splitting fast and slow tokenizers [WIP]

* First try

* Fix TF utils

* Handle authorized unexpected keys when loading weights

* Add several more authorized unexpected keys

* Apply style

* Fix test

* Address Patrick's comments.

* Update src/transformers/modeling_tf_utils.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/modeling_tf_utils.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Apply style

* Make return_dict the default behavior and display a warning message

* Revert

* Replace wrong keyword

* Revert code

* Add forgot key

* Fix bug in loading PT models from a TF one.

* Fix sort

* Add a test for custom load weights in BERT

* Apply style

* Remove unused import

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

.circleci/config.yml

@@ -248,6 +248,7 @@ jobs:
             - run: isort --check-only examples templates tests src utils
             - run: flake8 examples templates tests src utils
             - run: python utils/check_copies.py
+            - run: python utils/check_dummies.py
             - run: python utils/check_repo.py
     check_repository_consistency:
         working_directory: ~/transformers

.circleci/deploy.sh

@@ -49,4 +49,5 @@ deploy_doc "10d7239" v2.10.0
 deploy_doc "b42586e" v2.11.0
 deploy_doc "7fb8bdf" v3.0.2
 deploy_doc "4b3ee9c" v3.1.0
-deploy_doc "3ebb1b3" # v3.2.0 Latest stable release
+deploy_doc "3ebb1b3" v3.2.0
+deploy_doc "0613f05" # v3.3.0 Latest stable release

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,2 +1,61 @@
-<!-- This line specifies which issue to close after the pull request is merged. -->
-Fixes #{issue number}
+# What does this PR do?
+
+<!--
+Congratulations! You've made it this far! You're not quite done yet though.
+
+Once merged, your PR is going to appear in the release notes with the title you set, so make sure it's a great title that fully reflects the extent of your awesome contribution.
+
+Then, please replace this with a description of the change and which issue is fixed (if applicable). Please also include relevant motivation and context. List any dependencies (if any) that are required for this change.
+
+Once you're done, someone will review your PR shortly (see the section "Who can review?" below to tag some potential reviewers). They may suggest changes to make the code even better. If no one reviewed your PR after a week has passed, don't hesitate to post a new comment @-mentioning the same persons---sometimes notifications get lost.
+-->
+
+<!-- Remove if not applicable -->
+
+Fixes # (issue)
+
+
+## Before submitting
+- [ ] This PR fixes a typo or improves the docs (you can dimiss the other checks if that's the case).
+- [ ] Did you read the [contributor guideline](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md#start-contributing-pull-requests), 
+      Pull Request section?
+- [ ] Was this discussed/approved via a Github issue or the [forum](https://discuss.huggingface.co/)? Please add a link
+      to the it if that's the case.
+- [ ] Did you make sure to update the documentation with your changes? Here are the
+      [documentation guidelines](https://github.com/huggingface/transformers/tree/master/docs), and
+      [here are tips on formatting docstrings](https://github.com/huggingface/transformers/tree/master/docs#writing-source-documentation).
+- [ ] Did you write any new necessary tests? 
+
+
+## Who can review?
+
+Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
+members/contributors which may be interested in your PR.
+
+<!-- Your PR will be replied to more quickly if you can figure out the right person to tag with @
+
+ If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**.
+ Please tag fewer than 3 people.
+
+ albert, bert, GPT2, XLM: @LysandreJik 
+ tokenizers: @mfuntowicz
+ Trainer: @sgugger
+ Speed and Memory Benchmarks: @patrickvonplaten
+ Model Cards: @julien-c
+ Translation: @sshleifer
+ Summarization: @sshleifer
+ TextGeneration: @TevenLeScao 
+ examples/distillation: @VictorSanh
+ nlp datasets: [different repo](https://github.com/huggingface/nlp)
+ rust tokenizers: [different repo](https://github.com/huggingface/tokenizers)
+ Text Generation: @TevenLeScao
+ Blenderbot, Bart, Marian, Pegasus: @sshleifer
+ T5: @patrickvonplaten
+ Longformer/Reformer: @patrickvonplaten
+ TransfoXL/XLNet: @TevenLeScao 
+ examples/seq2seq: @sshleifer
+ examples/bert-loses-patience: @JetRunner
+ tensorflow: @jplu
+ examples/token-classification: @stefan-it
+ documentation: @sgugger
+ -->

.github/workflows/self-push.yml

@@ -14,7 +14,7 @@ on:
 
 jobs:
   run_tests_torch_and_tf_gpu:
-    runs-on: self-hosted
+    runs-on: [self-hosted, single-gpu]
     steps:
     - uses: actions/checkout@v2
     - name: Python version
@@ -51,7 +51,8 @@ jobs:
     - name: Are GPUs recognized by our DL frameworks
       run: |
         source .env/bin/activate
-        python -c "import torch; print(torch.cuda.is_available())"
+        python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
+        python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
 
     - name: Run all non-slow tests on GPU
       env:
@@ -62,3 +63,54 @@ jobs:
       run: |
         source .env/bin/activate
         python -m pytest -n 2 --dist=loadfile -s ./tests/
+
+  run_tests_torch_and_tf_multiple_gpu:
+    runs-on: [self-hosted, multi-gpu]
+    steps:
+      - uses: actions/checkout@v2
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v0-tests_tf_torch_multiple_gpu-${{ hashFiles('setup.py') }}
+
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install torch!=1.6.0
+          pip install .[sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
+
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
+          python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
+
+      - name: Run all non-slow tests on GPU
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          # TF_GPU_MEMORY_LIMIT: 4096
+          OMP_NUM_THREADS: 1
+          USE_CUDA: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 2 --dist=loadfile -s ./tests/

.github/workflows/self-scheduled.yml

@@ -10,7 +10,7 @@ on:
 
 jobs:
   run_all_tests_torch_and_tf_gpu:
-    runs-on: self-hosted
+    runs-on: [self-hosted, single-gpu]
     steps:
     - uses: actions/checkout@v2
 
@@ -48,7 +48,9 @@ jobs:
     - name: Are GPUs recognized by our DL frameworks
       run: |
         source .env/bin/activate
-        python -c "import torch; print(torch.cuda.is_available())"
+        python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
+        python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
+
 
     - name: Run all tests on GPU
       env:
@@ -70,3 +72,66 @@ jobs:
         source .env/bin/activate
         pip install -r examples/requirements.txt
         python -m pytest -n 1 --dist=loadfile -s examples
+
+  run_all_tests_torch_and_tf_multiple_gpu:
+    runs-on: [self-hosted, multi-gpu]
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v0-slow_tests_tf_torch_multi_gpu-${{ hashFiles('setup.py') }}
+
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        if: steps.cache.outputs.cache-hit != 'true'
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install torch!=1.6.0
+          pip install .[sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
+
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
+          python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
+
+      - name: Run all tests on GPU
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+          USE_CUDA: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --dist=loadfile -s ./tests/
+
+      - name: Run examples tests on GPU
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+          USE_CUDA: yes
+        run: |
+          source .env/bin/activate
+          pip install -r examples/requirements.txt
+          python -m pytest -n 1 --dist=loadfile -s examples

CODE_OF_CONDUCT.md

@@ -0,0 +1,129 @@
+
+# Contributor Covenant Code of Conduct
+
+## Our Pledge
+
+We as members, contributors, and leaders pledge to make participation in our
+community a harassment-free experience for everyone, regardless of age, body
+size, visible or invisible disability, ethnicity, sex characteristics, gender
+identity and expression, level of experience, education, socio-economic status,
+nationality, personal appearance, race, religion, or sexual identity
+and orientation.
+
+We pledge to act and interact in ways that contribute to an open, welcoming,
+diverse, inclusive, and healthy community.
+
+## Our Standards
+
+Examples of behavior that contributes to a positive environment for our
+community include:
+
+* Demonstrating empathy and kindness toward other people
+* Being respectful of differing opinions, viewpoints, and experiences
+* Giving and gracefully accepting constructive feedback
+* Accepting responsibility and apologizing to those affected by our mistakes,
+  and learning from the experience
+* Focusing on what is best not just for us as individuals, but for the
+  overall community
+
+Examples of unacceptable behavior include:
+
+* The use of sexualized language or imagery, and sexual attention or
+  advances of any kind
+* Trolling, insulting or derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or email
+  address, without their explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Enforcement Responsibilities
+
+Community leaders are responsible for clarifying and enforcing our standards of
+acceptable behavior and will take appropriate and fair corrective action in
+response to any behavior that they deem inappropriate, threatening, offensive,
+or harmful.
+
+Community leaders have the right and responsibility to remove, edit, or reject
+comments, commits, code, wiki edits, issues, and other contributions that are
+not aligned to this Code of Conduct, and will communicate reasons for moderation
+decisions when appropriate.
+
+## Scope
+
+This Code of Conduct applies within all community spaces, and also applies when
+an individual is officially representing the community in public spaces.
+Examples of representing our community include using an official e-mail address,
+posting via an official social media account, or acting as an appointed
+representative at an online or offline event.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported to the community leaders responsible for enforcement at
+feedback@huggingface.co.
+All complaints will be reviewed and investigated promptly and fairly.
+
+All community leaders are obligated to respect the privacy and security of the
+reporter of any incident.
+
+## Enforcement Guidelines
+
+Community leaders will follow these Community Impact Guidelines in determining
+the consequences for any action they deem in violation of this Code of Conduct:
+
+### 1. Correction
+
+**Community Impact**: Use of inappropriate language or other behavior deemed
+unprofessional or unwelcome in the community.
+
+**Consequence**: A private, written warning from community leaders, providing
+clarity around the nature of the violation and an explanation of why the
+behavior was inappropriate. A public apology may be requested.
+
+### 2. Warning
+
+**Community Impact**: A violation through a single incident or series
+of actions.
+
+**Consequence**: A warning with consequences for continued behavior. No
+interaction with the people involved, including unsolicited interaction with
+those enforcing the Code of Conduct, for a specified period of time. This
+includes avoiding interactions in community spaces as well as external channels
+like social media. Violating these terms may lead to a temporary or
+permanent ban.
+
+### 3. Temporary Ban
+
+**Community Impact**: A serious violation of community standards, including
+sustained inappropriate behavior.
+
+**Consequence**: A temporary ban from any sort of interaction or public
+communication with the community for a specified period of time. No public or
+private interaction with the people involved, including unsolicited interaction
+with those enforcing the Code of Conduct, is allowed during this period.
+Violating these terms may lead to a permanent ban.
+
+### 4. Permanent Ban
+
+**Community Impact**: Demonstrating a pattern of violation of community
+standards, including sustained inappropriate behavior,  harassment of an
+individual, or aggression toward or disparagement of classes of individuals.
+
+**Consequence**: A permanent ban from any sort of public interaction within
+the community.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage],
+version 2.0, available at
+https://www.contributor-covenant.org/version/2/0/code_of_conduct.html.
+
+Community Impact Guidelines were inspired by [Mozilla's code of conduct
+enforcement ladder](https://github.com/mozilla/diversity).
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see the FAQ at
+https://www.contributor-covenant.org/faq. Translations are available at
+https://www.contributor-covenant.org/translations.

CONTRIBUTING.md

@@ -9,6 +9,9 @@ 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".
 
+Whichever way you choose to contribute, please be mindful to respect our
+[code of conduct](https://github.com/huggingface/transformers/blob/master/CODE_OF_CONDUCT.md).
+
 ## You can contribute in so many ways!
 
 There are 4 ways you can contribute to transformers:
@@ -176,13 +179,14 @@ Follow these steps to start contributing:
    ```bash
    $ make quality
    ```
-
    You can do the automatic style corrections and code verifications that can't be automated in one go:
 
    ```bash
    $ make fixup
    ```
 
+   This target is also optimized to only work with files modified by the PR you're working on.
+
    If you're modifying documents under `docs/source`, make sure to validate that
    they can still be built. This check also runs in CI. To run a local check
    make sure you have installed the documentation builder requirements, by

Makefile

@@ -1,29 +1,53 @@
-.PHONY: quality_checks quality style fixup test test-examples docs
+.PHONY: modified_only_fixup extra_quality_checks quality style fixup fix-copies test test-examples docs
+
+
+check_dirs := examples templates tests src utils
+
+# get modified files since the branch was made
+fork_point_sha := $(shell git merge-base --fork-point master)
+joined_dirs    := $(shell echo $(check_dirs) | tr " " "|")
+modified_files := $(shell git diff --name-only $(fork_point_sha) | egrep '^($(joined_dirs))')
+#$(info modified files are: $(modified_files))
+
+modified_only_fixup:
+	@if [ -n "$(modified_files)" ]; then \
+		echo "Checking/fixing $(modified_files)"; \
+		black $(modified_files); \
+		isort $(modified_files); \
+		flake8 $(modified_files); \
+	else \
+		echo "No relevant files were modified"; \
+	fi
 
 # Check that source code meets quality standards
 
-quality_checks:
-	flake8 examples templates tests src utils
+extra_quality_checks:
 	python utils/check_copies.py
+	python utils/check_dummies.py
 	python utils/check_repo.py
 
+# this target runs checks on all files
 quality:
-	black --check examples templates tests src utils
-	isort --check-only examples templates tests src utils
-	${MAKE} quality_checks
+	black --check $(check_dirs)
+	isort --check-only $(check_dirs)
+	flake8 $(check_dirs)
+	${MAKE} extra_quality_checks
 
 # Format source code automatically and check is there are any problems left that need manual fixing
 
 style:
-	black examples templates tests src utils
-	isort examples templates tests src utils
+	black $(check_dirs)
+	isort $(check_dirs)
 
-fixup: style quality_checks
+# Super fast fix and check target that only works on relevant modified files since the branch was made
+
+fixup: modified_only_fixup extra_quality_checks
 
 # Make marked copies of snippets of codes conform to the original
 
 fix-copies:
 	python utils/check_copies.py --fix_and_overwrite
+	python utils/check_dummies.py --fix_and_overwrite
 
 # Run tests for the library
 

README.md

@@ -16,15 +16,18 @@
     <a href="https://github.com/huggingface/transformers/releases">
         <img alt="GitHub release" src="https://img.shields.io/github/release/huggingface/transformers.svg">
     </a>
+    <a href="https://github.com/huggingface/transformers/blob/master/CODE_OF_CONDUCT.md">
+        <img alt="Contributor Covenant" src="https://img.shields.io/badge/Contributor%20Covenant-v2.0%20adopted-ff69b4.svg">
+    </a>
 </p>
 
 <h3 align="center">
 <p>State-of-the-art Natural Language Processing for PyTorch and TensorFlow 2.0
 </h3>
 
-🤗 Transformers provides thousands of pretrained models to perform tasks on texts such as classification, information extraction, question answering, summarization, translation, text generation, etc in 100+ languages. Its aim is to make cutting-edge NLP easier to use for everyone. 
+🤗 Transformers provides thousands of pretrained models to perform tasks on texts such as classification, information extraction, question answering, summarization, translation, text generation, etc in 100+ languages. Its aim is to make cutting-edge NLP easier to use for everyone.
 
-🤗 Transformers provides APIs to quickly download and use those pretrained models on a given text, fine-tune them on your own datasets then share them with the community on our [model hub](https://huggingface.co/models). At the same time, each python module defining an architecture can be used as a standalone and modified to enable quick research experiments. 
+🤗 Transformers provides APIs to quickly download and use those pretrained models on a given text, fine-tune them on your own datasets then share them with the community on our [model hub](https://huggingface.co/models). At the same time, each python module defining an architecture can be used as a standalone and modified to enable quick research experiments.
 
 🤗 Transformers is backed by the two most popular deep learning libraries, [PyTorch](https://pytorch.org/) and [TensorFlow](https://www.tensorflow.org/), with a seamless integration between them, allowing you to train your models with one then load it for inference with the other.
 
@@ -35,7 +38,7 @@
 
 You can test most of our models directly on their pages from the [model hub](https://huggingface.co/models). We also offer an [inference API](https://huggingface.co/pricing) to use those models.
 
-Here are a few examples: 
+Here are a few examples:
 - [Masked word completion with BERT](https://huggingface.co/bert-base-uncased?text=Paris+is+the+%5BMASK%5D+of+France)
 - [Name Entity Recognition with Electra](https://huggingface.co/dbmdz/electra-large-discriminator-finetuned-conll03-english?text=My+name+is+Sarah+and+I+live+in+London+city)
 - [Text generation with GPT-2](https://huggingface.co/gpt2?text=A+long+time+ago%2C+)
@@ -48,7 +51,7 @@ Here are a few examples:
 
 ## Quick tour
 
-To immediately use a model on a given text, we provide the `pipeline` API. Pipelines group together a pretrained model with the preprocessing that was used during that model training. Here is how to quickly use a pipeline to classify positive versus negative texts 
+To immediately use a model on a given text, we provide the `pipeline` API. Pipelines group together a pretrained model with the preprocessing that was used during that model training. Here is how to quickly use a pipeline to classify positive versus negative texts
 
 ```python
 >>> from transformers import pipeline
@@ -59,7 +62,7 @@ To immediately use a model on a given text, we provide the `pipeline` API. Pipel
 [{'label': 'POSITIVE', 'score': 0.9978193640708923}]
 ```
 
-The second line of code downloads and caches the pretrained model used by the pipeline, the third line evaluates it on the given text. Here the answer is "positive" with a confidence of 99.8%. 
+The second line of code downloads and caches the pretrained model used by the pipeline, the third line evaluates it on the given text. Here the answer is "positive" with a confidence of 99.8%.
 
 This is another example of pipeline used for that can extract question answers from some context:
 
@@ -108,7 +111,7 @@ The model itself is a regular [Pytorch `nn.Module`](https://pytorch.org/docs/sta
 1. Easy-to-use state-of-the-art models:
     - High performance on NLU and NLG tasks.
     - Low barrier to entry for educators and practitioners.
-    - Few user-facing abastractions with just three classes to learn.
+    - Few user-facing abstractions with just three classes to learn.
     - A unified API for using all our pretrained models.
 
 1. Lower compute costs, smaller carbon footprint:
@@ -124,7 +127,7 @@ The model itself is a regular [Pytorch `nn.Module`](https://pytorch.org/docs/sta
 1. Easily customize a model or an example to your needs:
     - Examples for each architecture to reproduce the results by the official authors of said architecture.
     - Expose the models internal as consistently as possible.
-    - Model files can be used independently of the library for quick experiments. 
+    - Model files can be used independently of the library for quick experiments.
 
 ## Why shouldn't I use transformers?
 
@@ -155,37 +158,41 @@ If you'd like to play with the examples, you must [install the library from sour
 
 🤗 Transformers currently provides the following architectures (see [here](https://huggingface.co/transformers/model_summary.html) for a high-level summary of each them):
 
+1. **[ALBERT](https://huggingface.co/transformers/model_doc/albert.html)** (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.
+1. **[BART](https://huggingface.co/transformers/model_doc/bart.html)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
 1. **[BERT](https://huggingface.co/transformers/model_doc/bert.html)** (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://huggingface.co/transformers/model_doc/gpt.html)** (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.
-3. **[GPT-2](https://huggingface.co/transformers/model_doc/gpt2.html)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
-4. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
-5. **[XLNet](https://huggingface.co/transformers/model_doc/xlnet.html)** (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://huggingface.co/transformers/model_doc/xlm.html)** (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://huggingface.co/transformers/model_doc/roberta.html)** (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), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
-9. **[CTRL](https://huggingface.co/transformers/model_doc/ctrl.html)** (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://huggingface.co/transformers/model_doc/camembert.html)** (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://huggingface.co/transformers/model_doc/albert.html)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
-12. **[T5](https://huggingface.co/transformers/model_doc/t5.html)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
-13. **[XLM-RoBERTa](https://huggingface.co/transformers/model_doc/xlmroberta.html)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
-14. **[MMBT](https://github.com/facebookresearch/mmbt/)** (from Facebook), released together with the paper a [Supervised Multimodal Bitransformers for Classifying Images and Text](https://arxiv.org/pdf/1909.02950.pdf) by Douwe Kiela, Suvrat Bhooshan, Hamed Firooz, Davide Testuggine.
-15. **[FlauBERT](https://huggingface.co/transformers/model_doc/flaubert.html)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
-16. **[BART](https://huggingface.co/transformers/model_doc/bart.html)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
-17. **[ELECTRA](https://huggingface.co/transformers/model_doc/electra.html)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
-18. **[DialoGPT](https://huggingface.co/transformers/model_doc/dialogpt.html)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
-19. **[Reformer](https://huggingface.co/transformers/model_doc/reformer.html)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
-20. **[MarianMT](https://huggingface.co/transformers/model_doc/marian.html)** Machine translation models trained using [OPUS](http://opus.nlpl.eu/) data by Jörg Tiedemann. The [Marian Framework](https://marian-nmt.github.io/) is being developed by the Microsoft Translator Team.
-21. **[Longformer](https://huggingface.co/transformers/model_doc/longformer.html)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
-22. **[DPR](https://github.com/facebookresearch/DPR)** (from Facebook) released with the paper [Dense Passage Retrieval
+1. **[BERT For Sequence Generation](https://tfhub.dev/s?module-type=text-generation&subtype=module,placeholder)** (from Google) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
+1. **[CamemBERT](https://huggingface.co/transformers/model_doc/camembert.html)** (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.
+1. **[CTRL](https://huggingface.co/transformers/model_doc/ctrl.html)** (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.
+1. **[DeBERTa](https://huggingface.co/transformers/model_doc/deberta.html)** (from Microsoft Research) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
+1. **[DialoGPT](https://huggingface.co/transformers/model_doc/dialogpt.html)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
+1. **[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), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
+1. **[DPR](https://github.com/facebookresearch/DPR)** (from Facebook) released with the paper [Dense Passage Retrieval
 for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimir Karpukhin, Barlas Oğuz, Sewon
 Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
-23. **[Pegasus](https://github.com/google-research/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777)> by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
-24. **[MBart](https://github.com/pytorch/fairseq/tree/master/examples/mbart)** (from Facebook) released with the paper  [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.  
-25. **[LXMERT](https://github.com/airsplay/lxmert)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) by Hao Tan and Mohit Bansal.
-26. **[Funnel Transformer](https://github.com/laiguokun/Funnel-Transformer)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
-27. **[LayoutLM](https://github.com/microsoft/unilm/tree/master/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
-28. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
-29. 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.
+1. **[ELECTRA](https://huggingface.co/transformers/model_doc/electra.html)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+1. **[FlauBERT](https://huggingface.co/transformers/model_doc/flaubert.html)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
+1. **[Funnel Transformer](https://github.com/laiguokun/Funnel-Transformer)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
+1. **[GPT](https://huggingface.co/transformers/model_doc/gpt.html)** (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.
+1. **[GPT-2](https://huggingface.co/transformers/model_doc/gpt2.html)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
+1. **[LayoutLM](https://github.com/microsoft/unilm/tree/master/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
+1. **[Longformer](https://huggingface.co/transformers/model_doc/longformer.html)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
+1. **[LXMERT](https://github.com/airsplay/lxmert)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) by Hao Tan and Mohit Bansal.
+1. **[MarianMT](https://huggingface.co/transformers/model_doc/marian.html)** Machine translation models trained using [OPUS](http://opus.nlpl.eu/) data by Jörg Tiedemann. The [Marian Framework](https://marian-nmt.github.io/) is being developed by the Microsoft Translator Team.
+1. **[MBart](https://github.com/pytorch/fairseq/tree/master/examples/mbart)** (from Facebook) released with the paper  [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
+1. **[MMBT](https://github.com/facebookresearch/mmbt/)** (from Facebook), released together with the paper a [Supervised Multimodal Bitransformers for Classifying Images and Text](https://arxiv.org/pdf/1909.02950.pdf) by Douwe Kiela, Suvrat Bhooshan, Hamed Firooz, Davide Testuggine.
+1. **[Pegasus](https://github.com/google-research/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777)> by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
+1. **[Reformer](https://huggingface.co/transformers/model_doc/reformer.html)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
+1. **[RoBERTa](https://huggingface.co/transformers/model_doc/roberta.html)** (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.
+ultilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
+1. **[SqueezeBert](https://huggingface.co/transformers/model_doc/squeezebert.html)** released with the paper [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://arxiv.org/abs/2006.11316) by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
+1. **[T5](https://huggingface.co/transformers/model_doc/t5.html)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
+1. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
+1. **[XLM](https://huggingface.co/transformers/model_doc/xlm.html)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
+1. **[XLM-RoBERTa](https://huggingface.co/transformers/model_doc/xlmroberta.html)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
+1. **[XLNet](https://huggingface.co/transformers/model_doc/xlnet.html)** (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.
+1. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
+1. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
 
 These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations. You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
 

codecov.yml

@@ -4,7 +4,4 @@ coverage:
       default:
         informational: true
     patch: off
-comment:
-  require_changes: true    # only comment if there was change in coverage
-  require_head: yes        # don't report if there is no head coverage report
-  require_base: yes        # don't report if there is no base coverage report
+comment: false

docs/source/_static/css/huggingface.css

@@ -125,6 +125,12 @@ a.copybtn {
     background-color: #6670FF;
 }
 
+/* The section headers in the toc tree */
+.wy-menu-vertical p.caption{
+    background-color: #4d59ff;
+    line-height: 40px;
+}
+
 /* The selected items in the toc tree */
 .wy-menu-vertical li.current{
     background-color: #A6B0FF;

docs/source/_static/js/custom.js

@@ -1,10 +1,11 @@
 // These two things need to be updated at each release for the version selector.
 // Last stable version
-const stableVersion = "v3.2.0"
+const stableVersion = "v3.3.0"
 // Dictionary doc folder to label
 const versionMapping = {
     "master": "master",
-    "": "v3.2.0",
+    "": "v3.3.0/v3.3.1",
+    "v3.2.0": "v3.2.0",
     "v3.1.0": "v3.1.0 (stable)",
     "v3.0.2": "v3.0.0/v3.0.1/v3.0.2",
     "v2.11.0": "v2.11.0",
@@ -235,9 +236,11 @@ function platformToggle() {
 
     const createFrameworkButtons = sample => {
             const pytorchButton = document.createElement("button");
+            pytorchButton.classList.add('pytorch-button')
             pytorchButton.innerText = "PyTorch";
 
             const tensorflowButton = document.createElement("button");
+            tensorflowButton.classList.add('tensorflow-button')
             tensorflowButton.innerText = "TensorFlow";
 
             const selectorDiv = document.createElement("div");
@@ -252,22 +255,36 @@ function platformToggle() {
             tensorflowButton.classList.remove("selected");
 
             pytorchButton.addEventListener("click", () => {
-                sample.element.innerHTML = sample.pytorchSample;
-                pytorchButton.classList.add("selected");
-                tensorflowButton.classList.remove("selected");
+                for(const codeBlock of updatedCodeBlocks){
+                    codeBlock.element.innerHTML = codeBlock.pytorchSample;
+                }
+                Array.from(document.getElementsByClassName('pytorch-button')).forEach(button => {
+                    button.classList.add("selected");
+                })
+                Array.from(document.getElementsByClassName('tensorflow-button')).forEach(button => {
+                    button.classList.remove("selected");
+                })
             });
             tensorflowButton.addEventListener("click", () => {
-               sample.element.innerHTML = sample.tensorflowSample;
-                tensorflowButton.classList.add("selected");
-                pytorchButton.classList.remove("selected");
+                for(const codeBlock of updatedCodeBlocks){
+                    codeBlock.element.innerHTML = codeBlock.tensorflowSample;
+                }
+                Array.from(document.getElementsByClassName('tensorflow-button')).forEach(button => {
+                    button.classList.add("selected");
+                })
+                Array.from(document.getElementsByClassName('pytorch-button')).forEach(button => {
+                    button.classList.remove("selected");
+                })
             });
         };
 
-    codeBlocks
+    const updatedCodeBlocks = codeBlocks
         .map(element => {return {element: element.firstChild, innerText: element.innerText}})
         .filter(codeBlock => codeBlock.innerText.includes(pytorchIdentifier) && codeBlock.innerText.includes(tensorflowIdentifier))
         .map(getFrameworkSpans)
-        .forEach(createFrameworkButtons);
+
+    updatedCodeBlocks
+        .forEach(createFrameworkButtons)
 }
 
 

docs/source/conf.py

@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'3.2.0'
+release = u'3.3.1'
 
 
 # -- General configuration ---------------------------------------------------

docs/source/index.rst

@@ -46,102 +46,125 @@ The documentation is organized in five parts:
 - **ADVANCED GUIDES** contains more advanced guides that are more specific to a given script or part of the library.
 - **RESEARCH** focuses on tutorials that have less to do with how to use the library but more about general resarch in
   transformers model
-- **PACKAGE REFERENCE** contains the documentation of each public class and function.
+- The three last section contain the documentation of each public class and function, grouped in:
+    - **MAIN CLASSES** for the main classes exposing the important APIs of the library.
+    - **MODELS** for the classes and functions related to each model implemented in the library.
+    - **INTERNAL HELPERS** for the classes and functions we use internally.
 
 The library currently contains PyTorch and Tensorflow implementations, pre-trained model weights, usage scripts and
 conversion utilities for the following models:
 
-1. `BERT <https://github.com/google-research/bert>`_ (from Google) released with the paper `BERT: Pre-training of Deep
-   Bidirectional Transformers for Language Understanding <https://arxiv.org/abs/1810.04805>`_ by Jacob Devlin, Ming-Wei
-   Chang, Kenton Lee, and Kristina Toutanova.
-2. `GPT <https://github.com/openai/finetune-transformer-lm>`_ (from OpenAI) released with the paper `Improving Language
-   Understanding by Generative Pre-Training <https://blog.openai.com/language-unsupervised>`_ by Alec Radford, Karthik
-   Narasimhan, Tim Salimans, and Ilya Sutskever.
-3. `GPT-2 <https://blog.openai.com/better-language-models>`_ (from OpenAI) released with the paper `Language Models are
-   Unsupervised Multitask Learners <https://blog.openai.com/better-language-models>`_ by Alec Radford, Jeffrey Wu,
-   Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever.
-4. `Transformer-XL <https://github.com/kimiyoung/transformer-xl>`_ (from Google/CMU) released with the paper
-   `Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context <https://arxiv.org/abs/1901.02860>`_ by
-   Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V. Le, and Ruslan Salakhutdinov.
-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, and 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, and Veselin
-   Stoyanov.
-8. `DistilBERT <https://huggingface.co/transformers/model_doc/distilbert.html>`_ (from HuggingFace) released together
+..
+    This list is updated automatically from the README with `make fix-copies`. Do not update manually!
+
+1. `ALBERT <https://huggingface.co/transformers/model_doc/albert.html>`__ (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.
+2. `BART <https://huggingface.co/transformers/model_doc/bart.html>`__ (from Facebook) released with the paper `BART:
+   Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
+   <https://arxiv.org/pdf/1910.13461.pdf>`__ by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman
+   Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
+3. `BERT <https://huggingface.co/transformers/model_doc/bert.html>`__ (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.
+4. `BERT For Sequence Generation <https://tfhub.dev/s?module-type=text-generation&subtype=module,placeholder>`__ (from
+   Google) released with the paper `Leveraging Pre-trained Checkpoints for Sequence Generation Tasks
+   <https://arxiv.org/abs/1907.12461>`__ by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
+5. `CamemBERT <https://huggingface.co/transformers/model_doc/camembert.html>`__ (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.
+6. `CTRL <https://huggingface.co/transformers/model_doc/ctrl.html>`__ (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.
+7. `DeBERTa <https://huggingface.co/transformers/model_doc/deberta.html>`__ (from Microsoft Research) released with the
+   paper `DeBERTa: Decoding-enhanced BERT with Disentangled Attention <https://arxiv.org/abs/2006.03654>`__ by
+   Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
+8. `DialoGPT <https://huggingface.co/transformers/model_doc/dialogpt.html>`__ (from Microsoft Research) released with
+   the paper `DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation
+   <https://arxiv.org/abs/1911.00536>`__ by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang
+   Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
+9. `DistilBERT <https://huggingface.co/transformers/model_doc/distilbert.html>`__ (from HuggingFace), released together
    with the paper `DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter
-   <https://arxiv.org/abs/1910.01108>`_ by Victor Sanh, Lysandre Debut, and Thomas Wolf. The same method has been
-   applied to compress GPT2 into
-   `DistilGPT2 <https://github.com/huggingface/transformers/tree/master/examples/distillation>`_.
-9. `CTRL <https://github.com/pytorch/fairseq/tree/master/examples/ctrl>`_ (from Salesforce), released together with the
-   paper `CTRL: A Conditional Transformer Language Model for Controllable Generation
-   <https://www.github.com/salesforce/ctrl>`_ by Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney, Caiming Xiong,
-   and Richard Socher.
-10. `CamemBERT <https://huggingface.co/transformers/model_doc/camembert.html>`_ (from FAIR, Inria, Sorbonne Université)
-    released together with the paper `CamemBERT: a Tasty French Language Model <https://arxiv.org/abs/1911.03894>`_ by
-    Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suarez, Yoann Dupont, Laurent Romary, Eric Villemonte de la
-    Clergerie, Djame Seddah, and Benoît Sagot.
-11. `ALBERT <https://github.com/google-research/ALBERT>`_ (from Google Research), released together with the paper
-    `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, and Radu Soricut.
-12. `T5 <https://github.com/google-research/text-to-text-transfer-transformer>`_ (from Google) released with the paper
-    `Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
-    <https://arxiv.org/abs/1910.10683>`_ by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang,
-    Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu.
-13. `XLM-RoBERTa <https://github.com/pytorch/fairseq/tree/master/examples/xlmr>`_ (from Facebook AI), released together
-    with the paper `Unsupervised Cross-lingual Representation Learning at Scale <https://arxiv.org/abs/1911.02116>`_ by
-    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard
-    Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov.
-14. `MMBT <https://github.com/facebookresearch/mmbt/>`_ (from Facebook), released together with the paper a `Supervised
-    Multimodal Bitransformers for Classifying Images and Text <https://arxiv.org/pdf/1909.02950.pdf>`_ by Douwe Kiela,
-    Suvrat Bhooshan, Hamed Firooz, and Davide Testuggine.
-15. `FlauBERT <https://github.com/getalp/Flaubert>`_ (from CNRS) released with the paper `FlauBERT: Unsupervised
-    Language Model Pre-training for French <https://arxiv.org/abs/1912.05372>`_ by Hang Le, Loïc Vial, Jibril Frej,
-    Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, and
-    Didier Schwab.
-16. `BART <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_ (from Facebook) released with the paper
-    `BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
-    <https://arxiv.org/pdf/1910.13461.pdf>`_ by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman
-    Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer.
-17. `ELECTRA <https://github.com/google-research/electra>`_ (from Google Research/Stanford University) released with
-    the paper `ELECTRA: Pre-training text encoders as discriminators rather than generators
-    <https://arxiv.org/abs/2003.10555>`_ by Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning.
-18. `DialoGPT <https://github.com/microsoft/DialoGPT>`_ (from Microsoft Research) released with the paper `DialoGPT:
-    Large-Scale Generative Pre-training for Conversational Response Generation <https://arxiv.org/abs/1911.00536>`_ by
-    Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu,
-    and Bill Dolan.
-19. `Reformer <https://github.com/google/trax/tree/master/trax/models/reformer>`_ (from Google Research) released with
-    the paper `Reformer: The Efficient Transformer <https://arxiv.org/abs/2001.04451>`_ by Nikita Kitaev, Łukasz
-    Kaiser, and Anselm Levskaya.
-20. `MarianMT <https://marian-nmt.github.io/>`_ (developed by the Microsoft Translator Team) machine translation models
-    trained using `OPUS <http://opus.nlpl.eu/>`_ pretrained_models data by Jörg Tiedemann.
-21. `Longformer <https://github.com/allenai/longformer>`_ (from AllenAI) released with the paper `Longformer: The
-    Long-Document Transformer <https://arxiv.org/abs/2004.05150>`_ by Iz Beltagy, Matthew E. Peters, and Arman Cohan.
-22. `DPR <https://github.com/facebookresearch/DPR>`_ (from Facebook) released with the paper `Dense Passage Retrieval
-    for Open-Domain Question Answering <https://arxiv.org/abs/2004.04906>`_ by Vladimir Karpukhin, Barlas Oğuz, Sewon
+   <https://arxiv.org/abs/1910.01108>`__ by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been
+   applied to compress GPT2 into `DistilGPT2
+   <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__, RoBERTa into `DistilRoBERTa
+   <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__, Multilingual BERT into
+   `DistilmBERT <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__ and a German version
+   of DistilBERT.
+10. `DPR <https://github.com/facebookresearch/DPR>`__ (from Facebook) released with the paper `Dense Passage Retrieval
+    for Open-Domain Question Answering <https://arxiv.org/abs/2004.04906>`__ by Vladimir Karpukhin, Barlas Oğuz, Sewon
     Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
-23. `Pegasus <https://github.com/google-research/pegasus>`_ (from Google) released with the paper `PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization
-    <https://arxiv.org/abs/1912.08777>`_ by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
-24. `MBart <https://github.com/pytorch/fairseq/tree/master/examples/mbart>`_ (from Facebook) released with the paper  `Multilingual Denoising Pre-training for Neural Machine Translation <https://arxiv.org/abs/2001.08210>`_ by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov,
-    Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
-25. `LXMERT <https://github.com/airsplay/lxmert>`_ (from UNC Chapel Hill) released with the paper `LXMERT: Learning
-    Cross-Modality Encoder Representations from Transformers for Open-Domain Question
-    Answering <https://arxiv.org/abs/1908.07490>`_ by Hao Tan and Mohit Bansal.
-26. `Funnel Transformer <https://github.com/laiguokun/Funnel-Transformer>`_ (from CMU/Google Brain) released with the paper
-    `Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing
-    <https://arxiv.org/abs/2006.03236>`_ by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
-27. `Bert For Sequence Generation <https://tfhub.dev/s?module-type=text-generation&subtype=module,placeholder>`_ (from Google) released with the paper
-    `Leveraging Pre-trained Checkpoints for Sequence Generation Tasks
-    <https://arxiv.org/abs/1907.12461>`_ by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
-28. `LayoutLM <https://github.com/microsoft/unilm/tree/master/layoutlm>`_ (from Microsoft Research Asia) released with the paper
-    `LayoutLM: Pre-training of Text and Layout for Document Image Understanding
-    <https://arxiv.org/abs/1912.13318>`_ by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
-29. `Other community models <https://huggingface.co/models>`_, contributed by the `community
-    <https://huggingface.co/users>`_.
+11. `ELECTRA <https://huggingface.co/transformers/model_doc/electra.html>`__ (from Google Research/Stanford University)
+    released with the paper `ELECTRA: Pre-training text encoders as discriminators rather than generators
+    <https://arxiv.org/abs/2003.10555>`__ by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+12. `FlauBERT <https://huggingface.co/transformers/model_doc/flaubert.html>`__ (from CNRS) released with the paper
+    `FlauBERT: Unsupervised Language Model Pre-training for French <https://arxiv.org/abs/1912.05372>`__ by Hang Le,
+    Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé,
+    Laurent Besacier, Didier Schwab.
+13. `Funnel Transformer <https://github.com/laiguokun/Funnel-Transformer>`__ (from CMU/Google Brain) released with the
+    paper `Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing
+    <https://arxiv.org/abs/2006.03236>`__ by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
+14. `GPT <https://huggingface.co/transformers/model_doc/gpt.html>`__ (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.
+15. `GPT-2 <https://huggingface.co/transformers/model_doc/gpt2.html>`__ (from OpenAI) released with the paper `Language
+    Models are Unsupervised Multitask Learners <https://blog.openai.com/better-language-models/>`__ by Alec Radford*,
+    Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
+16. `LayoutLM <https://github.com/microsoft/unilm/tree/master/layoutlm>`__ (from Microsoft Research Asia) released with
+    the paper `LayoutLM: Pre-training of Text and Layout for Document Image Understanding
+    <https://arxiv.org/abs/1912.13318>`__ by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
+17. `Longformer <https://huggingface.co/transformers/model_doc/longformer.html>`__ (from AllenAI) released with the
+    paper `Longformer: The Long-Document Transformer <https://arxiv.org/abs/2004.05150>`__ by Iz Beltagy, Matthew E.
+    Peters, Arman Cohan.
+18. `LXMERT <https://github.com/airsplay/lxmert>`__ (from UNC Chapel Hill) released with the paper `LXMERT: Learning
+    Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering
+    <https://arxiv.org/abs/1908.07490>`__ by Hao Tan and Mohit Bansal.
+19. `MarianMT <https://huggingface.co/transformers/model_doc/marian.html>`__ Machine translation models trained using
+    `OPUS <http://opus.nlpl.eu/>`__ data by Jörg Tiedemann. The `Marian Framework <https://marian-nmt.github.io/>`__ is
+    being developed by the Microsoft Translator Team.
+20. `MBart <https://github.com/pytorch/fairseq/tree/master/examples/mbart>`__ (from Facebook) released with the paper 
+    `Multilingual Denoising Pre-training for Neural Machine Translation <https://arxiv.org/abs/2001.08210>`__ by Yinhan
+    Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
+21. `MMBT <https://github.com/facebookresearch/mmbt/>`__ (from Facebook), released together with the paper a
+    `Supervised Multimodal Bitransformers for Classifying Images and Text <https://arxiv.org/pdf/1909.02950.pdf>`__ by
+    Douwe Kiela, Suvrat Bhooshan, Hamed Firooz, Davide Testuggine.
+22. `Pegasus <https://github.com/google-research/pegasus>`__ (from Google) released with the paper `PEGASUS:
+    Pre-training with Extracted Gap-sentences for Abstractive Summarization <https://arxiv.org/abs/1912.08777>`__> by
+    Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
+23. `Reformer <https://huggingface.co/transformers/model_doc/reformer.html>`__ (from Google Research) released with the
+    paper `Reformer: The Efficient Transformer <https://arxiv.org/abs/2001.04451>`__ by Nikita Kitaev, Łukasz Kaiser,
+    Anselm Levskaya.
+24. `RoBERTa <https://huggingface.co/transformers/model_doc/roberta.html>`__ (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.
+    ultilingual BERT into `DistilmBERT
+    <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__ and a German version of
+    DistilBERT.
+25. `SqueezeBert <https://huggingface.co/transformers/model_doc/squeezebert.html>`__ released with the paper
+    `SqueezeBERT: What can computer vision teach NLP about efficient neural networks?
+    <https://arxiv.org/abs/2006.11316>`__ by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
+26. `T5 <https://huggingface.co/transformers/model_doc/t5.html>`__ (from Google AI) released with the paper `Exploring
+    the Limits of Transfer Learning with a Unified Text-to-Text Transformer <https://arxiv.org/abs/1910.10683>`__ by
+    Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi
+    Zhou and Wei Li and Peter J. Liu.
+27. `Transformer-XL <https://huggingface.co/transformers/model_doc/transformerxl.html>`__ (from Google/CMU) released
+    with the paper `Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context
+    <https://arxiv.org/abs/1901.02860>`__ by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le,
+    Ruslan Salakhutdinov.
+28. `XLM <https://huggingface.co/transformers/model_doc/xlm.html>`__ (from Facebook) released together with the paper
+    `Cross-lingual Language Model Pretraining <https://arxiv.org/abs/1901.07291>`__ by Guillaume Lample and Alexis
+    Conneau.
+29. `XLM-RoBERTa <https://huggingface.co/transformers/model_doc/xlmroberta.html>`__ (from Facebook AI), released
+    together with the paper `Unsupervised Cross-lingual Representation Learning at Scale
+    <https://arxiv.org/abs/1911.02116>`__ by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary,
+    Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
+30. `XLNet <https://huggingface.co/transformers/model_doc/xlnet.html>`__ (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.
+31. `Other community models <https://huggingface.co/models>`__, contributed by the `community
+    <https://huggingface.co/users>`__.
 
 .. toctree::
     :maxdepth: 2
@@ -188,49 +211,62 @@ conversion utilities for the following models:
 
 .. toctree::
     :maxdepth: 2
-    :caption: Package Reference
+    :caption: Main Classes
 
     main_classes/configuration
-    main_classes/output
-    main_classes/model
-    main_classes/tokenizer
-    main_classes/pipelines
-    main_classes/trainer
-    main_classes/optimizer_schedules
-    main_classes/processors
     main_classes/logging
-    model_doc/auto
-    model_doc/encoderdecoder
-    model_doc/bert
-    model_doc/gpt
-    model_doc/transformerxl
-    model_doc/gpt2
-    model_doc/xlm
-    model_doc/xlnet
-    model_doc/roberta
-    model_doc/distilbert
-    model_doc/ctrl
-    model_doc/camembert
+    main_classes/model
+    main_classes/optimizer_schedules
+    main_classes/output
+    main_classes/pipelines
+    main_classes/processors
+    main_classes/tokenizer
+    main_classes/trainer
+
+.. toctree::
+    :maxdepth: 2
+    :caption: Models
+
     model_doc/albert
-    model_doc/xlmroberta
-    model_doc/flaubert
+    model_doc/auto
     model_doc/bart
-    model_doc/t5
-    model_doc/electra
+    model_doc/bert
+    model_doc/bertgeneration
+    model_doc/camembert
+    model_doc/ctrl
+    model_doc/deberta
     model_doc/dialogpt
-    model_doc/reformer
-    model_doc/marian
-    model_doc/longformer
-    model_doc/retribert
-    model_doc/mobilebert
+    model_doc/distilbert
     model_doc/dpr
-    model_doc/pegasus
-    model_doc/mbart
+    model_doc/electra
+    model_doc/encoderdecoder
+    model_doc/flaubert
     model_doc/fsmt
     model_doc/funnel
-    model_doc/lxmert
-    model_doc/bertgeneration
     model_doc/layoutlm
+    model_doc/longformer
+    model_doc/lxmert
+    model_doc/marian
+    model_doc/mbart
+    model_doc/mobilebert
+    model_doc/gpt
+    model_doc/gpt2
+    model_doc/pegasus
+    model_doc/rag
+    model_doc/reformer
+    model_doc/retribert
+    model_doc/roberta
+    model_doc/squeezebert
+    model_doc/t5
+    model_doc/transformerxl
+    model_doc/xlm
+    model_doc/xlmroberta
+    model_doc/xlnet
+
+.. toctree::
+    :maxdepth: 2
+    :caption: Internal Helpers
+
     internal/modeling_utils
-    internal/tokenization_utils
     internal/pipelines_utils
+    internal/tokenization_utils

docs/source/installation.md

@@ -37,13 +37,13 @@ pip install transformers[tf-cpu]
 To check 🤗 Transformers is properly installed, run the following command:
 
 ```bash
-python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('I hate you'))"
+python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('we love you'))"
 ```
 
 It should download a pretrained model then print something like
 
 ```bash
-[{'label': 'NEGATIVE', 'score': 0.9991129040718079}]
+[{'label': 'POSITIVE', 'score': 0.9998704791069031}]
 ```
 
 (Note that TensorFlow will print additional stuff before that last statement.)

docs/source/model_doc/bart.rst

@@ -1,38 +1,46 @@
-Bart
+BART
 -----------------------------------------------------------------------------------------------------------------------
-**DISCLAIMER:** If you see something strange,
-file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
+
+**DISCLAIMER:** If you see something strange, file a `Github Issue
+<https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
 @sshleifer
 
 Overview
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-The Bart model was `proposed <https://arxiv.org/abs/1910.13461>`_ by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer on 29 Oct, 2019.
+The Bart model was proposed in `BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation,
+Translation, and Comprehension <https://arxiv.org/abs/1910.13461>`__ by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan
+Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer on 29 Oct, 2019.
+
 According to the abstract,
 
-- Bart uses a standard seq2seq/machine translation architecture with a bidirectional encoder (like BERT) and a left-to-right decoder (like GPT).
-- The pretraining task involves randomly shuffling the order of the original sentences and a novel in-filling scheme, where spans of text are replaced with a single mask token.
-- BART is particularly effective when fine tuned for text generation but also works well for comprehension tasks. It matches the performance of RoBERTa with comparable training resources on GLUE and SQuAD, achieves new state-of-the-art results on a range of abstractive dialogue, question answering, and summarization tasks, with gains of up to 6 ROUGE.
+- Bart uses a standard seq2seq/machine translation architecture with a bidirectional encoder (like BERT) and a
+  left-to-right decoder (like GPT).
+- The pretraining task involves randomly shuffling the order of the original sentences and a novel in-filling scheme,
+  where spans of text are replaced with a single mask token.
+- BART is particularly effective when fine tuned for text generation but also works well for comprehension tasks. It
+  matches the performance of RoBERTa with comparable training resources on GLUE and SQuAD, achieves new
+  state-of-the-art results on a range of abstractive dialogue, question answering, and summarization tasks, with gains
+  of up to 6 ROUGE.
 
-The Authors' code can be found `here <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_
+The Authors' code can be found `here <https://github.com/pytorch/fairseq/tree/master/examples/bart>`__.
 
 
 Implementation Notes
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-- Bart doesn't use :obj:`token_type_ids` for sequence classification. Use BartTokenizer.encode to get the proper splitting.
-- The forward pass of ``BartModel`` will create decoder inputs (using the helper function ``transformers.modeling_bart._prepare_bart_decoder_inputs``)  if they are not passed. This is different than some other modeling APIs.
-- Model predictions are intended to be identical to the original implementation. This only works, however, if the string you pass to ``fairseq.encode`` starts with a space.
-- ``BartForConditionalGeneration.generate`` should be used for conditional generation tasks like summarization, see the example in that docstrings
-- Models that load the ``"facebook/bart-large-cnn"`` weights will not have a ``mask_token_id``, or be able to perform mask filling tasks.
-- for training/forward passes that don't involve beam search, pass ``use_cache=False``
-
-
-BartForConditionalGeneration
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.BartForConditionalGeneration
-    :members: forward
+- Bart doesn't use :obj:`token_type_ids` for sequence classification. Use :class:`~transformers.BartTokenizer` 
+  or :meth:`~transformers.BartTokenizer.encode` to get the proper splitting.
+- The forward pass of :class:`~transformers.BartModel` will create decoder inputs (using the helper function
+  :func:`transformers.modeling_bart._prepare_bart_decoder_inputs`)  if they are not passed. This is different than some
+  other modeling APIs.
+- Model predictions are intended to be identical to the original implementation. This only works, however, if the
+  string you pass to :func:`fairseq.encode` starts with a space.
+- :meth:`~transformers.BartForConditionalGeneration.generate` should be used for conditional generation tasks like
+  summarization, see the example in that docstrings.
+- Models that load the `facebook/bart-large-cnn` weights will not have a :obj:`mask_token_id`, or be able to perform
+  mask-filling tasks.
+- For training/forward passes that don't involve beam search, pass :obj:`use_cache=False`.
 
 
 BartConfig
@@ -59,6 +67,13 @@ BartModel
 .. autofunction:: transformers.modeling_bart._prepare_bart_decoder_inputs
 
 
+BartForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BartForConditionalGeneration
+    :members: forward
+
+
 BartForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -71,5 +86,3 @@ BartForQuestionAnswering
 
 .. autoclass:: transformers.BartForQuestionAnswering
     :members: forward
-
-

docs/source/model_doc/deberta.rst

@@ -0,0 +1,62 @@
+DeBERTa
+----------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~
+
+The DeBERTa model was proposed in `DeBERTa: Decoding-enhanced BERT with Disentangled Attention <https://arxiv.org/abs/2006.03654>`__
+by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen
+It is based on Google's BERT model released in 2018 and Facebook's RoBERTa model released in 2019.
+
+It builds on RoBERTa with disentangled attention and enhanced mask decoder training with half of the data used in RoBERTa.
+
+The abstract from the paper is the following:
+
+*Recent progress in pre-trained neural language models has significantly improved the performance of many natural language processing (NLP) tasks. 
+In this paper we propose a new model architecture DeBERTa (Decoding-enhanced BERT with disentangled attention) that improves the BERT and RoBERTa 
+models using two novel techniques. The first is the disentangled attention mechanism, where each word is represented using two vectors that encode
+its content and position, respectively, and the attention weights among words are computed using disentangled matrices on their contents and 
+relative positions. Second, an enhanced mask decoder is used to replace the output softmax layer to predict the masked tokens for model pretraining.
+We show that these two techniques significantly improve the efficiency of model pre-training and performance of downstream tasks. Compared to 
+RoBERTa-Large, a DeBERTa model trained on half of the training data performs consistently better on a wide range of NLP tasks, achieving improvements 
+on MNLI by +0.9% (90.2% vs. 91.1%), on SQuAD v2.0 by +2.3% (88.4% vs. 90.7%) and RACE by +3.6% (83.2% vs. 86.8%). The DeBERTa code and pre-trained 
+models will be made publicly available at https://github.com/microsoft/DeBERTa.*
+
+
+The original code can be found `here <https://github.com/microsoft/DeBERTa>`__.
+
+
+DebertaConfig
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaConfig
+    :members:
+
+
+DebertaTokenizer
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+DebertaModel
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaModel
+    :members:
+
+
+DebertaPreTrainedModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaPreTrainedModel
+    :members:
+
+
+DebertaForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaForSequenceClassification
+    :members:

docs/source/model_doc/layoutlm.rst

@@ -4,8 +4,8 @@ LayoutLM
 Overview
 ~~~~~~~~~~~~~~~~~~~~~
 
-The LayoutLM model was proposed in `LayoutLM: Pre-training of Text and Layout for Document Image Understanding <https://arxiv.org/abs/1912.13318>`__
-by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. It's a simple but effective pre-training method 
+The LayoutLM model was proposed in the paper `LayoutLM: Pre-training of Text and Layout for Document Image Understanding <https://arxiv.org/abs/1912.13318>`__
+by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. It's a simple but effective pre-training method
 of text and layout for document image understanding and information extraction tasks, such as form understanding and receipt understanding.
 
 The abstract from the paper is the following:

docs/source/model_doc/marian.rst

@@ -1,36 +1,51 @@
 MarianMT
 -----------------------------------------------------------------------------------------------------------------------
-**Bugs:** If you see something strange,
-file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=sshleifer&labels=&template=bug-report.md&title>`__ and assign
-@sshleifer. Translations should be similar, but not identical to, output in the test set linked to in each model card.
+
+**Bugs:** If you see something strange, file a `Github Issue
+<https://github.com/huggingface/transformers/issues/new?assignees=sshleifer&labels=&template=bug-report.md&title>`__
+and assign @sshleifer. 
+
+Translations should be similar, but not identical to, output in the test set linked to in each model card.
 
 Implementation Notes
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-- Each model is about 298 MB on disk, there are 1,000+ models.
+
+- Each model is about 298 MB on disk, there are more than 1,000 models.
 - The list of supported language pairs can be found `here <https://huggingface.co/Helsinki-NLP>`__.
-- models were originally trained by `Jörg Tiedemann <https://researchportal.helsinki.fi/en/persons/j%C3%B6rg-tiedemann>`__ using the `Marian <https://marian-nmt.github.io/>`_ C++ library, which supports fast training and translation.
-- All models are transformer encoder-decoders with 6 layers in each component. Each model's performance is documented in a model card.
+- Models were originally trained by 
+  `Jörg Tiedemann <https://researchportal.helsinki.fi/en/persons/j%C3%B6rg-tiedemann>`__ using the
+  `Marian <https://marian-nmt.github.io/>`__ C++ library, which supports fast training and translation.
+- All models are transformer encoder-decoders with 6 layers in each component. Each model's performance is documented
+  in a model card.
 - The 80 opus models that require BPE preprocessing are not supported.
-- The modeling code is the same as ``BartForConditionalGeneration`` with a few minor modifications:
-    - static (sinusoid) positional embeddings (``MarianConfig.static_position_embeddings=True``)
-    - a new final_logits_bias (``MarianConfig.add_bias_logits=True``)
-    - no layernorm_embedding (``MarianConfig.normalize_embedding=False``)
-    - the model starts generating with pad_token_id (which has 0 token_embedding) as the prefix. (Bart uses <s/>)
-- Code to bulk convert models can be found in ``convert_marian_to_pytorch.py``
+- The modeling code is the same as :class:`~transformers.BartForConditionalGeneration` with a few minor modifications:
+    - static (sinusoid) positional embeddings (:obj:`MarianConfig.static_position_embeddings=True`)
+    - a new final_logits_bias (:obj:`MarianConfig.add_bias_logits=True`)
+    - no layernorm_embedding (:obj:`MarianConfig.normalize_embedding=False`)
+    - the model starts generating with :obj:`pad_token_id` (which has 0 as a token_embedding) as the prefix (Bart uses
+      :obj:`<s/>`),
+- Code to bulk convert models can be found in ``convert_marian_to_pytorch.py``.
 
 Naming
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-- All  model names use the following format: ``Helsinki-NLP/opus-mt-{src}-{tgt}``
-- The language codes used to name models are inconsistent. Two digit codes can usually be found `here <https://developers.google.com/admin-sdk/directory/v1/languages>`_, three digit codes require googling "language code {code}".
-- Codes formatted like ``es_AR`` are usually ``code_{region}``. That one is spanish documents from Argentina.
+
+- All  model names use the following format: :obj:`Helsinki-NLP/opus-mt-{src}-{tgt}`
+- The language codes used to name models are inconsistent. Two digit codes can usually be found `here
+  <https://developers.google.com/admin-sdk/directory/v1/languages>`__, three digit codes require googling
+  "language code {code}".
+- Codes formatted like :obj:`es_AR` are usually :obj:`code_{region}`. That one is Spanish from Argentina.
 
 
 Multilingual Models
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-All  model names use the following format: ``Helsinki-NLP/opus-mt-{src}-{tgt}``:
-    - if ``src`` is in all caps, the model supports multiple input languages, you can figure out which ones by looking at the model card, or the Group Members `mapping <https://gist.github.com/sshleifer/6d20e7761931b08e73c3219027b97b8a>`_ .
-    - if ``tgt`` is in all caps, the model can output multiple languages, and you should specify a language code by prepending the desired output language to the src_text
+All  model names use the following format: :obj:`Helsinki-NLP/opus-mt-{src}-{tgt}`:
+
+    - If :obj:`src` is in all caps, the model supports multiple input languages, you can figure out which ones by
+      looking at the model card, or the Group Members `mapping
+      <https://gist.github.com/sshleifer/6d20e7761931b08e73c3219027b97b8a>`_ .
+    - If :obj:`tgt` is in all caps, the model can output multiple languages, and you should specify a language code by
+      prepending the desired output language to the :obj:`src_text`.
     - You can see a tokenizer's supported language codes in ``tokenizer.supported_language_codes``
 
 Example of translating english to many romance languages, using language codes:
@@ -54,12 +69,20 @@ Example of translating english to many romance languages, using language codes:
     # 'Isto deve ir para o português.',
     # 'Y esto al español']
 
-Sometimes, models were trained on collections of languages that do not resolve to a group. In this case, _ is used as a separator for src or tgt, as in ``'Helsinki-NLP/opus-mt-en_el_es_fi-en_el_es_fi'``. These still require language codes.
-There are many supported regional language codes, like ``>>es_ES<<`` (Spain) and ``>>es_AR<<`` (Argentina), that do not seem to change translations. I have not found these to provide different results than just using ``>>es<<``.
+Sometimes, models were trained on collections of languages that do not resolve to a group. In this case, _ is used as a
+separator for src or tgt, as in :obj:`Helsinki-NLP/opus-mt-en_el_es_fi-en_el_es_fi`. These still require language
+codes.
 
-For Example:
-    - ``Helsinki-NLP/opus-mt-NORTH_EU-NORTH_EU``: translates from all NORTH_EU languages (see `mapping <https://gist.github.com/sshleifer/6d20e7761931b08e73c3219027b97b8a>`_) to all NORTH_EU languages. Use a special language code like ``>>de<<`` to specify output language.
-    - ``Helsinki-NLP/opus-mt-ROMANCE-en``: translates from many romance languages to english, no codes needed since there is only 1 tgt language.
+There are many supported regional language codes, like :obj:`>>es_ES<<` (Spain) and :obj:`>>es_AR<<` (Argentina), that
+do not seem to change translations. I have not found these to provide different results than just using :obj:`>>es<<`.
+
+For example:
+
+    - `Helsinki-NLP/opus-mt-NORTH_EU-NORTH_EU`: translates from all NORTH_EU languages (see `mapping
+      <https://gist.github.com/sshleifer/6d20e7761931b08e73c3219027b97b8a>`_) to all NORTH_EU languages. Use a special
+      language code like :obj:`>>de<<` to specify output language.
+    - `Helsinki-NLP/opus-mt-ROMANCE-en`: translates from many romance languages to english, no codes needed since there
+      is only one target language.
 
 
 
@@ -86,13 +109,6 @@ Code to see available pretrained models:
     suffix = [x.split('/')[1] for x in model_ids]
     multi_models = [f'{org}/{s}' for s in suffix if s != s.lower()]
 
-MarianMTModel
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-Pytorch version of marian-nmt's transformer.h (c++). Designed for the OPUS-NMT translation checkpoints.
-Model API is identical to BartForConditionalGeneration.
-Available models are listed at `Model List <https://huggingface.co/models?search=Helsinki-NLP>`__
-This class inherits nearly all functionality from ``BartForConditionalGeneration``, see that page for method signatures.
 
 MarianConfig
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -107,5 +123,7 @@ MarianTokenizer
     :members: prepare_seq2seq_batch
 
 
+MarianMTModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-
+.. autoclass:: transformers.MarianMTModel

docs/source/model_doc/mbart.rst

@@ -1,15 +1,20 @@
 MBart
 -----------------------------------------------------------------------------------------------------------------------
-**DISCLAIMER:** If you see something strange,
-file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
+
+**DISCLAIMER:** If you see something strange, file a `Github Issue
+<https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
 @sshleifer
 
 Overview
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-The MBart model was presented in `Multilingual Denoising Pre-training for Neural Machine Translation <https://arxiv.org/abs/2001.08210>`_ by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov
-Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer. According to the abstract,
+The MBart model was presented in `Multilingual Denoising Pre-training for Neural Machine Translation
+<https://arxiv.org/abs/2001.08210>`_ by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov
+Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
 
-MBART is a sequence-to-sequence denoising auto-encoder pre-trained on large-scale monolingual corpora in many languages using the BART objective. mBART is one of the first methods for pre-training a complete sequence-to-sequence model by denoising full texts in multiple languages, while previous approaches have focused only on the encoder, decoder, or reconstructing parts of the text.
+According to the abstract, MBART is a sequence-to-sequence denoising auto-encoder pretrained on large-scale monolingual
+corpora in many languages using the BART objective. mBART is one of the first methods for pre-training a complete
+sequence-to-sequence model by denoising full texts in multiple languages, while previous approaches have focused only
+on the encoder, decoder, or reconstructing parts of the text.
 
 The Authors' code can be found `here <https://github.com/pytorch/fairseq/tree/master/examples/mbart>`__
 
@@ -18,10 +23,11 @@ Training
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 MBart is a multilingual encoder-decoder (seq-to-seq) model primarily intended for translation task. 
 As the model is multilingual it expects the sequences in a different format. A special language id token 
-is added in both the source and target text. The source text format is ``X [eos, src_lang_code]`` 
-where ``X`` is the source text. The target text format is ```[tgt_lang_code] X [eos]```. ```bos``` is never used.
-The ```MBartTokenizer.prepare_seq2seq_batch``` handles this automatically and should be used to encode 
-the sequences for seq-2-seq fine-tuning.
+is added in both the source and target text. The source text format is :obj:`X [eos, src_lang_code]`
+where :obj:`X` is the source text. The target text format is :obj:`[tgt_lang_code] X [eos]`. :obj:`bos` is never used.
+
+The :meth:`~transformers.MBartTokenizer.prepare_seq2seq_batch` handles this automatically and should be used to encode 
+the sequences for sequence-to-sequence fine-tuning.
 
 - Supervised training
 
@@ -38,8 +44,8 @@ the sequences for seq-2-seq fine-tuning.
 
 - Generation
 
-    While generating the target text set the `decoder_start_token_id` to the target language id. 
-    The following example shows how to translate English to Romanian using the ```facebook/mbart-large-en-ro``` model.
+    While generating the target text set the :obj:`decoder_start_token_id` to the target language id. 
+    The following example shows how to translate English to Romanian using the `facebook/mbart-large-en-ro` model.
 
 .. code-block::
 
@@ -71,6 +77,4 @@ MBartForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.MBartForConditionalGeneration
-    :members: generate, forward
-
-
+    :members: forward

docs/source/model_doc/pegasus.rst

@@ -1,30 +1,40 @@
 Pegasus
 -----------------------------------------------------------------------------------------------------------------------
-**DISCLAIMER:** If you see something strange,
-file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=sshleifer&labels=&template=bug-report.md&title>`__ and assign
-@sshleifer.
+
+**DISCLAIMER:** If you see something strange, file a `Github Issue
+<https://github.com/huggingface/transformers/issues/new?assignees=sshleifer&labels=&template=bug-report.md&title>`__
+and assign @sshleifer.
 
 
 Overview
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 The Pegasus model was proposed in `PEGASUS: Pre-training with Extracted Gap-sentences for
-Abstractive Summarization <https://arxiv.org/pdf/1912.08777.pdf>`_ by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu on Dec 18, 2019.
+Abstractive Summarization <https://arxiv.org/pdf/1912.08777.pdf>`__ by Jingqing Zhang, Yao Zhao, Mohammad Saleh and
+Peter J. Liu on Dec 18, 2019.
+
 According to the abstract,
 
-- Pegasus' pretraining task is intentionally similar to summarization: important sentences are removed/masked from an input document and are generated together as one output sequence from the remaining sentences, similar to an extractive summary.
+- Pegasus' pretraining task is intentionally similar to summarization: important sentences are removed/masked from an
+  input document and are generated together as one output sequence from the remaining sentences, similar to an
+  extractive summary.
 - Pegasus achieves SOTA summarization performance on all 12 downstream tasks, as measured by ROUGE and human eval.
 
-The Authors' code can be found `here <https://github.com/google-research/pegasus>`_.
+The Authors' code can be found `here <https://github.com/google-research/pegasus>`__.
 
 
 Checkpoints
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-All the `checkpoints <https://huggingface.co/models?search=pegasus>`_ are finetuned for summarization, besides ``pegasus-large``, whence the other checkpoints are finetuned.
+
+All the `checkpoints <https://huggingface.co/models?search=pegasus>`__ are fine-tuned for summarization, besides 
+`pegasus-large`, whence the other checkpoints are fine-tuned:
+
 - Each checkpoint is 2.2 GB on disk and 568M parameters.
 - FP16 is not supported (help/ideas on this appreciated!).
 - Summarizing xsum in fp32 takes about 400ms/sample, with default parameters on a v100 GPU.
-- For XSUM, The paper reports rouge1,rouge2, rougeL of paper: 47.21/24.56/39.25. As of Aug 9, this port scores 46.91/24.34/39.1.
+- For XSUM, The paper reports rouge1,rouge2, rougeL of paper: 47.21/24.56/39.25. As of Aug 9, this port scores
+  46.91/24.34/39.1.
+
 The gap is likely because of different alpha/length_penalty implementations in beam search.
 
 
@@ -32,14 +42,16 @@ Implementation Notes
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 - All models are transformer encoder-decoders with 16 layers in each component.
-- The implementation is completely inherited from ``BartForConditionalGeneration``
+- The implementation is completely inherited from :class:`~transformers.BartForConditionalGeneration`
 - Some key configuration differences:
     - static, sinusoidal position embeddings
-    - no ``layernorm_embedding`` (``PegasusConfig.normalize_embedding=False``)
+    - no :obj:`layernorm_embedding` (:obj`PegasusConfig.normalize_embedding=False`)
     - the model starts generating with pad_token_id (which has 0 token_embedding) as the prefix.
-    - ``num_beams=8``
-- All pretrained pegasus checkpoints are the same besides three attributes: ``tokenizer.model_max_length`` (max input size),  ``max_length`` (max num tokens to generate) and ``length_penalty``
-- Code to convert checkpoints trained in the author's `repo <https://github.com/google-research/pegasus>`_ can be found in ``convert_pegasus_tf_to_pytorch.py``
+    - more beams are used (:obj:`num_beams=8`)
+- All pretrained pegasus checkpoints are the same besides three attributes: :obj:`tokenizer.model_max_length` (maximum
+  input size), :obj:`max_length` (the maximum number of tokens to generate) and :obj:`length_penalty`.
+- The code to convert checkpoints trained in the author's `repo <https://github.com/google-research/pegasus>`_ can be
+  found in ``convert_pegasus_tf_to_pytorch.py``.
 
 
 Usage Example
@@ -62,48 +74,12 @@ Usage Example
     tgt_text = tokenizer.batch_decode(translated, skip_special_tokens=True)
     assert tgt_text[0] == "California's largest electricity provider has turned off power to hundreds of thousands of customers."
 
-PegasusForConditionalGeneration
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-This class inherits all functionality from ``BartForConditionalGeneration``, see that page for method signatures.
-Available models are listed at `Model List <https://huggingface.co/models?search=pegasus>`__
-
-.. autoclass:: transformers.PegasusForConditionalGeneration
-    :members:
 
 
 PegasusConfig
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-This config fully inherits from ``BartConfig``, but pegasus uses different default values:
-Up to date parameter values can be seen in `S3 <https://s3.amazonaws.com/models.huggingface.co/bert/google/pegasus-xsum/config.json>`_.
-As of Aug 10, 2020, they are:
 
-.. code-block:: python
-
-    dict(
-    vocab_size=96103,
-    max_position_embeddings=512,
-    d_model=1024,
-    encoder_ffn_dim=4096,
-    decoder_ffn_dim=4096,
-    encoder_attention_heads=16,
-    decoder_attention_heads=16,
-    encoder_layers=16,
-    decoder_layers=16,
-    dropout=0.1,
-    attention_dropout=0.1,
-    activation_dropout=0.1,
-    pad_token_id=0,
-    eos_token_id=1,
-    is_encoder_decoder=True,
-    normalize_before=True,
-    scale_embedding=True,
-    normalize_embedding=False,
-    add_final_layer_norm=True,
-    static_position_embeddings=True,
-    num_beams=8,
-    activation_function="relu",
-    )
+.. autoclass:: transformers.PegasusConfig
 
 
 PegasusTokenizer
@@ -114,4 +90,7 @@ warning: ``add_tokens`` does not work at the moment.
     :members: __call__, prepare_seq2seq_batch
 
 
+PegasusForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
+.. autoclass:: transformers.PegasusForConditionalGeneration

docs/source/model_doc/rag.rst

@@ -0,0 +1,91 @@
+RAG
+----------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~
+
+Retrieval-augmented generation ("RAG") models combine the powers of pretrained dense retrieval (DPR) and
+sequence-to-sequence models. RAG models retrieve documents, pass them to a seq2seq model, then marginalize to generate
+outputs. The retriever and seq2seq modules are initialized from pretrained models, and fine-tuned jointly, allowing
+both retrieval and generation to adapt to downstream tasks.
+
+It is based on the paper `Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
+<https://arxiv.org/abs/2005.11401>`__ by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir
+Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
+
+The abstract from the paper is the following:
+
+*Large pre-trained language models have been shown to store factual knowledge
+in their parameters, and achieve state-of-the-art results when fine-tuned on
+downstream NLP tasks. However, their ability to access and precisely manipulate
+knowledge is still limited, and hence on knowledge-intensive tasks, their
+performance lags behind task-specific architectures. Additionally, providing
+provenance for their decisions and updating their world knowledge remain open
+research problems. Pre-trained models with a differentiable access mechanism to
+explicit nonparametric memory can overcome this issue, but have so far been only
+investigated for extractive downstream tasks. We explore a general-purpose
+fine-tuning recipe for retrieval-augmented generation (RAG) — models which combine
+pre-trained parametric and non-parametric memory for language generation. We
+introduce RAG models where the parametric memory is a pre-trained seq2seq model and
+the non-parametric memory is a dense vector index of Wikipedia, accessed with
+a pre-trained neural retriever. We compare two RAG formulations, one which
+conditions on the same retrieved passages across the whole generated sequence, the
+other can use different passages per token. We fine-tune and evaluate our models
+on a wide range of knowledge-intensive NLP tasks and set the state-of-the-art
+on three open domain QA tasks, outperforming parametric seq2seq models and
+task-specific retrieve-and-extract architectures. For language generation tasks, we
+find that RAG models generate more specific, diverse and factual language than a
+state-of-the-art parametric-only seq2seq baseline.*
+
+
+
+RagConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RagConfig
+    :members:
+
+
+RagTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RagTokenizer
+    :members: prepare_seq2seq_batch
+
+
+Rag specific outputs
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_rag.RetrievAugLMMarginOutput
+    :members:
+
+.. autoclass:: transformers.modeling_rag.RetrievAugLMOutput
+    :members:
+
+
+RAGRetriever
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RagRetriever
+    :members:
+
+
+RagModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RagModel
+    :members: forward
+
+
+RagSequenceForGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RagSequenceForGeneration
+    :members: forward, generate
+
+
+RagTokenForGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RagTokenForGeneration
+    :members: forward, generate

docs/source/model_doc/squeezebert.rst

@@ -0,0 +1,103 @@
+SqueezeBERT
+----------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~
+
+The SqueezeBERT model was proposed in
+`SqueezeBERT: What can computer vision teach NLP about efficient neural networks?
+<https://arxiv.org/abs/2006.11316>`__
+by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, Kurt W. Keutzer.
+It's a bidirectional transformer similar to the BERT model.
+The key difference between the BERT architecture and the SqueezeBERT architecture
+is that SqueezeBERT uses `grouped convolutions <https://blog.yani.io/filter-group-tutorial>`__
+instead of fully-connected layers for the Q, K, V and FFN layers.
+
+The abstract from the paper is the following:
+
+*Humans read and write hundreds of billions of messages every day. Further, due to the availability of
+large datasets, large computing systems, and better neural network models, natural language processing (NLP)
+technology has made significant strides in understanding, proofreading, and organizing these messages.
+Thus, there is a significant opportunity to deploy NLP in myriad applications to help web users,
+social networks, and businesses. In particular, we consider smartphones and other mobile devices as
+crucial platforms for deploying NLP models at scale. However, today's highly-accurate NLP neural network
+models such as BERT and RoBERTa are extremely computationally expensive, with BERT-base taking 1.7 seconds
+to classify a text snippet on a Pixel 3 smartphone. In this work, we observe that methods such as grouped
+convolutions have yielded significant speedups for computer vision networks, but many of these techniques
+have not been adopted by NLP neural network designers. We demonstrate how to replace several operations in
+self-attention layers with grouped convolutions, and we use this technique in a novel network architecture
+called SqueezeBERT, which runs 4.3x faster than BERT-base on the Pixel 3 while achieving competitive
+accuracy on the GLUE test set. The SqueezeBERT code will be released.*
+
+Tips:
+
+- SqueezeBERT is a model with absolute position embeddings so it's usually advised to pad the inputs on
+  the right rather than the left.
+- SqueezeBERT is similar to BERT and therefore relies on the masked language modeling (MLM) objective.
+  It is therefore efficient at predicting masked tokens and at NLU in general, but is not optimal for
+  text generation. Models trained with a causal language modeling (CLM) objective are better in that regard.
+- For best results when finetuning on sequence classification tasks, it is recommended to start with the
+  `squeezebert/squeezebert-mnli-headless` checkpoint.
+
+SqueezeBertConfig
+~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SqueezeBertConfig
+    :members:
+
+
+SqueezeBertTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SqueezeBertTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+SqueezeBertTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SqueezeBertTokenizerFast
+    :members:
+
+
+SqueezeBertModel
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SqueezeBertModel
+    :members:
+
+
+SqueezeBertForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SqueezeBertForMaskedLM
+    :members:
+
+
+SqueezeBertForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SqueezeBertForSequenceClassification
+    :members:
+
+
+SqueezeBertForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SqueezeBertForMultipleChoice
+    :members:
+
+
+SqueezeBertForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SqueezeBertForTokenClassification
+    :members:
+
+
+SqueezeBertForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SqueezeBertForQuestionAnswering
+    :members:

docs/source/model_doc/transformerxl.rst

@@ -46,13 +46,6 @@ TransfoXLTokenizer
     :members: save_vocabulary
 
 
-TransfoXLTokenizerFast
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TransfoXLTokenizerFast
-    :members:
-
-
 TransfoXL specific outputs
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

docs/source/model_sharing.rst

@@ -112,8 +112,7 @@ Make sure there are no garbage files in the directory you'll upload. It should o
 - a `tf_model.h5` file, which is the TensorFlow checkpoint (unless you can't have it for some reason) ;
 - a `special_tokens_map.json`, which is part of your :doc:`tokenizer <main_classes/tokenizer>` save;
 - a `tokenizer_config.json`, which is part of your :doc:`tokenizer <main_classes/tokenizer>` save;
-- a `vocab.txt`, which is the vocabulary of your tokenizer, part of your :doc:`tokenizer <main_classes/tokenizer>`
-  save;
+- files named `vocab.json`, `vocab.txt`, `merges.txt`, or similar, which contain the vocabulary of your tokenizer, part of your :doc:`tokenizer <main_classes/tokenizer>` save;
 - maybe a `added_tokens.json`, which is part of your :doc:`tokenizer <main_classes/tokenizer>` save.
 
 Other files can safely be deleted.
@@ -221,4 +220,3 @@ You can also delete unneeded files with
 .. code-block::
 
     transformers-cli s3 rm awesome-name-you-picked/filename
-

docs/source/model_summary.rst

@@ -672,6 +672,27 @@ DPR consists in three models:
 
 DPR's pipeline (not implemented yet) uses a retrieval step to find the top k contexts given a certain question, and then it calls the reader with the question and the retrieved documents to get the answer.
 
+RAG
+-----------------------------------------------------------------------------------------------------------------------
+
+.. raw:: html
+
+   <a href="https://huggingface.co/models?filter=rag">
+       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-rag-blueviolet">
+   </a>
+   <a href="model_doc/rag.html">
+       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-rag-blueviolet">
+   </a>
+
+`Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks <https://arxiv.org/abs/2005.11401>`_,
+Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela
+
+Retrieval-augmented generation ("RAG") models combine the powers of pretrained dense retrieval (DPR) and Seq2Seq models.
+RAG models retrieve docs, pass them to a seq2seq model, then marginalize to generate outputs.
+The retriever and seq2seq modules are initialized from pretrained models, and fine-tuned jointly, allowing both retrieval and generation to adapt to downstream tasks.
+
+The two models RAG-Token and RAG-Sequence are available for generation.
+
 More technical aspects
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 

docs/source/pretrained_models.rst

@@ -11,26 +11,26 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 | BERT               | ``bert-base-uncased``                                      | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                    |                                                            | | Trained on lower-cased English text.                                                                                                |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``bert-large-uncased``                                     | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                    | ``bert-large-uncased``                                     | | 24-layer, 1024-hidden, 16-heads, 336M parameters.                                                                                   |
 |                    |                                                            | | Trained on lower-cased English text.                                                                                                |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``bert-base-cased``                                        | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                    | ``bert-base-cased``                                        | | 12-layer, 768-hidden, 12-heads, 109M parameters.                                                                                    |
 |                    |                                                            | | Trained on cased English text.                                                                                                      |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``bert-large-cased``                                       | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                    | ``bert-large-cased``                                       | | 24-layer, 1024-hidden, 16-heads, 335M parameters.                                                                                   |
 |                    |                                                            | | Trained on cased English text.                                                                                                      |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``bert-base-multilingual-uncased``                         | | (Original, not recommended) 12-layer, 768-hidden, 12-heads, 110M parameters.                                                        |
+|                    | ``bert-base-multilingual-uncased``                         | | (Original, not recommended) 12-layer, 768-hidden, 12-heads, 168M parameters.                                                        |
 |                    |                                                            | | Trained on lower-cased text in the top 102 languages with the largest Wikipedias                                                    |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details <https://github.com/google-research/bert/blob/master/multilingual.md>`__).                                              |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``bert-base-multilingual-cased``                           | | (New, **recommended**) 12-layer, 768-hidden, 12-heads, 110M parameters.                                                             |
+|                    | ``bert-base-multilingual-cased``                           | | (New, **recommended**) 12-layer, 768-hidden, 12-heads, 179M parameters.                                                             |
 |                    |                                                            | | Trained on cased text in the top 104 languages with the largest Wikipedias                                                          |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details <https://github.com/google-research/bert/blob/master/multilingual.md>`__).                                              |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``bert-base-chinese``                                      | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                    | ``bert-base-chinese``                                      | | 12-layer, 768-hidden, 12-heads, 103M parameters.                                                                                    |
 |                    |                                                            | | Trained on cased Chinese Simplified and Traditional text.                                                                           |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                    | ``bert-base-german-cased``                                 | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
@@ -38,22 +38,22 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details on deepset.ai website <https://deepset.ai/german-bert>`__).                                                             |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``bert-large-uncased-whole-word-masking``                  | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                    | ``bert-large-uncased-whole-word-masking``                  | | 24-layer, 1024-hidden, 16-heads, 336M parameters.                                                                                   |
 |                    |                                                            | | Trained on lower-cased English text using Whole-Word-Masking                                                                        |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details <https://github.com/google-research/bert/#bert>`__).                                                                    |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``bert-large-cased-whole-word-masking``                    | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                    | ``bert-large-cased-whole-word-masking``                    | | 24-layer, 1024-hidden, 16-heads, 335M parameters.                                                                                   |
 |                    |                                                            | | Trained on cased English text using Whole-Word-Masking                                                                              |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details <https://github.com/google-research/bert/#bert>`__).                                                                    |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``bert-large-uncased-whole-word-masking-finetuned-squad``  | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                    | ``bert-large-uncased-whole-word-masking-finetuned-squad``  | | 24-layer, 1024-hidden, 16-heads, 336M 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>`__).           |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``bert-large-cased-whole-word-masking-finetuned-squad``    | | 24-layer, 1024-hidden, 16-heads, 340M parameters                                                                                    |
+|                    | ``bert-large-cased-whole-word-masking-finetuned-squad``    | | 24-layer, 1024-hidden, 16-heads, 335M 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>`__)                           |
@@ -73,31 +73,31 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``cl-tohoku/bert-base-japanese``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                    | ``cl-tohoku/bert-base-japanese``                           | | 12-layer, 768-hidden, 12-heads, 111M parameters.                                                                                    |
 |                    |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece and this requires some extra dependencies,                     |
 |                    |                                                            | | `fugashi <https://github.com/polm/fugashi>`__ which is a wrapper around `MeCab <https://taku910.github.io/mecab/>`__.               |
 |                    |                                                            | | Use ``pip install transformers["ja"]`` (or ``pip install -e .["ja"]`` if you install from source) to install them.                  |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``cl-tohoku/bert-base-japanese-whole-word-masking``        | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                    | ``cl-tohoku/bert-base-japanese-whole-word-masking``        | | 12-layer, 768-hidden, 12-heads, 111M parameters.                                                                                    |
 |                    |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece and this requires some extra dependencies,                     |
 |                    |                                                            | | `fugashi <https://github.com/polm/fugashi>`__ which is a wrapper around `MeCab <https://taku910.github.io/mecab/>`__.               |
 |                    |                                                            | | Use ``pip install transformers["ja"]`` (or ``pip install -e .["ja"]`` if you install from source) to install them.                  |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``cl-tohoku/bert-base-japanese-char``                      | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                    | ``cl-tohoku/bert-base-japanese-char``                      | | 12-layer, 768-hidden, 12-heads, 90M parameters.                                                                                     |
 |                    |                                                            | | Trained on Japanese text. Text is tokenized into characters.                                                                        |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``cl-tohoku/bert-base-japanese-char-whole-word-masking``   | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                    | ``cl-tohoku/bert-base-japanese-char-whole-word-masking``   | | 12-layer, 768-hidden, 12-heads, 90M parameters.                                                                                     |
 |                    |                                                            | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized into characters.                                               |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``TurkuNLP/bert-base-finnish-cased-v1``                    | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                    | ``TurkuNLP/bert-base-finnish-cased-v1``                    | | 12-layer, 768-hidden, 12-heads, 125M parameters.                                                                                    |
 |                    |                                                            | | Trained on cased Finnish text.                                                                                                      |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details on turkunlp.org <http://turkunlp.org/FinBERT/>`__).                                                                     |
@@ -415,4 +415,24 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                    | ``microsoft/layoutlm-large-uncased``                       | | 24 layers, 1024-hidden, 16-heads, 343M parameters                                                                                   |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details <https://github.com/microsoft/unilm/tree/master/layoutlm>`__)                                                           |
-+--------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
++--------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| DeBERTa            | ``microsoft/deberta-base``                                 | | 12-layer, 768-hidden, 12-heads, ~125M parameters                                                                                    |
+|                    |                                                            | | DeBERTa using the BERT-base architecture                                                                                            |
+|                    |                                                            |                                                                                                                                       |
+|                    |                                                            | (see `details <https://github.com/microsoft/DeBERTa>`__)                                                                              |
+|                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                    | ``microsoft/deberta-large``                                | | 24-layer, 1024-hidden, 16-heads, ~390M parameters                                                                                   |
+|                    |                                                            | | DeBERTa using the BERT-large architecture                                                                                           |
+|                    |                                                            |                                                                                                                                       |
+|                    |                                                            | (see `details <https://github.com/microsoft/DeBERTa>`__)                                                                              |
++--------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| SqueezeBERT        | ``squeezebert/squeezebert-uncased``                        | | 12-layer, 768-hidden, 12-heads, 51M parameters, 4.3x faster than bert-base-uncased on a smartphone.                                 |
+|                    |                                                            | | SqueezeBERT architecture pretrained from scratch on masked language model (MLM) and sentence order prediction (SOP) tasks.          |
+|                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                    | ``squeezebert/squeezebert-mnli``                           | | 12-layer, 768-hidden, 12-heads, 51M parameters, 4.3x faster than bert-base-uncased on a smartphone.                                 |
+|                    |                                                            | | This is the squeezebert-uncased model finetuned on MNLI sentence pair classification task with distillation from electra-base.      |
+|                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                    | ``squeezebert/squeezebert-mnli-headless``                  | | 12-layer, 768-hidden, 12-heads, 51M parameters, 4.3x faster than bert-base-uncased on a smartphone.                                 |
+|                    |                                                            | | This is the squeezebert-uncased model finetuned on MNLI sentence pair classification task with distillation from electra-base.      |
+|                    |                                                            | | The final classification layer is removed, so when you finetune, the final layer will be reinitialized.                             |
++--------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+

docs/source/task_summary.rst

@@ -89,7 +89,7 @@ of each other. The process is the following:
     >>> import torch
 
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased-finetuned-mrpc")
-    >>> model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased-finetuned-mrpc")
+    >>> model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased-finetuned-mrpc", return_dict=True)
 
     >>> classes = ["not paraphrase", "is paraphrase"]
 
@@ -122,7 +122,7 @@ of each other. The process is the following:
     >>> import tensorflow as tf
 
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased-finetuned-mrpc")
-    >>> model = TFAutoModelForSequenceClassification.from_pretrained("bert-base-cased-finetuned-mrpc")
+    >>> model = TFAutoModelForSequenceClassification.from_pretrained("bert-base-cased-finetuned-mrpc", return_dict=True)
 
     >>> classes = ["not paraphrase", "is paraphrase"]
 
@@ -213,7 +213,7 @@ Here is an example of question answering using a model and a tokenizer. The proc
     >>> import torch
 
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
-    >>> model = AutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
+    >>> model = AutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad", return_dict=True)
 
     >>> text = r"""
     ... 🤗 Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides general-purpose
@@ -255,7 +255,7 @@ Here is an example of question answering using a model and a tokenizer. The proc
     >>> import tensorflow as tf
 
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
-    >>> model = TFAutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
+    >>> model = TFAutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad", return_dict=True)
 
     >>> text = r"""
     ... 🤗 Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides general-purpose
@@ -378,7 +378,7 @@ Here is an example of doing masked language modeling using a model and a tokeniz
     >>> import torch
 
     >>> tokenizer = AutoTokenizer.from_pretrained("distilbert-base-cased")
-    >>> model = AutoModelWithLMHead.from_pretrained("distilbert-base-cased")
+    >>> model = AutoModelWithLMHead.from_pretrained("distilbert-base-cased", return_dict=True)
 
     >>> sequence = f"Distilled models are smaller than the models they mimic. Using them instead of the large versions would help {tokenizer.mask_token} our carbon footprint."
 
@@ -394,7 +394,7 @@ Here is an example of doing masked language modeling using a model and a tokeniz
     >>> import tensorflow as tf
 
     >>> tokenizer = AutoTokenizer.from_pretrained("distilbert-base-cased")
-    >>> model = TFAutoModelWithLMHead.from_pretrained("distilbert-base-cased")
+    >>> model = TFAutoModelWithLMHead.from_pretrained("distilbert-base-cased", return_dict=True)
 
     >>> sequence = f"Distilled models are smaller than the models they mimic. Using them instead of the large versions would help {tokenizer.mask_token} our carbon footprint."
 
@@ -439,7 +439,7 @@ Here is an example of using the tokenizer and model and leveraging the :func:`~t
     >>> from torch.nn import functional as F
 
     >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
-    >>> model = AutoModelWithLMHead.from_pretrained("gpt2")
+    >>> model = AutoModelWithLMHead.from_pretrained("gpt2", return_dict=True)
 
     >>> sequence = f"Hugging Face is based in DUMBO, New York City, and "
 
@@ -463,7 +463,7 @@ Here is an example of using the tokenizer and model and leveraging the :func:`~t
     >>> import tensorflow as tf
 
     >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
-    >>> model = TFAutoModelWithLMHead.from_pretrained("gpt2")
+    >>> model = TFAutoModelWithLMHead.from_pretrained("gpt2", return_dict=True)
 
     >>> sequence = f"Hugging Face is based in DUMBO, New York City, and "
 
@@ -517,7 +517,7 @@ Here is an example of text generation using ``XLNet`` and its tokenzier.
     >>> ## PYTORCH CODE
     >>> from transformers import AutoModelWithLMHead, AutoTokenizer
 
-    >>> model = AutoModelWithLMHead.from_pretrained("xlnet-base-cased")
+    >>> model = AutoModelWithLMHead.from_pretrained("xlnet-base-cased", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("xlnet-base-cased")
 
     >>> # Padding text helps XLNet with short prompts - proposed by Aman Rusia in https://github.com/rusiaaman/XLNet-gen#methodology
@@ -542,7 +542,7 @@ Here is an example of text generation using ``XLNet`` and its tokenzier.
     >>> ## TENSORFLOW CODE
     >>> from transformers import TFAutoModelWithLMHead, AutoTokenizer
 
-    >>> model = TFAutoModelWithLMHead.from_pretrained("xlnet-base-cased")
+    >>> model = TFAutoModelWithLMHead.from_pretrained("xlnet-base-cased", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("xlnet-base-cased")
 
     >>> # Padding text helps XLNet with short prompts - proposed by Aman Rusia in https://github.com/rusiaaman/XLNet-gen#methodology
@@ -659,7 +659,7 @@ Here is an example of doing named entity recognition, using a model and a tokeni
     >>> from transformers import AutoModelForTokenClassification, AutoTokenizer
     >>> import torch
 
-    >>> model = AutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english")
+    >>> model = AutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
 
     >>> label_list = [
@@ -687,7 +687,7 @@ Here is an example of doing named entity recognition, using a model and a tokeni
     >>> from transformers import TFAutoModelForTokenClassification, AutoTokenizer
     >>> import tensorflow as tf
 
-    >>> model = TFAutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english")
+    >>> model = TFAutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
 
     >>> label_list = [
@@ -781,7 +781,7 @@ In this example we use Google`s T5 model. Even though it was pre-trained only on
     >>> ## PYTORCH CODE
     >>> from transformers import AutoModelWithLMHead, AutoTokenizer
 
-    >>> model = AutoModelWithLMHead.from_pretrained("t5-base")
+    >>> model = AutoModelWithLMHead.from_pretrained("t5-base", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
 
     >>> # T5 uses a max_length of 512 so we cut the article to 512 tokens.
@@ -790,7 +790,7 @@ In this example we use Google`s T5 model. Even though it was pre-trained only on
     >>> ## TENSORFLOW CODE
     >>> from transformers import TFAutoModelWithLMHead, AutoTokenizer
 
-    >>> model = TFAutoModelWithLMHead.from_pretrained("t5-base")
+    >>> model = TFAutoModelWithLMHead.from_pretrained("t5-base", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
 
     >>> # T5 uses a max_length of 512 so we cut the article to 512 tokens.
@@ -834,7 +834,7 @@ Here is an example of doing translation using a model and a tokenizer. The proce
     >>> ## PYTORCH CODE
     >>> from transformers import AutoModelWithLMHead, AutoTokenizer
 
-    >>> model = AutoModelWithLMHead.from_pretrained("t5-base")
+    >>> model = AutoModelWithLMHead.from_pretrained("t5-base", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
 
     >>> inputs = tokenizer.encode("translate English to German: Hugging Face is a technology company based in New York and Paris", return_tensors="pt")
@@ -842,7 +842,7 @@ Here is an example of doing translation using a model and a tokenizer. The proce
     >>> ## TENSORFLOW CODE
     >>> from transformers import TFAutoModelWithLMHead, AutoTokenizer
 
-    >>> model = TFAutoModelWithLMHead.from_pretrained("t5-base")
+    >>> model = TFAutoModelWithLMHead.from_pretrained("t5-base", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
 
     >>> inputs = tokenizer.encode("translate English to German: Hugging Face is a technology company based in New York and Paris", return_tensors="tf")

examples/README.md

@@ -47,9 +47,7 @@ pip install -r ./examples/requirements.txt
 
 ## One-click Deploy to Cloud (wip)
 
-#### Azure
-
-[![Deploy to Azure](https://aka.ms/deploytoazurebutton)](https://portal.azure.com/#create/Microsoft.Template/uri/https%3A%2F%2Fraw.githubusercontent.com%2FAzure%2Fazure-quickstart-templates%2Fmaster%2F101-storage-account-create%2Fazuredeploy.json)
+**Coming soon!**
 
 ## Running on TPUs
 

examples/seq2seq/builtin_trainer/train_distil_marian_enro.sh

@@ -19,5 +19,4 @@ python finetune_trainer.py \
     --do_train --do_eval --do_predict --evaluate_during_training\
     --predict_with_generate --logging_first_step \
     --task translation --label_smoothing 0.1 \
-    --run_name marian_en_ro_6_3 \
     "$@"

examples/seq2seq/builtin_trainer/train_distil_marian_enro_tpu.sh

@@ -20,5 +20,4 @@ python xla_spawn.py --num_cores $TPU_NUM_CORES \
     --do_train --do_eval --evaluate_during_training \
     --prediction_loss_only \
     --task translation --label_smoothing 0.1 \
-    --run_name marian_en_ro_6_3 \
     "$@"

examples/seq2seq/builtin_trainer/train_distilbart_cnn.sh

@@ -19,8 +19,7 @@ python finetune_trainer.py \
     --num_train_epochs=2 \
     --save_steps 3000 --eval_steps 3000 \
     --logging_first_step \
-    --max_target_length $MAX_TGT_LEN --val_max_target_length $MAX_TGT_LEN --test_max_target_length $MAX_TGT_LEN \
+    --max_target_length 56 --val_max_target_length $MAX_TGT_LEN --test_max_target_length $MAX_TGT_LEN \
     --do_train --do_eval --do_predict --evaluate_during_training \
     --predict_with_generate \
-    --run_name distilbart-cnn-12-6 \
     "$@"

examples/seq2seq/builtin_trainer/train_mbart_cc25_enro.sh

@@ -18,5 +18,4 @@ python finetune_trainer.py \
     --do_train --do_eval --do_predict --evaluate_during_training \
     --predict_with_generate --logging_first_step 
     --task translation \
-    --run_name mbart_en_ro \
     "$@"

examples/seq2seq/finetune.py

@@ -26,6 +26,7 @@ from utils import (
     calculate_bleu,
     calculate_rouge,
     flatten_list,
+    freeze_embeds,
     freeze_params,
     get_git_info,
     label_smoothed_nll_loss,
@@ -90,7 +91,7 @@ class SummarizationModule(BaseTransformer):
         assert self.target_lens["train"] <= self.target_lens["val"], f"target_lens: {self.target_lens}"
         assert self.target_lens["train"] <= self.target_lens["test"], f"target_lens: {self.target_lens}"
         if self.hparams.freeze_embeds:
-            self.freeze_embeds()
+            freeze_embeds(self.model)
         if self.hparams.freeze_encoder:
             freeze_params(self.model.get_encoder())
             assert_all_frozen(self.model.get_encoder())
@@ -105,29 +106,12 @@ class SummarizationModule(BaseTransformer):
             Seq2SeqDataset if hasattr(self.tokenizer, "prepare_seq2seq_batch") else LegacySeq2SeqDataset
         )
         self.eval_beams = self.model.config.num_beams if self.hparams.eval_beams is None else self.hparams.eval_beams
-        assert self.eval_beams >= 1, f"got self.eval_beams={self.eval_beams}. Need an integer > 1"
         if self.hparams.eval_max_gen_length is not None:
             self.eval_max_length = self.hparams.eval_max_gen_length
         else:
             self.eval_max_length = self.model.config.max_length
         self.val_metric = self.default_val_metric if self.hparams.val_metric is None else self.hparams.val_metric
 
-    def freeze_embeds(self):
-        """Freeze token embeddings and positional embeddings for bart, just token embeddings for t5."""
-        if self.model_type == "t5":
-            freeze_params(self.model.shared)
-            for d in [self.model.encoder, self.model.decoder]:
-                freeze_params(d.embed_tokens)
-        elif self.model_type == "fsmt":
-            for d in [self.model.model.encoder, self.model.model.decoder]:
-                freeze_params(d.embed_positions)
-                freeze_params(d.embed_tokens)
-        else:
-            freeze_params(self.model.model.shared)
-            for d in [self.model.model.encoder, self.model.model.decoder]:
-                freeze_params(d.embed_positions)
-                freeze_params(d.embed_tokens)
-
     def forward(self, input_ids, **kwargs):
         return self.model(input_ids, **kwargs)
 

examples/seq2seq/finetune_trainer.py

@@ -1,111 +1,38 @@
-import json
 import logging
 import os
 import sys
 from dataclasses import dataclass, field
-from typing import Callable, Dict, List, Optional, Tuple
-
-import numpy as np
-import torch
+from typing import Optional
 
 from seq2seq_trainer import Seq2SeqTrainer
 from transformers import (
     AutoConfig,
     AutoModelForSeq2SeqLM,
     AutoTokenizer,
-    BartTokenizer,
-    EvalPrediction,
     HfArgumentParser,
     MBartTokenizer,
-    T5Tokenizer,
     TrainingArguments,
     set_seed,
 )
-from transformers.modeling_bart import shift_tokens_right
+from transformers.trainer_utils import EvaluationStrategy
 from utils import (
     LegacySeq2SeqDataset,
+    Seq2SeqDataCollator,
     Seq2SeqDataset,
     assert_all_frozen,
-    calculate_bleu,
-    calculate_rouge,
+    build_compute_metrics_fn,
+    freeze_embeds,
     freeze_params,
     lmap,
-    trim_batch,
+    save_json,
     use_task_specific_params,
+    write_txt_file,
 )
 
 
 logger = logging.getLogger(__name__)
 
 
-class Seq2SeqDataCollator:
-    def __init__(self, tokenizer, data_args, tpu_num_cores=None):
-        self.tokenizer = tokenizer
-        self.pad_token_id = tokenizer.pad_token_id
-        self.data_args = data_args
-        self.tpu_num_cores = tpu_num_cores
-        self.add_prefix_space = isinstance(tokenizer, BartTokenizer)
-
-    def __call__(self, batch) -> Dict[str, torch.Tensor]:
-        if hasattr(self.tokenizer, "prepare_seq2seq_batch"):
-            batch = self._encode(batch)
-            input_ids, attention_mask, labels = (
-                batch["input_ids"],
-                batch["attention_mask"],
-                batch["labels"],
-            )
-        else:
-            input_ids = torch.stack([x["input_ids"] for x in batch])
-            attention_mask = torch.stack([x["attention_mask"] for x in batch])
-            labels = torch.stack([x["labels"] for x in batch])
-
-            labels = trim_batch(labels, self.pad_token_id)
-            input_ids, attention_mask = trim_batch(input_ids, self.pad_token_id, attention_mask=attention_mask)
-
-        if isinstance(self.tokenizer, T5Tokenizer):
-            decoder_input_ids = self._shift_right_t5(labels)
-            labels = labels
-        else:
-            decoder_input_ids = shift_tokens_right(labels, self.pad_token_id)
-            labels = labels
-
-        batch = {
-            "input_ids": input_ids,
-            "attention_mask": attention_mask,
-            "decoder_input_ids": decoder_input_ids,
-            "labels": labels,
-        }
-        return batch
-
-    def _shift_right_t5(self, input_ids):
-        decoder_start_token_id = self.pad_token_id
-
-        assert (
-            decoder_start_token_id is not None
-        ), "self.model.config.decoder_start_token_id has to be defined. In T5 it is usually set to the pad_token_id. See T5 docs for more information"
-
-        # shift inputs to the right
-        shifted_input_ids = input_ids.new_zeros(input_ids.shape)
-        shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
-        shifted_input_ids[..., 0] = decoder_start_token_id
-
-        return shifted_input_ids
-
-    def _encode(self, batch) -> Dict[str, torch.Tensor]:
-        batch_encoding = self.tokenizer.prepare_seq2seq_batch(
-            [x["src_texts"] for x in batch],
-            src_lang=self.data_args.src_lang,
-            tgt_texts=[x["tgt_texts"] for x in batch],
-            tgt_lang=self.data_args.tgt_lang,
-            max_length=self.data_args.max_source_length,
-            max_target_length=self.data_args.max_target_length,
-            padding="max_length" if self.tpu_num_cores is not None else "longest",  # TPU hack
-            return_tensors="pt",
-            add_prefix_space=self.add_prefix_space,
-        )
-        return batch_encoding.data
-
-
 @dataclass
 class Seq2SeqTrainingArguments(TrainingArguments):
     """
@@ -125,6 +52,17 @@ class Seq2SeqTrainingArguments(TrainingArguments):
     predict_with_generate: bool = field(
         default=False, metadata={"help": "Whether to use generate to calculate generative metrics (ROUGE, BLEU)."}
     )
+    adafactor: bool = field(default=False, metadata={"help": "whether to use adafactor"})
+    encoder_layerdrop: Optional[float] = field(
+        default=None, metadata={"help": "Encoder layer dropout probability. Goes into model.config."}
+    )
+    decoder_layerdrop: Optional[float] = field(
+        default=None, metadata={"help": "Decoder layer dropout probability. Goes into model.config."}
+    )
+    dropout: Optional[float] = field(default=None, metadata={"help": "Dropout probability. Goes into model.config."})
+    attention_dropout: Optional[float] = field(
+        default=None, metadata={"help": "Attention dropout probability. Goes into model.config."}
+    )
 
 
 @dataclass
@@ -251,6 +189,13 @@ def main():
         model_args.config_name if model_args.config_name else model_args.model_name_or_path,
         cache_dir=model_args.cache_dir,
     )
+
+    extra_model_params = ("encoder_layerdrop", "decoder_layerdrop", "dropout", "attention_dropout")
+    for p in extra_model_params:
+        if getattr(training_args, p, None):
+            assert hasattr(config, p), f"({config.__class__.__name__}) doesn't have a `{p}` attribute"
+            setattr(config, p, getattr(training_args, p))
+
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
         cache_dir=model_args.cache_dir,
@@ -266,57 +211,15 @@ def main():
     use_task_specific_params(model, data_args.task)
 
     # set num_beams for evaluation
-    if data_args.eval_beams is not None:
-        model.config.num_beams = data_args.eval_beams
-    assert model.config.num_beams >= 1, f"got eval_beams={model.config.num_beams}. Need an integer >= 1"
-
-    # set max length for generation
-    model.config.max_generate_length = data_args.val_max_target_length
+    if data_args.eval_beams is None:
+        data_args.eval_beams = model.config.num_beams
 
     # set decoder_start_token_id for MBart
     if model.config.decoder_start_token_id is None and isinstance(tokenizer, MBartTokenizer):
-        decoder_start_token_id = tokenizer.lang_code_to_id[data_args.tgt_lang]
-        model.config.decoder_start_token_id = decoder_start_token_id
-
-    def build_compute_metrics_fn(task_name: str) -> Callable[[EvalPrediction], Dict]:
-        def non_pad_len(tokens: np.ndarray) -> int:
-            return np.count_nonzero(tokens != tokenizer.pad_token_id)
-
-        def decode_pred(pred: EvalPrediction) -> Tuple[List[str], List[str]]:
-            pred_str = tokenizer.batch_decode(pred.predictions, skip_special_tokens=True)
-            label_str = tokenizer.batch_decode(pred.label_ids, skip_special_tokens=True)
-            pred_str = lmap(str.strip, pred_str)
-            label_str = lmap(str.strip, label_str)
-            return pred_str, label_str
-
-        def summarization_metrics(pred: EvalPrediction) -> Dict:
-            pred_str, label_str = decode_pred(pred)
-            rouge: Dict = calculate_rouge(pred_str, label_str)
-            summ_len = np.mean(lmap(non_pad_len, pred.predictions))
-            rouge.update({"gen_len": summ_len})
-            return rouge
-
-        def translation_metrics(pred: EvalPrediction) -> Dict:
-            pred_str, label_str = decode_pred(pred)
-            bleu: Dict = calculate_bleu(pred_str, label_str)
-            gen_len = np.mean(lmap(non_pad_len, pred.predictions))
-            bleu.update({"gen_len": gen_len})
-            return bleu
-
-        compute_metrics_fn = summarization_metrics if "summarization" in task_name else translation_metrics
-        return compute_metrics_fn
-
-    def freeze_embeds(model: torch.nn.Module):
-        """Freeze token embeddings and positional embeddings for bart, just token embeddings for t5."""
-        try:
-            freeze_params(model.model.shared)
-            for d in [model.model.encoder, model.model.decoder]:
-                freeze_params(d.embed_positions)
-                freeze_params(d.embed_tokens)
-        except AttributeError:
-            freeze_params(model.shared)
-            for d in [model.encoder, model.decoder]:
-                freeze_params(d.embed_tokens)
+        assert (
+            data_args.tgt_lang is not None and data_args.src_lang is not None
+        ), "mBart requires --tgt_lang and --src_lang"
+        model.config.decoder_start_token_id = tokenizer.lang_code_to_id[data_args.tgt_lang]
 
     if model_args.freeze_embeds:
         freeze_embeds(model)
@@ -350,7 +253,7 @@ def main():
             max_source_length=data_args.max_source_length,
             prefix=model.config.prefix or "",
         )
-        if training_args.do_eval
+        if training_args.do_eval or training_args.evaluation_strategy != EvaluationStrategy.NO
         else None
     )
     test_dataset = (
@@ -368,13 +271,18 @@ def main():
     )
 
     # Initialize our Trainer
+    compute_metrics_fn = (
+        build_compute_metrics_fn(data_args.task, tokenizer) if training_args.predict_with_generate else None
+    )
     trainer = Seq2SeqTrainer(
         model=model,
+        config=config,
         args=training_args,
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,
         data_collator=Seq2SeqDataCollator(tokenizer, data_args, training_args.tpu_num_cores),
-        compute_metrics=build_compute_metrics_fn(data_args.task) if training_args.predict_with_generate else None,
+        compute_metrics=compute_metrics_fn,
+        data_args=data_args,
     )
 
     # Training
@@ -386,6 +294,7 @@ def main():
         # For convenience, we also re-save the tokenizer to the same directory,
         # so that you can share your model easily on huggingface.co/models =)
         if trainer.is_world_process_zero():
+            trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
             tokenizer.save_pretrained(training_args.output_dir)
 
     # Evaluation
@@ -395,41 +304,36 @@ def main():
 
         result = trainer.evaluate()
 
-        output_eval_file = os.path.join(training_args.output_dir, "eval_results.json")
         if trainer.is_world_process_zero():
             logger.info("***** Eval results *****")
             for key, value in result.items():
                 logger.info("  %s = %s", key, value)
-
-            with open(output_eval_file, "w") as f:
-                json.dump(result, f)
-
+            save_json(result, os.path.join(training_args.output_dir, "eval_results.json"))
             eval_results.update(result)
 
     if training_args.do_predict:
         logging.info("*** Test ***")
 
         test_output = trainer.predict(test_dataset=test_dataset)
-        test_metrics = test_output.metrics
-        test_metrics = {k.replace("eval", "test"): v for k, v in test_metrics.items()}
-
-        output_test_file = os.path.join(training_args.output_dir, "test_results.json")
+        test_metrics = {k.replace("eval", "test"): v for k, v in test_output.metrics.items()}
 
         if trainer.is_world_process_zero():
             logger.info("***** Test results *****")
             for key, value in test_metrics.items():
                 logger.info("  %s = %s", key, value)
 
-            with open(output_test_file, "w") as f:
-                json.dump(test_metrics, f)
+            save_json(test_metrics, os.path.join(training_args.output_dir, "test_results.json"))
+            eval_results.update(test_metrics)
 
             if training_args.predict_with_generate:
-                test_preds = tokenizer.batch_decode(test_output.predictions, skip_special_tokens=True)
+                test_preds = tokenizer.batch_decode(
+                    test_output.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
+                )
                 test_preds = lmap(str.strip, test_preds)
-                output_test_pred_file = os.path.join(training_args.output_dir, "test_generations.txt")
-                with open(output_test_pred_file, "w") as f:
-                    f.write("\n".join(test_preds))
+                write_txt_file(test_preds, os.path.join(training_args.output_dir, "test_generations.txt"))
 
+    if trainer.is_world_process_zero():
+        save_json(eval_results, "all_results.json")
     return eval_results
 
 

examples/seq2seq/make_student.py

@@ -32,7 +32,7 @@ LAYERS_TO_COPY = {
     },
     16: {  # maps  num layers in student -> which teacher layers to copy
         1: [0],
-        2: [0, 8],
+        2: [0, 15],
         3: [0, 8, 15],
         4: [0, 5, 10, 15],
         6: [0, 3, 6, 9, 12, 15],

examples/seq2seq/rouge_cli.py

@@ -9,7 +9,7 @@ def calculate_rouge_path(pred_path, tgt_path, save_path=None, **kwargs):
     tgt_lns = [x.strip() for x in open(tgt_path).readlines()][: len(pred_lns)]
     metrics = calculate_rouge(pred_lns, tgt_lns, **kwargs)
     if save_path is not None:
-        save_json(metrics, save_path)
+        save_json(metrics, save_path, indent=None)
     return metrics  # these print nicely
 
 

examples/seq2seq/run_distributed_eval.py

@@ -42,8 +42,7 @@ def eval_data_dir(
     task="summarization",
     local_rank=None,
     num_return_sequences=1,
-    src_lang=None,
-    tgt_lang=None,
+    dataset_kwargs: Dict = None,
     prefix="",
     **generate_kwargs,
 ) -> Dict:
@@ -78,9 +77,8 @@ def eval_data_dir(
         max_target_length=1024,
         type_path=type_path,
         n_obs=n_obs,
-        src_lang=src_lang,
-        tgt_lang=tgt_lang,
         prefix=prefix,
+        **dataset_kwargs,
     )
     # I set shuffle=True for a more accurate progress bar.
     # If all the longest samples are first, the prog bar estimate is too high at the beginning.
@@ -158,6 +156,11 @@ def run_generate():
     if intermediate_files:
         raise ValueError(f"Found files at {json_save_dir} please move or remove them.")
         # In theory, a node could finish and save before another node hits this. If this happens, we can address later.
+    dataset_kwargs = {}
+    if args.src_lang is not None:
+        dataset_kwargs["src_lang"] = args.src_lang
+    if args.tgt_lang is not None:
+        dataset_kwargs["tgt_lang"] = args.tgt_lang
 
     Path(args.save_dir).mkdir(exist_ok=True)
     results, num_replicas = eval_data_dir(
@@ -173,8 +176,7 @@ def run_generate():
         max_source_length=args.max_source_length,
         num_return_sequences=args.num_return_sequences,
         prefix=args.prefix,
-        src_lang=args.src_lang,
-        tgt_lang=args.tgt_lang,
+        dataset_kwargs=dataset_kwargs,
         **generate_kwargs,
     )
 

examples/seq2seq/run_eval.py

@@ -152,8 +152,7 @@ def run_generate(verbose=True):
         print(scores)
 
     if args.score_path is not None:
-        path = args.score_path
-        json.dump(scores, open(path, "w"))
+        json.dump(scores, open(args.score_path, "w"))
 
     return scores
 

examples/seq2seq/sentence_splitter.py

@@ -1,5 +1,7 @@
 import re
 
+from filelock import FileLock
+
 
 try:
     import nltk
@@ -9,13 +11,12 @@ except (ImportError, ModuleNotFoundError):
     NLTK_AVAILABLE = False
 
 if NLTK_AVAILABLE:
-    try:
+    with FileLock(".lock") as lock:
         nltk.download("punkt", quiet=True)
-    except FileExistsError:  # multiprocessing race condition
-        pass
 
 
 def add_newline_to_end_of_each_sentence(x: str) -> str:
+    """This was added to get rougeLsum scores matching published rougeL scores for BART and PEGASUS."""
     re.sub("<n>", "", x)  # remove pegasus newline char
-    assert NLTK_AVAILABLE, "nltk must be installed to separate newlines betwee sentences. (pip install nltk)"
+    assert NLTK_AVAILABLE, "nltk must be installed to separate newlines between sentences. (pip install nltk)"
     return "\n".join(nltk.sent_tokenize(x))

examples/seq2seq/seq2seq_trainer.py

@@ -6,7 +6,9 @@ from torch import nn
 from torch.utils.data import DistributedSampler, RandomSampler
 
 from transformers import Trainer
+from transformers.configuration_fsmt import FSMTConfig
 from transformers.file_utils import is_torch_tpu_available
+from transformers.optimization import Adafactor, AdamW, get_linear_schedule_with_warmup
 from transformers.trainer import get_tpu_sampler
 
 
@@ -20,6 +22,50 @@ logger = logging.getLogger(__name__)
 
 
 class Seq2SeqTrainer(Trainer):
+    def __init__(self, config, data_args, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.config = config
+        self.data_args = data_args
+        self.max_gen_length = data_args.val_max_target_length
+        self.vocab_size = self.config.tgt_vocab_size if isinstance(self.config, FSMTConfig) else self.config.vocab_size
+
+    def create_optimizer_and_scheduler(self, num_training_steps: int):
+        """
+        Setup the optimizer and the learning rate scheduler.
+
+        We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the
+        Trainer's init through :obj:`optimizers`, or subclass and override this method in a subclass.
+        """
+        if self.optimizer is None:
+            no_decay = ["bias", "LayerNorm.weight"]
+            optimizer_grouped_parameters = [
+                {
+                    "params": [p for n, p in self.model.named_parameters() if not any(nd in n for nd in no_decay)],
+                    "weight_decay": self.args.weight_decay,
+                },
+                {
+                    "params": [p for n, p in self.model.named_parameters() if any(nd in n for nd in no_decay)],
+                    "weight_decay": 0.0,
+                },
+            ]
+            if self.args.adafactor:
+                self.optimizer = Adafactor(
+                    optimizer_grouped_parameters,
+                    lr=self.args.learning_rate,
+                    scale_parameter=False,
+                    relative_step=False,
+                )
+
+            else:
+                self.optimizer = AdamW(
+                    optimizer_grouped_parameters, lr=self.args.learning_rate, eps=self.args.adam_epsilon
+                )
+
+        if self.lr_scheduler is None:
+            self.lr_scheduler = get_linear_schedule_with_warmup(
+                self.optimizer, num_warmup_steps=self.args.warmup_steps, num_training_steps=num_training_steps
+            )
+
     def _get_train_sampler(self) -> Optional[torch.utils.data.sampler.Sampler]:
         if isinstance(self.train_dataset, torch.utils.data.IterableDataset):
             return None
@@ -41,18 +87,18 @@ class Seq2SeqTrainer(Trainer):
         labels = inputs.pop("labels")
         outputs = model(**inputs, use_cache=False)
         logits = outputs[0]
-        return self._compute_loss(logits, labels, ignore_index=model.config.pad_token_id)
+        return self._compute_loss(logits, labels)
 
-    def _compute_loss(self, logits, labels, ignore_index):
+    def _compute_loss(self, logits, labels):
         if self.args.label_smoothing == 0:
             # Same behavior as modeling_bart.py
-            loss_fct = torch.nn.CrossEntropyLoss(ignore_index=ignore_index)
-            assert logits.shape[-1] == self.model.config.vocab_size
+            loss_fct = torch.nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)
+            assert logits.shape[-1] == self.vocab_size
             loss = loss_fct(logits.view(-1, logits.shape[-1]), labels.view(-1))
         else:
             lprobs = torch.nn.functional.log_softmax(logits, dim=-1)
             loss, nll_loss = label_smoothed_nll_loss(
-                lprobs, labels, self.args.label_smoothing, ignore_index=ignore_index
+                lprobs, labels, self.args.label_smoothing, ignore_index=self.config.pad_token_id
             )
         return loss
 
@@ -81,45 +127,34 @@ class Seq2SeqTrainer(Trainer):
         """
         inputs = self._prepare_inputs(inputs)
 
-        max_length = (
-            model.config.max_generate_length
-            if hasattr(model.config, "max_generate_length")
-            else model.config.max_position_embeddings
-        )
-
         with torch.no_grad():
             if self.args.predict_with_generate and not self.args.prediction_loss_only:
                 generated_tokens = model.generate(
                     inputs["input_ids"],
                     attention_mask=inputs["attention_mask"],
                     use_cache=True,
-                    num_beams=model.config.num_beams,
-                    max_length=max_length,
+                    num_beams=self.data_args.eval_beams,
+                    max_length=self.max_gen_length,
                 )
                 # in case the batch is shorter than max length, the output should be padded
-                generated_tokens = self._pad_tensors_to_max_len(
-                    generated_tokens, max_length, model.config.pad_token_id
-                )
+                generated_tokens = self._pad_tensors_to_max_len(generated_tokens, self.max_gen_length)
 
             labels_out = inputs.get("labels")
-            outputs = model(**inputs)
-            logits = outputs[1]
-            loss = self._compute_loss(logits, labels_out, model.config.pad_token_id)
+            # Call forward again to get loss # TODO: avoidable?
+            outputs = model(**inputs, use_cache=False)
+            loss = self._compute_loss(outputs[1], labels_out)
             loss = loss.mean().item()
             if self.args.prediction_loss_only:
-                logits = None
-            else:
-                logits = generated_tokens if self.args.predict_with_generate else logits
+                return (loss, None, None)
 
-        if self.args.prediction_loss_only:
-            return (loss, None, None)
+            logits = generated_tokens if self.args.predict_with_generate else outputs[1]
 
         labels_out = labels_out.detach()
-        labels = self._pad_tensors_to_max_len(labels_out, max_length, model.config.pad_token_id)
+        labels = self._pad_tensors_to_max_len(labels_out, self.max_gen_length)
         return (loss, logits.detach(), labels)
 
-    def _pad_tensors_to_max_len(self, tensor, max_length, pad_token_id):
-        padded_tensor = pad_token_id * torch.ones(
+    def _pad_tensors_to_max_len(self, tensor, max_length):
+        padded_tensor = self.config.pad_token_id * torch.ones(
             (tensor.shape[0], max_length), dtype=tensor.dtype, device=tensor.device
         )
         padded_tensor[:, : tensor.shape[-1]] = tensor

examples/seq2seq/test_datasets.py

@@ -185,3 +185,36 @@ def test_distributed_sortish_sampler_splits_indices_between_procs():
     ids1 = set(DistributedSortishSampler(ds, 256, num_replicas=2, rank=0, add_extra_examples=False))
     ids2 = set(DistributedSortishSampler(ds, 256, num_replicas=2, rank=1, add_extra_examples=False))
     assert ids1.intersection(ids2) == set()
+
+
+@pytest.mark.parametrize(
+    "tok_name",
+    [
+        MBART_TINY,
+        MARIAN_TINY,
+        T5_TINY,
+        BART_TINY,
+        PEGASUS_XSUM,
+    ],
+)
+def test_dataset_kwargs(tok_name):
+    tokenizer = AutoTokenizer.from_pretrained(tok_name)
+    if tok_name == MBART_TINY:
+        train_dataset = Seq2SeqDataset(
+            tokenizer,
+            data_dir=make_test_data_dir(),
+            type_path="train",
+            max_source_length=4,
+            max_target_length=8,
+            src_lang="EN",
+            tgt_lang="FR",
+        )
+        kwargs = train_dataset.dataset_kwargs
+        assert "src_lang" in kwargs and "tgt_lang" in kwargs
+    else:
+        train_dataset = Seq2SeqDataset(
+            tokenizer, data_dir=make_test_data_dir(), type_path="train", max_source_length=4, max_target_length=8
+        )
+        kwargs = train_dataset.dataset_kwargs
+        assert "add_prefix_space" not in kwargs if tok_name != BART_TINY else "add_prefix_space" in kwargs
+        assert len(kwargs) == 1 if tok_name == BART_TINY else len(kwargs) == 0

examples/seq2seq/test_finetune_trainer.py

@@ -3,36 +3,53 @@ import sys
 import tempfile
 from unittest.mock import patch
 
-from transformers import BartForConditionalGeneration, MarianMTModel
 from transformers.testing_utils import slow
+from transformers.trainer_utils import TrainerState, set_seed
 
 from .finetune_trainer import main
 from .test_seq2seq_examples import MBART_TINY
-from .utils import load_json
 
 
-MODEL_NAME = MBART_TINY
-# TODO(SS): MODEL_NAME = "sshleifer/student_mbart_en_ro_1_1"
+set_seed(42)
 MARIAN_MODEL = "sshleifer/student_marian_en_ro_6_1"
 
 
-@slow
-def test_model_download():
-    """This warms up the cache so that we can time the next test without including download time, which varies between machines."""
-    BartForConditionalGeneration.from_pretrained(MODEL_NAME)
-    MarianMTModel.from_pretrained(MARIAN_MODEL)
-
-
-@slow
 def test_finetune_trainer():
+    output_dir = run_trainer(1, "12", MBART_TINY, 1)
+    logs = TrainerState.load_from_json(os.path.join(output_dir, "trainer_state.json")).log_history
+    eval_metrics = [log for log in logs if "eval_loss" in log.keys()]
+    first_step_stats = eval_metrics[0]
+    assert "eval_bleu" in first_step_stats
+
+
+@slow
+def test_finetune_trainer_slow():
+    # TODO(SS): This will fail on devices with more than 1 GPU.
+    # There is a missing call to __init__process_group somewhere
+    output_dir = run_trainer(eval_steps=2, max_len="128", model_name=MARIAN_MODEL, num_train_epochs=3)
+
+    # Check metrics
+    logs = TrainerState.load_from_json(os.path.join(output_dir, "trainer_state.json")).log_history
+    eval_metrics = [log for log in logs if "eval_loss" in log.keys()]
+    first_step_stats = eval_metrics[0]
+    last_step_stats = eval_metrics[-1]
+
+    assert first_step_stats["eval_bleu"] < last_step_stats["eval_bleu"]  # model learned nothing
+    assert isinstance(last_step_stats["eval_bleu"], float)
+
+    # test if do_predict saves generations and metrics
+    contents = os.listdir(output_dir)
+    contents = {os.path.basename(p) for p in contents}
+    assert "test_generations.txt" in contents
+    assert "test_results.json" in contents
+
+
+def run_trainer(eval_steps: int, max_len: str, model_name: str, num_train_epochs: int):
     data_dir = "examples/seq2seq/test_data/wmt_en_ro"
-    output_dir = tempfile.mkdtemp(prefix="marian_output")
-    max_len = "128"
-    num_train_epochs = 4
-    eval_steps = 2
+    output_dir = tempfile.mkdtemp(prefix="test_output")
     argv = [
         "--model_name_or_path",
-        MARIAN_MODEL,
+        model_name,
         "--data_dir",
         data_dir,
         "--output_dir",
@@ -72,25 +89,18 @@ def test_finetune_trainer():
         "--sortish_sampler",
         "--label_smoothing",
         "0.1",
+        # "--eval_beams",
+        # "2",
+        "--adafactor",
         "--task",
         "translation",
+        "--tgt_lang",
+        "ro_RO",
+        "--src_lang",
+        "en_XX",
     ]
-
     testargs = ["finetune_trainer.py"] + argv
     with patch.object(sys, "argv", testargs):
         main()
 
-    # Check metrics
-    logs = load_json(os.path.join(output_dir, "log_history.json"))
-    eval_metrics = [log for log in logs if "eval_loss" in log.keys()]
-    first_step_stats = eval_metrics[0]
-    last_step_stats = eval_metrics[-1]
-
-    assert first_step_stats["eval_bleu"] < last_step_stats["eval_bleu"]  # model learned nothing
-    assert isinstance(last_step_stats["eval_bleu"], float)
-
-    # test if do_predict saves generations and metrics
-    contents = os.listdir(output_dir)
-    contents = {os.path.basename(p) for p in contents}
-    assert "test_generations.txt" in contents
-    assert "test_results.json" in contents
+    return output_dir

examples/seq2seq/utils.py

@@ -7,7 +7,7 @@ import pickle
 import socket
 from logging import getLogger
 from pathlib import Path
-from typing import Callable, Dict, Iterable, List, Union
+from typing import Callable, Dict, Iterable, List, Tuple, Union
 
 import git
 import numpy as np
@@ -19,8 +19,9 @@ from torch import nn
 from torch.utils.data import Dataset, Sampler
 
 from sentence_splitter import add_newline_to_end_of_each_sentence
-from transformers import BartTokenizer
+from transformers import BartTokenizer, EvalPrediction, PreTrainedTokenizer, T5Tokenizer
 from transformers.file_utils import cached_property
+from transformers.modeling_bart import shift_tokens_right
 
 
 try:
@@ -52,19 +53,6 @@ def label_smoothed_nll_loss(lprobs, target, epsilon, ignore_index=-100):
     return loss, nll_loss
 
 
-def encode_line(tokenizer, line, max_length, pad_to_max_length=True, return_tensors="pt"):
-    """Only used by LegacyDataset"""
-    extra_kw = {"add_prefix_space": True} if isinstance(tokenizer, BartTokenizer) else {}
-    return tokenizer(
-        [line],
-        max_length=max_length,
-        padding="max_length" if pad_to_max_length else None,
-        truncation=True,
-        return_tensors=return_tensors,
-        **extra_kw,
-    )
-
-
 def lmap(f: Callable, x: Iterable) -> List:
     """list(map(f, x))"""
     return list(map(f, x))
@@ -75,6 +63,35 @@ def calculate_bleu(output_lns, refs_lns, **kwargs) -> dict:
     return {"bleu": round(corpus_bleu(output_lns, [refs_lns], **kwargs).score, 4)}
 
 
+def build_compute_metrics_fn(task_name: str, tokenizer: PreTrainedTokenizer) -> Callable[[EvalPrediction], Dict]:
+    def non_pad_len(tokens: np.ndarray) -> int:
+        return np.count_nonzero(tokens != tokenizer.pad_token_id)
+
+    def decode_pred(pred: EvalPrediction) -> Tuple[List[str], List[str]]:
+        pred_str = tokenizer.batch_decode(pred.predictions, skip_special_tokens=True)
+        label_str = tokenizer.batch_decode(pred.label_ids, skip_special_tokens=True)
+        pred_str = lmap(str.strip, pred_str)
+        label_str = lmap(str.strip, label_str)
+        return pred_str, label_str
+
+    def summarization_metrics(pred: EvalPrediction) -> Dict:
+        pred_str, label_str = decode_pred(pred)
+        rouge: Dict = calculate_rouge(pred_str, label_str)
+        summ_len = np.round(np.mean(lmap(non_pad_len, pred.predictions)), 1)
+        rouge.update({"gen_len": summ_len})
+        return rouge
+
+    def translation_metrics(pred: EvalPrediction) -> Dict:
+        pred_str, label_str = decode_pred(pred)
+        bleu: Dict = calculate_bleu(pred_str, label_str)
+        gen_len = np.round(np.mean(lmap(non_pad_len, pred.predictions)), 1)
+        bleu.update({"gen_len": gen_len})
+        return bleu
+
+    compute_metrics_fn = summarization_metrics if "summarization" in task_name else translation_metrics
+    return compute_metrics_fn
+
+
 def trim_batch(
     input_ids,
     pad_token_id,
@@ -97,9 +114,8 @@ class AbstractSeq2SeqDataset(Dataset):
         max_target_length,
         type_path="train",
         n_obs=None,
-        src_lang=None,
-        tgt_lang=None,
         prefix="",
+        **dataset_kwargs
     ):
         super().__init__()
         self.src_file = Path(data_dir).joinpath(type_path + ".source")
@@ -120,9 +136,8 @@ class AbstractSeq2SeqDataset(Dataset):
         if n_obs is not None:
             self.src_lens = self.src_lens[:n_obs]
         self.pad_token_id = self.tokenizer.pad_token_id
-        self.src_lang = src_lang
-        self.tgt_lang = tgt_lang
-        self.add_prefix_space = isinstance(self.tokenizer, BartTokenizer)
+        self.dataset_kwargs = dataset_kwargs
+        dataset_kwargs.update({"add_prefix_space": True} if isinstance(self.tokenizer, BartTokenizer) else {})
 
     def __len__(self):
         return len(self.src_lens)
@@ -182,8 +197,8 @@ class LegacySeq2SeqDataset(AbstractSeq2SeqDataset):
         tgt_line = linecache.getline(str(self.tgt_file), index).rstrip("\n")
         assert source_line, f"empty source line for index {index}"
         assert tgt_line, f"empty tgt line for index {index}"
-        source_inputs = encode_line(self.tokenizer, source_line, self.max_source_length)
-        target_inputs = encode_line(self.tokenizer, tgt_line, self.max_target_length)
+        source_inputs = self.encode_line(self.tokenizer, source_line, self.max_source_length)
+        target_inputs = self.encode_line(self.tokenizer, tgt_line, self.max_target_length)
 
         source_ids = source_inputs["input_ids"].squeeze()
         target_ids = target_inputs["input_ids"].squeeze()
@@ -194,6 +209,17 @@ class LegacySeq2SeqDataset(AbstractSeq2SeqDataset):
             "labels": target_ids,
         }
 
+    def encode_line(self, tokenizer, line, max_length, pad_to_max_length=True, return_tensors="pt"):
+        """Only used by LegacyDataset"""
+        return tokenizer(
+            [line],
+            max_length=max_length,
+            padding="max_length" if pad_to_max_length else None,
+            truncation=True,
+            return_tensors=return_tensors,
+            **self.dataset_kwargs,
+        )
+
     def collate_fn(self, batch) -> Dict[str, torch.Tensor]:
         input_ids = torch.stack([x["input_ids"] for x in batch])
         masks = torch.stack([x["attention_mask"] for x in batch])
@@ -224,18 +250,80 @@ class Seq2SeqDataset(AbstractSeq2SeqDataset):
         """Call prepare_seq2seq_batch."""
         batch_encoding: Dict[str, torch.Tensor] = self.tokenizer.prepare_seq2seq_batch(
             [x["src_texts"] for x in batch],
-            src_lang=self.src_lang,
             tgt_texts=[x["tgt_texts"] for x in batch],
-            tgt_lang=self.tgt_lang,
             max_length=self.max_source_length,
             max_target_length=self.max_target_length,
             return_tensors="pt",
-            add_prefix_space=self.add_prefix_space,
+            **self.dataset_kwargs,
         ).data
         batch_encoding["ids"] = torch.tensor([x["id"] for x in batch])
         return batch_encoding
 
 
+class Seq2SeqDataCollator:
+    def __init__(self, tokenizer, data_args, tpu_num_cores=None):
+        self.tokenizer = tokenizer
+        self.pad_token_id = tokenizer.pad_token_id
+        assert (
+            self.pad_token_id is not None
+        ), f"pad_token_id is not defined for ({self.tokenizer.__class__.__name__}), it must be defined."
+        self.data_args = data_args
+        self.tpu_num_cores = tpu_num_cores
+        self.dataset_kwargs = {"add_prefix_space": isinstance(tokenizer, BartTokenizer)}
+        if data_args.src_lang is not None:
+            self.dataset_kwargs["src_lang"] = data_args.src_lang
+        if data_args.tgt_lang is not None:
+            self.dataset_kwargs["tgt_lang"] = data_args.tgt_lang
+
+    def __call__(self, batch) -> Dict[str, torch.Tensor]:
+        if hasattr(self.tokenizer, "prepare_seq2seq_batch"):
+            batch = self._encode(batch)
+            input_ids, attention_mask, labels = (
+                batch["input_ids"],
+                batch["attention_mask"],
+                batch["labels"],
+            )
+        else:
+            input_ids = torch.stack([x["input_ids"] for x in batch])
+            attention_mask = torch.stack([x["attention_mask"] for x in batch])
+            labels = torch.stack([x["labels"] for x in batch])
+
+            labels = trim_batch(labels, self.pad_token_id)
+            input_ids, attention_mask = trim_batch(input_ids, self.pad_token_id, attention_mask=attention_mask)
+
+        if isinstance(self.tokenizer, T5Tokenizer):
+            decoder_input_ids = self._shift_right_t5(labels)
+        else:
+            decoder_input_ids = shift_tokens_right(labels, self.pad_token_id)
+
+        batch = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "decoder_input_ids": decoder_input_ids,
+            "labels": labels,
+        }
+        return batch
+
+    def _shift_right_t5(self, input_ids):
+        # shift inputs to the right
+        shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+        shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
+        shifted_input_ids[..., 0] = self.pad_token_id
+        return shifted_input_ids
+
+    def _encode(self, batch) -> Dict[str, torch.Tensor]:
+        batch_encoding = self.tokenizer.prepare_seq2seq_batch(
+            [x["src_texts"] for x in batch],
+            tgt_texts=[x["tgt_texts"] for x in batch],
+            max_length=self.data_args.max_source_length,
+            max_target_length=self.data_args.max_target_length,
+            padding="max_length" if self.tpu_num_cores is not None else "longest",  # TPU hack
+            return_tensors="pt",
+            **self.dataset_kwargs,
+        )
+        return batch_encoding.data
+
+
 class SortishSampler(Sampler):
     "Go through the text data by order of src length with a bit of randomness. From fastai repo."
 
@@ -447,6 +535,25 @@ def freeze_params(model: nn.Module):
         par.requires_grad = False
 
 
+def freeze_embeds(model):
+    """Freeze token embeddings and positional embeddings for bart, just token embeddings for t5."""
+    model_type = model.config.model_type
+
+    if model_type == "t5":
+        freeze_params(model.shared)
+        for d in [model.encoder, model.decoder]:
+            freeze_params(d.embed_tokens)
+    elif model_type == "fsmt":
+        for d in [model.model.encoder, model.model.decoder]:
+            freeze_params(d.embed_positions)
+            freeze_params(d.embed_tokens)
+    else:
+        freeze_params(model.model.shared)
+        for d in [model.model.encoder, model.model.decoder]:
+            freeze_params(d.embed_positions)
+            freeze_params(d.embed_tokens)
+
+
 def grad_status(model: nn.Module) -> Iterable:
     return (par.requires_grad for par in model.parameters())
 

model_cards/TypicaAI/magbert-ner/README.md

@@ -0,0 +1,55 @@
+---
+language: fr
+widget:
+- text: "Je m'appelle Hicham et je vis a Fès"
+---
+
+# MagBERT-NER: a state-of-the-art NER model for Moroccan French language (Maghreb)
+
+## Introduction
+
+[MagBERT-NER] is a state-of-the-art NER model for Moroccan French language (Maghreb). The MagBERT-NER model was fine-tuned for NER Task based the language model for French Camembert (based on the RoBERTa architecture).
+
+For further information or requests, please go to [Typica.AI Website](https://typicasoft.io/)
+
+## How to use MagBERT-NER with HuggingFace
+
+##### Load MagBERT-NER and its sub-word tokenizer :
+```python
+from transformers import AutoTokenizer, AutoModelForTokenClassification
+
+tokenizer = AutoTokenizer.from_pretrained("TypicaAI/magbert-ner")
+model = AutoModelForTokenClassification.from_pretrained("TypicaAI/magbert-ner")
+
+
+##### Process text sample (from wikipedia about the current Prime Minister of Morocco) Using NER pipeline  
+
+from transformers import pipeline
+
+nlp = pipeline('ner', model=model, tokenizer=tokenizer, grouped_entities=True)
+nlp("Saad Dine El Otmani, né le 16 janvier 1956 à Inezgane, est un homme d'État marocain, chef du gouvernement du Maroc depuis le 5 avril 2017")
+
+
+#[{'entity_group': 'I-PERSON',
+#  'score': 0.8941445276141167,
+#  'word': 'Saad Dine El Otmani'},
+# {'entity_group': 'B-DATE',
+#  'score': 0.5967703461647034,
+#  'word': '16 janvier 1956'},
+# {'entity_group': 'B-GPE', 'score': 0.7160899192094803, 'word': 'Inezgane'},
+# {'entity_group': 'B-NORP', 'score': 0.7971733212471008, 'word': 'marocain'},
+# {'entity_group': 'B-GPE', 'score': 0.8921478390693665, 'word': 'Maroc'},
+# {'entity_group': 'B-DATE',
+#  'score': 0.5760444005330404,
+#  'word': '5 avril 2017'}]
+
+```
+
+```
+
+
+## Authors 
+
+MagBert-NER was trained and evaluated by Hicham Assoudi, Ph.D.
+
+

model_cards/adalbertojunior/PTT5-SMALL-SUM/README.md

@@ -0,0 +1,37 @@
+---
+language: pt
+---
+
+# PTT5-SMALL-SUM
+
+## Model description
+
+This model was trained to summarize texts in portuguese
+
+
+based on ```unicamp-dl/ptt5-small-portuguese-vocab```
+
+#### How to use
+
+```python
+from transformers import T5Tokenizer, T5ForConditionalGeneration
+
+tokenizer = T5Tokenizer.from_pretrained('adalbertojunior/PTT5-SMALL-SUM')
+
+t5 = T5ForConditionalGeneration.from_pretrained('adalbertojunior/PTT5-SMALL-SUM')
+
+text="Esse é um exemplo de sumarização."
+
+input_ids = tokenizer.encode(text, return_tensors="pt", add_special_tokens=True)
+
+generated_ids = t5.generate(
+        input_ids=input_ids,
+        num_beams=1,
+        max_length=40,
+        #repetition_penalty=2.5
+    ).squeeze()
+    
+predicted_span = tokenizer.decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
+
+
+```

model_cards/ahotrod/albert_xxlargev1_squad2_512/README.md

@@ -1,71 +1,60 @@
 ## Albert xxlarge version 1 language model fine-tuned on SQuAD2.0
 
-### with the following results:
+###  (updated 30Sept2020) with the following results:
 
 ```
-exact: 85.65653162637918
-f1: 89.260458954177
+exact: 86.11134506864315
+f1: 89.35371214945009
 total': 11873
-HasAns_exact': 82.6417004048583
-HasAns_f1': 89.8598902096736
+HasAns_exact': 83.56950067476383
+HasAns_f1': 90.06353312254078
 HasAns_total': 5928
-NoAns_exact': 88.66274179983179
-NoAns_f1': 88.66274179983179
+NoAns_exact': 88.64592094196804
+NoAns_f1': 88.64592094196804
 NoAns_total': 5945
-best_exact': 85.65653162637918
+best_exact': 86.11134506864315
 best_exact_thresh': 0.0
-best_f1': 89.2604589541768
+best_f1': 89.35371214944985
 best_f1_thresh': 0.0
 ```
 
 ### from script:
 
 ```
-python -m torch.distributed.launch --nproc_per_node=2 ${RUN_SQUAD_DIR}/run_squad.py \
---model_type albert \
---model_name_or_path albert-xxlarge-v1 \
---do_train \
---train_file ${SQUAD_DIR}/train-v2.0.json \
---predict_file ${SQUAD_DIR}/dev-v2.0.json \
---version_2_with_negative \
---num_train_epochs 3 \
---max_steps 8144 \
---warmup_steps 814 \
---do_lower_case \
---learning_rate 3e-5 \
---max_seq_length 512 \
---doc_stride 128 \
---save_steps 2000 \
---per_gpu_train_batch_size 1 \
---gradient_accumulation_steps 24 \
---output_dir ${MODEL_PATH}
-
-CUDA_VISIBLE_DEVICES=0 python ${RUN_SQUAD_DIR}/run_squad.py \
---model_type albert \
---model_name_or_path ${MODEL_PATH} \
---do_eval \
---train_file ${SQUAD_DIR}/train-v2.0.json \
---predict_file ${SQUAD_DIR}/dev-v2.0.json \
---version_2_with_negative \
---do_lower_case \
---max_seq_length 512 \
---per_gpu_eval_batch_size 48 \
---output_dir ${MODEL_PATH}
+python ${EXAMPLES}/run_squad.py \
+  --model_type albert \
+  --model_name_or_path albert-xxlarge-v1 \
+  --do_train \
+  --do_eval \
+  --train_file ${SQUAD}/train-v2.0.json \
+  --predict_file ${SQUAD}/dev-v2.0.json \
+  --version_2_with_negative \
+  --do_lower_case \
+  --num_train_epochs 3 \
+  --max_steps 8144 \
+  --warmup_steps 814 \
+  --learning_rate 3e-5 \
+  --max_seq_length 512 \
+  --doc_stride 128 \
+  --per_gpu_train_batch_size 6 \
+  --gradient_accumulation_steps 8 \
+  --per_gpu_eval_batch_size 48 \
+  --fp16 \
+  --fp16_opt_level O1 \
+  --threads 12 \
+  --logging_steps 50 \
+  --save_steps 3000 \
+  --overwrite_output_dir \
+  --output_dir ${MODEL_PATH}
 ```
 
-### using the following system & software:
+### using the following software & system:
 
 ```
-OS/Platform: Linux-4.15.0-76-generic-x86_64-with-debian-buster-sid
-GPU/CPU: 2 x NVIDIA 1080Ti / Intel i7-8700
-Transformers: 2.3.0
-PyTorch: 1.4.0
-TensorFlow: 2.1.0
-Python: 3.7.6
+Transformers: 3.1.0
+PyTorch: 1.6.0
+TensorFlow: 2.3.1
+Python: 3.8.1
+OS: Linux-5.4.0-48-generic-x86_64-with-glibc2.10
+CPU/GPU: Intel i9-9900K / NVIDIA Titan RTX 24GB
 ```
-
-### Access this albert_xxlargev1_sqd2_512 fine-tuned model with:
-
-```python
-tokenizer = AutoTokenizer.from_pretrained("ahotrod/albert_xxlargev1_squad2_512")
-model = AutoModelForQuestionAnswering.from_pretrained("ahotrod/albert_xxlargev1_squad2_512")

model_cards/akhooli/personachat-arabic/README.md

@@ -0,0 +1,12 @@
+---
+tags:
+- conversational
+language:
+- ar
+license: mit
+---
+## personachat-arabic (conversational AI)
+This is personachat-arabic, using a subset from the persona-chat validation dataset, machine translated to Arabic (from English) 
+and fine-tuned from [akhooli/gpt2-small-arabic](https://huggingface.co/akhooli/gpt2-small-arabic) which is a limited text generation model.  
+Usage: see the last section of this [example notebook](https://colab.research.google.com/drive/1I6RFOWMaTpPBX7saJYjnSTddW0TD6H1t?usp=sharing) 
+Note: model has limited training set which was machine translated (do not use for production). 

model_cards/allenyummy/chinese-bert-wwm-ehr-ner-sl/README.md

@@ -0,0 +1,15 @@
+---
+language: zh-tw
+---
+
+# Model name
+Chinese-bert-wwm-electrical-health-record-ner-sequence-labeling
+
+
+#### How to use
+
+```
+from transformers import AutoTokenizer, AutoModelForTokenClassification  
+tokenizer = AutoTokenizer.from_pretrained("chinese-bert-wwm-ehr-ner-sl")  
+model = AutoModelForTokenClassification.from_pretrained("chinese-bert-wwm-ehr-ner-sl") 
+```

model_cards/bayartsogt/bert-base-mongolian-cased/README.md

@@ -0,0 +1,60 @@
+---
+language: "mn"
+tags:
+- mongolian
+- cased
+---
+
+# BERT-BASE-MONGOLIAN-CASED
+[Link to Official Mongolian-BERT repo](https://github.com/tugstugi/mongolian-bert)
+
+## Model description
+This repository contains pre-trained Mongolian [BERT](https://arxiv.org/abs/1810.04805) models trained by [tugstugi](https://github.com/tugstugi), [enod](https://github.com/enod) and [sharavsambuu](https://github.com/sharavsambuu).
+Special thanks to [nabar](https://github.com/nabar) who provided 5x TPUs.
+
+This repository is based on the following open source projects: [google-research/bert](https://github.com/google-research/bert/),
+[huggingface/pytorch-pretrained-BERT](https://github.com/huggingface/pytorch-pretrained-BERT) and [yoheikikuta/bert-japanese](https://github.com/yoheikikuta/bert-japanese).
+
+#### How to use
+
+```python
+from transformers import pipeline, AlbertTokenizer, BertForMaskedLM
+
+tokenizer = AlbertTokenizer.from_pretrained('bayartsogt/bert-base-mongolian-cased')
+model = BertForMaskedLM.from_pretrained('bayartsogt/bert-base-mongolian-cased')
+
+## declare task ##
+pipe = pipeline(task="fill-mask", model=model, tokenizer=tokenizer)
+
+## example ##
+input_  = 'Миний [MASK] хоол идэх нь тун чухал.'
+
+output_ = pipe(input_)
+for i in range(len(output_)):
+    print(output_[i])
+    
+## Output ##
+# {'sequence': '[CLS] Миний хувьд хоол идэх нь тун чухал.[SEP]', 'score': 0.8734784722328186, 'token': 95, 'token_str': '▁хувьд'}
+# {'sequence': '[CLS] Миний бодлоор хоол идэх нь тун чухал.[SEP]', 'score': 0.09788835793733597, 'token': 6320, 'token_str': '▁бодлоор'}
+# {'sequence': '[CLS] Миний хүү хоол идэх нь тун чухал.[SEP]', 'score': 0.0027510314248502254, 'token': 590, 'token_str': '▁хүү'}
+# {'sequence': '[CLS] Миний бие хоол идэх нь тун чухал.[SEP]', 'score': 0.0014857524074614048, 'token': 267, 'token_str': '▁бие'}
+# {'sequence': '[CLS] Миний охин хоол идэх нь тун чухал.[SEP]', 'score': 0.0013575413031503558, 'token': 1116, 'token_str': '▁охин'}
+
+```
+
+
+## Training data
+Mongolian Wikipedia and the 700 million word Mongolian news data set  [[Pretraining Procedure](https://github.com/tugstugi/mongolian-bert#pre-training)]
+
+### BibTeX entry and citation info
+
+```bibtex
+@misc{mongolian-bert,
+  author = {Tuguldur, Erdene-Ochir and Gunchinish, Sharavsambuu and Bataa, Enkhbold},
+  title = {BERT Pretrained Models on Mongolian Datasets},
+  year = {2019},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  howpublished = {\url{https://github.com/tugstugi/mongolian-bert/}}
+}
+```

model_cards/bayartsogt/bert-base-mongolian-uncased/README.md

@@ -0,0 +1,54 @@
+---
+language: "mn"
+tags:
+- bert
+- mongolian
+- uncased
+---
+
+# BERT-BASE-MONGOLIAN-UNCASED
+[Link to Official Mongolian-BERT repo](https://github.com/tugstugi/mongolian-bert)
+
+## Model description
+This repository contains pre-trained Mongolian [BERT](https://arxiv.org/abs/1810.04805) models trained by [tugstugi](https://github.com/tugstugi), [enod](https://github.com/enod) and [sharavsambuu](https://github.com/sharavsambuu).
+Special thanks to [nabar](https://github.com/nabar) who provided 5x TPUs.
+
+This repository is based on the following open source projects: [google-research/bert](https://github.com/google-research/bert/),
+[huggingface/pytorch-pretrained-BERT](https://github.com/huggingface/pytorch-pretrained-BERT) and [yoheikikuta/bert-japanese](https://github.com/yoheikikuta/bert-japanese).
+
+#### How to use
+
+```python
+from transformers import pipeline, AlbertTokenizer, BertForMaskedLM
+
+tokenizer = AlbertTokenizer.from_pretrained('bayartsogt/bert-base-mongolian-uncased')
+model = BertForMaskedLM.from_pretrained('bayartsogt/bert-base-mongolian-uncased')
+
+## declare task ##
+pipe = pipeline(task="fill-mask", model=model, tokenizer=tokenizer)
+
+## example ##
+input_  = 'Миний [MASK] хоол идэх нь тун чухал.'
+
+output_ = pipe(input_)
+for i in range(len(output_)):
+    print(output_[i])
+
+```
+
+
+## Training data
+Mongolian Wikipedia and the 700 million word Mongolian news data set  [[Pretraining Procedure](https://github.com/tugstugi/mongolian-bert#pre-training)]
+
+### BibTeX entry and citation info
+
+```bibtex
+@misc{mongolian-bert,
+  author = {Tuguldur, Erdene-Ochir and Gunchinish, Sharavsambuu and Bataa, Enkhbold},
+  title = {BERT Pretrained Models on Mongolian Datasets},
+  year = {2019},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  howpublished = {\url{https://github.com/tugstugi/mongolian-bert/}}
+}
+```

model_cards/digitalepidemiologylab/covid-twitter-bert/README.md

@@ -1,5 +1,18 @@
-# COVID-Twitter-BERT (CT-BERT)
-BERT-large-uncased model, pretrained on a corpus of messages from Twitter about COVID-19
+---
+language: "en"
+thumbnail: "https://raw.githubusercontent.com/digitalepidemiologylab/covid-twitter-bert/master/images/COVID-Twitter-BERT_small.png"
+tags:
+- Twitter
+- COVID-19
+license: "MIT"
+---
+
+# COVID-Twitter-BERT (CT-BERT) v1
+
+:warning: _You may want to use the [v2 model](https://huggingface.co/digitalepidemiologylab/covid-twitter-bert-v2) which was trained on more recent data and yields better performance_ :warning: 
+
+
+BERT-large-uncased model, pretrained on a corpus of messages from Twitter about COVID-19. Find more info on our [GitHub page](https://github.com/digitalepidemiologylab/covid-twitter-bert).
 
 ## Overview
 This model was trained on 160M tweets collected between January 12 and April 16, 2020 containing at least one of the keywords "wuhan", "ncov", "coronavirus", "covid", or "sars-cov-2". These tweets were filtered and preprocessed to reach a final sample of 22.5M tweets (containing 40.7M sentences and 633M tokens) which were used for training.
@@ -14,5 +27,25 @@ tokenizer = AutoTokenizer.from_pretrained("digitalepidemiologylab/covid-twitter-
 model = AutoModel.from_pretrained("digitalepidemiologylab/covid-twitter-bert")
 ```
 
+You can also use the model with the `pipeline` interface:
+
+```python
+from transformers import pipeline
+import json
+
+pipe = pipeline(task='fill-mask', model='digitalepidemiologylab/covid-twitter-bert-v2')
+out = pipe(f"In places with a lot of people, it's a good idea to wear a {pipe.tokenizer.mask_token}")
+print(json.dumps(out, indent=4))
+[
+    {   
+        "sequence": "[CLS] in places with a lot of people, it's a good idea to wear a mask [SEP]",
+        "score": 0.9959408044815063,
+        "token": 7308,
+        "token_str": "mask"
+    },  
+    ... 
+]
+```
+
 ## References
 [1] Martin Müller, Marcel Salaté, Per E Kummervold. "COVID-Twitter-BERT: A Natural Language Processing Model to Analyse COVID-19 Content on Twitter" arXiv preprint arXiv:2005.07503 (2020).

model_cards/distilbert-base-german-cased-README.md

@@ -0,0 +1,4 @@
+---
+language: de
+---
+## distilbert-base-german-cased

model_cards/microsoft/DeBERTa-base/README.md

@@ -0,0 +1,36 @@
+---
+thumbnail: https://huggingface.co/front/thumbnails/microsoft.png
+license: mit
+---
+
+## DeBERTa: Decoding-enhanced BERT with Disentangled Attention
+
+[DeBERTa](https://arxiv.org/abs/2006.03654) improves the BERT and RoBERTa models using disentangled attention and enhanced mask decoder. With those two improvements, DeBERTa out perform RoBERTa on a majority of NLU tasks with 80GB training data. 
+
+Please check the [official repository](https://github.com/microsoft/DeBERTa) for more details and updates.
+
+
+#### Fine-tuning on NLU tasks
+
+We present the dev results on SQuAD 1.1/2.0 and MNLI tasks.
+
+| Model             | SQuAD 1.1 | SQuAD 2.0 | MNLI-m |
+|-------------------|-----------|-----------|--------|
+| RoBERTa-base      | 91.5/84.6 | 83.7/80.5 | 87.6   |
+| XLNet-Large       | -/-       | -/80.2    | 86.8   |
+| **DeBERTa-base**  | 93.1/87.2 | 86.2/83.1 | 88.8   |
+
+### Citation
+
+If you find DeBERTa useful for your work, please cite the following paper:
+
+``` latex
+@misc{he2020deberta,
+    title={DeBERTa: Decoding-enhanced BERT with Disentangled Attention},
+    author={Pengcheng He and Xiaodong Liu and Jianfeng Gao and Weizhu Chen},
+    year={2020},
+    eprint={2006.03654},
+    archivePrefix={arXiv},
+    primaryClass={cs.CL}
+		}
+```

model_cards/microsoft/DeBERTa-large/README.md

@@ -0,0 +1,37 @@
+---
+thumbnail: https://huggingface.co/front/thumbnails/microsoft.png
+license: mit
+---
+
+## DeBERTa: Decoding-enhanced BERT with Disentangled Attention
+
+[DeBERTa](https://arxiv.org/abs/2006.03654) improves the BERT and RoBERTa models using disentangled attention and enhanced mask decoder. With those two improvements, DeBERTa out perform RoBERTa on a majority of NLU tasks with 80GB training data. 
+
+Please check the [official repository](https://github.com/microsoft/DeBERTa) for more details and updates.
+
+
+#### Fine-tuning on NLU tasks
+
+We present the dev results on SQuAD 1.1/2.0 and several GLUE benchmark tasks.
+
+| Model             | SQuAD 1.1 | SQuAD 2.0 | MNLI-m | SST-2 | QNLI | CoLA | RTE  | MRPC | QQP  |STS-B|
+|-------------------|-----------|-----------|--------|-------|------|------|------|------|------|-----|
+| BERT-Large        | 90.9/84.1 | 81.8/79.0 | 86.6   | 93.2  | 92.3 | 60.6 | 70.4 | 88.0 | 91.3 |90.0 |
+| RoBERTa-Large     | 94.6/88.9 | 89.4/86.5 | 90.2   | 96.4  | 93.9 | 68.0 | 86.6 | 90.9 | 92.2 |92.4 |
+| XLNet-Large       | 95.1/89.7 | 90.6/87.9 | 90.8   | 97.0  | 94.9 | 69.0 | 85.9 | 90.8 | 92.3 |92.5 |
+| **DeBERTa-Large** | 95.5/90.1 | 90.7/88.0 | 91.1   | 96.5  | 95.3 | 69.5 | 88.1 | 92.5 | 92.3 |92.5 |
+
+### Citation
+
+If you find DeBERTa useful for your work, please cite the following paper:
+
+``` latex
+@misc{he2020deberta,
+    title={DeBERTa: Decoding-enhanced BERT with Disentangled Attention},
+    author={Pengcheng He and Xiaodong Liu and Jianfeng Gao and Weizhu Chen},
+    year={2020},
+    eprint={2006.03654},
+    archivePrefix={arXiv},
+    primaryClass={cs.CL}
+		}
+```

model_cards/microsoft/DialoGPT-large/README.md

@@ -31,12 +31,12 @@ ArXiv paper: [https://arxiv.org/abs/1911.00536](https://arxiv.org/abs/1911.00536
 Now we are ready to try out how the model works as a chatting partner!
 
 ```python
-from transformers import AutoModelWithLMHead, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
 
 
 tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-large")
-model = AutoModelWithLMHead.from_pretrained("microsoft/DialoGPT-large")
+model = AutoModelForCausalLM.from_pretrained("microsoft/DialoGPT-large")
 
 # Let's chat for 5 lines
 for step in range(5):

model_cards/microsoft/DialoGPT-medium/README.md

@@ -31,12 +31,12 @@ ArXiv paper: [https://arxiv.org/abs/1911.00536](https://arxiv.org/abs/1911.00536
 Now we are ready to try out how the model works as a chatting partner!
 
 ```python
-from transformers import AutoModelWithLMHead, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
 
 
 tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-medium")
-model = AutoModelWithLMHead.from_pretrained("microsoft/DialoGPT-medium")
+model = AutoModelForCausalLM.from_pretrained("microsoft/DialoGPT-medium")
 
 # Let's chat for 5 lines
 for step in range(5):

model_cards/microsoft/DialoGPT-small/README.md

@@ -31,12 +31,12 @@ ArXiv paper: [https://arxiv.org/abs/1911.00536](https://arxiv.org/abs/1911.00536
 Now we are ready to try out how the model works as a chatting partner!
 
 ```python
-from transformers import AutoModelWithLMHead, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
 
 
 tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-small")
-model = AutoModelWithLMHead.from_pretrained("microsoft/DialoGPT-small")
+model = AutoModelForCausalLM.from_pretrained("microsoft/DialoGPT-small")
 
 # Let's chat for 5 lines
 for step in range(5):

model_cards/mrm8488/electricidad-base-discriminator/README.md

@@ -59,9 +59,21 @@ predictions = torch.round((torch.sign(discriminator_outputs[0]) + 1) / 2)
 el rapido  zorro  marro    ##n   amar  sobre     el  perro   pere ##zoso    0.0    0.0    0.0    0.0    0.0    0.0    1.0    1.0    0.0    0.0    0.0    0.0    0.0[None, None, None, None, None, None, None, None, None, None, None, None, None
 '''
 ```
-
 As you can see there are **1s** in the places where the model detected a fake token. So, it works! 🎉
 
+
+### Some models fine-tuned on a downstream task 🛠️
+
+[Question Answering](https://huggingface.co/mrm8488/electricidad-base-finetuned-squadv1-es)
+
+[POS](https://huggingface.co/mrm8488/electricidad-base-finetuned-pos)
+
+[NER](https://huggingface.co/mrm8488/electricidad-base-finetuned-ner)
+
+[Paraphrase Identification](https://huggingface.co/mrm8488/RuPERTa-base-finetuned-pawsx-es)
+
+
+
 ## Acknowledgments
 
 I thank [🤗/transformers team](https://github.com/huggingface/transformers) for allowing me to train the model (specially to [Julien Chaumond](https://twitter.com/julien_c)).

model_cards/ncoop57/bart-base-code-summarizer-java-v0/README.md

@@ -0,0 +1,8 @@
+---
+tags:
+- summarization
+
+license: mit
+---
+
+## ncoop57/bart-base-code-summarizer-java-v0

model_cards/squeezebert/squeezebert-mnli-headless/README.md

@@ -0,0 +1,67 @@
+language: en
+license: bsd
+datasets:
+- bookcorpus
+- wikipedia
+---
+
+# SqueezeBERT pretrained model
+
+This model, `squeezebert-mnli-headless`, has been pretrained for the English language using a masked language modeling (MLM) and Sentence Order Prediction (SOP) objective and finetuned on the [Multi-Genre Natural Language Inference (MNLI)](https://cims.nyu.edu/~sbowman/multinli/) dataset. This is a "headless" model with the final classification layer removed, and this will allow Transformers to automatically reinitialize the final classification layer before you begin finetuning on your data.
+SqueezeBERT was introduced in [this paper](https://arxiv.org/abs/2006.11316). This model is case-insensitive. The model architecture is similar to BERT-base, but with the pointwise fully-connected layers replaced with [grouped convolutions](https://blog.yani.io/filter-group-tutorial/).
+The authors found that SqueezeBERT is 4.3x faster than `bert-base-uncased` on a Google Pixel 3 smartphone.
+
+
+## Pretraining
+
+### Pretraining data
+- [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of thousands of unpublished books
+- [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia)
+
+### Pretraining procedure
+The model is pretrained using the Masked Language Model (MLM) and Sentence Order Prediction (SOP) tasks.
+(Author's note: If you decide to pretrain your own model, and you prefer to train with MLM only, that should work too.)
+
+From the SqueezeBERT paper:
+> We pretrain SqueezeBERT from scratch (without distillation) using the [LAMB](https://arxiv.org/abs/1904.00962) optimizer, and we employ the hyperparameters recommended by the LAMB authors: a global batch size of 8192, a learning rate of 2.5e-3, and a warmup proportion of 0.28. Following the LAMB paper's recommendations, we pretrain for 56k steps with a maximum sequence length of 128 and then for 6k steps with a maximum sequence length of 512.
+
+## Finetuning
+
+The SqueezeBERT paper presents 2 approaches to finetuning the model:
+- "finetuning without bells and whistles" -- after pretraining the SqueezeBERT model, finetune it on each GLUE task
+- "finetuning with bells and whistles" -- after pretraining the SqueezeBERT model, finetune it on a MNLI with distillation from a teacher model. Then, use the MNLI-finetuned SqueezeBERT model as a student model to finetune on each of the other GLUE tasks (e.g. RTE, MRPC, …) with distillation from a task-specific teacher model.
+
+A detailed discussion of the hyperparameters used for finetuning is provided in the appendix of the [SqueezeBERT paper](https://arxiv.org/abs/2006.11316).
+Note that finetuning SqueezeBERT with distillation is not yet implemented in this repo. If the author (Forrest Iandola - forrest.dnn@gmail.com) gets enough encouragement from the user community, he will add example code to Transformers for finetuning SqueezeBERT with distillation.
+
+This model, `squeezebert/squeezebert-mnli-headless`, is the "finetuned with bells and whistles" MNLI-finetuned SqueezeBERT model. In this particular model, we have removed the final classification layer -- in other words, it is "headless." We recommend using this model if you intend to finetune the model on your own data. Using this model means that your final layer will automatically be reinitialized when you start finetuning on your data.
+
+### How to finetune
+To try finetuning SqueezeBERT on the [MRPC](https://www.microsoft.com/en-us/download/details.aspx?id=52398) text classification task, you can run the following command:
+```
+./utils/download_glue_data.py
+
+python examples/text-classification/run_glue.py \
+    --model_name_or_path squeezebert-base-headless \
+    --task_name mrpc \
+    --data_dir ./glue_data/MRPC \
+    --output_dir ./models/squeezebert_mrpc \
+    --overwrite_output_dir \
+    --do_train \
+    --do_eval \
+    --num_train_epochs 10 \
+    --learning_rate 3e-05 \
+    --per_device_train_batch_size 16 \
+    --save_steps 20000
+
+```
+
+## BibTeX entry and citation info
+```
+@article{2020_SqueezeBERT,
+     author = {Forrest N. Iandola and Albert E. Shaw and Ravi Krishna and Kurt W. Keutzer},
+     title = {{SqueezeBERT}: What can computer vision teach NLP about efficient neural networks?},
+     journal = {arXiv:2006.11316},
+     year = {2020}
+}
+```

model_cards/squeezebert/squeezebert-mnli/README.md

@@ -0,0 +1,67 @@
+language: en
+license: bsd
+datasets:
+- bookcorpus
+- wikipedia
+---
+
+# SqueezeBERT pretrained model
+
+This model, `squeezebert-mnli`, has been pretrained for the English language using a masked language modeling (MLM) and Sentence Order Prediction (SOP) objective and finetuned on the [Multi-Genre Natural Language Inference (MNLI)](https://cims.nyu.edu/~sbowman/multinli/) dataset.
+SqueezeBERT was introduced in [this paper](https://arxiv.org/abs/2006.11316). This model is case-insensitive. The model architecture is similar to BERT-base, but with the pointwise fully-connected layers replaced with [grouped convolutions](https://blog.yani.io/filter-group-tutorial/).
+The authors found that SqueezeBERT is 4.3x faster than `bert-base-uncased` on a Google Pixel 3 smartphone.
+
+
+## Pretraining
+
+### Pretraining data
+- [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of thousands of unpublished books
+- [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia)
+
+### Pretraining procedure
+The model is pretrained using the Masked Language Model (MLM) and Sentence Order Prediction (SOP) tasks.
+(Author's note: If you decide to pretrain your own model, and you prefer to train with MLM only, that should work too.)
+
+From the SqueezeBERT paper:
+> We pretrain SqueezeBERT from scratch (without distillation) using the [LAMB](https://arxiv.org/abs/1904.00962) optimizer, and we employ the hyperparameters recommended by the LAMB authors: a global batch size of 8192, a learning rate of 2.5e-3, and a warmup proportion of 0.28. Following the LAMB paper's recommendations, we pretrain for 56k steps with a maximum sequence length of 128 and then for 6k steps with a maximum sequence length of 512.
+
+## Finetuning
+
+The SqueezeBERT paper presents 2 approaches to finetuning the model:
+- "finetuning without bells and whistles" -- after pretraining the SqueezeBERT model, finetune it on each GLUE task
+- "finetuning with bells and whistles" -- after pretraining the SqueezeBERT model, finetune it on a MNLI with distillation from a teacher model. Then, use the MNLI-finetuned SqueezeBERT model as a student model to finetune on each of the other GLUE tasks (e.g. RTE, MRPC, …) with distillation from a task-specific teacher model.
+
+A detailed discussion of the hyperparameters used for finetuning is provided in the appendix of the [SqueezeBERT paper](https://arxiv.org/abs/2006.11316).
+Note that finetuning SqueezeBERT with distillation is not yet implemented in this repo. If the author (Forrest Iandola - forrest.dnn@gmail.com) gets enough encouragement from the user community, he will add example code to Transformers for finetuning SqueezeBERT with distillation.
+
+This model, `squeezebert/squeezebert-mnli`, is the "trained with bells and whistles" MNLI-finetuned SqueezeBERT model.
+
+### How to finetune
+To try finetuning SqueezeBERT on the [MRPC](https://www.microsoft.com/en-us/download/details.aspx?id=52398) text classification task, you can run the following command:
+```
+./utils/download_glue_data.py
+
+python examples/text-classification/run_glue.py \
+    --model_name_or_path squeezebert-base-headless \
+    --task_name mrpc \
+    --data_dir ./glue_data/MRPC \
+    --output_dir ./models/squeezebert_mrpc \
+    --overwrite_output_dir \
+    --do_train \
+    --do_eval \
+    --num_train_epochs 10 \
+    --learning_rate 3e-05 \
+    --per_device_train_batch_size 16 \
+    --save_steps 20000
+
+```
+
+## BibTeX entry and citation info
+```
+@article{2020_SqueezeBERT,
+     author = {Forrest N. Iandola and Albert E. Shaw and Ravi Krishna and Kurt W. Keutzer},
+     title = {{SqueezeBERT}: What can computer vision teach NLP about efficient neural networks?},
+     journal = {arXiv:2006.11316},
+     year = {2020}
+}
+```

model_cards/squeezebert/squeezebert-uncased/README.md

@@ -0,0 +1,67 @@
+language: en
+license: bsd
+datasets:
+- bookcorpus
+- wikipedia
+---
+
+# SqueezeBERT pretrained model
+
+This model, `squeezebert-uncased`, is a pretrained model for the English language using a masked language modeling (MLM) and Sentence Order Prediction (SOP) objective.
+SqueezeBERT was introduced in [this paper](https://arxiv.org/abs/2006.11316). This model is case-insensitive. The model architecture is similar to BERT-base, but with the pointwise fully-connected layers replaced with [grouped convolutions](https://blog.yani.io/filter-group-tutorial/).
+The authors found that SqueezeBERT is 4.3x faster than `bert-base-uncased` on a Google Pixel 3 smartphone.
+
+
+## Pretraining
+
+### Pretraining data
+- [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of thousands of unpublished books
+- [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia)
+
+### Pretraining procedure
+The model is pretrained using the Masked Language Model (MLM) and Sentence Order Prediction (SOP) tasks.
+(Author's note: If you decide to pretrain your own model, and you prefer to train with MLM only, that should work too.)
+
+The SqueezeBERT paper presents 2 approaches to finetuning the model:
+> We pretrain SqueezeBERT from scratch (without distillation) using the [LAMB](https://arxiv.org/abs/1904.00962) optimizer, and we employ the hyperparameters recommended by the LAMB authors: a global batch size of 8192, a learning rate of 2.5e-3, and a warmup proportion of 0.28. Following the LAMB paper's recommendations, we pretrain for 56k steps with a maximum sequence length of 128 and then for 6k steps with a maximum sequence length of 512.
+
+## Finetuning
+
+The SqueezeBERT paper results from 2 approaches to finetuning the model:
+- "finetuning without bells and whistles" -- after pretraining the SqueezeBERT model, finetune it on each GLUE task
+- "finetuning with bells and whistles" -- after pretraining the SqueezeBERT model, finetune it on a MNLI with distillation from a teacher model. Then, use the MNLI-finetuned SqueezeBERT model as a student model to finetune on each of the other GLUE tasks (e.g. RTE, MRPC, …) with distillation from a task-specific teacher model.
+
+A detailed discussion of the hyperparameters used for finetuning is provided in the appendix of the [SqueezeBERT paper](https://arxiv.org/abs/2006.11316).
+Note that finetuning SqueezeBERT with distillation is not yet implemented in this repo. If the author (Forrest Iandola - forrest.dnn@gmail.com) gets enough encouragement from the user community, he will add example code to Transformers for finetuning SqueezeBERT with distillation.
+
+This model, `squeezebert/squeezebert-uncased`, has been pretrained but not finetuned. For most text classification tasks, we recommend using squeezebert-mnli-headless as a starting point.
+
+### How to finetune
+To try finetuning SqueezeBERT on the [MRPC](https://www.microsoft.com/en-us/download/details.aspx?id=52398) text classification task, you can run the following command:
+```
+./utils/download_glue_data.py
+
+python examples/text-classification/run_glue.py \
+    --model_name_or_path squeezebert-base-headless \
+    --task_name mrpc \
+    --data_dir ./glue_data/MRPC \
+    --output_dir ./models/squeezebert_mrpc \
+    --overwrite_output_dir \
+    --do_train \
+    --do_eval \
+    --num_train_epochs 10 \
+    --learning_rate 3e-05 \
+    --per_device_train_batch_size 16 \
+    --save_steps 20000
+
+```
+
+## BibTeX entry and citation info
+```
+@article{2020_SqueezeBERT,
+     author = {Forrest N. Iandola and Albert E. Shaw and Ravi Krishna and Kurt W. Keutzer},
+     title = {{SqueezeBERT}: What can computer vision teach NLP about efficient neural networks?},
+     journal = {arXiv:2006.11316},
+     year = {2020}
+}
+```

model_cards/unideeplearning/polibert_sa/README.md

@@ -12,15 +12,22 @@ widget:
   
 ## Model description  
   
-This model performs sentiment analysis on Italian political twitter sentences. It was trained starting from an instance of "bert-base-italian-uncased-xxl" and fine-tuned on an Italian dataset of tweets.
+This model performs sentiment analysis on Italian political twitter sentences. It was trained starting from an instance of "bert-base-italian-uncased-xxl" and fine-tuned on an Italian dataset of tweets. You can try it out at https://www.unideeplearning.com/twitter_sa/ (in italian!)
   
 #### Hands-on  
   
 ```python
 import torch
 from torch import nn 
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
 
-text = "Giueseppe Rossi è un pessimo politico"
+tokenizer = AutoTokenizer.from_pretrained("unideeplearning/polibert_sa")
+model = AutoModelForSequenceClassification.from_pretrained("unideeplearning/polibert_sa")
+			
+
+
+
+text = "Giuseppe Rossi è un pessimo politico"
 input_ids = tokenizer.encode(text, add_special_tokens=True, return_tensors= 'pt')
 
 logits, = model(input_ids)
@@ -41,4 +48,6 @@ print(prob.argmax().tolist())
 ## Acknowledgments
 
 Thanks to the support from: 
-the [Hugging Face](https://huggingface.co/), Unione Professionisti (https://www.unioneprofessionisti.com/)
+the [Hugging Face](https://huggingface.co/), https://www.unioneprofessionisti.com
+
+https://www.unideeplearning.com/

notebooks/README.md

@@ -9,6 +9,7 @@ Pull Request so it can be included under the Community notebooks.
 
 ## Hugging Face's notebooks 🤗
 
+
 | Notebook     |      Description      |   |
 |:----------|:-------------|------:|
 | [Getting Started Tokenizers](https://github.com/huggingface/transformers/blob/master/notebooks/01-training-tokenizers.ipynb)  | How to train and use your very own tokenizer  |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/transformers/blob/master/notebooks/01-training-tokenizers.ipynb) |
@@ -25,6 +26,7 @@ Pull Request so it can be included under the Community notebooks.
 
 | Notebook     |      Description      |      Author      |      |
 |:----------|:-------------|:-------------|------:|
+| [Train T5 in Tensoflow 2 ](https://github.com/snapthat/TF-T5-text-to-text) | How to train T5 for any task using Tensorflow 2. This notebook demonstrates a Question & Answer task implemented in Tensorflow 2 using SQUAD | [Muhammad Harris](https://github.com/HarrisDePerceptron) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/snapthat/TF-T5-text-to-text/blob/master/snapthatT5/notebooks/TF-T5-Datasets%20Training.ipynb) |
 | [Train T5 on TPU](https://github.com/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb)  | How to train T5 on SQUAD with Transformers and Nlp | [Suraj Patil](https://github.com/patil-suraj) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb#scrollTo=QLGiFCDqvuil) |
 | [Fine-tune T5 for Classification and Multiple Choice](https://github.com/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb)  | How to fine-tune T5 for classification and multiple choice tasks using a text-to-text format with PyTorch Lightning |  [Suraj Patil](https://github.com/patil-suraj) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) |
 | [Fine-tune DialoGPT on New Datasets and Languages](https://github.com/ncoop57/i-am-a-nerd/blob/master/_notebooks/2020-05-12-chatbot-part-1.ipynb)  | How to fine-tune the DialoGPT model on a new dataset for open-dialog conversational chatbots |  [Nathan Cooper](https://github.com/ncoop57) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ncoop57/i-am-a-nerd/blob/master/_notebooks/2020-05-12-chatbot-part-1.ipynb) |
@@ -46,3 +48,4 @@ Pull Request so it can be included under the Community notebooks.
 |[fine-tune a non-English GPT-2 Model with Trainer class](https://github.com/philschmid/fine-tune-GPT-2/blob/master/Fine_tune_a_non_English_GPT_2_Model_with_Huggingface.ipynb) | How to fine-tune a non-English GPT-2 Model with Trainer class | [Philipp Schmid](https://www.philschmid.de) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/philschmid/fine-tune-GPT-2/blob/master/Fine_tune_a_non_English_GPT_2_Model_with_Huggingface.ipynb)|
 |[Fine-tune a DistilBERT Model for Multi Label Classification task](https://github.com/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb) | How to fine-tune a DistilBERT Model for Multi Label Classification task | [Dhaval Taunk](https://github.com/DhavalTaunk08) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb)|
 |[Fine-tune ALBERT for sentence-pair classification](https://github.com/NadirEM/nlp-notebooks/blob/master/Fine_tune_ALBERT_sentence_pair_classification.ipynb) | How to fine-tune an ALBERT model or another BERT-based model for the sentence-pair classification task | [Nadir El Manouzi](https://github.com/NadirEM) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NadirEM/nlp-notebooks/blob/master/Fine_tune_ALBERT_sentence_pair_classification.ipynb)|
+|[Fine-tune Roberta for sentiment analysis](https://github.com/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb) | How to fine-tune an Roberta model for sentiment analysis | [Dhaval Taunk](https://github.com/DhavalTaunk08) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb)|

setup.py

@@ -5,7 +5,7 @@ To create the package for pypi.
 
 1. Change the version in __init__.py, setup.py as well as docs/source/conf.py.
 
-2. Unpin specific versions from setup.py (like isort).
+2. Unpin specific versions from setup.py that use a git install.
 
 2. Commit these changes with the message: "Release: VERSION"
 
@@ -93,12 +93,12 @@ extras["retrieval"] = ["faiss-cpu", "datasets"]
 extras["testing"] = ["pytest", "pytest-xdist", "timeout-decorator", "parameterized", "psutil"] + extras["retrieval"]
 # sphinx-rtd-theme==0.5.0 introduced big changes in the style.
 extras["docs"] = ["recommonmark", "sphinx", "sphinx-markdown-tables", "sphinx-rtd-theme==0.4.3", "sphinx-copybutton"]
-extras["quality"] = ["black >= 20.8b1", "isort >= 5", "flake8 >= 3.8.3"]
+extras["quality"] = ["black >= 20.8b1", "isort >= 5.5.4", "flake8 >= 3.8.3"]
 extras["dev"] = extras["testing"] + extras["quality"] + extras["ja"] + ["scikit-learn", "tensorflow", "torch"]
 
 setup(
     name="transformers",
-    version="3.2.0",
+    version="3.3.1",
     author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Sam Shleifer, Patrick von Platen, Sylvain Gugger, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
     author_email="thomas@huggingface.co",
     description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch",
@@ -111,7 +111,7 @@ setup(
     packages=find_packages("src"),
     install_requires=[
         "numpy",
-        "tokenizers == 0.8.1.rc2",
+        "tokenizers == 0.9.0.rc1",
         # dataclasses for Python versions that don't have it
         "dataclasses;python_version<'3.7'",
         # utilities from PyPA to e.g. compare versions
@@ -124,8 +124,9 @@ setup(
         "tqdm >= 4.27",
         # for OpenAI GPT
         "regex != 2019.12.17",
-        # for XLNet
+        # for SentencePiece models
         "sentencepiece != 0.1.92",
+        "protobuf",
         # for XLM
         "sacremoses",
     ],

src/transformers/__init__.py

@@ -2,7 +2,7 @@
 # There's no way to ignore "F401 '...' imported but unused" warnings in this
 # module, but to preserve other warnings. So, don't check this module at all.
 
-__version__ = "3.2.0"
+__version__ = "3.3.1"
 
 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -35,6 +35,7 @@ from .configuration_bert import BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, BertConfig
 from .configuration_bert_generation import BertGenerationConfig
 from .configuration_camembert import CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, CamembertConfig
 from .configuration_ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig
+from .configuration_deberta import DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaConfig
 from .configuration_distilbert import DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DistilBertConfig
 from .configuration_dpr import DPR_PRETRAINED_CONFIG_ARCHIVE_MAP, DPRConfig
 from .configuration_electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, ElectraConfig
@@ -56,6 +57,7 @@ from .configuration_rag import RagConfig
 from .configuration_reformer import REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, ReformerConfig
 from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RetriBertConfig
 from .configuration_roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig
+from .configuration_squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig
 from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
 from .configuration_transfo_xl import TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP, TransfoXLConfig
 from .configuration_utils import PretrainedConfig
@@ -71,12 +73,13 @@ from .data import (
     SquadFeatures,
     SquadV1Processor,
     SquadV2Processor,
+    glue_compute_metrics,
     glue_convert_examples_to_features,
     glue_output_modes,
     glue_processors,
     glue_tasks_num_labels,
-    is_sklearn_available,
     squad_convert_examples_to_features,
+    xnli_compute_metrics,
     xnli_output_modes,
     xnli_processors,
     xnli_tasks_num_labels,
@@ -100,9 +103,12 @@ from .file_utils import (
     is_faiss_available,
     is_psutil_available,
     is_py3nvml_available,
+    is_sklearn_available,
     is_tf_available,
     is_torch_available,
     is_torch_tpu_available,
+    is_sentencepiece_available,
+    is_tokenizers_available,
 )
 from .hf_argparser import HfArgumentParser
 
@@ -147,15 +153,16 @@ from .pipelines import (
 from .retrieval_rag import RagRetriever
 
 # Tokenizers
-from .tokenization_albert import AlbertTokenizer
+from .tokenization_albert import AlbertTokenizer, AlbertTokenizerFast
 from .tokenization_auto import TOKENIZER_MAPPING, AutoTokenizer
 from .tokenization_bart import BartTokenizer, BartTokenizerFast
 from .tokenization_bert import BasicTokenizer, BertTokenizer, BertTokenizerFast, WordpieceTokenizer
 from .tokenization_bert_generation import BertGenerationTokenizer
 from .tokenization_bert_japanese import BertJapaneseTokenizer, CharacterTokenizer, MecabTokenizer
 from .tokenization_bertweet import BertweetTokenizer
-from .tokenization_camembert import CamembertTokenizer
+from .tokenization_camembert import CamembertTokenizer, CamembertTokenizerFast
 from .tokenization_ctrl import CTRLTokenizer
+from .tokenization_deberta import DebertaTokenizer
 from .tokenization_distilbert import DistilBertTokenizer, DistilBertTokenizerFast
 from .tokenization_dpr import (
     DPRContextEncoderTokenizer,
@@ -173,17 +180,18 @@ from .tokenization_gpt2 import GPT2Tokenizer, GPT2TokenizerFast
 from .tokenization_layoutlm import LayoutLMTokenizer, LayoutLMTokenizerFast
 from .tokenization_longformer import LongformerTokenizer, LongformerTokenizerFast
 from .tokenization_lxmert import LxmertTokenizer, LxmertTokenizerFast
-from .tokenization_mbart import MBartTokenizer
+from .tokenization_mbart import MBartTokenizer, MBartTokenizerFast
 from .tokenization_mobilebert import MobileBertTokenizer, MobileBertTokenizerFast
 from .tokenization_openai import OpenAIGPTTokenizer, OpenAIGPTTokenizerFast
-from .tokenization_pegasus import PegasusTokenizer
+from .tokenization_pegasus import PegasusTokenizer, PegasusTokenizerFast
 from .tokenization_phobert import PhobertTokenizer
 from .tokenization_rag import RagTokenizer
-from .tokenization_reformer import ReformerTokenizer
+from .tokenization_reformer import ReformerTokenizer, ReformerTokenizerFast
 from .tokenization_retribert import RetriBertTokenizer, RetriBertTokenizerFast
 from .tokenization_roberta import RobertaTokenizer, RobertaTokenizerFast
-from .tokenization_t5 import T5Tokenizer
-from .tokenization_transfo_xl import TransfoXLCorpus, TransfoXLTokenizer, TransfoXLTokenizerFast
+from .tokenization_squeezebert import SqueezeBertTokenizer, SqueezeBertTokenizerFast
+from .tokenization_t5 import T5Tokenizer, T5TokenizerFast
+from .tokenization_transfo_xl import TransfoXLCorpus, TransfoXLTokenizer
 from .tokenization_utils import PreTrainedTokenizer
 from .tokenization_utils_base import (
     BatchEncoding,
@@ -195,11 +203,11 @@ from .tokenization_utils_base import (
 )
 from .tokenization_utils_fast import PreTrainedTokenizerFast
 from .tokenization_xlm import XLMTokenizer
-from .tokenization_xlm_roberta import XLMRobertaTokenizer
-from .tokenization_xlnet import SPIECE_UNDERLINE, XLNetTokenizer
+from .tokenization_xlm_roberta import XLMRobertaTokenizer, XLMRobertaTokenizerFast
+from .tokenization_xlnet import SPIECE_UNDERLINE, XLNetTokenizer, XLNetTokenizerFast
 
 # Trainer
-from .trainer_utils import EvalPrediction, set_seed
+from .trainer_utils import EvalPrediction, TrainerState, set_seed
 from .training_args import TrainingArguments
 from .training_args_tf import TFTrainingArguments
 from .utils import logging
@@ -208,10 +216,6 @@ from .utils import logging
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-if is_sklearn_available():
-    from .data import glue_compute_metrics, xnli_compute_metrics
-
-
 # Modeling
 if is_torch_available():
     # Benchmarks
@@ -310,6 +314,12 @@ if is_torch_available():
         CamembertModel,
     )
     from .modeling_ctrl import CTRL_PRETRAINED_MODEL_ARCHIVE_LIST, CTRLLMHeadModel, CTRLModel, CTRLPreTrainedModel
+    from .modeling_deberta import (
+        DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
+        DebertaForSequenceClassification,
+        DebertaModel,
+        DebertaPreTrainedModel,
+    )
     from .modeling_distilbert import (
         DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
         DistilBertForMaskedLM,
@@ -445,6 +455,17 @@ if is_torch_available():
         RobertaForTokenClassification,
         RobertaModel,
     )
+    from .modeling_squeezebert import (
+        SQUEEZEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        SqueezeBertForMaskedLM,
+        SqueezeBertForMultipleChoice,
+        SqueezeBertForQuestionAnswering,
+        SqueezeBertForSequenceClassification,
+        SqueezeBertForTokenClassification,
+        SqueezeBertModel,
+        SqueezeBertModule,
+        SqueezeBertPreTrainedModel,
+    )
     from .modeling_t5 import (
         T5_PRETRAINED_MODEL_ARCHIVE_LIST,
         T5ForConditionalGeneration,
@@ -510,6 +531,8 @@ if is_torch_available():
 
     # Trainer
     from .trainer import EvalPrediction, Trainer, set_seed, torch_distributed_zero_first
+else:
+    from .utils.dummy_pt_objects import *
 
 # TensorFlow
 if is_tf_available():
@@ -732,6 +755,11 @@ if is_tf_available():
     # Trainer
     from .trainer_tf import TFTrainer
 
+else:
+    # Import the same objects as dummies to get them in the namespace.
+    # They will raise an import error if the user tries to instantiate / use them.
+    from .utils.dummy_tf_objects import *
+
 
 if not is_tf_available() and not is_torch_available():
     logger.warning(

src/transformers/activations.py

@@ -44,6 +44,10 @@ def mish(x):
     return x * torch.tanh(torch.nn.functional.softplus(x))
 
 
+def linear_act(x):
+    return x
+
+
 ACT2FN = {
     "relu": F.relu,
     "swish": swish,
@@ -52,6 +56,8 @@ ACT2FN = {
     "gelu_new": gelu_new,
     "gelu_fast": gelu_fast,
     "mish": mish,
+    "linear": linear_act,
+    "sigmoid": torch.sigmoid,
 }
 
 

src/transformers/configuration_auto.py

@@ -23,6 +23,7 @@ from .configuration_bert import BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, BertConfig
 from .configuration_bert_generation import BertGenerationConfig
 from .configuration_camembert import CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, CamembertConfig
 from .configuration_ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig
+from .configuration_deberta import DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaConfig
 from .configuration_distilbert import DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DistilBertConfig
 from .configuration_dpr import DPR_PRETRAINED_CONFIG_ARCHIVE_MAP, DPRConfig
 from .configuration_electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, ElectraConfig
@@ -43,6 +44,7 @@ from .configuration_rag import RagConfig
 from .configuration_reformer import ReformerConfig
 from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RetriBertConfig
 from .configuration_roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig
+from .configuration_squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig
 from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
 from .configuration_transfo_xl import TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP, TransfoXLConfig
 from .configuration_utils import PretrainedConfig
@@ -78,6 +80,8 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
         LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
         LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
         DPR_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ]
     for key, value, in pretrained_map.items()
 )
@@ -100,8 +104,10 @@ CONFIG_MAPPING = OrderedDict(
         ("reformer", ReformerConfig),
         ("longformer", LongformerConfig),
         ("roberta", RobertaConfig),
+        ("deberta", DebertaConfig),
         ("flaubert", FlaubertConfig),
         ("fsmt", FSMTConfig),
+        ("squeezebert", SqueezeBertConfig),
         ("bert", BertConfig),
         ("openai-gpt", OpenAIGPTConfig),
         ("gpt2", GPT2Config),
@@ -138,6 +144,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("roberta", "RoBERTa"),
         ("flaubert", "FlauBERT"),
         ("fsmt", "FairSeq Machine-Translation"),
+        ("squeezebert", "SqueezeBERT"),
         ("bert", "BERT"),
         ("openai-gpt", "OpenAI GPT"),
         ("gpt2", "OpenAI GPT-2"),
@@ -149,6 +156,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("encoder-decoder", "Encoder decoder"),
         ("funnel", "Funnel Transformer"),
         ("lxmert", "LXMERT"),
+        ("deberta", "DeBERTa"),
         ("layoutlm", "LayoutLM"),
         ("dpr", "DPR"),
         ("rag", "RAG"),

src/transformers/configuration_bart.py

@@ -15,7 +15,6 @@
 """ BART configuration """
 
 from .configuration_utils import PretrainedConfig
-from .file_utils import add_start_docstrings_to_callable
 from .utils import logging
 
 
@@ -31,78 +30,83 @@ BART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     "yjernite/bart_eli5": "https://s3.amazonaws.com/models.huggingface.co/bert/yjernite/bart_eli5/config.json",
 }
 
-BART_CONFIG_ARGS_DOC = r"""
-    Args:
-        vocab_size (:obj:`int`, optional, defaults to 50265):
-            defines the different tokens that can be represented by `inputs_ids` passed to the forward method.
-        d_model (:obj:`int`, optional, defaults to 1024):
-            Dimensionality of the layers and the pooler layer.
-        encoder_layers (:obj:`int`, optional, defaults to 12):
-            Number of encoder layers, 16 for pegasus, 6 for bart-base and marian
-        decoder_layers (:obj:`int`, optional, defaults to 12):
-            Number of decoder layers, 16 for pegasus, 6 for bart-base and marian
-        encoder_attention_heads (:obj:`int`, optional, defaults to 16):
-            Number of attention heads for each attention layer in the Transformer encoder.
-        decoder_attention_heads (:obj:`int`, optional, defaults to 16):
-            Number of attention heads for each attention layer in the Transformer decoder.
-        decoder_ffn_dim (:obj:`int`, optional, defaults to 4096):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
-        encoder_ffn_dim (:obj:`int`, optional, defaults to 4096):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
-        activation_function (:obj:`str` or :obj:`function`, optional, defaults to "gelu"):
-            The non-linear activation function (function or string) in the encoder and pooler.
-            If string, "gelu", "relu", "swish" and "gelu_new" are supported.
-        dropout (:obj:`float`, optional, defaults to 0.1):
-            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
-        attention_dropout (:obj:`float`, optional, defaults to 0.0):
-            The dropout ratio for the attention probabilities.
-        activation_dropout (:obj:`float`, optional, defaults to 0.0):
-            The dropout ratio for activations inside the fully connected layer.
-        classifier_dropout (:obj:`float`, optional, defaults to 0.0):
-            The dropout ratio for classifier.
-        max_position_embeddings (:obj:`int`, optional, defaults to 1024):
-            The maximum sequence length that this model might ever be used with.
-            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
-        init_std (:obj:`float`, optional, defaults to 0.02):
-            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        add_bias_logits (:obj:`bool`, optional, defaults to :obj:`False`):
-            True for marian only.
-        normalize_before (:obj:`bool`, optional, defaults to :obj:`False`):
-            Call layernorm before attention ops. True for pegasus, mbart. False for bart. FIXME: marian?
-        normalize_embedding (:obj:`bool`, optional, defaults to :obj:`True`):
-            Call layernorm after embeddings. Only True for Bart.
-        static_position_embeddings (:obj:`bool`, optional, defaults to :obj:`False`):
-            Don't learn positional embeddings, use sinusoidal. True for marian, pegasus.
-        add_final_layer_norm (:obj:`bool`, optional, defaults to :obj:`False`):
-            Why not add another layernorm?
-        scale_embedding (:obj:`bool`, optional, defaults to :obj:`False`):
-            Scale embeddings by diving by sqrt(d_model).
-        eos_token_id (:obj:`int`, optional, defaults to 2)
-            End of stream token id.
-        pad_token_id (:obj:`int`, optional, defaults to 1)
-            Padding token id.
-        bos_token_id (:obj:`int`, optional, defaults to 0)
-            Beginning of stream token id.
-        encoder_layerdrop: (:obj:`float`, optional, defaults to 0.0):
-            Google "layerdrop arxiv", as its not explainable in one line.
-        decoder_layerdrop: (:obj:`float`, optional, defaults to 0.0):
-            Google "layerdrop arxiv", as its not explainable in one line.
-        extra_pos_embeddings: (:obj:`int`, optional, defaults to 2):
-            How many extra learned positional embeddings to use. Should be pad_token_id+1 for bart.
-        num_labels: (:obj:`int`, optional, defaults to 3):
-            for SequenceClassification
-        is_encoder_decoder (:obj:`bool`, optional, defaults to :obj:`True`):
-            Whether this is an encoder/decoder model
-        force_bos_token_to_be_generated (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether or not to force BOS token to be generated at step 1 (after ``decoder_start_token_id``), only true for `bart-large-cnn`.
 
-"""
-
-
-@add_start_docstrings_to_callable(BART_CONFIG_ARGS_DOC)
 class BartConfig(PretrainedConfig):
     r"""
-    Configuration class for Bart. Parameters are renamed from the fairseq implementation
+    This is the configuration class to store the configuration of a :class:`~transformers.BartModel`. It is used to
+    instantiate a BART model according to the specified arguments, defining the model architecture.
+
+    Configuration objects inherit from  :class:`~transformers.PretrainedConfig` and can be used
+    to control the model outputs. Read the documentation from  :class:`~transformers.PretrainedConfig`
+    for more information.
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 50265):
+            Vocabulary size of the BERT model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.BartModel`.
+        d_model (:obj:`int`, `optional`, defaults to 1024):
+            Dimensionality of the layers and the pooler layer.
+        encoder_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of encoder layers, 6 are used for the `bart-base` model.
+        decoder_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of decoder layers, 6 are used for the `bart-base` model.
+        encoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler.
+            If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+        dropout (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        classifier_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for classifier.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 1024):
+            The maximum sequence length that this model might ever be used with.
+            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
+        init_std (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        add_bias_logits (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            This should be completed, specific to marian.
+        normalize_before (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Call layernorm before attention ops.
+        normalize_embedding (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Call layernorm after embeddings.
+        static_position_embeddings (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Don't learn positional embeddings, use sinusoidal.
+        add_final_layer_norm (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Why not add another layernorm?
+        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Scale embeddings by diving by sqrt(d_model).
+        eos_token_id (:obj:`int`, `optional`, defaults to 2)
+            End of stream token id.
+        pad_token_id (:obj:`int`, `optional`, defaults to 1)
+            Padding token id.
+        bos_token_id (:obj:`int`, `optional`, defaults to 0)
+            Beginning of stream token id.
+        encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the `LayerDrop paper
+            <see https://arxiv.org/abs/1909.11556>`__ for more details.
+        decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the `LayerDrop paper
+            <see https://arxiv.org/abs/1909.11556>`__ for more details.
+        extra_pos_embeddings: (:obj:`int`, `optional`, defaults to 2):
+            How many extra learned positional embeddings to use. Should be set to :obj:`pad_token_id+1`.
+        num_labels: (:obj:`int`, `optional`, defaults to 3):
+            The number of labels to use in :class:`~transformers.BartForSequenceClassification`.
+        is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether this is an encoder/decoder model.
+        force_bos_token_to_be_generated (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to force BOS token to be generated at step 1 (after ``decoder_start_token_id``),
+            only :obj:`True` for `bart-large-cnn`.
     """
     model_type = "bart"
 

src/transformers/configuration_deberta.py

@@ -0,0 +1,132 @@
+# coding=utf-8
+# Copyright 2020, Microsoft and the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" DeBERTa model configuration """
+
+from .configuration_utils import PretrainedConfig
+from .utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "microsoft/deberta-base": "https://s3.amazonaws.com/models.huggingface.co/bert/microsoft/deberta-base/config.json",
+    "microsoft/deberta-large": "https://s3.amazonaws.com/models.huggingface.co/bert/microsoft/deberta-large/config.json",
+}
+
+
+class DebertaConfig(PretrainedConfig):
+    r"""
+    :class:`~transformers.DebertaConfig` is the configuration class to store the configuration of a
+    :class:`~transformers.DebertaModel`.
+
+    Arguments:
+        vocab_size (:obj:`int`, `optional`, defaults to 30522):
+            Vocabulary size of the DeBERTa model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.DebertaModel` or
+            :class:`~transformers.TFDebertaModel`.
+        hidden_size (:obj:`int`, `optional`, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (:obj:`int`, `optional`, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (:obj:`int`, `optional`, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler.
+            If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"`, :obj:`"gelu"`, :obj:`"tanh"`, :obj:`"gelu_fast"`,
+            :obj:`"mish"`, :obj:`"linear"`, :obj:`"sigmoid"` and :obj:`"gelu_new"` are supported.
+        hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 512):
+            The maximum sequence length that this model might ever be used with.
+            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (:obj:`int`, `optional`, defaults to 2):
+            The vocabulary size of the :obj:`token_type_ids` passed when calling :class:`~transformers.DebertaModel` or
+            :class:`~transformers.TFDebertaModel`.
+        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (:obj:`float`, `optional`, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        relative_attention (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether use relative position encoding.
+        max_relative_positions (:obj:`int`, `optional`, defaults to 1):
+            The range of relative positions :obj:`[-max_position_embeddings, max_position_embeddings]`.
+            Use the same value as :obj:`max_position_embeddings`.
+        pad_token_id (:obj:`int`, `optional`, defaults to 0):
+            The value used to pad input_ids.
+        position_biased_input (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether add absolute position embedding to content embedding.
+        pos_att_type (:obj:`List[str]`, `optional`):
+            The type of relative position attention, it can be a combination of :obj:`["p2c", "c2p", "p2p"]`,
+            e.g. :obj:`["p2c"]`, :obj:`["p2c", "c2p"]`, :obj:`["p2c", "c2p", 'p2p"]`.
+        layer_norm_eps (:obj:`float`, optional, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+    """
+    model_type = "deberta"
+
+    def __init__(
+        self,
+        vocab_size=50265,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-7,
+        relative_attention=False,
+        max_relative_positions=-1,
+        pad_token_id=0,
+        position_biased_input=True,
+        pos_att_type=None,
+        pooler_dropout=0,
+        pooler_hidden_act="gelu",
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.relative_attention = relative_attention
+        self.max_relative_positions = max_relative_positions
+        self.pad_token_id = pad_token_id
+        self.position_biased_input = position_biased_input
+
+        # Backwards compatibility
+        if type(pos_att_type) == str:
+            pos_att_type = [x.strip() for x in pos_att_type.lower().split("|")]
+
+        self.pos_att_type = pos_att_type
+        self.vocab_size = vocab_size
+        self.layer_norm_eps = layer_norm_eps
+
+        self.pooler_hidden_size = kwargs.get("pooler_hidden_size", hidden_size)
+        self.pooler_dropout = pooler_dropout
+        self.pooler_hidden_act = pooler_hidden_act

src/transformers/configuration_gpt2.py

@@ -103,6 +103,8 @@ class GPT2Config(PretrainedConfig):
             :class:`~transformers.GPT2DoubleHeadsModel` and :class:`~transformers.TFGPT2DoubleHeadsModel`.
 
             The dropout ratio to be used after the projection and activation.
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.
 
     Example::
 
@@ -142,6 +144,7 @@ class GPT2Config(PretrainedConfig):
         summary_first_dropout=0.1,
         bos_token_id=50256,
         eos_token_id=50256,
+        gradient_checkpointing=False,
         **kwargs
     ):
         super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
@@ -164,6 +167,7 @@ class GPT2Config(PretrainedConfig):
         self.summary_activation = summary_activation
         self.summary_first_dropout = summary_first_dropout
         self.summary_proj_to_labels = summary_proj_to_labels
+        self.gradient_checkpointing = gradient_checkpointing
 
         self.bos_token_id = bos_token_id
         self.eos_token_id = eos_token_id

src/transformers/configuration_marian.py

@@ -23,4 +23,78 @@ PRETRAINED_CONFIG_ARCHIVE_MAP = {
 
 
 class MarianConfig(BartConfig):
+    """
+    This is the configuration class to store the configuration of a :class:`~transformers.MarianMTModel`. It is used to
+    instantiate a Marian model according to the specified arguments, defining the model architecture.
+
+    Configuration objects inherit from  :class:`~transformers.PretrainedConfig` and can be used
+    to control the model outputs. Read the documentation from  :class:`~transformers.PretrainedConfig`
+    for more information.
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 58101):
+            Vocabulary size of the Marian model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.MarianMTModel`.
+        d_model (:obj:`int`, `optional`, defaults to 512):
+            Dimensionality of the layers and the pooler layer.
+        encoder_layers (:obj:`int`, `optional`, defaults to 6):
+            Number of encoder layers.
+        decoder_layers (:obj:`int`, `optional`, defaults to 6):
+            Number of decoder layers.
+        encoder_attention_heads (:obj:`int`, `optional`, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (:obj:`int`, `optional`, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (:obj:`int`, `optional`, defaults to 2048):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+        encoder_ffn_dim (:obj:`int`, `optional`, defaults to 2048):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+        activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler.
+            If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+        dropout (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        classifier_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for classifier.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 512):
+            The maximum sequence length that this model might ever be used with.
+            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
+        init_std (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        add_bias_logits (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            This should be completed, specific to marian.
+        normalize_before (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Call layernorm before attention ops.
+        normalize_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Call layernorm after embeddings.
+        static_position_embeddings (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Don't learn positional embeddings, use sinusoidal.
+        add_final_layer_norm (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Why not add another layernorm?
+        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Scale embeddings by diving by sqrt(d_model).
+        eos_token_id (:obj:`int`, `optional`, defaults to 2)
+            End of stream token id.
+        pad_token_id (:obj:`int`, `optional`, defaults to 1)
+            Padding token id.
+        bos_token_id (:obj:`int`, `optional`, defaults to 0)
+            Beginning of stream token id.
+        encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the `LayerDrop paper
+            <see https://arxiv.org/abs/1909.11556>`__ for more details.
+        decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the `LayerDrop paper
+            <see https://arxiv.org/abs/1909.11556>`__ for more details.
+        extra_pos_embeddings: (:obj:`int`, `optional`, defaults to 2):
+            How many extra learned positional embeddings to use.
+        is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether this is an encoder/decoder model
+        force_bos_token_to_be_generated (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to force BOS token to be generated at step 1 (after ``decoder_start_token_id``).
+    """
+
     model_type = "marian"

src/transformers/configuration_mbart.py

@@ -27,5 +27,79 @@ MBART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
 
 
 class MBartConfig(BartConfig):
+    """
+    This is the configuration class to store the configuration of a
+    :class:`~transformers.MBartForConditionalGeneration`. It is used to
+    instantiate a BART model according to the specified arguments, defining the model architecture.
+
+    Configuration objects inherit from  :class:`~transformers.PretrainedConfig` and can be used
+    to control the model outputs. Read the documentation from  :class:`~transformers.PretrainedConfig`
+    for more information.
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 250027):
+            Vocabulary size of the MBART model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.MBartForConditionalGeneration`.
+        d_model (:obj:`int`, `optional`, defaults to 1024):
+            Dimensionality of the layers and the pooler layer.
+        encoder_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of encoder layers.
+        decoder_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of decoder layers.
+        encoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+        encoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+        activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler.
+            If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+        dropout (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        classifier_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for classifier.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 1024):
+            The maximum sequence length that this model might ever be used with.
+            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
+        init_std (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        add_bias_logits (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            This should be completed, specific to marian.
+        normalize_before (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Call layernorm before attention ops.
+        normalize_embedding (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Call layernorm after embeddings. Only True for Bart.
+        static_position_embeddings (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Don't learn positional embeddings, use sinusoidal.
+        add_final_layer_norm (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Why not add another layernorm?
+        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Scale embeddings by diving by sqrt(d_model).
+        eos_token_id (:obj:`int`, `optional`, defaults to 2)
+            End of stream token id.
+        pad_token_id (:obj:`int`, `optional`, defaults to 1)
+            Padding token id.
+        bos_token_id (:obj:`int`, `optional`, defaults to 0)
+            Beginning of stream token id.
+        encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the `LayerDrop paper
+            <see https://arxiv.org/abs/1909.11556>`__ for more details.
+        decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the `LayerDrop paper
+            <see https://arxiv.org/abs/1909.11556>`__ for more details.
+        extra_pos_embeddings: (:obj:`int`, `optional`, defaults to 2):
+            How many extra learned positional embeddings to use. Should be equal to :obj:`pad_token_id+1`.
+        is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether this is an encoder/decoder model
+        force_bos_token_to_be_generated (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to force BOS token to be generated at step 1 (after ``decoder_start_token_id``).
+    """
+
     model_type = "mbart"
-    """See real config values at https://s3.amazonaws.com/models.huggingface.co/bert/facebook/mbart-large-en-ro/config.json."""

src/transformers/configuration_pegasus.py

@@ -14,8 +14,7 @@
 # limitations under the License.
 """ PEGASUS model configuration """
 
-from .configuration_bart import BART_CONFIG_ARGS_DOC, BartConfig
-from .file_utils import add_start_docstrings_to_callable
+from .configuration_bart import BartConfig
 from .utils import logging
 
 
@@ -66,11 +65,81 @@ task_specific_params = {
 }
 
 
-@add_start_docstrings_to_callable(BART_CONFIG_ARGS_DOC)
 class PegasusConfig(BartConfig):
-    r"""
-    :class:`~transformers.PegasusConfig` is the configuration class to store the configuration of a
-    `PegasusModel`.
     """
+    This is the configuration class to store the configuration of a
+    :class:`~transformers.PegasusForConditionalGeneration`. It is used to
+    instantiate a Pegasus model according to the specified arguments, defining the model architecture.
+
+    Configuration objects inherit from  :class:`~transformers.PretrainedConfig` and can be used
+    to control the model outputs. Read the documentation from  :class:`~transformers.PretrainedConfig`
+    for more information.
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 96103):
+            Vocabulary size of the Pegasus model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.PegasusForConditionalGeneration`.
+        d_model (:obj:`int`, `optional`, defaults to 1024):
+            Dimensionality of the layers and the pooler layer.
+        encoder_layers (:obj:`int`, `optional`, defaults to 16):
+            Number of encoder layers.
+        decoder_layers (:obj:`int`, `optional`, defaults to 16):
+            Number of decoder layers.
+        encoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+        encoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
+        activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler.
+            If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+        dropout (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        classifier_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for classifier.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 1024):
+            The maximum sequence length that this model might ever be used with.
+            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
+        init_std (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        add_bias_logits (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            This should be completed, specific to marian.
+        normalize_before (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Call layernorm before attention ops.
+        normalize_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Call layernorm after embeddings.
+        static_position_embeddings (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Don't learn positional embeddings, use sinusoidal.
+        add_final_layer_norm (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Why not add another layernorm?
+        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Scale embeddings by diving by sqrt(d_model).
+        eos_token_id (:obj:`int`, `optional`, defaults to 2)
+            End of stream token id.
+        pad_token_id (:obj:`int`, `optional`, defaults to 1)
+            Padding token id.
+        bos_token_id (:obj:`int`, `optional`, defaults to 0)
+            Beginning of stream token id.
+        encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the `LayerDrop paper
+            <see https://arxiv.org/abs/1909.11556>`__ for more details.
+        decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the `LayerDrop paper
+            <see https://arxiv.org/abs/1909.11556>`__ for more details.
+        extra_pos_embeddings: (:obj:`int`, `optional`, defaults to 2):
+            How many extra learned positional embeddings to use. Should be pad_token_id+1 for bart.
+        is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether this is an encoder/decoder model
+        force_bos_token_to_be_generated (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to force BOS token to be generated at step 1 (after ``decoder_start_token_id``).
+    """
+
     model_type = "pegasus"
     # The implementation of the config object is in BartConfig

src/transformers/configuration_squeezebert.py

@@ -0,0 +1,153 @@
+# coding=utf-8
+# Copyright 2020 The SqueezeBert authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" SqueezeBERT model configuration """
+
+from .configuration_utils import PretrainedConfig
+from .utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "squeezebert/squeezebert-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/squeezebert/squeezebert-uncased/config.json",
+    "squeezebert/squeezebert-mnli": "https://s3.amazonaws.com/models.huggingface.co/bert/squeezebert/squeezebert-mnli/config.json",
+    "squeezebert/squeezebert-mnli-headless": "https://s3.amazonaws.com/models.huggingface.co/bert/squeezebert/squeezebert-mnli-headless/config.json",
+}
+
+
+class SqueezeBertConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.SqueezeBertModel`.
+    It is used to instantiate a SqueezeBERT model according to the specified arguments, defining the model
+    architecture.
+
+    Configuration objects inherit from  :class:`~transformers.PretrainedConfig` and can be used
+    to control the model outputs. Read the documentation from  :class:`~transformers.PretrainedConfig`
+    for more information.
+
+
+    Args:
+        vocab_size (:obj:`int`, optional, defaults to 30522):
+            Vocabulary size of the SqueezeBERT model. Defines the different tokens that
+            can be represented by the :obj:`inputs_ids` passed to the forward method of :class:`~transformers.SqueezeBertModel`.
+        hidden_size (:obj:`int`, optional, defaults to 512):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (:obj:`int`, optional, defaults to 24):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (:obj:`int`, optional, defaults to 4):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (:obj:`int`, optional, defaults to 512):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        hidden_act (:obj:`str` or :obj:`function`, optional, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler.
+            If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"`, :obj:`"gelu_new"`, :obj:`"gelu_fast"`,
+            and :obj:`"mish"` are supported.
+        hidden_dropout_prob (:obj:`float`, optional, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (:obj:`float`, optional, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (:obj:`int`, optional, defaults to 512):
+            The maximum sequence length that this model might ever be used with.
+            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (:obj:`int`, optional, defaults to 2):
+            The vocabulary size of the :obj:`token_type_ids` passed into :class:`~transformers.SqueezeBertModel`.
+        initializer_range (:obj:`float`, optional, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (:obj:`float`, optional, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+
+        pad_token_id (:obj:`int`, optional, defaults to 0):
+            The ID of the token in the word embedding to use as padding.
+        embedding_size (:obj:`int`, optional, defaults to 128):
+            The dimension of the word embedding vectors.
+
+        q_groups (:obj:`int`, optional, defaults to 4):
+            The number of groups in Q layer
+        k_groups (:obj:`int`, optional, defaults to 4):
+            The number of groups in K layer
+        v_groups (:obj:`int`, optional, defaults to 4):
+            The number of groups in V layer
+        post_attention_groups (:obj:`int`, optional, defaults to 1):
+            The number of groups in the first feed forward network layer
+        intermediate_groups (:obj:`int`, optional, defaults to 4):
+            The number of groups in the second feed forward network layer
+        output_groups (:obj:`int`, optional, defaults to 4):
+            The number of groups in the third feed forward network layer
+
+    Example:
+
+        >>> from transformers import SqueezeBertModel, SqueezeBertConfig
+
+        >>> # Initializing a SqueezeBERT configuration
+        >>> configuration = SqueezeBertConfig()
+
+        >>> # Initializing a model from the configuration above
+        >>> model = SqueezeBertModel(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+
+    Attributes:
+        pretrained_config_archive_map (Dict[str, str]):
+            A dictionary containing all the available pre-trained checkpoints.
+    """
+    pretrained_config_archive_map = SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+    model_type = "squeezebert"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        embedding_size=768,
+        q_groups=4,
+        k_groups=4,
+        v_groups=4,
+        post_attention_groups=1,
+        intermediate_groups=4,
+        output_groups=4,
+        **kwargs
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.embedding_size = embedding_size
+        self.q_groups = q_groups
+        self.k_groups = k_groups
+        self.v_groups = v_groups
+        self.post_attention_groups = post_attention_groups
+        self.intermediate_groups = intermediate_groups
+        self.output_groups = output_groups

src/transformers/configuration_utils.py

@@ -114,8 +114,8 @@ class PretrainedConfig(object):
           model pretrained weights.
         - **finetuning_task** (:obj:`str`, `optional`) -- Name of the task used to fine-tune the model. This can be
           used when converting from an original (TensorFlow or PyTorch) checkpoint.
-        - **id2label** (:obj:`List[str]`, `optional`) -- A map from index (for instance prediction index, or target
-          index) to label.
+        - **id2label** (:obj:`Dict[int, str]`, `optional`) -- A map from index (for instance prediction index, or
+          target index) to label.
         - **label2id** (:obj:`Dict[str, int]`, `optional`) -- A map from label to index for the model.
         - **num_labels** (:obj:`int`, `optional`) -- Number of labels to use in the last layer added to the model,
           typically for a classification task.

src/transformers/convert_slow_tokenizer.py

@@ -0,0 +1,546 @@
+from typing import Dict, List, Tuple
+
+from sentencepiece import SentencePieceProcessor
+from tokenizers import Tokenizer, decoders, normalizers, pre_tokenizers, processors
+from tokenizers.models import BPE, Unigram, WordPiece
+
+# from transformers.tokenization_openai import OpenAIGPTTokenizer
+from transformers.utils import sentencepiece_model_pb2 as model
+
+
+class SentencePieceExtractor:
+    """
+    Extractor implementation for SentencePiece trained models.
+    https://github.com/google/sentencepiece
+    """
+
+    def __init__(self, model: str):
+        # Get SentencePiece
+        self.sp = SentencePieceProcessor()
+        self.sp.Load(model)
+
+    def extract(self) -> Tuple[Dict[str, int], List[Tuple]]:
+        sp = self.sp
+        vocab = {sp.id_to_piece(index): index for index in range(sp.GetPieceSize())}
+
+        # Merges
+        merges = []
+        for piece_l in vocab.keys():
+            for piece_r in vocab.keys():
+                merge = f"{piece_l}{piece_r}"
+                piece_id = vocab.get(merge, None)
+                if piece_id:
+                    merges += [(piece_l, piece_r, piece_id)]
+        merges = sorted(merges, key=lambda val: val[2])
+        merges = [(val[0], val[1]) for val in merges]
+
+        return vocab, merges
+
+
+def check_number_comma(piece: str) -> bool:
+    return len(piece) < 2 or piece[-1] != "," or not piece[-2].isdigit()
+
+
+def get_proto(filename: str):
+    m = model.ModelProto()
+    m.ParseFromString(open(filename, "rb").read())
+    return m
+
+
+class Converter:
+    def __init__(self, original_tokenizer):
+        self.original_tokenizer = original_tokenizer
+
+    def converted(self) -> Tokenizer:
+        raise NotImplementedError()
+
+
+class BertConverter(Converter):
+    def converted(self) -> Tokenizer:
+        vocab = self.original_tokenizer.vocab
+        tokenizer = Tokenizer(WordPiece(vocab, unk_token=str(self.original_tokenizer.unk_token)))
+
+        # # Let the tokenizer know about special tokens if they are part of the vocab
+        # if tokenizer.token_to_id(str(self.original_tokenizer.unk_token)) is not None:
+        #     tokenizer.add_special_tokens([str(self.original_tokenizer.unk_token)])
+        # if tokenizer.token_to_id(str(self.original_tokenizer.sep_token)) is not None:
+        #     tokenizer.add_special_tokens([str(self.original_tokenizer.sep_token)])
+        # if tokenizer.token_to_id(str(self.original_tokenizer.cls_token)) is not None:
+        #     tokenizer.add_special_tokens([str(self.original_tokenizer.cls_token)])
+        # if tokenizer.token_to_id(str(self.original_tokenizer.pad_token)) is not None:
+        #     tokenizer.add_special_tokens([str(self.original_tokenizer.pad_token)])
+        # if tokenizer.token_to_id(str(self.original_tokenizer.mask_token)) is not None:
+        #     tokenizer.add_special_tokens([str(self.original_tokenizer.mask_token)])
+
+        tokenize_chinese_chars = False
+        strip_accents = False
+        do_lower_case = False
+        if hasattr(self.original_tokenizer, "basic_tokenizer"):
+            tokenize_chinese_chars = self.original_tokenizer.basic_tokenizer.tokenize_chinese_chars
+            strip_accents = self.original_tokenizer.basic_tokenizer.strip_accents
+            do_lower_case = self.original_tokenizer.basic_tokenizer.do_lower_case
+
+        tokenizer.normalizer = normalizers.BertNormalizer(
+            clean_text=True,
+            handle_chinese_chars=tokenize_chinese_chars,
+            strip_accents=strip_accents,
+            lowercase=do_lower_case,
+        )
+        tokenizer.pre_tokenizer = pre_tokenizers.BertPreTokenizer()
+
+        cls = str(self.original_tokenizer.cls_token)
+        sep = str(self.original_tokenizer.sep_token)
+        cls_token_id = self.original_tokenizer.cls_token_id
+        sep_token_id = self.original_tokenizer.sep_token_id
+
+        tokenizer.post_processor = processors.TemplateProcessing(
+            single=f"{cls}:0 $A:0 {sep}:0",
+            pair=f"{cls}:0 $A:0 {sep}:0 $B:1 {sep}:1",
+            special_tokens=[
+                (cls, cls_token_id),
+                (sep, sep_token_id),
+            ],
+        )
+        tokenizer.decoder = decoders.WordPiece(prefix="##")
+
+        return tokenizer
+
+
+class FunnelConverter(Converter):
+    def converted(self) -> Tokenizer:
+        vocab = self.original_tokenizer.vocab
+        tokenizer = Tokenizer(WordPiece(vocab, unk_token=str(self.original_tokenizer.unk_token)))
+
+        # # Let the tokenizer know about special tokens if they are part of the vocab
+        # if tokenizer.token_to_id(str(self.original_tokenizer.unk_token)) is not None:
+        #     tokenizer.add_special_tokens([str(self.original_tokenizer.unk_token)])
+        # if tokenizer.token_to_id(str(self.original_tokenizer.sep_token)) is not None:
+        #     tokenizer.add_special_tokens([str(self.original_tokenizer.sep_token)])
+        # if tokenizer.token_to_id(str(self.original_tokenizer.cls_token)) is not None:
+        #     tokenizer.add_special_tokens([str(self.original_tokenizer.cls_token)])
+        # if tokenizer.token_to_id(str(self.original_tokenizer.pad_token)) is not None:
+        #     tokenizer.add_special_tokens([str(self.original_tokenizer.pad_token)])
+        # if tokenizer.token_to_id(str(self.original_tokenizer.mask_token)) is not None:
+        #     tokenizer.add_special_tokens([str(self.original_tokenizer.mask_token)])
+
+        tokenize_chinese_chars = False
+        strip_accents = False
+        do_lower_case = False
+        if hasattr(self.original_tokenizer, "basic_tokenizer"):
+            tokenize_chinese_chars = self.original_tokenizer.basic_tokenizer.tokenize_chinese_chars
+            strip_accents = self.original_tokenizer.basic_tokenizer.strip_accents
+            do_lower_case = self.original_tokenizer.basic_tokenizer.do_lower_case
+
+        tokenizer.normalizer = normalizers.BertNormalizer(
+            clean_text=True,
+            handle_chinese_chars=tokenize_chinese_chars,
+            strip_accents=strip_accents,
+            lowercase=do_lower_case,
+        )
+        tokenizer.pre_tokenizer = pre_tokenizers.BertPreTokenizer()
+
+        cls = str(self.original_tokenizer.cls_token)
+        sep = str(self.original_tokenizer.sep_token)
+        cls_token_id = self.original_tokenizer.cls_token_id
+        sep_token_id = self.original_tokenizer.sep_token_id
+
+        tokenizer.post_processor = processors.TemplateProcessing(
+            single=f"{cls}:2 $A:0 {sep}:0",  # token_type_id is 2 for Funnel transformer
+            pair=f"{cls}:2 $A:0 {sep}:0 $B:1 {sep}:1",
+            special_tokens=[
+                (cls, cls_token_id),
+                (sep, sep_token_id),
+            ],
+        )
+        tokenizer.decoder = decoders.WordPiece(prefix="##")
+
+        return tokenizer
+
+
+class OpenAIGPTConverter(Converter):
+    def converted(self) -> Tokenizer:
+        vocab = self.original_tokenizer.encoder
+        merges = list(self.original_tokenizer.bpe_ranks.keys())
+        unk_token = self.original_tokenizer.unk_token
+
+        tokenizer = Tokenizer(
+            BPE(
+                vocab=vocab,
+                merges=merges,
+                dropout=None,
+                unk_token=str(unk_token),
+                end_of_word_suffix="</w>",
+                fuse_unk=False,
+            )
+        )
+
+        if tokenizer.token_to_id(str(unk_token)) is not None:
+            tokenizer.add_special_tokens([str(unk_token)])
+
+        tokenizer.normalizer = normalizers.BertNormalizer(lowercase=True)
+        tokenizer.pre_tokenizer = pre_tokenizers.BertPreTokenizer()
+        tokenizer.decoder = decoders.BPEDecoder(suffix="</w>")
+
+        return tokenizer
+
+
+class GPT2Converter(Converter):
+    def converted(self) -> Tokenizer:
+        vocab = self.original_tokenizer.encoder
+        merges = list(self.original_tokenizer.bpe_ranks.keys())
+
+        tokenizer = Tokenizer(
+            BPE(
+                vocab=vocab,
+                merges=merges,
+                dropout=None,
+                continuing_subword_prefix="",
+                end_of_word_suffix="",
+                fuse_unk=False,
+            )
+        )
+
+        tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=self.original_tokenizer.add_prefix_space)
+        tokenizer.decoder = decoders.ByteLevel()
+        tokenizer.post_processor = processors.ByteLevel(trim_offsets=False)
+
+        return tokenizer
+
+
+class RobertaConverter(Converter):
+    def converted(self) -> Tokenizer:
+        ot = self.original_tokenizer
+        vocab = ot.encoder
+        merges = list(ot.bpe_ranks.keys())
+
+        tokenizer = Tokenizer(
+            BPE(
+                vocab=vocab,
+                merges=merges,
+                dropout=None,
+                continuing_subword_prefix="",
+                end_of_word_suffix="",
+                fuse_unk=False,
+            )
+        )
+
+        tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=ot.add_prefix_space)
+        tokenizer.decoder = decoders.ByteLevel()
+        tokenizer.post_processor = processors.RobertaProcessing(
+            sep=(ot.sep_token, ot.sep_token_id),
+            cls=(ot.cls_token, ot.cls_token_id),
+            add_prefix_space=ot.add_prefix_space,
+            trim_offsets=True,  # True by default on Roberta (historical)
+        )
+
+        return tokenizer
+
+
+class SpmConverter(Converter):
+    def __init__(self, *args):
+        super().__init__(*args)
+        self.proto = get_proto(self.original_tokenizer.vocab_file)
+
+    def vocab(self, proto):
+        return [(piece.piece, piece.score) for piece in proto.pieces]
+
+    def unk_id(self, proto):
+        return proto.trainer_spec.unk_id
+
+    def tokenizer(self, proto):
+        model_type = proto.trainer_spec.model_type
+        vocab = self.vocab(proto)
+        unk_id = self.unk_id(proto)
+
+        if model_type == 1:
+            tokenizer = Tokenizer(Unigram(vocab, unk_id))
+        elif model_type == 2:
+            vocab, merges = SentencePieceExtractor(self.original_tokenizer.vocab_file).extract()
+            tokenizer = Tokenizer(
+                BPE(
+                    vocab,
+                    merges,
+                    unk_token=proto.trainer_spec.unk_piece,
+                    fuse_unk=True,
+                )
+            )
+        else:
+            raise Exception(
+                "You're trying to run a `Unigram` model but you're file was trained with a different algorithm"
+            )
+
+        return tokenizer
+
+    def normalizer(self, proto):
+        precompiled_charsmap = proto.normalizer_spec.precompiled_charsmap
+        return normalizers.Precompiled(precompiled_charsmap)
+
+    def post_processor(self):
+        return None
+
+    def converted(self) -> Tokenizer:
+        tokenizer = self.tokenizer(self.proto)
+
+        # Tokenizer assemble
+        tokenizer.normalizer = self.normalizer(self.proto)
+
+        replacement = "▁"
+        add_prefix_space = True
+        tokenizer.pre_tokenizer = pre_tokenizers.Sequence(
+            [
+                pre_tokenizers.WhitespaceSplit(),
+                pre_tokenizers.Metaspace(replacement=replacement, add_prefix_space=add_prefix_space),
+            ]
+        )
+        tokenizer.decoder = decoders.Metaspace(replacement=replacement, add_prefix_space=add_prefix_space)
+        post_processor = self.post_processor()
+        if post_processor:
+            tokenizer.post_processor = post_processor
+
+        return tokenizer
+
+
+class AlbertConverter(SpmConverter):
+    def vocab(self, proto):
+        return [
+            (piece.piece, piece.score) if check_number_comma(piece.piece) else (piece.piece, piece.score - 100)
+            for piece in proto.pieces
+        ]
+
+    def normalizer(self, proto):
+        list_normalizers = [normalizers.Replace("``", '"'), normalizers.Replace("''", '"')]
+        if not self.original_tokenizer.keep_accents:
+            list_normalizers.append(normalizers.NFKD())
+            list_normalizers.append(normalizers.StripAccents())
+        if self.original_tokenizer.do_lower_case:
+            list_normalizers.append(normalizers.Lowercase())
+
+        precompiled_charsmap = proto.normalizer_spec.precompiled_charsmap
+        list_normalizers.append(normalizers.Precompiled(precompiled_charsmap))
+        return normalizers.Sequence(list_normalizers)
+
+    def post_processor(self):
+        return processors.TemplateProcessing(
+            single="[CLS]:0 $A:0 [SEP]:0",
+            pair="[CLS]:0 $A:0 [SEP]:0 $B:1 [SEP]:1",
+            special_tokens=[
+                ("[CLS]", self.original_tokenizer.convert_tokens_to_ids("[CLS]")),
+                ("[SEP]", self.original_tokenizer.convert_tokens_to_ids("[SEP]")),
+            ],
+        )
+
+
+class CamembertConverter(SpmConverter):
+    def vocab(self, proto):
+        vocab = [
+            ("<s>NOTUSED", 0.0),
+            ("<pad>", 0.0),
+            ("</s>NOTUSED", 0.0),
+            ("<unk>", 0.0),
+        ]
+        # We down-grade the original SentencePiece by -100 to avoid using it and use our added token instead
+        vocab += [(piece.piece, piece.score if i != 0 else piece.score - 100) for i, piece in enumerate(proto.pieces)]
+        vocab += [("<mask>", 0.0)]
+        return vocab
+
+    def unk_id(self, proto):
+        # See vocab unk position
+        return 3
+
+    def post_processor(self):
+        return processors.TemplateProcessing(
+            single="<s> $A </s>",
+            pair="<s> $A </s> </s> $B </s>",
+            special_tokens=[
+                ("<s>", self.original_tokenizer.convert_tokens_to_ids("<s>")),
+                ("</s>", self.original_tokenizer.convert_tokens_to_ids("</s>")),
+            ],
+        )
+
+
+class MBartConverter(SpmConverter):
+    def vocab(self, proto):
+        vocab = [
+            ("<s>", 0.0),
+            ("<pad>", 0.0),
+            ("</s>", 0.0),
+            ("<unk>", 0.0),
+        ]
+        vocab += [(piece.piece, piece.score) for piece in proto.pieces[3:]]
+        vocab += [
+            ("ar_AR", 0.0),
+            ("cs_CZ", 0.0),
+            ("de_DE", 0.0),
+            ("en_XX", 0.0),
+            ("es_XX", 0.0),
+            ("et_EE", 0.0),
+            ("fi_FI", 0.0),
+            ("fr_XX", 0.0),
+            ("gu_IN", 0.0),
+            ("hi_IN", 0.0),
+            ("it_IT", 0.0),
+            ("ja_XX", 0.0),
+            ("kk_KZ", 0.0),
+            ("ko_KR", 0.0),
+            ("lt_LT", 0.0),
+            ("lv_LV", 0.0),
+            ("my_MM", 0.0),
+            ("ne_NP", 0.0),
+            ("nl_XX", 0.0),
+            ("ro_RO", 0.0),
+            ("ru_RU", 0.0),
+            ("si_LK", 0.0),
+            ("tr_TR", 0.0),
+            ("vi_VN", 0.0),
+            ("zh_CN", 0.0),
+        ]
+        vocab += [("<mask>", 0.0)]
+        return vocab
+
+    def unk_id(self, proto):
+        return 3
+
+    def post_processor(self):
+        return processors.TemplateProcessing(
+            single="$A </s> en_XX",
+            pair="$A $B </s> en_XX",
+            special_tokens=[
+                ("en_XX", self.original_tokenizer.convert_tokens_to_ids("en_XX")),
+                ("</s>", self.original_tokenizer.convert_tokens_to_ids("</s>")),
+            ],
+        )
+
+
+class XLMRobertaConverter(SpmConverter):
+    def vocab(self, proto):
+        vocab = [
+            ("<s>", 0.0),
+            ("<pad>", 0.0),
+            ("</s>", 0.0),
+            ("<unk>", 0.0),
+        ]
+        vocab += [(piece.piece, piece.score) for piece in proto.pieces[3:]]
+        vocab += [("<mask>", 0.0)]
+        return vocab
+
+    def unk_id(self, proto):
+        unk_id = 3
+        return unk_id
+
+    def post_processor(self):
+        return processors.TemplateProcessing(
+            single="<s> $A </s>",
+            pair="<s> $A </s> </s> $B </s>",
+            special_tokens=[
+                ("<s>", self.original_tokenizer.convert_tokens_to_ids("<s>")),
+                ("</s>", self.original_tokenizer.convert_tokens_to_ids("</s>")),
+            ],
+        )
+
+
+class XLNetConverter(SpmConverter):
+    def vocab(self, proto):
+        return [
+            (piece.piece, piece.score) if check_number_comma(piece.piece) else (piece.piece, piece.score - 100)
+            for piece in proto.pieces
+        ]
+
+    def normalizer(self, proto):
+        list_normalizers = [normalizers.Replace("``", '"'), normalizers.Replace("''", '"')]
+        if not self.original_tokenizer.keep_accents:
+            list_normalizers.append(normalizers.NFKD())
+            list_normalizers.append(normalizers.StripAccents())
+        if self.original_tokenizer.do_lower_case:
+            list_normalizers.append(normalizers.Lowercase())
+
+        precompiled_charsmap = proto.normalizer_spec.precompiled_charsmap
+        list_normalizers.append(normalizers.Precompiled(precompiled_charsmap))
+        return normalizers.Sequence(list_normalizers)
+
+    def post_processor(self):
+        return processors.TemplateProcessing(
+            single="$A:0 <sep>:0 <cls>:2",
+            pair="$A:0 <sep>:0 $B:1 <sep>:1 <cls>:2",
+            special_tokens=[
+                ("<sep>", self.original_tokenizer.convert_tokens_to_ids("<sep>")),
+                ("<cls>", self.original_tokenizer.convert_tokens_to_ids("<cls>")),
+            ],
+        )
+
+
+class ReformerConverter(SpmConverter):
+    pass
+
+
+class BertGenerationConverter(SpmConverter):
+    pass
+
+
+class PegasusConverter(SpmConverter):
+    def vocab(self, proto):
+        vocab = [
+            (self.original_tokenizer.pad_token, 0),
+            (self.original_tokenizer.eos_token, 0),
+        ]
+        vocab += [(f"unk_{i}", -100) for i in range(2, 2 + self.original_tokenizer.offset)]
+        vocab += [(piece.piece, piece.score) for piece in proto.pieces[2:]]
+        return vocab
+
+    def unk_id(self, proto):
+        return proto.trainer_spec.unk_id + self.original_tokenizer.offset
+
+    def post_processor(self):
+        eos = self.original_tokenizer.eos_token
+        return processors.TemplateProcessing(
+            single=["$A", eos],
+            pair=["$A", "$B", eos],
+            special_tokens=[
+                (eos, self.original_tokenizer.eos_token_id),
+            ],
+        )
+
+
+class T5Converter(SpmConverter):
+    def vocab(self, proto):
+        num_extra_ids = self.original_tokenizer._extra_ids
+        vocab = [(piece.piece, piece.score) for piece in proto.pieces]
+        vocab += [("<extra_id_{}>".format(i), 0.0) for i in range(num_extra_ids - 1, -1, -1)]
+        return vocab
+
+    def post_processor(self):
+        return processors.TemplateProcessing(
+            single=["$A", "</s>"],
+            pair=["$A", "</s>", "$B", "</s>"],
+            special_tokens=[
+                ("</s>", self.original_tokenizer.convert_tokens_to_ids("</s>")),
+            ],
+        )
+
+
+CONVERTERS = {
+    "AlbertTokenizer": AlbertConverter,
+    "BertTokenizer": BertConverter,
+    "BertGenerationTokenizer": BertGenerationConverter,
+    "BartTokenizer": RobertaConverter,
+    "CamembertTokenizer": CamembertConverter,
+    "DistilBertTokenizer": BertConverter,
+    "DPRReaderTokenizer": BertConverter,
+    "DPRQuestionEncoderTokenizer": BertConverter,
+    "DPRContextEncoderTokenizer": BertConverter,
+    "FunnelTokenizer": FunnelConverter,
+    "GPT2Tokenizer": GPT2Converter,
+    "LxmertTokenizer": BertConverter,
+    "MBartTokenizer": MBartConverter,
+    "OpenAIGPTTokenizer": OpenAIGPTConverter,
+    "PegasusTokenizer": PegasusConverter,
+    "ReformerTokenizer": ReformerConverter,
+    "RobertaTokenizer": RobertaConverter,
+    "T5Tokenizer": T5Converter,
+    "XLMRobertaTokenizer": XLMRobertaConverter,
+    "XLNetTokenizer": XLNetConverter,
+}
+
+
+def convert_slow_tokenizer(transformer_tokenizer) -> Tokenizer:
+    converter_class = CONVERTERS[transformer_tokenizer.__class__.__name__]
+    return converter_class(transformer_tokenizer).converted()

src/transformers/data/__init__.py

@@ -2,7 +2,7 @@
 # There's no way to ignore "F401 '...' imported but unused" warnings in this
 # module, but to preserve other warnings. So, don't check this module at all.
 
-from .metrics import is_sklearn_available
+from .metrics import glue_compute_metrics, xnli_compute_metrics
 from .processors import (
     DataProcessor,
     InputExample,
@@ -21,7 +21,3 @@ from .processors import (
     xnli_processors,
     xnli_tasks_num_labels,
 )
-
-
-if is_sklearn_available():
-    from .metrics import glue_compute_metrics, xnli_compute_metrics

src/transformers/data/metrics/__init__.py

@@ -14,77 +14,75 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-try:
+from ...file_utils import is_sklearn_available, requires_sklearn
+
+
+if is_sklearn_available():
     from sklearn.metrics import f1_score, matthews_corrcoef
 
     from scipy.stats import pearsonr, spearmanr
 
-    _has_sklearn = True
-except (AttributeError, ImportError):
-    _has_sklearn = False
+
+def simple_accuracy(preds, labels):
+    requires_sklearn(simple_accuracy)
+    return (preds == labels).mean()
 
 
-def is_sklearn_available():
-    return _has_sklearn
+def acc_and_f1(preds, labels):
+    requires_sklearn(acc_and_f1)
+    acc = simple_accuracy(preds, labels)
+    f1 = f1_score(y_true=labels, y_pred=preds)
+    return {
+        "acc": acc,
+        "f1": f1,
+        "acc_and_f1": (acc + f1) / 2,
+    }
 
 
-if _has_sklearn:
+def pearson_and_spearman(preds, labels):
+    requires_sklearn(pearson_and_spearman)
+    pearson_corr = pearsonr(preds, labels)[0]
+    spearman_corr = spearmanr(preds, labels)[0]
+    return {
+        "pearson": pearson_corr,
+        "spearmanr": spearman_corr,
+        "corr": (pearson_corr + spearman_corr) / 2,
+    }
 
-    def simple_accuracy(preds, labels):
-        return (preds == labels).mean()
 
-    def acc_and_f1(preds, labels):
-        acc = simple_accuracy(preds, labels)
-        f1 = f1_score(y_true=labels, y_pred=preds)
-        return {
-            "acc": acc,
-            "f1": f1,
-            "acc_and_f1": (acc + f1) / 2,
-        }
+def glue_compute_metrics(task_name, preds, labels):
+    requires_sklearn(glue_compute_metrics)
+    assert len(preds) == len(labels), f"Predictions and labels have mismatched lengths {len(preds)} and {len(labels)}"
+    if task_name == "cola":
+        return {"mcc": matthews_corrcoef(labels, preds)}
+    elif task_name == "sst-2":
+        return {"acc": simple_accuracy(preds, labels)}
+    elif task_name == "mrpc":
+        return acc_and_f1(preds, labels)
+    elif task_name == "sts-b":
+        return pearson_and_spearman(preds, labels)
+    elif task_name == "qqp":
+        return acc_and_f1(preds, labels)
+    elif task_name == "mnli":
+        return {"mnli/acc": simple_accuracy(preds, labels)}
+    elif task_name == "mnli-mm":
+        return {"mnli-mm/acc": simple_accuracy(preds, labels)}
+    elif task_name == "qnli":
+        return {"acc": simple_accuracy(preds, labels)}
+    elif task_name == "rte":
+        return {"acc": simple_accuracy(preds, labels)}
+    elif task_name == "wnli":
+        return {"acc": simple_accuracy(preds, labels)}
+    elif task_name == "hans":
+        return {"acc": simple_accuracy(preds, labels)}
+    else:
+        raise KeyError(task_name)
 
-    def pearson_and_spearman(preds, labels):
-        pearson_corr = pearsonr(preds, labels)[0]
-        spearman_corr = spearmanr(preds, labels)[0]
-        return {
-            "pearson": pearson_corr,
-            "spearmanr": spearman_corr,
-            "corr": (pearson_corr + spearman_corr) / 2,
-        }
 
-    def glue_compute_metrics(task_name, preds, labels):
-        assert len(preds) == len(
-            labels
-        ), f"Predictions and labels have mismatched lengths {len(preds)} and {len(labels)}"
-        if task_name == "cola":
-            return {"mcc": matthews_corrcoef(labels, preds)}
-        elif task_name == "sst-2":
-            return {"acc": simple_accuracy(preds, labels)}
-        elif task_name == "mrpc":
-            return acc_and_f1(preds, labels)
-        elif task_name == "sts-b":
-            return pearson_and_spearman(preds, labels)
-        elif task_name == "qqp":
-            return acc_and_f1(preds, labels)
-        elif task_name == "mnli":
-            return {"mnli/acc": simple_accuracy(preds, labels)}
-        elif task_name == "mnli-mm":
-            return {"mnli-mm/acc": simple_accuracy(preds, labels)}
-        elif task_name == "qnli":
-            return {"acc": simple_accuracy(preds, labels)}
-        elif task_name == "rte":
-            return {"acc": simple_accuracy(preds, labels)}
-        elif task_name == "wnli":
-            return {"acc": simple_accuracy(preds, labels)}
-        elif task_name == "hans":
-            return {"acc": simple_accuracy(preds, labels)}
-        else:
-            raise KeyError(task_name)
-
-    def xnli_compute_metrics(task_name, preds, labels):
-        assert len(preds) == len(
-            labels
-        ), f"Predictions and labels have mismatched lengths {len(preds)} and {len(labels)}"
-        if task_name == "xnli":
-            return {"acc": simple_accuracy(preds, labels)}
-        else:
-            raise KeyError(task_name)
+def xnli_compute_metrics(task_name, preds, labels):
+    requires_sklearn(xnli_compute_metrics)
+    assert len(preds) == len(labels), f"Predictions and labels have mismatched lengths {len(preds)} and {len(labels)}"
+    if task_name == "xnli":
+        return {"acc": simple_accuracy(preds, labels)}
+    else:
+        raise KeyError(task_name)

src/transformers/data/processors/squad.py

@@ -7,7 +7,10 @@ import numpy as np
 from tqdm import tqdm
 
 from ...file_utils import is_tf_available, is_torch_available
+from ...tokenization_bart import BartTokenizer
 from ...tokenization_bert import whitespace_tokenize
+from ...tokenization_longformer import LongformerTokenizer
+from ...tokenization_roberta import RobertaTokenizer
 from ...tokenization_utils_base import TruncationStrategy
 from ...utils import logging
 from .utils import DataProcessor
@@ -109,7 +112,10 @@ def squad_convert_example_to_features(
     all_doc_tokens = []
     for (i, token) in enumerate(example.doc_tokens):
         orig_to_tok_index.append(len(all_doc_tokens))
-        sub_tokens = tokenizer.tokenize(token)
+        if isinstance(tokenizer, (RobertaTokenizer, LongformerTokenizer, BartTokenizer)):
+            sub_tokens = tokenizer.tokenize(token, add_prefix_space=True)
+        else:
+            sub_tokens = tokenizer.tokenize(token)
         for sub_token in sub_tokens:
             tok_to_orig_index.append(i)
             all_doc_tokens.append(sub_token)

src/transformers/file_utils.py

@@ -68,8 +68,12 @@ except (ImportError, AssertionError):
 try:
     import datasets  # noqa: F401
 
-    _datasets_available = True
-    logger.debug(f"Succesfully imported datasets version {datasets.__version__}")
+    # Check we're not importing a "datasets" directory somewhere
+    _datasets_available = hasattr(datasets, "__version__") and hasattr(datasets, "load_dataset")
+    if _datasets_available:
+        logger.debug(f"Succesfully imported datasets version {datasets.__version__}")
+    else:
+        logger.debug("Imported a datasets object but this doesn't seem to be the 🤗 datasets library.")
 
 except ImportError:
     _datasets_available = False
@@ -129,6 +133,33 @@ try:
 except ImportError:
     _faiss_available = False
 
+try:
+    import sklearn.metrics  # noqa: F401
+
+    import scipy.stats  # noqa: F401
+
+    _has_sklearn = True
+except (AttributeError, ImportError):
+    _has_sklearn = False
+
+
+try:
+    import sentencepiece  # noqa: F401
+
+    _sentencepiece_available = True
+
+except ImportError:
+    _sentencepiece_available = False
+
+
+try:
+    import tokenizers  # noqa: F401
+
+    _tokenizers_available = True
+
+except ImportError:
+    _tokenizers_available = False
+
 
 default_cache_path = os.path.join(torch_cache_home, "transformers")
 
@@ -190,6 +221,96 @@ def is_faiss_available():
     return _faiss_available
 
 
+def is_sklearn_available():
+    return _has_sklearn
+
+
+def is_sentencepiece_available():
+    return _sentencepiece_available
+
+
+def is_tokenizers_available():
+    return _tokenizers_available
+
+
+DATASETS_IMPORT_ERROR = """
+{0} requires the 🤗 Datasets library but it was not found in your enviromnent. You can install it with:
+```
+pip install datasets
+```
+In a notebook or a colab, you can install it by executing a cell with
+```
+!pip install datasets
+```
+then restarting your kernel.
+
+Note that if you have a local folder named `datasets` or a local python file named `datasets.py` in your current
+working directory, python may try to import this instead of the 🤗 Datasets library. You should rename this folder or
+that python file if that's the case.
+"""
+
+
+FAISS_IMPORT_ERROR = """
+{0} requires the faiss library but it was not found in your enviromnent. Checkout the instructions on the
+installation page of its repo: https://github.com/facebookresearch/faiss/blob/master/INSTALL.md and follow the ones
+that match your enviromnent.
+"""
+
+
+PYTORCH_IMPORT_ERROR = """
+{0} requires the PyTorch library but it was not found in your enviromnent. Checkout the instructions on the
+installation page: https://pytorch.org/get-started/locally/ and follow the ones that match your enviromnent.
+"""
+
+
+SKLEARN_IMPORT_ERROR = """
+{0} requires the scikit-learn library but it was not found in your enviromnent. You can install it with:
+```
+pip install -U scikit-learn
+```
+In a notebook or a colab, you can install it by executing a cell with
+```
+!pip install -U scikit-learn
+```
+"""
+
+
+TENSORFLOW_IMPORT_ERROR = """
+{0} requires the TensorFlow library but it was not found in your enviromnent. Checkout the instructions on the
+installation page: https://www.tensorflow.org/install and follow the ones that match your enviromnent.
+"""
+
+
+def requires_datasets(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_datasets_available():
+        raise ImportError(DATASETS_IMPORT_ERROR.format(name))
+
+
+def requires_faiss(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_faiss_available():
+        raise ImportError(FAISS_IMPORT_ERROR.format(name))
+
+
+def requires_pytorch(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_torch_available():
+        raise ImportError(PYTORCH_IMPORT_ERROR.format(name))
+
+
+def requires_sklearn(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_sklearn_available():
+        raise ImportError(SKLEARN_IMPORT_ERROR.format(name))
+
+
+def requires_tf(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_tf_available():
+        raise ImportError(TENSORFLOW_IMPORT_ERROR.format(name))
+
+
 def add_start_docstrings(*docstr):
     def docstring_decorator(fn):
         fn.__doc__ = "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
@@ -360,11 +481,12 @@ PT_MASKED_LM_SAMPLE = r"""
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
         >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
-        >>> input_ids = tokenizer("Hello, my dog is cute", return_tensors="pt")["input_ids"]
+        >>> inputs = tokenizer("The capital of France is {mask}.", return_tensors="pt")
+        >>> labels = tokenizer("The capital of France is Paris.", return_tensors="pt")["input_ids"]
 
-        >>> outputs = model(input_ids, labels=input_ids)
+        >>> outputs = model(**inputs, labels=labels)
         >>> loss = outputs.loss
-        >>> prediction_logits = outputs.logits
+        >>> logits = outputs.logits
 """
 
 PT_BASE_MODEL_SAMPLE = r"""
@@ -426,14 +548,15 @@ TF_TOKEN_CLASSIFICATION_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True))
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
         >>> input_ids = inputs["input_ids"]
         >>> inputs["labels"] = tf.reshape(tf.constant([1] * tf.size(input_ids).numpy()), (-1, tf.size(input_ids))) # Batch size 1
 
         >>> outputs = model(inputs)
-        >>> loss, scores = outputs[:2]
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
 """
 
 TF_QUESTION_ANSWERING_SAMPLE = r"""
@@ -443,14 +566,16 @@ TF_QUESTION_ANSWERING_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True))
 
         >>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
         >>> input_dict = tokenizer(question, text, return_tensors='tf')
-        >>> start_scores, end_scores = model(input_dict)
+        >>> outputs = model(input_dict)
+        >>> start_logits = outputs.start_logits
+        >>> end_logits = outputs.end_logits
 
         >>> all_tokens = tokenizer.convert_ids_to_tokens(input_dict["input_ids"].numpy()[0])
-        >>> answer = ' '.join(all_tokens[tf.math.argmax(start_scores, 1)[0] : tf.math.argmax(end_scores, 1)[0]+1])
+        >>> answer = ' '.join(all_tokens[tf.math.argmax(start_logits, 1)[0] : tf.math.argmax(end_logits, 1)[0]+1])
 """
 
 TF_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
@@ -460,13 +585,14 @@ TF_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True))
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
         >>> inputs["labels"] = tf.reshape(tf.constant(1), (-1, 1)) # Batch size 1
 
         >>> outputs = model(inputs)
-        >>> loss, logits = outputs[:2]
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
 """
 
 TF_MASKED_LM_SAMPLE = r"""
@@ -476,12 +602,14 @@ TF_MASKED_LM_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True))
 
-        >>> input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
+        >>> inputs = tokenizer("The capital of France is {mask}.", return_tensors="tf")
+        >>> inputs["labels"] = tokenizer("The capital of France is Paris.", return_tensors="tf")["input_ids"]
 
-        >>> outputs = model(input_ids)
-        >>> prediction_scores = outputs[0]
+        >>> outputs = model(inputs)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
 """
 
 TF_BASE_MODEL_SAMPLE = r"""
@@ -491,12 +619,12 @@ TF_BASE_MODEL_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True))
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
         >>> outputs = model(inputs)
 
-        >>> last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+        >>> last_hidden_states = outputs.last_hidden_states
 """
 
 TF_MULTIPLE_CHOICE_SAMPLE = r"""
@@ -506,7 +634,7 @@ TF_MULTIPLE_CHOICE_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True))
 
         >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
         >>> choice0 = "It is eaten with a fork and a knife."
@@ -517,7 +645,7 @@ TF_MULTIPLE_CHOICE_SAMPLE = r"""
         >>> outputs = model(inputs)  # batch size is 1
 
         >>> # the linear classifier still needs to be trained
-        >>> logits = outputs[0]
+        >>> logits = outputs.logits
 """
 
 TF_CAUSAL_LM_SAMPLE = r"""
@@ -527,18 +655,21 @@ TF_CAUSAL_LM_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True))
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
         >>> outputs = model(inputs)
-        >>> logits = outputs[0]
+        >>> logits = outputs.logits
 """
 
 
-def add_code_sample_docstrings(*docstr, tokenizer_class=None, checkpoint=None, output_type=None, config_class=None):
+def add_code_sample_docstrings(
+    *docstr, tokenizer_class=None, checkpoint=None, output_type=None, config_class=None, mask=None
+):
     def docstring_decorator(fn):
         model_class = fn.__qualname__.split(".")[0]
         is_tf_class = model_class[:2] == "TF"
+        doc_kwargs = dict(model_class=model_class, tokenizer_class=tokenizer_class, checkpoint=checkpoint)
 
         if "SequenceClassification" in model_class:
             code_sample = TF_SEQUENCE_CLASSIFICATION_SAMPLE if is_tf_class else PT_SEQUENCE_CLASSIFICATION_SAMPLE
@@ -548,7 +679,8 @@ def add_code_sample_docstrings(*docstr, tokenizer_class=None, checkpoint=None, o
             code_sample = TF_TOKEN_CLASSIFICATION_SAMPLE if is_tf_class else PT_TOKEN_CLASSIFICATION_SAMPLE
         elif "MultipleChoice" in model_class:
             code_sample = TF_MULTIPLE_CHOICE_SAMPLE if is_tf_class else PT_MULTIPLE_CHOICE_SAMPLE
-        elif "MaskedLM" in model_class:
+        elif "MaskedLM" in model_class or model_class in ["FlaubertWithLMHeadModel", "XLMWithLMHeadModel"]:
+            doc_kwargs["mask"] = "[MASK]" if mask is None else mask
             code_sample = TF_MASKED_LM_SAMPLE if is_tf_class else PT_MASKED_LM_SAMPLE
         elif "LMHead" in model_class:
             code_sample = TF_CAUSAL_LM_SAMPLE if is_tf_class else PT_CAUSAL_LM_SAMPLE
@@ -558,7 +690,7 @@ def add_code_sample_docstrings(*docstr, tokenizer_class=None, checkpoint=None, o
             raise ValueError(f"Docstring can't be built for model {model_class}")
 
         output_doc = _prepare_output_docstrings(output_type, config_class) if output_type is not None else ""
-        built_doc = code_sample.format(model_class=model_class, tokenizer_class=tokenizer_class, checkpoint=checkpoint)
+        built_doc = code_sample.format(**doc_kwargs)
         fn.__doc__ = (fn.__doc__ or "") + "".join(docstr) + output_doc + built_doc
         return fn
 
@@ -843,7 +975,7 @@ def get_from_cache(
         else:
             matching_files = [
                 file
-                for file in fnmatch.filter(os.listdir(cache_dir), filename + ".*")
+                for file in fnmatch.filter(os.listdir(cache_dir), filename.split(".")[0] + ".*")
                 if not file.endswith(".json") and not file.endswith(".lock")
             ]
             if len(matching_files) > 0:

src/transformers/integrations.py

@@ -124,8 +124,7 @@ def run_hp_search_ray(trainer, n_trials: int, direction: str, **kwargs) -> BestR
             metrics = trainer.evaluate()
             trainer.objective = trainer.compute_objective(metrics)
             trainer._tune_save_checkpoint()
-            ray.tune.report(objective=trainer.objective)
-        return trainer.objective
+            ray.tune.report(objective=trainer.objective, **metrics, done=True)
 
     # The model and TensorBoard writer do not pickle so we have to remove them (if they exists)
     # while doing the ray hp search.
@@ -142,7 +141,7 @@ def run_hp_search_ray(trainer, n_trials: int, direction: str, **kwargs) -> BestR
             num_gpus_per_trial = int(math.ceil(num_gpus_per_trial / n_jobs))
         kwargs["resources_per_trial"] = {"gpu": num_gpus_per_trial}
 
-    if "reporter" not in kwargs:
+    if "progress_reporter" not in kwargs:
         from ray.tune import CLIReporter
 
         kwargs["progress_reporter"] = CLIReporter(metric_columns=["objective"])

src/transformers/modeling_auto.py

@@ -26,6 +26,7 @@ from .configuration_auto import (
     BertGenerationConfig,
     CamembertConfig,
     CTRLConfig,
+    DebertaConfig,
     DistilBertConfig,
     DPRConfig,
     ElectraConfig,
@@ -44,6 +45,7 @@ from .configuration_auto import (
     ReformerConfig,
     RetriBertConfig,
     RobertaConfig,
+    SqueezeBertConfig,
     T5Config,
     TransfoXLConfig,
     XLMConfig,
@@ -90,6 +92,7 @@ from .modeling_camembert import (
     CamembertModel,
 )
 from .modeling_ctrl import CTRLLMHeadModel, CTRLModel
+from .modeling_deberta import DebertaForSequenceClassification, DebertaModel
 from .modeling_distilbert import (
     DistilBertForMaskedLM,
     DistilBertForMultipleChoice,
@@ -171,6 +174,14 @@ from .modeling_roberta import (
     RobertaForTokenClassification,
     RobertaModel,
 )
+from .modeling_squeezebert import (
+    SqueezeBertForMaskedLM,
+    SqueezeBertForMultipleChoice,
+    SqueezeBertForQuestionAnswering,
+    SqueezeBertForSequenceClassification,
+    SqueezeBertForTokenClassification,
+    SqueezeBertModel,
+)
 from .modeling_t5 import T5ForConditionalGeneration, T5Model
 from .modeling_transfo_xl import TransfoXLLMHeadModel, TransfoXLModel
 from .modeling_xlm import (
@@ -216,6 +227,7 @@ MODEL_MAPPING = OrderedDict(
         (LongformerConfig, LongformerModel),
         (RobertaConfig, RobertaModel),
         (LayoutLMConfig, LayoutLMModel),
+        (SqueezeBertConfig, SqueezeBertModel),
         (BertConfig, BertModel),
         (OpenAIGPTConfig, OpenAIGPTModel),
         (GPT2Config, GPT2Model),
@@ -231,6 +243,7 @@ MODEL_MAPPING = OrderedDict(
         (FunnelConfig, FunnelModel),
         (LxmertConfig, LxmertModel),
         (BertGenerationConfig, BertGenerationEncoder),
+        (DebertaConfig, DebertaModel),
         (DPRConfig, DPRQuestionEncoder),
     ]
 )
@@ -248,6 +261,7 @@ MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
         (FSMTConfig, FSMTForConditionalGeneration),
         (LongformerConfig, LongformerForMaskedLM),
         (RobertaConfig, RobertaForMaskedLM),
+        (SqueezeBertConfig, SqueezeBertForMaskedLM),
         (BertConfig, BertForPreTraining),
         (OpenAIGPTConfig, OpenAIGPTLMHeadModel),
         (GPT2Config, GPT2LMHeadModel),
@@ -275,6 +289,7 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
         (BartConfig, BartForConditionalGeneration),
         (LongformerConfig, LongformerForMaskedLM),
         (RobertaConfig, RobertaForMaskedLM),
+        (SqueezeBertConfig, SqueezeBertForMaskedLM),
         (BertConfig, BertForMaskedLM),
         (OpenAIGPTConfig, OpenAIGPTLMHeadModel),
         (GPT2Config, GPT2LMHeadModel),
@@ -321,6 +336,7 @@ MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
         (XLMRobertaConfig, XLMRobertaForMaskedLM),
         (LongformerConfig, LongformerForMaskedLM),
         (RobertaConfig, RobertaForMaskedLM),
+        (SqueezeBertConfig, SqueezeBertForMaskedLM),
         (BertConfig, BertForMaskedLM),
         (MobileBertConfig, MobileBertForMaskedLM),
         (FlaubertConfig, FlaubertWithLMHeadModel),
@@ -352,6 +368,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
         (BartConfig, BartForSequenceClassification),
         (LongformerConfig, LongformerForSequenceClassification),
         (RobertaConfig, RobertaForSequenceClassification),
+        (SqueezeBertConfig, SqueezeBertForSequenceClassification),
         (BertConfig, BertForSequenceClassification),
         (XLNetConfig, XLNetForSequenceClassification),
         (MobileBertConfig, MobileBertForSequenceClassification),
@@ -359,6 +376,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
         (XLMConfig, XLMForSequenceClassification),
         (ElectraConfig, ElectraForSequenceClassification),
         (FunnelConfig, FunnelForSequenceClassification),
+        (DebertaConfig, DebertaForSequenceClassification),
     ]
 )
 
@@ -371,6 +389,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
         (LongformerConfig, LongformerForQuestionAnswering),
         (XLMRobertaConfig, XLMRobertaForQuestionAnswering),
         (RobertaConfig, RobertaForQuestionAnswering),
+        (SqueezeBertConfig, SqueezeBertForQuestionAnswering),
         (BertConfig, BertForQuestionAnswering),
         (XLNetConfig, XLNetForQuestionAnsweringSimple),
         (FlaubertConfig, FlaubertForQuestionAnsweringSimple),
@@ -392,6 +411,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
         (XLMRobertaConfig, XLMRobertaForTokenClassification),
         (LongformerConfig, LongformerForTokenClassification),
         (RobertaConfig, RobertaForTokenClassification),
+        (SqueezeBertConfig, SqueezeBertForTokenClassification),
         (BertConfig, BertForTokenClassification),
         (MobileBertConfig, MobileBertForTokenClassification),
         (XLNetConfig, XLNetForTokenClassification),
@@ -409,6 +429,7 @@ MODEL_FOR_MULTIPLE_CHOICE_MAPPING = OrderedDict(
         (XLMRobertaConfig, XLMRobertaForMultipleChoice),
         (LongformerConfig, LongformerForMultipleChoice),
         (RobertaConfig, RobertaForMultipleChoice),
+        (SqueezeBertConfig, SqueezeBertForMultipleChoice),
         (BertConfig, BertForMultipleChoice),
         (DistilBertConfig, DistilBertForMultipleChoice),
         (MobileBertConfig, MobileBertForMultipleChoice),

src/transformers/modeling_bart.py

@@ -64,8 +64,13 @@ BART_PRETRAINED_MODEL_ARCHIVE_LIST = [
 
 BART_START_DOCSTRING = r"""
 
-    This model is a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`_ sub-class. Use it as a regular PyTorch Module and
-    refer to the PyTorch documentation for all matters related to general usage and behavior.
+    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
+    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
+    pruning heads etc.)
+
+    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__ subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general
+    usage and behavior.
 
     Parameters:
         config (:class:`~transformers.BartConfig`): Model configuration class with all the parameters of the model.
@@ -73,6 +78,7 @@ BART_START_DOCSTRING = r"""
             Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 
 """
+
 BART_GENERATION_EXAMPLE = r"""
     Summarization example::
 
@@ -94,39 +100,54 @@ BART_GENERATION_EXAMPLE = r"""
 BART_INPUTS_DOCSTRING = r"""
     Args:
         input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
-               Indices of input sequence tokens in the vocabulary. Use BartTokenizer.encode to produce them.
-            Padding will be ignored by default should you provide it.
-            Indices can be obtained using :class:`transformers.BartTokenizer.encode(text)`.
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using :class:`~transformers.BartTokenizer`.
+            See :meth:`transformers.PreTrainedTokenizer.encode` and
+            :meth:`transformers.PreTrainedTokenizer.__call__` for details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
         attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
-            Mask to avoid performing attention on padding token indices in input_ids.
+            Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
-            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`):
-            Tuple consists of (`last_hidden_state`, `optional`: `hidden_states`, `optional`: `attentions`)
-            `last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`) is a sequence of hidden-states at the output of the last layer of the encoder.
-            Used in the cross-attention of the decoder.
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **maked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
         decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
-            Provide for translation and summarization training. By default, the model will create this tensor by shifting the input_ids right, following the paper.
+            Provide for translation and summarization training. By default, the model will create this tensor by
+            shifting the :obj:`input_ids` to the right, following the paper.
         decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, tgt_seq_len)`, `optional`):
-            Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also be used by default.
-            If you want to change padding behavior, you should read :func:`~transformers.modeling_bart._prepare_decoder_inputs` and modify.
-            See diagram 1 in the paper for more info on the default strategy
+            Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
+            also be used by default.
+
+            If you want to change padding behavior, you should read :func:`modeling_bart._prepare_decoder_inputs` and
+            modify to your needs. See diagram 1 in `the paper <https://arxiv.org/abs/1910.13461>`__ for more
+            information on the default strategy.
+        encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`):
+            Tuple consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`: :obj:`attentions`)
+            :obj:`last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`) is a
+            sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of
+            the decoder.
         past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
-            Contains pre-computed key and value hidden-states of the attention blocks.
-            Can be used to speed up decoding.
-            If ``past_key_values`` are used, the user can optionally input only the last
+            Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last
             ``decoder_input_ids`` (those that don't have their past key value states given to this model) of shape
             :obj:`(batch_size, 1)` instead of all ``decoder_input_ids`` of shape :obj:`(batch_size, sequence_length)`.
-        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            If `use_cache` is True, ``past_key_values`` are returned and can be used to speed up decoding (see
-            ``past_key_values``).
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
         output_attentions (:obj:`bool`, `optional`):
-            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
         output_hidden_states (:obj:`bool`, `optional`):
-            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
         return_dict (:obj:`bool`, `optional`):
-            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
-            plain tuple.
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
 """
 
 
@@ -269,6 +290,9 @@ class EncoderLayer(nn.Module):
         x = residual + x
         if not self.normalize_before:
             x = self.final_layer_norm(x)
+        if torch.isinf(x).any() or torch.isnan(x).any():
+            clamp_value = torch.finfo(x.dtype).max - 1000
+            x = torch.clamp(x, min=-clamp_value, max=clamp_value)
         return x, attn_weights
 
 
@@ -871,8 +895,8 @@ class BartModel(PretrainedBartModel):
         input_ids,
         attention_mask=None,
         decoder_input_ids=None,
-        encoder_outputs: Optional[Tuple] = None,
         decoder_attention_mask=None,
+        encoder_outputs: Optional[Tuple] = None,
         past_key_values=None,
         use_cache=None,
         output_attentions=None,
@@ -1001,9 +1025,9 @@ class BartForConditionalGeneration(PretrainedBartModel):
         self,
         input_ids,
         attention_mask=None,
-        encoder_outputs=None,
         decoder_input_ids=None,
         decoder_attention_mask=None,
+        encoder_outputs=None,
         past_key_values=None,
         labels=None,
         use_cache=None,
@@ -1013,31 +1037,31 @@ class BartForConditionalGeneration(PretrainedBartModel):
         **unused,
     ):
         r"""
-            labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
-                Labels for computing the masked language modeling loss.
-                Indices should either be in ``[0, ..., config.vocab_size]`` or -100 (see ``input_ids`` docstring).
-                Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens
-                with labels in ``[0, ..., config.vocab_size]``.
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the masked language modeling loss.
+            Indices should either be in ``[0, ..., config.vocab_size]`` or -100 (see ``input_ids`` docstring).
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens
+            with labels in ``[0, ..., config.vocab_size]``.
 
         Returns:
 
         Conditional generation example::
 
-                >>> # Mask filling only works for bart-large
-                >>> from transformers import BartTokenizer, BartForConditionalGeneration
-                >>> tokenizer = BartTokenizer.from_pretrained('facebook/bart-large')
-                >>> TXT = "My friends are <mask> but they eat too many carbs."
+            >>> # Mask filling only works for bart-large
+            >>> from transformers import BartTokenizer, BartForConditionalGeneration
+            >>> tokenizer = BartTokenizer.from_pretrained('facebook/bart-large')
+            >>> TXT = "My friends are <mask> but they eat too many carbs."
 
-                >>> model = BartForConditionalGeneration.from_pretrained('facebook/bart-large')
-                >>> input_ids = tokenizer([TXT], return_tensors='pt')['input_ids']
-                >>> logits = model(input_ids).logits
+            >>> model = BartForConditionalGeneration.from_pretrained('facebook/bart-large')
+            >>> input_ids = tokenizer([TXT], return_tensors='pt')['input_ids']
+            >>> logits = model(input_ids).logits
 
-                >>> masked_index = (input_ids[0] == tokenizer.mask_token_id).nonzero().item()
-                >>> probs = logits[0, masked_index].softmax(dim=0)
-                >>> values, predictions = probs.topk(5)
+            >>> masked_index = (input_ids[0] == tokenizer.mask_token_id).nonzero().item()
+            >>> probs = logits[0, masked_index].softmax(dim=0)
+            >>> values, predictions = probs.topk(5)
 
-                >>> tokenizer.decode(predictions).split()
-                >>> # ['good', 'great', 'all', 'really', 'very']
+            >>> tokenizer.decode(predictions).split()
+            >>> # ['good', 'great', 'all', 'really', 'very']
         """
         if "lm_labels" in unused:
             warnings.warn(
@@ -1168,9 +1192,9 @@ class BartForSequenceClassification(PretrainedBartModel):
         self,
         input_ids,
         attention_mask=None,
-        encoder_outputs=None,
         decoder_input_ids=None,
         decoder_attention_mask=None,
+        encoder_outputs=None,
         labels=None,
         use_cache=None,
         output_attentions=None,
@@ -1254,9 +1278,9 @@ class BartForQuestionAnswering(PretrainedBartModel):
         self,
         input_ids,
         attention_mask=None,
-        encoder_outputs=None,
         decoder_input_ids=None,
         decoder_attention_mask=None,
+        encoder_outputs=None,
         start_positions=None,
         end_positions=None,
         use_cache=None,

src/transformers/modeling_ctrl.py

@@ -251,11 +251,11 @@ CTRL_START_DOCSTRING = r"""
 CTRL_INPUTS_DOCSTRING = r"""
     Args:
         input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
-            :obj:`input_ids_length` = ``sequence_length`` if ``past_key_values`` is ``None`` else
+            :obj:`input_ids_length` = ``sequence_length`` if :obj:`past_key_values` is ``None`` else
             ``past_key_values[0].shape[-2]`` (``sequence_length`` of input past key value states).
             Indices of input sequence tokens in the vocabulary.
 
-            If ``past_key_values`` is used, only input IDs that do not have their past calculated should be passed as
+            If :obj:`past_key_values` is used, only input IDs that do not have their past calculated should be passed as
             ``input_ids``.
 
             Indices can be obtained using :class:`~transformers.CTRLTokenizer`.
@@ -265,7 +265,7 @@ CTRL_INPUTS_DOCSTRING = r"""
             `What are input IDs? <../glossary.html#input-ids>`__
         past_key_values (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
             Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
-            (see ``past_key_values`` output below). Can be used to speed up sequential decoding.
+            (see :obj:`past_key_values` output below). Can be used to speed up sequential decoding.
             The ``input_ids`` which have their past given to this model should not be passed as input ids as they have
             already been computed.
         attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
@@ -301,8 +301,8 @@ CTRL_INPUTS_DOCSTRING = r"""
             This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
             vectors than the model's internal embedding lookup matrix.
         use_cache (:obj:`bool`, `optional`):
-            If set to :obj:`True`, ``past_key_values`` key value states are returned and can be used to speed up
-            decoding (see ``past_key_values``).
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
         output_attentions (:obj:`bool`, `optional`):
             Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
             tensors for more detail.

src/transformers/modeling_electra.py

@@ -535,6 +535,7 @@ class ElectraPreTrainedModel(PreTrainedModel):
     load_tf_weights = load_tf_weights_in_electra
     base_model_prefix = "electra"
     authorized_missing_keys = [r"position_ids"]
+    authorized_unexpected_keys = [r"electra\.embeddings_project\.weight", r"electra\.embeddings_project\.bias"]
 
     # Copied from transformers.modeling_bert.BertPreTrainedModel._init_weights
     def _init_weights(self, module):

src/transformers/modeling_encoder_decoder.py

@@ -69,10 +69,6 @@ ENCODER_DECODER_INPUTS_DOCSTRING = r"""
             :meth:`transformers.PreTrainedTokenizer.__call__` for details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
-        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
-            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
-            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
-            vectors than the model's internal embedding lookup matrix.
         attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -81,11 +77,6 @@ ENCODER_DECODER_INPUTS_DOCSTRING = r"""
             - 0 for tokens that are **maked**.
 
             `What are attention masks? <../glossary.html#attention-mask>`__
-        encoder_outputs (:obj:`tuple(torch.FloatTensor)`, `optional`):
-            This tuple must consist of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`: :obj:`attentions`)
-            :obj:`last_hidden_state` (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`)
-            is a tensor of hidden-states at the output of the last layer of the encoder.
-            Used in the cross-attention of the decoder.
         decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
             Provide for sequence to sequence training to the decoder.
             Indices can be obtained using :class:`~transformers.PretrainedTokenizer`.
@@ -94,6 +85,21 @@ ENCODER_DECODER_INPUTS_DOCSTRING = r"""
         decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, tgt_seq_len)`, `optional`):
             Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
             also be used by default.
+        encoder_outputs (:obj:`tuple(torch.FloatTensor)`, `optional`):
+            This tuple must consist of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`: :obj:`attentions`)
+            :obj:`last_hidden_state` (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`)
+            is a tensor of hidden-states at the output of the last layer of the encoder.
+            Used in the cross-attention of the decoder.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
         decoder_inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`):
             Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded
             representation. This is useful if you want more control over how to convert :obj:`decoder_input_ids`
@@ -103,6 +109,15 @@ ENCODER_DECODER_INPUTS_DOCSTRING = r"""
             Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
             Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with
             labels in ``[0, ..., config.vocab_size]``
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
         return_dict (:obj:`bool`, `optional`):
             If set to ``True``, the model will return a :class:`~transformers.file_utils.Seq2SeqLMOutput` instead of a
             plain tuple.
@@ -328,13 +343,17 @@ class EncoderDecoderModel(PreTrainedModel):
     def forward(
         self,
         input_ids=None,
-        inputs_embeds=None,
         attention_mask=None,
-        encoder_outputs=None,
         decoder_input_ids=None,
         decoder_attention_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,  # TODO: (PVP) implement :obj:`use_cache`
+        inputs_embeds=None,
         decoder_inputs_embeds=None,
         labels=None,
+        use_cache=None,  # TODO: (PVP) implement :obj:`use_cache`
+        output_attentions=None,
+        output_hidden_states=None,
         return_dict=None,
         **kwargs,
     ):
@@ -378,20 +397,24 @@ class EncoderDecoderModel(PreTrainedModel):
                 input_ids=input_ids,
                 attention_mask=attention_mask,
                 inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
                 return_dict=return_dict,
                 **kwargs_encoder,
             )
 
-        hidden_states = encoder_outputs[0]
+        encoder_hidden_states = encoder_outputs[0]
 
         # Decode
         decoder_outputs = self.decoder(
             input_ids=decoder_input_ids,
-            inputs_embeds=decoder_inputs_embeds,
             attention_mask=decoder_attention_mask,
-            encoder_hidden_states=hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=attention_mask,
+            inputs_embeds=decoder_inputs_embeds,
             labels=labels,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
             return_dict=return_dict,
             **kwargs_decoder,
         )
@@ -423,7 +446,7 @@ class EncoderDecoderModel(PreTrainedModel):
             "encoder_outputs": encoder_outputs,
         }
 
-        # Ideally all models should have a `use_cache`
+        # Ideally all models should have a :obj:`use_cache`
         # leave following to ifs until all have it implemented
         if "use_cache" in decoder_inputs:
             input_dict["decoder_use_cache"] = decoder_inputs["use_cache"]

src/transformers/modeling_fsmt.py

@@ -227,10 +227,6 @@ FSMT_INPUTS_DOCSTRING = r"""
             - 0 for tokens that are **maked**.
 
             `What are attention masks? <../glossary.html#attention-mask>`__
-        encoder_outputs (:obj:`Tuple(torch.FloatTensor)`, `optional`):
-            Tuple consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`: :obj:`attentions`)
-            :obj:`last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)` is a sequence of
-            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
         decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
             Provide for translation and summarization training. By default, the model will create this tensor by
             shifting the input_ids right, following the paper.
@@ -240,6 +236,10 @@ FSMT_INPUTS_DOCSTRING = r"""
             If you want to change padding behavior, you should read
             :func:`modeling_fstm._prepare_fstm_decoder_inputs` and modify.
             See diagram 1 in the paper for more info on the default strategy
+        encoder_outputs (:obj:`Tuple(torch.FloatTensor)`, `optional`):
+            Tuple consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`: :obj:`attentions`)
+            :obj:`last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)` is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
         past_key_values (:obj:`Tuple(torch.FloatTensor)` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
             Contains precomputed key and value hidden-states of the attention blocks.
             Can be used to speed up decoding.
@@ -248,8 +248,8 @@ FSMT_INPUTS_DOCSTRING = r"""
             :obj:`(batch_size, 1)` instead of all :obj:`decoder_input_ids` of shape
             :obj:`(batch_size, sequence_length)`.
         use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            If set to :obj:`True`, ``past_key_values`` key value states are returned and can be used to speed up
-            decoding (see ``past_key_values``).
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
         output_attentions (:obj:`bool`, `optional`):
             Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
             tensors for more detail.
@@ -910,8 +910,8 @@ class FSMTModel(PretrainedFSMTModel):
         input_ids,
         attention_mask=None,
         decoder_input_ids=None,
-        encoder_outputs: Optional[Tuple] = None,
         decoder_attention_mask=None,
+        encoder_outputs: Optional[Tuple] = None,
         past_key_values=None,
         use_cache=None,
         output_attentions=None,
@@ -1045,9 +1045,9 @@ class FSMTForConditionalGeneration(PretrainedFSMTModel):
         self,
         input_ids,
         attention_mask=None,
-        encoder_outputs=None,
         decoder_input_ids=None,
         decoder_attention_mask=None,
+        encoder_outputs=None,
         past_key_values=None,
         labels=None,
         use_cache=None,

src/transformers/modeling_funnel.py

@@ -187,16 +187,16 @@ class FunnelAttentionStructure(nn.Module):
         # dividide.
         self.pooling_mult = None
 
-    def init_attention_inputs(self, input_embeds, attention_mask=None, token_type_ids=None):
+    def init_attention_inputs(self, inputs_embeds, attention_mask=None, token_type_ids=None):
         """ Returns the attention inputs associated to the inputs of the model. """
-        # input_embeds has shape batch_size x seq_len x d_model
+        # inputs_embeds has shape batch_size x seq_len x d_model
         # attention_mask and token_type_ids have shape batch_size x seq_len
         self.pooling_mult = 1
-        self.seq_len = seq_len = input_embeds.size(1)
-        position_embeds = self.get_position_embeds(seq_len, input_embeds.dtype, input_embeds.device)
+        self.seq_len = seq_len = inputs_embeds.size(1)
+        position_embeds = self.get_position_embeds(seq_len, inputs_embeds.dtype, inputs_embeds.device)
         token_type_mat = self.token_type_ids_to_mat(token_type_ids) if token_type_ids is not None else None
         cls_mask = (
-            F.pad(input_embeds.new_ones([seq_len - 1, seq_len - 1]), (1, 0, 1, 0))
+            F.pad(inputs_embeds.new_ones([seq_len - 1, seq_len - 1]), (1, 0, 1, 0))
             if self.config.separate_cls
             else None
         )
@@ -1169,6 +1169,7 @@ class FunnelForMaskedLM(FunnelPreTrainedModel):
         checkpoint="funnel-transformer/small",
         output_type=MaskedLMOutput,
         config_class=_CONFIG_FOR_DOC,
+        mask="<mask>",
     )
     def forward(
         self,

src/transformers/modeling_gpt2.py

@@ -15,7 +15,6 @@
 # limitations under the License.
 """PyTorch OpenAI GPT-2 model."""
 
-
 import os
 import warnings
 from dataclasses import dataclass
@@ -366,7 +365,7 @@ class GPT2DoubleHeadsModelOutput(ModelOutput):
             :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
 
             Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
-            ``past_key_values`` input) to speed up sequential decoding.
+            :obj:`past_key_values` input) to speed up sequential decoding.
         hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
@@ -408,11 +407,11 @@ GPT2_START_DOCSTRING = r"""
 GPT2_INPUTS_DOCSTRING = r"""
     Args:
         input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
-            :obj:`input_ids_length` = ``sequence_length`` if ``past_key_values`` is ``None`` else
+            :obj:`input_ids_length` = ``sequence_length`` if :obj:`past_key_values` is ``None`` else
             ``past_key_values[0].shape[-2]`` (``sequence_length`` of input past key value states).
             Indices of input sequence tokens in the vocabulary.
 
-            If ``past_key_values`` is used, only ``input_ids`` that do not have their past calculated should be passed
+            If :obj:`past_key_values` is used, only ``input_ids`` that do not have their past calculated should be passed
             as ``input_ids``.
 
             Indices can be obtained using :class:`~transformers.GPT2Tokenizer`.
@@ -422,7 +421,7 @@ GPT2_INPUTS_DOCSTRING = r"""
             `What are input IDs? <../glossary.html#input-ids>`__
         past_key_values (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
             Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model
-            (see ``past_key_values`` output below). Can be used to speed up sequential decoding.
+            (see :obj:`past_key_values` output below). Can be used to speed up sequential decoding.
             The ``input_ids`` which have their past given to this model should not be passed as ``input_ids`` as they
             have already been computed.
         attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
@@ -458,11 +457,11 @@ GPT2_INPUTS_DOCSTRING = r"""
             This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
             vectors than the model's internal embedding lookup matrix.
 
-            If ``past_key_values`` is used, optionally only the last :obj:`inputs_embeds` have to be input (see
-            ``past_key_values``).
+            If :obj:`past_key_values` is used, optionally only the last :obj:`inputs_embeds` have to be input (see
+            :obj:`past_key_values`).
         use_cache (:obj:`bool`, `optional`):
-            If set to :obj:`True`, ``past_key_values`` key value states are returned and can be used to speed up
-            decoding (see ``past_key_values``).
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
         output_attentions (:obj:`bool`, `optional`):
             Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
             tensors for more detail.
@@ -624,16 +623,35 @@ class GPT2Model(GPT2PreTrainedModel):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)
 
-            outputs = block(
-                hidden_states,
-                layer_past=layer_past,
-                attention_mask=attention_mask,
-                head_mask=head_mask[i],
-                encoder_hidden_states=encoder_hidden_states,
-                encoder_attention_mask=encoder_attention_mask,
-                use_cache=use_cache,
-                output_attentions=output_attentions,
-            )
+            if getattr(self.config, "gradient_checkpointing", False):
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # checkpointing only works with tuple returns, not with lists
+                        return tuple(output for output in module(*inputs, use_cache, output_attentions))
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    layer_past,
+                    attention_mask,
+                    head_mask[i],
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=attention_mask,
+                    head_mask=head_mask[i],
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
 
             hidden_states, present = outputs[:2]
             if use_cache is True:

src/transformers/modeling_layoutlm.py

@@ -94,10 +94,14 @@ class LayoutLMEmbeddings(nn.Module):
 
         words_embeddings = inputs_embeds
         position_embeddings = self.position_embeddings(position_ids)
-        left_position_embeddings = self.x_position_embeddings(bbox[:, :, 0])
-        upper_position_embeddings = self.y_position_embeddings(bbox[:, :, 1])
-        right_position_embeddings = self.x_position_embeddings(bbox[:, :, 2])
-        lower_position_embeddings = self.y_position_embeddings(bbox[:, :, 3])
+        try:
+            left_position_embeddings = self.x_position_embeddings(bbox[:, :, 0])
+            upper_position_embeddings = self.y_position_embeddings(bbox[:, :, 1])
+            right_position_embeddings = self.x_position_embeddings(bbox[:, :, 2])
+            lower_position_embeddings = self.y_position_embeddings(bbox[:, :, 3])
+        except IndexError as e:
+            raise IndexError("The :obj:`bbox`coordinate values should be within 0-1000 range.") from e
+
         h_position_embeddings = self.h_position_embeddings(bbox[:, :, 3] - bbox[:, :, 1])
         w_position_embeddings = self.w_position_embeddings(bbox[:, :, 2] - bbox[:, :, 0])
         token_type_embeddings = self.token_type_embeddings(token_type_ids)

src/transformers/modeling_lxmert.py

@@ -958,7 +958,7 @@ class LxmertModel(LxmertPreTrainedModel):
         # positions we want to attend and -10000.0 for masked positions.
         # Since we are adding it to the raw scores before the softmax, this is
         # effectively the same as removing these entirely.
-        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype)
+        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)
         extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
 
         # Process the visual attention mask

src/transformers/modeling_marian.py

@@ -23,11 +23,12 @@ from .modeling_bart import BartForConditionalGeneration
 
 
 class MarianMTModel(BartForConditionalGeneration):
-    config_class = MarianConfig
     r"""
     Pytorch version of marian-nmt's transformer.h (c++). Designed for the OPUS-NMT translation checkpoints.
-    Model API is identical to BartForConditionalGeneration.
-    Available models are listed at `Model List <https://huggingface.co/models?search=Helsinki-NLP>`__
+    Available models are listed `here <https://huggingface.co/models?search=Helsinki-NLP>`__.
+
+    This class overrides :class:`~transformers.BartForConditionalGeneration`. Please check the
+    superclass for the appropriate documentation alongside usage examples.
 
     Examples::
 
@@ -45,6 +46,7 @@ class MarianMTModel(BartForConditionalGeneration):
         >>> words: List[str] = tok.batch_decode(gen, skip_special_tokens=True)  # returns "Where is the bus stop ?"
 
     """
+    config_class = MarianConfig
 
     def adjust_logits_during_generation(self, logits, cur_len, max_length):
         logits[:, self.config.pad_token_id] = float("-inf")  # never predict pad token.

src/transformers/modeling_mbart.py

@@ -1,5 +1,4 @@
 from .configuration_mbart import MBartConfig
-from .file_utils import add_start_docstrings
 from .modeling_bart import BartForConditionalGeneration
 
 
@@ -12,23 +11,7 @@ MBART_PRETRAINED_MODEL_ARCHIVE_LIST = [
     # See all multilingual BART models at https://huggingface.co/models?filter=mbart
 ]
 
-MBART_START_DOCSTRING = r"""
 
-    This model is a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__ sub-class.
-    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general
-    usage and behavior.
-
-    Parameters:
-        config (:class:`~transformers.MBartConfig`): Model configuration class with all the parameters of the
-            model. Initializing with a config file does not load the weights associated with the model, only the
-            configuration.
-            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
-"""
-
-
-@add_start_docstrings(
-    "The BART Model with a language modeling head. Can be used for machine translation.", MBART_START_DOCSTRING
-)
 class MBartForConditionalGeneration(BartForConditionalGeneration):
     r"""
     This class overrides :class:`~transformers.BartForConditionalGeneration`. Please check the

src/transformers/modeling_outputs.py

@@ -80,7 +80,7 @@ class BaseModelOutputWithPast(ModelOutput):
             :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
 
             Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
-            ``past_key_values`` input) to speed up sequential decoding.
+            :obj:`past_key_values` input) to speed up sequential decoding.
         hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
@@ -110,13 +110,13 @@ class Seq2SeqModelOutput(ModelOutput):
         last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
             Sequence of hidden-states at the output of the last layer of the decoder of the model.
 
-            If ``past_key_values`` is used only the last hidden-state of the sequences of shape :obj:`(batch_size, 1, hidden_size)` is output.
+            If :obj:`past_key_values` is used only the last hidden-state of the sequences of shape :obj:`(batch_size, 1, hidden_size)` is output.
         past_key_values (:obj:`List[torch.FloatTensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
             List of :obj:`torch.FloatTensor` of length :obj:`config.n_layers`,  with each tensor of shape
             :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
 
             Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
-            used (see ``past_key_values`` input) to speed up sequential decoding.
+            used (see :obj:`past_key_values` input) to speed up sequential decoding.
         decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
@@ -196,7 +196,7 @@ class CausalLMOutputWithPast(ModelOutput):
             :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
 
             Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
-            ``past_key_values`` input) to speed up sequential decoding.
+            :obj:`past_key_values` input) to speed up sequential decoding.
         hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
@@ -261,7 +261,7 @@ class Seq2SeqLMOutput(ModelOutput):
             :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
 
             Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
-            used (see ``past_key_values`` input) to speed up sequential decoding.
+            used (see :obj:`past_key_values` input) to speed up sequential decoding.
         decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
@@ -371,7 +371,7 @@ class Seq2SeqSequenceClassifierOutput(ModelOutput):
             :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
 
             Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
-            used (see ``past_key_values`` input) to speed up sequential decoding.
+            used (see :obj:`past_key_values` input) to speed up sequential decoding.
         decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
@@ -517,7 +517,7 @@ class Seq2SeqQuestionAnsweringModelOutput(ModelOutput):
             :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
 
             Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
-            used (see ``past_key_values`` input) to speed up sequential decoding.
+            used (see :obj:`past_key_values` input) to speed up sequential decoding.
         decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.

src/transformers/modeling_pegasus.py

@@ -22,18 +22,12 @@ from .modeling_bart import BART_START_DOCSTRING, BartForConditionalGeneration
 
 @add_start_docstrings("The Pegasus Model for summarization ", BART_START_DOCSTRING)
 class PegasusForConditionalGeneration(BartForConditionalGeneration):
-    config_class = PegasusConfig
-    authorized_missing_keys = [
-        r"final_logits_bias",
-        r"encoder\.version",
-        r"decoder\.version",
-        r"model.encoder.embed_positions",
-        "model.decoder.embed_positions",
-    ]
     r"""
     Pytorch version of google's pegasus model for summarization.
-    Model API is identical to BartForConditionalGeneration.
-    Available models are listed at `Model List <https://huggingface.co/models?search=pegasus>`__
+    Available models are listed `here <https://huggingface.co/models?search=pegasus>`__.
+
+    This class overrides :class:`~transformers.BartForConditionalGeneration`. Please check the
+    superclass for the appropriate documentation alongside usage examples.
 
     Examples::
 
@@ -51,3 +45,11 @@ class PegasusForConditionalGeneration(BartForConditionalGeneration):
 
     """
     # All the code is in src/transformers/modeling_bart.py
+    config_class = PegasusConfig
+    authorized_missing_keys = [
+        r"final_logits_bias",
+        r"encoder\.version",
+        r"decoder\.version",
+        r"model.encoder.embed_positions",
+        "model.decoder.embed_positions",
+    ]

src/transformers/modeling_rag.py

@@ -52,7 +52,7 @@ class RetrievAugLMMarginOutput(ModelOutput):
             :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
 
             Contains precomputed hidden-states (key and values in the attention blocks) of the decoder that can be used
-            (see ``past_key_values`` input) to speed up sequential decoding.
+            (see :obj:`past_key_values` input) to speed up sequential decoding.
         retrieved_doc_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.n_docs, hidden_size)`, `optional`, returned when `output_retrieved=True`):
             Embedded documents retrieved by the retriever.
             Is used with ``question_encoder_last_hidden_state`` to compute the ``doc_scores``.
@@ -137,7 +137,7 @@ class RetrievAugLMOutput(ModelOutput):
             :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
 
             Contains precomputed hidden-states (key and values in the attention blocks) of the decoder that can be used
-            (see ``past_key_values`` input) to speed up sequential decoding.
+            (see :obj:`past_key_values` input) to speed up sequential decoding.
         retrieved_doc_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.n_docs, hidden_size)`, `optional`, returned when `output_retrieved=True`):
             Embedded documents retrieved by the retriever.
             Is used with ``question_encoder_last_hidden_state`` to compute the ``doc_scores``.
@@ -447,8 +447,8 @@ RAG_FORWARD_INPUTS_DOCSTRING = r"""
             to the forward pass. :obj:`context_attention_mask` are returned by
             :meth:`~transformers.RagRetriever.__call__`.
         use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            If set to :obj:`True`, ``past_key_values`` key value states are returned and can be used to speed up
-            decoding (see ``past_key_values``).
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
         output_attentions (:obj:`bool`, `optional`):
             Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
             tensors for more detail.
@@ -537,7 +537,7 @@ class RagModel(RagPreTrainedModel):
 
             >>> input_dict = tokenizer.prepare_seq2seq_batch("How many people live in Paris?", "In Paris, there are 10 million people.", return_tensors="pt")
             >>> input_ids = input_dict["input_ids"]
-            >>> outputs = model(input_ids=input_ids, labels=input_dict["labels"])
+            >>> outputs = model(input_ids=input_ids)
 
         """
         use_cache = use_cache if use_cache is not None else self.config.use_cache

src/transformers/modeling_reformer.py

@@ -1959,8 +1959,8 @@ REFORMER_INPUTS_DOCSTRING = r"""
             Contains precomputed hidden-states and buckets (only relevant for LSH Self-Attention). Can be used to speed
             up sequential decoding.
         use_cache (:obj:`bool`, `optional`):
-            If set to :obj:`True`, ``past_key_values`` key value states are returned and can be used to speed up
-            decoding (see ``past_key_values``).
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
         output_attentions (:obj:`bool`, `optional`):
             Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
             tensors for more detail.