
-                                          finetuning_task=args.task_name,
-                                          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)
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
+    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)

    if args.local_rank == 0:
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
@@ -525,15 +508,13 @@ def main():
        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 logging
+            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
-            prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
-            
            model = model_class.from_pretrained(checkpoint)
            model.to(args.device)
-            result = evaluate(args, model, tokenizer, prefix=prefix)
+            result = evaluate(args, model, tokenizer, prefix=global_step)
            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
            results.update(result)

@@ -543,15 +524,13 @@ def main():
        checkpoints = [args.output_dir]
        # if args.eval_all_checkpoints: # can not use this to do test!!
        #     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 logging
+        #     logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
-            prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
-            
            model = model_class.from_pretrained(checkpoint)
            model.to(args.device)
-            result = evaluate(args, model, tokenizer, prefix=prefix, test=True)
+            result = evaluate(args, model, tokenizer, prefix=global_step, test=True)
            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
            results.update(result)
    if best_steps:
@@ -1,532 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" Fine-tuning the library models for named entity recognition on CoNLL-2003 (Bert or Roberta). """
-
-from __future__ import absolute_import, division, print_function
-
-import argparse
-import glob
-import logging
-import os
-import random
-
-import numpy as np
-import torch
-from seqeval.metrics import precision_score, recall_score, f1_score
-from tensorboardX import SummaryWriter
-from torch.nn import CrossEntropyLoss
-from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
-from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm, trange
-from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file
-
-from transformers import AdamW, get_linear_schedule_with_warmup
-from transformers import WEIGHTS_NAME, BertConfig, BertForTokenClassification, BertTokenizer
-from transformers import RobertaConfig, RobertaForTokenClassification, RobertaTokenizer
-from transformers import DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer
-from transformers import CamembertConfig, CamembertForTokenClassification, CamembertTokenizer
-
-logger = logging.getLogger(__name__)
-
-ALL_MODELS = sum(
-    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)),
-    ())
-
-MODEL_CLASSES = {
-    "bert": (BertConfig, BertForTokenClassification, BertTokenizer),
-    "roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer),
-    "distilbert": (DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer),
-    "camembert": (CamembertConfig, CamembertForTokenClassification, CamembertTokenizer),
-}
-
-
-def set_seed(args):
-    random.seed(args.seed)
-    np.random.seed(args.seed)
-    torch.manual_seed(args.seed)
-    if args.n_gpu > 0:
-        torch.cuda.manual_seed_all(args.seed)
-
-
-def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
-    """ Train the model """
-    if args.local_rank in [-1, 0]:
-        tb_writer = SummaryWriter()
-
-    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
-    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
-    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
-
-    if args.max_steps > 0:
-        t_total = args.max_steps
-        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
-    else:
-        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
-
-    # Prepare optimizer and schedule (linear warmup and decay)
-    no_decay = ["bias", "LayerNorm.weight"]
-    optimizer_grouped_parameters = [
-        {"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
-         "weight_decay": args.weight_decay},
-        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0}
-    ]
-    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
-    if args.fp16:
-        try:
-            from apex import amp
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
-        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
-
-    # multi-gpu training (should be after apex fp16 initialization)
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    # Distributed training (should be after apex fp16 initialization)
-    if args.local_rank != -1:
-        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
-                                                          output_device=args.local_rank,
-                                                          find_unused_parameters=True)
-
-    # Train!
-    logger.info("***** Running training *****")
-    logger.info("  Num examples = %d", len(train_dataset))
-    logger.info("  Num Epochs = %d", args.num_train_epochs)
-    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
-    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
-                args.train_batch_size * args.gradient_accumulation_steps * (
-                    torch.distributed.get_world_size() if args.local_rank != -1 else 1))
-    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
-    logger.info("  Total optimization steps = %d", t_total)
-
-    global_step = 0
-    tr_loss, logging_loss = 0.0, 0.0
-    model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
-    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
-    for _ in train_iterator:
-        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
-        for step, batch in enumerate(epoch_iterator):
-            model.train()
-            batch = tuple(t.to(args.device) for t in batch)
-            inputs = {"input_ids": batch[0],
-                      "attention_mask": batch[1],
-                      "labels": batch[3]}
-            if args.model_type != "distilbert":
-                inputs["token_type_ids"] = batch[2] if args.model_type in ["bert", "xlnet"] else None  # XLM and RoBERTa don"t use segment_ids
-
-            outputs = model(**inputs)
-            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)
-
-            if args.n_gpu > 1:
-                loss = loss.mean()  # mean() to average on multi-gpu parallel training
-            if args.gradient_accumulation_steps > 1:
-                loss = loss / args.gradient_accumulation_steps
-
-            if args.fp16:
-                with amp.scale_loss(loss, optimizer) as scaled_loss:
-                    scaled_loss.backward()
-            else:
-                loss.backward()
-
-            tr_loss += loss.item()
-            if (step + 1) % args.gradient_accumulation_steps == 0:
-                if args.fp16:
-                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
-                else:
-                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
-
-                scheduler.step()  # Update learning rate schedule
-                optimizer.step()
-                model.zero_grad()
-                global_step += 1
-
-                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
-                    # Log metrics
-                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
-                        results, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev")
-                        for key, value in results.items():
-                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
-                    tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
-                    tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
-                    logging_loss = tr_loss
-
-                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
-                    # Save model checkpoint
-                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
-                    if not os.path.exists(output_dir):
-                        os.makedirs(output_dir)
-                    model_to_save = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
-                    model_to_save.save_pretrained(output_dir)
-                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
-                    logger.info("Saving model checkpoint to %s", output_dir)
-
-            if args.max_steps > 0 and global_step > args.max_steps:
-                epoch_iterator.close()
-                break
-        if args.max_steps > 0 and global_step > args.max_steps:
-            train_iterator.close()
-            break
-
-    if args.local_rank in [-1, 0]:
-        tb_writer.close()
-
-    return global_step, tr_loss / global_step
-
-
-def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""):
-    eval_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode=mode)
-
-    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
-    # Note that DistributedSampler samples randomly
-    eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
-    eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
-
-    # multi-gpu evaluate
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    # Eval!
-    logger.info("***** Running evaluation %s *****", prefix)
-    logger.info("  Num examples = %d", len(eval_dataset))
-    logger.info("  Batch size = %d", args.eval_batch_size)
-    eval_loss = 0.0
-    nb_eval_steps = 0
-    preds = None
-    out_label_ids = None
-    model.eval()
-    for batch in tqdm(eval_dataloader, desc="Evaluating"):
-        batch = tuple(t.to(args.device) for t in batch)
-
-        with torch.no_grad():
-            inputs = {"input_ids": batch[0],
-                      "attention_mask": batch[1],
-                      "labels": batch[3]}
-            if args.model_type != "distilbert":
-                inputs["token_type_ids"] = batch[2] if args.model_type in ["bert", "xlnet"] else None  # XLM and RoBERTa don"t use segment_ids
-            outputs = model(**inputs)
-            tmp_eval_loss, logits = outputs[:2]
-
-            if args.n_gpu > 1:
-                tmp_eval_loss = tmp_eval_loss.mean()  # mean() to average on multi-gpu parallel evaluating
-
-            eval_loss += tmp_eval_loss.item()
-        nb_eval_steps += 1
-        if preds is None:
-            preds = logits.detach().cpu().numpy()
-            out_label_ids = inputs["labels"].detach().cpu().numpy()
-        else:
-            preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
-            out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
-
-    eval_loss = eval_loss / nb_eval_steps
-    preds = np.argmax(preds, axis=2)
-
-    label_map = {i: label for i, label in enumerate(labels)}
-
-    out_label_list = [[] for _ in range(out_label_ids.shape[0])]
-    preds_list = [[] for _ in range(out_label_ids.shape[0])]
-
-    for i in range(out_label_ids.shape[0]):
-        for j in range(out_label_ids.shape[1]):
-            if out_label_ids[i, j] != pad_token_label_id:
-                out_label_list[i].append(label_map[out_label_ids[i][j]])
-                preds_list[i].append(label_map[preds[i][j]])
-
-    results = {
-        "loss": eval_loss,
-        "precision": precision_score(out_label_list, preds_list),
-        "recall": recall_score(out_label_list, preds_list),
-        "f1": f1_score(out_label_list, preds_list)
-    }
-
-    logger.info("***** Eval results %s *****", prefix)
-    for key in sorted(results.keys()):
-        logger.info("  %s = %s", key, str(results[key]))
-
-    return results, preds_list
-
-
-def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode):
-    if args.local_rank not in [-1, 0] and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    # Load data features from cache or dataset file
-    cached_features_file = os.path.join(args.data_dir, "cached_{}_{}_{}".format(mode,
-        list(filter(None, args.model_name_or_path.split("/"))).pop(),
-        str(args.max_seq_length)))
-    if os.path.exists(cached_features_file) and not args.overwrite_cache:
-        logger.info("Loading features from cached file %s", cached_features_file)
-        features = torch.load(cached_features_file)
-    else:
-        logger.info("Creating features from dataset file at %s", args.data_dir)
-        examples = read_examples_from_file(args.data_dir, mode)
-        features = convert_examples_to_features(examples, labels, args.max_seq_length, tokenizer,
-                                                cls_token_at_end=bool(args.model_type in ["xlnet"]),
-                                                # xlnet has a cls token at the end
-                                                cls_token=tokenizer.cls_token,
-                                                cls_token_segment_id=2 if args.model_type in ["xlnet"] else 0,
-                                                sep_token=tokenizer.sep_token,
-                                                sep_token_extra=bool(args.model_type in ["roberta"]),
-                                                # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
-                                                pad_on_left=bool(args.model_type in ["xlnet"]),
-                                                # pad on the left for xlnet
-                                                pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
-                                                pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
-                                                pad_token_label_id=pad_token_label_id
-                                                )
-        if args.local_rank in [-1, 0]:
-            logger.info("Saving features into cached file %s", cached_features_file)
-            torch.save(features, cached_features_file)
-
-    if args.local_rank == 0 and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    # Convert to Tensors and build dataset
-    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
-    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
-    all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
-    all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long)
-
-    dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
-    return dataset
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    ## Required parameters
-    parser.add_argument("--data_dir", default=None, type=str, required=True,
-                        help="The input data dir. Should contain the training files for the CoNLL-2003 NER task.")
-    parser.add_argument("--model_type", default=None, type=str, required=True,
-                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
-    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
-                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
-    parser.add_argument("--output_dir", default=None, type=str, required=True,
-                        help="The output directory where the model predictions and checkpoints will be written.")
-
-    ## Other parameters
-    parser.add_argument("--labels", default="", type=str,
-                        help="Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.")
-    parser.add_argument("--config_name", default="", type=str,
-                        help="Pretrained config name or path if not the same as model_name")
-    parser.add_argument("--tokenizer_name", default="", type=str,
-                        help="Pretrained tokenizer name or path if not the same as model_name")
-    parser.add_argument("--cache_dir", default="", type=str,
-                        help="Where do you want to store the pre-trained models downloaded from s3")
-    parser.add_argument("--max_seq_length", default=128, type=int,
-                        help="The maximum total input sequence length after tokenization. Sequences longer "
-                             "than this will be truncated, sequences shorter will be padded.")
-    parser.add_argument("--do_train", action="store_true",
-                        help="Whether to run training.")
-    parser.add_argument("--do_eval", action="store_true",
-                        help="Whether to run eval on the dev set.")
-    parser.add_argument("--do_predict", action="store_true",
-                        help="Whether to run predictions on the test set.")
-    parser.add_argument("--evaluate_during_training", action="store_true",
-                        help="Whether to run evaluation during training at each logging step.")
-    parser.add_argument("--do_lower_case", action="store_true",
-                        help="Set this flag if you are using an uncased model.")
-
-    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
-                        help="Batch size per GPU/CPU for training.")
-    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
-                        help="Batch size per GPU/CPU for evaluation.")
-    parser.add_argument("--gradient_accumulation_steps", type=int, default=1,
-                        help="Number of updates steps to accumulate before performing a backward/update pass.")
-    parser.add_argument("--learning_rate", default=5e-5, type=float,
-                        help="The initial learning rate for Adam.")
-    parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight decay if we apply some.")
-    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
-                        help="Epsilon for Adam optimizer.")
-    parser.add_argument("--max_grad_norm", default=1.0, type=float,
-                        help="Max gradient norm.")
-    parser.add_argument("--num_train_epochs", default=3.0, type=float,
-                        help="Total number of training epochs to perform.")
-    parser.add_argument("--max_steps", default=-1, type=int,
-                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
-    parser.add_argument("--warmup_steps", default=0, type=int,
-                        help="Linear warmup over warmup_steps.")
-
-    parser.add_argument("--logging_steps", type=int, default=50,
-                        help="Log every X updates steps.")
-    parser.add_argument("--save_steps", type=int, default=50,
-                        help="Save checkpoint every X updates steps.")
-    parser.add_argument("--eval_all_checkpoints", action="store_true",
-                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
-    parser.add_argument("--no_cuda", action="store_true",
-                        help="Avoid using CUDA when available")
-    parser.add_argument("--overwrite_output_dir", action="store_true",
-                        help="Overwrite the content of the output directory")
-    parser.add_argument("--overwrite_cache", action="store_true",
-                        help="Overwrite the cached training and evaluation sets")
-    parser.add_argument("--seed", type=int, default=42,
-                        help="random seed for initialization")
-
-    parser.add_argument("--fp16", action="store_true",
-                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
-    parser.add_argument("--fp16_opt_level", type=str, default="O1",
-                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
-                             "See details at https://nvidia.github.io/apex/amp.html")
-    parser.add_argument("--local_rank", type=int, default=-1,
-                        help="For distributed training: local_rank")
-    parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
-    parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
-    args = parser.parse_args()
-
-    if os.path.exists(args.output_dir) and os.listdir(
-            args.output_dir) and args.do_train and not args.overwrite_output_dir:
-        raise ValueError(
-            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
-                args.output_dir))
-
-    # Setup distant debugging if needed
-    if args.server_ip and args.server_port:
-        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
-        import ptvsd
-        print("Waiting for debugger attach")
-        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
-        ptvsd.wait_for_attach()
-
-    # Setup CUDA, GPU & distributed training
-    if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
-    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
-        torch.cuda.set_device(args.local_rank)
-        device = torch.device("cuda", args.local_rank)
-        torch.distributed.init_process_group(backend="nccl")
-        args.n_gpu = 1
-    args.device = device
-
-    # Setup logging
-    logging.basicConfig(format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
-                        datefmt="%m/%d/%Y %H:%M:%S",
-                        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
-    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
-                   args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
-
-    # Set seed
-    set_seed(args)
-
-    # Prepare CONLL-2003 task
-    labels = get_labels(args.labels)
-    num_labels = len(labels)
-    # Use cross entropy ignore index as padding label id so that only real label ids contribute to the loss later
-    pad_token_label_id = CrossEntropyLoss().ignore_index
-
-    # Load pretrained model and tokenizer
-    if args.local_rank not in [-1, 0]:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    args.model_type = args.model_type.lower()
-    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
-                                          num_labels=num_labels,
-                                          cache_dir=args.cache_dir if args.cache_dir else None)
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
-                                                do_lower_case=args.do_lower_case,
-                                                cache_dir=args.cache_dir if args.cache_dir else None)
-    model = model_class.from_pretrained(args.model_name_or_path,
-                                        from_tf=bool(".ckpt" in args.model_name_or_path),
-                                        config=config,
-                                        cache_dir=args.cache_dir if args.cache_dir else None)
-
-    if args.local_rank == 0:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    model.to(args.device)
-
-    logger.info("Training/evaluation parameters %s", args)
-
-    # Training
-    if args.do_train:
-        train_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode="train")
-        global_step, tr_loss = train(args, train_dataset, model, tokenizer, labels, pad_token_label_id)
-        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
-
-    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
-    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
-        logger.info("Saving model checkpoint to %s", args.output_dir)
-        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
-        # They can then be reloaded using `from_pretrained()`
-        model_to_save = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
-        model_to_save.save_pretrained(args.output_dir)
-        tokenizer.save_pretrained(args.output_dir)
-
-        # Good practice: save your training arguments together with the trained model
-        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
-
-    # Evaluation
-    results = {}
-    if args.do_eval and args.local_rank in [-1, 0]:
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
-        checkpoints = [args.output_dir]
-        if args.eval_all_checkpoints:
-            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True)))
-            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
-        logger.info("Evaluate the following checkpoints: %s", checkpoints)
-        for checkpoint in checkpoints:
-            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
-            model = model_class.from_pretrained(checkpoint)
-            model.to(args.device)
-            result, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev", prefix=global_step)
-            if global_step:
-                result = {"{}_{}".format(global_step, k): v for k, v in result.items()}
-            results.update(result)
-        output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
-        with open(output_eval_file, "w") as writer:
-            for key in sorted(results.keys()):
-                writer.write("{} = {}\n".format(key, str(results[key])))
-
-    if args.do_predict and args.local_rank in [-1, 0]:
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
-        model = model_class.from_pretrained(args.output_dir)
-        model.to(args.device)
-        result, predictions = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="test")
-        # Save results
-        output_test_results_file = os.path.join(args.output_dir, "test_results.txt")
-        with open(output_test_results_file, "w") as writer:
-            for key in sorted(result.keys()):
-                writer.write("{} = {}\n".format(key, str(result[key])))
-        # Save predictions
-        output_test_predictions_file = os.path.join(args.output_dir, "test_predictions.txt")
-        with open(output_test_predictions_file, "w") as writer:
-            with open(os.path.join(args.data_dir, "test.txt"), "r") as f:
-                example_id = 0
-                for line in f:
-                    if line.startswith("-DOCSTART-") or line == "" or line == "\n":
-                        writer.write(line)
-                        if not predictions[example_id]:
-                            example_id += 1
-                    elif predictions[example_id]:
-                        output_line = line.split()[0] + " " + predictions[example_id].pop(0) + "\n"
-                        writer.write(output_line)
-                    else:
-                        logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
-
-    return results
-
-
-if __name__ == "__main__":
-    main()
-
@@ -13,7 +13,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-""" Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet)."""
+""" Finetuning the library models for question-answering on SQuAD (Bert, XLM, XLNet)."""

 from __future__ import absolute_import, division, print_function

@@ -22,31 +22,25 @@ import logging
 import os
 import random
 import glob
-import timeit

 import numpy as np
 import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-
-try:
-    from torch.utils.tensorboard import SummaryWriter
-except:
-    from tensorboardX import SummaryWriter
-
 from tqdm import tqdm, trange

-from transformers import (WEIGHTS_NAME, BertConfig,
+from tensorboardX import SummaryWriter
+
+from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
                                  BertForQuestionAnswering, BertTokenizer,
                                  XLMConfig, XLMForQuestionAnswering,
                                  XLMTokenizer, XLNetConfig,
                                  XLNetForQuestionAnswering,
                                  XLNetTokenizer,
-                                  DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer,
-                                  AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer)
+                                  DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)

-from transformers import AdamW, get_linear_schedule_with_warmup
+from pytorch_transformers import AdamW, WarmupLinearSchedule

 from utils_squad import (read_squad_examples, convert_examples_to_features,
                         RawResult, write_predictions,
@@ -66,8 +60,7 @@ MODEL_CLASSES = {
    'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
    'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
    'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
-    'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
-    'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer)
+    'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
 }

 def set_seed(args):
@@ -102,7 +95,7 @@ def train(args, train_dataset, model, tokenizer):
        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
    if args.fp16:
        try:
            from apex import amp
@@ -130,7 +123,7 @@ def train(args, train_dataset, model, tokenizer):
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", t_total)

-    global_step = 1
+    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])
@@ -141,16 +134,15 @@ def train(args, train_dataset, model, tokenizer):
            model.train()
            batch = tuple(t.to(args.device) for t in batch)
            inputs = {'input_ids':       batch[0],
-                      'attention_mask':  batch[1],
-                      'start_positions': batch[3],
+                      'attention_mask':  batch[1], 
+                      'token_type_ids':  None if args.model_type == 'xlm' else batch[2],  
+                      'start_positions': batch[3], 
                      'end_positions':   batch[4]}
-            if args.model_type != 'distilbert':
-                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
            if args.model_type in ['xlnet', 'xlm']:
                inputs.update({'cls_index': batch[5],
                               'p_mask':       batch[6]})
            outputs = model(**inputs)
-            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
+            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)

            if args.n_gpu > 1:
                loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
@@ -160,16 +152,13 @@ def train(args, train_dataset, model, tokenizer):
            if args.fp16:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
+                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
            else:
                loss.backward()
+                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
-                if args.fp16:
-                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
-                else:
-                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
-
                optimizer.step()
                scheduler.step()  # Update learning rate schedule
                model.zero_grad()
@@ -219,25 +208,19 @@ def evaluate(args, model, tokenizer, prefix=""):
    eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

-    # multi-gpu evaluate
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
    # Eval!
    logger.info("***** Running evaluation {} *****".format(prefix))
    logger.info("  Num examples = %d", len(dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)
    all_results = []
-    start_time = timeit.default_timer()
    for batch in tqdm(eval_dataloader, desc="Evaluating"):
        model.eval()
        batch = tuple(t.to(args.device) for t in batch)
        with torch.no_grad():
            inputs = {'input_ids':      batch[0],
-                      'attention_mask': batch[1]
+                      'attention_mask': batch[1],
+                      'token_type_ids': None if args.model_type == 'xlm' else batch[2]  # XLM don't use segment_ids
                      }
-            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],
@@ -261,9 +244,6 @@ def evaluate(args, model, tokenizer, prefix=""):
                                   end_logits   = to_list(outputs[1][i]))
            all_results.append(result)

-    evalTime = timeit.default_timer() - start_time
-    logger.info("  Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))
-
    # Compute predictions
    output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
    output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
@@ -316,11 +296,7 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
                                                max_seq_length=args.max_seq_length,
                                                doc_stride=args.doc_stride,
                                                max_query_length=args.max_query_length,
-                                                is_training=not evaluate,
-                                                cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
-                                                pad_token_segment_id=3 if args.model_type in ['xlnet'] else 0,
-                                                cls_token_at_end=True if args.model_type in ['xlnet'] else False,
-                                                sequence_a_is_doc=True if args.model_type in ['xlnet'] else False)
+                                                is_training=not evaluate)
        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
            torch.save(features, cached_features_file)
@@ -404,7 +380,7 @@ def main():
    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
                        help="Number of updates steps to accumulate before performing a backward/update pass.")
    parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight decay if we apply some.")
+                        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,
@@ -488,15 +464,9 @@ def main():

    args.model_type = args.model_type.lower()
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
-                                          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)
+    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
@@ -505,16 +475,6 @@ def main():

    logger.info("Training/evaluation parameters %s", args)

-    # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
-    # Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
-    # remove the need for this code, but it is still valid.
-    if args.fp16:
-        try:
-            import apex
-            apex.amp.register_half_function(torch, 'einsum')
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
-
    # Training
    if args.do_train:
        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
@@ -539,7 +499,7 @@ def main():
        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))

        # Load a trained model and vocabulary that you have fine-tuned
-        model = model_class.from_pretrained(args.output_dir, force_download=True)
+        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)

@@ -550,14 +510,14 @@ def main():
        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
+            logging.getLogger("pytorch_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, force_download=True)
+            model = model_class.from_pretrained(checkpoint)
            model.to(args.device)

            # Evaluate
@@ -1,492 +0,0 @@
-# coding=utf-8
-# Copyright 2019 The HuggingFace Inc. team.
-# Copyright (c) 2019 The HuggingFace Inc.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" Finetuning seq2seq models for sequence generation."""
-
-import argparse
-import functools
-import logging
-import os
-import random
-import sys
-
-import numpy as np
-from tqdm import tqdm, trange
-import torch
-from torch.optim import Adam
-from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
-
-from transformers import (
-    AutoTokenizer,
-    BertForMaskedLM,
-    BertConfig,
-    PreTrainedEncoderDecoder,
-    Model2Model,
-)
-
-from utils_summarization import (
-    CNNDailyMailDataset,
-    encode_for_summarization,
-    fit_to_block_size,
-    build_lm_labels,
-    build_mask,
-    compute_token_type_ids,
-)
-
-logger = logging.getLogger(__name__)
-logging.basicConfig(stream=sys.stdout, level=logging.INFO)
-
-
-def set_seed(args):
-    random.seed(args.seed)
-    np.random.seed(args.seed)
-    torch.manual_seed(args.seed)
-
-
-# ------------
-# Load dataset
-# ------------
-
-
-def load_and_cache_examples(args, tokenizer):
-    dataset = CNNDailyMailDataset(tokenizer, data_dir=args.data_dir)
-    return dataset
-
-
-def collate(data, tokenizer, block_size):
-    """ List of tuple as an input. """
-    # remove the files with empty an story/summary, encode and fit to block
-    data = filter(lambda x: not (len(x[0]) == 0 or len(x[1]) == 0), data)
-    data = [
-        encode_for_summarization(story, summary, tokenizer) for story, summary in data
-    ]
-    data = [
-        (
-            fit_to_block_size(story, block_size, tokenizer.pad_token_id),
-            fit_to_block_size(summary, block_size, tokenizer.pad_token_id),
-        )
-        for story, summary in data
-    ]
-
-    stories = torch.tensor([story for story, summary in data])
-    summaries = torch.tensor([summary for story, summary in data])
-    encoder_token_type_ids = compute_token_type_ids(stories, tokenizer.cls_token_id)
-    encoder_mask = build_mask(stories, tokenizer.pad_token_id)
-    decoder_mask = build_mask(summaries, tokenizer.pad_token_id)
-    lm_labels = build_lm_labels(summaries, tokenizer.pad_token_id)
-
-    return (
-        stories,
-        summaries,
-        encoder_token_type_ids,
-        encoder_mask,
-        decoder_mask,
-        lm_labels,
-    )
-
-
-# ----------
-# Optimizers
-# ----------
-
-
-class BertSumOptimizer(object):
-    """ Specific optimizer for BertSum.
-
-    As described in [1], the authors fine-tune BertSum for abstractive
-    summarization using two Adam Optimizers with different warm-up steps and
-    learning rate. They also use a custom learning rate scheduler.
-
-    [1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders."
-        arXiv preprint arXiv:1908.08345 (2019).
-    """
-
-    def __init__(self, model, lr, warmup_steps, beta_1=0.99, beta_2=0.999, eps=1e-8):
-        self.encoder = model.encoder
-        self.decoder = model.decoder
-        self.lr = lr
-        self.warmup_steps = warmup_steps
-
-        self.optimizers = {
-            "encoder": Adam(
-                model.encoder.parameters(),
-                lr=lr["encoder"],
-                betas=(beta_1, beta_2),
-                eps=eps,
-            ),
-            "decoder": Adam(
-                model.decoder.parameters(),
-                lr=lr["decoder"],
-                betas=(beta_1, beta_2),
-                eps=eps,
-            ),
-        }
-
-        self._step = 0
-
-    def _update_rate(self, stack):
-        return self.lr[stack] * min(
-            self._step ** (-0.5), self._step * self.warmup_steps[stack] ** (-0.5)
-        )
-
-    def zero_grad(self):
-        self.optimizer_decoder.zero_grad()
-        self.optimizer_encoder.zero_grad()
-
-    def step(self):
-        self._step += 1
-        for stack, optimizer in self.optimizers.items():
-            new_rate = self._update_rate(stack)
-            for param_group in optimizer.param_groups:
-                param_group["lr"] = new_rate
-            optimizer.step()
-
-
-# ------------
-# Train
-# ------------
-
-
-def train(args, model, tokenizer):
-    """ Fine-tune the pretrained model on the corpus. """
-    set_seed(args)
-
-    # Load the data
-    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
-    train_dataset = load_and_cache_examples(args, tokenizer)
-    train_sampler = RandomSampler(train_dataset)
-    model_collate_fn = functools.partial(collate, tokenizer=tokenizer, block_size=512)
-    train_dataloader = DataLoader(
-        train_dataset,
-        sampler=train_sampler,
-        batch_size=args.train_batch_size,
-        collate_fn=model_collate_fn,
-    )
-
-    # Training schedule
-    if args.max_steps > 0:
-        t_total = args.max_steps
-        args.num_train_epochs = t_total // (
-            len(train_dataloader) // args.gradient_accumulation_steps + 1
-        )
-    else:
-        t_total = (
-            len(train_dataloader)
-            // args.gradient_accumulation_steps
-            * args.num_train_epochs
-        )
-
-    # Prepare the optimizer
-    lr = {"encoder": 0.002, "decoder": 0.2}
-    warmup_steps = {"encoder": 20000, "decoder": 10000}
-    optimizer = BertSumOptimizer(model, lr, warmup_steps)
-
-    # Train
-    logger.info("***** Running training *****")
-    logger.info("  Num examples = %d", len(train_dataset))
-    logger.info("  Num Epochs = %d", args.num_train_epochs)
-    logger.info(
-        "  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size
-    )
-    logger.info(
-        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
-        args.train_batch_size * args.gradient_accumulation_steps
-        # * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
-    )
-    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
-    logger.info("  Total optimization steps = %d", t_total)
-
-    model.zero_grad()
-    train_iterator = trange(args.num_train_epochs, desc="Epoch", disable=True)
-
-    global_step = 0
-    tr_loss = 0.0
-    for _ in train_iterator:
-        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=True)
-        for step, batch in enumerate(epoch_iterator):
-            source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch
-
-            source = source.to(args.device)
-            target = target.to(args.device)
-            encoder_token_type_ids = encoder_token_type_ids.to(args.device)
-            encoder_mask = encoder_mask.to(args.device)
-            decoder_mask = decoder_mask.to(args.device)
-            lm_labels = lm_labels.to(args.device)
-
-            model.train()
-            outputs = model(
-                source,
-                target,
-                encoder_token_type_ids=encoder_token_type_ids,
-                encoder_attention_mask=encoder_mask,
-                decoder_attention_mask=decoder_mask,
-                decoder_lm_labels=lm_labels,
-            )
-
-            loss = outputs[0]
-            print(loss)
-            if args.gradient_accumulation_steps > 1:
-                loss /= args.gradient_accumulation_steps
-
-            loss.backward()
-
-            tr_loss += loss.item()
-            if (step + 1) % args.gradient_accumulation_steps == 0:
-                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
-                optimizer.step()
-                model.zero_grad()
-                global_step += 1
-
-            if args.max_steps > 0 and global_step > args.max_steps:
-                epoch_iterator.close()
-                break
-
-        if args.max_steps > 0 and global_step > args.max_steps:
-            train_iterator.close()
-            break
-
-    return global_step, tr_loss / global_step
-
-
-# ------------
-# Train
-# ------------
-
-
-def evaluate(args, model, tokenizer, prefix=""):
-    set_seed(args)
-
-    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
-    eval_dataset = load_and_cache_examples(args, tokenizer, evaluate=True)
-    eval_sampler = SequentialSampler(eval_dataset)
-    eval_dataloader = DataLoader(
-        eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size
-    )
-
-    # multi-gpu evaluate
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-        
-    logger.info("***** Running evaluation {} *****".format(prefix))
-    logger.info("  Num examples = %d", len(eval_dataset))
-    logger.info("  Batch size = %d", args.eval_batch_size)
-    eval_loss = 0.0
-    nb_eval_steps = 0
-    model.eval()
-
-    for batch in tqdm(eval_dataloader, desc="Evaluating"):
-        source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch
-
-        source = source.to(args.device)
-        target = target.to(args.device)
-        encoder_token_type_ids = encoder_token_type_ids.to(args.device)
-        encoder_mask = encoder_mask.to(args.device)
-        decoder_mask = decoder_mask.to(args.device)
-        lm_labels = lm_labels.to(args.device)
-
-        with torch.no_grad():
-            outputs = model(
-                source,
-                target,
-                encoder_token_type_ids=encoder_token_type_ids,
-                encoder_attention_mask=encoder_mask,
-                decoder_attention_mask=decoder_mask,
-                decoder_lm_labels=lm_labels,
-            )
-            lm_loss = outputs[0]
-            eval_loss += lm_loss.mean().item()
-        nb_eval_steps += 1
-
-    eval_loss = eval_loss / nb_eval_steps
-    perplexity = torch.exp(torch.tensor(eval_loss))
-
-    result = {"perplexity": perplexity}
-
-    # Save the evaluation's results
-    output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
-    if not os.path.exists(args.output_dir):
-        os.makedirs(args.output_dir)
-
-    with open(output_eval_file, "w") as writer:
-        logger.info("***** Eval results {} *****".format(prefix))
-        for key in sorted(result.keys()):
-            logger.info("  %s = %s", key, str(result[key]))
-            writer.write("%s = %s\n" % (key, str(result[key])))
-
-    return result
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    # Required parameters
-    parser.add_argument(
-        "--data_dir",
-        default=None,
-        type=str,
-        required=True,
-        help="The input training data file (a text file).",
-    )
-    parser.add_argument(
-        "--output_dir",
-        default=None,
-        type=str,
-        required=True,
-        help="The output directory where the model predictions and checkpoints will be written.",
-    )
-
-    # Optional parameters
-    parser.add_argument(
-        "--gradient_accumulation_steps",
-        type=int,
-        default=1,
-        help="Number of updates steps to accumulate before performing a backward/update pass.",
-    )
-    parser.add_argument(
-        "--do_evaluate",
-        type=bool,
-        default=False,
-        help="Run model evaluation on out-of-sample data.",
-    )
-    parser.add_argument("--do_train", type=bool, default=False, help="Run training.")
-    parser.add_argument(
-        "--do_overwrite_output_dir",
-        type=bool,
-        default=False,
-        help="Whether to overwrite the output dir.",
-    )
-    parser.add_argument(
-        "--model_name_or_path",
-        default="bert-base-cased",
-        type=str,
-        help="The model checkpoint to initialize the encoder and decoder's weights with.",
-    )
-    parser.add_argument(
-        "--model_type",
-        default="bert",
-        type=str,
-        help="The decoder architecture to be fine-tuned.",
-    )
-    parser.add_argument(
-        "--max_grad_norm", default=1.0, type=float, help="Max gradient norm."
-    )
-    parser.add_argument(
-        "--max_steps",
-        default=-1,
-        type=int,
-        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
-    )
-    parser.add_argument(
-        "--to_cpu", default=False, type=bool, help="Whether to force training on CPU."
-    )
-    parser.add_argument(
-        "--num_train_epochs",
-        default=10,
-        type=int,
-        help="Total number of training epochs to perform.",
-    )
-    parser.add_argument(
-        "--per_gpu_train_batch_size",
-        default=4,
-        type=int,
-        help="Batch size per GPU/CPU for training.",
-    )
-    parser.add_argument("--seed", default=42, type=int)
-    args = parser.parse_args()
-
-    if (
-        os.path.exists(args.output_dir)
-        and os.listdir(args.output_dir)
-        and args.do_train
-        and not args.do_overwrite_output_dir
-    ):
-        raise ValueError(
-            "Output directory ({}) already exists and is not empty. Use --do_overwrite_output_dir to overwrite.".format(
-                args.output_dir
-            )
-        )
-
-    # Set up training device
-    if args.to_cpu or not torch.cuda.is_available():
-        args.device = torch.device("cpu")
-        args.n_gpu = 0
-    else:
-        args.device = torch.device("cuda")
-        args.n_gpu = torch.cuda.device_count()
-
-    # Load pretrained model and tokenizer. The decoder's weights are randomly initialized.
-    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
-    config = BertConfig.from_pretrained(args.model_name_or_path)
-    decoder_model = BertForMaskedLM(config)
-    model = Model2Model.from_pretrained(
-        args.model_name_or_path, decoder_model=decoder_model
-    )
-
-    # Setup logging
-    logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
-        datefmt="%m/%d/%Y %H:%M:%S",
-        level=logging.INFO,
-    )
-    logger.warning(
-        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
-        0,
-        args.device,
-        args.n_gpu,
-        False,
-        False,
-    )
-
-    logger.info("Training/evaluation parameters %s", args)
-
-    # Train the model
-    model.to(args.device)
-    if args.do_train:
-        global_step, tr_loss = train(args, model, tokenizer)
-        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
-
-        if not os.path.exists(args.output_dir):
-            os.makedirs(args.output_dir)
-
-        logger.info("Saving model checkpoint to %s", args.output_dir)
-
-        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
-        # They can then be reloaded using `from_pretrained()`
-        model_to_save = (
-            model.module if hasattr(model, "module") else model
-        )  # Take care of distributed/parallel training
-        model_to_save.save_pretrained(args.output_dir)
-        tokenizer.save_pretrained(args.output_dir)
-        torch.save(args, os.path.join(args.output_dir, "training_arguments.bin"))
-
-    # Evaluate the model
-    results = {}
-    if args.do_evaluate:
-        checkpoints = []
-        logger.info("Evaluate the following checkpoints: %s", checkpoints)
-        for checkpoint in checkpoints:
-            encoder_checkpoint = os.path.join(checkpoint, "encoder")
-            decoder_checkpoint = os.path.join(checkpoint, "decoder")
-            model = PreTrainedEncoderDecoder.from_pretrained(
-                encoder_checkpoint, decoder_checkpoint
-            )
-            model.to(args.device)
-            results = "placeholder"
-
-    return results
-
-
-if __name__ == "__main__":
-    main()
@@ -1,93 +1,23 @@
-import os
 import tensorflow as tf
 import tensorflow_datasets
-from transformers import BertTokenizer, TFBertForSequenceClassification, BertConfig, glue_convert_examples_to_features, BertForSequenceClassification, glue_processors
+from pytorch_transformers import BertTokenizer, TFBertForSequenceClassification, glue_convert_examples_to_features

-# script parameters
-BATCH_SIZE = 32
-EVAL_BATCH_SIZE = BATCH_SIZE * 2
-USE_XLA = False
-USE_AMP = False
-EPOCHS = 3
+# Load tokenizer, model, dataset
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+model = TFBertForSequenceClassification.from_pretrained('bert-base-uncased')
+dataset, dataset_info = tensorflow_datasets.load("glue/mrpc", split="train", with_info=True)
+print(dataset_info)

-TASK = "mrpc"
+# Prepare dataset for GLUE
+dataset = glue_convert_examples_to_features(dataset, tokenizer, task='mrpc', max_length=64)
+dataset = dataset.batch(32)

-if TASK == "sst-2":
-    TFDS_TASK = "sst2"
-elif TASK == "sts-b":
-    TFDS_TASK = "stsb"
-else: 
-    TFDS_TASK = TASK
+# Compile model for training
+learning_rate = tf.keras.optimizers.schedules.PolynomialDecay(5e-5, 300, 0)
+model.compile(optimizer=tf.keras.optimizers.Adam(
+              learning_rate=learning_rate),
+              loss=lambda x, y: tf.keras.backend.sparse_categorical_crossentropy(x, y, from_logits=True),
+              metrics=['accuracy'])

-num_labels = len(glue_processors[TASK]().get_labels())
-print(num_labels)
-
-tf.config.optimizer.set_jit(USE_XLA)
-tf.config.optimizer.set_experimental_options({"auto_mixed_precision": USE_AMP})
-
-# Load tokenizer and model from pretrained model/vocabulary. Specify the number of labels to classify (2+: classification, 1: regression)
-config = BertConfig.from_pretrained("bert-base-cased", num_labels=num_labels)
-tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
-model = TFBertForSequenceClassification.from_pretrained('bert-base-cased', config=config)
-
-# Load dataset via TensorFlow Datasets
-data, info = tensorflow_datasets.load(f'glue/{TFDS_TASK}', with_info=True)
-train_examples = info.splits['train'].num_examples
-
-# MNLI expects either validation_matched or validation_mismatched
-valid_examples = info.splits['validation'].num_examples
-
-# Prepare dataset for GLUE as a tf.data.Dataset instance
-train_dataset = glue_convert_examples_to_features(data['train'], tokenizer, 128, TASK)
-
-# MNLI expects either validation_matched or validation_mismatched
-valid_dataset = glue_convert_examples_to_features(data['validation'], tokenizer, 128, TASK)
-train_dataset = train_dataset.shuffle(128).batch(BATCH_SIZE).repeat(-1)
-valid_dataset = valid_dataset.batch(EVAL_BATCH_SIZE)
-
-# Prepare training: Compile tf.keras model with optimizer, loss and learning rate schedule 
-opt = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08)
-if USE_AMP:
-    # loss scaling is currently required when using mixed precision
-    opt = tf.keras.mixed_precision.experimental.LossScaleOptimizer(opt, 'dynamic')
-
-
-if num_labels == 1:
-    loss = tf.keras.losses.MeanSquaredError()
-else:
-    loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
-
-metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')
-model.compile(optimizer=opt, loss=loss, metrics=[metric])
-
-# Train and evaluate using tf.keras.Model.fit()
-train_steps = train_examples//BATCH_SIZE
-valid_steps = valid_examples//EVAL_BATCH_SIZE
-
-history = model.fit(train_dataset, epochs=EPOCHS, steps_per_epoch=train_steps,
-                    validation_data=valid_dataset, validation_steps=valid_steps)
-
-# Save TF2 model
-os.makedirs('./save/', exist_ok=True)
-model.save_pretrained('./save/')
-
-if TASK == "mrpc":
-    # Load the TensorFlow model in PyTorch for inspection
-    # This is to demo the interoperability between the two frameworks, you don't have to 
-    # do this in real life (you can run the inference on the TF model).
-    pytorch_model = BertForSequenceClassification.from_pretrained('./save/', from_tf=True)
-
-    # Quickly test a few predictions - MRPC is a paraphrasing task, let's see if our model learned the task
-    sentence_0 = 'This research was consistent with his findings.'
-    sentence_1 = 'His findings were compatible with this research.'
-    sentence_2 = 'His findings were not compatible with this research.'
-    inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
-    inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
-
-    del inputs_1["special_tokens_mask"]
-    del inputs_2["special_tokens_mask"]
-
-    pred_1 = pytorch_model(**inputs_1)[0].argmax().item()
-    pred_2 = pytorch_model(**inputs_2)[0].argmax().item()
-    print('sentence_1 is', 'a paraphrase' if pred_1 else 'not a paraphrase', 'of sentence_0')
-    print('sentence_2 is', 'a paraphrase' if pred_2 else 'not a paraphrase', 'of sentence_0')
+# Train model
+model.fit(dataset, epochs=3)
@@ -1,515 +0,0 @@
-# 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.
-""" Finetuning multi-lingual models on XNLI (Bert, DistilBERT, XLM).
-    Adapted from `examples/run_glue.py`"""
-
-from __future__ import absolute_import, division, print_function
-
-import argparse
-import glob
-import logging
-import os
-import random
-
-import numpy as np
-import torch
-from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
-                              TensorDataset)
-from torch.utils.data.distributed import DistributedSampler
-
-try:
-    from torch.utils.tensorboard import SummaryWriter
-except:
-    from tensorboardX import SummaryWriter
-
-from tqdm import tqdm, trange
-
-from transformers import (WEIGHTS_NAME, 
-                          BertConfig, BertForSequenceClassification, BertTokenizer,
-                          XLMConfig, XLMForSequenceClassification, XLMTokenizer,
-                          DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
-
-from transformers import AdamW, get_linear_schedule_with_warmup
-
-from transformers import xnli_compute_metrics as compute_metrics
-from transformers import xnli_output_modes as output_modes
-from transformers import xnli_processors as processors
-
-from transformers import glue_convert_examples_to_features as convert_examples_to_features
-
-logger = logging.getLogger(__name__)
-
-ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, DistilBertConfig, XLMConfig)), ())
-
-MODEL_CLASSES = {
-    'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
-    'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
-    'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
-}
-
-
-def set_seed(args):
-    random.seed(args.seed)
-    np.random.seed(args.seed)
-    torch.manual_seed(args.seed)
-    if args.n_gpu > 0:
-        torch.cuda.manual_seed_all(args.seed)
-
-
-def train(args, train_dataset, model, tokenizer):
-    """ Train the model """
-    if args.local_rank in [-1, 0]:
-        tb_writer = SummaryWriter()
-
-    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
-    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
-    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
-
-    if args.max_steps > 0:
-        t_total = args.max_steps
-        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
-    else:
-        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
-
-    # Prepare optimizer and schedule (linear warmup and decay)
-    no_decay = ['bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
-        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
-        ]
-    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
-    if args.fp16:
-        try:
-            from apex import amp
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
-        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
-
-    # multi-gpu training (should be after apex fp16 initialization)
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    # Distributed training (should be after apex fp16 initialization)
-    if args.local_rank != -1:
-        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
-                                                          output_device=args.local_rank,
-                                                          find_unused_parameters=True)
-
-    # Train!
-    logger.info("***** Running training *****")
-    logger.info("  Num examples = %d", len(train_dataset))
-    logger.info("  Num Epochs = %d", args.num_train_epochs)
-    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
-    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
-                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
-    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
-    logger.info("  Total optimization steps = %d", t_total)
-
-    global_step = 0
-    tr_loss, logging_loss = 0.0, 0.0
-    model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
-    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
-    for _ in train_iterator:
-        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
-        for step, batch in enumerate(epoch_iterator):
-            model.train()
-            batch = tuple(t.to(args.device) for t in batch)
-            inputs = {'input_ids':      batch[0],
-                      'attention_mask': batch[1],
-                      'labels':         batch[3]}
-            if args.model_type != 'distilbert':
-                inputs['token_type_ids'] = batch[2] if args.model_type in ['bert'] else None  # XLM and DistilBERT don't use segment_ids
-            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 training
-            if args.gradient_accumulation_steps > 1:
-                loss = loss / args.gradient_accumulation_steps
-
-            if args.fp16:
-                with amp.scale_loss(loss, optimizer) as scaled_loss:
-                    scaled_loss.backward()
-            else:
-                loss.backward()
-
-            tr_loss += loss.item()
-            if (step + 1) % args.gradient_accumulation_steps == 0:
-                if args.fp16:
-                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
-                else:
-                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
-
-                optimizer.step()
-                scheduler.step()  # Update learning rate schedule
-                model.zero_grad()
-                global_step += 1
-
-                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
-                    # Log metrics
-                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
-                        results = evaluate(args, model, tokenizer)
-                        for key, value in results.items():
-                            tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
-                    tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
-                    tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
-                    logging_loss = tr_loss
-
-                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
-                    # Save model checkpoint
-                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
-                    if not os.path.exists(output_dir):
-                        os.makedirs(output_dir)
-                    model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
-                    model_to_save.save_pretrained(output_dir)
-                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
-                    logger.info("Saving model checkpoint to %s", output_dir)
-
-            if args.max_steps > 0 and global_step > args.max_steps:
-                epoch_iterator.close()
-                break
-        if args.max_steps > 0 and global_step > args.max_steps:
-            train_iterator.close()
-            break
-
-    if args.local_rank in [-1, 0]:
-        tb_writer.close()
-
-    return global_step, tr_loss / global_step
-
-
-def evaluate(args, model, tokenizer, prefix=""):
-    eval_task_names = (args.task_name,)
-    eval_outputs_dirs = (args.output_dir,)
-
-    results = {}
-    for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
-        eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)
-
-        if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(eval_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(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
-        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
-
-        # multi-gpu eval
-        if args.n_gpu > 1:
-            model = torch.nn.DataParallel(model)
-
-        # Eval!
-        logger.info("***** Running evaluation {} *****".format(prefix))
-        logger.info("  Num examples = %d", len(eval_dataset))
-        logger.info("  Batch size = %d", args.eval_batch_size)
-        eval_loss = 0.0
-        nb_eval_steps = 0
-        preds = None
-        out_label_ids = None
-        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],
-                          'labels':         batch[3]}
-                if args.model_type != 'distilbert':
-                    inputs['token_type_ids'] = batch[2] if args.model_type in ['bert'] else None  # XLM and DistilBERT don't use segment_ids
-                outputs = model(**inputs)
-                tmp_eval_loss, logits = outputs[:2]
-
-                eval_loss += tmp_eval_loss.mean().item()
-            nb_eval_steps += 1
-            if preds is None:
-                preds = logits.detach().cpu().numpy()
-                out_label_ids = inputs['labels'].detach().cpu().numpy()
-            else:
-                preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
-                out_label_ids = np.append(out_label_ids, inputs['labels'].detach().cpu().numpy(), axis=0)
-
-        eval_loss = eval_loss / nb_eval_steps
-        if args.output_mode == "classification":
-            preds = np.argmax(preds, axis=1)
-        else:
-            raise ValueError('No other `output_mode` for XNLI.')
-        result = compute_metrics(eval_task, preds, out_label_ids)
-        results.update(result)
-
-        output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
-        with open(output_eval_file, "w") as writer:
-            logger.info("***** Eval results {} *****".format(prefix))
-            for key in sorted(result.keys()):
-                logger.info("  %s = %s", key, str(result[key]))
-                writer.write("%s = %s\n" % (key, str(result[key])))
-
-    return results
-
-
-def load_and_cache_examples(args, task, tokenizer, evaluate=False):
-    if args.local_rank not in [-1, 0] and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    processor = processors[task](language=args.language, train_language=args.train_language)
-    output_mode = output_modes[task]
-    # Load data features from cache or dataset file
-    cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}_{}'.format(
-        'test' if evaluate else 'train',
-        list(filter(None, args.model_name_or_path.split('/'))).pop(),
-        str(args.max_seq_length),
-        str(task),
-        str(args.train_language if (not evaluate and args.train_language is not None) else args.language)))
-    if os.path.exists(cached_features_file) and not args.overwrite_cache:
-        logger.info("Loading features from cached file %s", cached_features_file)
-        features = torch.load(cached_features_file)
-    else:
-        logger.info("Creating features from dataset file at %s", args.data_dir)
-        label_list = processor.get_labels()
-        examples = processor.get_test_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
-        features = convert_examples_to_features(examples,
-                                                tokenizer,
-                                                label_list=label_list,
-                                                max_length=args.max_seq_length,
-                                                output_mode=output_mode,
-                                                pad_on_left=False,
-                                                pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
-                                                pad_token_segment_id=0,
-        )
-        if args.local_rank in [-1, 0]:
-            logger.info("Saving features into cached file %s", cached_features_file)
-            torch.save(features, cached_features_file)
-
-    if args.local_rank == 0 and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    # Convert to Tensors and build dataset
-    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
-    all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
-    all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
-    if output_mode == "classification":
-        all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
-    else:
-        raise ValueError('No other `output_mode` for XNLI.')
-
-    dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
-    return dataset
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    ## Required parameters
-    parser.add_argument("--data_dir", default=None, type=str, required=True,
-                        help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
-    parser.add_argument("--model_type", default=None, type=str, required=True,
-                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
-    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
-                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
-    parser.add_argument("--language", default=None, type=str, required=True,
-                        help="Evaluation language. Also train language if `train_language` is set to None.")
-    parser.add_argument("--train_language", default=None, type=str,
-                        help="Train language if is different of the evaluation language.")
-    parser.add_argument("--output_dir", default=None, type=str, required=True,
-                        help="The output directory where the model predictions and checkpoints will be written.")
-
-    ## Other parameters
-    parser.add_argument("--config_name", default="", type=str,
-                        help="Pretrained config name or path if not the same as model_name")
-    parser.add_argument("--tokenizer_name", default="", type=str,
-                        help="Pretrained tokenizer name or path if not the same as model_name")
-    parser.add_argument("--cache_dir", default="", type=str,
-                        help="Where do you want to store the pre-trained models downloaded from s3")
-    parser.add_argument("--max_seq_length", default=128, type=int,
-                        help="The maximum total input sequence length after tokenization. Sequences longer "
-                             "than this will be truncated, sequences shorter will be padded.")
-    parser.add_argument("--do_train", action='store_true',
-                        help="Whether to run training.")
-    parser.add_argument("--do_eval", action='store_true',
-                        help="Whether to run eval on the test 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('--gradient_accumulation_steps', type=int, default=1,
-                        help="Number of updates steps to accumulate before performing a backward/update pass.")
-    parser.add_argument("--learning_rate", default=5e-5, type=float,
-                        help="The initial learning rate for Adam.")
-    parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight 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="Avoid using CUDA when available")
-    parser.add_argument('--overwrite_output_dir', action='store_true',
-                        help="Overwrite the content of the output directory")
-    parser.add_argument('--overwrite_cache', action='store_true',
-                        help="Overwrite the cached training and evaluation sets")
-    parser.add_argument('--seed', type=int, default=42,
-                        help="random seed for initialization")
-
-    parser.add_argument('--fp16', action='store_true',
-                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
-    parser.add_argument('--fp16_opt_level', type=str, default='O1',
-                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
-                             "See details at https://nvidia.github.io/apex/amp.html")
-    parser.add_argument("--local_rank", type=int, default=-1,
-                        help="For distributed training: local_rank")
-    parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.")
-    parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
-    args = parser.parse_args()
-
-    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
-        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
-
-    # Setup distant debugging if needed
-    if args.server_ip and args.server_port:
-        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
-        import ptvsd
-        print("Waiting for debugger attach")
-        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
-        ptvsd.wait_for_attach()
-
-    # Setup CUDA, GPU & distributed training
-    if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
-    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
-        torch.cuda.set_device(args.local_rank)
-        device = torch.device("cuda", args.local_rank)
-        torch.distributed.init_process_group(backend='nccl')
-        args.n_gpu = 1
-    args.device = device
-
-    # Setup logging
-    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
-                        datefmt = '%m/%d/%Y %H:%M:%S',
-                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
-    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
-                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
-
-    # Set seed
-    set_seed(args)
-
-    # Prepare XNLI task
-    args.task_name = 'xnli'
-    if args.task_name not in processors:
-        raise ValueError("Task not found: %s" % (args.task_name))
-    processor = processors[args.task_name](language=args.language, train_language=args.train_language)
-    args.output_mode = output_modes[args.task_name]
-    label_list = processor.get_labels()
-    num_labels = len(label_list)
-
-    # Load pretrained model and tokenizer
-    if args.local_rank not in [-1, 0]:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    args.model_type = args.model_type.lower()
-    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
-                                          num_labels=num_labels,
-                                          finetuning_task=args.task_name,
-                                          cache_dir=args.cache_dir if args.cache_dir else None)
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
-                                                do_lower_case=args.do_lower_case,
-                                                cache_dir=args.cache_dir if args.cache_dir else None)
-    model = model_class.from_pretrained(args.model_name_or_path,
-                                        from_tf=bool('.ckpt' in args.model_name_or_path),
-                                        config=config,
-                                        cache_dir=args.cache_dir if args.cache_dir else None)
-
-    if args.local_rank == 0:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    model.to(args.device)
-
-    logger.info("Training/evaluation parameters %s", args)
-
-
-    # Training
-    if args.do_train:
-        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
-        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
-        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
-
-
-    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
-    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
-        logger.info("Saving model checkpoint to %s", args.output_dir)
-        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
-        # They can then be reloaded using `from_pretrained()`
-        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
-        model_to_save.save_pretrained(args.output_dir)
-        tokenizer.save_pretrained(args.output_dir)
-
-        # Good practice: save your training arguments together with the trained model
-        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
-
-        # 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
-    results = {}
-    if args.do_eval and args.local_rank in [-1, 0]:
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
-        checkpoints = [args.output_dir]
-        if args.eval_all_checkpoints:
-            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
-            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
-        logger.info("Evaluate the following checkpoints: %s", checkpoints)
-        for checkpoint in checkpoints:
-            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
-            prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
-            
-            model = model_class.from_pretrained(checkpoint)
-            model.to(args.device)
-            result = evaluate(args, model, tokenizer, prefix=prefix)
-            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
-            results.update(result)
-
-    return results
-
-
-if __name__ == "__main__":
-    main()
@@ -13,7 +13,7 @@
 # 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.
-""" Multiple choice fine-tuning: utilities to work with multiple choice tasks of reading comprehension  """
+""" BERT multiple choice fine-tuning: utilities to work with multiple choice tasks of reading comprehension  """

 from __future__ import absolute_import, division, print_function

@@ -26,8 +26,6 @@ import json
 import csv
 import glob
 import tqdm
-from typing import List
-from transformers import PreTrainedTokenizer


 logger = logging.getLogger(__name__)
@@ -36,13 +34,13 @@ logger = logging.getLogger(__name__)
 class InputExample(object):
    """A single training/test example for multiple choice"""

-    def __init__(self, example_id, question, contexts, endings, label=None):
+    def __init__(self, example_id, question,  contexts, endings, label=None):
        """Constructs a InputExample.

        Args:
            example_id: Unique id for the example.
            contexts: list of str. The untokenized text of the first sequence (context of corresponding question).
-            question: string. The untokenized text of the second sequence (question).
+            question: string. The untokenized text of the second sequence (qustion).
            endings: list of str. multiple choice's options. Its length must be equal to contexts' length.
            label: (Optional) string. The label of the example. This should be
            specified for train and dev examples, but not for test examples.
@@ -68,7 +66,7 @@ class InputFeatures(object):
                'input_mask': input_mask,
                'segment_ids': segment_ids
            }
-            for input_ids, input_mask, segment_ids in choices_features
+            for _, input_ids, input_mask, segment_ids in choices_features
        ]
        self.label = label

@@ -194,7 +192,7 @@ class SwagProcessor(DataProcessor):
            return lines


-    def _create_examples(self, lines: List[List[str]], type: str):
+    def _create_examples(self, lines, type):
        """Creates examples for the training and dev sets."""
        if type == "train" and lines[0][-1] != 'label':
            raise ValueError(
@@ -302,18 +300,24 @@ class ArcProcessor(DataProcessor):
        return examples


-def convert_examples_to_features(
-    examples: List[InputExample],
-    label_list: List[str],
-    max_length: int,
-    tokenizer: PreTrainedTokenizer,
-    pad_token_segment_id=0,
-    pad_on_left=False,
-    pad_token=0,
-    mask_padding_with_zero=True,
-) -> List[InputFeatures]:
-    """
-    Loads a data file into a list of `InputFeatures`
+def convert_examples_to_features(examples, label_list, max_seq_length,
+                                 tokenizer,
+                                 cls_token_at_end=False,
+                                 cls_token='[CLS]',
+                                 cls_token_segment_id=1,
+                                 sep_token='[SEP]',
+                                 sequence_a_segment_id=0,
+                                 sequence_b_segment_id=1,
+                                 sep_token_extra=False,
+                                 pad_token_segment_id=0,
+                                 pad_on_left=False,
+                                 pad_token=0,
+                                 mask_padding_with_zero=True):
+    """ Loads a data file into a list of `InputBatch`s
+        `cls_token_at_end` define the location of the CLS token:
+            - False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]
+            - True (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS]
+        `cls_token_segment_id` define the segment id associated to the CLS token (0 for BERT, 2 for XLNet)
    """

    label_map = {label : i for i, label in enumerate(label_list)}
@@ -324,70 +328,125 @@ def convert_examples_to_features(
            logger.info("Writing example %d of %d" % (ex_index, len(examples)))
        choices_features = []
        for ending_idx, (context, ending) in enumerate(zip(example.contexts, example.endings)):
-            text_a = context
+            tokens_a = tokenizer.tokenize(context)
+            tokens_b = None
            if example.question.find("_") != -1:
-                # this is for cloze question
-                text_b = example.question.replace("_", ending)
+                #this is for cloze question
+                tokens_b = tokenizer.tokenize(example.question.replace("_", ending))
            else:
-                text_b = example.question + " " + ending
+                tokens_b = tokenizer.tokenize(example.question + " " + ending)
+                # you can add seq token between quesiotn and ending. This does not make too much difference.
+                # tokens_b = tokenizer.tokenize(example.question)
+                # tokens_b += [sep_token]
+                # if sep_token_extra:
+                #     tokens_b += [sep_token]
+                # tokens_b += tokenizer.tokenize(ending)

-            inputs = tokenizer.encode_plus(
-                text_a,
-                text_b,
-                add_special_tokens=True,
-                max_length=max_length,
-            )
-            if 'num_truncated_tokens' in inputs and inputs['num_truncated_tokens'] > 0:
-                logger.info('Attention! you are cropping tokens (swag task is ok). '
-                        'If you are training ARC and RACE and you are poping question + options,'
-                        'you need to try to use a bigger max seq length!')
+            special_tokens_count = 4 if sep_token_extra else 3
+            _truncate_seq_pair(tokens_a, tokens_b, max_seq_length - special_tokens_count)

-            input_ids, token_type_ids = inputs["input_ids"], inputs["token_type_ids"]
+            # The convention in BERT is:
+            # (a) For sequence pairs:
+            #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
+            #  type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1
+            # (b) For single sequences:
+            #  tokens:   [CLS] the dog is hairy . [SEP]
+            #  type_ids:   0   0   0   0  0     0   0
+            #
+            # Where "type_ids" are used to indicate whether this is the first
+            # sequence or the second sequence. The embedding vectors for `type=0` and
+            # `type=1` were learned during pre-training and are added to the wordpiece
+            # embedding vector (and position vector). This is not *strictly* necessary
+            # since the [SEP] token unambiguously separates the sequences, but it makes
+            # it easier for the model to learn the concept of sequences.
+            #
+            # For classification tasks, the first vector (corresponding to [CLS]) is
+            # used as as the "sentence vector". Note that this only makes sense because
+            # the entire model is fine-tuned.
+            tokens = tokens_a + [sep_token]
+            if sep_token_extra:
+                # roberta uses an extra separator b/w pairs of sentences
+                tokens += [sep_token]
+
+            segment_ids = [sequence_a_segment_id] * len(tokens)
+
+            if tokens_b:
+                tokens += tokens_b + [sep_token]
+                segment_ids += [sequence_b_segment_id] * (len(tokens_b) + 1)
+
+            if cls_token_at_end:
+                tokens = tokens + [cls_token]
+                segment_ids = segment_ids + [cls_token_segment_id]
+            else:
+                tokens = [cls_token] + tokens
+                segment_ids = [cls_token_segment_id] + segment_ids
+
+            input_ids = tokenizer.convert_tokens_to_ids(tokens)

            # The mask has 1 for real tokens and 0 for padding tokens. Only real
            # tokens are attended to.
-            attention_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
+            input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)

            # Zero-pad up to the sequence length.
-            padding_length = max_length - len(input_ids)
+            padding_length = max_seq_length - len(input_ids)
            if pad_on_left:
                input_ids = ([pad_token] * padding_length) + input_ids
-                attention_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + attention_mask
-                token_type_ids = ([pad_token_segment_id] * padding_length) + token_type_ids
+                input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask
+                segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids
            else:
                input_ids = input_ids + ([pad_token] * padding_length)
-                attention_mask = attention_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
-                token_type_ids = token_type_ids + ([pad_token_segment_id] * padding_length)
-
-            assert len(input_ids) == max_length
-            assert len(attention_mask) == max_length
-            assert len(token_type_ids) == max_length
-            choices_features.append((input_ids, attention_mask, token_type_ids))
-
+                input_mask = input_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
+                segment_ids = segment_ids + ([pad_token_segment_id] * padding_length)

+            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 = label_map[example.label]

        if ex_index < 2:
            logger.info("*** Example ***")
            logger.info("race_id: {}".format(example.example_id))
-            for choice_idx, (input_ids, attention_mask, token_type_ids) in enumerate(choices_features):
+            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("attention_mask: {}".format(' '.join(map(str, attention_mask))))
-                logger.info("token_type_ids: {}".format(' '.join(map(str, token_type_ids))))
+                logger.info("input_mask: {}".format(' '.join(map(str, input_mask))))
+                logger.info("segment_ids: {}".format(' '.join(map(str, segment_ids))))
                logger.info("label: {}".format(label))

        features.append(
            InputFeatures(
-                example_id=example.example_id,
-                choices_features=choices_features,
-                label=label,
+                example_id = example.example_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.
+
+    # However, since we'd better not to remove tokens of options and questions, you can choose to use a bigger
+    # length or only pop from context
+    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:
+            logger.info('Attention! you are removing from token_b (swag task is ok). '
+                        'If you are training ARC and RACE (you are poping question + options), '
+                        'you need to try to use a bigger max seq length!')
+            tokens_b.pop()


 processors = {
@@ -397,7 +456,7 @@ processors = {
 }


-MULTIPLE_CHOICE_TASKS_NUM_LABELS = {
+GLUE_TASKS_NUM_LABELS = {
    "race", 4,
    "swag", 4,
    "arc", 4
@@ -1,212 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" Named entity recognition fine-tuning: utilities to work with CoNLL-2003 task. """
-
-from __future__ import absolute_import, division, print_function
-
-import logging
-import os
-from io import open
-
-logger = logging.getLogger(__name__)
-
-
-class InputExample(object):
-    """A single training/test example for token classification."""
-
-    def __init__(self, guid, words, labels):
-        """Constructs a InputExample.
-
-        Args:
-            guid: Unique id for the example.
-            words: list. The words of the sequence.
-            labels: (Optional) list. The labels for each word of the sequence. This should be
-            specified for train and dev examples, but not for test examples.
-        """
-        self.guid = guid
-        self.words = words
-        self.labels = labels
-
-
-class InputFeatures(object):
-    """A single set of features of data."""
-
-    def __init__(self, input_ids, input_mask, segment_ids, label_ids):
-        self.input_ids = input_ids
-        self.input_mask = input_mask
-        self.segment_ids = segment_ids
-        self.label_ids = label_ids
-
-
-def read_examples_from_file(data_dir, mode):
-    file_path = os.path.join(data_dir, "{}.txt".format(mode))
-    guid_index = 1
-    examples = []
-    with open(file_path, encoding="utf-8") as f:
-        words = []
-        labels = []
-        for line in f:
-            if line.startswith("-DOCSTART-") or line == "" or line == "\n":
-                if words:
-                    examples.append(InputExample(guid="{}-{}".format(mode, guid_index),
-                                                 words=words,
-                                                 labels=labels))
-                    guid_index += 1
-                    words = []
-                    labels = []
-            else:
-                splits = line.split(" ")
-                words.append(splits[0])
-                if len(splits) > 1:
-                    labels.append(splits[-1].replace("\n", ""))
-                else:
-                    # Examples could have no label for mode = "test"
-                    labels.append("O")
-        if words:
-            examples.append(InputExample(guid="%s-%d".format(mode, guid_index),
-                                         words=words,
-                                         labels=labels))
-    return examples
-
-
-def convert_examples_to_features(examples,
-                                 label_list,
-                                 max_seq_length,
-                                 tokenizer,
-                                 cls_token_at_end=False,
-                                 cls_token="[CLS]",
-                                 cls_token_segment_id=1,
-                                 sep_token="[SEP]",
-                                 sep_token_extra=False,
-                                 pad_on_left=False,
-                                 pad_token=0,
-                                 pad_token_segment_id=0,
-                                 pad_token_label_id=-1,
-                                 sequence_a_segment_id=0,
-                                 mask_padding_with_zero=True):
-    """ Loads a data file into a list of `InputBatch`s
-        `cls_token_at_end` define the location of the CLS token:
-            - False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]
-            - True (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS]
-        `cls_token_segment_id` define the segment id associated to the CLS token (0 for BERT, 2 for XLNet)
-    """
-
-    label_map = {label: i for i, label in enumerate(label_list)}
-
-    features = []
-    for (ex_index, example) in enumerate(examples):
-        if ex_index % 10000 == 0:
-            logger.info("Writing example %d of %d", ex_index, len(examples))
-
-        tokens = []
-        label_ids = []
-        for word, label in zip(example.words, example.labels):
-            word_tokens = tokenizer.tokenize(word)
-            tokens.extend(word_tokens)
-            # Use the real label id for the first token of the word, and padding ids for the remaining tokens
-            label_ids.extend([label_map[label]] + [pad_token_label_id] * (len(word_tokens) - 1))
-
-        # Account for [CLS] and [SEP] with "- 2" and with "- 3" for RoBERTa.
-        special_tokens_count = 3 if sep_token_extra else 2
-        if len(tokens) > max_seq_length - special_tokens_count:
-            tokens = tokens[:(max_seq_length - special_tokens_count)]
-            label_ids = label_ids[:(max_seq_length - special_tokens_count)]
-
-        # The convention in BERT is:
-        # (a) For sequence pairs:
-        #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
-        #  type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1
-        # (b) For single sequences:
-        #  tokens:   [CLS] the dog is hairy . [SEP]
-        #  type_ids:   0   0   0   0  0     0   0
-        #
-        # Where "type_ids" are used to indicate whether this is the first
-        # sequence or the second sequence. The embedding vectors for `type=0` and
-        # `type=1` were learned during pre-training and are added to the wordpiece
-        # embedding vector (and position vector). This is not *strictly* necessary
-        # since the [SEP] token unambiguously separates the sequences, but it makes
-        # it easier for the model to learn the concept of sequences.
-        #
-        # For classification tasks, the first vector (corresponding to [CLS]) is
-        # used as as the "sentence vector". Note that this only makes sense because
-        # the entire model is fine-tuned.
-        tokens += [sep_token]
-        label_ids += [pad_token_label_id]
-        if sep_token_extra:
-            # roberta uses an extra separator b/w pairs of sentences
-            tokens += [sep_token]
-            label_ids += [pad_token_label_id]
-        segment_ids = [sequence_a_segment_id] * len(tokens)
-
-        if cls_token_at_end:
-            tokens += [cls_token]
-            label_ids += [pad_token_label_id]
-            segment_ids += [cls_token_segment_id]
-        else:
-            tokens = [cls_token] + tokens
-            label_ids = [pad_token_label_id] + label_ids
-            segment_ids = [cls_token_segment_id] + segment_ids
-
-        input_ids = tokenizer.convert_tokens_to_ids(tokens)
-
-        # The mask has 1 for real tokens and 0 for padding tokens. Only real
-        # tokens are attended to.
-        input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
-
-        # Zero-pad up to the sequence length.
-        padding_length = max_seq_length - len(input_ids)
-        if pad_on_left:
-            input_ids = ([pad_token] * padding_length) + input_ids
-            input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask
-            segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids
-            label_ids = ([pad_token_label_id] * padding_length) + label_ids
-        else:
-            input_ids += ([pad_token] * padding_length)
-            input_mask += ([0 if mask_padding_with_zero else 1] * padding_length)
-            segment_ids += ([pad_token_segment_id] * padding_length)
-            label_ids += ([pad_token_label_id] * padding_length)
-
-        assert len(input_ids) == max_seq_length
-        assert len(input_mask) == max_seq_length
-        assert len(segment_ids) == max_seq_length
-        assert len(label_ids) == max_seq_length
-
-        if ex_index < 5:
-            logger.info("*** Example ***")
-            logger.info("guid: %s", example.guid)
-            logger.info("tokens: %s", " ".join([str(x) for x in tokens]))
-            logger.info("input_ids: %s", " ".join([str(x) for x in input_ids]))
-            logger.info("input_mask: %s", " ".join([str(x) for x in input_mask]))
-            logger.info("segment_ids: %s", " ".join([str(x) for x in segment_ids]))
-            logger.info("label_ids: %s", " ".join([str(x) for x in label_ids]))
-
-        features.append(
-                InputFeatures(input_ids=input_ids,
-                              input_mask=input_mask,
-                              segment_ids=segment_ids,
-                              label_ids=label_ids))
-    return features
-
-
-def get_labels(path):
-    if path:
-        with open(path, "r") as f:
-            labels = f.read().splitlines()
-        if "O" not in labels:
-            labels = ["O"] + labels
-        return labels
-    else:
-        return ["O", "B-MISC", "I-MISC",  "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"]
@@ -23,9 +23,8 @@ import logging
 import math
 import collections
 from io import open
-from tqdm import tqdm

-from transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize
+from pytorch_transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize

 # Required by XLNet evaluation method to compute optimal threshold (see write_predictions_extended() method)
 from utils_squad_evaluate import find_all_best_thresh_v2, make_qid_to_has_ans, get_raw_scores
@@ -193,8 +192,7 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
                                 cls_token='[CLS]', sep_token='[SEP]', pad_token=0,
                                 sequence_a_segment_id=0, sequence_b_segment_id=1,
                                 cls_token_segment_id=0, pad_token_segment_id=0,
-                                 mask_padding_with_zero=True,
-                                 sequence_a_is_doc=False):
+                                 mask_padding_with_zero=True):
    """Loads a data file into a list of `InputBatch`s."""

    unique_id = 1000000000
@@ -203,7 +201,7 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
    # f = np.zeros((max_N, max_M), dtype=np.float32)

    features = []
-    for (example_index, example) in enumerate(tqdm(examples)):
+    for (example_index, example) in enumerate(examples):

        # if example_index % 100 == 0:
        #     logger.info('Converting %s/%s pos %s neg %s', example_index, len(examples), cnt_pos, cnt_neg)
@@ -240,7 +238,6 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,

        # The -3 accounts for [CLS], [SEP] and [SEP]
        max_tokens_for_doc = max_seq_length - len(query_tokens) - 3
-        assert max_tokens_for_doc > 0

        # We can have documents that are longer than the maximum sequence length.
        # To deal with this we do a sliding window approach, where we take chunks
@@ -275,19 +272,17 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
                p_mask.append(0)
                cls_index = 0

-            # XLNet: P SEP Q SEP CLS
-            # Others: CLS Q SEP P SEP
-            if not sequence_a_is_doc:
-                # Query
-                tokens += query_tokens
-                segment_ids += [sequence_a_segment_id] * len(query_tokens)
-                p_mask += [1] * len(query_tokens)
-
-                # SEP token
-                tokens.append(sep_token)
+            # Query
+            for token in query_tokens:
+                tokens.append(token)
                segment_ids.append(sequence_a_segment_id)
                p_mask.append(1)

+            # SEP token
+            tokens.append(sep_token)
+            segment_ids.append(sequence_a_segment_id)
+            p_mask.append(1)
+
            # Paragraph
            for i in range(doc_span.length):
                split_token_index = doc_span.start + i
@@ -297,23 +292,10 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
                                                       split_token_index)
                token_is_max_context[len(tokens)] = is_max_context
                tokens.append(all_doc_tokens[split_token_index])
-                if not sequence_a_is_doc:
-                    segment_ids.append(sequence_b_segment_id)
-                else:
-                    segment_ids.append(sequence_a_segment_id)
+                segment_ids.append(sequence_b_segment_id)
                p_mask.append(0)
            paragraph_len = doc_span.length

-            if sequence_a_is_doc:
-                # SEP token
-                tokens.append(sep_token)
-                segment_ids.append(sequence_a_segment_id)
-                p_mask.append(1)
-
-                tokens += query_tokens
-                segment_ids += [sequence_b_segment_id] * len(query_tokens)
-                p_mask += [1] * len(query_tokens)
-
            # SEP token
            tokens.append(sep_token)
            segment_ids.append(sequence_b_segment_id)
@@ -360,10 +342,7 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
                    end_position = 0
                    span_is_impossible = True
                else:
-                    if sequence_a_is_doc:
-                        doc_offset = 0
-                    else:
-                        doc_offset = len(query_tokens) + 2
+                    doc_offset = len(query_tokens) + 2
                    start_position = tok_start_position - doc_start + doc_offset
                    end_position = tok_end_position - doc_start + doc_offset

@@ -1,184 +0,0 @@
-from collections import deque
-import os
-
-import torch
-from torch.utils.data import Dataset
-
-
-# ------------
-# Data loading
-# ------------
-
-
-class CNNDailyMailDataset(Dataset):
-    """ Abstracts the dataset used to train seq2seq models.
-
-    CNN/Daily News:
-
-    The CNN/Daily News raw datasets are downloaded from [1]. The stories are
-    stored in different files; the summary appears at the end of the story as
-    sentences that are prefixed by the special `@highlight` line. To process
-    the data, untar both datasets in the same folder, and pass the path to this
-    folder as the "data_dir argument. The formatting code was inspired by [2].
-
-    [1] https://cs.nyu.edu/~kcho/
-    [2] https://github.com/abisee/cnn-dailymail/
-    """
-
-    def __init__(self, tokenizer, prefix="train", data_dir=""):
-        assert os.path.isdir(data_dir)
-        self.tokenizer = tokenizer
-
-        # We initialize the class by listing all the files that contain
-        # stories and summaries. Files are not read in memory given
-        # the size of the corpus.
-        self.stories_path = []
-        datasets = ("cnn", "dailymail")
-        for dataset in datasets:
-            path_to_stories = os.path.join(data_dir, dataset, "stories")
-            story_filenames_list = os.listdir(path_to_stories)
-            for story_filename in story_filenames_list:
-                path_to_story = os.path.join(path_to_stories, story_filename)
-                if not os.path.isfile(path_to_story):
-                    continue
-                self.stories_path.append(path_to_story)
-
-    def __len__(self):
-        return len(self.stories_path)
-
-    def __getitem__(self, idx):
-        story_path = self.stories_path[idx]
-        with open(story_path, encoding="utf-8") as source:
-            raw_story = source.read()
-            story_lines, summary_lines = process_story(raw_story)
-        return story_lines, summary_lines
-
-
-def process_story(raw_story):
-    """ Extract the story and summary from a story file.
-
-    Attributes:
-        raw_story (str): content of the story file as an utf-8 encoded string.
-
-    Raises:
-        IndexError: If the stoy is empty or contains no highlights.
-    """
-    nonempty_lines = list(
-        filter(lambda x: len(x) != 0, [line.strip() for line in raw_story.split("\n")])
-    )
-
-    # for some unknown reason some lines miss a period, add it
-    nonempty_lines = [_add_missing_period(line) for line in nonempty_lines]
-
-    # gather article lines
-    story_lines = []
-    lines = deque(nonempty_lines)
-    while True:
-        try:
-            element = lines.popleft()
-            if element.startswith("@highlight"):
-                break
-            story_lines.append(element)
-        except IndexError:
-            # if "@highlight" is absent from the file we pop
-            # all elements until there is None.
-            return story_lines, []
-
-    # gather summary lines
-    summary_lines = list(filter(lambda t: not t.startswith("@highlight"), lines))
-
-    return story_lines, summary_lines
-
-
-def _add_missing_period(line):
-    END_TOKENS = [".", "!", "?", "...", "'", "`", '"', u"\u2019", u"\u2019", ")"]
-    if line.startswith("@highlight"):
-        return line
-    if line[-1] in END_TOKENS:
-        return line
-    return line + "."
-
-
-# --------------------------
-# Encoding and preprocessing
-# --------------------------
-
-
-def fit_to_block_size(sequence, block_size, pad_token):
-    """ Adapt the source and target sequences' lengths to the block size.
-    If the sequence is shorter than the block size we pad it with -1 ids
-    which correspond to padding tokens.
-    """
-    if len(sequence) > block_size:
-        return sequence[:block_size]
-    else:
-        sequence.extend([pad_token] * (block_size - len(sequence)))
-        return sequence
-
-
-def build_lm_labels(sequence, pad_token):
-    """ Padding token, encoded as 0, are represented by the value -1 so they
-    are not taken into account in the loss computation. """
-    padded = sequence.clone()
-    padded[padded == pad_token] = -1
-    return padded
-
-
-def build_mask(sequence, pad_token):
-    """ Builds the mask. The attention mechanism will only attend to positions
-    with value 1. """
-    mask = torch.ones_like(sequence)
-    idx_pad_tokens = sequence == pad_token
-    mask[idx_pad_tokens] = 0
-    return mask
-
-
-def encode_for_summarization(story_lines, summary_lines, tokenizer):
-    """ Encode the story and summary lines, and join them
-    as specified in [1] by using `[SEP] [CLS]` tokens to separate
-    sentences.
-    """
-    story_lines_token_ids = [
-        tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line))
-        for line in story_lines
-    ]
-    summary_lines_token_ids = [
-        tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line))
-        for line in summary_lines
-    ]
-
-    story_token_ids = [
-        token for sentence in story_lines_token_ids for token in sentence
-    ]
-    summary_token_ids = [
-        token for sentence in summary_lines_token_ids for token in sentence
-    ]
-
-    return story_token_ids, summary_token_ids
-
-
-def compute_token_type_ids(batch, separator_token_id):
-    """ Segment embeddings as described in [1]
-
-    The values {0,1} were found in the repository [2].
-
-    Attributes:
-        batch: torch.Tensor, size [batch_size, block_size]
-            Batch of input.
-        separator_token_id: int
-            The value of the token that separates the segments.
-
-    [1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders."
-        arXiv preprint arXiv:1908.08345 (2019).
-    [2] https://github.com/nlpyang/PreSumm (/src/prepro/data_builder.py, commit fac1217)
-    """
-    batch_embeddings = []
-    for sequence in batch:
-        sentence_num = 0
-        embeddings = []
-        for s in sequence:
-            if s == separator_token_id:
-                sentence_num += 1
-            embeddings.append(sentence_num % 2)
-        batch_embeddings.append(embeddings)
-    return torch.tensor(batch_embeddings)
@@ -1,136 +0,0 @@
-# coding=utf-8
-# Copyright 2019 HuggingFace Inc.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-import unittest
-
-import numpy as np
-import torch
-
-from utils_summarization import (
-    compute_token_type_ids,
-    fit_to_block_size,
-    build_mask,
-    build_lm_labels,
-    process_story,
-)
-
-
-class SummarizationDataProcessingTest(unittest.TestCase):
-    def setUp(self):
-        self.block_size = 10
-
-    def test_fit_to_block_sequence_too_small(self):
-        """ Pad the sequence with 0 if the sequence is smaller than the block size."""
-        sequence = [1, 2, 3, 4]
-        expected_output = [1, 2, 3, 4, 0, 0, 0, 0, 0, 0]
-        self.assertEqual(
-            fit_to_block_size(sequence, self.block_size, 0), expected_output
-        )
-
-    def test_fit_to_block_sequence_fit_exactly(self):
-        """ Do nothing if the sequence is the right size. """
-        sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
-        expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
-        self.assertEqual(
-            fit_to_block_size(sequence, self.block_size, 0), expected_output
-        )
-
-    def test_fit_to_block_sequence_too_big(self):
-        """ Truncate the sequence if it is too long. """
-        sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
-        expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
-        self.assertEqual(
-            fit_to_block_size(sequence, self.block_size, 0), expected_output
-        )
-
-    def test_process_story_no_highlights(self):
-        """ Processing a story with no highlights returns an empty list for the summary.
-        """
-        raw_story = """It was the year of Our Lord one thousand seven hundred and
-        seventy-five.\n\nSpiritual revelations were conceded to England at that
-        favoured period, as at this."""
-        _, summary_lines = process_story(raw_story)
-        self.assertEqual(summary_lines, [])
-
-    def test_process_empty_story(self):
-        """ An empty story returns an empty collection of lines.
-        """
-        raw_story = ""
-        story_lines, summary_lines = process_story(raw_story)
-        self.assertEqual(story_lines, [])
-        self.assertEqual(summary_lines, [])
-
-    def test_process_story_with_missing_period(self):
-        raw_story = (
-            "It was the year of Our Lord one thousand seven hundred and "
-            "seventy-five\n\nSpiritual revelations were conceded to England "
-            "at that favoured period, as at this.\n@highlight\n\nIt was the best of times"
-        )
-        story_lines, summary_lines = process_story(raw_story)
-
-        expected_story_lines = [
-            "It was the year of Our Lord one thousand seven hundred and seventy-five.",
-            "Spiritual revelations were conceded to England at that favoured period, as at this.",
-        ]
-        self.assertEqual(expected_story_lines, story_lines)
-
-        expected_summary_lines = ["It was the best of times."]
-        self.assertEqual(expected_summary_lines, summary_lines)
-
-    def test_build_lm_labels_no_padding(self):
-        sequence = torch.tensor([1, 2, 3, 4])
-        expected = sequence
-        np.testing.assert_array_equal(
-            build_lm_labels(sequence, 0).numpy(), expected.numpy()
-        )
-
-    def test_build_lm_labels(self):
-        sequence = torch.tensor([1, 2, 3, 4, 0, 0, 0])
-        expected = torch.tensor([1, 2, 3, 4, -1, -1, -1])
-        np.testing.assert_array_equal(
-            build_lm_labels(sequence, 0).numpy(), expected.numpy()
-        )
-
-    def test_build_mask_no_padding(self):
-        sequence = torch.tensor([1, 2, 3, 4])
-        expected = torch.tensor([1, 1, 1, 1])
-        np.testing.assert_array_equal(build_mask(sequence, 0).numpy(), expected.numpy())
-
-    def test_build_mask(self):
-        sequence = torch.tensor([1, 2, 3, 4, 23, 23, 23])
-        expected = torch.tensor([1, 1, 1, 1, 0, 0, 0])
-        np.testing.assert_array_equal(
-            build_mask(sequence, 23).numpy(), expected.numpy()
-        )
-
-    def test_build_mask_with_padding_equal_to_one(self):
-        sequence = torch.tensor([8, 2, 3, 4, 1, 1, 1])
-        expected = torch.tensor([1, 1, 1, 1, 0, 0, 0])
-        np.testing.assert_array_equal(build_mask(sequence, 1).numpy(), expected.numpy())
-
-    def test_compute_token_type_ids(self):
-        separator = 101
-        batch = torch.tensor(
-            [[1, 2, 3, 4, 5, 6], [1, 2, 3, 101, 5, 6], [1, 101, 3, 4, 101, 6]]
-        )
-        expected = torch.tensor(
-            [[0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1], [0, 1, 1, 1, 0, 0]]
-        )
-
-        result = compute_token_type_ids(batch, separator)
-        np.testing.assert_array_equal(result, expected)
-
-
-if __name__ == "__main__":
-    unittest.main()
@@ -1,7 +1,7 @@
-from transformers import (
+from pytorch_transformers import (
    AutoTokenizer, AutoConfig, AutoModel, AutoModelWithLMHead, AutoModelForSequenceClassification, AutoModelForQuestionAnswering
 )
-from transformers.file_utils import add_start_docstrings
+from pytorch_transformers.file_utils import add_start_docstrings

 dependencies = ['torch', 'tqdm', 'boto3', 'requests', 'regex', 'sentencepiece', 'sacremoses']

@@ -11,12 +11,12 @@ def config(*args, **kwargs):
                # Using torch.hub !
                import torch

-                config = torch.hub.load('huggingface/transformers', 'config', 'bert-base-uncased')  # Download configuration from S3 and cache.
-                config = torch.hub.load('huggingface/transformers', 'config', './test/bert_saved_model/')  # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')`
-                config = torch.hub.load('huggingface/transformers', 'config', './test/bert_saved_model/my_configuration.json')
-                config = torch.hub.load('huggingface/transformers', 'config', 'bert-base-uncased', output_attention=True, foo=False)
+                config = torch.hub.load('huggingface/pytorch-transformers', 'config', 'bert-base-uncased')  # Download configuration from S3 and cache.
+                config = torch.hub.load('huggingface/pytorch-transformers', 'config', './test/bert_saved_model/')  # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')`
+                config = torch.hub.load('huggingface/pytorch-transformers', 'config', './test/bert_saved_model/my_configuration.json')
+                config = torch.hub.load('huggingface/pytorch-transformers', 'config', 'bert-base-uncased', output_attention=True, foo=False)
                assert config.output_attention == True
-                config, unused_kwargs = torch.hub.load('huggingface/transformers', 'config', 'bert-base-uncased', output_attention=True, foo=False, return_unused_kwargs=True)
+                config, unused_kwargs = torch.hub.load('huggingface/pytorch-transformers', 'config', 'bert-base-uncased', output_attention=True, foo=False, return_unused_kwargs=True)
                assert config.output_attention == True
                assert unused_kwargs == {'foo': False}

@@ -31,8 +31,8 @@ def tokenizer(*args, **kwargs):
        # Using torch.hub !
        import torch

-        tokenizer = torch.hub.load('huggingface/transformers', 'tokenizer', 'bert-base-uncased')    # Download vocabulary from S3 and cache.
-        tokenizer = torch.hub.load('huggingface/transformers', 'tokenizer', './test/bert_saved_model/')  # E.g. tokenizer was saved using `save_pretrained('./test/saved_model/')`
+        tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'tokenizer', 'bert-base-uncased')    # Download vocabulary from S3 and cache.
+        tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'tokenizer', './test/bert_saved_model/')  # E.g. tokenizer was saved using `save_pretrained('./test/saved_model/')`

    """

@@ -45,13 +45,13 @@ def model(*args, **kwargs):
            # Using torch.hub !
            import torch

-            model = torch.hub.load('huggingface/transformers', 'model', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
-            model = torch.hub.load('huggingface/transformers', 'model', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
-            model = torch.hub.load('huggingface/transformers', 'model', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
+            model = torch.hub.load('huggingface/pytorch-transformers', 'model', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = torch.hub.load('huggingface/pytorch-transformers', 'model', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = torch.hub.load('huggingface/pytorch-transformers', 'model', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
            assert model.config.output_attention == True
            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
            config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
-            model = torch.hub.load('huggingface/transformers', 'model', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+            model = torch.hub.load('huggingface/pytorch-transformers', 'model', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)

        """

@@ -63,13 +63,13 @@ def modelWithLMHead(*args, **kwargs):
        # Using torch.hub !
        import torch

-        model = torch.hub.load('huggingface/transformers', 'modelWithLMHead', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
-        model = torch.hub.load('huggingface/transformers', 'modelWithLMHead', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
-        model = torch.hub.load('huggingface/transformers', 'modelWithLMHead', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelWithLMHead', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelWithLMHead', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelWithLMHead', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
        assert model.config.output_attention == True
        # Loading from a TF checkpoint file instead of a PyTorch model (slower)
        config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
-        model = torch.hub.load('huggingface/transformers', 'modelWithLMHead', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelWithLMHead', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)

    """
    return AutoModelWithLMHead.from_pretrained(*args, **kwargs)
@@ -81,13 +81,13 @@ def modelForSequenceClassification(*args, **kwargs):
            # Using torch.hub !
            import torch

-            model = torch.hub.load('huggingface/transformers', 'modelForSequenceClassification', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
-            model = torch.hub.load('huggingface/transformers', 'modelForSequenceClassification', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
-            model = torch.hub.load('huggingface/transformers', 'modelForSequenceClassification', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
+            model = torch.hub.load('huggingface/pytorch-transformers', 'modelForSequenceClassification', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = torch.hub.load('huggingface/pytorch-transformers', 'modelForSequenceClassification', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = torch.hub.load('huggingface/pytorch-transformers', 'modelForSequenceClassification', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
            assert model.config.output_attention == True
            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
            config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
-            model = torch.hub.load('huggingface/transformers', 'modelForSequenceClassification', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+            model = torch.hub.load('huggingface/pytorch-transformers', 'modelForSequenceClassification', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)

        """

@@ -100,13 +100,13 @@ def modelForQuestionAnswering(*args, **kwargs):
        # Using torch.hub !
        import torch

-        model = torch.hub.load('huggingface/transformers', 'modelForQuestionAnswering', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
-        model = torch.hub.load('huggingface/transformers', 'modelForQuestionAnswering', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
-        model = torch.hub.load('huggingface/transformers', 'modelForQuestionAnswering', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelForQuestionAnswering', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelForQuestionAnswering', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelForQuestionAnswering', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
        assert model.config.output_attention == True
        # Loading from a TF checkpoint file instead of a PyTorch model (slower)
        config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
-        model = torch.hub.load('huggingface/transformers', 'modelForQuestionAnswering', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelForQuestionAnswering', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)

    """
    return AutoModelForQuestionAnswering.from_pretrained(*args, **kwargs)
@@ -1,4 +1,4 @@
-__version__ = "2.2.0"
+__version__ = "1.2.0"

 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -16,35 +16,17 @@ import logging

 logger = logging.getLogger(__name__)  # pylint: disable=invalid-name

-# Files and general utilities
-from .file_utils import (TRANSFORMERS_CACHE, PYTORCH_TRANSFORMERS_CACHE, PYTORCH_PRETRAINED_BERT_CACHE,
-                         cached_path, add_start_docstrings, add_end_docstrings,
-                         WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, CONFIG_NAME,
-                         is_tf_available, is_torch_available)
-
-from .data import (is_sklearn_available,
-                   InputExample, InputFeatures, DataProcessor,
-                   glue_output_modes, glue_convert_examples_to_features,
-                   glue_processors, glue_tasks_num_labels,
-                   xnli_output_modes, xnli_processors, xnli_tasks_num_labels)
-
-if is_sklearn_available():
-    from .data import glue_compute_metrics, xnli_compute_metrics
-
-# Tokenizers
+# Tokenizer
 from .tokenization_utils import (PreTrainedTokenizer)
 from .tokenization_auto import AutoTokenizer
 from .tokenization_bert import BertTokenizer, BasicTokenizer, WordpieceTokenizer
 from .tokenization_openai import OpenAIGPTTokenizer
 from .tokenization_transfo_xl import (TransfoXLTokenizer, TransfoXLCorpus)
 from .tokenization_gpt2 import GPT2Tokenizer
-from .tokenization_ctrl import CTRLTokenizer
 from .tokenization_xlnet import XLNetTokenizer, SPIECE_UNDERLINE
 from .tokenization_xlm import XLMTokenizer
 from .tokenization_roberta import RobertaTokenizer
 from .tokenization_distilbert import DistilBertTokenizer
-from .tokenization_albert import AlbertTokenizer
-from .tokenization_camembert import CamembertTokenizer

 # Configurations
 from .configuration_utils import PretrainedConfig
@@ -53,17 +35,21 @@ from .configuration_bert import BertConfig, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_openai import OpenAIGPTConfig, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_transfo_xl import TransfoXLConfig, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_gpt2 import GPT2Config, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP
-from .configuration_ctrl import CTRLConfig, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_xlnet import XLNetConfig, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP
-from .configuration_ctrl import CTRLConfig, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_xlm import XLMConfig, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_roberta import RobertaConfig, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_distilbert import DistilBertConfig, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
-from .configuration_albert import AlbertConfig, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
-from .configuration_camembert import CamembertConfig, CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP

 # Modeling
-if is_torch_available():
+try:
+    import torch
+    _torch_available = True  # pylint: disable=invalid-name
+except ImportError:
+    _torch_available = False  # pylint: disable=invalid-name
+
+if _torch_available:
+    logger.info("PyTorch version {} available.".format(torch.__version__))
+
    from .modeling_utils import (PreTrainedModel, prune_layer, Conv1D)
    from .modeling_auto import (AutoModel, AutoModelForSequenceClassification, AutoModelForQuestionAnswering,
                                AutoModelWithLMHead)
@@ -77,48 +63,41 @@ if is_torch_available():
                                OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
                                load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
    from .modeling_transfo_xl import (TransfoXLPreTrainedModel, TransfoXLModel, TransfoXLLMHeadModel,
-                                    AdaptiveEmbedding,
                                    load_tf_weights_in_transfo_xl, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
    from .modeling_gpt2 import (GPT2PreTrainedModel, GPT2Model,
                                GPT2LMHeadModel, GPT2DoubleHeadsModel,
                                load_tf_weights_in_gpt2, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP)
-    from .modeling_ctrl import (CTRLPreTrainedModel, CTRLModel,
-                                CTRLLMHeadModel,
-                                CTRL_PRETRAINED_MODEL_ARCHIVE_MAP)
    from .modeling_xlnet import (XLNetPreTrainedModel, XLNetModel, XLNetLMHeadModel,
-                                XLNetForSequenceClassification, XLNetForMultipleChoice,
-                                XLNetForQuestionAnsweringSimple, XLNetForQuestionAnswering,
+                                XLNetForSequenceClassification, XLNetForQuestionAnsweringSimple,
+                                XLNetForQuestionAnswering,
                                load_tf_weights_in_xlnet, XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)
    from .modeling_xlm import (XLMPreTrainedModel , XLMModel,
                            XLMWithLMHeadModel, XLMForSequenceClassification,
                            XLMForQuestionAnswering, XLMForQuestionAnsweringSimple,
                            XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
-    from .modeling_roberta import (RobertaForMaskedLM, RobertaModel,
-                                RobertaForSequenceClassification, RobertaForMultipleChoice,
-                                RobertaForTokenClassification,
+    from .modeling_roberta import (RobertaForMaskedLM, RobertaModel, RobertaForSequenceClassification,
                                ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
    from .modeling_distilbert import (DistilBertForMaskedLM, DistilBertModel,
                                DistilBertForSequenceClassification, DistilBertForQuestionAnswering,
-                                DistilBertForTokenClassification,
                                DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
-    from .modeling_camembert import (CamembertForMaskedLM, CamembertModel,
-                                CamembertForSequenceClassification, CamembertForMultipleChoice,
-                                CamembertForTokenClassification,
-                                CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
-    from .modeling_encoder_decoder import PreTrainedEncoderDecoder, Model2Model
-
-    from .modeling_albert import (AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification,
-                                AlbertForQuestionAnswering,
-                                load_tf_weights_in_albert, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)

    # Optimization
-    from .optimization import (AdamW, get_constant_schedule, get_constant_schedule_with_warmup, get_cosine_schedule_with_warmup,
-                               get_cosine_with_hard_restarts_schedule_with_warmup, get_linear_schedule_with_warmup)
+    from .optimization import (AdamW, ConstantLRSchedule, WarmupConstantSchedule, WarmupCosineSchedule,
+                               WarmupCosineWithHardRestartsSchedule, WarmupLinearSchedule)


 # TensorFlow
-if is_tf_available():
-    from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list
+try:
+    import tensorflow as tf
+    assert int(tf.__version__[0]) >= 2
+    _tf_available = True  # pylint: disable=invalid-name
+except (ImportError, AssertionError):
+    _tf_available = False  # pylint: disable=invalid-name
+
+if _tf_available:
+    logger.info("TensorFlow version {} available.".format(tf.__version__))
+
+    from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary
    from .modeling_tf_auto import (TFAutoModel, TFAutoModelForSequenceClassification, TFAutoModelForQuestionAnswering,
                                   TFAutoModelWithLMHead)

@@ -127,62 +106,70 @@ if is_tf_available():
                                   TFBertForMaskedLM, TFBertForNextSentencePrediction,
                                   TFBertForSequenceClassification, TFBertForMultipleChoice,
                                   TFBertForTokenClassification, TFBertForQuestionAnswering,
+                                   load_bert_pt_weights_in_tf2,
                                   TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP)

    from .modeling_tf_gpt2 import (TFGPT2PreTrainedModel, TFGPT2MainLayer,
                                   TFGPT2Model, TFGPT2LMHeadModel, TFGPT2DoubleHeadsModel,
+                                   load_gpt2_pt_weights_in_tf2,
                                   TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP)

    from .modeling_tf_openai import (TFOpenAIGPTPreTrainedModel, TFOpenAIGPTMainLayer,
                                     TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel, TFOpenAIGPTDoubleHeadsModel,
+                                     load_openai_gpt_pt_weights_in_tf2,
                                     TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)

    from .modeling_tf_transfo_xl import (TFTransfoXLPreTrainedModel, TFTransfoXLMainLayer,
                                         TFTransfoXLModel, TFTransfoXLLMHeadModel,
+                                         load_transfo_xl_pt_weights_in_tf2,
                                         TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)

    from .modeling_tf_xlnet import (TFXLNetPreTrainedModel, TFXLNetMainLayer,
                                    TFXLNetModel, TFXLNetLMHeadModel,
                                    TFXLNetForSequenceClassification,
                                    TFXLNetForQuestionAnsweringSimple,
+                                    load_xlnet_pt_weights_in_tf2,
                                    TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)

    from .modeling_tf_xlm import (TFXLMPreTrainedModel, TFXLMMainLayer,
                                  TFXLMModel, TFXLMWithLMHeadModel,
                                  TFXLMForSequenceClassification,
                                  TFXLMForQuestionAnsweringSimple,
+                                  load_xlm_pt_weights_in_tf2,
                                  TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP)

    from .modeling_tf_roberta import (TFRobertaPreTrainedModel, TFRobertaMainLayer,
                                      TFRobertaModel, TFRobertaForMaskedLM,
                                      TFRobertaForSequenceClassification,
-                                      TFRobertaForTokenClassification,
+                                      load_roberta_pt_weights_in_tf2,
                                      TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)

    from .modeling_tf_distilbert import (TFDistilBertPreTrainedModel, TFDistilBertMainLayer,
                                         TFDistilBertModel, TFDistilBertForMaskedLM,
                                         TFDistilBertForSequenceClassification,
                                         TFDistilBertForQuestionAnswering,
+                                         load_distilbert_pt_weights_in_tf2,
                                         TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)

-    from .modeling_tf_ctrl import (TFCTRLPreTrainedModel, TFCTRLModel,
-                                    TFCTRLLMHeadModel,
-                                    TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP)
+if _tf_available and _torch_available:
+    from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name,
+                                            load_pytorch_checkpoint_in_tf2_model,
+                                            load_pytorch_weights_in_tf2_model,
+                                            load_pytorch_model_in_tf2_model,
+                                            load_tf2_checkpoint_in_pytorch_model,
+                                            load_tf2_weights_in_pytorch_model,
+                                            load_tf2_model_in_pytorch_model)

-    from .modeling_tf_albert import (TFAlbertPreTrainedModel, TFAlbertModel, TFAlbertForMaskedLM,
-                                     TFAlbertForSequenceClassification,
-                                    TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
+# Files and general utilities
+from .file_utils import (PYTORCH_TRANSFORMERS_CACHE, PYTORCH_PRETRAINED_BERT_CACHE,
+                         cached_path, add_start_docstrings, add_end_docstrings,
+                         WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, CONFIG_NAME,
+                         is_tf_available, is_torch_available)

-# TF 2.0 <=> PyTorch conversion utilities
-from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name,
-                                        load_pytorch_checkpoint_in_tf2_model,
-                                        load_pytorch_weights_in_tf2_model,
-                                        load_pytorch_model_in_tf2_model,
-                                        load_tf2_checkpoint_in_pytorch_model,
-                                        load_tf2_weights_in_pytorch_model,
-                                        load_tf2_model_in_pytorch_model)
+from .data import (is_sklearn_available,
+                   InputExample, InputFeatures, DataProcessor,
+                   glue_output_modes, glue_convert_examples_to_features,
+                   glue_processors, glue_tasks_num_labels)

-if not is_tf_available() and not is_torch_available():
-    logger.warning("Neither PyTorch nor TensorFlow >= 2.0 have been found."
-                   "Models won't be available and only tokenizers, configuration"
-                   "and file/data utilities can be used.")
+if is_sklearn_available():
+    from .data import glue_compute_metrics
@@ -5,25 +5,25 @@ def main():
        print(
        "This command line utility let you convert original (author released) model checkpoint to pytorch.\n"
        "It should be used as one of: \n"
-        ">> transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT, \n"
-        ">> transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG], \n"
-        ">> transformers transfo_xl TF_CHECKPOINT_OR_DATASET PYTORCH_DUMP_OUTPUT [TF_CONFIG] or \n"
-        ">> transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [GPT2_CONFIG] or \n"
-        ">> transformers xlnet TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT [FINETUNING_TASK_NAME] or \n"
-        ">> transformers xlm XLM_CHECKPOINT_PATH PYTORCH_DUMP_OUTPUT")
+        ">> pytorch_transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT, \n"
+        ">> pytorch_transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG], \n"
+        ">> pytorch_transformers transfo_xl TF_CHECKPOINT_OR_DATASET PYTORCH_DUMP_OUTPUT [TF_CONFIG] or \n"
+        ">> pytorch_transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [GPT2_CONFIG] or \n"
+        ">> pytorch_transformers xlnet TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT [FINETUNING_TASK_NAME] or \n"
+        ">> pytorch_transformers xlm XLM_CHECKPOINT_PATH PYTORCH_DUMP_OUTPUT")
    else:
        if sys.argv[1] == "bert":
            try:
                from .convert_bert_original_tf_checkpoint_to_pytorch import convert_tf_checkpoint_to_pytorch
            except ImportError:
-                print("transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
                    "In that case, it requires TensorFlow to be installed. Please see "
                    "https://www.tensorflow.org/install/ for installation instructions.")
                raise

            if len(sys.argv) != 5:
                # pylint: disable=line-too-long
-                print("Should be used as `transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")
+                print("Should be used as `pytorch_transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")
            else:
                PYTORCH_DUMP_OUTPUT = sys.argv.pop()
                TF_CONFIG = sys.argv.pop()
@@ -33,7 +33,7 @@ def main():
            from .convert_openai_original_tf_checkpoint_to_pytorch import convert_openai_checkpoint_to_pytorch
            if len(sys.argv) < 4 or len(sys.argv) > 5:
                # pylint: disable=line-too-long
-                print("Should be used as `transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG]`")
+                print("Should be used as `pytorch_transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG]`")
            else:
                OPENAI_GPT_CHECKPOINT_FOLDER_PATH = sys.argv[2]
                PYTORCH_DUMP_OUTPUT = sys.argv[3]
@@ -48,13 +48,13 @@ def main():
            try:
                from .convert_transfo_xl_original_tf_checkpoint_to_pytorch import convert_transfo_xl_checkpoint_to_pytorch
            except ImportError:
-                print("transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
                    "In that case, it requires TensorFlow to be installed. Please see "
                    "https://www.tensorflow.org/install/ for installation instructions.")
                raise
            if len(sys.argv) < 4 or len(sys.argv) > 5:
                # pylint: disable=line-too-long
-                print("Should be used as `transformers transfo_xl TF_CHECKPOINT/TF_DATASET_FILE PYTORCH_DUMP_OUTPUT [TF_CONFIG]`")
+                print("Should be used as `pytorch_transformers transfo_xl TF_CHECKPOINT/TF_DATASET_FILE PYTORCH_DUMP_OUTPUT [TF_CONFIG]`")
            else:
                if 'ckpt' in sys.argv[2].lower():
                    TF_CHECKPOINT = sys.argv[2]
@@ -72,14 +72,14 @@ def main():
            try:
                from .convert_gpt2_original_tf_checkpoint_to_pytorch import convert_gpt2_checkpoint_to_pytorch
            except ImportError:
-                print("transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
                    "In that case, it requires TensorFlow to be installed. Please see "
                    "https://www.tensorflow.org/install/ for installation instructions.")
                raise

            if len(sys.argv) < 4 or len(sys.argv) > 5:
                # pylint: disable=line-too-long
-                print("Should be used as `transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [TF_CONFIG]`")
+                print("Should be used as `pytorch_transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [TF_CONFIG]`")
            else:
                TF_CHECKPOINT = sys.argv[2]
                PYTORCH_DUMP_OUTPUT = sys.argv[3]
@@ -92,14 +92,14 @@ def main():
            try:
                from .convert_xlnet_original_tf_checkpoint_to_pytorch import convert_xlnet_checkpoint_to_pytorch
            except ImportError:
-                print("transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
                    "In that case, it requires TensorFlow to be installed. Please see "
                    "https://www.tensorflow.org/install/ for installation instructions.")
                raise

            if len(sys.argv) < 5 or len(sys.argv) > 6:
                # pylint: disable=line-too-long
-                print("Should be used as `transformers xlnet TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT [FINETUNING_TASK_NAME]`")
+                print("Should be used as `pytorch_transformers xlnet TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT [FINETUNING_TASK_NAME]`")
            else:
                TF_CHECKPOINT = sys.argv[2]
                TF_CONFIG = sys.argv[3]
@@ -118,7 +118,7 @@ def main():

            if len(sys.argv) != 4:
                # pylint: disable=line-too-long
-                print("Should be used as `transformers xlm XLM_CHECKPOINT_PATH PYTORCH_DUMP_OUTPUT`")
+                print("Should be used as `pytorch_transformers xlm XLM_CHECKPOINT_PATH PYTORCH_DUMP_OUTPUT`")
            else:
                XLM_CHECKPOINT_PATH = sys.argv[2]
                PYTORCH_DUMP_OUTPUT = sys.argv[3]
@@ -26,14 +26,12 @@ from .configuration_xlnet import XLNetConfig
 from .configuration_xlm import XLMConfig
 from .configuration_roberta import RobertaConfig
 from .configuration_distilbert import DistilBertConfig
-from .configuration_ctrl import CTRLConfig
-from .configuration_camembert import CamembertConfig

 logger = logging.getLogger(__name__)


 class AutoConfig(object):
-    r""":class:`~transformers.AutoConfig` is a generic configuration class
+    r""":class:`~pytorch_transformers.AutoConfig` is a generic configuration class
        that will be instantiated as one of the configuration classes of the library
        when created with the `AutoConfig.from_pretrained(pretrained_model_name_or_path)`
        class method.
@@ -51,8 +49,7 @@ class AutoConfig(object):
            - contains `xlnet`: XLNetConfig (XLNet model)
            - contains `xlm`: XLMConfig (XLM model)
            - contains `roberta`: RobertaConfig (RoBERTa model)
-            - contains `camembert`: CamembertConfig (CamemBERT model)
-            - contains `ctrl` : CTRLConfig (CTRL model)
+
        This class cannot be instantiated using `__init__()` (throw an error).
    """
    def __init__(self):
@@ -74,13 +71,12 @@ class AutoConfig(object):
            - contains `xlnet`: XLNetConfig (XLNet model)
            - contains `xlm`: XLMConfig (XLM model)
            - contains `roberta`: RobertaConfig (RoBERTa model)
-            - contains `camembert`: CamembertConfig (CamemBERT model)
-            - contains `ctrl` : CTRLConfig (CTRL model)
+
        Params:
            pretrained_model_name_or_path: either:

                - a string with the `shortcut name` of a pre-trained model configuration to load from cache or download, e.g.: ``bert-base-uncased``.
-                - a path to a `directory` containing a configuration file saved using the :func:`~transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
+                - a path to a `directory` containing a configuration file saved using the :func:`~pytorch_transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
                - a path or url to a saved configuration JSON `file`, e.g.: ``./my_model_directory/configuration.json``.

            cache_dir: (`optional`) string:
@@ -95,9 +91,6 @@ class AutoConfig(object):
            force_download: (`optional`) boolean, default False:
                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.

-            resume_download: (`optional`) boolean, default False:
-                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
-
            proxies: (`optional`) dict, default None:
                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
                The proxies are used on each request.
@@ -122,8 +115,6 @@ class AutoConfig(object):
        """
        if 'distilbert' in pretrained_model_name_or_path:
            return DistilBertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
-        elif 'camembert' in pretrained_model_name_or_path:
-            return CamembertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
        elif 'roberta' in pretrained_model_name_or_path:
            return RobertaConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
        elif 'bert' in pretrained_model_name_or_path:
@@ -138,8 +129,7 @@ class AutoConfig(object):
            return XLNetConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
        elif 'xlm' in pretrained_model_name_or_path:
            return XLMConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
-        elif 'ctrl' in pretrained_model_name_or_path:
-            return CTRLConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
+
        raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                         "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
-                         "'xlm', 'roberta', 'camembert', 'ctrl'".format(pretrained_model_name_or_path))
+                         "'xlm', 'roberta'".format(pretrained_model_name_or_path))
@@ -40,14 +40,12 @@ BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-config.json",
    'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-config.json",
    'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-config.json",
-    'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-config.json",
-    'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-config.json",
 }


 class BertConfig(PretrainedConfig):
    r"""
-        :class:`~transformers.BertConfig` is the configuration class to store the configuration of a
+        :class:`~pytorch_transformers.BertConfig` is the configuration class to store the configuration of a
        `BertModel`.


@@ -27,8 +27,7 @@ logger = logging.getLogger(__name__)

 DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    'distilbert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-config.json",
-    'distilbert-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-distilled-squad-config.json",
-    'distilbert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-multilingual-cased-config.json",
+    'distilbert-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-distilled-squad-config.json"
 }


@@ -28,9 +28,7 @@ logger = logging.getLogger(__name__)

 GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-config.json",
                                      "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-config.json",
-                                      "gpt2-large": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-large-config.json",
-                                      "gpt2-xl": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-xl-config.json",
-                                      "distilgpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/distilgpt2-config.json",}
+                                      "gpt2-large": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-large-config.json"}

 class GPT2Config(PretrainedConfig):
    """Configuration class to store the configuration of a `GPT2Model`.
@@ -36,6 +36,7 @@ class OpenAIGPTConfig(PretrainedConfig):

    Args:
        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `OpenAIGPTModel` or a configuration json file.
+        n_special: The number of special tokens to learn during fine-tuning ('[SEP]', '[CLF]', ...)
        n_positions: Number of positional embeddings.
        n_ctx: Size of the causal mask (usually same as n_positions).
        n_embd: Dimensionality of the embeddings and hidden states.
@@ -28,9 +28,6 @@ ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    'roberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-base-config.json",
    'roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-config.json",
    'roberta-large-mnli': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-mnli-config.json",
-    'distilroberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/distilroberta-base-config.json",
-    'roberta-base-openai-detector': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-base-openai-detector-config.json",
-    'roberta-large-openai-detector': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-openai-detector-config.json",
 }


@@ -53,15 +53,13 @@ class PretrainedConfig(object):
        self.num_labels = kwargs.pop('num_labels', 2)
        self.output_attentions = kwargs.pop('output_attentions', False)
        self.output_hidden_states = kwargs.pop('output_hidden_states', False)
-        self.output_past = kwargs.pop('output_past', True)  # Not used by all models
-        self.torchscript = kwargs.pop('torchscript', False)  # Only used by PyTorch models
+        self.torchscript = kwargs.pop('torchscript', False)
        self.use_bfloat16 = kwargs.pop('use_bfloat16', False)
        self.pruned_heads = kwargs.pop('pruned_heads', {})
-        self.is_decoder = kwargs.pop('is_decoder', False)

    def save_pretrained(self, save_directory):
        """ Save a configuration object to the directory `save_directory`, so that it
-            can be re-loaded using the :func:`~transformers.PretrainedConfig.from_pretrained` class method.
+            can be re-loaded using the :func:`~pytorch_transformers.PretrainedConfig.from_pretrained` class method.
        """
        assert os.path.isdir(save_directory), "Saving path should be a directory where the model and configuration can be saved"

@@ -69,17 +67,16 @@ class PretrainedConfig(object):
        output_config_file = os.path.join(save_directory, CONFIG_NAME)

        self.to_json_file(output_config_file)
-        logger.info("Configuration saved in {}".format(output_config_file))

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
-        r""" Instantiate a :class:`~transformers.PretrainedConfig` (or a derived class) from a pre-trained model configuration.
+        r""" Instantiate a :class:`~pytorch_transformers.PretrainedConfig` (or a derived class) from a pre-trained model configuration.

        Parameters:
            pretrained_model_name_or_path: either:

                - a string with the `shortcut name` of a pre-trained model configuration to load from cache or download, e.g.: ``bert-base-uncased``.
-                - a path to a `directory` containing a configuration file saved using the :func:`~transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
+                - a path to a `directory` containing a configuration file saved using the :func:`~pytorch_transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
                - a path or url to a saved configuration JSON `file`, e.g.: ``./my_model_directory/configuration.json``.

            cache_dir: (`optional`) string:
@@ -94,9 +91,6 @@ class PretrainedConfig(object):
            force_download: (`optional`) boolean, default False:
                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.

-            resume_download: (`optional`) boolean, default False:
-                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
-
            proxies: (`optional`) dict, default None:
                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
                The proxies are used on each request.
@@ -123,7 +117,6 @@ class PretrainedConfig(object):
        """
        cache_dir = kwargs.pop('cache_dir', None)
        force_download = kwargs.pop('force_download', False)
-        resume_download = kwargs.pop('resume_download', False)
        proxies = kwargs.pop('proxies', None)
        return_unused_kwargs = kwargs.pop('return_unused_kwargs', False)

@@ -135,21 +128,21 @@ class PretrainedConfig(object):
            config_file = pretrained_model_name_or_path
        # redirect to the cache, if necessary
        try:
-            resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download,
-                                               proxies=proxies, resume_download=resume_download)
-        except EnvironmentError:
+            resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+        except EnvironmentError as e:
            if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
-                msg = "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
-                        config_file)
+                logger.error(
+                    "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
+                        config_file))
            else:
-                msg = "Model name '{}' was not found in model name list ({}). " \
-                      "We assumed '{}' was a path or url to a configuration file named {} or " \
-                      "a directory containing such a file but couldn't find any such file at this path or url.".format(
+                logger.error(
+                    "Model name '{}' was not found in model name list ({}). "
+                    "We assumed '{}' was a path or url but couldn't find any file "
+                    "associated to this path or url.".format(
                        pretrained_model_name_or_path,
                        ', '.join(cls.pretrained_config_archive_map.keys()),
-                        config_file, CONFIG_NAME)
-            raise EnvironmentError(msg)
-
+                        config_file))
+            raise e
        if resolved_config_file == config_file:
            logger.info("loading configuration file {}".format(config_file))
        else:
@@ -160,7 +153,7 @@ class PretrainedConfig(object):
        config = cls.from_json_file(resolved_config_file)

        if hasattr(config, 'pruned_heads'):
-            config.pruned_heads = dict((int(key), value) for key, value in config.pruned_heads.items())
+            config.pruned_heads = dict((int(key), set(value)) for key, value in config.pruned_heads.items())

        # Update config with kwargs if needed
        to_remove = []
@@ -171,7 +164,7 @@ class PretrainedConfig(object):
        for key in to_remove:
            kwargs.pop(key, None)

-        logger.info("Model config %s", str(config))
+        logger.info("Model config %s", config)
        if return_unused_kwargs:
            return config, kwargs
        else:
@@ -21,7 +21,7 @@ from __future__ import print_function
 import argparse
 import torch

-from transformers import BertConfig, BertForPreTraining, load_tf_weights_in_bert
+from pytorch_transformers import BertConfig, BertForPreTraining, load_tf_weights_in_bert

 import logging
 logging.basicConfig(level=logging.INFO)
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
LysandreJik	3bf90e9e6b	switch to Thom script	2019-09-25 09:38:37 -04:00
LysandreJik	3321480e5c	Learning rate and perdiction	2019-09-25 09:38:37 -04:00
LysandreJik	9c39853032	Beginning of GLUE script	2019-09-25 09:38:37 -04:00