adding positional embeds masking to TFRoBERTa

closes #1960 Add saving and resuming functionality for remaining examples

.circleci/config.yml

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

CONTRIBUTING.md

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

README.md

@@ -55,10 +55,12 @@ Choose the right framework for every part of a model's lifetime
 | [Online demo](#online-demo) | Experimenting with this repo’s text generation capabilities |
 | [Quick tour: Usage](#quick-tour) | Tokenizers & models usage: Bert and GPT-2 |
 | [Quick tour: TF 2.0 and PyTorch ](#Quick-tour-TF-20-training-and-PyTorch-interoperability) | Train a TF 2.0 model in 10 lines of code, load it in PyTorch |
+| [Quick tour: pipelines](#quick-tour-of-pipelines) | Using Pipelines: Wrapper around tokenizer and models to use finetuned models |
 | [Quick tour: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
+| [Quick tour: Share your models ](#Quick-tour-of-model-sharing) | Upload and share your fine-tuned models with the community |
 | [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-transformers to transformers |
 | [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
-| [Documentation][(v2.2.0/v2.2.1)](https://huggingface.co/transformers/v2.2.0) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) [(master)](https://huggingface.co/transformers) | Full API documentation and more |
+| [Documentation][(v2.3.0)](https://huggingface.co/transformers/v2.3.0)[(v2.2.0/v2.2.1/v2.2.2)](https://huggingface.co/transformers/v2.2.0) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) [(master)](https://huggingface.co/transformers) | Full API documentation and more |
 
 ## Installation
 
@@ -101,17 +103,26 @@ pip install [--editable] .
 
 A series of tests are included for the library and the example scripts. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples).
 
-These tests can be run using `pytest` (install pytest if needed with `pip install pytest`).
+These tests can be run using `unittest` or `pytest` (install pytest if needed with `pip install pytest`).
 
 Depending on which framework is installed (TensorFlow 2.0 and/or PyTorch), the irrelevant tests will be skipped. Ensure that both frameworks are installed if you want to execute all tests.
 
 You can run the tests from the root of the cloned repository with the commands:
 
+```bash
+python -m unittest discover -s transformers/tests -p "*test.py" -t .
+python -m unittest discover -s examples -p "*test.py" -t examples
+```
+
+or
+
 ```bash
 python -m pytest -sv ./transformers/tests/
 python -m pytest -sv ./examples/
 ```
 
+By default, slow tests are skipped. Set the `RUN_SLOW` environment variable to `yes` to run them.
+
 ### Do you want to run a Transformer model on a mobile device?
 
 You should check out our [`swift-coreml-transformers`](https://github.com/huggingface/swift-coreml-transformers) repo.
@@ -122,7 +133,7 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 
 ## Model architectures
 
-🤗 Transformers currently provides 10 NLU/NLG architectures:
+🤗 Transformers currently provides the following NLU/NLG architectures:
 
 1. **[BERT](https://github.com/google-research/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
 2. **[GPT](https://github.com/openai/finetune-transformer-lm)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
@@ -131,11 +142,13 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 5. **[XLNet](https://github.com/zihangdai/xlnet/)** (from Google/CMU) released with the paper [​XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
 6. **[XLM](https://github.com/facebookresearch/XLM/)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 7. **[RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/roberta)** (from Facebook), released together with the paper a [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
-8. **[DistilBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation).
+8. **[DistilBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
 9. **[CTRL](https://github.com/salesforce/ctrl/)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
 10. **[CamemBERT](https://camembert-model.fr)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
-11. **[ALBERT](https://github.com/google-research/google-research/tree/master/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
-11. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
+11. **[ALBERT](https://github.com/google-research/ALBERT)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+12. **[T5](https://github.com/google-research/text-to-text-transfer-transformer)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
+13. **[XLM-RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/xlmr)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
+14. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
 
 These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Peason R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
 
@@ -157,7 +170,7 @@ import torch
 from transformers import *
 
 # Transformers has a unified API
-# for 8 transformer architectures and 30 pretrained weights.
+# for 10 transformer architectures and 30 pretrained weights.
 #          Model          | Tokenizer          | Pretrained weights shortcut
 MODELS = [(BertModel,       BertTokenizer,       'bert-base-uncased'),
           (OpenAIGPTModel,  OpenAIGPTTokenizer,  'openai-gpt'),
@@ -167,7 +180,9 @@ MODELS = [(BertModel,       BertTokenizer,       'bert-base-uncased'),
           (XLNetModel,      XLNetTokenizer,      'xlnet-base-cased'),
           (XLMModel,        XLMTokenizer,        'xlm-mlm-enfr-1024'),
           (DistilBertModel, DistilBertTokenizer, 'distilbert-base-uncased'),
-          (RobertaModel,    RobertaTokenizer,    'roberta-base')]
+          (RobertaModel,    RobertaTokenizer,    'roberta-base'),
+          (XLMRobertaModel, XLMRobertaTokenizer, 'xlm-roberta-base'),
+         ]
 
 # To use TensorFlow 2.0 versions of the models, simply prefix the class names with 'TF', e.g. `TFRobertaModel` is the TF 2.0 counterpart of the PyTorch model `RobertaModel`
 
@@ -436,6 +451,76 @@ python ./examples/run_generation.py \
     --repetition_penalty=1.2 \
 ```
 
+## Quick tour of model sharing
+
+New in `v2.2.2`: you can now upload and share your fine-tuned models with the community, using the <abbr title="Command-line interface">CLI</abbr> that's built-in to the library.
+
+**First, create an account on [https://huggingface.co/join](https://huggingface.co/join)**. Then:
+
+```shell
+transformers-cli login
+# log in using the same credentials as on huggingface.co
+```
+Upload your model:
+```shell
+transformers-cli upload ./path/to/pretrained_model/
+
+# ^^ Upload folder containing weights/tokenizer/config
+# saved via `.save_pretrained()`
+
+transformers-cli upload ./config.json [--filename folder/foobar.json]
+
+# ^^ Upload a single file
+# (you can optionally override its filename, which can be nested inside a folder)
+```
+
+Your model will then be accessible through its identifier, a concatenation of your username and the folder name above:
+```python
+"username/model_name"
+```
+
+Anyone can load it from code:
+```python
+tokenizer = AutoTokenizer.from_pretrained("username/pretrained_model")
+model = AutoModel.from_pretrained("username/pretrained_model")
+```
+
+Finally, list all your files on S3:
+```shell
+transformers-cli ls
+# List all your S3 objects.
+```
+
+## Quick tour of pipelines
+
+New in version `v2.3`: `Pipeline` are high-level objects which automatically handle tokenization, running your data through a transformers model
+and outputting the result in a structured object. 
+
+You can create `Pipeline` objects for the following down-stream tasks:
+
+ - `feature-extraction`: Generates a tensor representation for the input sequence
+ - `ner`: Generates named entity mapping for each word in the input sequence.
+ - `sentiment-analysis`: Gives the polarity (positive / negative) of the whole input sequence. 
+ - `question-answering`: Provided some context and a question refering to the context, it will extract the answer to the question
+ in the context.
+
+```python
+from transformers import pipeline
+
+# Allocate a pipeline for sentiment-analysis
+nlp = pipeline('sentiment-analysis')
+nlp('We are very happy to include pipeline into the transformers repository.')
+>>> {'label': 'POSITIVE', 'score': 0.99893874}
+
+# Allocate a pipeline for question-answering
+nlp = pipeline('question-answering')
+nlp({
+    'question': 'What is the name of the repository ?', 
+    'context': 'Pipeline have been included in the huggingface/transformers repository'
+})
+>>> {'score': 0.28756016668193496, 'start': 35, 'end': 59, 'answer': 'huggingface/transformers'}
+```
+
 ## Migrating from pytorch-transformers to transformers
 
 Here is a quick summary of what you should take care of when migrating from `pytorch-transformers` to `transformers`.

docs/source/conf.py

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

docs/source/index.rst

@@ -49,7 +49,8 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
 8. `DistilBERT <https://huggingface.co/transformers/model_doc/distilbert.html>`_ (from HuggingFace) released together with the paper `DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter <https://arxiv.org/abs/1910.01108>`_ by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into `DistilGPT2 <https://github.com/huggingface/transformers/tree/master/examples/distillation>`_.
 9. `CTRL <https://github.com/pytorch/fairseq/tree/master/examples/ctrl>`_ (from Salesforce), released together with the paper `CTRL: A Conditional Transformer Language Model for Controllable Generation <https://www.github.com/salesforce/ctrl>`_ by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
 10. `CamemBERT <https://huggingface.co/transformers/model_doc/camembert.html>`_ (from FAIR, Inria, Sorbonne Université) released together with the paper `CamemBERT: a Tasty French Language Model <https://arxiv.org/abs/1911.03894>`_ by Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suarez, Yoann Dupont, Laurent Romary, Eric Villemonte de la Clergerie, Djame Seddah, and Benoît Sagot.
-11. `ALBERT <https://github.com/pytorch/fairseq/tree/master/examples/albert>`_ (from Google Research), released together with the paper a `ALBERT: A Lite BERT for Self-supervised Learning of Language Representations <https://arxiv.org/abs/1909.11942>`_ by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+11. `ALBERT <https://github.com/google-research/ALBERT>`_ (from Google Research), released together with the paper a `ALBERT: A Lite BERT for Self-supervised Learning of Language Representations <https://arxiv.org/abs/1909.11942>`_ by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+12. `XLM-RoBERTa <https://github.com/pytorch/fairseq/tree/master/examples/xlmr>`_ (from Facebook AI), released together with the paper `Unsupervised Cross-lingual Representation Learning at Scale <https://arxiv.org/abs/1911.02116>`_ by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
 
 .. toctree::
     :maxdepth: 2
@@ -58,6 +59,7 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
     installation
     quickstart
     pretrained_models
+    model_sharing
     examples
     notebooks
     serialization

docs/source/installation.md

@@ -24,15 +24,24 @@ pip install [--editable] .
 
 An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples).
 
-Tests can be run using `pytest` (install pytest if needed with `pip install pytest`).
+Tests can be run using `unittest` or `pytest` (install pytest if needed with `pip install pytest`).
 
 Run all the tests from the root of the cloned repository with the commands:
 
+```bash
+python -m unittest discover -s transformers/tests -p "*test.py" -t .
+python -m unittest discover -s examples -p "*test.py" -t examples
+```
+
+or
+
 ``` bash
 python -m pytest -sv ./transformers/tests/
 python -m pytest -sv ./examples/
 ```
 
+By default, slow tests are skipped. Set the `RUN_SLOW` environment variable to `yes` to run them.
+
 ## OpenAI GPT original tokenization workflow
 
 If you want to reproduce the original tokenization process of the `OpenAI GPT` paper, you will need to install `ftfy` (use version 4.4.3 if you are using Python 2) and `SpaCy`:

docs/source/main_classes/optimizer_schedules.rst

@@ -5,6 +5,7 @@ The ``.optimization`` module provides:
 
 - an optimizer with weight decay fixed that can be used to fine-tuned models, and
 - several schedules in the form of schedule objects that inherit from ``_LRSchedule``:
+- a gradient accumulation class to accumulate the gradients of multiple batches
 
 ``AdamW``
 ~~~~~~~~~~~~~~~~
@@ -12,6 +13,15 @@ The ``.optimization`` module provides:
 .. autoclass:: transformers.AdamW
     :members:
 
+``AdamWeightDecay``
+~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AdamWeightDecay
+    :members:
+
+.. autofunction:: transformers.create_optimizer
+    :members:
+
 Schedules
 ----------------------------------------------------
 
@@ -49,3 +59,17 @@ Learning Rate Schedules
 .. image:: /imgs/warmup_linear_schedule.png
     :target: /imgs/warmup_linear_schedule.png
     :alt:
+
+``Warmup``
+~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.Warmup
+    :members:
+
+Gradient Strategies
+----------------------------------------------------
+
+``GradientAccumulator``
+~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.GradientAccumulator

docs/source/main_classes/processors.rst

@@ -54,8 +54,7 @@ Additionally, the following method  can be used to load values from a data file
 Example usage
 ^^^^^^^^^^^^^^^^^^^^^^^^^
 
-An example using these processors is given in the
-`run_glue.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_glue.py>`__ script.
+An example using these processors is given in the `run_glue.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_glue.py>`__ script.
 
 
 XNLI
@@ -74,8 +73,81 @@ This library hosts the processor to load the XNLI data:
 
 Please note that since the gold labels are available on the test set, evaluation is performed on the test set.
 
-Example usage
+An example using these processors is given in the
+`run_xnli.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_xnli.py>`__ script.
+
+
+SQuAD
+~~~~~~~~~~~~~~~~~~~~~
+
+`The Stanford Question Answering Dataset (SQuAD) <https://rajpurkar.github.io/SQuAD-explorer//>`__ is a benchmark that evaluates
+the performance of models on question answering. Two versions are available, v1.1 and v2.0. The first version (v1.1) was released together with the paper
+`SQuAD: 100,000+ Questions for Machine Comprehension of Text <https://arxiv.org/abs/1606.05250>`__. The second version (v2.0) was released alongside 
+the paper `Know What You Don't Know: Unanswerable Questions for SQuAD <https://arxiv.org/abs/1806.03822>`__.
+
+This library hosts a processor for each of the two versions:
+
+Processors
 ^^^^^^^^^^^^^^^^^^^^^^^^^
 
-An example using these processors is given in the
-`run_xnli.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_xnli.py>`__ script.
+Those processors are:
+    - :class:`~transformers.data.processors.utils.SquadV1Processor`
+    - :class:`~transformers.data.processors.utils.SquadV2Processor`
+
+They both inherit from the abstract class :class:`~transformers.data.processors.utils.SquadProcessor`
+
+.. autoclass:: transformers.data.processors.squad.SquadProcessor
+    :members:
+
+Additionally, the following method can be used to convert SQuAD examples into :class:`~transformers.data.processors.utils.SquadFeatures`
+that can be used as model inputs.
+
+.. automethod:: transformers.data.processors.squad.squad_convert_examples_to_features
+
+These processors as well as the aforementionned method can be used with files containing the data as well as with the `tensorflow_datasets` package.
+Examples are given below.
+
+
+Example usage
+^^^^^^^^^^^^^^^^^^^^^^^^^
+Here is an example using the processors as well as the conversion method using data files:
+
+Example::
+
+    # Loading a V2 processor
+    processor = SquadV2Processor()
+    examples = processor.get_dev_examples(squad_v2_data_dir)
+
+    # Loading a V1 processor
+    processor = SquadV1Processor()
+    examples = processor.get_dev_examples(squad_v1_data_dir)
+
+    features = squad_convert_examples_to_features( 
+        examples=examples,
+        tokenizer=tokenizer,
+        max_seq_length=max_seq_length,
+        doc_stride=args.doc_stride,
+        max_query_length=max_query_length,
+        is_training=not evaluate,
+    )
+
+Using `tensorflow_datasets` is as easy as using a data file:
+
+Example::
+
+    # tensorflow_datasets only handle Squad V1.
+    tfds_examples = tfds.load("squad")
+    examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
+
+    features = squad_convert_examples_to_features( 
+        examples=examples,
+        tokenizer=tokenizer,
+        max_seq_length=max_seq_length,
+        doc_stride=args.doc_stride,
+        max_query_length=max_query_length,
+        is_training=not evaluate,
+    )
+
+
+Another example using these processors is given in the
+`run_squad.py <https://github.com/huggingface/transformers/blob/master/examples/run_squad.py>`__ script.

docs/source/model_sharing.md

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

docs/source/pretrained_models.rst

@@ -3,6 +3,7 @@ Pretrained models
 
 Here is the full list of the currently provided pretrained models together with a short presentation of each model.
 
+For a list that includes community-uploaded models, refer to `https://huggingface.co/models <https://huggingface.co/models>`__.
 
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | Architecture      | Shortcut name                                              | Details of the model                                                                                                                  |
@@ -61,6 +62,32 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``bert-base-german-dbmdz-uncased``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | Trained on uncased German text by DBMDZ                                                                                             |
 |                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-japanese``                                     | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece.                                                               |
+|                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                          |
+|                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-japanese-whole-word-masking``                  | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized with MeCab and WordPiece.                                      |
+|                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                          |
+|                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-japanese-char``                                | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on Japanese text. Text is tokenized into characters.                                                                        |
+|                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-japanese-char-whole-word-masking``             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized into characters.                                               |
+|                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-finnish-cased-v1``                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on cased Finnish text.                                                                                                      |
+|                   |                                                            | (see `details on turkunlp.org <http://turkunlp.org/FinBERT/>`__).                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-finnish-uncased-v1``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on uncased Finnish text.                                                                                                    |
+|                   |                                                            | (see `details on turkunlp.org <http://turkunlp.org/FinBERT/>`__).                                                                     |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | GPT               | ``openai-gpt``                                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | OpenAI GPT English model                                                                                                            |
@@ -128,6 +155,10 @@ Here is the full list of the currently provided pretrained models together with
 |                   |                                                            | | ``roberta-large`` fine-tuned on `MNLI <http://www.nyu.edu/projects/bowman/multinli/>`__.                                            |
 |                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__)                                                   |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilroberta-base``                                     | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
+|                   |                                                            | | The DistilRoBERTa model distilled from the RoBERTa model `roberta-base` checkpoint.                                                 |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``roberta-base-openai-detector``                           | | 12-layer, 768-hidden, 12-heads, 125M parameters                                                                                     |
 |                   |                                                            | | ``roberta-base`` fine-tuned by OpenAI on the outputs of the 1.5B-parameter GPT-2 model.                                             |
 |                   |                                                            | (see `details <https://github.com/openai/gpt-2-output-dataset/tree/master/detector>`__)                                               |
@@ -148,13 +179,13 @@ Here is the full list of the currently provided pretrained models together with
 |                   |                                                            | | The DistilGPT2 model distilled from the GPT2 model `gpt2` checkpoint.                                                               |
 |                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``distilroberta-base``                                     | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
-|                   |                                                            | | The DistilRoBERTa model distilled from the RoBERTa model `roberta-base` checkpoint.                                                 |
-|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
-|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``distilbert-base-german-cased``                           | | 6-layer, 768-hidden, 12-heads, 66M parameters                                                                                       |
 |                   |                                                            | | The German DistilBERT model distilled from the German DBMDZ BERT model `bert-base-german-dbmdz-cased` checkpoint.                   |
 |                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilbert-base-multilingual-cased``                     | | 6-layer, 768-hidden, 12-heads, 134M parameters                                                                                      |
+|                   |                                                            | | The multilingual DistilBERT model distilled from the Multilingual BERT model `bert-base-multilingual-cased` checkpoint.             |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | CTRL              | ``ctrl``                                                   | | 48-layer, 1280-hidden, 16-heads, 1.6B parameters                                                                                    |
 |                   |                                                            | | Salesforce's Large-sized CTRL English model                                                                                         |
@@ -165,35 +196,56 @@ Here is the full list of the currently provided pretrained models together with
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | ALBERT            | ``albert-base-v1``                                         | | 12 repeating layers, 128 embedding, 768-hidden, 12-heads, 11M parameters                                                            |
 |                   |                                                            | | ALBERT base model                                                                                                                   |
-|                   |                                                            | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__)                                             |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``albert-large-v1``                                        | | 24 repeating layers, 128 embedding, 1024-hidden, 16-heads, 17M parameters                                                           |
 |                   |                                                            | | ALBERT large model                                                                                                                  |
-|                   |                                                            | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__)                                             |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``albert-xlarge-v1``                                       | | 24 repeating layers, 128 embedding, 2048-hidden, 16-heads, 58M parameters                                                           |
 |                   |                                                            | | ALBERT xlarge model                                                                                                                 |
-|                   |                                                            | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__)                                             |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``albert-xxlarge-v1``                                      | | 12 repeating layer, 128 embedding, 4096-hidden, 64-heads, 223M parameters                                                           |
 |                   |                                                            | | ALBERT xxlarge model                                                                                                                |
-|                   |                                                            | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__)                                             |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``albert-base-v2``                                         | | 12 repeating layers, 128 embedding, 768-hidden, 12-heads, 11M parameters                                                            |
 |                   |                                                            | | ALBERT base model with no dropout, additional training data and longer training                                                     |
-|                   |                                                            | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__)                                             |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``albert-large-v2``                                        | | 24 repeating layers, 128 embedding, 1024-hidden, 16-heads, 17M parameters                                                           |
 |                   |                                                            | | ALBERT large model with no dropout, additional training data and longer training                                                    |
-|                   |                                                            | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__)                                             |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``albert-xlarge-v2``                                       | | 24 repeating layers, 128 embedding, 2048-hidden, 16-heads, 58M parameters                                                           |
 |                   |                                                            | | ALBERT xlarge model with no dropout, additional training data and longer training                                                   |
-|                   |                                                            | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__)                                             |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``albert-xxlarge-v2``                                      | | 12 repeating layer, 128 embedding, 4096-hidden, 64-heads, 223M parameters                                                           |
 |                   |                                                            | | ALBERT xxlarge model with no dropout, additional training data and longer training                                                  |
-|                   |                                                            | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__)                                             |
+|                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| T5                | ``t5-small``                                               | | ~60M parameters with 6-layers, 512-hidden-state, 2048 feed-forward hidden-state, 8-heads,                                           |
+|                   |                                                            | | Trained on English text: the Colossal Clean Crawled Corpus (C4)                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``t5-base``                                                | | ~220M parameters with 12-layers, 768-hidden-state, 3072 feed-forward hidden-state, 12-heads,                                        |
+|                   |                                                            | | Trained on English text: the Colossal Clean Crawled Corpus (C4)                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``t5-large``                                               | | ~770M parameters with 24-layers, 1024-hidden-state, 4096 feed-forward hidden-state, 16-heads,                                       |
+|                   |                                                            | | Trained on English text: the Colossal Clean Crawled Corpus (C4)                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``t5-3B``                                                  | | ~2.8B parameters with 24-layers, 1024-hidden-state, 16384 feed-forward hidden-state, 32-heads,                                      |
+|                   |                                                            | | Trained on English text: the Colossal Clean Crawled Corpus (C4)                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``t5-11B``                                                 | | ~11B parameters with 24-layers, 1024-hidden-state, 65536 feed-forward hidden-state, 128-heads,                                      |
+|                   |                                                            | | Trained on English text: the Colossal Clean Crawled Corpus (C4)                                                                     |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| XLM-RoBERTa       | ``xlm-roberta-base``                                       | | ~125M parameters with 12-layers, 768-hidden-state, 3072 feed-forward hidden-state, 8-heads,                                         |
+|                   |                                                            | | Trained on on 2.5 TB of newly created clean CommonCrawl data in 100 languages                                                       |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlm-roberta-large``                                      | | ~355M parameters with 24-layers, 1027-hidden-state, 4096 feed-forward hidden-state, 16-heads,                                       |
+|                   |                                                            | | Trained on 2.5 TB of newly created clean CommonCrawl data in 100 languages                                                          |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 
 

docs/source/quickstart.md

@@ -219,4 +219,97 @@ sequence = tokenizer.decode(generated)
 print(sequence)
 ```
 
-The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.
+The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.
+
+### Model2Model example
+
+Encoder-decoder architectures require two tokenized inputs: one for the encoder and the other one for the decoder. Let's assume that we want to use `Model2Model` for generative question answering, and start by tokenizing the question and answer that will be fed to the model.
+
+```python
+import torch
+from transformers import BertTokenizer, Model2Model
+
+# OPTIONAL: if you want to have more information on what's happening under the hood, activate the logger as follows
+import logging
+logging.basicConfig(level=logging.INFO)
+
+# Load pre-trained model tokenizer (vocabulary)
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+# Encode the input to the encoder (the question)
+question = "Who was Jim Henson?"
+encoded_question = tokenizer.encode(question)
+
+# Encode the input to the decoder (the answer)
+answer = "Jim Henson was a puppeteer"
+encoded_answer = tokenizer.encode(answer)
+
+# Convert inputs to PyTorch tensors
+question_tensor = torch.tensor([encoded_question])
+answer_tensor = torch.tensor([encoded_answer])
+```
+
+Let's see how we can use `Model2Model` to get the value of the loss associated with this (question, answer) pair:
+
+```python
+# In order to compute the loss we need to provide language model
+# labels (the token ids that the model should have produced) to
+# the decoder.
+lm_labels =  encoded_answer
+labels_tensor = torch.tensor([lm_labels])
+
+# Load pre-trained model (weights)
+model = Model2Model.from_pretrained('bert-base-uncased')
+
+# Set the model in evaluation mode to deactivate the DropOut modules
+# This is IMPORTANT to have reproducible results during evaluation!
+model.eval()
+
+# If you have a GPU, put everything on cuda
+question_tensor = question_tensor.to('cuda')
+answer_tensor = answer_tensor.to('cuda')
+labels_tensor = labels_tensor.to('cuda')
+model.to('cuda')
+
+# Predict hidden states features for each layer
+with torch.no_grad():
+    # See the models docstrings for the detail of the inputs
+    outputs = model(question_tensor, answer_tensor, decoder_lm_labels=labels_tensor)
+    # Transformers models always output tuples.
+    # See the models docstrings for the detail of all the outputs
+    # In our case, the first element is the value of the LM loss 
+    lm_loss = outputs[0]
+```
+
+This loss can be used to fine-tune `Model2Model` on the question answering task. Assuming that we fine-tuned the model, let us now see how to generate an answer:
+
+```python
+# Let's re-use the previous question
+question = "Who was Jim Henson?"
+encoded_question = tokenizer.encode(question)
+question_tensor = torch.tensor([encoded_question])
+
+# This time we try to generate the answer, so we start with an empty sequence
+answer = "[CLS]"
+encoded_answer = tokenizer.encode(answer, add_special_tokens=False)
+answer_tensor = torch.tensor([encoded_answer])
+
+# Load pre-trained model (weights)
+model = Model2Model.from_pretrained('fine-tuned-weights')
+model.eval()
+
+# If you have a GPU, put everything on cuda
+question_tensor = encoded_question.to('cuda')
+answer_tensor = encoded_answer.to('cuda')
+model.to('cuda')
+
+# Predict all tokens
+with torch.no_grad():
+    outputs = model(question_tensor, answer_tensor)
+    predictions = outputs[0]
+
+# confirm we were able to predict 'jim'
+predicted_index = torch.argmax(predictions[0, -1]).item()
+predicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]
+assert predicted_token == 'jim'
+```

examples/README.md

@@ -24,7 +24,6 @@ pip install -r ./examples/requirements.txt
 | [Multiple Choice](#multiple-choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks. 
 | [Named Entity Recognition](#named-entity-recognition) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training.                                                                                  |
 | [XNLI](#xnli) | Examples running BERT/XLM on the XNLI benchmark. |
-| [Abstractive summarization](#abstractive-summarization) | Fine-tuning the library models for abstractive summarization tasks on the CNN/Daily Mail dataset. |
 
 ## TensorFlow 2.0 Bert models on GLUE
 
@@ -44,7 +43,7 @@ Quick benchmarks from the script (no other modifications):
 | Titan V | AMP | 26s | 0.8281/0.8568/0.8411 |
 | V100    | FP32 | 35s | 0.8646/0.8359/0.8464 |
 | V100    | AMP | 22s | 0.8646/0.8385/0.8411 |
-| 1080 Ti | FP32 | 55s | - | 
+| 1080 Ti | FP32 | 55s | - |
 
 Mixed precision (AMP) reduces the training time considerably for the same hardware and hyper-parameters (same batch size was used).
 
@@ -358,9 +357,9 @@ eval_loss = 0.44457291918821606
 
 Based on the script [`run_squad.py`](https://github.com/huggingface/transformers/blob/master/examples/run_squad.py).
 
-#### Fine-tuning on SQuAD
+#### Fine-tuning BERT on SQuAD1.0
 
-This example code fine-tunes BERT on the SQuAD dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) 
+This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) 
 on a single tesla V100 16GB. The data for SQuAD can be downloaded with the following links and should be saved in a 
 $SQUAD_DIR directory.
 
@@ -368,6 +367,12 @@ $SQUAD_DIR directory.
 * [dev-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json)
 * [evaluate-v1.1.py](https://github.com/allenai/bi-att-flow/blob/master/squad/evaluate-v1.1.py)
 
+And for SQuAD2.0, you need to download:
+
+- [train-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json)
+- [dev-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json)
+- [evaluate-v2.0.py](https://worksheets.codalab.org/rest/bundles/0x6b567e1cf2e041ec80d7098f031c5c9e/contents/blob/)
+
 ```bash
 export SQUAD_DIR=/path/to/SQUAD
 
@@ -397,7 +402,7 @@ exact_match = 81.22
 #### Distributed training
 
 
-Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD:
+Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD1.0:
 
 ```bash
 python -m torch.distributed.launch --nproc_per_node=8 run_squad.py \
@@ -427,79 +432,11 @@ exact_match = 86.91
 This fine-tuned model is available as a checkpoint under the reference
 `bert-large-uncased-whole-word-masking-finetuned-squad`.
 
-### `run_semeval.py`: relationship classification using R-Bert
-
-R-BERT is a relationship classification head for BERT and RoBERTa, described [here](https://arxiv.org/pdf/1905.08284.pdf").
-
-This example code fine-tunes R-BERT on the semeval 2010 Task 8 dataset:
- 
-```bash
-python ./examples/run_semeval.py \
---data_dir $SEMEVAL_DIR \
---output_dir $RESULTS_DIR \
---model_name_or_path bert-base-uncased \
---do_train \
---do_eval \
---overwrite_output_dir \
---num_train_epochs 8.0 \
---per_gpu_train_batch_size 16 \
---per_gpu_eval_batch_size 16 \
---learning_rate 2e-5 \
---max_seq_length 128 \
---task_name semeval2010_task8 \
---train_on_other_labels \
---eval_on_other_labels \
---include_directionality
-
-```
-The ```$SEMEVAL_DIR``` should point to the extracted archive.
-
-The ```--include_directionality``` flag trains a classifier using all 18 semeval classes. The 
-```--train_on_other_labels``` and ```--eval_on_other_labels``` flags also include instances labeled as 'Other' in the 
-training and evaluation respectively. Include all of these to be able to use the official evaluation script. 
-
-Note, although an F1 score is calculated in the python code, two additional files are also written out at the checkpoint 
-intervals ```{global_step}_semeval_results.tsv``` that may be used with the official Semeval evaluation script 
-(supplied in the semeval data archive). The dataset is available under Creative Commons Atrribution 3.0 
-Unported Licence (http://creativecommons.org/licenses/by/3.0/) and is available 
-[here](http://docs.google.com/leaf?id=0B_jQiLugGTAkMDQ5ZjZiMTUtMzQ1Yy00YWNmLWJlZDYtOWY1ZDMwY2U4YjFk&sort=name&layout=list&num=50). 
-
-for example:
-
-```bash
-./semeval2010_task8_scorer-v1.2.pl $SEMEVAL_DIR/{global_step}_semeval_results.tsv $SEMEVAL_DIR/TEST_FILE_SEMEVAL_SCRIPT_FORMAT.tsv 
-```
-
-
-Using the ```bert-base-uncased``` should give around 87.6 undirected F1 (excluding other), with the official script. 
-However, the RoBERTa model can also be used with this head:
-
-```bash
-python ./examples/run_semeval.py \
---data_dir $SEMEVAL_DIR \
---output_dir $RESULTS_DIR \
---model_name_or_path roberta-large \
---model_type roberta \
---do_train \
---do_eval \
---overwrite_output_dir \
---num_train_epochs 8.0 \
---per_gpu_train_batch_size 16 \
---per_gpu_eval_batch_size 16 \
---learning_rate 2e-5 \
---max_seq_length 128 \
---task_name semeval2010_task8 \
---train_on_other_labels \
---eval_on_other_labels \
---include_directionality
-
-```
-
-This should give an undirected F1 of aroubd 90.8 (excluding other) with the official script.
-
 #### Fine-tuning XLNet on SQuAD
 
-This example code fine-tunes XLNet on the SQuAD dataset. See above to download the data for SQuAD .
+This example code fine-tunes XLNet on both SQuAD1.0 and SQuAD2.0 dataset. See above to download the data for SQuAD .
+
+##### Command for SQuAD1.0:
 
 ```bash
 export SQUAD_DIR=/path/to/SQUAD
@@ -522,7 +459,32 @@ python /data/home/hlu/transformers/examples/run_squad.py \
     --save_steps 5000
 ```
 
-Training with the previously defined hyper-parameters yields the following results:
+##### Command for SQuAD2.0:
+
+```bash
+export SQUAD_DIR=/path/to/SQUAD
+
+python run_squad.py \
+    --model_type xlnet \
+    --model_name_or_path xlnet-large-cased \
+    --do_train \
+    --do_eval \
+    --version_2_with_negative \
+    --train_file $SQUAD_DIR/train-v2.0.json \
+    --predict_file $SQUAD_DIR/dev-v2.0.json \
+    --learning_rate 3e-5 \
+    --num_train_epochs 4 \
+    --max_seq_length 384 \
+    --doc_stride 128 \
+    --output_dir ./wwm_cased_finetuned_squad/ \
+    --per_gpu_eval_batch_size=2  \
+    --per_gpu_train_batch_size=2   \
+    --save_steps 5000
+```
+
+Larger batch size may improve the performance while costing more memory.
+
+##### Results for SQuAD1.0 with the previously defined hyper-parameters:
 
 ```python
 {
@@ -535,9 +497,28 @@ Training with the previously defined hyper-parameters yields the following resul
 }
 ```
 
+##### Results for SQuAD2.0 with the previously defined hyper-parameters:
+
+```python
+{
+"exact": 80.4177545691906,
+"f1": 84.07154997729623,
+"total": 11873,
+"HasAns_exact": 76.73751686909581,
+"HasAns_f1": 84.05558584352873,
+"HasAns_total": 5928,
+"NoAns_exact": 84.0874684608915,
+"NoAns_f1": 84.0874684608915,
+"NoAns_total": 5945
+}
+```
+
+
+
 ## Named Entity Recognition
 
-Based on the script [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py).
+Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py) for Pytorch and
+[`run_tf_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_tf_ner.py) for Tensorflow 2.
 This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
 Details and results for the fine-tuning provided by @stefan-it.
 
@@ -582,7 +563,7 @@ The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so
 cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
 ```
 
-### Training
+### Prepare the run
 
 Additional environment variables must be set:
 
@@ -594,6 +575,8 @@ export SAVE_STEPS=750
 export SEED=1
 ```
 
+### Run the Pytorch version
+
 To start training, just run:
 
 ```bash
@@ -614,7 +597,7 @@ python3 run_ner.py --data_dir ./ \
 
 If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
 
-### Evaluation
+#### Evaluation
 
 Evaluation on development dataset outputs the following for our example:
 
@@ -636,7 +619,7 @@ On the test dataset the following results could be achieved:
 10/04/2019 00:42:42 - INFO - __main__ -     recall = 0.8624150210424085
 ```
 
-### Comparing BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased)
+#### Comparing BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased)
 
 Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased) with the same hyperparameters as specified in the [example documentation](https://huggingface.co/transformers/examples.html#named-entity-recognition) (one run):
 
@@ -646,32 +629,70 @@ Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) a
 | `roberta-large`                  | 95.96 | 91.87
 | `distilbert-base-uncased` | 94.34 | 90.32
 
-## Abstractive summarization
+### Run the Tensorflow 2 version
 
-Based on the script
-[`run_summarization_finetuning.py`](https://github.com/huggingface/transformers/blob/master/examples/run_summarization_finetuning.py).
-
-Before running this script you should download **both** CNN and Daily Mail
-datasets from [Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/)  (the
-links next to "Stories") in the same folder. Then uncompress the archives by running:
+To start training, just run:
 
 ```bash
-tar -xvf cnn_stories.tgz && tar -xvf dailymail_stories.tgz
+python3 run_tf_ner.py --data_dir ./ \
+--model_type bert \
+--labels ./labels.txt \
+--model_name_or_path $BERT_MODEL \
+--output_dir $OUTPUT_DIR \
+--max_seq_length  $MAX_LENGTH \
+--num_train_epochs $NUM_EPOCHS \
+--per_device_train_batch_size $BATCH_SIZE \
+--save_steps $SAVE_STEPS \
+--seed $SEED \
+--do_train \
+--do_eval \
+--do_predict
 ```
 
-note that the finetuning script **will not work** if you do not download both
-datasets. We will refer as `$DATA_PATH` the path to where you uncompressed both
-archive.
+Such as the Pytorch version, if your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
 
+#### Evaluation
+
+Evaluation on development dataset outputs the following for our example:
 ```bash
-export DATA_PATH=/path/to/dataset/
+           precision    recall  f1-score   support
 
-python run_summarization_finetuning.py \
-    --output_dir=output \
-    --model_type=bert2bert \
-    --model_name_or_path=bert2bert \
-    --do_train \
-    --data_path=$DATA_PATH \
+ LOCderiv     0.7619    0.6154    0.6809        52
+  PERpart     0.8724    0.8997    0.8858      4057
+  OTHpart     0.9360    0.9466    0.9413       711
+  ORGpart     0.7015    0.6989    0.7002       269
+  LOCpart     0.7668    0.8488    0.8057       496
+      LOC     0.8745    0.9191    0.8963       235
+ ORGderiv     0.7723    0.8571    0.8125        91
+ OTHderiv     0.4800    0.6667    0.5581        18
+      OTH     0.5789    0.6875    0.6286        16
+ PERderiv     0.5385    0.3889    0.4516        18
+      PER     0.5000    0.5000    0.5000         2
+      ORG     0.0000    0.0000    0.0000         3
+
+micro avg     0.8574    0.8862    0.8715      5968
+macro avg     0.8575    0.8862    0.8713      5968
+```
+
+On the test dataset the following results could be achieved:
+```bash
+           precision    recall  f1-score   support
+
+  PERpart     0.8847    0.8944    0.8896      9397
+  OTHpart     0.9376    0.9353    0.9365      1639
+  ORGpart     0.7307    0.7044    0.7173       697
+      LOC     0.9133    0.9394    0.9262       561
+  LOCpart     0.8058    0.8157    0.8107      1150
+      ORG     0.0000    0.0000    0.0000         8
+ OTHderiv     0.5882    0.4762    0.5263        42
+ PERderiv     0.6571    0.5227    0.5823        44
+      OTH     0.4906    0.6667    0.5652        39
+ ORGderiv     0.7016    0.7791    0.7383       172
+ LOCderiv     0.8256    0.6514    0.7282       109
+      PER     0.0000    0.0000    0.0000        11
+
+micro avg     0.8722    0.8774    0.8748     13869
+macro avg     0.8712    0.8774    0.8740     13869
 ```
 
 ## XNLI

examples/contrib/run_openai_gpt.py

@@ -75,7 +75,7 @@ def pre_process_datasets(encoded_datasets, input_len, cap_length, start_token, d
         n_batch = len(dataset)
         input_ids = np.zeros((n_batch, 2, input_len), dtype=np.int64)
         mc_token_ids = np.zeros((n_batch, 2), dtype=np.int64)
-        lm_labels = np.full((n_batch, 2, input_len), fill_value=-1, dtype=np.int64)
+        lm_labels = np.full((n_batch, 2, input_len), fill_value=-100, dtype=np.int64)
         mc_labels = np.zeros((n_batch,), dtype=np.int64)
         for i, (story, cont1, cont2, mc_label), in enumerate(dataset):
             with_cont1 = [start_token] + story[:cap_length] + [delimiter_token] + cont1[:cap_length] + [clf_token]

examples/distillation/README.md

@@ -2,6 +2,8 @@
 
 This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT, DistilRoBERTa and DistilGPT2.
 
+**December 6th, 2019 - Update** We release **DistilmBERT**: 92% of `bert-base-multilingual-cased` on XNLI. The model supports 104 different languages listed [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages).
+
 **November 19th, 2019 - Update** We release German **DistilBERT**: 98.8% of `bert-base-german-dbmdz-cased` on NER tasks.
 
 **October 23rd, 2019 - Update** We release **DistilRoBERTa**: 95% of `RoBERTa-base`'s performance on GLUE, twice as fast as RoBERTa while being 35% smaller.
@@ -17,8 +19,9 @@ Distil* is a class of compressed models that started with DistilBERT. DistilBERT
 
 We have applied the same method to other Transformer architectures and released the weights:
 - GPT2: on the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 15.0 compared to 18.5 for **DistilGPT2** (after fine-tuning on the train set).
-- RoBERTa: **DistilRoBERTa** reaches 95% of `RoBERTa-base` performance on GLUE while being twice faster and 35% smaller.
-- and more to come! 🤗🤗🤗
+- RoBERTa: **DistilRoBERTa** reaches 95% of `RoBERTa-base`'s performance on GLUE while being twice faster and 35% smaller.
+- German BERT: **German DistilBERT** reaches 99% of `bert-base-german-dbmdz-cased`'s performance on German NER (CoNLL-2003).
+- Multilingual BERT: **DistilmBERT** reaches 92% of Multilingual BERT's performance on XNLI while being twice faster and 25% smaller. The model supports 104 languages listed [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages).
 
 For more information on DistilBERT, please refer to our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108).
 
@@ -29,7 +32,7 @@ Here are the results on the dev sets of GLUE:
 | BERT-base                 |  **77.6**                      | 48.9 | 84.3 | 88.6 | 89.3 | 89.5 | 71.3 | 91.7 | 91.2 | 43.7              |
 | DistilBERT                |  **76.8**                      | 49.1 | 81.8 | 90.2 | 90.2 | 89.2 | 62.9 | 92.7 | 90.7 | 44.4              |
 | ---                       |    ---                         |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  ---              |
-| RoBERTa-base (reported)   |  **83.2**/**86.4**<sup>2</sup> | 63.6 | 87.6 | 90.2 | 92.8 | 91.9 | 78.7 | 94.8 | 91.2 | 57.7<sup>3</sup> |
+| RoBERTa-base (reported)   |  **83.2**/**86.4**<sup>2</sup> | 63.6 | 87.6 | 90.2 | 92.8 | 91.9 | 78.7 | 94.8 | 91.2 | 57.7<sup>3</sup>  |
 | DistilRoBERTa<sup>1</sup> |  **79.0**/**82.3**<sup>2</sup> | 59.4 | 83.9 | 86.6 | 90.8 | 89.4 | 67.9 | 92.5 | 88.3 | 52.1              |
 
 <sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directy perform transfer learning on the pre-trained DistilRoBERTa.
@@ -38,6 +41,14 @@ Here are the results on the dev sets of GLUE:
 
 <sup>3</sup> We compute this score ourselves for completeness.
 
+Here are the results on the *test* sets for 6 of the languages available in XNLI. The results are computed in the zero shot setting (trained on the English portion and evaluated on the target language portion):
+
+| Model                        | English | Spanish | Chinese | German | Arabic  | Urdu |
+| :---:                        | :---:   | :---:   | :---:   | :---:  | :---:   | :---:|
+| mBERT base cased (computed)  | 82.1    | 74.6    | 69.1    | 72.3   | 66.4    | 58.5 |
+| mBERT base uncased (reported)| 81.4    | 74.3    | 63.8    | 70.5   | 62.1    | 58.3 |
+| DistilmBERT                  | 78.2    | 69.1    | 64.0    | 66.3   | 59.1    | 54.7 |
+
 ## Setup
 
 This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`. 
@@ -54,7 +65,7 @@ Transformers includes five pre-trained Distil* models, currently only provided f
 - `distilbert-base-german-cased`: DistilBERT German language model pretrained on 1/2 of the data used to pretrain Bert using distillation with the supervision of the `bert-base-german-dbmdz-cased` version of German DBMDZ Bert. For NER tasks the model reaches a F1 score of 83.49 on the CoNLL-2003 test set (for comparison, `bert-base-german-dbmdz-cased` reaches a 84.52 F1 score), and a F1 score of 85.23 on the GermEval 2014 test set (`bert-base-german-dbmdz-cased` reaches a 86.89 F1 score).
 - `distilgpt2`: DistilGPT2 English language model pretrained with the supervision of `gpt2` (the smallest version of GPT2) on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset. The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
 - `distilroberta-base`: DistilRoBERTa English language model pretrained with the supervision of `roberta-base` solely on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset (it is ~4 times less training data than the teacher RoBERTa). The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 125M parameters for RoBERTa-base). On average DistilRoBERTa is twice as fast as Roberta-base.
-- and more to come! 🤗🤗🤗
+- `distilbert-base-multilingual-cased`: DistilmBERT multilingual model pretrained with the supervision of `bert-base-multilingual-cased` on the concatenation of Wikipedia in 104 different languages. The model supports the 104 languages listed [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages). The model has 6 layers, 768 dimension and 12 heads, totalizing 134M parameters (compared to 177M parameters for mBERT-base). On average DistilmBERT is twice as fast as mBERT-base.
 
 Using DistilBERT is very similar to using BERT. DistilBERT share the same tokenizer as BERT's `bert-base-uncased` even though we provide a link to this tokenizer under the `DistilBertTokenizer` name to have a consistent naming between the library models.
 
@@ -70,6 +81,7 @@ last_hidden_states = outputs[0]  # The last hidden-state is the first element of
 Similarly, using the other Distil* models simply consists in calling the base classes with a different pretrained checkpoint:
 - DistilGPT2: `model = GPT2Model.from_pretrained('distilgpt2')`
 - DistilRoBERTa: `model = RobertaModel.from_pretrained('distilroberta-base')`
+- DistilmBERT: `model = DistilBertModel.from_pretrained('distilbert-base-multilingual-cased')`
 
 
 ## How to train Distil*

examples/distillation/distiller.py

@@ -21,7 +21,6 @@ import psutil
 import time
 from tqdm import trange, tqdm
 import numpy as np
-import psutil
 
 import torch
 import torch.nn as nn
@@ -113,7 +112,7 @@ class Distiller:
         self.last_log = 0
 
         self.ce_loss_fct = nn.KLDivLoss(reduction='batchmean')
-        self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
+        self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
         if self.alpha_mse > 0.:
             self.mse_loss_fct = nn.MSELoss(reduction='sum')
         if self.alpha_cos > 0.:
@@ -187,7 +186,7 @@ class Distiller:
         -------
             token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
             attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
-            mlm_labels: `torch.tensor(bs, seq_length)` - The masked languge modeling labels. There is a -1 where there is nothing to predict.
+            mlm_labels: `torch.tensor(bs, seq_length)` - The masked languge modeling labels. There is a -100 where there is nothing to predict.
         """
         token_ids, lengths = batch
         token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
@@ -225,7 +224,7 @@ class Distiller:
         _token_ids = _token_ids_mask * (probs == 0).long() + _token_ids_real * (probs == 1).long() + _token_ids_rand * (probs == 2).long()
         token_ids = token_ids.masked_scatter(pred_mask, _token_ids)
 
-        mlm_labels[~pred_mask] = -1 # previously `mlm_labels[1-pred_mask] = -1`, cf pytorch 1.2.0 compatibility
+        mlm_labels[~pred_mask] = -100 # previously `mlm_labels[1-pred_mask] = -1`, cf pytorch 1.2.0 compatibility
 
         # sanity checks
         assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size
@@ -247,7 +246,7 @@ class Distiller:
         -------
             token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
             attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
-            clm_labels: `torch.tensor(bs, seq_length)` - The causal languge modeling labels. There is a -1 where there is nothing to predict.
+            clm_labels: `torch.tensor(bs, seq_length)` - The causal languge modeling labels. There is a -100 where there is nothing to predict.
         """
         token_ids, lengths = batch
         token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
@@ -255,7 +254,7 @@ class Distiller:
 
         attn_mask = (torch.arange(token_ids.size(1), dtype=torch.long, device=lengths.device) < lengths[:, None])
         clm_labels = token_ids.new(token_ids.size()).copy_(token_ids)
-        clm_labels[~attn_mask] = -1 # previously `clm_labels[1-attn_mask] = -1`, cf pytorch 1.2.0 compatibility
+        clm_labels[~attn_mask] = -100 # previously `clm_labels[1-attn_mask] = -1`, cf pytorch 1.2.0 compatibility
 
         # sanity checks
         assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size

examples/distillation/requirements.txt

@@ -3,4 +3,4 @@ tensorboard>=1.14.0
 tensorboardX==1.8
 psutil==5.6.3
 scipy==1.3.1
-transformers==2.0.0
+transformers

examples/pplm/README.md

@@ -1,17 +1,15 @@
-# PPLM
+# Plug and Play Language Models: a Simple Approach to Controlled Text Generation
+
+Authors: [Sumanth Dathathri](https://dathath.github.io/), [Andrea Madotto](https://andreamad8.github.io/), Janice Lan, Jane Hung, Eric Frank, [Piero Molino](https://w4nderlu.st/), [Jason Yosinski](http://yosinski.com/), and [Rosanne Liu](http://www.rosanneliu.com/)
 
 This folder contains the original code used to run the Plug and Play Language Model (PPLM).
-![header image](./imgs/headfigure.png)
 
-## Plug and Play Language Models: a Simple Approach to Steerable Text Generation
-Authors: [Sumanth Dathathri](https://dathath.github.io/), Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, [Piero Molino](https://w4nderlu.st/), [Jason Yosinski](http://yosinski.com/), and [Rosanne Liu](http://www.rosanneliu.com/)
-
-PPLM allows a user to flexibly plug in one or more tiny attribute models representing the desired steering objective into a large, unconditional LM. The method has the key property that it uses the LM _as is_---no training or fine-tuning is required---which enables researchers to leverage best-in-class LMs even if they do not have the extensive hardware required to train them.
-
-Paper link: 
+Paper link: https://arxiv.org/abs/1912.02164
 
 Blog link: https://eng.uber.com/pplm
 
+Please check out the repo under uber-research for more information: https://github.com/uber-research/PPLM
+
 
 ## Setup
 
@@ -27,7 +25,7 @@ cd examples/pplm
 ### Example command for bag-of-words control
 
 ```bash
-python run_pplm.py -B space --cond_text "The president" --length 100 --gamma 1.5 --num_iterations 3 --num_samples 1 --stepsize 0.01 --window_length 5 --kl_scale 0.01 --gm_scale 0.95
+python run_pplm.py -B military --cond_text "The potato" --length 50 --gamma 1.5 --num_iterations 3 --num_samples 10 --stepsize 0.03 --window_length 5 --kl_scale 0.01 --gm_scale 0.99 --colorama --sample
 ```
 
 ### Tuning hyperparameters for bag-of-words control
@@ -45,7 +43,7 @@ python run_pplm.py -B space --cond_text "The president" --length 100 --gamma 1.5
 ### Example command for discriminator based sentiment control
 
 ```bash
-python run_pplm.py -D sentiment --class_label 3 --cond_text "The lake" --length 10 --gamma 1.0 --num_iterations 10 --num_samples 1 --stepsize 0.03 --kl_scale 0.01 --gm_scale 0.95
+python run_pplm.py -D sentiment --class_label 2 --cond_text "My dog died" --length 50 --gamma 1.0 --num_iterations 10 --num_samples 10 --stepsize 0.04 --kl_scale 0.01 --gm_scale 0.95 --sample
 ```
 
 ### Tuning hyperparameters for discriminator control
@@ -54,8 +52,3 @@ python run_pplm.py -D sentiment --class_label 3 --cond_text "The lake" --length
 
 2. Use `--class_label 3` for negative, and `--class_label 2` for positive
 
-### Example command for detoxificiation:
-
-```bash
-python run_pplm.py -D toxicity --length 100 --num_iterations 10 --cond-text 'TH PEOPLEMan goddreams Blacks' --gamma 1.0 --num_samples 10 --stepsize 0.02
-```

examples/pplm/run_pplm.py

@@ -1,18 +1,19 @@
 #! /usr/bin/env python3
 # coding=utf-8
-# Copyright 2018 The Uber AI Team Authors.
+
+#Copyright (c) 2019 Uber Technologies, Inc.
 #
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
+#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
+#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.
+#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.
 
 """
 Example command with bag of words:
@@ -45,12 +46,9 @@ SMALL_CONST = 1e-15
 BIG_CONST = 1e10
 
 BAG_OF_WORDS_ARCHIVE_MAP = {
-    'kitchen': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/kitchen.txt",
     'legal': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/legal.txt",
     'military': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/military.txt",
-    'monsters': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/monsters.txt",
     'politics': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/politics.txt",
-    'positive_words': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/positive_words.txt",
     'religion': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/religion.txt",
     'science': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/science.txt",
     'space': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/space.txt",
@@ -74,14 +72,6 @@ DISCRIMINATOR_MODELS_PARAMS = {
         "default_class": 3,
         "pretrained_model": "gpt2-medium",
     },
-    "toxicity": {
-        "url": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/discriminators/toxic_classifier_head.pt",
-        "class_size": 2,
-        "embed_size": 1024,
-        "class_vocab": {"non_toxic": 0, "toxic": 1},
-        "default_class": 0,
-        "pretrained_model": "gpt2-medium",
-    },
 }
 
 

examples/pplm/run_pplm_discrim_train.py

@@ -1,7 +1,19 @@
 #! /usr/bin/env python3
 # coding=utf-8
 
-# This code is licensed under a non-commercial license.
+#Copyright (c) 2019 Uber Technologies, Inc.
+#
+#Licensed under the Apache License, Version 2.0 (the "License");
+#you may not use this file except in compliance with the License.
+#You may obtain a copy of the License at
+#
+#http://www.apache.org/licenses/LICENSE-2.0
+#
+#Unless required by applicable law or agreed to in writing, software
+#distributed under the License is distributed on an "AS IS" BASIS,
+#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#See the License for the specific language governing permissions and
+#limitations under the License.
 
 import argparse
 import csv

examples/run_generation.py

@@ -20,14 +20,10 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import argparse
 import logging
-from tqdm import trange
 
 import torch
-import torch.nn.functional as F
 import numpy as np
 
-from transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig
-
 from transformers import GPT2LMHeadModel, GPT2Tokenizer
 from transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer
 from transformers import XLNetLMHeadModel, XLNetTokenizer
@@ -36,22 +32,22 @@ from transformers import CTRLLMHeadModel, CTRLTokenizer
 from transformers import XLMWithLMHeadModel, XLMTokenizer
 
 
-logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
-                    datefmt = '%m/%d/%Y %H:%M:%S',
-                    level = logging.INFO)
+logging.basicConfig(
+    format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+    datefmt="%m/%d/%Y %H:%M:%S",
+    level=logging.INFO,
+)
 logger = logging.getLogger(__name__)
 
 MAX_LENGTH = int(10000)  # Hardcoded max length to avoid infinite loop
 
-ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig)), ())
-
 MODEL_CLASSES = {
-    'gpt2': (GPT2LMHeadModel, GPT2Tokenizer),
-    'ctrl': (CTRLLMHeadModel, CTRLTokenizer),
-    'openai-gpt': (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
-    'xlnet': (XLNetLMHeadModel, XLNetTokenizer),
-    'transfo-xl': (TransfoXLLMHeadModel, TransfoXLTokenizer),
-    'xlm': (XLMWithLMHeadModel, XLMTokenizer),
+    "gpt2": (GPT2LMHeadModel, GPT2Tokenizer),
+    "ctrl": (CTRLLMHeadModel, CTRLTokenizer),
+    "openai-gpt": (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
+    "xlnet": (XLNetLMHeadModel, XLNetTokenizer),
+    "transfo-xl": (TransfoXLLMHeadModel, TransfoXLTokenizer),
+    "xlm": (XLMWithLMHeadModel, XLMTokenizer),
 }
 
 # Padding text to help Transformer-XL and XLNet with short prompts as proposed by Aman Rusia
@@ -75,81 +71,79 @@ def set_seed(args):
     if args.n_gpu > 0:
         torch.cuda.manual_seed_all(args.seed)
 
-
-def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
-    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
-        Args:
-            logits: logits distribution shape (batch size x vocabulary size)
-            top_k > 0: keep only top k tokens with highest probability (top-k filtering).
-            top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
-                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
-        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
-    """
-    top_k = min(top_k, logits.size(-1))  # Safety check
-    if top_k > 0:
-        # Remove all tokens with a probability less than the last token of the top-k
-        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
-        logits[indices_to_remove] = filter_value
-
-    if top_p > 0.0:
-        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
-        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
-
-        # Remove tokens with cumulative probability above the threshold
-        sorted_indices_to_remove = cumulative_probs > top_p
-        # Shift the indices to the right to keep also the first token above the threshold
-        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
-        sorted_indices_to_remove[..., 0] = 0
-
-        # scatter sorted tensors to original indexing
-        indices_to_remove = sorted_indices_to_remove.scatter(dim=1, index=sorted_indices, src=sorted_indices_to_remove)
-        logits[indices_to_remove] = filter_value
-    return logits
+#
+# Functions to prepare models' input
+#
 
 
-def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0,
-                    is_xlnet=False, is_xlm_mlm=False, xlm_mask_token=None, xlm_lang=None, device='cpu'):
-    context = torch.tensor(context, dtype=torch.long, device=device)
-    context = context.unsqueeze(0).repeat(num_samples, 1)
-    generated = context
-    with torch.no_grad():
-        for _ in trange(length):
+def prepare_ctrl_input(args, _, tokenizer, prompt_text):
+    if args.temperature > 0.7:
+        logger.info(
+            "CTRL typically works better with lower temperatures (and lower top_k)."
+        )
 
-            inputs = {'input_ids': generated}
-            if is_xlnet: 
-                # XLNet is a direct (predict same token, not next token) and bi-directional model by default
-                # => need one additional dummy token in the input (will be masked), attention mask and target mapping (see model docstring)
-                input_ids = torch.cat((generated, torch.zeros((1, 1), dtype=torch.long, device=device)), dim=1)
-                perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float, device=device)
-                perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
-                target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float, device=device)
-                target_mapping[0, 0, -1] = 1.0  # predict last token
-                inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping}
+    encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False)
+    if not any(encoded_prompt[0] == x for x in tokenizer.control_codes.values()):
+        logger.info(
+            "WARNING! You are not starting your generation from a control code so you won't get good results"
+        )
+    return prompt_text
 
-            if is_xlm_mlm and xlm_mask_token:
-                # XLM MLM models are direct models (predict same token, not next token)
-                # => need one additional dummy token in the input (will be masked and guessed)
-                input_ids = torch.cat((generated, torch.full((1, 1), xlm_mask_token, dtype=torch.long, device=device)), dim=1)
-                inputs = {'input_ids': input_ids}
 
-            if xlm_lang is not None:
-                inputs["langs"] = torch.tensor([xlm_lang] * inputs["input_ids"].shape[1], device=device).view(1, -1)
+def prepare_xlm_input(args, model, tokenizer, prompt_text):
+    # kwargs = {"language": None, "mask_token_id": None}
 
-            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states)
-            next_token_logits = outputs[0][:, -1, :] / (temperature if temperature > 0 else 1.)
+    # Set the language
+    use_lang_emb = hasattr(model.config, "use_lang_emb") and model.config.use_lang_emb
+    if hasattr(model.config, "lang2id") and use_lang_emb:
+        available_languages = model.config.lang2id.keys()
+        if args.xlm_language in available_languages:
+            language = args.xlm_language
+        else:
+            language = None
+            while language not in available_languages:
+                language = input(
+                    "Using XLM. Select language in "
+                    + str(list(available_languages))
+                    + " >>> "
+                )
+        # kwargs["language"] = tokenizer.lang2id[language]
 
-            # repetition penalty from CTRL (https://arxiv.org/abs/1909.05858)
-            for i in range(num_samples):
-                for _ in set(generated[i].tolist()):
-                    next_token_logits[i, _] /= repetition_penalty
-                
-            filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
-            if temperature == 0: # greedy sampling:
-                next_token = torch.argmax(filtered_logits, dim=-1).unsqueeze(-1)
-            else:
-                next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
-            generated = torch.cat((generated, next_token), dim=1)
-    return generated
+    # TODO fix mask_token_id setup when configurations will be synchronized between models and tokenizers
+    # XLM masked-language modeling (MLM) models need masked token
+    # is_xlm_mlm = "mlm" in args.model_name_or_path
+    # if is_xlm_mlm:
+    #     kwargs["mask_token_id"] = tokenizer.mask_token_id
+
+    return prompt_text
+
+
+def prepare_xlnet_input(args, _, tokenizer, prompt_text):
+    prompt_text = (args.padding_text if args.padding_text else PADDING_TEXT) + prompt_text
+    return prompt_text, {}
+
+
+def prepare_transfoxl_input(args, _, tokenizer, prompt_text):
+    prompt_text = (args.padding_text if args.padding_text else PADDING_TEXT) + prompt_text
+    return prompt_text, {}
+
+
+PREPROCESSING_FUNCTIONS = {
+    "ctrl": prepare_ctrl_input,
+    "xlm": prepare_xlm_input,
+    "xlnet": prepare_xlnet_input,
+    "transfo-xl": prepare_transfoxl_input,
+}
+
+
+def adjust_length_to_model(length, max_sequence_length):
+    if length < 0 and max_sequence_length > 0:
+        length = max_sequence_length
+    elif 0 < max_sequence_length < length:
+        length = max_sequence_length  # No generation bigger than model size
+    elif length < 0:
+        length = MAX_LENGTH  # avoid infinite loop
+    return length
 
 
 def main():
@@ -157,104 +151,76 @@ def main():
     parser.add_argument("--model_type", default=None, type=str, required=True,
                         help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
     parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
-                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+
     parser.add_argument("--prompt", type=str, default="")
-    parser.add_argument("--padding_text", type=str, default="")
-    parser.add_argument("--xlm_lang", type=str, default="", help="Optional language when used with the XLM model.")
     parser.add_argument("--length", type=int, default=20)
-    parser.add_argument("--num_samples", type=int, default=1)
-    parser.add_argument("--temperature", type=float, default=1.0,
-                        help="temperature of 0 implies greedy sampling")
-    parser.add_argument("--repetition_penalty", type=float, default=1.0,
-                        help="primarily useful for CTRL model; in that case, use 1.2")
-    parser.add_argument("--top_k", type=int, default=0)
-    parser.add_argument("--top_p", type=float, default=0.9)
-    parser.add_argument("--no_cuda", action='store_true',
-                        help="Avoid using CUDA when available")
-    parser.add_argument('--seed', type=int, default=42,
-                        help="random seed for initialization")
-    parser.add_argument('--stop_token', type=str, default=None,
-                        help="Token at which text generation is stopped")
+    parser.add_argument("--stop_token", type=str, default=None, help="Token at which text generation is stopped")
+
+    parser.add_argument("--temperature", type=float, default=1.0, help="temperature of 1.0 has no effect, lower tend toward greedy sampling")
+    parser.add_argument("--repetition_penalty", type=float, default=1.0, help="primarily useful for CTRL model; in that case, use 1.2")
+    parser.add_argument("--k", type=int, default=0)
+    parser.add_argument("--p", type=float, default=0.9)
+
+    parser.add_argument("--padding_text", type=str, default="", help="Padding text for Transfo-XL and XLNet.")
+    parser.add_argument("--xlm_language", type=str, default="", help="Optional language when used with the XLM model.")
+
+    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
+    parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
     args = parser.parse_args()
 
-    args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+    args.device = torch.device(
+        "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu"
+    )
     args.n_gpu = torch.cuda.device_count()
 
     set_seed(args)
 
-    args.model_type = args.model_type.lower()
-    model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    # Initialize the model and tokenizer
+    try:
+        args.model_type = args.model_type.lower()
+        model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    except KeyError:
+        raise KeyError(
+            "the model {} you specified is not supported. You are welcome to add it and open a PR :)"
+        )
+
     tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
     model = model_class.from_pretrained(args.model_name_or_path)
     model.to(args.device)
-    model.eval()
-
-    if args.length < 0 and model.config.max_position_embeddings > 0:
-        args.length = model.config.max_position_embeddings
-    elif 0 < model.config.max_position_embeddings < args.length:
-        args.length = model.config.max_position_embeddings  # No generation bigger than model size 
-    elif args.length < 0:
-        args.length = MAX_LENGTH  # avoid infinite loop
 
+    args.length = adjust_length_to_model(
+        args.length, max_sequence_length=model.config.max_position_embeddings
+    )
     logger.info(args)
-    if args.model_type in ["ctrl"]:
-        if args.temperature > 0.7:
-            logger.info('CTRL typically works better with lower temperatures (and lower top_k).')
 
-    while True:
-        xlm_lang = None
-        # XLM Language usage detailed in the issues #1414
-        if args.model_type in ["xlm"] and hasattr(tokenizer, 'lang2id') and hasattr(model.config, 'use_lang_emb') \
-                and model.config.use_lang_emb:
-            if args.xlm_lang:
-                language = args.xlm_lang
-            else:
-                language = None
-                while language not in tokenizer.lang2id.keys():
-                    language = input("Using XLM. Select language in " + str(list(tokenizer.lang2id.keys())) + " >>> ")
-            xlm_lang = tokenizer.lang2id[language]
+    prompt_text = args.prompt if args.prompt else input("Model prompt >>> ")
 
-        # XLM masked-language modeling (MLM) models need masked token (see details in sample_sequence)
-        is_xlm_mlm = args.model_type in ["xlm"] and 'mlm' in args.model_name_or_path
-        if is_xlm_mlm:
-            xlm_mask_token = tokenizer.mask_token_id
-        else:
-            xlm_mask_token = None
+    # Different models need different input formatting and/or extra arguments
+    requires_preprocessing = args.model_type in PREPROCESSING_FUNCTIONS.keys()
+    if requires_preprocessing:
+        prepare_input = PREPROCESSING_FUNCTIONS.get(args.model_type)
+        prompt_text = prepare_input(args, model, tokenizer, prompt_text)
+    encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False, return_tensors='pt')
 
-        raw_text = args.prompt if args.prompt else input("Model prompt >>> ")
-        if args.model_type in ["transfo-xl", "xlnet"]:
-            # Models with memory likes to have a long prompt for short inputs.
-            raw_text = (args.padding_text if args.padding_text else PADDING_TEXT) + raw_text
-        context_tokens = tokenizer.encode(raw_text, add_special_tokens=False)
-        if args.model_type == "ctrl":
-            if not any(context_tokens[0] == x for x in tokenizer.control_codes.values()):
-                logger.info("WARNING! You are not starting your generation from a control code so you won't get good results")
-        out = sample_sequence(
-            model=model,
-            context=context_tokens,
-            num_samples=args.num_samples,
-            length=args.length,
-            temperature=args.temperature,
-            top_k=args.top_k,
-            top_p=args.top_p,
-            repetition_penalty=args.repetition_penalty,
-            is_xlnet=bool(args.model_type == "xlnet"),
-            is_xlm_mlm=is_xlm_mlm,
-            xlm_mask_token=xlm_mask_token,
-            xlm_lang=xlm_lang,
-            device=args.device,
-        )
-        out = out[:, len(context_tokens):].tolist()
-        for o in out:
-            text = tokenizer.decode(o, clean_up_tokenization_spaces=True)
-            text = text[: text.find(args.stop_token) if args.stop_token else None]
+    output_sequences = model.generate(
+        input_ids=encoded_prompt,
+        max_length=args.length,
+        temperature=args.temperature,
+        top_k=args.k,
+        top_p=args.p,
+        repetition_penalty=args.repetition_penalty,
+    )
 
-            print(text)
+    # Batch size == 1. to add more examples please use num_return_sequences > 1 
+    generated_sequence = output_sequences[0].tolist()
+    text = tokenizer.decode(generated_sequence, clean_up_tokenization_spaces=True)
+    text = text[: t.find(args.stop_token) if args.stop_token else None]
+
+    print(text)
 
-        if args.prompt:
-            break
     return text
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

examples/run_glue.py

@@ -52,6 +52,9 @@ from transformers import (WEIGHTS_NAME, BertConfig,
                                   AlbertConfig,
                                   AlbertForSequenceClassification, 
                                   AlbertTokenizer,
+                                  XLMRobertaConfig,
+                                  XLMRobertaForSequenceClassification,
+                                  XLMRobertaTokenizer,
                                 )
 
 from transformers import AdamW, get_linear_schedule_with_warmup
@@ -72,7 +75,8 @@ MODEL_CLASSES = {
     'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
     'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
     'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer),
-    'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer)
+    'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer),
+    'xlmroberta': (XLMRobertaConfig, XLMRobertaForSequenceClassification, XLMRobertaTokenizer),
 }
 
 
@@ -108,6 +112,13 @@ def train(args, train_dataset, model, tokenizer):
 
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
     scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
+
     if args.fp16:
         try:
             from apex import amp
@@ -136,13 +147,33 @@ def train(args, train_dataset, model, tokenizer):
     logger.info("  Total optimization steps = %d", t_total)
 
     global_step = 0
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to gobal_step of last saved checkpoint from model path
+        global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
+        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
+
+        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
+        logger.info("  Continuing training from epoch %d", epochs_trained)
+        logger.info("  Continuing training from global step %d", global_step)
+        logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
+
     tr_loss, logging_loss = 0.0, 0.0
     model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
     set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
     for _ in train_iterator:
         epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
         for step, batch in enumerate(epoch_iterator):
+
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
             model.train()
             batch = tuple(t.to(args.device) for t in batch)
             inputs = {'input_ids':      batch[0],
@@ -201,9 +232,15 @@ def train(args, train_dataset, model, tokenizer):
                         os.makedirs(output_dir)
                     model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
                     model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+
                     torch.save(args, os.path.join(output_dir, 'training_args.bin'))
                     logger.info("Saving model checkpoint to %s", output_dir)
 
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
+                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
+
             if args.max_steps > 0 and global_step > args.max_steps:
                 epoch_iterator.close()
                 break
@@ -304,9 +341,9 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
     else:
         logger.info("Creating features from dataset file at %s", args.data_dir)
         label_list = processor.get_labels()
-        if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta']:
+        if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta', 'xlmroberta']:
             # HACK(label indices are swapped in RoBERTa pretrained model)
-            label_list[1], label_list[2] = label_list[2], label_list[1] 
+            label_list[1], label_list[2] = label_list[2], label_list[1]
         examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
         features = convert_examples_to_features(examples,
                                                 tokenizer,
@@ -380,7 +417,7 @@ def main():
     parser.add_argument("--learning_rate", default=5e-5, type=float,
                         help="The initial learning rate for Adam.")
     parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight deay if we apply some.")
+                        help="Weight decay if we apply some.")
     parser.add_argument("--adam_epsilon", default=1e-8, type=float,
                         help="Epsilon for Adam optimizer.")
     parser.add_argument("--max_grad_norm", default=1.0, type=float,

examples/run_lm_finetuning.py

@@ -150,7 +150,7 @@ def mask_tokens(inputs, tokenizer, args):
     special_tokens_mask = [tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()]
     probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
     masked_indices = torch.bernoulli(probability_matrix).bool()
-    labels[~masked_indices] = -1  # We only compute loss on masked tokens
+    labels[~masked_indices] = -100  # We only compute loss on masked tokens
 
     # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
     indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices
@@ -188,6 +188,13 @@ def train(args, train_dataset, model, tokenizer):
         ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
     scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
+
     if args.fp16:
         try:
             from apex import amp
@@ -216,17 +223,37 @@ def train(args, train_dataset, model, tokenizer):
     logger.info("  Total optimization steps = %d", t_total)
 
     global_step = 0
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to gobal_step of last saved checkpoint from model path
+        global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
+        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
+
+        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
+        logger.info("  Continuing training from epoch %d", epochs_trained)
+        logger.info("  Continuing training from global step %d", global_step)
+        logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
+
     tr_loss, logging_loss = 0.0, 0.0
 
     model_to_resize = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
     model_to_resize.resize_token_embeddings(len(tokenizer))
 
     model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
     set_seed(args)  # Added here for reproducibility (even between python 2 and 3)
     for _ in train_iterator:
         epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
         for step, batch in enumerate(epoch_iterator):
+            
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
             inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
             inputs = inputs.to(args.device)
             labels = labels.to(args.device)
@@ -274,11 +301,17 @@ def train(args, train_dataset, model, tokenizer):
                         os.makedirs(output_dir)
                     model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
                     model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+
                     torch.save(args, os.path.join(output_dir, 'training_args.bin'))
                     logger.info("Saving model checkpoint to %s", output_dir)
 
                     _rotate_checkpoints(args, checkpoint_prefix)
 
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
+                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
+
             if args.max_steps > 0 and global_step > args.max_steps:
                 epoch_iterator.close()
                 break
@@ -397,7 +430,7 @@ def main():
     parser.add_argument("--learning_rate", default=5e-5, type=float,
                         help="The initial learning rate for Adam.")
     parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight deay if we apply some.")
+                        help="Weight decay if we apply some.")
     parser.add_argument("--adam_epsilon", default=1e-8, type=float,
                         help="Epsilon for Adam optimizer.")
     parser.add_argument("--max_grad_norm", default=1.0, type=float,

examples/run_ner.py

@@ -38,11 +38,13 @@ from transformers import WEIGHTS_NAME, BertConfig, BertForTokenClassification, B
 from transformers import RobertaConfig, RobertaForTokenClassification, RobertaTokenizer
 from transformers import DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer
 from transformers import CamembertConfig, CamembertForTokenClassification, CamembertTokenizer
+from transformers import XLMRobertaConfig, XLMRobertaForTokenClassification, XLMRobertaTokenizer
 
 logger = logging.getLogger(__name__)
 
 ALL_MODELS = sum(
-    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)),
+    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig,
+                                                                   CamembertConfig, XLMRobertaConfig)),
     ())
 
 MODEL_CLASSES = {
@@ -50,6 +52,7 @@ MODEL_CLASSES = {
     "roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer),
     "distilbert": (DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer),
     "camembert": (CamembertConfig, CamembertForTokenClassification, CamembertTokenizer),
+    "xlmroberta": (XLMRobertaConfig, XLMRobertaForTokenClassification, XLMRobertaTokenizer),
 }
 
 
@@ -85,6 +88,13 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
     ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
     scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
+
     if args.fp16:
         try:
             from apex import amp
@@ -114,13 +124,33 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
     logger.info("  Total optimization steps = %d", t_total)
 
     global_step = 0
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to gobal_step of last saved checkpoint from model path
+        global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
+        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
+
+        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
+        logger.info("  Continuing training from epoch %d", epochs_trained)
+        logger.info("  Continuing training from global step %d", global_step)
+        logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
+
     tr_loss, logging_loss = 0.0, 0.0
     model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
     set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
     for _ in train_iterator:
         epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
         for step, batch in enumerate(epoch_iterator):
+
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
             model.train()
             batch = tuple(t.to(args.device) for t in batch)
             inputs = {"input_ids": batch[0],
@@ -172,9 +202,15 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
                         os.makedirs(output_dir)
                     model_to_save = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
                     model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+
                     torch.save(args, os.path.join(output_dir, "training_args.bin"))
                     logger.info("Saving model checkpoint to %s", output_dir)
 
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
+                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
+
             if args.max_steps > 0 and global_step > args.max_steps:
                 epoch_iterator.close()
                 break

examples/run_semeval.py

@@ -1,483 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" Finetuning the library models for sequence classification on Semeval (Bert)."""
-
-from __future__ import absolute_import, division, print_function
-
-import argparse
-import glob
-import logging
-import os
-import random
-
-import numpy as np
-import torch
-from tensorboardX import SummaryWriter
-from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler)
-from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm, trange
-
-from transformers import AdamW, WarmupLinearSchedule
-from transformers import (WEIGHTS_NAME, BertConfig, RobertaConfig,
-                          BertForRelationshipClassification, RobertaForRelationshipClassification,
-                          BertTokenizer, RobertaTokenizer)
-from examples.utils_semeval import (processors, output_modes, convert_examples_to_features,
-                                    convert_features_to_dataset, compute_metrics, RBertUtils,
-                                    get_entity_seperator_token_ids)
-
-logger = logging.getLogger(__name__)
-
-MODEL_CLASSES = {
-    'bert': (BertConfig, BertForRelationshipClassification, BertTokenizer),
-    'roberta': (RobertaConfig, RobertaForRelationshipClassification, RobertaTokenizer)
-}
-
-RBERT_UTILS = RBertUtils()
-
-def set_seed(args):
-    random.seed(args.seed)
-    np.random.seed(args.seed)
-    torch.manual_seed(args.seed)
-    if args.n_gpu > 0:
-        torch.cuda.manual_seed_all(args.seed)
-
-
-def train(args, train_dataset, model, tokenizer):
-    """ Train the model """
-    if args.local_rank in [-1, 0]:
-        tb_writer = SummaryWriter()
-
-    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
-    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
-    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
-
-    if args.max_steps > 0:
-        t_total = args.max_steps
-        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
-    else:
-        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
-
-    # Prepare optimizer and schedule (linear warmup and decay)
-    no_decay = ['bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
-         'weight_decay': args.weight_decay},
-        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
-    ]
-    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
-    if args.fp16:
-        try:
-            from apex import amp
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
-        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
-
-    # multi-gpu training (should be after apex fp16 initialization)
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    # Distributed training (should be after apex fp16 initialization)
-    if args.local_rank != -1:
-        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
-                                                          output_device=args.local_rank,
-                                                          find_unused_parameters=True)
-
-    # Train!
-    logger.info("***** Running training *****")
-    logger.info("  Num examples = %d", len(train_dataset))
-    logger.info("  Num Epochs = %d", args.num_train_epochs)
-    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
-    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
-                args.train_batch_size * args.gradient_accumulation_steps * (
-                    torch.distributed.get_world_size() if args.local_rank != -1 else 1))
-    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
-    logger.info("  Total optimization steps = %d", t_total)
-
-    global_step = 0
-    tr_loss, logging_loss = 0.0, 0.0
-    model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
-    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
-    for _ in train_iterator:
-        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
-        for step, batch in enumerate(epoch_iterator):
-            model.train()
-            batch = tuple(t.to(args.device) for t in batch)
-            inputs = {'input_ids': batch[0],
-                      'attention_mask': batch[1],
-                      'token_type_ids': batch[2],
-                      'labels': batch[3]}
-            outputs = model(**inputs)
-            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)
-
-            if args.n_gpu > 1:
-                loss = loss.mean()  # mean() to average on multi-gpu parallel training
-            if args.gradient_accumulation_steps > 1:
-                loss = loss / args.gradient_accumulation_steps
-
-            if args.fp16:
-                with amp.scale_loss(loss, optimizer) as scaled_loss:
-                    scaled_loss.backward()
-                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
-            else:
-                loss.backward()
-                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
-
-            tr_loss += loss.item()
-            if (step + 1) % args.gradient_accumulation_steps == 0:
-                scheduler.step()  # Update learning rate schedule
-                optimizer.step()
-                model.zero_grad()
-                global_step += 1
-
-                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
-                    # Log metrics
-                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
-                        results = evaluate(args, model, tokenizer)
-                        for key, value in results.items():
-                            tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
-                    tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
-                    tb_writer.add_scalar('loss', (tr_loss - logging_loss) / args.logging_steps, global_step)
-                    logging_loss = tr_loss
-
-                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
-                    # Save model checkpoint
-                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
-                    if not os.path.exists(output_dir):
-                        os.makedirs(output_dir)
-                    model_to_save = model.module if hasattr(model,
-                                                            'module') else model  # Take care of distributed/parallel training
-                    model_to_save.save_pretrained(output_dir)
-                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
-                    logger.info("Saving model checkpoint to %s", output_dir)
-
-            if args.max_steps > 0 and global_step > args.max_steps:
-                epoch_iterator.close()
-                break
-        if args.max_steps > 0 and global_step > args.max_steps:
-            train_iterator.close()
-            break
-
-    if args.local_rank in [-1, 0]:
-        tb_writer.close()
-
-    return global_step, tr_loss / global_step
-
-
-def evaluate(args, model, eval_dataset, prefix=""):
-    eval_task = args.task_name
-    eval_output_dir = args.output_dir
-
-    results = {}
-
-    if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
-        os.makedirs(eval_output_dir)
-
-    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
-    # Note that DistributedSampler samples randomly
-    eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
-    eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
-
-    # Eval!
-    logger.info("***** Running evaluation {} *****".format(prefix))
-    logger.info("  Num examples = %d", len(eval_dataset))
-    logger.info("  Batch size = %d", args.eval_batch_size)
-    eval_loss = 0.0
-    nb_eval_steps = 0
-    preds = None
-    out_label_ids = None
-    out_instance_ids = None
-    for batch in tqdm(eval_dataloader, desc="Evaluating"):
-        model.eval()
-        batch = tuple(t.to(args.device) for t in batch)
-
-        with torch.no_grad():
-            inputs = {'input_ids': batch[0],
-                      'attention_mask': batch[1],
-                      'token_type_ids': batch[2],
-                      'labels': batch[3]}
-            outputs = model(**inputs)
-            tmp_eval_loss, logits = outputs[:2]
-
-            eval_loss += tmp_eval_loss.mean().item()
-        nb_eval_steps += 1
-
-        if preds is None:
-            preds = logits.detach().cpu().numpy()
-            out_label_ids = inputs['labels'].detach().cpu().numpy()
-            out_instance_ids = batch[4].detach().cpu().numpy()
-        else:
-            preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
-            out_label_ids = np.append(out_label_ids, inputs['labels'].detach().cpu().numpy(), axis=0)
-            out_instance_ids = np.append(out_instance_ids, batch[4].detach().cpu().numpy(), axis=0)
-
-    eval_loss = eval_loss / nb_eval_steps
-    preds = np.argmax(preds, axis=1)
-    result = compute_metrics(eval_task, preds, out_label_ids)
-    results.update(result)
-
-    output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
-    with open(output_eval_file, "w") as writer:
-        logger.info("***** Eval results {} *****".format(prefix))
-        for key in sorted(result.keys()):
-            logger.info("  %s = %s", key, str(result[key]))
-            writer.write("%s = %s\n" % (key, str(result[key])))
-    return results, np.stack([out_instance_ids,preds],axis=1)
-
-
-def load_and_cache_examples(args, task, tokenizer, evaluate=False):
-    if args.local_rank not in [-1, 0] and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    processor = processors[task](include_directionality=args.include_directionality)
-    output_mode = output_modes[task]
-    # Load data features from cache or dataset file
-    cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}'.format(
-        'dev' if evaluate else 'train',
-        list(filter(None, args.model_name_or_path.split('/'))).pop(),
-        str(args.max_seq_length),
-        str(task)))
-    if os.path.exists(cached_features_file):
-        logger.info("Loading features from cached file %s", cached_features_file)
-        features = torch.load(cached_features_file)
-    else:
-        logger.info("Creating features from dataset file at %s", args.data_dir)
-        label_list = processor.get_labels()
-        if evaluate:
-            examples = list(processor.get_dev_examples(args.data_dir,include_other=args.eval_on_other_labels))
-            # so that the official scoring script can be used
-            processor.write_dev_examples_to_official_format(args.output_dir,examples=examples)
-        else:
-            examples = list(processor.get_train_examples(args.data_dir,include_other=args.train_on_other_labels))
-        features = convert_examples_to_features(examples, label_list,args.max_seq_length, tokenizer,RBERT_UTILS)
-        if args.local_rank in [-1, 0]:
-            logger.info("Saving features into cached file %s", cached_features_file)
-            torch.save(features, cached_features_file)
-
-    if args.local_rank == 0 and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    # Convert to Tensors and build dataset
-    dataset = convert_features_to_dataset(features, output_mode=output_mode)
-    return dataset
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    ## Required parameters
-    parser.add_argument("--data_dir", default=None, type=str, required=True,
-                        help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
-    parser.add_argument("--model_name_or_path", default='bert-base-uncased', type=str)
-    parser.add_argument("--model_type", default='bert', type=str)
-    parser.add_argument("--output_dir", default=None, type=str, required=True,
-                        help="The output directory where the model predictions and checkpoints will be written.")
-    parser.add_argument("--task_name", default='semeval2010_task8', type=str)
-    ## Other parameters
-    parser.add_argument("--config_name", default="", type=str,
-                        help="Pretrained config name or path if not the same as model_name")
-    parser.add_argument("--tokenizer_name", default="", type=str,
-                        help="Pretrained tokenizer name or path if not the same as model_name")
-    parser.add_argument("--cache_dir", default="", type=str,
-                        help="Where do you want to store the pre-trained models downloaded from s3")
-    parser.add_argument("--max_seq_length", default=128, type=int,
-                        help="The maximum total input sequence length after tokenization. Sequences longer "
-                             "than this will be truncated, sequences shorter will be padded.")
-    parser.add_argument("--do_train", action='store_true',
-                        help="Whether to run training.")
-    parser.add_argument("--do_eval", action='store_true',
-                        help="Whether to run eval on the dev set.")
-    parser.add_argument("--evaluate_during_training", action='store_true',
-                        help="Rul evaluation during training at each logging step.")
-    parser.add_argument("--do_lower_case", action='store_true',
-                        help="Set this flag if you are using an uncased model.")
-
-    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
-                        help="Batch size per GPU/CPU for training.")
-    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
-                        help="Batch size per GPU/CPU for evaluation.")
-    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
-                        help="Number of updates steps to accumulate before performing a backward/update pass.")
-    parser.add_argument("--learning_rate", default=5e-5, type=float,
-                        help="The initial learning rate for Adam.")
-    parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight deay if we apply some.")
-    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
-                        help="Epsilon for Adam optimizer.")
-    parser.add_argument("--max_grad_norm", default=1.0, type=float,
-                        help="Max gradient norm.")
-    parser.add_argument("--num_train_epochs", default=3.0, type=float,
-                        help="Total number of training epochs to perform.")
-    parser.add_argument("--max_steps", default=-1, type=int,
-                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
-    parser.add_argument("--warmup_steps", default=0, type=int,
-                        help="Linear warmup over warmup_steps.")
-
-    parser.add_argument('--logging_steps', type=int, default=50,
-                        help="Log every X updates steps.")
-    parser.add_argument('--save_steps', type=int, default=50,
-                        help="Save checkpoint every X updates steps.")
-    parser.add_argument("--eval_all_checkpoints", action='store_true',
-                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
-    parser.add_argument("--no_cuda", action='store_true',
-                        help="Avoid using CUDA when available")
-    parser.add_argument('--overwrite_output_dir', action='store_true',
-                        help="Overwrite the content of the output directory")
-    parser.add_argument('--overwrite_cache', action='store_true',
-                        help="Overwrite the cached training and evaluation sets")
-    parser.add_argument('--seed', type=int, default=42,
-                        help="random seed for initialization")
-
-    parser.add_argument('--fp16', action='store_true',
-                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
-    parser.add_argument('--fp16_opt_level', type=str, default='O1',
-                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
-                             "See details at https://nvidia.github.io/apex/amp.html")
-    parser.add_argument("--local_rank", type=int, default=-1,
-                        help="For distributed training: local_rank")
-    parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.")
-    parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
-    parser.add_argument("--include_directionality", action='store_true',help='train on all 19 classes for semeval 2010 task 8. Exclude this flag to train on the 9 relation labels only')
-    parser.add_argument("--train_on_other_labels", action='store_true',help='For semeval 2010 task 8, do not include instances with the label "Other" in the training data')
-    parser.add_argument("--eval_on_other_labels", action='store_true',help='For semeval 2010 task 8, Do not include instances with the label "Other" in the evaluation data')
-
-    args = parser.parse_args()
-
-    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
-        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
-
-    # Setup distant debugging if needed
-    if args.server_ip and args.server_port:
-        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
-        import ptvsd
-        print("Waiting for debugger attach")
-        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
-        ptvsd.wait_for_attach()
-
-    # Setup CUDA, GPU & distributed training
-    if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
-    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
-        torch.cuda.set_device(args.local_rank)
-        device = torch.device("cuda", args.local_rank)
-        torch.distributed.init_process_group(backend='nccl')
-        args.n_gpu = 1
-    args.device = device
-
-    # Setup logging
-    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
-                        datefmt = '%m/%d/%Y %H:%M:%S',
-                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
-    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
-                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
-
-    # Set seed
-    set_seed(args)
-
-    # Prepare semeval task
-    args.task_name = args.task_name.lower()
-    if args.task_name not in processors:
-        raise ValueError("Task not found: %s" % (args.task_name))
-    processor = processors[args.task_name](include_directionality=args.include_directionality)
-    args.output_mode = output_modes[args.task_name]
-    label_list = processor.get_labels()
-    logger.info(f"labels are:  {','.join(label_list)}")
-    num_labels = len(label_list)
-
-    # Load pretrained model and tokenizer
-    if args.local_rank not in [-1, 0]:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    args.model_type = args.model_type.lower()
-    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
-                                                do_lower_case=args.do_lower_case)
-
-    entity_1_token_id,entity_2_token_id = get_entity_seperator_token_ids(RBERT_UTILS, tokenizer)
-
-    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
-                                          num_labels=num_labels, finetuning_task=args.task_name,
-                                          entity_1_token_id=entity_1_token_id,
-                                          entity_2_token_id=entity_2_token_id,
-                                          )
-
-    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
-
-    if args.local_rank == 0:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    model.to(args.device)
-
-    logger.info("Training/evaluation parameters %s", args)
-
-
-    # Training
-    if args.do_train:
-        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
-        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
-        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
-
-
-    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
-    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
-        logger.info("Saving model checkpoint to %s", args.output_dir)
-        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
-        # They can then be reloaded using `from_pretrained()`
-        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
-        model_to_save.save_pretrained(args.output_dir)
-        tokenizer.save_pretrained(args.output_dir)
-
-        # Good practice: save your training arguments together with the trained model
-        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
-
-        # Load a trained model and vocabulary that you have fine-tuned
-        model = model_class.from_pretrained(args.output_dir)
-        tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path, do_lower_case=args.do_lower_case)
-        model.to(args.device)
-
-
-    # Evaluation
-    results = {}
-    if args.do_eval and args.local_rank in [-1, 0]:
-        checkpoints = [args.output_dir]
-        if args.eval_all_checkpoints:
-            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
-            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
-        logger.info("Evaluate the following checkpoints: %s", checkpoints)
-        for checkpoint in checkpoints:
-            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else "best"
-            model = model_class.from_pretrained(checkpoint)
-            model.to(args.device)
-            eval_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=True)
-            result , instance_result_array = evaluate(args, model, eval_dataset, prefix=global_step)
-            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
-            results.update(result)
-            # output file in the format for the official semeval script
-            output_semeval_file = os.path.join(args.output_dir , f"{global_step}_semeval_results.tsv")
-            with open(output_semeval_file, "w") as writer:
-                label_map = {i: label for i, label in enumerate(label_list)}
-                for x in instance_result_array:
-                    writer.write(str(x[0])+'\t'+label_map[x[1]]+'\n')
-    return results
-
-if __name__ == "__main__":
-    main()

examples/run_squad.py

@@ -16,6 +16,8 @@
 """ Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet)."""
 
 from __future__ import absolute_import, division, print_function
+from transformers.data.processors.squad import SquadV1Processor, SquadV2Processor, SquadResult
+from transformers.data.metrics.squad_metrics import compute_predictions_logits, compute_predictions_log_probs, squad_evaluate
 
 import argparse
 import logging
@@ -23,11 +25,10 @@ import os
 import random
 import glob
 import timeit
-
 import numpy as np
 import torch
-from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
-                              TensorDataset)
+from torch.utils.data import (
+    DataLoader, RandomSampler, SequentialSampler, TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
 
 try:
@@ -39,37 +40,33 @@ from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
                                   BertForQuestionAnswering, BertTokenizer,
+                                  RobertaForQuestionAnswering, RobertaTokenizer, RobertaConfig,
                                   XLMConfig, XLMForQuestionAnswering,
                                   XLMTokenizer, XLNetConfig,
                                   XLNetForQuestionAnswering,
                                   XLNetTokenizer,
                                   DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer,
-                                  AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer)
+                                  AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer,
+                                  XLMConfig, XLMForQuestionAnswering, XLMTokenizer,
+                                  )
 
-from transformers import AdamW, get_linear_schedule_with_warmup
-
-from utils_squad import (read_squad_examples, convert_examples_to_features,
-                         RawResult, write_predictions,
-                         RawResultExtended, write_predictions_extended)
-
-# The follwing import is the official SQuAD evaluation script (2.0).
-# You can remove it from the dependencies if you are using this script outside of the library
-# We've added it here for automated tests (see examples/test_examples.py file)
-from utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad
+from transformers import AdamW, get_linear_schedule_with_warmup, squad_convert_examples_to_features
 
 logger = logging.getLogger(__name__)
 
 ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
-                  for conf in (BertConfig, XLNetConfig, XLMConfig)), ())
+                  for conf in (BertConfig, RobertaConfig, XLNetConfig, XLMConfig)), ())
 
 MODEL_CLASSES = {
     'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
+    'roberta': (RobertaConfig, RobertaForQuestionAnswering, RobertaTokenizer),
     'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
     'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
     'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
-    'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer)
+    'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer),
 }
 
+
 def set_seed(args):
     random.seed(args.seed)
     np.random.seed(args.seed)
@@ -77,38 +74,60 @@ def set_seed(args):
     if args.n_gpu > 0:
         torch.cuda.manual_seed_all(args.seed)
 
+
 def to_list(tensor):
     return tensor.detach().cpu().tolist()
 
+
 def train(args, train_dataset, model, tokenizer):
     """ Train the model """
     if args.local_rank in [-1, 0]:
         tb_writer = SummaryWriter()
 
     args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
-    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
-    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+    train_sampler = RandomSampler(
+        train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(
+        train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
 
     if args.max_steps > 0:
         t_total = args.max_steps
-        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+        args.num_train_epochs = args.max_steps // (
+            len(train_dataloader) // args.gradient_accumulation_steps) + 1
     else:
-        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+        t_total = len(
+            train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
 
     # Prepare optimizer and schedule (linear warmup and decay)
     no_decay = ['bias', 'LayerNorm.weight']
     optimizer_grouped_parameters = [
-        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
-        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
-        ]
-    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+        {'params': [p for n, p in model.named_parameters() if not any(
+            nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
+        {'params': [p for n, p in model.named_parameters() if any(
+            nd in n for nd in no_decay)], 'weight_decay': 0.0}
+    ]
+    optimizer = AdamW(optimizer_grouped_parameters,
+                      lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = get_linear_schedule_with_warmup(
+        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(
+            os.path.join(args.model_name_or_path, 'optimizer.pt')))
+        scheduler.load_state_dict(torch.load(
+            os.path.join(args.model_name_or_path, 'scheduler.pt')))
+
     if args.fp16:
         try:
             from apex import amp
         except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
-        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+            raise ImportError(
+                "Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+
+        model, optimizer = amp.initialize(
+            model, optimizer, opt_level=args.fp16_opt_level)
 
     # multi-gpu training (should be after apex fp16 initialization)
     if args.n_gpu > 1:
@@ -124,36 +143,72 @@ def train(args, train_dataset, model, tokenizer):
     logger.info("***** Running training *****")
     logger.info("  Num examples = %d", len(train_dataset))
     logger.info("  Num Epochs = %d", args.num_train_epochs)
-    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info("  Instantaneous batch size per GPU = %d",
+                args.per_gpu_train_batch_size)
     logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
-                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
-    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+                args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
+    logger.info("  Gradient Accumulation steps = %d",
+                args.gradient_accumulation_steps)
     logger.info("  Total optimization steps = %d", t_total)
 
     global_step = 1
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to gobal_step of last saved checkpoint from model path
+        global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
+        epochs_trained = global_step // (len(train_dataloader) //
+                                         args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (
+            len(train_dataloader) // args.gradient_accumulation_steps)
+
+        logger.info(
+            "  Continuing training from checkpoint, will skip to saved global_step")
+        logger.info("  Continuing training from epoch %d", epochs_trained)
+        logger.info("  Continuing training from global step %d", global_step)
+        logger.info("  Will skip the first %d steps in the first epoch",
+                    steps_trained_in_current_epoch)
+
     tr_loss, logging_loss = 0.0, 0.0
     model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
-    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
+    train_iterator = trange(epochs_trained, int(
+        args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    # Added here for reproductibility (even between python 2 and 3)
+    set_seed(args)
+
     for _ in train_iterator:
-        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration",
+                              disable=args.local_rank not in [-1, 0])
         for step, batch in enumerate(epoch_iterator):
+
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
             model.train()
             batch = tuple(t.to(args.device) for t in batch)
-            inputs = {'input_ids':       batch[0],
-                      'attention_mask':  batch[1],
-                      'start_positions': batch[3],
-                      'end_positions':   batch[4]}
-            if args.model_type != 'distilbert':
-                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
+
+            inputs = {
+                'input_ids':       batch[0],
+                'attention_mask':  batch[1],
+                'token_type_ids': None if args.model_type in ['xlm', 'roberta', 'distilbert'] else batch[2],
+                'start_positions': batch[3],
+                'end_positions':   batch[4],
+            }
+
             if args.model_type in ['xlnet', 'xlm']:
                 inputs.update({'cls_index': batch[5],
                                'p_mask':       batch[6]})
+                if args.version_2_with_negative:
+                    inputs.update({'is_impossible': batch[7]})
             outputs = model(**inputs)
-            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
+            # model outputs are always tuple in transformers (see doc)
+            loss = outputs[0]
 
             if args.n_gpu > 1:
-                loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
+                loss = loss.mean()  # mean() to average on multi-gpu parallel (not distributed) training
             if args.gradient_accumulation_steps > 1:
                 loss = loss / args.gradient_accumulation_steps
 
@@ -166,35 +221,54 @@ def train(args, train_dataset, model, tokenizer):
             tr_loss += loss.item()
             if (step + 1) % args.gradient_accumulation_steps == 0:
                 if args.fp16:
-                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                    torch.nn.utils.clip_grad_norm_(
+                        amp.master_params(optimizer), args.max_grad_norm)
                 else:
-                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+                    torch.nn.utils.clip_grad_norm_(
+                        model.parameters(), args.max_grad_norm)
 
                 optimizer.step()
                 scheduler.step()  # Update learning rate schedule
                 model.zero_grad()
                 global_step += 1
 
+                # Log metrics
                 if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
-                    # Log metrics
-                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
+                    # Only evaluate when single GPU otherwise metrics may not average well
+                    if args.local_rank == -1 and args.evaluate_during_training:
                         results = evaluate(args, model, tokenizer)
                         for key, value in results.items():
-                            tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
-                    tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
-                    tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
+                            tb_writer.add_scalar(
+                                'eval_{}'.format(key), value, global_step)
+                    tb_writer.add_scalar(
+                        'lr', scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar(
+                        'loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
                     logging_loss = tr_loss
 
+                # Save model checkpoint
                 if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
-                    # Save model checkpoint
-                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
+                    output_dir = os.path.join(
+                        args.output_dir, 'checkpoint-{}'.format(global_step))
                     if not os.path.exists(output_dir):
                         os.makedirs(output_dir)
-                    model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+                    # Take care of distributed/parallel training
+                    model_to_save = model.module if hasattr(
+                        model, 'module') else model
                     model_to_save.save_pretrained(output_dir)
-                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
+                    tokenizer.save_pretrained(output_dir)
+
+                    torch.save(args, os.path.join(
+                        output_dir, 'training_args.bin'))
                     logger.info("Saving model checkpoint to %s", output_dir)
 
+                    torch.save(optimizer.state_dict(), os.path.join(
+                        output_dir, 'optimizer.pt'))
+                    torch.save(scheduler.state_dict(), os.path.join(
+                        output_dir, 'scheduler.pt'))
+                    logger.info(
+                        "Saving optimizer and scheduler states to %s", output_dir)
+
             if args.max_steps > 0 and global_step > args.max_steps:
                 epoch_iterator.close()
                 break
@@ -209,141 +283,181 @@ def train(args, train_dataset, model, tokenizer):
 
 
 def evaluate(args, model, tokenizer, prefix=""):
-    dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
+    dataset, examples, features = load_and_cache_examples(
+        args, tokenizer, evaluate=True, output_examples=True)
 
     if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
         os.makedirs(args.output_dir)
 
     args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+
     # Note that DistributedSampler samples randomly
     eval_sampler = SequentialSampler(dataset)
-    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+    eval_dataloader = DataLoader(
+        dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
 
     # multi-gpu evaluate
-    if args.n_gpu > 1:
+    if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
         model = torch.nn.DataParallel(model)
 
     # Eval!
     logger.info("***** Running evaluation {} *****".format(prefix))
     logger.info("  Num examples = %d", len(dataset))
     logger.info("  Batch size = %d", args.eval_batch_size)
+
     all_results = []
     start_time = timeit.default_timer()
+
     for batch in tqdm(eval_dataloader, desc="Evaluating"):
         model.eval()
         batch = tuple(t.to(args.device) for t in batch)
+
         with torch.no_grad():
-            inputs = {'input_ids':      batch[0],
-                      'attention_mask': batch[1]
-                      }
-            if args.model_type != 'distilbert':
-                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]  # XLM don't use segment_ids
+            inputs = {
+                'input_ids':      batch[0],
+                'attention_mask': batch[1],
+                'token_type_ids': None if args.model_type in ['xlm', 'roberta', 'distilbert'] else batch[2],
+            }
             example_indices = batch[3]
+
+            # XLNet and XLM use more arguments for their predictions
             if args.model_type in ['xlnet', 'xlm']:
-                inputs.update({'cls_index': batch[4],
-                               'p_mask':    batch[5]})
+                inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
+
             outputs = model(**inputs)
 
         for i, example_index in enumerate(example_indices):
             eval_feature = features[example_index.item()]
             unique_id = int(eval_feature.unique_id)
-            if args.model_type in ['xlnet', 'xlm']:
-                # XLNet uses a more complex post-processing procedure
-                result = RawResultExtended(unique_id            = unique_id,
-                                           start_top_log_probs  = to_list(outputs[0][i]),
-                                           start_top_index      = to_list(outputs[1][i]),
-                                           end_top_log_probs    = to_list(outputs[2][i]),
-                                           end_top_index        = to_list(outputs[3][i]),
-                                           cls_logits           = to_list(outputs[4][i]))
+
+            output = [to_list(output[i]) for output in outputs]
+
+            # Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
+            # models only use two.
+            if len(output) >= 5:
+                start_logits = output[0]
+                start_top_index = output[1]
+                end_logits = output[2]
+                end_top_index = output[3]
+                cls_logits = output[4]
+
+                result = SquadResult(
+                    unique_id, start_logits, end_logits,
+                    start_top_index=start_top_index,
+                    end_top_index=end_top_index,
+                    cls_logits=cls_logits
+                )
+
             else:
-                result = RawResult(unique_id    = unique_id,
-                                   start_logits = to_list(outputs[0][i]),
-                                   end_logits   = to_list(outputs[1][i]))
+                start_logits, end_logits = output
+                result = SquadResult(
+                    unique_id, start_logits, end_logits
+                )
+
             all_results.append(result)
 
     evalTime = timeit.default_timer() - start_time
-    logger.info("  Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))
+    logger.info("  Evaluation done in total %f secs (%f sec per example)",
+                evalTime, evalTime / len(dataset))
 
     # Compute predictions
-    output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
-    output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
+    output_prediction_file = os.path.join(
+        args.output_dir, "predictions_{}.json".format(prefix))
+    output_nbest_file = os.path.join(
+        args.output_dir, "nbest_predictions_{}.json".format(prefix))
+
     if args.version_2_with_negative:
-        output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
+        output_null_log_odds_file = os.path.join(
+            args.output_dir, "null_odds_{}.json".format(prefix))
     else:
         output_null_log_odds_file = None
 
+    # XLNet and XLM use a more complex post-processing procedure
     if args.model_type in ['xlnet', 'xlm']:
-        # XLNet uses a more complex post-processing procedure
-        write_predictions_extended(examples, features, all_results, args.n_best_size,
-                        args.max_answer_length, output_prediction_file,
-                        output_nbest_file, output_null_log_odds_file, args.predict_file,
-                        model.config.start_n_top, model.config.end_n_top,
-                        args.version_2_with_negative, tokenizer, args.verbose_logging)
+        start_n_top = model.config.start_n_top if hasattr(
+            model, "config") else model.module.config.start_n_top
+        end_n_top = model.config.end_n_top if hasattr(
+            model, "config") else model.module.config.end_n_top
+
+        predictions = compute_predictions_log_probs(examples, features, all_results, args.n_best_size,
+                                                    args.max_answer_length, output_prediction_file,
+                                                    output_nbest_file, output_null_log_odds_file,
+                                                    start_n_top, end_n_top,
+                                                    args.version_2_with_negative, tokenizer, args.verbose_logging)
     else:
-        write_predictions(examples, features, all_results, args.n_best_size,
+        predictions = compute_predictions_logits(examples, features, all_results, args.n_best_size,
                         args.max_answer_length, args.do_lower_case, output_prediction_file,
                         output_nbest_file, output_null_log_odds_file, args.verbose_logging,
-                        args.version_2_with_negative, args.null_score_diff_threshold)
+                        args.version_2_with_negative, args.null_score_diff_threshold, tokenizer)
 
-    # Evaluate with the official SQuAD script
-    evaluate_options = EVAL_OPTS(data_file=args.predict_file,
-                                 pred_file=output_prediction_file,
-                                 na_prob_file=output_null_log_odds_file)
-    results = evaluate_on_squad(evaluate_options)
+    # Compute the F1 and exact scores.
+    results = squad_evaluate(examples, predictions)
     return results
 
 
 def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
     if args.local_rank not in [-1, 0] and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        torch.distributed.barrier()
 
     # Load data features from cache or dataset file
-    input_file = args.predict_file if evaluate else args.train_file
-    cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
+    input_dir = args.data_dir if args.data_dir else "."
+    cached_features_file = os.path.join(input_dir, 'cached_{}_{}_{}'.format(
         'dev' if evaluate else 'train',
         list(filter(None, args.model_name_or_path.split('/'))).pop(),
-        str(args.max_seq_length)))
+        str(args.max_seq_length))
+    )
+
+    # Init features and dataset from cache if it exists
     if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
-        logger.info("Loading features from cached file %s", cached_features_file)
-        features = torch.load(cached_features_file)
+        logger.info("Loading features from cached file %s",
+                    cached_features_file)
+        features_and_dataset = torch.load(cached_features_file)
+        features, dataset = features_and_dataset["features"], features_and_dataset["dataset"]
     else:
-        logger.info("Creating features from dataset file at %s", input_file)
-        examples = read_squad_examples(input_file=input_file,
-                                                is_training=not evaluate,
-                                                version_2_with_negative=args.version_2_with_negative)
-        features = convert_examples_to_features(examples=examples,
-                                                tokenizer=tokenizer,
-                                                max_seq_length=args.max_seq_length,
-                                                doc_stride=args.doc_stride,
-                                                max_query_length=args.max_query_length,
-                                                is_training=not evaluate,
-                                                cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
-                                                pad_token_segment_id=3 if args.model_type in ['xlnet'] else 0,
-                                                cls_token_at_end=True if args.model_type in ['xlnet'] else False,
-                                                sequence_a_is_doc=True if args.model_type in ['xlnet'] else False)
+        logger.info("Creating features from dataset file at %s", input_dir)
+
+        if not args.data_dir and ((evaluate and not args.predict_file) or (not evaluate and not args.train_file)):
+            try:
+                import tensorflow_datasets as tfds
+            except ImportError:
+                raise ImportError(
+                    "If not data_dir is specified, tensorflow_datasets needs to be installed.")
+
+            if args.version_2_with_negative:
+                logger.warn(
+                    "tensorflow_datasets does not handle version 2 of SQuAD.")
+
+            tfds_examples = tfds.load("squad")
+            examples = SquadV1Processor().get_examples_from_dataset(
+                tfds_examples, evaluate=evaluate)
+        else:
+            processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
+            if evaluate:
+                examples = processor.get_dev_examples(args.data_dir, filename=args.predict_file)
+            else:
+                examples = processor.get_train_examples(args.data_dir, filename=args.train_file)
+
+        features, dataset = squad_convert_examples_to_features(
+            examples=examples,
+            tokenizer=tokenizer,
+            max_seq_length=args.max_seq_length,
+            doc_stride=args.doc_stride,
+            max_query_length=args.max_query_length,
+            is_training=not evaluate,
+            return_dataset='pt',
+            threads=args.threads,
+        )
+
         if args.local_rank in [-1, 0]:
-            logger.info("Saving features into cached file %s", cached_features_file)
-            torch.save(features, cached_features_file)
+            logger.info("Saving features into cached file %s",
+                        cached_features_file)
+            torch.save({"features": features, "dataset": dataset},
+                       cached_features_file)
 
     if args.local_rank == 0 and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    # Convert to Tensors and build dataset
-    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
-    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
-    all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
-    all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
-    all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
-    if evaluate:
-        all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
-        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
-                                all_example_index, all_cls_index, all_p_mask)
-    else:
-        all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
-        all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
-        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
-                                all_start_positions, all_end_positions,
-                                all_cls_index, all_p_mask)
+        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        torch.distributed.barrier()
 
     if output_examples:
         return dataset, examples, features
@@ -353,11 +467,7 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
 def main():
     parser = argparse.ArgumentParser()
 
-    ## Required parameters
-    parser.add_argument("--train_file", default=None, type=str, required=True,
-                        help="SQuAD json for training. E.g., train-v1.1.json")
-    parser.add_argument("--predict_file", default=None, type=str, required=True,
-                        help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
+    # Required parameters
     parser.add_argument("--model_type", default=None, type=str, required=True,
                         help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
     parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
@@ -365,7 +475,16 @@ def main():
     parser.add_argument("--output_dir", default=None, type=str, required=True,
                         help="The output directory where the model checkpoints and predictions will be written.")
 
-    ## Other parameters
+    # Other parameters
+    parser.add_argument("--data_dir", default=None, type=str,
+                        help="The input data dir. Should contain the .json files for the task." +
+                             "If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
+    parser.add_argument("--train_file", default=None, type=str,
+                        help="The input training file. If a data dir is specified, will look for the file there" +
+                             "If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
+    parser.add_argument("--predict_file", default=None, type=str,
+                        help="The input evaluation file. If a data dir is specified, will look for the file there" +
+                             "If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
     parser.add_argument("--config_name", default="", type=str,
                         help="Pretrained config name or path if not the same as model_name")
     parser.add_argument("--tokenizer_name", default="", type=str,
@@ -448,22 +567,27 @@ def main():
                              "See details at https://nvidia.github.io/apex/amp.html")
     parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
     parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
+
+    parser.add_argument('--threads', type=int, default=1, help='multiple threads for converting example to features')
     args = parser.parse_args()
 
     if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
-        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
 
     # Setup distant debugging if needed
     if args.server_ip and args.server_port:
         # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
         import ptvsd
         print("Waiting for debugger attach")
-        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.enable_attach(
+            address=(args.server_ip, args.server_port), redirect_output=True)
         ptvsd.wait_for_attach()
 
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        device = torch.device(
+            "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
         args.n_gpu = torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
@@ -473,18 +597,19 @@ def main():
     args.device = device
 
     # Setup logging
-    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
-                        datefmt = '%m/%d/%Y %H:%M:%S',
-                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
+    logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+                        datefmt='%m/%d/%Y %H:%M:%S',
+                        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
     logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
-                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
+                   args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
 
     # Set seed
     set_seed(args)
 
     # Load pretrained model and tokenizer
     if args.local_rank not in [-1, 0]:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+        # Make sure only the first process in distributed training will download model & vocab
+        torch.distributed.barrier()
 
     args.model_type = args.model_type.lower()
     config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
@@ -494,12 +619,14 @@ def main():
                                                 do_lower_case=args.do_lower_case,
                                                 cache_dir=args.cache_dir if args.cache_dir else None)
     model = model_class.from_pretrained(args.model_name_or_path,
-                                        from_tf=bool('.ckpt' in args.model_name_or_path),
+                                        from_tf=bool(
+                                            '.ckpt' in args.model_name_or_path),
                                         config=config,
                                         cache_dir=args.cache_dir if args.cache_dir else None)
 
     if args.local_rank == 0:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+        # Make sure only the first process in distributed training will download model & vocab
+        torch.distributed.barrier()
 
     model.to(args.device)
 
@@ -513,14 +640,16 @@ def main():
             import apex
             apex.amp.register_half_function(torch, 'einsum')
         except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+            raise ImportError(
+                "Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
 
     # Training
     if args.do_train:
-        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
+        train_dataset = load_and_cache_examples(
+            args, tokenizer, evaluate=False, output_examples=False)
         global_step, tr_loss = train(args, train_dataset, model, tokenizer)
-        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
-
+        logger.info(" global_step = %s, average loss = %s",
+                    global_step, tr_loss)
 
     # Save the trained model and the tokenizer
     if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
@@ -531,7 +660,8 @@ def main():
         logger.info("Saving model checkpoint to %s", args.output_dir)
         # Save a trained model, configuration and tokenizer using `save_pretrained()`.
         # They can then be reloaded using `from_pretrained()`
-        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+        # Take care of distributed/parallel training
+        model_to_save = model.module if hasattr(model, 'module') else model
         model_to_save.save_pretrained(args.output_dir)
         tokenizer.save_pretrained(args.output_dir)
 
@@ -539,31 +669,40 @@ def main():
         torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
 
         # Load a trained model and vocabulary that you have fine-tuned
-        model = model_class.from_pretrained(args.output_dir, force_download=True)
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model = model_class.from_pretrained(
+            args.output_dir, force_download=True)
+        tokenizer = tokenizer_class.from_pretrained(
+            args.output_dir, do_lower_case=args.do_lower_case)
         model.to(args.device)
 
-
     # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
     results = {}
     if args.do_eval and args.local_rank in [-1, 0]:
-        checkpoints = [args.output_dir]
-        if args.eval_all_checkpoints:
-            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
-            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs
+        if args.do_train:
+            logger.info("Loading checkpoints saved during training for evaluation")
+            checkpoints = [args.output_dir]
+            if args.eval_all_checkpoints:
+                checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
+                logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs
+        else:
+            logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
+            checkpoints = [args.model_name_or_path]
 
         logger.info("Evaluate the following checkpoints: %s", checkpoints)
 
         for checkpoint in checkpoints:
             # Reload the model
-            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
-            model = model_class.from_pretrained(checkpoint, force_download=True)
+            global_step = checkpoint.split(
+                '-')[-1] if len(checkpoints) > 1 else ""
+            model = model_class.from_pretrained(
+                checkpoint, force_download=True)
             model.to(args.device)
 
             # Evaluate
             result = evaluate(args, model, tokenizer, prefix=global_step)
 
-            result = dict((k + ('_{}'.format(global_step) if global_step else ''), v) for k, v in result.items())
+            result = dict((k + ('_{}'.format(global_step) if global_step else ''), v)
+                          for k, v in result.items())
             results.update(result)
 
     logger.info("Results: {}".format(results))

examples/run_summarization_finetuning.py

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

examples/run_tf_ner.py

@@ -0,0 +1,615 @@
+# coding=utf-8
+import datetime
+import os
+import math
+import glob
+import re
+import tensorflow as tf
+import collections
+import numpy as np
+from seqeval import metrics
+import _pickle as pickle
+from absl import logging
+from transformers import TF2_WEIGHTS_NAME, BertConfig, BertTokenizer, TFBertForTokenClassification
+from transformers import RobertaConfig, RobertaTokenizer, TFRobertaForTokenClassification
+from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForTokenClassification
+from transformers import create_optimizer, GradientAccumulator
+from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file
+from fastprogress import master_bar, progress_bar
+from absl import flags
+from absl import app
+
+
+ALL_MODELS = sum(
+    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)),
+    ())
+
+MODEL_CLASSES = {
+    "bert": (BertConfig, TFBertForTokenClassification, BertTokenizer),
+    "roberta": (RobertaConfig, TFRobertaForTokenClassification, RobertaTokenizer),
+    "distilbert": (DistilBertConfig, TFDistilBertForTokenClassification, DistilBertTokenizer)
+}
+
+
+flags.DEFINE_string(
+    "data_dir", None,
+    "The input data dir. Should contain the .conll files (or other data files) "
+    "for the task.")
+
+flags.DEFINE_string(
+    "model_type", None,
+    "Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+
+flags.DEFINE_string(
+    "model_name_or_path", None,
+    "Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+
+flags.DEFINE_string(
+    "output_dir", None,
+    "The output directory where the model checkpoints will be written.")
+
+flags.DEFINE_string(
+    "labels", "",
+    "Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.")
+
+flags.DEFINE_string(
+    "config_name", "",
+    "Pretrained config name or path if not the same as model_name")
+
+flags.DEFINE_string(
+    "tokenizer_name", "",
+    "Pretrained tokenizer name or path if not the same as model_name")
+
+flags.DEFINE_string(
+    "cache_dir", "",
+    "Where do you want to store the pre-trained models downloaded from s3")
+
+flags.DEFINE_integer(
+    "max_seq_length", 128,
+    "The maximum total input sentence length after tokenization. "
+    "Sequences longer than this will be truncated, sequences shorter "
+    "will be padded.")
+
+flags.DEFINE_string(
+    "tpu", None,
+    "The Cloud TPU to use for training. This should be either the name "
+    "used when creating the Cloud TPU, or a grpc://ip.address.of.tpu:8470 "
+    "url.")
+
+flags.DEFINE_integer(
+    "num_tpu_cores", 8,
+    "Total number of TPU cores to use.")
+
+flags.DEFINE_boolean(
+    "do_train", False,
+    "Whether to run training.")
+
+flags.DEFINE_boolean(
+    "do_eval", False,
+    "Whether to run eval on the dev set.")
+
+flags.DEFINE_boolean(
+    "do_predict", False,
+    "Whether to run predictions on the test set.")
+
+flags.DEFINE_boolean(
+    "evaluate_during_training", False,
+    "Whether to run evaluation during training at each logging step.")
+
+flags.DEFINE_boolean(
+    "do_lower_case", False,
+    "Set this flag if you are using an uncased model.")
+
+flags.DEFINE_integer(
+    "per_device_train_batch_size", 8,
+    "Batch size per GPU/CPU/TPU for training.")
+
+flags.DEFINE_integer(
+    "per_device_eval_batch_size", 8,
+    "Batch size per GPU/CPU/TPU for evaluation.")
+
+flags.DEFINE_integer(
+    "gradient_accumulation_steps", 1,
+    "Number of updates steps to accumulate before performing a backward/update pass.")
+
+flags.DEFINE_float(
+    "learning_rate", 5e-5,
+    "The initial learning rate for Adam.")
+
+flags.DEFINE_float(
+    "weight_decay", 0.0,
+    "Weight decay if we apply some.")
+
+flags.DEFINE_float(
+    "adam_epsilon", 1e-8,
+    "Epsilon for Adam optimizer.")
+
+flags.DEFINE_float(
+    "max_grad_norm", 1.0,
+    "Max gradient norm.")
+
+flags.DEFINE_integer(
+    "num_train_epochs", 3,
+    "Total number of training epochs to perform.")
+
+flags.DEFINE_integer(
+    "max_steps", -1,
+    "If > 0: set total number of training steps to perform. Override num_train_epochs.")
+
+flags.DEFINE_integer(
+    "warmup_steps", 0,
+    "Linear warmup over warmup_steps.")
+
+flags.DEFINE_integer(
+    "logging_steps", 50,
+    "Log every X updates steps.")
+
+flags.DEFINE_integer(
+    "save_steps", 50,
+    "Save checkpoint every X updates steps.")
+
+flags.DEFINE_boolean(
+    "eval_all_checkpoints", False,
+    "Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
+
+flags.DEFINE_boolean(
+    "no_cuda", False,
+    "Avoid using CUDA when available")
+
+flags.DEFINE_boolean(
+    "overwrite_output_dir", False,
+    "Overwrite the content of the output directory")
+
+flags.DEFINE_boolean(
+    "overwrite_cache", False,
+    "Overwrite the cached training and evaluation sets")
+
+flags.DEFINE_integer(
+    "seed", 42,
+    "random seed for initialization")
+
+flags.DEFINE_boolean(
+    "fp16", False,
+    "Whether to use 16-bit (mixed) precision instead of 32-bit")
+
+flags.DEFINE_string(
+    "gpus", "0",
+    "Comma separated list of gpus devices. If only one, switch to single "
+    "gpu strategy, if None takes all the gpus available.")
+
+
+def train(args, strategy, train_dataset, tokenizer, model, num_train_examples, labels, train_batch_size, pad_token_label_id):
+    if args['max_steps'] > 0:
+        num_train_steps = args['max_steps'] * args['gradient_accumulation_steps']
+        args['num_train_epochs'] = 1
+    else:
+        num_train_steps = math.ceil(num_train_examples / train_batch_size) // args['gradient_accumulation_steps'] * args['num_train_epochs']
+
+    writer = tf.summary.create_file_writer("/tmp/mylogs")
+
+    with strategy.scope():
+        loss_fct = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
+        optimizer = create_optimizer(args['learning_rate'], num_train_steps, args['warmup_steps'])
+
+        if args['fp16']:
+            optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(optimizer, 'dynamic')
+
+        loss_metric = tf.keras.metrics.Mean(name='loss', dtype=tf.float32)
+        gradient_accumulator = GradientAccumulator()
+        
+    logging.info("***** Running training *****")
+    logging.info("  Num examples = %d", num_train_examples)
+    logging.info("  Num Epochs = %d", args['num_train_epochs'])
+    logging.info("  Instantaneous batch size per device = %d", args['per_device_train_batch_size'])
+    logging.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
+                train_batch_size * args['gradient_accumulation_steps'])
+    logging.info("  Gradient Accumulation steps = %d", args['gradient_accumulation_steps'])
+    logging.info("  Total training steps = %d", num_train_steps)
+
+    model.summary()
+
+    @tf.function
+    def apply_gradients():
+        grads_and_vars = []
+
+        for gradient, variable in zip(gradient_accumulator.gradients, model.trainable_variables):
+            if gradient is not None:
+                scaled_gradient = gradient / (args['n_device'] * args['gradient_accumulation_steps'])
+                grads_and_vars.append((scaled_gradient, variable))
+            else:
+                grads_and_vars.append((gradient, variable))
+
+        optimizer.apply_gradients(grads_and_vars, args['max_grad_norm'])
+        gradient_accumulator.reset()
+
+    @tf.function
+    def train_step(train_features, train_labels):
+        def step_fn(train_features, train_labels):
+            inputs = {'attention_mask': train_features['input_mask'], 'training': True}
+
+            if args['model_type'] != "distilbert":
+                inputs["token_type_ids"] = train_features['segment_ids'] if args['model_type'] in ["bert", "xlnet"] else None
+
+            with tf.GradientTape() as tape:
+                logits = model(train_features['input_ids'], **inputs)[0]
+                logits = tf.reshape(logits, (-1, len(labels) + 1))
+                active_loss = tf.reshape(train_features['input_mask'], (-1,))
+                active_logits = tf.boolean_mask(logits, active_loss)
+                train_labels = tf.reshape(train_labels, (-1,))
+                active_labels = tf.boolean_mask(train_labels, active_loss)
+                cross_entropy = loss_fct(active_labels, active_logits)
+                loss = tf.reduce_sum(cross_entropy) * (1.0 / train_batch_size)
+                grads = tape.gradient(loss, model.trainable_variables)
+
+                gradient_accumulator(grads)
+
+            return cross_entropy
+
+        per_example_losses = strategy.experimental_run_v2(step_fn, args=(train_features, train_labels))
+        mean_loss = strategy.reduce(tf.distribute.ReduceOp.MEAN, per_example_losses, axis=0)
+
+        return mean_loss
+
+    current_time = datetime.datetime.now()
+    train_iterator = master_bar(range(args['num_train_epochs']))
+    global_step = 0
+    logging_loss = 0.0
+
+    for epoch in train_iterator:
+        epoch_iterator = progress_bar(train_dataset, total=num_train_steps, parent=train_iterator, display=args['n_device'] > 1)
+        step = 1
+
+        with strategy.scope():
+            for train_features, train_labels in epoch_iterator:
+                loss = train_step(train_features, train_labels)
+
+                if step % args['gradient_accumulation_steps'] == 0:
+                    strategy.experimental_run_v2(apply_gradients)
+
+                    loss_metric(loss)
+
+                    global_step += 1
+
+                    if args['logging_steps'] > 0 and global_step % args['logging_steps'] == 0:
+                        # Log metrics
+                        if args['n_device'] == 1 and args['evaluate_during_training']:  # Only evaluate when single GPU otherwise metrics may not average well
+                            y_true, y_pred, eval_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="dev")
+                            report = metrics.classification_report(y_true, y_pred, digits=4)
+                            
+                            logging.info("Eval at step " + str(global_step) + "\n" + report)
+                            logging.info("eval_loss: " + str(eval_loss))
+                            
+                            precision = metrics.precision_score(y_true, y_pred)
+                            recall = metrics.recall_score(y_true, y_pred)
+                            f1 = metrics.f1_score(y_true, y_pred)
+
+                            with writer.as_default():
+                                tf.summary.scalar("eval_loss", eval_loss, global_step)
+                                tf.summary.scalar("precision", precision, global_step)
+                                tf.summary.scalar("recall", recall, global_step)
+                                tf.summary.scalar("f1", f1, global_step)
+                        
+                        lr = optimizer.learning_rate
+                        learning_rate = lr(step)
+
+                        with writer.as_default():
+                            tf.summary.scalar("lr", learning_rate, global_step)
+                            tf.summary.scalar("loss", (loss_metric.result() - logging_loss) / args['logging_steps'], global_step)
+                        
+                        logging_loss = loss_metric.result()
+
+                    with writer.as_default():
+                        tf.summary.scalar("loss", loss_metric.result(), step=step)
+
+                    if args['save_steps'] > 0 and global_step % args['save_steps'] == 0:
+                        # Save model checkpoint
+                        output_dir = os.path.join(args['output_dir'], "checkpoint-{}".format(global_step))
+
+                        if not os.path.exists(output_dir):
+                            os.makedirs(output_dir)
+                        
+                        model.save_pretrained(output_dir)
+                        logging.info("Saving model checkpoint to %s", output_dir)
+                
+                train_iterator.child.comment = f'loss : {loss_metric.result()}'
+                step += 1
+
+        train_iterator.write(f'loss epoch {epoch + 1}: {loss_metric.result()}')
+
+        loss_metric.reset_states()
+
+    logging.info("  Training took time = {}".format(datetime.datetime.now() - current_time))
+
+
+def evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode):
+    eval_batch_size = args['per_device_eval_batch_size'] * args['n_device']
+    eval_dataset, size = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, eval_batch_size, mode=mode)
+    eval_dataset = strategy.experimental_distribute_dataset(eval_dataset)
+    preds = None
+    num_eval_steps = math.ceil(size / eval_batch_size)
+    master = master_bar(range(1))
+    eval_iterator = progress_bar(eval_dataset, total=num_eval_steps, parent=master, display=args['n_device'] > 1)
+    loss_fct = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
+    loss = 0.0
+
+    logging.info("***** Running evaluation *****")
+    logging.info("  Num examples = %d", size)
+    logging.info("  Batch size = %d", eval_batch_size)
+
+    for eval_features, eval_labels in eval_iterator:
+        inputs = {'attention_mask': eval_features['input_mask'], 'training': False}
+
+        if args['model_type'] != "distilbert":
+            inputs["token_type_ids"] = eval_features['segment_ids'] if args['model_type'] in ["bert", "xlnet"] else None
+
+        with strategy.scope():
+            logits = model(eval_features['input_ids'], **inputs)[0]
+            tmp_logits = tf.reshape(logits, (-1, len(labels) + 1))
+            active_loss = tf.reshape(eval_features['input_mask'], (-1,))
+            active_logits = tf.boolean_mask(tmp_logits, active_loss)
+            tmp_eval_labels = tf.reshape(eval_labels, (-1,))
+            active_labels = tf.boolean_mask(tmp_eval_labels, active_loss)
+            cross_entropy = loss_fct(active_labels, active_logits)
+            loss += tf.reduce_sum(cross_entropy) * (1.0 / eval_batch_size)
+
+        if preds is None:
+            preds = logits.numpy()
+            label_ids = eval_labels.numpy()
+        else:
+            preds = np.append(preds, logits.numpy(), axis=0)
+            label_ids = np.append(label_ids, eval_labels.numpy(), axis=0)
+
+    preds = np.argmax(preds, axis=2)
+    y_pred = [[] for _ in range(label_ids.shape[0])]
+    y_true = [[] for _ in range(label_ids.shape[0])]
+    loss = loss / num_eval_steps
+
+    for i in range(label_ids.shape[0]):
+        for j in range(label_ids.shape[1]):
+            if label_ids[i, j] != pad_token_label_id:
+                y_pred[i].append(labels[preds[i, j] - 1])
+                y_true[i].append(labels[label_ids[i, j] - 1])
+
+    return y_true, y_pred, loss.numpy()
+
+
+def load_cache(cached_file, max_seq_length):
+    name_to_features = {
+        "input_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
+        "input_mask": tf.io.FixedLenFeature([max_seq_length], tf.int64),
+        "segment_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
+        "label_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
+    }
+
+    def _decode_record(record):
+        example = tf.io.parse_single_example(record, name_to_features)
+        features = {}
+        features['input_ids'] = example['input_ids']
+        features['input_mask'] = example['input_mask']
+        features['segment_ids'] = example['segment_ids']
+
+        return features, example['label_ids']
+
+    d = tf.data.TFRecordDataset(cached_file)
+    d = d.map(_decode_record, num_parallel_calls=4)
+    count = d.reduce(0, lambda x, _: x + 1)
+
+    return d, count.numpy()
+
+
+def save_cache(features, cached_features_file):
+    writer = tf.io.TFRecordWriter(cached_features_file)
+
+    for (ex_index, feature) in enumerate(features):
+        if ex_index % 5000 == 0:
+            logging.info("Writing example %d of %d" % (ex_index, len(features)))
+
+        def create_int_feature(values):
+            f = tf.train.Feature(int64_list=tf.train.Int64List(value=list(values)))
+            return f
+
+        record_feature = collections.OrderedDict()
+        record_feature["input_ids"] = create_int_feature(feature.input_ids)
+        record_feature["input_mask"] = create_int_feature(feature.input_mask)
+        record_feature["segment_ids"] = create_int_feature(feature.segment_ids)
+        record_feature["label_ids"] = create_int_feature(feature.label_ids)
+
+        tf_example = tf.train.Example(features=tf.train.Features(feature=record_feature))
+
+        writer.write(tf_example.SerializeToString())
+
+    writer.close()
+
+
+def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, batch_size, mode):
+    drop_remainder = True if args['tpu'] or mode == 'train' else False
+
+    # Load data features from cache or dataset file
+    cached_features_file = os.path.join(args['data_dir'], "cached_{}_{}_{}.tf_record".format(mode,
+        list(filter(None, args['model_name_or_path'].split("/"))).pop(),
+        str(args['max_seq_length'])))
+    if os.path.exists(cached_features_file) and not args['overwrite_cache']:
+        logging.info("Loading features from cached file %s", cached_features_file)
+        dataset, size = load_cache(cached_features_file, args['max_seq_length'])
+    else:
+        logging.info("Creating features from dataset file at %s", args['data_dir'])
+        examples = read_examples_from_file(args['data_dir'], mode)
+        features = convert_examples_to_features(examples, labels, args['max_seq_length'], tokenizer,
+                                                cls_token_at_end=bool(args['model_type'] in ["xlnet"]),
+                                                # xlnet has a cls token at the end
+                                                cls_token=tokenizer.cls_token,
+                                                cls_token_segment_id=2 if args['model_type'] in ["xlnet"] else 0,
+                                                sep_token=tokenizer.sep_token,
+                                                sep_token_extra=bool(args['model_type'] in ["roberta"]),
+                                                # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
+                                                pad_on_left=bool(args['model_type'] in ["xlnet"]),
+                                                # pad on the left for xlnet
+                                                pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
+                                                pad_token_segment_id=4 if args['model_type'] in ["xlnet"] else 0,
+                                                pad_token_label_id=pad_token_label_id
+                                                )
+        logging.info("Saving features into cached file %s", cached_features_file)
+        save_cache(features, cached_features_file)
+        dataset, size = load_cache(cached_features_file, args['max_seq_length'])
+
+    if mode == 'train':
+        dataset = dataset.repeat()
+        dataset = dataset.shuffle(buffer_size=8192, seed=args['seed'])
+
+    dataset = dataset.batch(batch_size, drop_remainder)
+    dataset = dataset.prefetch(buffer_size=batch_size)
+
+    return dataset, size
+
+
+def main(_):
+    logging.set_verbosity(logging.INFO)
+    args = flags.FLAGS.flag_values_dict()
+
+    if os.path.exists(args['output_dir']) and os.listdir(
+            args['output_dir']) and args['do_train'] and not args['overwrite_output_dir']:
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
+                args['output_dir']))
+
+    if args['fp16']:
+        tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})
+
+    if args['tpu']:
+        resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu=args['tpu'])
+        tf.config.experimental_connect_to_cluster(resolver)
+        tf.tpu.experimental.initialize_tpu_system(resolver)
+        strategy = tf.distribute.experimental.TPUStrategy(resolver)
+        args['n_device'] = args['num_tpu_cores']
+    elif len(args['gpus'].split(',')) > 1:
+        args['n_device'] = len([f"/gpu:{gpu}" for gpu in args['gpus'].split(',')])
+        strategy = tf.distribute.MirroredStrategy(devices=[f"/gpu:{gpu}" for gpu in args['gpus'].split(',')])
+    elif args['no_cuda']:
+        args['n_device'] = 1
+        strategy = tf.distribute.OneDeviceStrategy(device="/cpu:0")
+    else:
+        args['n_device'] = len(args['gpus'].split(','))
+        strategy = tf.distribute.OneDeviceStrategy(device="/gpu:" + args['gpus'].split(',')[0])
+
+    logging.warning("n_device: %s, distributed training: %s, 16-bits training: %s",
+                   args['n_device'], bool(args['n_device'] > 1), args['fp16'])
+
+    labels = get_labels(args['labels'])
+    num_labels = len(labels) + 1
+    pad_token_label_id = 0
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args['model_type']]
+    config = config_class.from_pretrained(args['config_name'] if args['config_name'] else args['model_name_or_path'],
+                                          num_labels=num_labels,
+                                          cache_dir=args['cache_dir'] if args['cache_dir'] else None)
+
+    logging.info("Training/evaluation parameters %s", args)
+
+    # Training
+    if args['do_train']:
+        tokenizer = tokenizer_class.from_pretrained(args['tokenizer_name'] if args['tokenizer_name'] else args['model_name_or_path'],
+                                                    do_lower_case=args['do_lower_case'],
+                                                    cache_dir=args['cache_dir'] if args['cache_dir'] else None)
+
+        with strategy.scope():
+            model = model_class.from_pretrained(args['model_name_or_path'],
+                                                from_pt=bool(".bin" in args['model_name_or_path']),
+                                                config=config,
+                                                cache_dir=args['cache_dir'] if args['cache_dir'] else None)
+            model.layers[-1].activation = tf.keras.activations.softmax
+
+        train_batch_size = args['per_device_train_batch_size'] * args['n_device']
+        train_dataset, num_train_examples = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, train_batch_size, mode="train")
+        train_dataset = strategy.experimental_distribute_dataset(train_dataset)
+        train(args, strategy, train_dataset, tokenizer, model, num_train_examples, labels, train_batch_size, pad_token_label_id)
+
+        if not os.path.exists(args['output_dir']):
+            os.makedirs(args['output_dir'])
+
+        logging.info("Saving model to %s", args['output_dir'])
+
+        model.save_pretrained(args['output_dir'])
+        tokenizer.save_pretrained(args['output_dir'])
+
+    # Evaluation
+    if args['do_eval']:
+        tokenizer = tokenizer_class.from_pretrained(args['output_dir'], do_lower_case=args['do_lower_case'])
+        checkpoints = []
+        results = []
+
+        if args['eval_all_checkpoints']:
+            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args['output_dir'] + "/**/" + TF2_WEIGHTS_NAME, recursive=True), key=lambda f: int(''.join(filter(str.isdigit, f)) or -1)))
+        
+        logging.info("Evaluate the following checkpoints: %s", checkpoints)
+
+        if len(checkpoints) == 0:
+            checkpoints.append(args['output_dir'])
+        
+        for checkpoint in checkpoints:
+            global_step = checkpoint.split("-")[-1] if re.match(".*checkpoint-[0-9]", checkpoint) else "final"
+
+            with strategy.scope():
+                model = model_class.from_pretrained(checkpoint)
+
+            y_true, y_pred, eval_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="dev")
+            report = metrics.classification_report(y_true, y_pred, digits=4)
+
+            if global_step:
+                results.append({global_step + "_report": report, global_step + "_loss": eval_loss})
+
+        output_eval_file = os.path.join(args['output_dir'], "eval_results.txt")
+        
+        with tf.io.gfile.GFile(output_eval_file, "w") as writer:
+            for res in results:
+                for key, val in res.items():
+                    if "loss" in key:
+                        logging.info(key + " = " + str(val))
+                        writer.write(key + " = " + str(val))
+                        writer.write("\n")
+                    else:
+                        logging.info(key)
+                        logging.info("\n" + report)
+                        writer.write(key + "\n")
+                        writer.write(report)
+                        writer.write("\n")
+
+    if args['do_predict']:
+        tokenizer = tokenizer_class.from_pretrained(args['output_dir'], do_lower_case=args['do_lower_case'])
+        model = model_class.from_pretrained(args['output_dir'])
+        eval_batch_size = args['per_device_eval_batch_size'] * args['n_device']
+        predict_dataset, _ = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, eval_batch_size, mode="test")
+        y_true, y_pred, pred_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="test")
+        output_test_results_file = os.path.join(args['output_dir'], "test_results.txt")
+        output_test_predictions_file = os.path.join(args['output_dir'], "test_predictions.txt")
+        report = metrics.classification_report(y_true, y_pred, digits=4)
+
+        with tf.io.gfile.GFile(output_test_results_file, "w") as writer:
+            report = metrics.classification_report(y_true, y_pred, digits=4)
+            
+            logging.info("\n" + report)
+            
+            writer.write(report)
+            writer.write("\n\nloss = " + str(pred_loss))
+
+        with tf.io.gfile.GFile(output_test_predictions_file, "w") as writer:
+            with tf.io.gfile.GFile(os.path.join(args['data_dir'], "test.txt"), "r") as f:
+                example_id = 0
+
+                for line in f:
+                    if line.startswith("-DOCSTART-") or line == "" or line == "\n":
+                        writer.write(line)
+
+                        if not y_pred[example_id]:
+                            example_id += 1
+                    elif y_pred[example_id]:
+                        output_line = line.split()[0] + " " + y_pred[example_id].pop(0) + "\n"
+                        writer.write(output_line)
+                    else:
+                        logging.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
+
+
+if __name__ == "__main__":
+    flags.mark_flag_as_required("data_dir")
+    flags.mark_flag_as_required("output_dir")
+    flags.mark_flag_as_required("model_name_or_path")
+    flags.mark_flag_as_required("model_type")
+    app.run(main)

examples/run_xnli.py

@@ -92,6 +92,13 @@ def train(args, train_dataset, model, tokenizer):
         ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
     scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
+
     if args.fp16:
         try:
             from apex import amp
@@ -120,13 +127,32 @@ def train(args, train_dataset, model, tokenizer):
     logger.info("  Total optimization steps = %d", t_total)
 
     global_step = 0
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to gobal_step of last saved checkpoint from model path
+        global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
+        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
+
+        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
+        logger.info("  Continuing training from epoch %d", epochs_trained)
+        logger.info("  Continuing training from global step %d", global_step)
+        logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
+
     tr_loss, logging_loss = 0.0, 0.0
     model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
     set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
     for _ in train_iterator:
         epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
         for step, batch in enumerate(epoch_iterator):
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
             model.train()
             batch = tuple(t.to(args.device) for t in batch)
             inputs = {'input_ids':      batch[0],
@@ -177,9 +203,15 @@ def train(args, train_dataset, model, tokenizer):
                         os.makedirs(output_dir)
                     model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
                     model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+
                     torch.save(args, os.path.join(output_dir, 'training_args.bin'))
                     logger.info("Saving model checkpoint to %s", output_dir)
 
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
+                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
+
             if args.max_steps > 0 and global_step > args.max_steps:
                 epoch_iterator.close()
                 break

examples/summarization/README.md

@@ -0,0 +1,61 @@
+# Text Summarization with Pretrained Encoders
+
+This folder contains part of the code necessary to reproduce the results on abstractive summarization from the article [Text Summarization with Pretrained Encoders](https://arxiv.org/pdf/1908.08345.pdf) by [Yang Liu](https://nlp-yang.github.io/) and [Mirella Lapata](https://homepages.inf.ed.ac.uk/mlap/). It can also be used to summarize any document.
+
+The original code can be found on the Yang Liu's [github repository](https://github.com/nlpyang/PreSumm).
+
+The model is loaded with the pre-trained weights for the abstractive summarization model trained on the CNN/Daily Mail dataset with an extractive and then abstractive tasks.
+
+## Setup
+
+```
+git clone https://github.com/huggingface/transformers && cd transformers
+pip install [--editable] .
+pip install nltk py-rouge
+cd examples/summarization
+```
+
+## Reproduce the authors' results on ROUGE
+
+To be able to reproduce the authors' results on the CNN/Daily Mail dataset you first need to download both CNN and Daily Mail datasets [from Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/) (the links next to "Stories") in the same folder. Then uncompress the archives by running:
+
+```bash
+tar -xvf cnn_stories.tgz && tar -xvf dailymail_stories.tgz
+```
+
+And move all the stories to the same folder. We will refer as `$DATA_PATH` the path to where you uncompressed both archive. Then run the following in the same folder as `run_summarization.py`:
+
+```bash
+python run_summarization.py \
+    --documents_dir $DATA_PATH \
+    --summaries_output_dir $SUMMARIES_PATH \ # optional
+    --no_cuda false \
+    --batch_size 4 \
+    --min_length 50 \
+    --max_length 200 \
+    --beam_size 5 \
+    --alpha 0.95 \
+    --block_trigram true \
+    --compute_rouge true
+```
+
+The scripts executes on GPU if one is available and if `no_cuda` is not set to `true`. Inference on multiple GPUs is not suported yet. The ROUGE scores will be displayed in the console at the end of evaluation and written in a `rouge_scores.txt` file. The script takes 30 hours to compute with a single Tesla V100 GPU and a batch size of 10 (300,000 texts to summarize).
+
+## Summarize any text
+
+Put the documents that you would like to summarize in a folder (the path to which is referred to as `$DATA_PATH` below) and run the following in the same folder as `run_summarization.py`:
+
+```bash
+python run_summarization.py \
+    --documents_dir $DATA_PATH \
+    --summaries_output_dir $SUMMARIES_PATH \ # optional
+    --no_cuda false \
+    --batch_size 4 \
+    --min_length 50 \
+    --max_length 200 \
+    --beam_size 5 \
+    --alpha 0.95 \
+    --block_trigram true \
+```
+
+You may want to play around with `min_length`, `max_length` and `alpha` to suit your use case. If you want to compute ROUGE on another dataset you will need to tweak the stories/summaries import in `utils_summarization.py` and tell it where to fetch the reference summaries.

examples/summarization/configuration_bertabs.py

@@ -0,0 +1,99 @@
+# coding=utf-8
+# Copyright 2019 The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" BertAbs configuration """
+import json
+import logging
+import sys
+
+from transformers import PretrainedConfig
+
+
+logger = logging.getLogger(__name__)
+
+
+BERTABS_FINETUNED_CONFIG_MAP = {
+    "bertabs-finetuned-cnndm": "https://s3.amazonaws.com/models.huggingface.co/bert/remi/bertabs-finetuned-cnndm-extractive-abstractive-summarization-config.json",
+}
+
+
+class BertAbsConfig(PretrainedConfig):
+    r""" Class to store the configuration of the BertAbs model.
+
+    Arguments:
+        vocab_size: int
+            Number of tokens in the vocabulary.
+        max_pos: int
+            The maximum sequence length that this model will be used with.
+        enc_layer: int
+            The numner of hidden layers in the Transformer encoder.
+        enc_hidden_size: int
+            The size of the encoder's layers.
+        enc_heads: int
+            The number of attention heads for each attention layer in the encoder.
+        enc_ff_size: int
+            The size of the encoder's feed-forward layers.
+        enc_dropout: int
+            The dropout probabilitiy for all fully connected layers in the
+            embeddings, layers, pooler and also the attention probabilities in
+            the encoder.
+        dec_layer: int
+            The numner of hidden layers in the decoder.
+        dec_hidden_size: int
+            The size of the decoder's layers.
+        dec_heads: int
+            The number of attention heads for each attention layer in the decoder.
+        dec_ff_size: int
+            The size of the decoder's feed-forward layers.
+        dec_dropout: int
+            The dropout probabilitiy for all fully connected layers in the
+            embeddings, layers, pooler and also the attention probabilities in
+            the decoder.
+    """
+
+    pretrained_config_archive_map = BERTABS_FINETUNED_CONFIG_MAP
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        max_pos=512,
+        enc_layers=6,
+        enc_hidden_size=512,
+        enc_heads=8,
+        enc_ff_size=512,
+        enc_dropout=0.2,
+        dec_layers=6,
+        dec_hidden_size=768,
+        dec_heads=8,
+        dec_ff_size=2048,
+        dec_dropout=0.2,
+        **kwargs,
+    ):
+        super(BertAbsConfig, self).__init__(**kwargs)
+
+        self.vocab_size = vocab_size
+        self.max_pos = max_pos
+
+        self.enc_layers = enc_layers
+        self.enc_hidden_size = enc_hidden_size
+        self.enc_heads = enc_heads
+        self.enc_ff_size = enc_ff_size
+        self.enc_dropout = enc_dropout
+
+        self.dec_layers = dec_layers
+        self.dec_hidden_size = dec_hidden_size
+        self.dec_heads = dec_heads
+        self.dec_ff_size = dec_ff_size
+        self.dec_dropout = dec_dropout

examples/summarization/convert_bertabs_original_pytorch_checkpoint.py

@@ -0,0 +1,163 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Convert BertExtAbs's checkpoints.
+
+The script looks like it is doing something trivial but it is not. The "weights"
+proposed by the authors are actually the entire model pickled. We need to load
+the model within the original codebase to be able to only save its `state_dict`.
+"""
+
+import argparse
+from collections import namedtuple
+import logging
+import torch
+
+from models.model_builder import AbsSummarizer  # The authors' implementation
+from model_bertabs import BertAbsSummarizer
+
+from transformers import BertTokenizer
+
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+SAMPLE_TEXT = 'Hello world! cécé herlolip'
+
+
+BertAbsConfig = namedtuple(
+    "BertAbsConfig",
+    ["temp_dir", "large", "use_bert_emb", "finetune_bert", "encoder", "share_emb", "max_pos", "enc_layers", "enc_hidden_size", "enc_heads", "enc_ff_size", "enc_dropout", "dec_layers", "dec_hidden_size", "dec_heads", "dec_ff_size", "dec_dropout"],
+)
+
+
+def convert_bertabs_checkpoints(path_to_checkpoints, dump_path):
+    """ Copy/paste and tweak the pre-trained weights provided by the creators
+    of BertAbs for the internal architecture.
+    """
+
+    # Instantiate the authors' model with the pre-trained weights
+    config = BertAbsConfig(
+        temp_dir=".",
+        finetune_bert=False,
+        large=False,
+        share_emb=True,
+        use_bert_emb=False,
+        encoder="bert",
+        max_pos=512,
+        enc_layers=6,
+        enc_hidden_size=512,
+        enc_heads=8,
+        enc_ff_size=512,
+        enc_dropout=0.2,
+        dec_layers=6,
+        dec_hidden_size=768,
+        dec_heads=8,
+        dec_ff_size=2048,
+        dec_dropout=0.2,
+    )
+    checkpoints = torch.load(path_to_checkpoints, lambda storage, loc: storage)
+    original = AbsSummarizer(config, torch.device("cpu"), checkpoints)
+    original.eval()
+
+    new_model = BertAbsSummarizer(config, torch.device("cpu"))
+    new_model.eval()
+
+    # -------------------
+    # Convert the weights
+    # -------------------
+
+    logging.info("convert the model")
+    new_model.bert.load_state_dict(original.bert.state_dict())
+    new_model.decoder.load_state_dict(original.decoder.state_dict())
+    new_model.generator.load_state_dict(original.generator.state_dict())
+
+    # ----------------------------------
+    # Make sure the outpus are identical
+    # ----------------------------------
+
+    logging.info("Make sure that the models' outputs are identical")
+    tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+
+    # prepare the model inputs
+    encoder_input_ids = tokenizer.encode("This is sample éàalj'-.")
+    encoder_input_ids.extend([tokenizer.pad_token_id] * (512 - len(encoder_input_ids)))
+    encoder_input_ids = torch.tensor(encoder_input_ids).unsqueeze(0)
+    decoder_input_ids = tokenizer.encode("This is sample 3 éàalj'-.")
+    decoder_input_ids.extend([tokenizer.pad_token_id] * (512 - len(decoder_input_ids)))
+    decoder_input_ids = torch.tensor(decoder_input_ids).unsqueeze(0)
+
+    # failsafe to make sure the weights reset does not affect the
+    # loaded weights.
+    assert torch.max(torch.abs(original.generator[0].weight - new_model.generator[0].weight)) == 0
+
+    # forward pass
+    src = encoder_input_ids
+    tgt = decoder_input_ids
+    segs = token_type_ids = None
+    clss = None
+    mask_src = encoder_attention_mask = None
+    mask_tgt = decoder_attention_mask = None
+    mask_cls = None
+
+    # The original model does not apply the geneator layer immediatly but rather in
+    # the beam search (where it combines softmax + linear layer). Since we already
+    # apply the softmax in our generation process we only apply the linear layer here.
+    # We make sure that the outputs of the full stack are identical
+    output_original_model = original(src, tgt, segs, clss, mask_src, mask_tgt, mask_cls)[0]
+    output_original_generator = original.generator(output_original_model)
+
+    output_converted_model = new_model(encoder_input_ids, decoder_input_ids, token_type_ids, encoder_attention_mask, decoder_attention_mask)[0]
+    output_converted_generator = new_model.generator(output_converted_model)
+
+    maximum_absolute_difference = torch.max(torch.abs(output_converted_model - output_original_model)).item()
+    print("Maximum absolute difference beween weights: {:.2f}".format(maximum_absolute_difference))
+    maximum_absolute_difference = torch.max(torch.abs(output_converted_generator - output_original_generator)).item()
+    print("Maximum absolute difference beween weights: {:.2f}".format(maximum_absolute_difference))
+
+    are_identical = torch.allclose(output_converted_model, output_original_model, atol=1e-3)
+    if are_identical:
+        logging.info("all weights are equal up to 1e-3")
+    else:
+        raise ValueError("the weights are different. The new model is likely different from the original one.")
+
+    # The model has been saved with torch.save(model) and this is bound to the exact
+    # directory structure. We save the state_dict instead.
+    logging.info("saving the model's state dictionary")
+    torch.save(new_model.state_dict(), "bertabs-finetuned-cnndm-extractive-abstractive-summarization-pytorch_model.bin")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--bertabs_checkpoint_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path the official PyTorch dump.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to the output PyTorch model.",
+    )
+    args = parser.parse_args()
+
+    convert_bertabs_checkpoints(
+        args.bertabs_checkpoint_path,
+        args.pytorch_dump_folder_path,
+    )

examples/summarization/requirements.txt

@@ -0,0 +1,9 @@
+# progress bars in model download and training scripts
+tqdm
+# Accessing files from S3 directly.
+boto3
+# Used for downloading models over HTTP
+requests
+# For ROUGE
+nltk
+py-rouge

examples/summarization/run_summarization.py

@@ -0,0 +1,344 @@
+#! /usr/bin/python3
+import argparse
+from collections import namedtuple
+import logging
+import os
+import sys
+
+import torch
+from torch.utils.data import DataLoader, SequentialSampler
+from tqdm import tqdm
+
+from transformers import BertTokenizer
+
+from modeling_bertabs import BertAbs, build_predictor
+
+from utils_summarization import (
+    SummarizationDataset,
+    encode_for_summarization,
+    build_mask,
+    fit_to_block_size,
+    compute_token_type_ids,
+)
+
+logger = logging.getLogger(__name__)
+logging.basicConfig(stream=sys.stdout, level=logging.INFO)
+
+
+Batch = namedtuple(
+    "Batch", ["document_names", "batch_size", "src", "segs", "mask_src", "tgt_str"]
+)
+
+
+def evaluate(args):
+    tokenizer = BertTokenizer.from_pretrained("bert-base-uncased", do_lower_case=True)
+    model = BertAbs.from_pretrained("bertabs-finetuned-cnndm")
+    model.to(args.device)
+    model.eval()
+
+    symbols = {
+        "BOS": tokenizer.vocab["[unused0]"],
+        "EOS": tokenizer.vocab["[unused1]"],
+        "PAD": tokenizer.vocab["[PAD]"],
+    }
+
+    if args.compute_rouge:
+        reference_summaries = []
+        generated_summaries = []
+
+        import rouge
+        import nltk
+        nltk.download('punkt')
+        rouge_evaluator = rouge.Rouge(
+            metrics=['rouge-n', 'rouge-l'],
+            max_n=2,
+            limit_length=True,
+            length_limit=args.beam_size,
+            length_limit_type='words',
+            apply_avg=True,
+            apply_best=False,
+            alpha=0.5,  # Default F1_score
+            weight_factor=1.2,
+            stemming=True,
+        )
+
+    # these (unused) arguments are defined to keep the compatibility
+    # with the legacy code and will be deleted in a next iteration.
+    args.result_path = ""
+    args.temp_dir = ""
+
+    data_iterator = build_data_iterator(args, tokenizer)
+    predictor = build_predictor(args, tokenizer, symbols, model)
+
+    logger.info("***** Running evaluation *****")
+    logger.info("  Number examples = %d", len(data_iterator.dataset))
+    logger.info("  Batch size = %d", args.batch_size)
+    logger.info("")
+    logger.info("***** Beam Search parameters *****")
+    logger.info("  Beam size = %d", args.beam_size)
+    logger.info("  Minimum length = %d", args.min_length)
+    logger.info("  Maximum length = %d", args.max_length)
+    logger.info("  Alpha (length penalty) = %.2f", args.alpha)
+    logger.info("  Trigrams %s be blocked", ("will" if args.block_trigram else "will NOT"))
+
+    for batch in tqdm(data_iterator):
+        batch_data = predictor.translate_batch(batch)
+        translations = predictor.from_batch(batch_data)
+        summaries = [format_summary(t) for t in translations]
+        save_summaries(summaries, args.summaries_output_dir, batch.document_names)
+
+        if args.compute_rouge:
+            reference_summaries += batch.tgt_str
+            generated_summaries += summaries
+
+    if args.compute_rouge:
+        scores = rouge_evaluator.get_scores(generated_summaries, reference_summaries)
+        str_scores = format_rouge_scores(scores)
+        save_rouge_scores(str_scores)
+        print(str_scores)
+
+
+def save_summaries(summaries, path, original_document_name):
+    """ Write the summaries in fies that are prefixed by the original
+    files' name with the `_summary` appended.
+
+    Attributes:
+        original_document_names: List[string]
+            Name of the document that was summarized.
+        path: string
+            Path were the summaries will be written
+        summaries: List[string]
+            The summaries that we produced.
+    """
+    for summary, document_name in zip(summaries, original_document_name):
+        # Prepare the summary file's name
+        if "." in document_name:
+            bare_document_name = ".".join(document_name.split(".")[:-1])
+            extension = document_name.split(".")[-1]
+            name = bare_document_name + "_summary." + extension
+        else:
+            name = document_name + "_summary"
+
+        file_path = os.path.join(path, name)
+        with open(file_path, "w") as output:
+            output.write(summary)
+
+
+def format_summary(translation):
+    """ Transforms the output of the `from_batch` function
+    into nicely formatted summaries.
+    """
+    raw_summary, _, _ = translation
+    summary = (
+        raw_summary.replace("[unused0]", "")
+        .replace("[unused3]", "")
+        .replace("[PAD]", "")
+        .replace("[unused1]", "")
+        .replace(r" +", " ")
+        .replace(" [unused2] ", ". ")
+        .replace("[unused2]", "")
+        .strip()
+    )
+
+    return summary
+
+
+def format_rouge_scores(scores):
+    return """\n
+****** ROUGE SCORES ******
+
+** ROUGE 1
+F1        >> {:.3f}
+Precision >> {:.3f}
+Recall    >> {:.3f}
+
+** ROUGE 2
+F1        >> {:.3f}
+Precision >> {:.3f}
+Recall    >> {:.3f}
+
+** ROUGE L
+F1        >> {:.3f}
+Precision >> {:.3f}
+Recall    >> {:.3f}""".format(
+        scores['rouge-1']['f'],
+        scores['rouge-1']['p'],
+        scores['rouge-1']['r'],
+        scores['rouge-2']['f'],
+        scores['rouge-2']['p'],
+        scores['rouge-2']['r'],
+        scores['rouge-l']['f'],
+        scores['rouge-l']['p'],
+        scores['rouge-l']['r'],
+    )
+
+
+def save_rouge_scores(str_scores):
+    with open("rouge_scores.txt", "w") as output:
+        output.write(str_scores)
+
+
+#
+# LOAD the dataset
+#
+
+
+def build_data_iterator(args, tokenizer):
+    dataset = load_and_cache_examples(args, tokenizer)
+    sampler = SequentialSampler(dataset)
+    collate_fn = lambda data: collate(data, tokenizer, block_size=512, device=args.device)
+    iterator = DataLoader(
+        dataset, sampler=sampler, batch_size=args.batch_size, collate_fn=collate_fn,
+    )
+
+    return iterator
+
+
+def load_and_cache_examples(args, tokenizer):
+    dataset = SummarizationDataset(args.documents_dir)
+    return dataset
+
+
+def collate(data, tokenizer, block_size, device):
+    """ Collate formats the data passed to the data loader.
+
+    In particular we tokenize the data batch after batch to avoid keeping them
+    all in memory. We output the data as a namedtuple to fit the original BertAbs's
+    API.
+    """
+    data = [x for x in data if not len(x[1]) == 0]  # remove empty_files
+    names = [name for name, _, _ in data]
+    summaries = [" ".join(summary_list) for _, _, summary_list in data]
+
+    encoded_text = [
+        encode_for_summarization(story, summary, tokenizer) for _, story, summary in data
+    ]
+    encoded_stories = torch.tensor(
+        [
+            fit_to_block_size(story, block_size, tokenizer.pad_token_id)
+            for story, _ in encoded_text
+        ]
+    )
+    encoder_token_type_ids = compute_token_type_ids(encoded_stories, tokenizer.cls_token_id)
+    encoder_mask = build_mask(encoded_stories, tokenizer.pad_token_id)
+
+    batch = Batch(
+        document_names=names,
+        batch_size=len(encoded_stories),
+        src=encoded_stories.to(device),
+        segs=encoder_token_type_ids.to(device),
+        mask_src=encoder_mask.to(device),
+        tgt_str=summaries,
+    )
+
+    return batch
+
+
+def decode_summary(summary_tokens, tokenizer):
+    """ Decode the summary and return it in a format
+    suitable for evaluation.
+    """
+    summary_tokens = summary_tokens.to("cpu").numpy()
+    summary = tokenizer.decode(summary_tokens)
+    sentences = summary.split(".")
+    sentences = [s + "." for s in sentences]
+    return sentences
+
+
+def main():
+    """ The main function defines the interface with the users.
+    """
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--documents_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The folder where the documents to summarize are located.",
+    )
+    parser.add_argument(
+        "--summaries_output_dir",
+        default=None,
+        type=str,
+        required=False,
+        help="The folder in wich the summaries should be written. Defaults to the folder where the documents are",
+    )
+    parser.add_argument(
+        "--compute_rouge",
+        default=False,
+        type=bool,
+        required=False,
+        help="Compute the ROUGE metrics during evaluation. Only available for the CNN/DailyMail dataset.",
+    )
+    # EVALUATION options
+    parser.add_argument(
+        "--no_cuda",
+        default=False,
+        type=bool,
+        help="Whether to force the execution on CPU.",
+    )
+    parser.add_argument(
+        "--batch_size", default=4, type=int, help="Batch size per GPU/CPU for training.",
+    )
+    # BEAM SEARCH arguments
+    parser.add_argument(
+        "--min_length",
+        default=50,
+        type=int,
+        help="Minimum number of tokens for the summaries.",
+    )
+    parser.add_argument(
+        "--max_length",
+        default=200,
+        type=int,
+        help="Maixmum number of tokens for the summaries.",
+    )
+    parser.add_argument(
+        "--beam_size",
+        default=5,
+        type=int,
+        help="The number of beams to start with for each example.",
+    )
+    parser.add_argument(
+        "--alpha",
+        default=0.95,
+        type=float,
+        help="The value of alpha for the length penalty in the beam search.",
+    )
+    parser.add_argument(
+        "--block_trigram",
+        default=True,
+        type=bool,
+        help="Whether to block the existence of repeating trigrams in the text generated by beam search.",
+    )
+    args = parser.parse_args()
+
+    # Select device (distibuted not available)
+    args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+
+    # Check the existence of directories
+    if not args.summaries_output_dir:
+        args.summaries_output_dir = args.documents_dir
+
+    if not documents_dir_is_valid(args.documents_dir):
+        raise FileNotFoundError(
+            "We could not find the directory you specified for the documents to summarize, or it was empty. Please specify a valid path."
+        )
+    os.makedirs(args.summaries_output_dir, exist_ok=True)
+
+    evaluate(args)
+
+
+def documents_dir_is_valid(path):
+    if not os.path.exists(path):
+        return False
+
+    file_list = os.listdir(path)
+    if len(file_list) == 0:
+        return False
+
+    return True
+
+
+if __name__ == "__main__":
+    main()

examples/summarization/utils_summarization.py

@@ -10,9 +10,14 @@ from torch.utils.data import Dataset
 # ------------
 
 
-class CNNDailyMailDataset(Dataset):
+class SummarizationDataset(Dataset):
     """ Abstracts the dataset used to train seq2seq models.
 
+    The class will process the documents that are located in the specified
+    folder. The preprocessing will work on any document that is reasonably
+    formatted. On the CNN/DailyMail dataset it will extract both the story
+    and the summary.
+
     CNN/Daily News:
 
     The CNN/Daily News raw datasets are downloaded from [1]. The stories are
@@ -25,33 +30,33 @@ class CNNDailyMailDataset(Dataset):
     [2] https://github.com/abisee/cnn-dailymail/
     """
 
-    def __init__(self, tokenizer, prefix="train", data_dir=""):
-        assert os.path.isdir(data_dir)
-        self.tokenizer = tokenizer
+    def __init__(self, path="", prefix="train"):
+        """ We initialize the class by listing all the documents to summarize.
+        Files are not read in memory due to the size of some datasets (like CNN/DailyMail).
+        """
+        assert os.path.isdir(path)
 
-        # We initialize the class by listing all the files that contain
-        # stories and summaries. Files are not read in memory given
-        # the size of the corpus.
-        self.stories_path = []
-        datasets = ("cnn", "dailymail")
-        for dataset in datasets:
-            path_to_stories = os.path.join(data_dir, dataset, "stories")
-            story_filenames_list = os.listdir(path_to_stories)
-            for story_filename in story_filenames_list:
-                path_to_story = os.path.join(path_to_stories, story_filename)
-                if not os.path.isfile(path_to_story):
-                    continue
-                self.stories_path.append(path_to_story)
+        self.documents = []
+        story_filenames_list = os.listdir(path)
+        for story_filename in story_filenames_list:
+            if "summary" in story_filename:
+                continue
+            path_to_story = os.path.join(path, story_filename)
+            if not os.path.isfile(path_to_story):
+                continue
+            self.documents.append(path_to_story)
 
     def __len__(self):
-        return len(self.stories_path)
+        """ Returns the number of documents. """
+        return len(self.documents)
 
     def __getitem__(self, idx):
-        story_path = self.stories_path[idx]
-        with open(story_path, encoding="utf-8") as source:
+        document_path = self.documents[idx]
+        document_name = document_path.split("/")[-1]
+        with open(document_path, encoding="utf-8") as source:
             raw_story = source.read()
             story_lines, summary_lines = process_story(raw_story)
-        return story_lines, summary_lines
+        return document_name, story_lines, summary_lines
 
 
 def process_story(raw_story):
@@ -81,7 +86,7 @@ def process_story(raw_story):
             story_lines.append(element)
         except IndexError:
             # if "@highlight" is absent from the file we pop
-            # all elements until there is None.
+            # all elements until there is None, raising an exception.
             return story_lines, []
 
     # gather summary lines
@@ -104,31 +109,22 @@ def _add_missing_period(line):
 # --------------------------
 
 
-def fit_to_block_size(sequence, block_size, pad_token):
+def fit_to_block_size(sequence, block_size, pad_token_id):
     """ Adapt the source and target sequences' lengths to the block size.
-    If the sequence is shorter than the block size we pad it with -1 ids
-    which correspond to padding tokens.
+    If the sequence is shorter we append padding token to the right of the sequence.
     """
     if len(sequence) > block_size:
         return sequence[:block_size]
     else:
-        sequence.extend([pad_token] * (block_size - len(sequence)))
+        sequence.extend([pad_token_id] * (block_size - len(sequence)))
         return sequence
 
 
-def build_lm_labels(sequence, pad_token):
-    """ Padding token, encoded as 0, are represented by the value -1 so they
-    are not taken into account in the loss computation. """
-    padded = sequence.clone()
-    padded[padded == pad_token] = -1
-    return padded
-
-
-def build_mask(sequence, pad_token):
+def build_mask(sequence, pad_token_id):
     """ Builds the mask. The attention mechanism will only attend to positions
     with value 1. """
     mask = torch.ones_like(sequence)
-    idx_pad_tokens = sequence == pad_token
+    idx_pad_tokens = sequence == pad_token_id
     mask[idx_pad_tokens] = 0
     return mask
 
@@ -138,18 +134,11 @@ def encode_for_summarization(story_lines, summary_lines, tokenizer):
     as specified in [1] by using `[SEP] [CLS]` tokens to separate
     sentences.
     """
-    story_lines_token_ids = [
-        tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line))
-        for line in story_lines
-    ]
-    summary_lines_token_ids = [
-        tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line))
-        for line in summary_lines
-    ]
-
+    story_lines_token_ids = [tokenizer.encode(line) for line in story_lines]
     story_token_ids = [
         token for sentence in story_lines_token_ids for token in sentence
     ]
+    summary_lines_token_ids = [tokenizer.encode(line) for line in summary_lines]
     summary_token_ids = [
         token for sentence in summary_lines_token_ids for token in sentence
     ]
@@ -174,7 +163,7 @@ def compute_token_type_ids(batch, separator_token_id):
     """
     batch_embeddings = []
     for sequence in batch:
-        sentence_num = 0
+        sentence_num = -1
         embeddings = []
         for s in sequence:
             if s == separator_token_id:

examples/summarization/utils_summarization_test.py

@@ -21,7 +21,6 @@ from utils_summarization import (
     compute_token_type_ids,
     fit_to_block_size,
     build_mask,
-    build_lm_labels,
     process_story,
 )
 
@@ -88,20 +87,6 @@ class SummarizationDataProcessingTest(unittest.TestCase):
         expected_summary_lines = ["It was the best of times."]
         self.assertEqual(expected_summary_lines, summary_lines)
 
-    def test_build_lm_labels_no_padding(self):
-        sequence = torch.tensor([1, 2, 3, 4])
-        expected = sequence
-        np.testing.assert_array_equal(
-            build_lm_labels(sequence, 0).numpy(), expected.numpy()
-        )
-
-    def test_build_lm_labels(self):
-        sequence = torch.tensor([1, 2, 3, 4, 0, 0, 0])
-        expected = torch.tensor([1, 2, 3, 4, -1, -1, -1])
-        np.testing.assert_array_equal(
-            build_lm_labels(sequence, 0).numpy(), expected.numpy()
-        )
-
     def test_build_mask_no_padding(self):
         sequence = torch.tensor([1, 2, 3, 4])
         expected = torch.tensor([1, 1, 1, 1])
@@ -125,7 +110,7 @@ class SummarizationDataProcessingTest(unittest.TestCase):
             [[1, 2, 3, 4, 5, 6], [1, 2, 3, 101, 5, 6], [1, 101, 3, 4, 101, 6]]
         )
         expected = torch.tensor(
-            [[0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1], [0, 1, 1, 1, 0, 0]]
+            [[1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 0, 0], [1, 0, 0, 0, 1, 1]]
         )
 
         result = compute_token_type_ids(batch, separator)

examples/test_examples.py

@@ -72,8 +72,7 @@ class ExamplesTests(unittest.TestCase):
         logger.addHandler(stream_handler)
 
         testargs = ["run_squad.py",
-                    "--train_file=./examples/tests_samples/SQUAD/dev-v2.0-small.json",
-                    "--predict_file=./examples/tests_samples/SQUAD/dev-v2.0-small.json",
+                    "--data_dir=./examples/tests_samples/SQUAD",
                     "--model_name=bert-base-uncased",
                     "--output_dir=./examples/tests_samples/temp_dir",
                     "--max_steps=10",

examples/tests_samples/SQUAD/train-v2.0.json

@@ -0,0 +1,140 @@
+{
+    "version": "v2.0",
+    "data": [{
+        "title": "Normans",
+        "paragraphs": [{
+            "qas": [{
+                "question": "In what country is Normandy located?",
+                "id": "56ddde6b9a695914005b9628",
+                "answers": [{
+                    "text": "France",
+                    "answer_start": 159
+                }],
+                "is_impossible": false
+            }, {
+                "question": "When were the Normans in Normandy?",
+                "id": "56ddde6b9a695914005b9629",
+                "answers": [{
+                    "text": "10th and 11th centuries",
+                    "answer_start": 94
+                }],
+                "is_impossible": false
+            }, {
+                "question": "From which countries did the Norse originate?",
+                "id": "56ddde6b9a695914005b962a",
+                "answers": [{
+                    "text": "Denmark, Iceland and Norway",
+                    "answer_start": 256
+                }],
+                "is_impossible": false
+            }, {
+                "plausible_answers": [{
+                    "text": "Rollo",
+                    "answer_start": 308
+                }],
+                "question": "Who did King Charles III swear fealty to?",
+                "id": "5ad39d53604f3c001a3fe8d3",
+                "answers": [],
+                "is_impossible": true
+            }, {
+                "plausible_answers": [{
+                    "text": "10th century",
+                    "answer_start": 671
+                }],
+                "question": "When did the Frankish identity emerge?",
+                "id": "5ad39d53604f3c001a3fe8d4",
+                "answers": [],
+                "is_impossible": true
+            }],
+            "context": "The Normans (Norman: Nourmands; French: Normands; Latin: Normanni) were the people who in the 10th and 11th centuries gave their name to Normandy, a region in France. They were descended from Norse (\"Norman\" comes from \"Norseman\") raiders and pirates from Denmark, Iceland and Norway who, under their leader Rollo, agreed to swear fealty to King Charles III of West Francia. Through generations of assimilation and mixing with the native Frankish and Roman-Gaulish populations, their descendants would gradually merge with the Carolingian-based cultures of West Francia. The distinct cultural and ethnic identity of the Normans emerged initially in the first half of the 10th century, and it continued to evolve over the succeeding centuries."
+        }, {
+            "qas": [{
+                "question": "Who was the duke in the battle of Hastings?",
+                "id": "56dddf4066d3e219004dad5f",
+                "answers": [{
+                    "text": "William the Conqueror",
+                    "answer_start": 1022
+                }],
+                "is_impossible": false
+            }, {
+                "plausible_answers": [{
+                    "text": "Antioch",
+                    "answer_start": 1295
+                }],
+                "question": "What principality did William the conquerer found?",
+                "id": "5ad3a266604f3c001a3fea2b",
+                "answers": [],
+                "is_impossible": true
+            }],
+            "context": "The Norman dynasty had a major political, cultural and military impact on medieval Europe and even the Near East. The Normans were famed for their martial spirit and eventually for their Christian piety, becoming exponents of the Catholic orthodoxy into which they assimilated. They adopted the Gallo-Romance language of the Frankish land they settled, their dialect becoming known as Norman, Normaund or Norman French, an important literary language. The Duchy of Normandy, which they formed by treaty with the French crown, was a great fief of medieval France, and under Richard I of Normandy was forged into a cohesive and formidable principality in feudal tenure. The Normans are noted both for their culture, such as their unique Romanesque architecture and musical traditions, and for their significant military accomplishments and innovations. Norman adventurers founded the Kingdom of Sicily under Roger II after conquering southern Italy on the Saracens and Byzantines, and an expedition on behalf of their duke, William the Conqueror, led to the Norman conquest of England at the Battle of Hastings in 1066. Norman cultural and military influence spread from these new European centres to the Crusader states of the Near East, where their prince Bohemond I founded the Principality of Antioch in the Levant, to Scotland and Wales in Great Britain, to Ireland, and to the coasts of north Africa and the Canary Islands."
+        }]
+    }, {
+        "title": "Computational_complexity_theory",
+        "paragraphs": [{
+            "qas": [{
+                "question": "What branch of theoretical computer science deals with broadly classifying computational problems by difficulty and class of relationship?",
+                "id": "56e16182e3433e1400422e28",
+                "answers": [{
+                    "text": "Computational complexity theory",
+                    "answer_start": 0
+                }],
+                "is_impossible": false
+            }, {
+                "plausible_answers": [{
+                    "text": "algorithm",
+                    "answer_start": 472
+                }],
+                "question": "What is a manual application of mathematical steps?",
+                "id": "5ad5316b5b96ef001a10ab76",
+                "answers": [],
+                "is_impossible": true
+            }],
+            "context": "Computational complexity theory is a branch of the theory of computation in theoretical computer science that focuses on classifying computational problems according to their inherent difficulty, and relating those classes to each other. A computational problem is understood to be a task that is in principle amenable to being solved by a computer, which is equivalent to stating that the problem may be solved by mechanical application of mathematical steps, such as an algorithm."
+        }, {
+            "qas": [{
+                "question": "What measure of a computational problem broadly defines the inherent difficulty of the solution?",
+                "id": "56e16839cd28a01900c67887",
+                "answers": [{
+                    "text": "if its solution requires significant resources",
+                    "answer_start": 46
+                }],
+                "is_impossible": false
+            }, {
+                "question": "What method is used to intuitively assess or quantify the amount of resources required to solve a computational problem?",
+                "id": "56e16839cd28a01900c67888",
+                "answers": [{
+                    "text": "mathematical models of computation",
+                    "answer_start": 176
+                }],
+                "is_impossible": false
+            }, {
+                "question": "What are two basic primary resources used to guage complexity?",
+                "id": "56e16839cd28a01900c67889",
+                "answers": [{
+                    "text": "time and storage",
+                    "answer_start": 305
+                }],
+                "is_impossible": false
+            }, {
+                "plausible_answers": [{
+                    "text": "the number of gates in a circuit",
+                    "answer_start": 436
+                }],
+                "question": "What unit is measured to determine circuit simplicity?",
+                "id": "5ad532575b96ef001a10ab7f",
+                "answers": [],
+                "is_impossible": true
+            }, {
+                "plausible_answers": [{
+                    "text": "the number of processors",
+                    "answer_start": 502
+                }],
+                "question": "What number is used in perpendicular computing?",
+                "id": "5ad532575b96ef001a10ab80",
+                "answers": [],
+                "is_impossible": true
+            }],
+            "context": "A problem is regarded as inherently difficult if its solution requires significant resources, whatever the algorithm used. The theory formalizes this intuition, by introducing mathematical models of computation to study these problems and quantifying the amount of resources needed to solve them, such as time and storage. Other complexity measures are also used, such as the amount of communication (used in communication complexity), the number of gates in a circuit (used in circuit complexity) and the number of processors (used in parallel computing). One of the roles of computational complexity theory is to determine the practical limits on what computers can and cannot do."
+        }]
+    }]
+}

examples/utils_ner.py

@@ -94,7 +94,7 @@ def convert_examples_to_features(examples,
                                  pad_on_left=False,
                                  pad_token=0,
                                  pad_token_segment_id=0,
-                                 pad_token_label_id=-1,
+                                 pad_token_label_id=-100,
                                  sequence_a_segment_id=0,
                                  mask_padding_with_zero=True):
     """ Loads a data file into a list of `InputBatch`s

examples/utils_semeval.py

@@ -1,309 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" Finetuning the library models for sequence classification on GLUE (Bert, XLM, XLNet, RoBERTa)."""
-
-from __future__ import absolute_import, division, print_function
-
-import logging
-import os
-import re
-import torch
-from torch.utils.data import (TensorDataset)
-from sklearn.metrics import  f1_score
-
-from transformers import RobertaTokenizer
-
-logger = logging.getLogger(__name__)
-
-
-class RBertUtils:
-    """
-    Utility class to prepare text for RBERT
-
-
-    Args:
-    same as BertTokenizer, apart from
-    replace_conflicting_entity_offset_char: since the forward method of RBert classification models detects
-                                            the special characters that bound each entity, any pre-existing characters
-                                            of this type will cause a conflict. To prevent this, we replace any affected
-                                            characters with this, prior to encoding. Defaults to '|'
-    ent1_sep_token: the char to delimit entity 1. Default is '$'
-    ent2_sep_token: the char to delimit entity 2. Default is '#'
-
-
-    """
-
-    def __init__(self, replace_conflicting_entity_offset_char='|',
-                 ent1_sep_token="$", ent2_sep_token="#", ):
-        self.ent1_sep_token = ent1_sep_token
-        self.ent2_sep_token = ent2_sep_token
-        self.replace_conflicting_entity_offset_char = replace_conflicting_entity_offset_char
-
-    def _clean_special_chars(self, text: str):
-        return text.replace(self.ent1_sep_token,
-                            self.replace_conflicting_entity_offset_char).replace(self.ent2_sep_token,
-                                                                                 self.replace_conflicting_entity_offset_char)
-
-    def _insert_relationship_special_chars(self, text, e1_offset_tup, e2_offset_tup):
-        reverse_ordered_tups = sorted([e1_offset_tup, e2_offset_tup], reverse=True, key=lambda x: x[1])
-        last_token_to_process = True
-        for start, end in reverse_ordered_tups:
-            before = text[:start]
-            during = text[start:end]
-            after = text[end:]
-            if last_token_to_process:
-                text = before + " " + self.ent2_sep_token + " " + during + " " + self.ent2_sep_token + " " + after
-            else:
-                text = before + " " + self.ent1_sep_token + " " + during + " " + self.ent1_sep_token + " " + after
-            last_token_to_process = False
-        return text
-
-    def insert_special_chars_into_text(self, text, e1_offset_tup, e2_offset_tup):
-        text = self._clean_special_chars(text)
-        text = self._insert_relationship_special_chars(text, e1_offset_tup, e2_offset_tup)
-        return text
-
-class InputExample(object):
-    """A single SemEval 2010 Task 8 example"""
-
-    def __init__(self, id: int, text: str, label: str, comment: str):
-        self.comment = comment
-        self.label = label
-        self.text = text
-        self.id = id
-
-
-class InputFeatures(object):
-    """A single set of features of data."""
-
-    def __init__(self, input_ids, input_mask, segment_ids, label_id, instance_id):
-        self.instance_id = instance_id
-        self.input_ids = input_ids
-        self.input_mask = input_mask
-        self.segment_ids = segment_ids
-        self.label_id = label_id
-
-
-
-class SemEval2010Task8DataProcessor():
-    """Processor for the SemEval 2010 Task 8 Dataset. Note, it's not clear from the RBert paper how many classes the
-    model is trained on. Clearly this has a big impact on the result.
-
-
-    """
-    def __init__(self,include_directionality=True):
-        self.include_directionality = include_directionality
-        self.e1_e2_labels = [
-                         'Product-Producer(e1,e2)',
-                         'Entity-Origin(e1,e2)',
-                         'Entity-Destination(e1,e2)',
-                         'Message-Topic(e1,e2)',
-                         'Component-Whole(e1,e2)',
-                         'Content-Container(e1,e2)',
-                         'Instrument-Agency(e1,e2)',
-                         'Cause-Effect(e1,e2)',
-                         'Member-Collection(e1,e2)'
-                       ]
-        self.e2_e1_labels = [
-                         'Product-Producer(e2,e1)',
-                         'Entity-Origin(e2,e1)',
-                         'Entity-Destination(e2,e1)',
-                         'Message-Topic(e2,e1)',
-                         'Component-Whole(e2,e1)',
-                         'Content-Container(e2,e1)',
-                         'Instrument-Agency(e2,e1)',
-                         'Cause-Effect(e2,e1)',
-                         'Member-Collection(e2,e1)'
-                       ]
-        self.undirected_labels_mapping = dict(zip(self.e2_e1_labels,self.e1_e2_labels))
-        self.other_label = ['Other']
-        self.undirected_labels_mapping[self.other_label[0]] = self.other_label[0]
-
-        if self.include_directionality:
-            self.all_labels = self.e1_e2_labels + self.e2_e1_labels + self.other_label
-        else:
-            self.all_labels = self.e1_e2_labels  + self.other_label
-
-    def _instance_generator(self, data,include_other):
-        for i in range(0, len(data), 4):
-            id, text = data[i].split('\t')
-            text = text.strip()
-            label = str(data[i + 1]).strip()
-            # map all labels to single direction if not using directionality
-            if not self.include_directionality and label in self.undirected_labels_mapping:
-                label = self.undirected_labels_mapping[label]
-            # Don't include Other label in training set
-            if label == 'Other' and not include_other:
-                logger.info(f'Skipping Other labeled instance at {id}')
-            else:
-                comment = data[i + 2].strip()
-                yield InputExample(id=id, text=text, label=label, comment=comment)
-
-
-    def _strip_direction(self,label):
-        return re.sub(r'\(.*', "", label)
-
-    def _create_examples(self, path,include_other):
-        with open(path, 'r') as f:
-            data = f.readlines()
-        return self._instance_generator(data,include_other)
-
-    def get_train_examples(self, data_dir,include_other):
-        return self._create_examples(os.path.join(data_dir, "SemEval2010_task8_training","TRAIN_FILE.TXT"),
-                                     include_other)
-
-    def get_dev_examples(self, data_dir,include_other):
-        return self._create_examples(os.path.join(data_dir, "SemEval2010_task8_testing_keys","TEST_FILE_FULL.TXT"),
-                                     include_other)
-
-    def write_dev_examples_to_official_format(self,data_dir,examples):
-        with open(os.path.join(data_dir,"TEST_FILE_SEMEVAL_SCRIPT_FORMAT.tsv"),'w') as f:
-            for example in examples:
-                f.write(str(example.id)+'\t'+example.label+'\n')
-
-
-    def get_labels(self):
-        return self.all_labels
-
-def mod_text(text, hit):
-    before = text[:hit.start()]
-    during = hit.group(2)
-    after = text[hit.end():]
-
-    start = len(before)
-    end = len(before) + len(during)
-    return (before + during + after, [start, end],)
-
-
-def find_ents_and_modify_string(text):
-    e1_hits = re.search("(<e1>)(.*)(</e1>)", text)
-    new_text, e1_offsets = mod_text(text, e1_hits)
-    e2_hits = re.search("(<e2>)(.*)(</e2>)", new_text)
-    new_text, e2_offsets = mod_text(new_text, e2_hits)
-
-    return new_text, e1_offsets, e2_offsets
-
-def get_entity_seperator_token_ids(rbert_utils,tokenizer):
-    if isinstance(tokenizer, RobertaTokenizer):
-        entity_1_token_id = \
-        tokenizer.encode(rbert_utils.ent1_sep_token, text_pair=None, add_special_tokens=True, add_prefix_space=True)[1]
-        entity_2_token_id = \
-        tokenizer.encode(rbert_utils.ent2_sep_token, text_pair=None, add_special_tokens=True, add_prefix_space=True)[1]
-    else:
-        entity_1_token_id = tokenizer.encode(rbert_utils.ent1_sep_token)[0]
-        entity_2_token_id = tokenizer.encode(rbert_utils.ent2_sep_token)[0]
-
-    return entity_1_token_id, entity_2_token_id
-
-def get_input_ids(text, tokenizer):
-    if isinstance(tokenizer, RobertaTokenizer):
-        input_ids = tokenizer.encode(text, text_pair=None, add_special_tokens=True, add_prefix_space=True)
-    else:
-        input_ids = tokenizer.encode(text, text_pair=None, add_special_tokens=True)
-    return input_ids
-
-def convert_examples_to_features(examples, label_list, max_seq_length,
-                                 tokenizer, rbert_utils,pad_token=0,
-                                 mask_padding_with_zero=True):
-    label_map = {label: i for i, label in enumerate(label_list)}
-    features = []
-    for (ex_index, example) in enumerate(examples):
-        if ex_index % 5000 == 0:
-            logger.info("Writing example %d" % (ex_index))
-
-
-        text_with_tags_removed, e1_offsets, e2_offsets = find_ents_and_modify_string(example.text)
-        new_text = rbert_utils.insert_special_chars_into_text(text_with_tags_removed,e1_offsets,e2_offsets)
-
-        entity_1_token_id, entity_2_token_id = get_entity_seperator_token_ids(rbert_utils,tokenizer)
-        input_ids = get_input_ids(new_text,tokenizer)
-
-
-
-
-        #check that the special tokens have been encoded in the right order
-        special_tokens_ordered = list(filter(lambda n: n in [entity_1_token_id, entity_2_token_id], input_ids))
-        assert special_tokens_ordered[0] == entity_1_token_id
-        assert special_tokens_ordered[1] == entity_1_token_id
-        assert special_tokens_ordered[2] == entity_2_token_id
-        assert special_tokens_ordered[3] == entity_2_token_id
-
-        # Account for [CLS] and [SEP] with "- 2"
-        if len(input_ids) > max_seq_length - 2:
-            input_ids = input_ids[:(max_seq_length - 2)]
-
-        # The mask has 1 for real tokens and 0 for padding tokens. Only real
-        # tokens are attended to.
-        input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
-
-        # Zero-pad up to the sequence length.
-        padding_length = max_seq_length - len(input_ids)
-
-        input_ids = input_ids + ([pad_token] * padding_length)
-        input_mask = input_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
-
-        assert len(input_ids) == max_seq_length
-        assert len(input_mask) == max_seq_length
-
-        label_id = label_map[example.label]
-        features.append(
-            InputFeatures(input_ids=input_ids,
-                          input_mask=input_mask,
-                          segment_ids=[0] * len(input_ids),
-                          label_id=label_id,
-                          instance_id=int(example.id)
-                          ))
-    return features
-
-
-def convert_features_to_dataset(features, output_mode='classification'):
-    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
-    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
-    all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
-    if output_mode == "classification":
-        all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.long)
-    else:
-        all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.float)
-    instance_ids = torch.tensor([f.instance_id for f in features], dtype=torch.int)
-    dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids,instance_ids)
-    return dataset
-
-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,average='macro')
-    return {
-        "acc": acc,
-        "f1": f1,
-        "acc_and_f1": (acc + f1) / 2,
-    }
-def compute_metrics(task_name, preds, labels):
-    assert len(preds) == len(labels)
-    if task_name == "semeval2010_task8":
-        return acc_and_f1(labels, preds)
-    else:
-        raise KeyError(task_name)
-
-
-processors = {
-    "semeval2010_task8": SemEval2010Task8DataProcessor
-}
-
-output_modes = {
-    "semeval2010_task8": "classification"
-}

examples/utils_squad_evaluate.py

@@ -1,330 +0,0 @@
-""" Official evaluation script for SQuAD version 2.0.
-    Modified by XLNet authors to update `find_best_threshold` scripts for SQuAD V2.0
-
-In addition to basic functionality, we also compute additional statistics and
-plot precision-recall curves if an additional na_prob.json file is provided.
-This file is expected to map question ID's to the model's predicted probability
-that a question is unanswerable.
-"""
-import argparse
-import collections
-import json
-import numpy as np
-import os
-import re
-import string
-import sys
-
-class EVAL_OPTS():
-  def __init__(self, data_file, pred_file, out_file="",
-               na_prob_file="na_prob.json", na_prob_thresh=1.0,
-               out_image_dir=None, verbose=False):
-    self.data_file = data_file
-    self.pred_file = pred_file
-    self.out_file = out_file
-    self.na_prob_file = na_prob_file
-    self.na_prob_thresh = na_prob_thresh
-    self.out_image_dir = out_image_dir
-    self.verbose = verbose
-
-OPTS = None
-
-def parse_args():
-  parser = argparse.ArgumentParser('Official evaluation script for SQuAD version 2.0.')
-  parser.add_argument('data_file', metavar='data.json', help='Input data JSON file.')
-  parser.add_argument('pred_file', metavar='pred.json', help='Model predictions.')
-  parser.add_argument('--out-file', '-o', metavar='eval.json',
-                      help='Write accuracy metrics to file (default is stdout).')
-  parser.add_argument('--na-prob-file', '-n', metavar='na_prob.json',
-                      help='Model estimates of probability of no answer.')
-  parser.add_argument('--na-prob-thresh', '-t', type=float, default=1.0,
-                      help='Predict "" if no-answer probability exceeds this (default = 1.0).')
-  parser.add_argument('--out-image-dir', '-p', metavar='out_images', default=None,
-                      help='Save precision-recall curves to directory.')
-  parser.add_argument('--verbose', '-v', action='store_true')
-  if len(sys.argv) == 1:
-    parser.print_help()
-    sys.exit(1)
-  return parser.parse_args()
-
-def make_qid_to_has_ans(dataset):
-  qid_to_has_ans = {}
-  for article in dataset:
-    for p in article['paragraphs']:
-      for qa in p['qas']:
-        qid_to_has_ans[qa['id']] = bool(qa['answers'])
-  return qid_to_has_ans
-
-def normalize_answer(s):
-  """Lower text and remove punctuation, articles and extra whitespace."""
-  def remove_articles(text):
-    regex = re.compile(r'\b(a|an|the)\b', re.UNICODE)
-    return re.sub(regex, ' ', text)
-  def white_space_fix(text):
-    return ' '.join(text.split())
-  def remove_punc(text):
-    exclude = set(string.punctuation)
-    return ''.join(ch for ch in text if ch not in exclude)
-  def lower(text):
-    return text.lower()
-  return white_space_fix(remove_articles(remove_punc(lower(s))))
-
-def get_tokens(s):
-  if not s: return []
-  return normalize_answer(s).split()
-
-def compute_exact(a_gold, a_pred):
-  return int(normalize_answer(a_gold) == normalize_answer(a_pred))
-
-def compute_f1(a_gold, a_pred):
-  gold_toks = get_tokens(a_gold)
-  pred_toks = get_tokens(a_pred)
-  common = collections.Counter(gold_toks) & collections.Counter(pred_toks)
-  num_same = sum(common.values())
-  if len(gold_toks) == 0 or len(pred_toks) == 0:
-    # If either is no-answer, then F1 is 1 if they agree, 0 otherwise
-    return int(gold_toks == pred_toks)
-  if num_same == 0:
-    return 0
-  precision = 1.0 * num_same / len(pred_toks)
-  recall = 1.0 * num_same / len(gold_toks)
-  f1 = (2 * precision * recall) / (precision + recall)
-  return f1
-
-def get_raw_scores(dataset, preds):
-  exact_scores = {}
-  f1_scores = {}
-  for article in dataset:
-    for p in article['paragraphs']:
-      for qa in p['qas']:
-        qid = qa['id']
-        gold_answers = [a['text'] for a in qa['answers']
-                        if normalize_answer(a['text'])]
-        if not gold_answers:
-          # For unanswerable questions, only correct answer is empty string
-          gold_answers = ['']
-        if qid not in preds:
-          print('Missing prediction for %s' % qid)
-          continue
-        a_pred = preds[qid]
-        # Take max over all gold answers
-        exact_scores[qid] = max(compute_exact(a, a_pred) for a in gold_answers)
-        f1_scores[qid] = max(compute_f1(a, a_pred) for a in gold_answers)
-  return exact_scores, f1_scores
-
-def apply_no_ans_threshold(scores, na_probs, qid_to_has_ans, na_prob_thresh):
-  new_scores = {}
-  for qid, s in scores.items():
-    pred_na = na_probs[qid] > na_prob_thresh
-    if pred_na:
-      new_scores[qid] = float(not qid_to_has_ans[qid])
-    else:
-      new_scores[qid] = s
-  return new_scores
-
-def make_eval_dict(exact_scores, f1_scores, qid_list=None):
-  if not qid_list:
-    total = len(exact_scores)
-    return collections.OrderedDict([
-        ('exact', 100.0 * sum(exact_scores.values()) / total),
-        ('f1', 100.0 * sum(f1_scores.values()) / total),
-        ('total', total),
-    ])
-  else:
-    total = len(qid_list)
-    return collections.OrderedDict([
-        ('exact', 100.0 * sum(exact_scores[k] for k in qid_list) / total),
-        ('f1', 100.0 * sum(f1_scores[k] for k in qid_list) / total),
-        ('total', total),
-    ])
-
-def merge_eval(main_eval, new_eval, prefix):
-  for k in new_eval:
-    main_eval['%s_%s' % (prefix, k)] = new_eval[k]
-
-def plot_pr_curve(precisions, recalls, out_image, title):
-  plt.step(recalls, precisions, color='b', alpha=0.2, where='post')
-  plt.fill_between(recalls, precisions, step='post', alpha=0.2, color='b')
-  plt.xlabel('Recall')
-  plt.ylabel('Precision')
-  plt.xlim([0.0, 1.05])
-  plt.ylim([0.0, 1.05])
-  plt.title(title)
-  plt.savefig(out_image)
-  plt.clf()
-
-def make_precision_recall_eval(scores, na_probs, num_true_pos, qid_to_has_ans,
-                               out_image=None, title=None):
-  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
-  true_pos = 0.0
-  cur_p = 1.0
-  cur_r = 0.0
-  precisions = [1.0]
-  recalls = [0.0]
-  avg_prec = 0.0
-  for i, qid in enumerate(qid_list):
-    if qid_to_has_ans[qid]:
-      true_pos += scores[qid]
-    cur_p = true_pos / float(i+1)
-    cur_r = true_pos / float(num_true_pos)
-    if i == len(qid_list) - 1 or na_probs[qid] != na_probs[qid_list[i+1]]:
-      # i.e., if we can put a threshold after this point
-      avg_prec += cur_p * (cur_r - recalls[-1])
-      precisions.append(cur_p)
-      recalls.append(cur_r)
-  if out_image:
-    plot_pr_curve(precisions, recalls, out_image, title)
-  return {'ap': 100.0 * avg_prec}
-
-def run_precision_recall_analysis(main_eval, exact_raw, f1_raw, na_probs, 
-                                  qid_to_has_ans, out_image_dir):
-  if out_image_dir and not os.path.exists(out_image_dir):
-    os.makedirs(out_image_dir)
-  num_true_pos = sum(1 for v in qid_to_has_ans.values() if v)
-  if num_true_pos == 0:
-    return
-  pr_exact = make_precision_recall_eval(
-      exact_raw, na_probs, num_true_pos, qid_to_has_ans,
-      out_image=os.path.join(out_image_dir, 'pr_exact.png'),
-      title='Precision-Recall curve for Exact Match score')
-  pr_f1 = make_precision_recall_eval(
-      f1_raw, na_probs, num_true_pos, qid_to_has_ans,
-      out_image=os.path.join(out_image_dir, 'pr_f1.png'),
-      title='Precision-Recall curve for F1 score')
-  oracle_scores = {k: float(v) for k, v in qid_to_has_ans.items()}
-  pr_oracle = make_precision_recall_eval(
-      oracle_scores, na_probs, num_true_pos, qid_to_has_ans,
-      out_image=os.path.join(out_image_dir, 'pr_oracle.png'),
-      title='Oracle Precision-Recall curve (binary task of HasAns vs. NoAns)')
-  merge_eval(main_eval, pr_exact, 'pr_exact')
-  merge_eval(main_eval, pr_f1, 'pr_f1')
-  merge_eval(main_eval, pr_oracle, 'pr_oracle')
-
-def histogram_na_prob(na_probs, qid_list, image_dir, name):
-  if not qid_list:
-    return
-  x = [na_probs[k] for k in qid_list]
-  weights = np.ones_like(x) / float(len(x))
-  plt.hist(x, weights=weights, bins=20, range=(0.0, 1.0))
-  plt.xlabel('Model probability of no-answer')
-  plt.ylabel('Proportion of dataset')
-  plt.title('Histogram of no-answer probability: %s' % name)
-  plt.savefig(os.path.join(image_dir, 'na_prob_hist_%s.png' % name))
-  plt.clf()
-
-def find_best_thresh(preds, scores, na_probs, qid_to_has_ans):
-  num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
-  cur_score = num_no_ans
-  best_score = cur_score
-  best_thresh = 0.0
-  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
-  for i, qid in enumerate(qid_list):
-    if qid not in scores: continue
-    if qid_to_has_ans[qid]:
-      diff = scores[qid]
-    else:
-      if preds[qid]:
-        diff = -1
-      else:
-        diff = 0
-    cur_score += diff
-    if cur_score > best_score:
-      best_score = cur_score
-      best_thresh = na_probs[qid]
-  return 100.0 * best_score / len(scores), best_thresh
-
-def find_best_thresh_v2(preds, scores, na_probs, qid_to_has_ans):
-  num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
-  cur_score = num_no_ans
-  best_score = cur_score
-  best_thresh = 0.0
-  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
-  for i, qid in enumerate(qid_list):
-    if qid not in scores: continue
-    if qid_to_has_ans[qid]:
-      diff = scores[qid]
-    else:
-      if preds[qid]:
-        diff = -1
-      else:
-        diff = 0
-    cur_score += diff
-    if cur_score > best_score:
-      best_score = cur_score
-      best_thresh = na_probs[qid]
-
-  has_ans_score, has_ans_cnt = 0, 0
-  for qid in qid_list:
-    if not qid_to_has_ans[qid]: continue
-    has_ans_cnt += 1
-
-    if qid not in scores: continue
-    has_ans_score += scores[qid]
-
-  return 100.0 * best_score / len(scores), best_thresh, 1.0 * has_ans_score / has_ans_cnt
-
-def find_all_best_thresh(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
-  best_exact, exact_thresh = find_best_thresh(preds, exact_raw, na_probs, qid_to_has_ans)
-  best_f1, f1_thresh = find_best_thresh(preds, f1_raw, na_probs, qid_to_has_ans)
-  main_eval['best_exact'] = best_exact
-  main_eval['best_exact_thresh'] = exact_thresh
-  main_eval['best_f1'] = best_f1
-  main_eval['best_f1_thresh'] = f1_thresh
-
-def find_all_best_thresh_v2(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
-  best_exact, exact_thresh, has_ans_exact = find_best_thresh_v2(preds, exact_raw, na_probs, qid_to_has_ans)
-  best_f1, f1_thresh, has_ans_f1 = find_best_thresh_v2(preds, f1_raw, na_probs, qid_to_has_ans)
-  main_eval['best_exact'] = best_exact
-  main_eval['best_exact_thresh'] = exact_thresh
-  main_eval['best_f1'] = best_f1
-  main_eval['best_f1_thresh'] = f1_thresh
-  main_eval['has_ans_exact'] = has_ans_exact
-  main_eval['has_ans_f1'] = has_ans_f1
-
-def main(OPTS):
-  with open(OPTS.data_file) as f:
-    dataset_json = json.load(f)
-    dataset = dataset_json['data']
-  with open(OPTS.pred_file) as f:
-    preds = json.load(f)
-  if OPTS.na_prob_file:
-    with open(OPTS.na_prob_file) as f:
-      na_probs = json.load(f)
-  else:
-    na_probs = {k: 0.0 for k in preds}
-  qid_to_has_ans = make_qid_to_has_ans(dataset)  # maps qid to True/False
-  has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
-  no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
-  exact_raw, f1_raw = get_raw_scores(dataset, preds)
-  exact_thresh = apply_no_ans_threshold(exact_raw, na_probs, qid_to_has_ans,
-                                        OPTS.na_prob_thresh)
-  f1_thresh = apply_no_ans_threshold(f1_raw, na_probs, qid_to_has_ans,
-                                     OPTS.na_prob_thresh)
-  out_eval = make_eval_dict(exact_thresh, f1_thresh)
-  if has_ans_qids:
-    has_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=has_ans_qids)
-    merge_eval(out_eval, has_ans_eval, 'HasAns')
-  if no_ans_qids:
-    no_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=no_ans_qids)
-    merge_eval(out_eval, no_ans_eval, 'NoAns')
-  if OPTS.na_prob_file:
-    find_all_best_thresh(out_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans)
-  if OPTS.na_prob_file and OPTS.out_image_dir:
-    run_precision_recall_analysis(out_eval, exact_raw, f1_raw, na_probs, 
-                                  qid_to_has_ans, OPTS.out_image_dir)
-    histogram_na_prob(na_probs, has_ans_qids, OPTS.out_image_dir, 'hasAns')
-    histogram_na_prob(na_probs, no_ans_qids, OPTS.out_image_dir, 'noAns')
-  if OPTS.out_file:
-    with open(OPTS.out_file, 'w') as f:
-      json.dump(out_eval, f)
-  else:
-    print(json.dumps(out_eval, indent=2))
-  return out_eval
-
-if __name__ == '__main__':
-  OPTS = parse_args()
-  if OPTS.out_image_dir:
-    import matplotlib
-    matplotlib.use('Agg')
-    import matplotlib.pyplot as plt 
-  main(OPTS)

requirements.txt

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

setup.py

@@ -38,13 +38,15 @@ from setuptools import find_packages, setup
 
 
 extras = {
-    'serving': ['uvicorn', 'fastapi']
+    'serving': ['pydantic', 'uvicorn', 'fastapi'],
+    'serving-tf': ['pydantic', 'uvicorn', 'fastapi', 'tensorflow'],
+    'serving-torch': ['pydantic', 'uvicorn', 'fastapi', 'torch']
 }
 extras['all'] = [package for package in extras.values()]
 
 setup(
     name="transformers",
-    version="2.2.1",
+    version="2.3.0",
     author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
     author_email="thomas@huggingface.co",
     description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch",
@@ -57,22 +59,17 @@ setup(
                                     "tests.*", "tests"]),
     install_requires=['numpy',
                       'boto3',
+                      'filelock',
                       'requests',
                       'tqdm',
-                      'regex',
+                      'regex != 2019.12.17',
                       'sentencepiece',
                       'sacremoses'],
-    entry_points={
-      'console_scripts': [
-        "transformers=transformers.__main__:main",
-      ]
-    },
     extras_require=extras,
     scripts=[
         'transformers-cli'
     ],
     # python_requires='>=3.5.0',
-    tests_require=['pytest'],
     classifiers=[
           'Intended Audience :: Science/Research',
           'License :: OSI Approved :: Apache Software License',

templates/adding_a_new_model/configuration_xxx.py

@@ -39,7 +39,7 @@ class XxxConfig(PretrainedConfig):
 
 
         Arguments:
-            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `XxxModel`.
+            vocab_size: Vocabulary size of `inputs_ids` in `XxxModel`.
             hidden_size: Size of the encoder layers and the pooler layer.
             num_hidden_layers: Number of hidden layers in the Transformer encoder.
             num_attention_heads: Number of attention heads for each attention layer in
@@ -64,7 +64,7 @@ class XxxConfig(PretrainedConfig):
     pretrained_config_archive_map = XXX_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=50257,
+                 vocab_size=50257,
                  n_positions=1024,
                  n_ctx=1024,
                  n_embd=768,
@@ -75,8 +75,6 @@ class XxxConfig(PretrainedConfig):
                  attn_pdrop=0.1,
                  layer_norm_epsilon=1e-5,
                  initializer_range=0.02,
-
-                 num_labels=1,
                  summary_type='cls_index',
                  summary_use_proj=True,
                  summary_activation=None,
@@ -84,7 +82,7 @@ class XxxConfig(PretrainedConfig):
                  summary_first_dropout=0.1,
                  **kwargs):
         super(XxxConfig, self).__init__(**kwargs)
-        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, six.string_types) else -1
+        self.vocab_size = vocab_size
         self.n_ctx = n_ctx
         self.n_positions = n_positions
         self.n_embd = n_embd
@@ -95,23 +93,11 @@ class XxxConfig(PretrainedConfig):
         self.attn_pdrop = attn_pdrop
         self.layer_norm_epsilon = layer_norm_epsilon
         self.initializer_range = initializer_range
-
-        self.num_labels = num_labels
         self.summary_type = summary_type
         self.summary_use_proj = summary_use_proj
         self.summary_activation = summary_activation
         self.summary_first_dropout = summary_first_dropout
         self.summary_proj_to_labels = summary_proj_to_labels
-        if isinstance(vocab_size_or_config_json_file, six.string_types):
-            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif not isinstance(vocab_size_or_config_json_file, int):
-            raise ValueError(
-                "First argument must be either a vocabulary size (int)"
-                "or the path to a pretrained model config file (str)"
-            )
 
     @property
     def max_position_embeddings(self):

templates/adding_a_new_model/convert_xxx_original_tf_checkpoint_to_pytorch.py

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

templates/adding_a_new_model/modeling_tf_xxx.py

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

templates/adding_a_new_model/modeling_xxx.py

@@ -25,6 +25,8 @@ import logging
 import math
 import os
 import sys
+import copy
+import itertools
 from io import open
 
 import torch
@@ -362,7 +364,7 @@ class XxxForMaskedLM(XxxPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -413,7 +415,7 @@ class XxxForMaskedLM(XxxPreTrainedModel):
 
         outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
         if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             outputs = (masked_lm_loss,) + outputs
 

templates/adding_a_new_model/tests/modeling_tf_xxx_test.py

@@ -17,12 +17,11 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
-import pytest
 import sys
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import CACHE_DIR, require_tf, slow
 
 from transformers import XxxConfig, is_tf_available
 
@@ -33,10 +32,9 @@ if is_tf_available():
                                                TFXxxForTokenClassification,
                                                TFXxxForQuestionAnswering,
                                                TF_XXX_PRETRAINED_MODEL_ARCHIVE_MAP)
-else:
-    pytestmark = pytest.mark.skip("Require TensorFlow")
 
 
+@require_tf
 class TFXxxModelTest(TFCommonTestCases.TFCommonModelTester):
 
     all_model_classes = (TFXxxModel, TFXxxForMaskedLM, TFXxxForQuestionAnswering,
@@ -112,7 +110,7 @@ class TFXxxModelTest(TFCommonTestCases.TFCommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = XxxConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 hidden_size=self.hidden_size,
                 num_hidden_layers=self.num_hidden_layers,
                 num_attention_heads=self.num_attention_heads,
@@ -244,12 +242,10 @@ class TFXxxModelTest(TFCommonTestCases.TFCommonModelTester):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_xxx_for_token_classification(*config_and_inputs)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in ['xxx-base-uncased']:
-            model = TFXxxModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFXxxModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

templates/adding_a_new_model/tests/modeling_xxx_test.py

@@ -17,13 +17,12 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
-import pytest
 
 from transformers import is_torch_available
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 if is_torch_available():
     from transformers import (XxxConfig, XxxModel, XxxForMaskedLM,
@@ -31,10 +30,9 @@ if is_torch_available():
                                         XxxForQuestionAnswering, XxxForSequenceClassification,
                                         XxxForTokenClassification, XxxForMultipleChoice)
     from transformers.modeling_xxx import XXX_PRETRAINED_MODEL_ARCHIVE_MAP
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 
+@require_torch
 class XxxModelTest(CommonTestCases.CommonModelTester):
 
     all_model_classes = (XxxModel, XxxForMaskedLM, XxxForQuestionAnswering,
@@ -110,7 +108,7 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = XxxConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 hidden_size=self.hidden_size,
                 num_hidden_layers=self.num_hidden_layers,
                 num_attention_heads=self.num_attention_heads,
@@ -131,6 +129,7 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_xxx_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = XxxModel(config=config)
+            model.to(torch_device)
             model.eval()
             sequence_output, pooled_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
             sequence_output, pooled_output = model(input_ids, token_type_ids=token_type_ids)
@@ -148,6 +147,7 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_xxx_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = XxxForMaskedLM(config=config)
+            model.to(torch_device)
             model.eval()
             loss, prediction_scores = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, masked_lm_labels=token_labels)
             result = {
@@ -162,6 +162,7 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_xxx_for_question_answering(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = XxxForQuestionAnswering(config=config)
+            model.to(torch_device)
             model.eval()
             loss, start_logits, end_logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids,
                                                    start_positions=sequence_labels, end_positions=sequence_labels)
@@ -182,6 +183,7 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
         def create_and_check_xxx_for_sequence_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             config.num_labels = self.num_labels
             model = XxxForSequenceClassification(config)
+            model.to(torch_device)
             model.eval()
             loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
             result = {
@@ -197,6 +199,7 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
         def create_and_check_xxx_for_token_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             config.num_labels = self.num_labels
             model = XxxForTokenClassification(config=config)
+            model.to(torch_device)
             model.eval()
             loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
             result = {
@@ -243,12 +246,10 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_xxx_for_token_classification(*config_and_inputs)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(XXX_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = XxxModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = XxxModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 if __name__ == "__main__":

templates/adding_a_new_model/tokenization_xxx.py

@@ -85,7 +85,7 @@ class XxxTokenizer(PreTrainedTokenizer):
 
     Args:
         vocab_file: Path to a one-wordpiece-per-line vocabulary file
-        do_lower_case: Whether to lower case the input. Only has an effect when do_wordpiece_only=False
+        do_lower_case: Whether to lower case the input. Only has an effect when do_basic_tokenize=True
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

transformers-cli

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

transformers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "2.2.1"
+__version__ = "2.3.0"
 
 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -19,22 +19,29 @@ logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
 # Files and general utilities
 from .file_utils import (TRANSFORMERS_CACHE, PYTORCH_TRANSFORMERS_CACHE, PYTORCH_PRETRAINED_BERT_CACHE,
                          cached_path, add_start_docstrings, add_end_docstrings,
-                         WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, CONFIG_NAME,
+                         WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, CONFIG_NAME, MODEL_CARD_NAME,
                          is_tf_available, is_torch_available)
 
 from .data import (is_sklearn_available,
                    InputExample, InputFeatures, DataProcessor,
+                   SingleSentenceClassificationProcessor,
                    glue_output_modes, glue_convert_examples_to_features,
                    glue_processors, glue_tasks_num_labels,
-                   xnli_output_modes, xnli_processors, xnli_tasks_num_labels)
+                   xnli_output_modes, xnli_processors, xnli_tasks_num_labels,
+                   squad_convert_examples_to_features, SquadFeatures, 
+                   SquadExample, SquadV1Processor, SquadV2Processor)
 
 if is_sklearn_available():
     from .data import glue_compute_metrics, xnli_compute_metrics
 
+# Model Cards
+from .modelcard import ModelCard
+
 # Tokenizers
 from .tokenization_utils import (PreTrainedTokenizer)
 from .tokenization_auto import AutoTokenizer
 from .tokenization_bert import BertTokenizer, BasicTokenizer, WordpieceTokenizer
+from .tokenization_bert_japanese import BertJapaneseTokenizer, MecabTokenizer, CharacterTokenizer
 from .tokenization_openai import OpenAIGPTTokenizer
 from .tokenization_transfo_xl import (TransfoXLTokenizer, TransfoXLCorpus)
 from .tokenization_gpt2 import GPT2Tokenizer
@@ -45,37 +52,40 @@ from .tokenization_roberta import RobertaTokenizer
 from .tokenization_distilbert import DistilBertTokenizer
 from .tokenization_albert import AlbertTokenizer
 from .tokenization_camembert import CamembertTokenizer
+from .tokenization_t5 import T5Tokenizer
+from .tokenization_xlm_roberta import XLMRobertaTokenizer
 
 # Configurations
 from .configuration_utils import PretrainedConfig
-from .configuration_auto import AutoConfig
+from .configuration_auto import AutoConfig, ALL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_bert import BertConfig, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_openai import OpenAIGPTConfig, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_transfo_xl import TransfoXLConfig, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_gpt2 import GPT2Config, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_ctrl import CTRLConfig, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_xlnet import XLNetConfig, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP
-from .configuration_ctrl import CTRLConfig, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_xlm import XLMConfig, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_roberta import RobertaConfig, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_distilbert import DistilBertConfig, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_albert import AlbertConfig, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_camembert import CamembertConfig, CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+from .configuration_t5 import T5Config, T5_PRETRAINED_CONFIG_ARCHIVE_MAP
+from .configuration_xlm_roberta import XLMRobertaConfig, XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
 
 # Modeling
 if is_torch_available():
     from .modeling_utils import (PreTrainedModel, prune_layer, Conv1D)
     from .modeling_auto import (AutoModel, AutoModelForSequenceClassification, AutoModelForQuestionAnswering,
-                                AutoModelWithLMHead)
+                                AutoModelWithLMHead, AutoModelForTokenClassification, ALL_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_bert import (BertPreTrainedModel, BertModel, BertForPreTraining,
                                 BertForMaskedLM, BertForNextSentencePrediction,
                                 BertForSequenceClassification, BertForMultipleChoice,
-                                BertForTokenClassification, BertForQuestionAnswering, BertForRelationshipClassification,
+                                BertForTokenClassification, BertForQuestionAnswering,
                                 load_tf_weights_in_bert, BERT_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_openai import (OpenAIGPTPreTrainedModel, OpenAIGPTModel,
-                                OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
-                                load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
+                                  OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
+                                  load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_transfo_xl import (TransfoXLPreTrainedModel, TransfoXLModel, TransfoXLLMHeadModel,
                                     AdaptiveEmbedding,
                                     load_tf_weights_in_transfo_xl, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
@@ -96,8 +106,7 @@ if is_torch_available():
                             XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_roberta import (RobertaForMaskedLM, RobertaModel,
                                 RobertaForSequenceClassification, RobertaForMultipleChoice,
-                                RobertaForRelationshipClassification,
-                                RobertaForTokenClassification,
+                                RobertaForTokenClassification, RobertaForQuestionAnswering,
                                 ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_distilbert import (DistilBertPreTrainedModel, DistilBertForMaskedLM, DistilBertModel,
                                 DistilBertForSequenceClassification, DistilBertForQuestionAnswering,
@@ -108,11 +117,17 @@ if is_torch_available():
                                 CamembertForTokenClassification,
                                 CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_encoder_decoder import PreTrainedEncoderDecoder, Model2Model
+    from .modeling_t5 import (T5PreTrainedModel, T5Model, T5WithLMHeadModel,
+                              load_tf_weights_in_t5,
+                              T5_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_albert import (AlbertPreTrainedModel, AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification,
                                 AlbertForQuestionAnswering,
                                 load_tf_weights_in_albert, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
+    from .modeling_xlm_roberta import (XLMRobertaForMaskedLM, XLMRobertaModel, XLMRobertaForMultipleChoice,
+                                       XLMRobertaForSequenceClassification, XLMRobertaForTokenClassification)
+
     # Optimization
     from .optimization import (AdamW, get_constant_schedule, get_constant_schedule_with_warmup, get_cosine_schedule_with_warmup,
                                get_cosine_with_hard_restarts_schedule_with_warmup, get_linear_schedule_with_warmup)
@@ -122,7 +137,7 @@ if is_torch_available():
 if is_tf_available():
     from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list
     from .modeling_tf_auto import (TFAutoModel, TFAutoModelForSequenceClassification, TFAutoModelForQuestionAnswering,
-                                   TFAutoModelWithLMHead)
+                                   TFAutoModelWithLMHead, TFAutoModelForTokenClassification, TF_ALL_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_bert import (TFBertPreTrainedModel, TFBertMainLayer, TFBertEmbeddings,
                                    TFBertModel, TFBertForPreTraining,
@@ -165,6 +180,7 @@ if is_tf_available():
     from .modeling_tf_distilbert import (TFDistilBertPreTrainedModel, TFDistilBertMainLayer,
                                          TFDistilBertModel, TFDistilBertForMaskedLM,
                                          TFDistilBertForSequenceClassification,
+                                         TFDistilBertForTokenClassification,
                                          TFDistilBertForQuestionAnswering,
                                          TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
@@ -176,6 +192,12 @@ if is_tf_available():
                                      TFAlbertForSequenceClassification,
                                     TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
+    from .modeling_tf_t5 import (TFT5PreTrainedModel, TFT5Model, TFT5WithLMHeadModel,
+                                 TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP)
+
+    # Optimization
+    from .optimization_tf import (WarmUp, create_optimizer, AdamWeightDecay, GradientAccumulator)
+
 # TF 2.0 <=> PyTorch conversion utilities
 from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name,
                                         load_pytorch_checkpoint_in_tf2_model,
@@ -185,6 +207,10 @@ from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name
                                         load_tf2_weights_in_pytorch_model,
                                         load_tf2_model_in_pytorch_model)
 
+# Pipelines
+from .pipelines import pipeline, PipelineDataFormat, CsvPipelineDataFormat, JsonPipelineDataFormat, PipedPipelineDataFormat, \
+    Pipeline, FeatureExtractionPipeline, QuestionAnsweringPipeline, NerPipeline, TextClassificationPipeline
+
 if not is_tf_available() and not is_torch_available():
     logger.warning("Neither PyTorch nor TensorFlow >= 2.0 have been found."
                    "Models won't be available and only tokenizers, configuration"

transformers/__main__.py

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

transformers/commands/convert.py

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

transformers/commands/download.py

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

transformers/commands/run.py

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

transformers/commands/serving.py

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

transformers/commands/train.py

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

transformers/commands/user.py

@@ -19,8 +19,8 @@ class UserCommands(BaseTransformersCLICommand):
         list_parser.set_defaults(func=lambda args: ListObjsCommand(args))
         # upload
         upload_parser = parser.add_parser('upload')
-        upload_parser.add_argument('file', type=str, help='Local filepath of the file to upload.')
-        upload_parser.add_argument('--filename', type=str, default=None, help='Optional: override object filename on S3.')
+        upload_parser.add_argument('path', type=str, help='Local path of the folder or individual file to upload.')
+        upload_parser.add_argument('--filename', type=str, default=None, help='Optional: override individual object filename on S3.')
         upload_parser.set_defaults(func=lambda args: UploadCommand(args))
 
 
@@ -138,28 +138,57 @@ class ListObjsCommand(BaseUserCommand):
 
 
 class UploadCommand(BaseUserCommand):
+    def walk_dir(self, rel_path):
+        """
+        Recursively list all files in a folder.
+        """
+        entries: List[os.DirEntry] = list(os.scandir(rel_path))
+        files = [
+            (
+                os.path.join(os.getcwd(), f.path),  # filepath
+                f.path  # filename
+            )
+            for f in entries if f.is_file()
+        ]
+        for f in entries:
+            if f.is_dir():
+                files += self.walk_dir(f.path)
+        return files
+
     def run(self):
         token = HfFolder.get_token()
         if token is None:
             print("Not logged in")
             exit(1)
-        filepath = os.path.join(os.getcwd(), self.args.file)
-        filename = self.args.filename if self.args.filename is not None else os.path.basename(filepath)
-        print(
-            "About to upload file {} to S3 under filename {}".format(
-                ANSI.bold(filepath), ANSI.bold(filename)
+        local_path = os.path.abspath(self.args.path)
+        if os.path.isdir(local_path):
+            if self.args.filename is not None:
+                raise ValueError("Cannot specify a filename override when uploading a folder.")
+            rel_path = os.path.basename(local_path)
+            files = self.walk_dir(rel_path)
+        elif os.path.isfile(local_path):
+            filename = self.args.filename if self.args.filename is not None else os.path.basename(local_path)
+            files = [(local_path, filename)]
+        else:
+            raise ValueError("Not a valid file or directory: {}".format(local_path))
+
+        for filepath, filename in files:
+            print(
+                "About to upload file {} to S3 under filename {}".format(
+                    ANSI.bold(filepath), ANSI.bold(filename)
+                )
             )
-        )
 
         choice = input("Proceed? [Y/n] ").lower()
         if not(choice == "" or choice == "y" or choice == "yes"):
             print("Abort")
             exit()
         print(
-            ANSI.bold("Uploading... This might take a while if file is large")
+            ANSI.bold("Uploading... This might take a while if files are large")
         )
-        access_url = self._api.presign_and_upload(
-            token=token, filename=filename, filepath=filepath
-        )
-        print("Your file now lives at:")
-        print(access_url)
+        for filepath, filename in files:
+            access_url = self._api.presign_and_upload(
+                token=token, filename=filename, filepath=filepath
+            )
+            print("Your file now lives at:")
+            print(access_url)

transformers/configuration_albert.py

@@ -37,7 +37,7 @@ class AlbertConfig(PretrainedConfig):
     pretrained_config_archive_map = ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=30000,
+                 vocab_size=30000,
                  embedding_size=128,
                  hidden_size=4096,
                  num_hidden_layers=12,
@@ -83,7 +83,7 @@ class AlbertConfig(PretrainedConfig):
         """
         super(AlbertConfig, self).__init__(**kwargs)
 
-        self.vocab_size = vocab_size_or_config_json_file
+        self.vocab_size = vocab_size
         self.embedding_size = embedding_size
         self.hidden_size = hidden_size
         self.num_hidden_layers = num_hidden_layers
@@ -97,4 +97,4 @@ class AlbertConfig(PretrainedConfig):
         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.layer_norm_eps = layer_norm_eps

transformers/configuration_auto.py

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

transformers/configuration_bert.py

@@ -42,6 +42,12 @@ BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-config.json",
     'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-config.json",
     'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-config.json",
+    'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-config.json",
+    'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-config.json",
+    'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-config.json",
+    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-config.json",
+    'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/config.json",
+    'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/config.json",
 }
 
 
@@ -52,7 +58,7 @@ class BertConfig(PretrainedConfig):
 
 
         Arguments:
-            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `BertModel`.
+            vocab_size: Vocabulary size of `inputs_ids` in `BertModel`.
             hidden_size: Size of the encoder layers and the pooler layer.
             num_hidden_layers: Number of hidden layers in the Transformer encoder.
             num_attention_heads: Number of attention heads for each attention layer in
@@ -77,7 +83,7 @@ class BertConfig(PretrainedConfig):
     pretrained_config_archive_map = BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=30522,
+                 vocab_size=30522,
                  hidden_size=768,
                  num_hidden_layers=12,
                  num_attention_heads=12,
@@ -89,11 +95,9 @@ class BertConfig(PretrainedConfig):
                  type_vocab_size=2,
                  initializer_range=0.02,
                  layer_norm_eps=1e-12,
-                 entity_1_token_id=1001,
-                 entity_2_token_id=1002,
                  **kwargs):
         super(BertConfig, self).__init__(**kwargs)
-        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.vocab_size = vocab_size
         self.hidden_size = hidden_size
         self.num_hidden_layers = num_hidden_layers
         self.num_attention_heads = num_attention_heads
@@ -105,14 +109,3 @@ class BertConfig(PretrainedConfig):
         self.type_vocab_size = type_vocab_size
         self.initializer_range = initializer_range
         self.layer_norm_eps = layer_norm_eps
-        self.entity_2_token_id = entity_2_token_id
-        self.entity_1_token_id = entity_1_token_id
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif not isinstance(vocab_size_or_config_json_file, int):
-            raise ValueError("First argument must be either a vocabulary size (int)"
-                             " or the path to a pretrained model config file (str)")

transformers/configuration_ctrl.py

@@ -31,7 +31,7 @@ class CTRLConfig(PretrainedConfig):
     """Configuration class to store the configuration of a `CTRLModel`.
 
     Args:
-        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
+        vocab_size: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
         n_positions: Number of positional embeddings.
         n_ctx: Size of the causal mask (usually same as n_positions).
         dff: Size of the inner dimension of the FFN.
@@ -52,7 +52,7 @@ class CTRLConfig(PretrainedConfig):
 
     def __init__(
         self,
-        vocab_size_or_config_json_file=246534,
+        vocab_size=246534,
         n_positions=256,
         n_ctx=256,
         n_embd=1280,
@@ -64,8 +64,6 @@ class CTRLConfig(PretrainedConfig):
         attn_pdrop=0.1,
         layer_norm_epsilon=1e-6,
         initializer_range=0.02,
-
-        num_labels=1,
         summary_type='cls_index',
         summary_use_proj=True,
         summary_activation=None,
@@ -76,7 +74,7 @@ class CTRLConfig(PretrainedConfig):
         """Constructs CTRLConfig.
 
         Args:
-            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
+            vocab_size: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
             n_positions: Number of positional embeddings.
             n_ctx: Size of the causal mask (usually same as n_positions).
             dff: Size of the inner dimension of the FFN.
@@ -94,8 +92,7 @@ class CTRLConfig(PretrainedConfig):
                 initializing all weight matrices.
         """
         super(CTRLConfig, self).__init__(**kwargs)
-
-        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.vocab_size = vocab_size
         self.n_ctx = n_ctx
         self.n_positions = n_positions
         self.n_embd = n_embd
@@ -108,23 +105,11 @@ class CTRLConfig(PretrainedConfig):
         self.layer_norm_epsilon = layer_norm_epsilon
         self.initializer_range = initializer_range
 
-        self.num_labels = num_labels
         self.summary_type = summary_type
         self.summary_use_proj = summary_use_proj
         self.summary_activation = summary_activation
         self.summary_first_dropout = summary_first_dropout
         self.summary_proj_to_labels = summary_proj_to_labels
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif not isinstance(vocab_size_or_config_json_file, int):
-            raise ValueError(
-                "First argument must be either a vocabulary size (int)"
-                "or the path to a pretrained model config file (str)"
-            )
 
     @property
     def max_position_embeddings(self):

transformers/configuration_distilbert.py

@@ -37,7 +37,7 @@ class DistilBertConfig(PretrainedConfig):
     pretrained_config_archive_map = DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=30522,
+                 vocab_size=30522,
                  max_position_embeddings=512,
                  sinusoidal_pos_embds=False,
                  n_layers=6,
@@ -53,7 +53,7 @@ class DistilBertConfig(PretrainedConfig):
                  seq_classif_dropout=0.2,
                  **kwargs):
         super(DistilBertConfig, self).__init__(**kwargs)
-        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.vocab_size = vocab_size
         self.max_position_embeddings = max_position_embeddings
         self.sinusoidal_pos_embds = sinusoidal_pos_embds
         self.n_layers = n_layers
@@ -67,15 +67,7 @@ class DistilBertConfig(PretrainedConfig):
         self.tie_weights_ = tie_weights_
         self.qa_dropout = qa_dropout
         self.seq_classif_dropout = seq_classif_dropout
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif not isinstance(vocab_size_or_config_json_file, int):
-            raise ValueError("First argument must be either a vocabulary size (int)"
-                             " or the path to a pretrained model config file (str)")
+
     @property
     def hidden_size(self):
         return self.dim

transformers/configuration_gpt2.py

@@ -36,7 +36,7 @@ class GPT2Config(PretrainedConfig):
     """Configuration class to store the configuration of a `GPT2Model`.
 
     Args:
-        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
+        vocab_size: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
         n_positions: Number of positional embeddings.
         n_ctx: Size of the causal mask (usually same as n_positions).
         n_embd: Dimensionality of the embeddings and hidden states.
@@ -56,7 +56,7 @@ class GPT2Config(PretrainedConfig):
 
     def __init__(
         self,
-        vocab_size_or_config_json_file=50257,
+        vocab_size=50257,
         n_positions=1024,
         n_ctx=1024,
         n_embd=768,
@@ -67,8 +67,6 @@ class GPT2Config(PretrainedConfig):
         attn_pdrop=0.1,
         layer_norm_epsilon=1e-5,
         initializer_range=0.02,
-
-        num_labels=1,
         summary_type='cls_index',
         summary_use_proj=True,
         summary_activation=None,
@@ -79,7 +77,7 @@ class GPT2Config(PretrainedConfig):
         """Constructs GPT2Config.
 
         Args:
-            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
+            vocab_size: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
             n_positions: Number of positional embeddings.
             n_ctx: Size of the causal mask (usually same as n_positions).
             n_embd: Dimensionality of the embeddings and hidden states.
@@ -96,7 +94,7 @@ class GPT2Config(PretrainedConfig):
                 initializing all weight matrices.
         """
         super(GPT2Config, self).__init__(**kwargs)
-        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.vocab_size = vocab_size
         self.n_ctx = n_ctx
         self.n_positions = n_positions
         self.n_embd = n_embd
@@ -107,24 +105,11 @@ class GPT2Config(PretrainedConfig):
         self.attn_pdrop = attn_pdrop
         self.layer_norm_epsilon = layer_norm_epsilon
         self.initializer_range = initializer_range
-
-        self.num_labels = num_labels
         self.summary_type = summary_type
         self.summary_use_proj = summary_use_proj
         self.summary_activation = summary_activation
         self.summary_first_dropout = summary_first_dropout
         self.summary_proj_to_labels = summary_proj_to_labels
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif not isinstance(vocab_size_or_config_json_file, int):
-            raise ValueError(
-                "First argument must be either a vocabulary size (int)"
-                "or the path to a pretrained model config file (str)"
-            )
 
     @property
     def max_position_embeddings(self):

transformers/configuration_openai.py

@@ -35,7 +35,7 @@ class OpenAIGPTConfig(PretrainedConfig):
     Configuration class to store the configuration of a `OpenAIGPTModel`.
 
     Args:
-        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `OpenAIGPTModel` or a configuration json file.
+        vocab_size: Vocabulary size of `inputs_ids` in `OpenAIGPTModel` or a configuration json file.
         n_positions: Number of positional embeddings.
         n_ctx: Size of the causal mask (usually same as n_positions).
         n_embd: Dimensionality of the embeddings and hidden states.
@@ -58,7 +58,7 @@ class OpenAIGPTConfig(PretrainedConfig):
 
     def __init__(
         self,
-        vocab_size_or_config_json_file=40478,
+        vocab_size=40478,
         n_positions=512,
         n_ctx=512,
         n_embd=768,
@@ -71,8 +71,6 @@ class OpenAIGPTConfig(PretrainedConfig):
         layer_norm_epsilon=1e-5,
         initializer_range=0.02,
         predict_special_tokens=True,
-
-        num_labels=1,
         summary_type='cls_index',
         summary_use_proj=True,
         summary_activation=None,
@@ -83,7 +81,7 @@ class OpenAIGPTConfig(PretrainedConfig):
         """Constructs OpenAIGPTConfig.
         """
         super(OpenAIGPTConfig, self).__init__(**kwargs)
-        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.vocab_size = vocab_size
         self.n_ctx = n_ctx
         self.n_positions = n_positions
         self.n_embd = n_embd
@@ -96,24 +94,11 @@ class OpenAIGPTConfig(PretrainedConfig):
         self.layer_norm_epsilon = layer_norm_epsilon
         self.initializer_range = initializer_range
         self.predict_special_tokens = predict_special_tokens
-
-        self.num_labels = num_labels
         self.summary_type = summary_type
         self.summary_use_proj = summary_use_proj
         self.summary_activation = summary_activation
         self.summary_first_dropout = summary_first_dropout
         self.summary_proj_to_labels = summary_proj_to_labels
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif not isinstance(vocab_size_or_config_json_file, int):
-            raise ValueError(
-                "First argument must be either a vocabulary size (int)"
-                "or the path to a pretrained model config file (str)"
-            )
 
     @property
     def max_position_embeddings(self):

transformers/configuration_t5.py

@@ -0,0 +1,108 @@
+# coding=utf-8
+# Copyright 2010, The T5 Authors and HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" T5 model configuration """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import sys
+import six
+from io import open
+
+from .configuration_utils import PretrainedConfig
+
+logger = logging.getLogger(__name__)
+
+T5_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-small-config.json",
+    't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-base-config.json",
+    't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-large-config.json",
+    't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-3b-config.json",
+    't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-11b-config.json",
+}
+
+
+class T5Config(PretrainedConfig):
+    r"""
+        :class:`~transformers.T5Config` is the configuration class to store the configuration of a
+        `T5Model`.
+
+
+        Arguments:
+            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `T5Model`.
+            hidden_size: Size of the encoder layers and the pooler layer.
+            num_hidden_layers: Number of hidden layers in the Transformer encoder.
+            num_attention_heads: Number of attention heads for each attention layer in
+                the Transformer encoder.
+            intermediate_size: The size of the "intermediate" (i.e., feed-forward)
+                layer in the Transformer encoder.
+            hidden_act: The non-linear activation function (function or string) in the
+                encoder and pooler. If string, "gelu", "relu", "swish" and "gelu_new" are supported.
+            hidden_dropout_prob: The dropout probabilitiy for all fully connected
+                layers in the embeddings, encoder, and pooler.
+            attention_probs_dropout_prob: The dropout ratio for the attention
+                probabilities.
+            max_position_embeddings: The maximum sequence length that this model might
+                ever be used with. Typically set this to something large just in case
+                (e.g., 512 or 1024 or 2048).
+            type_vocab_size: The vocabulary size of the `token_type_ids` passed into
+                `T5Model`.
+            initializer_factor: A factor for initializing all weight matrices (should be kept to 1.0, used for initialization testing).
+            layer_norm_eps: The epsilon used by LayerNorm.
+    """
+    pretrained_config_archive_map = T5_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+    def __init__(self,
+                 vocab_size=32128,
+                 n_positions=512,
+                 d_model=512,
+                 d_kv=64,
+                 d_ff=2048,
+                 num_layers=6,
+                 num_heads=8,
+                 relative_attention_num_buckets=32,
+                 dropout_rate=0.1,
+                 layer_norm_epsilon=1e-6,
+                 initializer_factor=1.0,
+                 **kwargs):
+        super(T5Config, self).__init__(**kwargs)
+        self.vocab_size = vocab_size
+        self.n_positions = n_positions
+        self.d_model = d_model
+        self.d_kv = d_kv
+        self.d_ff = d_ff
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.relative_attention_num_buckets = relative_attention_num_buckets
+        self.dropout_rate = dropout_rate
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_factor = initializer_factor
+
+    @property
+    def max_position_embeddings(self):
+        return self.n_positions
+
+    @property
+    def hidden_size(self):
+        return self.d_model
+
+    @property
+    def num_attention_heads(self):
+        return self.num_heads
+
+    @property
+    def num_hidden_layers(self):
+        return self.num_layers

transformers/configuration_transfo_xl.py

@@ -34,7 +34,7 @@ class TransfoXLConfig(PretrainedConfig):
     """Configuration class to store the configuration of a `TransfoXLModel`.
 
         Args:
-            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `TransfoXLModel` or a configuration json file.
+            vocab_size: Vocabulary size of `inputs_ids` in `TransfoXLModel` or a configuration json file.
             cutoffs: cutoffs for the adaptive softmax
             d_model: Dimensionality of the model's hidden states.
             d_embed: Dimensionality of the embeddings
@@ -68,7 +68,7 @@ class TransfoXLConfig(PretrainedConfig):
     pretrained_config_archive_map = TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=267735,
+                 vocab_size=267735,
                  cutoffs=[20000, 40000, 200000],
                  d_model=1024,
                  d_embed=1024,
@@ -100,7 +100,7 @@ class TransfoXLConfig(PretrainedConfig):
         """Constructs TransfoXLConfig.
         """
         super(TransfoXLConfig, self).__init__(**kwargs)
-        self.n_token = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.vocab_size = vocab_size
         self.cutoffs = []
         self.cutoffs.extend(cutoffs)
         self.tie_weight = tie_weight
@@ -133,27 +133,17 @@ class TransfoXLConfig(PretrainedConfig):
         self.init_std = init_std
         self.layer_norm_epsilon = layer_norm_epsilon
 
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif not isinstance(vocab_size_or_config_json_file, int):
-            raise ValueError("First argument must be either a vocabulary size (int)"
-                             " or the path to a pretrained model config file (str)")
-
     @property
     def max_position_embeddings(self):
         return self.tgt_len + self.ext_len + self.mem_len
 
     @property
-    def vocab_size(self):
-        return self.n_token
+    def n_token(self):  # Backward compatibility
+        return self.vocab_size
 
-    @vocab_size.setter
-    def vocab_size(self, value):
-        self.n_token = value
+    @n_token.setter
+    def n_token(self, value):  # Backward compatibility
+        self.vocab_size = value
 
     @property
     def hidden_size(self):

transformers/configuration_utils.py

@@ -24,7 +24,7 @@ import logging
 import os
 from io import open
 
-from .file_utils import cached_path, CONFIG_NAME
+from .file_utils import CONFIG_NAME, cached_path, is_remote_url, hf_bucket_url
 
 logger = logging.getLogger(__name__)
 
@@ -49,16 +49,47 @@ class PretrainedConfig(object):
     pretrained_config_archive_map = {}
 
     def __init__(self, **kwargs):
-        self.finetuning_task = kwargs.pop('finetuning_task', None)
-        self.num_labels = kwargs.pop('num_labels', 2)
+        # Attributes with defaults
         self.output_attentions = kwargs.pop('output_attentions', False)
         self.output_hidden_states = kwargs.pop('output_hidden_states', False)
         self.output_past = kwargs.pop('output_past', True)  # Not used by all models
         self.torchscript = kwargs.pop('torchscript', False)  # Only used by PyTorch models
         self.use_bfloat16 = kwargs.pop('use_bfloat16', False)
         self.pruned_heads = kwargs.pop('pruned_heads', {})
+
+        # Is decoder is used in encoder-decoder models to differentiate encoder from decoder
         self.is_decoder = kwargs.pop('is_decoder', False)
 
+        # Parameters for sequence generation
+        self.max_length = kwargs.pop('max_length', 20)
+        self.do_sample = kwargs.pop('do_sample', False)
+        self.num_beams = kwargs.pop('num_beams', 1)
+        self.temperature = kwargs.pop('temperature', 1.0)
+        self.top_k = kwargs.pop('top_k', 50)
+        self.top_p = kwargs.pop('top_p', 1.0)
+        self.repetition_penalty = kwargs.pop('repetition_penalty', 1.0)
+        self.bos_token_id = kwargs.pop('bos_token_id', 0)
+        self.pad_token_id = kwargs.pop('pad_token_id', 0)
+        self.eos_token_ids = kwargs.pop('eos_token_ids', 0)
+        self.length_penalty = kwargs.pop('length_penalty', 1.)
+        self.num_return_sequences = kwargs.pop('num_return_sequences', 1)
+
+        # Fine-tuning task arguments
+        self.finetuning_task = kwargs.pop('finetuning_task', None)
+        self.num_labels = kwargs.pop('num_labels', 2)
+        self.id2label = kwargs.pop('id2label', {i: 'LABEL_{}'.format(i) for i in range(self.num_labels)})
+        self.id2label = dict((int(key), value) for key, value in self.id2label.items())
+        self.label2id = kwargs.pop('label2id', dict(zip(self.id2label.values(), self.id2label.keys())))
+        self.label2id = dict((key, int(value)) for key, value in self.label2id.items())
+
+        # Additional attributes without default values
+        for key, value in kwargs.items():
+            try:
+                setattr(self, key, value)
+            except AttributeError as err:
+                logger.error("Can't set {} with value {} for {}".format(key, value, self))
+                raise err
+
     def save_pretrained(self, save_directory):
         """ Save a configuration object to the directory `save_directory`, so that it
             can be re-loaded using the :func:`~transformers.PretrainedConfig.from_pretrained` class method.
@@ -79,6 +110,7 @@ class PretrainedConfig(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model configuration to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model configuration that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing a configuration file saved using the :func:`~transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
                 - a path or url to a saved configuration JSON `file`, e.g.: ``./my_model_directory/configuration.json``.
 
@@ -131,12 +163,18 @@ class PretrainedConfig(object):
             config_file = cls.pretrained_config_archive_map[pretrained_model_name_or_path]
         elif os.path.isdir(pretrained_model_name_or_path):
             config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
-        else:
+        elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
             config_file = pretrained_model_name_or_path
-        # redirect to the cache, if necessary
+        else:
+            config_file = hf_bucket_url(pretrained_model_name_or_path, postfix=CONFIG_NAME)
+
         try:
+            # Load from URL or cache if already cached
             resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download,
                                                proxies=proxies, resume_download=resume_download)
+            # Load config
+            config = cls.from_json_file(resolved_config_file)
+
         except EnvironmentError:
             if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
                 msg = "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
@@ -150,15 +188,18 @@ class PretrainedConfig(object):
                         config_file, CONFIG_NAME)
             raise EnvironmentError(msg)
 
+        except json.JSONDecodeError:
+            msg = "Couldn't reach server at '{}' to download configuration file or " \
+                  "configuration file is not a valid JSON file. " \
+                  "Please check network or file content here: {}.".format(config_file, resolved_config_file)
+            raise EnvironmentError(msg)
+
         if resolved_config_file == config_file:
             logger.info("loading configuration file {}".format(config_file))
         else:
             logger.info("loading configuration file {} from cache at {}".format(
                 config_file, resolved_config_file))
 
-        # Load config
-        config = cls.from_json_file(resolved_config_file)
-
         if hasattr(config, 'pruned_heads'):
             config.pruned_heads = dict((int(key), value) for key, value in config.pruned_heads.items())
 
@@ -180,17 +221,15 @@ class PretrainedConfig(object):
     @classmethod
     def from_dict(cls, json_object):
         """Constructs a `Config` from a Python dictionary of parameters."""
-        config = cls(vocab_size_or_config_json_file=-1)
-        for key, value in json_object.items():
-            setattr(config, key, value)
-        return config
+        return cls(**json_object)
 
     @classmethod
     def from_json_file(cls, json_file):
-        """Constructs a `BertConfig` from a json file of parameters."""
+        """Constructs a `Config` from a json file of parameters."""
         with open(json_file, "r", encoding='utf-8') as reader:
             text = reader.read()
-        return cls.from_dict(json.loads(text))
+        dict_obj = json.loads(text)
+        return cls(**dict_obj)
 
     def __eq__(self, other):
         return self.__dict__ == other.__dict__

transformers/configuration_xlm.py

@@ -42,7 +42,7 @@ class XLMConfig(PretrainedConfig):
     """Configuration class to store the configuration of a `XLMModel`.
 
     Args:
-        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `XLMModel`.
+        vocab_size: Vocabulary size of `inputs_ids` in `XLMModel`.
         d_model: Size of the encoder layers and the pooler layer.
         n_layer: Number of hidden layers in the Transformer encoder.
         n_head: Number of attention heads for each attention layer in
@@ -81,7 +81,7 @@ class XLMConfig(PretrainedConfig):
     pretrained_config_archive_map = XLM_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=30145,
+                 vocab_size=30145,
                  emb_dim=2048,
                  n_layers=12,
                  n_heads=16,
@@ -103,9 +103,6 @@ class XLMConfig(PretrainedConfig):
                  unk_index=3,
                  mask_index=5,
                  is_encoder=True,
-
-                 finetuning_task=None,
-                 num_labels=2,
                  summary_type='first',
                  summary_use_proj=True,
                  summary_activation=None,
@@ -113,11 +110,13 @@ class XLMConfig(PretrainedConfig):
                  summary_first_dropout=0.1,
                  start_n_top=5,
                  end_n_top=5,
+                 mask_token_id=0,
+                 lang_id=0,
                  **kwargs):
         """Constructs XLMConfig.
         """
         super(XLMConfig, self).__init__(**kwargs)
-        self.n_words = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.vocab_size = vocab_size
         self.emb_dim = emb_dim
         self.n_layers = n_layers
         self.n_heads = n_heads
@@ -139,8 +138,6 @@ class XLMConfig(PretrainedConfig):
         self.max_position_embeddings = max_position_embeddings
         self.embed_init_std = embed_init_std
         self.init_std = init_std
-        self.finetuning_task = finetuning_task
-        self.num_labels = num_labels
         self.summary_type = summary_type
         self.summary_use_proj = summary_use_proj
         self.summary_activation = summary_activation
@@ -148,23 +145,19 @@ class XLMConfig(PretrainedConfig):
         self.summary_first_dropout = summary_first_dropout
         self.start_n_top = start_n_top
         self.end_n_top = end_n_top
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif not isinstance(vocab_size_or_config_json_file, int):
-            raise ValueError("First argument must be either a vocabulary size (int)"
-                             " or the path to a pretrained model config file (str)")
+        self.mask_token_id = mask_token_id
+        self.lang_id = lang_id
+
+        if "n_words" in kwargs:
+            self.n_words = kwargs["n_words"]
 
     @property
-    def vocab_size(self):
-        return self.n_words
+    def n_words(self):  # For backward compatibility
+        return self.vocab_size
 
-    @vocab_size.setter
-    def vocab_size(self, value):
-        self.n_words = value
+    @n_words.setter
+    def n_words(self, value):  # For backward compatibility
+        self.vocab_size = value
 
     @property
     def hidden_size(self):

transformers/configuration_xlm_roberta.py

@@ -0,0 +1,38 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" XLM-RoBERTa configuration """
+
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import logging
+
+from .configuration_roberta import RobertaConfig
+
+logger = logging.getLogger(__name__)
+
+XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    'xlm-roberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-base-config.json",
+    'xlm-roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-config.json",
+    'xlm-roberta-large-finetuned-conll02-dutch': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll02-dutch-config.json",
+    'xlm-roberta-large-finetuned-conll02-spanish': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll02-spanish-config.json",
+    'xlm-roberta-large-finetuned-conll03-english': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll03-english-config.json",
+    'xlm-roberta-large-finetuned-conll03-german': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll03-german-config.json",
+}
+
+
+class XLMRobertaConfig(RobertaConfig):
+    pretrained_config_archive_map = XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP

transformers/configuration_xlnet.py

@@ -35,7 +35,7 @@ class XLNetConfig(PretrainedConfig):
     """Configuration class to store the configuration of a ``XLNetModel``.
 
     Args:
-        vocab_size_or_config_json_file: Vocabulary size of ``inputs_ids`` in ``XLNetModel``.
+        vocab_size: Vocabulary size of ``inputs_ids`` in ``XLNetModel``.
         d_model: Size of the encoder layers and the pooler layer.
         n_layer: Number of hidden layers in the Transformer encoder.
         n_head: Number of attention heads for each attention layer in
@@ -72,7 +72,7 @@ class XLNetConfig(PretrainedConfig):
     pretrained_config_archive_map = XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=32000,
+                 vocab_size=32000,
                  d_model=1024,
                  n_layer=24,
                  n_head=16,
@@ -80,19 +80,14 @@ class XLNetConfig(PretrainedConfig):
                  ff_activation="gelu",
                  untie_r=True,
                  attn_type="bi",
-
                  initializer_range=0.02,
                  layer_norm_eps=1e-12,
-
                  dropout=0.1,
                  mem_len=None,
                  reuse_len=None,
                  bi_data=False,
                  clamp_len=-1,
                  same_length=False,
-
-                 finetuning_task=None,
-                 num_labels=2,
                  summary_type='last',
                  summary_use_proj=True,
                  summary_activation='tanh',
@@ -103,7 +98,7 @@ class XLNetConfig(PretrainedConfig):
         """Constructs XLNetConfig.
         """
         super(XLNetConfig, self).__init__(**kwargs)
-        self.n_token = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.vocab_size = vocab_size
         self.d_model = d_model
         self.n_layer = n_layer
         self.n_head = n_head
@@ -124,35 +119,24 @@ class XLNetConfig(PretrainedConfig):
         self.clamp_len = clamp_len
         self.same_length = same_length
 
-        self.finetuning_task = finetuning_task
-        self.num_labels = num_labels
         self.summary_type = summary_type
         self.summary_use_proj = summary_use_proj
         self.summary_activation = summary_activation
         self.summary_last_dropout = summary_last_dropout
         self.start_n_top = start_n_top
         self.end_n_top = end_n_top
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif not isinstance(vocab_size_or_config_json_file, int):
-            raise ValueError("First argument must be either a vocabulary size (int)"
-                             " or the path to a pretrained model config file (str)")
 
     @property
     def max_position_embeddings(self):
         return -1
 
     @property
-    def vocab_size(self):
-        return self.n_token
+    def n_token(self):  # Backward compatibility
+        return self.vocab_size
 
-    @vocab_size.setter
-    def vocab_size(self, value):
-        self.n_token = value
+    @n_token.setter
+    def n_token(self, value):  # Backward compatibility
+        self.vocab_size = value
 
     @property
     def hidden_size(self):

transformers/convert_pytorch_checkpoint_to_tf2.py

@@ -32,9 +32,10 @@ from transformers import (load_pytorch_checkpoint_in_tf2_model,
                                   TransfoXLConfig, TFTransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
                                   OpenAIGPTConfig, TFOpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
                                   RobertaConfig, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  DistilBertConfig, TFDistilBertForMaskedLM, TFDistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  DistilBertConfig, TFDistilBertForMaskedLM, TFDistilBertForQuestionAnswering, TFDistilBertForSequenceClassification, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
                                   CTRLConfig, TFCTRLLMHeadModel, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  AlbertConfig, TFAlbertForMaskedLM, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
+                                  AlbertConfig, TFAlbertForMaskedLM, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  T5Config, TFT5WithLMHeadModel, T5_PRETRAINED_CONFIG_ARCHIVE_MAP)
 
 if is_torch_available():
     import torch
@@ -46,9 +47,10 @@ if is_torch_available():
                                       TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
                                       OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                       RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
-                                      DistilBertForMaskedLM, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+                                      DistilBertForMaskedLM, DistilBertForQuestionAnswering, DistilBertForSequenceClassification, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                       CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
-                                      AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
+                                      AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+                                      T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP)
 else:
     (BertForPreTraining, BertForQuestionAnswering, BertForSequenceClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
@@ -57,9 +59,10 @@ else:
     TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
     OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
     RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
-    DistilBertForMaskedLM, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+    DistilBertForMaskedLM, DistilBertForSequenceClassification, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
-    AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP) = (
+    AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+    T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP) = (
         None, None, None, None,
         None, None,
         None, None,
@@ -67,7 +70,8 @@ else:
         None, None,
         None, None,
         None, None, None,
-        None, None, None,
+        None, None, None, None,
+        None, None,
         None, None,
         None, None)
 
@@ -89,8 +93,10 @@ MODEL_CLASSES = {
     'roberta-large-mnli': (RobertaConfig, TFRobertaForSequenceClassification, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP),
     'distilbert': (DistilBertConfig, TFDistilBertForMaskedLM, DistilBertForMaskedLM, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
     'distilbert-base-uncased-distilled-squad': (DistilBertConfig, TFDistilBertForQuestionAnswering, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'distilbert-base-uncased-distilled-squad': (DistilBertConfig, TFDistilBertForQuestionAnswering, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
     'ctrl': (CTRLConfig, TFCTRLLMHeadModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'albert': (AlbertConfig, TFAlbertForMaskedLM, AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
+    'albert': (AlbertConfig, TFAlbertForMaskedLM, AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    't5': (T5Config, TFT5WithLMHeadModel, T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP, T5_PRETRAINED_CONFIG_ARCHIVE_MAP),
 }
 
 def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file, tf_dump_path, compare_with_pt_model=False, use_cached_models=True):
@@ -115,23 +121,21 @@ def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file
     tf_model = load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path)
 
     if compare_with_pt_model:
-        inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-        tf_inputs = tf.constant(inputs_list)
-        tfo = tf_model(tf_inputs, training=False)  # build the network
+        tfo = tf_model(tf_model.dummy_inputs, training=False)  # build the network
 
-        pt_model = pt_model_class.from_pretrained(None,
+        state_dict = torch.load(pytorch_checkpoint_path, map_location='cpu')
+        pt_model = pt_model_class.from_pretrained(pretrained_model_name_or_path=None,
                                                   config=config,
-                                                  state_dict=torch.load(pytorch_checkpoint_path,
-                                                                        map_location='cpu'))
-        pt_inputs = torch.tensor(inputs_list)
-        with torch.no_grad():
-            pto = pt_model(pt_inputs)
+                                                  state_dict=state_dict)
 
-        np_pt = pto[0].detach().numpy()
+        with torch.no_grad():
+            pto = pt_model(**pt_model.dummy_inputs)
+
+        np_pt = pto[0].numpy()
         np_tf = tfo[0].numpy()
         diff = np.amax(np.abs(np_pt - np_tf))
         print("Max absolute difference between models outputs {}".format(diff))
-        assert diff <= 2e-2, "Error, model absolute difference is >2e-2"
+        assert diff <= 2e-2, "Error, model absolute difference is >2e-2: {}".format(diff)
 
     # Save pytorch-model
     print("Save TensorFlow model to {}".format(tf_dump_path))
@@ -139,7 +143,7 @@ def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file
 
 
 def convert_all_pt_checkpoints_to_tf(args_model_type, tf_dump_path, model_shortcut_names_or_path=None, config_shortcut_names_or_path=None,
-                                     compare_with_pt_model=False, use_cached_models=False, only_convert_finetuned_models=False):
+                                     compare_with_pt_model=False, use_cached_models=False, remove_cached_files=False, only_convert_finetuned_models=False):
     assert os.path.isdir(args.tf_dump_path), "--tf_dump_path should be a directory"
 
     if args_model_type is None:
@@ -187,13 +191,15 @@ def convert_all_pt_checkpoints_to_tf(args_model_type, tf_dump_path, model_shortc
 
             if os.path.isfile(model_shortcut_name):
                 model_shortcut_name = 'converted_model'
+
             convert_pt_checkpoint_to_tf(model_type=model_type,
                                         pytorch_checkpoint_path=model_file,
                                         config_file=config_file,
                                         tf_dump_path=os.path.join(tf_dump_path, model_shortcut_name + '-tf_model.h5'),
                                         compare_with_pt_model=compare_with_pt_model)
-            os.remove(config_file)
-            os.remove(model_file)
+            if remove_cached_files:
+                os.remove(config_file)
+                os.remove(model_file)
 
 
 if __name__ == "__main__":
@@ -226,6 +232,9 @@ if __name__ == "__main__":
     parser.add_argument("--use_cached_models",
                         action='store_true',
                         help = "Use cached models if possible instead of updating to latest checkpoint versions.")
+    parser.add_argument("--remove_cached_files",
+                        action='store_true',
+                        help = "Remove pytorch models after conversion (save memory when converting in batches).")
     parser.add_argument("--only_convert_finetuned_models",
                         action='store_true',
                         help = "Only convert finetuned models.")
@@ -245,4 +254,5 @@ if __name__ == "__main__":
                                         config_shortcut_names_or_path=[args.config_file] if args.config_file is not None else None,
                                         compare_with_pt_model=args.compare_with_pt_model,
                                         use_cached_models=args.use_cached_models,
+                                        remove_cached_files=args.remove_cached_files,
                                         only_convert_finetuned_models=args.only_convert_finetuned_models)

transformers/convert_roberta_original_pytorch_checkpoint_to_pytorch.py

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

transformers/convert_t5_original_tf_checkpoint_to_pytorch.py

@@ -0,0 +1,65 @@
+# coding=utf-8
+# Copyright 2018 The T5 authors and HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert T5 checkpoint."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import torch
+
+from transformers import T5Config, T5Model, load_tf_weights_in_t5
+
+import logging
+logging.basicConfig(level=logging.INFO)
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file, pytorch_dump_path):
+    # Initialise PyTorch model
+    config = T5Config.from_json_file(config_file)
+    print("Building PyTorch model from configuration: {}".format(str(config)))
+    model = T5Model(config)
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_t5(model, config, tf_checkpoint_path)
+
+    # Save pytorch-model
+    print("Save PyTorch model to {}".format(pytorch_dump_path))
+    torch.save(model.state_dict(), pytorch_dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    ## Required parameters
+    parser.add_argument("--tf_checkpoint_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the TensorFlow checkpoint path.")
+    parser.add_argument("--config_file",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "The config json file corresponding to the pre-trained T5 model. \n"
+                            "This specifies the model architecture.")
+    parser.add_argument("--pytorch_dump_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the output PyTorch model.")
+    args = parser.parse_args()
+    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path,
+                                     args.config_file,
+                                     args.pytorch_dump_path)

transformers/data/__init__.py

@@ -1,5 +1,6 @@
-from .processors import InputExample, InputFeatures, DataProcessor
+from .processors import InputExample, InputFeatures, DataProcessor, SquadFeatures, SingleSentenceClassificationProcessor
 from .processors import glue_output_modes, glue_processors, glue_tasks_num_labels, glue_convert_examples_to_features
+from .processors import squad_convert_examples_to_features, SquadExample, SquadV1Processor, SquadV2Processor
 from .processors import xnli_output_modes, xnli_processors, xnli_tasks_num_labels
 
 from .metrics import is_sklearn_available

transformers/data/metrics/squad_metrics.py

@@ -0,0 +1,767 @@
+""" Very heavily inspired by the official evaluation script for SQuAD version 2.0 which was
+modified by XLNet authors to update `find_best_threshold` scripts for SQuAD V2.0
+
+In addition to basic functionality, we also compute additional statistics and
+plot precision-recall curves if an additional na_prob.json file is provided.
+This file is expected to map question ID's to the model's predicted probability
+that a question is unanswerable.
+"""
+
+
+import json
+import logging
+import math
+import collections
+from io import open
+from tqdm import tqdm
+import string
+import re
+
+from transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize
+
+logger = logging.getLogger(__name__)
+
+
+def normalize_answer(s):
+    """Lower text and remove punctuation, articles and extra whitespace."""
+    def remove_articles(text):
+        regex = re.compile(r'\b(a|an|the)\b', re.UNICODE)
+        return re.sub(regex, ' ', text)
+
+    def white_space_fix(text):
+        return ' '.join(text.split())
+
+    def remove_punc(text):
+        exclude = set(string.punctuation)
+        return ''.join(ch for ch in text if ch not in exclude)
+
+    def lower(text):
+        return text.lower()
+    return white_space_fix(remove_articles(remove_punc(lower(s))))
+
+
+def get_tokens(s):
+    if not s:
+        return []
+    return normalize_answer(s).split()
+
+
+def compute_exact(a_gold, a_pred):
+    return int(normalize_answer(a_gold) == normalize_answer(a_pred))
+
+
+def compute_f1(a_gold, a_pred):
+    gold_toks = get_tokens(a_gold)
+    pred_toks = get_tokens(a_pred)
+    common = collections.Counter(gold_toks) & collections.Counter(pred_toks)
+    num_same = sum(common.values())
+    if len(gold_toks) == 0 or len(pred_toks) == 0:
+        # If either is no-answer, then F1 is 1 if they agree, 0 otherwise
+        return int(gold_toks == pred_toks)
+    if num_same == 0:
+        return 0
+    precision = 1.0 * num_same / len(pred_toks)
+    recall = 1.0 * num_same / len(gold_toks)
+    f1 = (2 * precision * recall) / (precision + recall)
+    return f1
+
+
+def get_raw_scores(examples, preds):
+    """
+    Computes the exact and f1 scores from the examples and the model predictions
+    """
+    exact_scores = {}
+    f1_scores = {}
+
+    for example in examples:
+        qas_id = example.qas_id
+        gold_answers = [answer['text'] for answer in example.answers if normalize_answer(answer['text'])]
+
+        if not gold_answers:
+            # For unanswerable questions, only correct answer is empty string
+            gold_answers = ['']
+
+        if qas_id not in preds:
+            print('Missing prediction for %s' % qas_id)
+            continue
+
+        prediction = preds[qas_id]
+        exact_scores[qas_id] = max(compute_exact(a, prediction) for a in gold_answers)
+        f1_scores[qas_id] = max(compute_f1(a, prediction) for a in gold_answers)
+
+    return exact_scores, f1_scores
+
+
+def apply_no_ans_threshold(scores, na_probs, qid_to_has_ans, na_prob_thresh):
+    new_scores = {}
+    for qid, s in scores.items():
+        pred_na = na_probs[qid] > na_prob_thresh
+        if pred_na:
+            new_scores[qid] = float(not qid_to_has_ans[qid])
+        else:
+            new_scores[qid] = s
+    return new_scores
+
+
+def make_eval_dict(exact_scores, f1_scores, qid_list=None):
+    if not qid_list:
+        total = len(exact_scores)
+        return collections.OrderedDict([
+            ('exact', 100.0 * sum(exact_scores.values()) / total),
+            ('f1', 100.0 * sum(f1_scores.values()) / total),
+            ('total', total),
+        ])
+    else:
+        total = len(qid_list)
+        return collections.OrderedDict([
+            ('exact', 100.0 * sum(exact_scores[k] for k in qid_list) / total),
+            ('f1', 100.0 * sum(f1_scores[k] for k in qid_list) / total),
+            ('total', total),
+        ])
+
+
+def merge_eval(main_eval, new_eval, prefix):
+    for k in new_eval:
+        main_eval['%s_%s' % (prefix, k)] = new_eval[k]
+
+
+def find_best_thresh_v2(preds, scores, na_probs, qid_to_has_ans):
+    num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
+    cur_score = num_no_ans
+    best_score = cur_score
+    best_thresh = 0.0
+    qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+    for i, qid in enumerate(qid_list):
+        if qid not in scores:
+            continue
+        if qid_to_has_ans[qid]:
+            diff = scores[qid]
+        else:
+            if preds[qid]:
+                diff = -1
+            else:
+                diff = 0
+        cur_score += diff
+        if cur_score > best_score:
+            best_score = cur_score
+            best_thresh = na_probs[qid]
+
+    has_ans_score, has_ans_cnt = 0, 0
+    for qid in qid_list:
+        if not qid_to_has_ans[qid]:
+            continue
+        has_ans_cnt += 1
+
+        if qid not in scores:
+            continue
+        has_ans_score += scores[qid]
+
+    return 100.0 * best_score / len(scores), best_thresh, 1.0 * has_ans_score / has_ans_cnt
+
+
+def find_all_best_thresh_v2(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
+    best_exact, exact_thresh, has_ans_exact = find_best_thresh_v2(
+        preds, exact_raw, na_probs, qid_to_has_ans)
+    best_f1, f1_thresh, has_ans_f1 = find_best_thresh_v2(
+        preds, f1_raw, na_probs, qid_to_has_ans)
+    main_eval['best_exact'] = best_exact
+    main_eval['best_exact_thresh'] = exact_thresh
+    main_eval['best_f1'] = best_f1
+    main_eval['best_f1_thresh'] = f1_thresh
+    main_eval['has_ans_exact'] = has_ans_exact
+    main_eval['has_ans_f1'] = has_ans_f1
+
+
+def find_best_thresh(preds, scores, na_probs, qid_to_has_ans):
+    num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
+    cur_score = num_no_ans
+    best_score = cur_score
+    best_thresh = 0.0
+    qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+    for _, qid in enumerate(qid_list):
+        if qid not in scores:
+            continue
+        if qid_to_has_ans[qid]:
+            diff = scores[qid]
+        else:
+            if preds[qid]:
+                diff = -1
+            else:
+                diff = 0
+        cur_score += diff
+        if cur_score > best_score:
+            best_score = cur_score
+            best_thresh = na_probs[qid]
+    return 100.0 * best_score / len(scores), best_thresh
+
+
+def find_all_best_thresh(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
+    best_exact, exact_thresh = find_best_thresh(preds, exact_raw, na_probs, qid_to_has_ans)
+    best_f1, f1_thresh = find_best_thresh(preds, f1_raw, na_probs, qid_to_has_ans)
+
+    main_eval['best_exact'] = best_exact
+    main_eval['best_exact_thresh'] = exact_thresh
+    main_eval['best_f1'] = best_f1
+    main_eval['best_f1_thresh'] = f1_thresh
+
+
+def squad_evaluate(examples, preds, no_answer_probs=None, no_answer_probability_threshold=1.0):
+    qas_id_to_has_answer = {example.qas_id: bool(example.answers) for example in examples}
+    has_answer_qids = [qas_id for qas_id, has_answer in qas_id_to_has_answer.items() if has_answer]
+    no_answer_qids = [qas_id for qas_id, has_answer in qas_id_to_has_answer.items() if not has_answer]
+
+    if no_answer_probs is None:
+        no_answer_probs = {k: 0.0 for k in preds}
+
+    exact, f1 = get_raw_scores(examples, preds)
+
+    exact_threshold = apply_no_ans_threshold(exact, no_answer_probs, qas_id_to_has_answer, no_answer_probability_threshold)
+    f1_threshold = apply_no_ans_threshold(f1, no_answer_probs, qas_id_to_has_answer, no_answer_probability_threshold)
+
+    evaluation = make_eval_dict(exact_threshold, f1_threshold)
+
+    if has_answer_qids:
+        has_ans_eval = make_eval_dict(exact_threshold, f1_threshold, qid_list=has_answer_qids)
+        merge_eval(evaluation, has_ans_eval, 'HasAns')
+
+    if no_answer_qids:
+        no_ans_eval = make_eval_dict(exact_threshold, f1_threshold, qid_list=no_answer_qids)
+        merge_eval(evaluation, no_ans_eval, 'NoAns')
+
+    if no_answer_probs:
+        find_all_best_thresh(evaluation, preds, exact, f1, no_answer_probs, qas_id_to_has_answer)
+
+    return evaluation
+
+
+def get_final_text(pred_text, orig_text, do_lower_case, verbose_logging=False):
+    """Project the tokenized prediction back to the original text."""
+
+    # When we created the data, we kept track of the alignment between original
+    # (whitespace tokenized) tokens and our WordPiece tokenized tokens. So
+    # now `orig_text` contains the span of our original text corresponding to the
+    # span that we predicted.
+    #
+    # However, `orig_text` may contain extra characters that we don't want in
+    # our prediction.
+    #
+    # For example, let's say:
+    #   pred_text = steve smith
+    #   orig_text = Steve Smith's
+    #
+    # We don't want to return `orig_text` because it contains the extra "'s".
+    #
+    # We don't want to return `pred_text` because it's already been normalized
+    # (the SQuAD eval script also does punctuation stripping/lower casing but
+    # our tokenizer does additional normalization like stripping accent
+    # characters).
+    #
+    # What we really want to return is "Steve Smith".
+    #
+    # Therefore, we have to apply a semi-complicated alignment heuristic between
+    # `pred_text` and `orig_text` to get a character-to-character alignment. This
+    # can fail in certain cases in which case we just return `orig_text`.
+
+    def _strip_spaces(text):
+        ns_chars = []
+        ns_to_s_map = collections.OrderedDict()
+        for (i, c) in enumerate(text):
+            if c == " ":
+                continue
+            ns_to_s_map[len(ns_chars)] = i
+            ns_chars.append(c)
+        ns_text = "".join(ns_chars)
+        return (ns_text, ns_to_s_map)
+
+    # We first tokenize `orig_text`, strip whitespace from the result
+    # and `pred_text`, and check if they are the same length. If they are
+    # NOT the same length, the heuristic has failed. If they are the same
+    # length, we assume the characters are one-to-one aligned.
+    tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
+
+    tok_text = " ".join(tokenizer.tokenize(orig_text))
+
+    start_position = tok_text.find(pred_text)
+    if start_position == -1:
+        if verbose_logging:
+            logger.info(
+                "Unable to find text: '%s' in '%s'" % (pred_text, orig_text))
+        return orig_text
+    end_position = start_position + len(pred_text) - 1
+
+    (orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text)
+    (tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text)
+
+    if len(orig_ns_text) != len(tok_ns_text):
+        if verbose_logging:
+            logger.info("Length not equal after stripping spaces: '%s' vs '%s'",
+                        orig_ns_text, tok_ns_text)
+        return orig_text
+
+    # We then project the characters in `pred_text` back to `orig_text` using
+    # the character-to-character alignment.
+    tok_s_to_ns_map = {}
+    for (i, tok_index) in tok_ns_to_s_map.items():
+        tok_s_to_ns_map[tok_index] = i
+
+    orig_start_position = None
+    if start_position in tok_s_to_ns_map:
+        ns_start_position = tok_s_to_ns_map[start_position]
+        if ns_start_position in orig_ns_to_s_map:
+            orig_start_position = orig_ns_to_s_map[ns_start_position]
+
+    if orig_start_position is None:
+        if verbose_logging:
+            logger.info("Couldn't map start position")
+        return orig_text
+
+    orig_end_position = None
+    if end_position in tok_s_to_ns_map:
+        ns_end_position = tok_s_to_ns_map[end_position]
+        if ns_end_position in orig_ns_to_s_map:
+            orig_end_position = orig_ns_to_s_map[ns_end_position]
+
+    if orig_end_position is None:
+        if verbose_logging:
+            logger.info("Couldn't map end position")
+        return orig_text
+
+    output_text = orig_text[orig_start_position:(orig_end_position + 1)]
+    return output_text
+
+
+def _get_best_indexes(logits, n_best_size):
+    """Get the n-best logits from a list."""
+    index_and_score = sorted(enumerate(logits), key=lambda x: x[1], reverse=True)
+
+    best_indexes = []
+    for i in range(len(index_and_score)):
+        if i >= n_best_size:
+            break
+        best_indexes.append(index_and_score[i][0])
+    return best_indexes
+
+
+def _compute_softmax(scores):
+    """Compute softmax probability over raw logits."""
+    if not scores:
+        return []
+
+    max_score = None
+    for score in scores:
+        if max_score is None or score > max_score:
+            max_score = score
+
+    exp_scores = []
+    total_sum = 0.0
+    for score in scores:
+        x = math.exp(score - max_score)
+        exp_scores.append(x)
+        total_sum += x
+
+    probs = []
+    for score in exp_scores:
+        probs.append(score / total_sum)
+    return probs
+
+
+def compute_predictions_logits(
+    all_examples,
+    all_features,
+    all_results,
+    n_best_size,
+    max_answer_length,
+    do_lower_case,
+    output_prediction_file,
+    output_nbest_file,
+    output_null_log_odds_file,
+    verbose_logging,
+    version_2_with_negative,
+    null_score_diff_threshold,
+    tokenizer,
+):
+    """Write final predictions to the json file and log-odds of null if needed."""
+    logger.info("Writing predictions to: %s" % (output_prediction_file))
+    logger.info("Writing nbest to: %s" % (output_nbest_file))
+
+    example_index_to_features = collections.defaultdict(list)
+    for feature in all_features:
+        example_index_to_features[feature.example_index].append(feature)
+
+    unique_id_to_result = {}
+    for result in all_results:
+        unique_id_to_result[result.unique_id] = result
+
+    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "PrelimPrediction",
+        ["feature_index", "start_index", "end_index", "start_logit", "end_logit"])
+
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict()
+
+    for (example_index, example) in enumerate(all_examples):
+        features = example_index_to_features[example_index]
+
+        prelim_predictions = []
+        # keep track of the minimum score of null start+end of position 0
+        score_null = 1000000  # large and positive
+        min_null_feature_index = 0  # the paragraph slice with min null score
+        null_start_logit = 0  # the start logit at the slice with min null score
+        null_end_logit = 0  # the end logit at the slice with min null score
+        for (feature_index, feature) in enumerate(features):
+            result = unique_id_to_result[feature.unique_id]
+            start_indexes = _get_best_indexes(result.start_logits, n_best_size)
+            end_indexes = _get_best_indexes(result.end_logits, n_best_size)
+            # if we could have irrelevant answers, get the min score of irrelevant
+            if version_2_with_negative:
+                feature_null_score = result.start_logits[0] + result.end_logits[0]
+                if feature_null_score < score_null:
+                    score_null = feature_null_score
+                    min_null_feature_index = feature_index
+                    null_start_logit = result.start_logits[0]
+                    null_end_logit = result.end_logits[0]
+            for start_index in start_indexes:
+                for end_index in end_indexes:
+                    # We could hypothetically create invalid predictions, e.g., predict
+                    # that the start of the span is in the question. We throw out all
+                    # invalid predictions.
+                    if start_index >= len(feature.tokens):
+                        continue
+                    if end_index >= len(feature.tokens):
+                        continue
+                    if start_index not in feature.token_to_orig_map:
+                        continue
+                    if end_index not in feature.token_to_orig_map:
+                        continue
+                    if not feature.token_is_max_context.get(start_index, False):
+                        continue
+                    if end_index < start_index:
+                        continue
+                    length = end_index - start_index + 1
+                    if length > max_answer_length:
+                        continue
+                    prelim_predictions.append(
+                        _PrelimPrediction(
+                            feature_index=feature_index,
+                            start_index=start_index,
+                            end_index=end_index,
+                            start_logit=result.start_logits[start_index],
+                            end_logit=result.end_logits[end_index]))
+        if version_2_with_negative:
+            prelim_predictions.append(
+                _PrelimPrediction(
+                    feature_index=min_null_feature_index,
+                    start_index=0,
+                    end_index=0,
+                    start_logit=null_start_logit,
+                    end_logit=null_end_logit))
+        prelim_predictions = sorted(
+            prelim_predictions,
+            key=lambda x: (x.start_logit + x.end_logit),
+            reverse=True)
+
+        _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+            "NbestPrediction", ["text", "start_logit", "end_logit"])
+
+        seen_predictions = {}
+        nbest = []
+        for pred in prelim_predictions:
+            if len(nbest) >= n_best_size:
+                break
+            feature = features[pred.feature_index]
+            if pred.start_index > 0:  # this is a non-null prediction
+                tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
+                orig_doc_start = feature.token_to_orig_map[pred.start_index]
+                orig_doc_end = feature.token_to_orig_map[pred.end_index]
+                orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
+
+                tok_text = tokenizer.convert_tokens_to_string(tok_tokens)
+
+                # tok_text = " ".join(tok_tokens)
+                #
+                # # De-tokenize WordPieces that have been split off.
+                # tok_text = tok_text.replace(" ##", "")
+                # tok_text = tok_text.replace("##", "")
+
+                # Clean whitespace
+                tok_text = tok_text.strip()
+                tok_text = " ".join(tok_text.split())
+                orig_text = " ".join(orig_tokens)
+
+                final_text = get_final_text(tok_text, orig_text, do_lower_case, verbose_logging)
+                if final_text in seen_predictions:
+                    continue
+
+                seen_predictions[final_text] = True
+            else:
+                final_text = ""
+                seen_predictions[final_text] = True
+
+            nbest.append(
+                _NbestPrediction(
+                    text=final_text,
+                    start_logit=pred.start_logit,
+                    end_logit=pred.end_logit))
+        # if we didn't include the empty option in the n-best, include it
+        if version_2_with_negative:
+            if "" not in seen_predictions:
+                nbest.append(
+                    _NbestPrediction(
+                        text="",
+                        start_logit=null_start_logit,
+                        end_logit=null_end_logit))
+
+            # In very rare edge cases we could only have single null prediction.
+            # So we just create a nonce prediction in this case to avoid failure.
+            if len(nbest) == 1:
+                nbest.insert(0,
+                             _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
+
+        # In very rare edge cases we could have no valid predictions. So we
+        # just create a nonce prediction in this case to avoid failure.
+        if not nbest:
+            nbest.append(
+                _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
+
+        assert len(nbest) >= 1
+
+        total_scores = []
+        best_non_null_entry = None
+        for entry in nbest:
+            total_scores.append(entry.start_logit + entry.end_logit)
+            if not best_non_null_entry:
+                if entry.text:
+                    best_non_null_entry = entry
+
+        probs = _compute_softmax(total_scores)
+
+        nbest_json = []
+        for (i, entry) in enumerate(nbest):
+            output = collections.OrderedDict()
+            output["text"] = entry.text
+            output["probability"] = probs[i]
+            output["start_logit"] = entry.start_logit
+            output["end_logit"] = entry.end_logit
+            nbest_json.append(output)
+
+        assert len(nbest_json) >= 1
+
+        if not version_2_with_negative:
+            all_predictions[example.qas_id] = nbest_json[0]["text"]
+        else:
+            # predict "" iff the null score - the score of best non-null > threshold
+            score_diff = score_null - best_non_null_entry.start_logit - (
+                best_non_null_entry.end_logit)
+            scores_diff_json[example.qas_id] = score_diff
+            if score_diff > null_score_diff_threshold:
+                all_predictions[example.qas_id] = ""
+            else:
+                all_predictions[example.qas_id] = best_non_null_entry.text
+        all_nbest_json[example.qas_id] = nbest_json
+
+    with open(output_prediction_file, "w") as writer:
+        writer.write(json.dumps(all_predictions, indent=4) + "\n")
+
+    with open(output_nbest_file, "w") as writer:
+        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+
+    if version_2_with_negative:
+        with open(output_null_log_odds_file, "w") as writer:
+            writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    return all_predictions
+
+
+def compute_predictions_log_probs(
+    all_examples,
+    all_features,
+    all_results,
+    n_best_size,
+    max_answer_length,
+    output_prediction_file,
+    output_nbest_file,
+    output_null_log_odds_file,
+    start_n_top,
+    end_n_top,
+    version_2_with_negative,
+    tokenizer,
+    verbose_logging
+):
+    """ XLNet write prediction logic (more complex than Bert's).
+        Write final predictions to the json file and log-odds of null if needed.
+
+        Requires utils_squad_evaluate.py
+    """
+    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "PrelimPrediction",
+        ["feature_index", "start_index", "end_index",
+         "start_log_prob", "end_log_prob"])
+
+    _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "NbestPrediction", ["text", "start_log_prob", "end_log_prob"])
+
+    logger.info("Writing predictions to: %s", output_prediction_file)
+    # logger.info("Writing nbest to: %s" % (output_nbest_file))
+
+    example_index_to_features = collections.defaultdict(list)
+    for feature in all_features:
+        example_index_to_features[feature.example_index].append(feature)
+
+    unique_id_to_result = {}
+    for result in all_results:
+        unique_id_to_result[result.unique_id] = result
+
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict()
+
+    for (example_index, example) in enumerate(all_examples):
+        features = example_index_to_features[example_index]
+
+        prelim_predictions = []
+        # keep track of the minimum score of null start+end of position 0
+        score_null = 1000000  # large and positive
+
+        for (feature_index, feature) in enumerate(features):
+            result = unique_id_to_result[feature.unique_id]
+
+            cur_null_score = result.cls_logits
+
+            # if we could have irrelevant answers, get the min score of irrelevant
+            score_null = min(score_null, cur_null_score)
+
+            for i in range(start_n_top):
+                for j in range(end_n_top):
+                    start_log_prob = result.start_logits[i]
+                    start_index = result.start_top_index[i]
+
+                    j_index = i * end_n_top + j
+
+                    end_log_prob = result.end_logits[j_index]
+                    end_index = result.end_top_index[j_index]
+
+                    # We could hypothetically create invalid predictions, e.g., predict
+                    # that the start of the span is in the question. We throw out all
+                    # invalid predictions.
+                    if start_index >= feature.paragraph_len - 1:
+                        continue
+                    if end_index >= feature.paragraph_len - 1:
+                        continue
+
+                    if not feature.token_is_max_context.get(start_index, False):
+                        continue
+                    if end_index < start_index:
+                        continue
+                    length = end_index - start_index + 1
+                    if length > max_answer_length:
+                        continue
+
+                    prelim_predictions.append(
+                        _PrelimPrediction(
+                            feature_index=feature_index,
+                            start_index=start_index,
+                            end_index=end_index,
+                            start_log_prob=start_log_prob,
+                            end_log_prob=end_log_prob))
+
+        prelim_predictions = sorted(
+            prelim_predictions,
+            key=lambda x: (x.start_log_prob + x.end_log_prob),
+            reverse=True)
+
+        seen_predictions = {}
+        nbest = []
+        for pred in prelim_predictions:
+            if len(nbest) >= n_best_size:
+                break
+            feature = features[pred.feature_index]
+
+            # XLNet un-tokenizer
+            # Let's keep it simple for now and see if we need all this later.
+            #
+            # tok_start_to_orig_index = feature.tok_start_to_orig_index
+            # tok_end_to_orig_index = feature.tok_end_to_orig_index
+            # start_orig_pos = tok_start_to_orig_index[pred.start_index]
+            # end_orig_pos = tok_end_to_orig_index[pred.end_index]
+            # paragraph_text = example.paragraph_text
+            # final_text = paragraph_text[start_orig_pos: end_orig_pos + 1].strip()
+
+            # Previously used Bert untokenizer
+            tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
+            orig_doc_start = feature.token_to_orig_map[pred.start_index]
+            orig_doc_end = feature.token_to_orig_map[pred.end_index]
+            orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
+            tok_text = tokenizer.convert_tokens_to_string(tok_tokens)
+
+            # Clean whitespace
+            tok_text = tok_text.strip()
+            tok_text = " ".join(tok_text.split())
+            orig_text = " ".join(orig_tokens)
+
+            if hasattr(tokenizer, "do_lower_case"):
+                do_lower_case = tokenizer.do_lower_case
+            else:
+                do_lower_case = tokenizer.do_lowercase_and_remove_accent
+
+            final_text = get_final_text(tok_text, orig_text, do_lower_case,
+                                        verbose_logging)
+
+            if final_text in seen_predictions:
+                continue
+
+            seen_predictions[final_text] = True
+
+            nbest.append(
+                _NbestPrediction(
+                    text=final_text,
+                    start_log_prob=pred.start_log_prob,
+                    end_log_prob=pred.end_log_prob))
+
+        # In very rare edge cases we could have no valid predictions. So we
+        # just create a nonce prediction in this case to avoid failure.
+        if not nbest:
+            nbest.append(
+                _NbestPrediction(text="", start_log_prob=-1e6,
+                end_log_prob=-1e6))
+
+        total_scores = []
+        best_non_null_entry = None
+        for entry in nbest:
+            total_scores.append(entry.start_log_prob + entry.end_log_prob)
+            if not best_non_null_entry:
+                best_non_null_entry = entry
+
+        probs = _compute_softmax(total_scores)
+
+        nbest_json = []
+        for (i, entry) in enumerate(nbest):
+            output = collections.OrderedDict()
+            output["text"] = entry.text
+            output["probability"] = probs[i]
+            output["start_log_prob"] = entry.start_log_prob
+            output["end_log_prob"] = entry.end_log_prob
+            nbest_json.append(output)
+
+        assert len(nbest_json) >= 1
+        assert best_non_null_entry is not None
+
+        score_diff = score_null
+        scores_diff_json[example.qas_id] = score_diff
+        # note(zhiliny): always predict best_non_null_entry
+        # and the evaluation script will search for the best threshold
+        all_predictions[example.qas_id] = best_non_null_entry.text
+
+        all_nbest_json[example.qas_id] = nbest_json
+
+    with open(output_prediction_file, "w") as writer:
+        writer.write(json.dumps(all_predictions, indent=4) + "\n")
+
+    with open(output_nbest_file, "w") as writer:
+        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+
+    if version_2_with_negative:
+        with open(output_null_log_odds_file, "w") as writer:
+            writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    return all_predictions

transformers/data/processors/__init__.py

@@ -1,3 +1,4 @@
-from .utils import InputExample, InputFeatures, DataProcessor
+from .utils import InputExample, InputFeatures, DataProcessor, SingleSentenceClassificationProcessor
 from .glue import glue_output_modes, glue_processors, glue_tasks_num_labels, glue_convert_examples_to_features
-from .xnli import xnli_output_modes, xnli_processors, xnli_tasks_num_labels
+from .squad import squad_convert_examples_to_features, SquadFeatures, SquadExample, SquadV1Processor, SquadV2Processor
+from .xnli import xnli_output_modes, xnli_processors, xnli_tasks_num_labels

transformers/data/processors/glue.py

@@ -133,7 +133,7 @@ def glue_convert_examples_to_features(examples, tokenizer,
     if is_tf_available() and is_tf_dataset:
         def gen():
             for ex in features:
-                yield  ({'input_ids': ex.input_ids,
+                yield ({'input_ids': ex.input_ids,
                          'attention_mask': ex.attention_mask,
                          'token_type_ids': ex.token_type_ids},
                         ex.label)

transformers/data/processors/squad.py

@@ -0,0 +1,674 @@
+from tqdm import tqdm
+import collections
+import logging
+import os
+import json
+import numpy as np
+from multiprocessing import Pool
+from multiprocessing import cpu_count
+from functools import partial
+
+from ...tokenization_bert import BasicTokenizer, whitespace_tokenize
+from .utils import DataProcessor, InputExample, InputFeatures
+from ...file_utils import is_tf_available, is_torch_available
+
+if is_torch_available():
+    import torch
+    from torch.utils.data import TensorDataset
+
+if is_tf_available():
+    import tensorflow as tf
+
+logger = logging.getLogger(__name__)
+
+
+def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer, orig_answer_text):
+    """Returns tokenized answer spans that better match the annotated answer."""
+    tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text))
+
+    for new_start in range(input_start, input_end + 1):
+        for new_end in range(input_end, new_start - 1, -1):
+            text_span = " ".join(doc_tokens[new_start : (new_end + 1)])
+            if text_span == tok_answer_text:
+                return (new_start, new_end)
+
+    return (input_start, input_end)
+
+
+def _check_is_max_context(doc_spans, cur_span_index, position):
+    """Check if this is the 'max context' doc span for the token."""
+    best_score = None
+    best_span_index = None
+    for (span_index, doc_span) in enumerate(doc_spans):
+        end = doc_span.start + doc_span.length - 1
+        if position < doc_span.start:
+            continue
+        if position > end:
+            continue
+        num_left_context = position - doc_span.start
+        num_right_context = end - position
+        score = min(num_left_context, num_right_context) + 0.01 * doc_span.length
+        if best_score is None or score > best_score:
+            best_score = score
+            best_span_index = span_index
+
+    return cur_span_index == best_span_index
+
+
+def _new_check_is_max_context(doc_spans, cur_span_index, position):
+    """Check if this is the 'max context' doc span for the token."""
+    # if len(doc_spans) == 1:
+    # return True
+    best_score = None
+    best_span_index = None
+    for (span_index, doc_span) in enumerate(doc_spans):
+        end = doc_span["start"] + doc_span["length"] - 1
+        if position < doc_span["start"]:
+            continue
+        if position > end:
+            continue
+        num_left_context = position - doc_span["start"]
+        num_right_context = end - position
+        score = min(num_left_context, num_right_context) + 0.01 * doc_span["length"]
+        if best_score is None or score > best_score:
+            best_score = score
+            best_span_index = span_index
+
+    return cur_span_index == best_span_index
+
+
+def _is_whitespace(c):
+    if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F:
+        return True
+    return False
+
+def squad_convert_example_to_features(example, max_seq_length,
+                                       doc_stride, max_query_length, is_training):
+    features = []
+    if is_training and not example.is_impossible:
+        # Get start and end position
+        start_position = example.start_position
+        end_position = example.end_position
+
+        # If the answer cannot be found in the text, then skip this example.
+        actual_text = " ".join(example.doc_tokens[start_position:(end_position + 1)])
+        cleaned_answer_text = " ".join(whitespace_tokenize(example.answer_text))
+        if actual_text.find(cleaned_answer_text) == -1:
+            logger.warning("Could not find answer: '%s' vs. '%s'", actual_text, cleaned_answer_text)
+            return []
+
+    tok_to_orig_index = []
+    orig_to_tok_index = []
+    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)
+        for sub_token in sub_tokens:
+            tok_to_orig_index.append(i)
+            all_doc_tokens.append(sub_token)
+
+    if is_training and not example.is_impossible:
+        tok_start_position = orig_to_tok_index[example.start_position]
+        if example.end_position < len(example.doc_tokens) - 1:
+            tok_end_position = orig_to_tok_index[example.end_position + 1] - 1
+        else:
+            tok_end_position = len(all_doc_tokens) - 1
+
+        (tok_start_position, tok_end_position) = _improve_answer_span(
+            all_doc_tokens, tok_start_position, tok_end_position, tokenizer, example.answer_text
+        )
+
+    spans = []
+
+    truncated_query = tokenizer.encode(example.question_text, add_special_tokens=False, max_length=max_query_length)
+    sequence_added_tokens = tokenizer.max_len - tokenizer.max_len_single_sentence + 1 \
+        if 'roberta' in str(type(tokenizer)) else tokenizer.max_len - tokenizer.max_len_single_sentence
+    sequence_pair_added_tokens = tokenizer.max_len - tokenizer.max_len_sentences_pair
+
+    span_doc_tokens = all_doc_tokens
+    while len(spans) * doc_stride < len(all_doc_tokens):
+
+        encoded_dict = tokenizer.encode_plus(
+            truncated_query if tokenizer.padding_side == "right" else span_doc_tokens,
+            span_doc_tokens if tokenizer.padding_side == "right" else truncated_query,
+            max_length=max_seq_length,
+            return_overflowing_tokens=True,
+            pad_to_max_length=True,
+            stride=max_seq_length - doc_stride - len(truncated_query) - sequence_pair_added_tokens,
+            truncation_strategy='only_second' if tokenizer.padding_side == "right" else 'only_first'
+        )
+
+        paragraph_len = min(len(all_doc_tokens) - len(spans) * doc_stride,
+                            max_seq_length - len(truncated_query) - sequence_pair_added_tokens)
+
+        if tokenizer.pad_token_id in encoded_dict['input_ids']:
+            non_padded_ids = encoded_dict['input_ids'][:encoded_dict['input_ids'].index(tokenizer.pad_token_id)]
+        else:
+            non_padded_ids = encoded_dict['input_ids']
+
+        tokens = tokenizer.convert_ids_to_tokens(non_padded_ids)
+
+        token_to_orig_map = {}
+        for i in range(paragraph_len):
+            index = len(truncated_query) + sequence_added_tokens + i if tokenizer.padding_side == "right" else i
+            token_to_orig_map[index] = tok_to_orig_index[len(spans) * doc_stride + i]
+
+        encoded_dict["paragraph_len"] = paragraph_len
+        encoded_dict["tokens"] = tokens
+        encoded_dict["token_to_orig_map"] = token_to_orig_map
+        encoded_dict["truncated_query_with_special_tokens_length"] = len(truncated_query) + sequence_added_tokens
+        encoded_dict["token_is_max_context"] = {}
+        encoded_dict["start"] = len(spans) * doc_stride
+        encoded_dict["length"] = paragraph_len
+
+        spans.append(encoded_dict)
+
+        if "overflowing_tokens" not in encoded_dict:
+            break
+        span_doc_tokens = encoded_dict["overflowing_tokens"]
+
+    for doc_span_index in range(len(spans)):
+        for j in range(spans[doc_span_index]["paragraph_len"]):
+            is_max_context = _new_check_is_max_context(spans, doc_span_index, doc_span_index * doc_stride + j)
+            index = j if tokenizer.padding_side == "left" else spans[doc_span_index][
+                                                                   "truncated_query_with_special_tokens_length"] + j
+            spans[doc_span_index]["token_is_max_context"][index] = is_max_context
+
+    for span in spans:
+        # Identify the position of the CLS token
+        cls_index = span['input_ids'].index(tokenizer.cls_token_id)
+
+        # p_mask: mask with 1 for token than cannot be in the answer (0 for token which can be in an answer)
+        # Original TF implem also keep the classification token (set to 0) (not sure why...)
+        p_mask = np.array(span['token_type_ids'])
+
+        p_mask = np.minimum(p_mask, 1)
+
+        if tokenizer.padding_side == "right":
+            # Limit positive values to one
+            p_mask = 1 - p_mask
+
+        p_mask[np.where(np.array(span["input_ids"]) == tokenizer.sep_token_id)[0]] = 1
+
+        # Set the CLS index to '0'
+        p_mask[cls_index] = 0
+
+        span_is_impossible = example.is_impossible
+        start_position = 0
+        end_position = 0
+        if is_training and not span_is_impossible:
+            # For training, if our document chunk does not contain an annotation
+            # we throw it out, since there is nothing to predict.
+            doc_start = span["start"]
+            doc_end = span["start"] + span["length"] - 1
+            out_of_span = False
+
+            if not (tok_start_position >= doc_start and tok_end_position <= doc_end):
+                out_of_span = True
+
+            if out_of_span:
+                start_position = cls_index
+                end_position = cls_index
+                span_is_impossible = True
+            else:
+                if tokenizer.padding_side == "left":
+                    doc_offset = 0
+                else:
+                    doc_offset = len(truncated_query) + sequence_added_tokens
+
+                start_position = tok_start_position - doc_start + doc_offset
+                end_position = tok_end_position - doc_start + doc_offset
+
+        features.append(SquadFeatures(
+            span['input_ids'],
+            span['attention_mask'],
+            span['token_type_ids'],
+            cls_index,
+            p_mask.tolist(),
+            example_index=0, # Can not set unique_id and example_index here. They will be set after multiple processing.
+            unique_id=0,
+            paragraph_len=span['paragraph_len'],
+            token_is_max_context=span["token_is_max_context"],
+            tokens=span["tokens"],
+            token_to_orig_map=span["token_to_orig_map"],
+
+            start_position=start_position,
+            end_position=end_position
+        ))
+    return features
+
+def squad_convert_example_to_features_init(tokenizer_for_convert):
+    global tokenizer
+    tokenizer = tokenizer_for_convert
+
+def squad_convert_examples_to_features(examples, tokenizer, max_seq_length,
+                                       doc_stride, max_query_length, is_training, 
+                                       return_dataset=False, threads=1):
+    """
+    Converts a list of examples into a list of features that can be directly given as input to a model.
+    It is model-dependant and takes advantage of many of the tokenizer's features to create the model's inputs.
+
+    Args:
+        examples: list of :class:`~transformers.data.processors.squad.SquadExample`
+        tokenizer: an instance of a child of :class:`~transformers.PreTrainedTokenizer`
+        max_seq_length: The maximum sequence length of the inputs.
+        doc_stride: The stride used when the context is too large and is split across several features.
+        max_query_length: The maximum length of the query.
+        is_training: whether to create features for model evaluation or model training.
+        return_dataset: Default False. Either 'pt' or 'tf'.
+            if 'pt': returns a torch.data.TensorDataset,
+            if 'tf': returns a tf.data.Dataset
+        threads: multiple processing threadsa-smi
+
+
+    Returns:
+        list of :class:`~transformers.data.processors.squad.SquadFeatures`
+
+    Example::
+
+        processor = SquadV2Processor()
+        examples = processor.get_dev_examples(data_dir)
+
+        features = squad_convert_examples_to_features( 
+            examples=examples,
+            tokenizer=tokenizer,
+            max_seq_length=args.max_seq_length,
+            doc_stride=args.doc_stride,
+            max_query_length=args.max_query_length,
+            is_training=not evaluate,
+        )
+    """
+
+    # Defining helper methods    
+    features = []
+    threads = min(threads, cpu_count())
+    with Pool(threads, initializer=squad_convert_example_to_features_init, initargs=(tokenizer,)) as p:
+        annotate_ = partial(squad_convert_example_to_features, max_seq_length=max_seq_length,
+                                       doc_stride=doc_stride, max_query_length=max_query_length, is_training=is_training)
+        features = list(tqdm(p.imap(annotate_, examples, chunksize=32), total=len(examples), desc='convert squad examples to features'))
+    new_features = []
+    unique_id = 1000000000
+    example_index = 0
+    for example_features in tqdm(features, total=len(features), desc='add example index and unique id'):
+        if not example_features:
+            continue
+        for example_feature in example_features:
+            example_feature.example_index = example_index
+            example_feature.unique_id = unique_id
+            new_features.append(example_feature)
+            unique_id += 1
+        example_index += 1
+    features = new_features
+    del new_features
+    if return_dataset == 'pt':
+        if not is_torch_available():
+            raise ImportError("Pytorch must be installed to return a pytorch dataset.")
+
+        # Convert to Tensors and build dataset
+        all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
+        all_attention_masks = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
+        all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
+        all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
+        all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
+
+        if not is_training:
+            all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
+            dataset = TensorDataset(
+                all_input_ids, all_attention_masks, all_token_type_ids, all_example_index, all_cls_index, all_p_mask
+            )
+        else:
+            all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
+            all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
+            dataset = TensorDataset(
+                all_input_ids,
+                all_attention_masks,
+                all_token_type_ids,
+                all_start_positions,
+                all_end_positions,
+                all_cls_index,
+                all_p_mask,
+            )
+
+        return features, dataset
+    elif return_dataset == "tf":
+        if not is_tf_available():
+            raise ImportError("TensorFlow must be installed to return a TensorFlow dataset.")
+
+        def gen():
+            for ex in features:
+                yield (
+                    {
+                        "input_ids": ex.input_ids,
+                        "attention_mask": ex.attention_mask,
+                        "token_type_ids": ex.token_type_ids,
+                    }, {
+                        "start_position": ex.start_position,
+                        "end_position": ex.end_position,
+                        "cls_index": ex.cls_index,
+                        "p_mask": ex.p_mask,
+                    }
+                )
+
+        return tf.data.Dataset.from_generator(
+            gen,
+            (
+                {"input_ids": tf.int32, "attention_mask": tf.int32, "token_type_ids": tf.int32},
+                {"start_position": tf.int64, "end_position": tf.int64, "cls_index": tf.int64, "p_mask": tf.int32},
+            ),
+            (
+                {
+                    "input_ids": tf.TensorShape([None]),
+                    "attention_mask": tf.TensorShape([None]),
+                    "token_type_ids": tf.TensorShape([None]),
+                },
+                {
+                    "start_position": tf.TensorShape([]),
+                    "end_position": tf.TensorShape([]),
+                    "cls_index": tf.TensorShape([]),
+                    "p_mask": tf.TensorShape([None]),
+                },
+            ),
+        )
+
+    return features
+
+
+class SquadProcessor(DataProcessor):
+    """
+    Processor for the SQuAD data set.
+    Overriden by SquadV1Processor and SquadV2Processor, used by the version 1.1 and version 2.0 of SQuAD, respectively.
+    """
+
+    train_file = None
+    dev_file = None
+
+    def _get_example_from_tensor_dict(self, tensor_dict, evaluate=False):
+        if not evaluate:
+            answer = tensor_dict["answers"]["text"][0].numpy().decode("utf-8")
+            answer_start = tensor_dict["answers"]["answer_start"][0].numpy()
+            answers = []
+        else:
+            answers = [
+                {"answer_start": start.numpy(), "text": text.numpy().decode("utf-8")}
+                for start, text in zip(tensor_dict["answers"]["answer_start"], tensor_dict["answers"]["text"])
+            ]
+
+            answer = None
+            answer_start = None
+
+        return SquadExample(
+            qas_id=tensor_dict["id"].numpy().decode("utf-8"),
+            question_text=tensor_dict["question"].numpy().decode("utf-8"),
+            context_text=tensor_dict["context"].numpy().decode("utf-8"),
+            answer_text=answer,
+            start_position_character=answer_start,
+            title=tensor_dict["title"].numpy().decode("utf-8"),
+            answers=answers,
+        )
+
+    def get_examples_from_dataset(self, dataset, evaluate=False):
+        """
+        Creates a list of :class:`~transformers.data.processors.squad.SquadExample` using a TFDS dataset.
+
+        Args:
+            dataset: The tfds dataset loaded from `tensorflow_datasets.load("squad")`
+            evaluate: boolean specifying if in evaluation mode or in training mode
+
+        Returns:
+            List of SquadExample
+
+        Examples::
+
+            import tensorflow_datasets as tfds
+            dataset = tfds.load("squad")
+
+            training_examples = get_examples_from_dataset(dataset, evaluate=False)
+            evaluation_examples = get_examples_from_dataset(dataset, evaluate=True)
+        """
+
+        if evaluate:
+            dataset = dataset["validation"]
+        else:
+            dataset = dataset["train"]
+
+        examples = []
+        for tensor_dict in tqdm(dataset):
+            examples.append(self._get_example_from_tensor_dict(tensor_dict, evaluate=evaluate))
+
+        return examples
+
+    def get_train_examples(self, data_dir, filename=None):
+        """
+        Returns the training examples from the data directory.
+
+        Args:
+            data_dir: Directory containing the data files used for training and evaluating.
+            filename: None by default, specify this if the training file has a different name than the original one
+                which is `train-v1.1.json` and `train-v2.0.json` for squad versions 1.1 and 2.0 respectively.
+
+        """
+        if data_dir is None:
+            data_dir = ""
+
+        if self.train_file is None:
+            raise ValueError("SquadProcessor should be instantiated via SquadV1Processor or SquadV2Processor")
+
+        with open(
+            os.path.join(data_dir, self.train_file if filename is None else filename), "r", encoding="utf-8"
+        ) as reader:
+            input_data = json.load(reader)["data"]
+        return self._create_examples(input_data, "train")
+
+    def get_dev_examples(self, data_dir, filename=None):
+        """
+        Returns the evaluation example from the data directory.
+
+        Args:
+            data_dir: Directory containing the data files used for training and evaluating.
+            filename: None by default, specify this if the evaluation file has a different name than the original one
+                which is `train-v1.1.json` and `train-v2.0.json` for squad versions 1.1 and 2.0 respectively.
+        """
+        if data_dir is None:
+            data_dir = ""
+
+        if self.dev_file is None:
+            raise ValueError("SquadProcessor should be instantiated via SquadV1Processor or SquadV2Processor")
+
+        with open(
+            os.path.join(data_dir, self.dev_file if filename is None else filename), "r", encoding="utf-8"
+        ) as reader:
+            input_data = json.load(reader)["data"]
+        return self._create_examples(input_data, "dev")
+
+    def _create_examples(self, input_data, set_type):
+        is_training = set_type == "train"
+        examples = []
+        for entry in tqdm(input_data):
+            title = entry["title"]
+            for paragraph in entry["paragraphs"]:
+                context_text = paragraph["context"]
+                for qa in paragraph["qas"]:
+                    qas_id = qa["id"]
+                    question_text = qa["question"]
+                    start_position_character = None
+                    answer_text = None
+                    answers = []
+
+                    if "is_impossible" in qa:
+                        is_impossible = qa["is_impossible"]
+                    else:
+                        is_impossible = False
+
+                    if not is_impossible:
+                        if is_training:
+                            answer = qa["answers"][0]
+                            answer_text = answer["text"]
+                            start_position_character = answer["answer_start"]
+                        else:
+                            answers = qa["answers"]
+
+                    example = SquadExample(
+                        qas_id=qas_id,
+                        question_text=question_text,
+                        context_text=context_text,
+                        answer_text=answer_text,
+                        start_position_character=start_position_character,
+                        title=title,
+                        is_impossible=is_impossible,
+                        answers=answers,
+                    )
+
+                    examples.append(example)
+        return examples
+
+
+class SquadV1Processor(SquadProcessor):
+    train_file = "train-v1.1.json"
+    dev_file = "dev-v1.1.json"
+
+
+class SquadV2Processor(SquadProcessor):
+    train_file = "train-v2.0.json"
+    dev_file = "dev-v2.0.json"
+
+
+class SquadExample(object):
+    """
+    A single training/test example for the Squad dataset, as loaded from disk.
+
+    Args:
+        qas_id: The example's unique identifier
+        question_text: The question string
+        context_text: The context string
+        answer_text: The answer string
+        start_position_character: The character position of the start of the answer
+        title: The title of the example
+        answers: None by default, this is used during evaluation. Holds answers as well as their start positions.
+        is_impossible: False by default, set to True if the example has no possible answer.
+    """
+
+    def __init__(
+        self,
+        qas_id,
+        question_text,
+        context_text,
+        answer_text,
+        start_position_character,
+        title,
+        answers=[],
+        is_impossible=False,
+    ):
+        self.qas_id = qas_id
+        self.question_text = question_text
+        self.context_text = context_text
+        self.answer_text = answer_text
+        self.title = title
+        self.is_impossible = is_impossible
+        self.answers = answers
+
+        self.start_position, self.end_position = 0, 0
+
+        doc_tokens = []
+        char_to_word_offset = []
+        prev_is_whitespace = True
+
+        # Split on whitespace so that different tokens may be attributed to their original position.
+        for c in self.context_text:
+            if _is_whitespace(c):
+                prev_is_whitespace = True
+            else:
+                if prev_is_whitespace:
+                    doc_tokens.append(c)
+                else:
+                    doc_tokens[-1] += c
+                prev_is_whitespace = False
+            char_to_word_offset.append(len(doc_tokens) - 1)
+
+        self.doc_tokens = doc_tokens
+        self.char_to_word_offset = char_to_word_offset
+
+        # Start end end positions only has a value during evaluation.
+        if start_position_character is not None and not is_impossible:
+            self.start_position = char_to_word_offset[start_position_character]
+            self.end_position = char_to_word_offset[
+                min(start_position_character + len(answer_text) - 1, len(char_to_word_offset) - 1)
+            ]
+
+
+class SquadFeatures(object):
+    """
+    Single squad example features to be fed to a model.
+    Those features are model-specific and can be crafted from :class:`~transformers.data.processors.squad.SquadExample`
+    using the :method:`~transformers.data.processors.squad.squad_convert_examples_to_features` method.
+
+    Args:
+        input_ids: Indices of input sequence tokens in the vocabulary.
+        attention_mask: Mask to avoid performing attention on padding token indices.
+        token_type_ids: Segment token indices to indicate first and second portions of the inputs.
+        cls_index: the index of the CLS token.
+        p_mask: Mask identifying tokens that can be answers vs. tokens that cannot.
+            Mask with 1 for tokens than cannot be in the answer and 0 for token that can be in an answer
+        example_index: the index of the example
+        unique_id: The unique Feature identifier
+        paragraph_len: The length of the context
+        token_is_max_context: List of booleans identifying which tokens have their maximum context in this feature object.
+            If a token does not have their maximum context in this feature object, it means that another feature object
+            has more information related to that token and should be prioritized over this feature for that token.
+        tokens: list of tokens corresponding to the input ids
+        token_to_orig_map: mapping between the tokens and the original text, needed in order to identify the answer.
+        start_position: start of the answer token index 
+        end_position: end of the answer token index 
+    """
+
+    def __init__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        cls_index,
+        p_mask,
+        example_index,
+        unique_id,
+        paragraph_len,
+        token_is_max_context,
+        tokens,
+        token_to_orig_map,
+        start_position,
+        end_position,
+    ):
+        self.input_ids = input_ids
+        self.attention_mask = attention_mask
+        self.token_type_ids = token_type_ids
+        self.cls_index = cls_index
+        self.p_mask = p_mask
+
+        self.example_index = example_index
+        self.unique_id = unique_id
+        self.paragraph_len = paragraph_len
+        self.token_is_max_context = token_is_max_context
+        self.tokens = tokens
+        self.token_to_orig_map = token_to_orig_map
+
+        self.start_position = start_position
+        self.end_position = end_position
+
+
+class SquadResult(object):
+    """
+    Constructs a SquadResult which can be used to evaluate a model's output on the SQuAD dataset.
+
+    Args:
+        unique_id: The unique identifier corresponding to that example.
+        start_logits: The logits corresponding to the start of the answer
+        end_logits: The logits corresponding to the end of the answer
+    """
+
+    def __init__(self, unique_id, start_logits, end_logits, start_top_index=None, end_top_index=None, cls_logits=None):
+        self.start_logits = start_logits
+        self.end_logits = end_logits
+        self.unique_id = unique_id
+
+        if start_top_index:
+            self.start_top_index = start_top_index
+            self.end_top_index = end_top_index
+            self.cls_logits = cls_logits

transformers/data/processors/utils.py

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

transformers/file_utils.py

@@ -10,10 +10,9 @@ import json
 import logging
 import os
 import six
-import shutil
 import tempfile
 import fnmatch
-from functools import wraps
+from functools import partial, wraps
 from hashlib import sha256
 from io import open
 
@@ -21,26 +20,38 @@ import boto3
 from botocore.config import Config
 from botocore.exceptions import ClientError
 import requests
-from tqdm import tqdm
+from tqdm.auto import tqdm
 from contextlib import contextmanager
+from . import __version__
+
+from filelock import FileLock
 
 logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
 
 try:
-    import tensorflow as tf
-    assert hasattr(tf, '__version__') and int(tf.__version__[0]) >= 2
-    _tf_available = True  # pylint: disable=invalid-name
-    logger.info("TensorFlow version {} available.".format(tf.__version__))
-except (ImportError, AssertionError):
-    _tf_available = False  # pylint: disable=invalid-name
-
-try:
-    import torch
-    _torch_available = True  # pylint: disable=invalid-name
-    logger.info("PyTorch version {} available.".format(torch.__version__))
+    os.environ.setdefault('USE_TORCH', 'YES')
+    if os.environ['USE_TORCH'].upper() in ('1', 'ON', 'YES'):
+        import torch
+        _torch_available = True  # pylint: disable=invalid-name
+        logger.info("PyTorch version {} available.".format(torch.__version__))
+    else:
+        logger.info("USE_TORCH override through env variable, disabling PyTorch")
+        _torch_available = False
 except ImportError:
     _torch_available = False  # pylint: disable=invalid-name
 
+try:
+    os.environ.setdefault('USE_TF', 'YES')
+    if os.environ['USE_TF'].upper() in ('1', 'ON', 'YES'):
+        import tensorflow as tf
+        assert hasattr(tf, '__version__') and int(tf.__version__[0]) >= 2
+        _tf_available = True  # pylint: disable=invalid-name
+        logger.info("TensorFlow version {} available.".format(tf.__version__))
+    else:
+        logger.info("USE_TF override through env variable, disabling Tensorflow")
+        _tf_available = False
+except (ImportError, AssertionError):
+    _tf_available = False  # pylint: disable=invalid-name
 
 try:
     from torch.hub import _get_torch_home
@@ -72,11 +83,20 @@ WEIGHTS_NAME = "pytorch_model.bin"
 TF2_WEIGHTS_NAME = 'tf_model.h5'
 TF_WEIGHTS_NAME = 'model.ckpt'
 CONFIG_NAME = "config.json"
+MODEL_CARD_NAME = "modelcard.json"
+
+DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
+DUMMY_MASK = [[1, 1, 1, 1, 1], [1, 1, 1, 0, 0], [0, 0, 0, 1, 1]]
+
+S3_BUCKET_PREFIX = "https://s3.amazonaws.com/models.huggingface.co/bert"
+CLOUDFRONT_DISTRIB_PREFIX = "https://d2ws9o8vfrpkyk.cloudfront.net"
+
 
 def is_torch_available():
     return _torch_available
 
 def is_tf_available():
+
     return _tf_available
 
 if not six.PY2:
@@ -103,12 +123,25 @@ else:
             return fn
         return docstring_decorator
 
+
+def is_remote_url(url_or_filename):
+    parsed = urlparse(url_or_filename)
+    return parsed.scheme in ('http', 'https', 's3')
+
+def hf_bucket_url(identifier, postfix=None, cdn=False):
+    endpoint = CLOUDFRONT_DISTRIB_PREFIX if cdn else S3_BUCKET_PREFIX
+    if postfix is None:
+        return "/".join((endpoint, identifier))
+    else:
+        return "/".join((endpoint, identifier, postfix))
+
+
 def url_to_filename(url, etag=None):
     """
     Convert `url` into a hashed filename in a repeatable way.
     If `etag` is specified, append its hash to the url's, delimited
     by a period.
-    If the url ends with .h5 (Keras HDF5 weights) ands '.h5' to the name
+    If the url ends with .h5 (Keras HDF5 weights) adds '.h5' to the name
     so that TF 2.0 can identify it as a HDF5 file
     (see https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/tensorflow/python/keras/engine/network.py#L1380)
     """
@@ -153,7 +186,7 @@ def filename_to_url(filename, cache_dir=None):
     return url, etag
 
 
-def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=None, resume_download=False):
+def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=None, resume_download=False, user_agent=None):
     """
     Given something that might be a URL (or might be a local path),
     determine which. If it's a URL, download the file and cache it, and
@@ -163,6 +196,7 @@ def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=N
         cache_dir: specify a cache directory to save the file to (overwrite the default cache dir).
         force_download: if True, re-dowload the file even if it's already cached in the cache dir.
         resume_download: if True, resume the download if incompletly recieved file is found.
+        user_agent: Optional string or dict that will be appended to the user-agent on remote requests.
     """
     if cache_dir is None:
         cache_dir = TRANSFORMERS_CACHE
@@ -171,17 +205,15 @@ def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=N
     if sys.version_info[0] == 3 and isinstance(cache_dir, Path):
         cache_dir = str(cache_dir)
 
-    parsed = urlparse(url_or_filename)
-
-    if parsed.scheme in ('http', 'https', 's3'):
+    if is_remote_url(url_or_filename):
         # URL, so get it from the cache (downloading if necessary)
         return get_from_cache(url_or_filename, cache_dir=cache_dir,
             force_download=force_download, proxies=proxies,
-            resume_download=resume_download)
+            resume_download=resume_download, user_agent=user_agent)
     elif os.path.exists(url_or_filename):
         # File, and it exists.
         return url_or_filename
-    elif parsed.scheme == '':
+    elif urlparse(url_or_filename).scheme == '':
         # File, but it doesn't exist.
         raise EnvironmentError("file {} not found".format(url_or_filename))
     else:
@@ -238,14 +270,26 @@ def s3_get(url, temp_file, proxies=None):
     s3_resource.Bucket(bucket_name).download_fileobj(s3_path, temp_file)
 
 
-def http_get(url, temp_file, proxies=None, resume_size=0):
-    headers={'Range':'bytes=%d-'%(resume_size,)} if resume_size > 0 else None
+def http_get(url, temp_file, proxies=None, resume_size=0, user_agent=None):
+    ua = "transformers/{}; python/{}".format(__version__, sys.version.split()[0])
+    if isinstance(user_agent, dict):
+        ua += "; " + "; ".join(
+            "{}/{}".format(k, v) for k, v in user_agent.items()
+        )
+    elif isinstance(user_agent, six.string_types):
+        ua += "; "+ user_agent
+    headers = {
+        "user-agent": ua
+    }
+    if resume_size > 0:
+        headers['Range'] = 'bytes=%d-' % (resume_size,)
     response = requests.get(url, stream=True, proxies=proxies, headers=headers)
     if response.status_code == 416:  # Range not satisfiable
         return
     content_length = response.headers.get('Content-Length')
     total = resume_size + int(content_length) if content_length is not None else None
-    progress = tqdm(unit="B", total=total, initial=resume_size)
+    progress = tqdm(unit="B", unit_scale=True, total=total, initial=resume_size,
+                    desc="Downloading", disable=bool(logger.level<=logging.INFO))
     for chunk in response.iter_content(chunk_size=1024):
         if chunk: # filter out keep-alive new chunks
             progress.update(len(chunk))
@@ -253,7 +297,7 @@ def http_get(url, temp_file, proxies=None, resume_size=0):
     progress.close()
 
 
-def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag_timeout=10, resume_download=False):
+def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag_timeout=10, resume_download=False, user_agent=None):
     """
     Given a URL, look for the corresponding dataset in the local cache.
     If it's not there, download it. Then return the path to the cached file.
@@ -291,59 +335,60 @@ def get_from_cache(url, cache_dir=None, force_download=False, proxies=None, etag
     # If we don't have a connection (etag is None) and can't identify the file
     # try to get the last downloaded one
     if not os.path.exists(cache_path) and etag is None:
-        matching_files = fnmatch.filter(os.listdir(cache_dir), filename + '.*')
-        matching_files = list(filter(lambda s: not s.endswith('.json'), matching_files))
+        matching_files = [
+            file
+            for file in fnmatch.filter(os.listdir(cache_dir), filename + '.*')
+            if not file.endswith('.json') and not file.endswith('.lock')
+        ]
         if matching_files:
             cache_path = os.path.join(cache_dir, matching_files[-1])
 
-    if resume_download:
-        incomplete_path = cache_path + '.incomplete'
-        @contextmanager
-        def _resumable_file_manager():
-            with open(incomplete_path,'a+b') as f:
-                yield f
-            os.remove(incomplete_path)
-        temp_file_manager = _resumable_file_manager
-        if os.path.exists(incomplete_path):
-            resume_size = os.stat(incomplete_path).st_size
-        else:
-            resume_size = 0
-    else:
-        temp_file_manager = tempfile.NamedTemporaryFile
-        resume_size = 0
+    # Prevent parallel downloads of the same file with a lock.
+    lock_path = cache_path + '.lock'
+    with FileLock(lock_path):
 
-    if not os.path.exists(cache_path) or force_download:
-        # Download to temporary file, then copy to cache dir once finished.
-        # Otherwise you get corrupt cache entries if the download gets interrupted.
-        with temp_file_manager() as temp_file:
-            logger.info("%s not found in cache or force_download set to True, downloading to %s", url, temp_file.name)
-
-            # GET file object
-            if url.startswith("s3://"):
-                if resume_download:
-                    logger.warn('Warning: resumable downloads are not implemented for "s3://" urls')
-                s3_get(url, temp_file, proxies=proxies)
+        if resume_download:
+            incomplete_path = cache_path + '.incomplete'
+            @contextmanager
+            def _resumable_file_manager():
+                with open(incomplete_path,'a+b') as f:
+                    yield f
+            temp_file_manager = _resumable_file_manager
+            if os.path.exists(incomplete_path):
+                resume_size = os.stat(incomplete_path).st_size
             else:
-                http_get(url, temp_file, proxies=proxies, resume_size=resume_size)
+                resume_size = 0
+        else:
+            temp_file_manager = partial(tempfile.NamedTemporaryFile, dir=cache_dir, delete=False)
+            resume_size = 0
 
-            # we are copying the file before closing it, so flush to avoid truncation
-            temp_file.flush()
-            # shutil.copyfileobj() starts at the current position, so go to the start
-            temp_file.seek(0)
+        if etag is not None and (not os.path.exists(cache_path) or force_download):
+            # Download to temporary file, then copy to cache dir once finished.
+            # Otherwise you get corrupt cache entries if the download gets interrupted.
+            with temp_file_manager() as temp_file:
+                logger.info("%s not found in cache or force_download set to True, downloading to %s", url, temp_file.name)
 
-            logger.info("copying %s to cache at %s", temp_file.name, cache_path)
-            with open(cache_path, 'wb') as cache_file:
-                shutil.copyfileobj(temp_file, cache_file)
+                # GET file object
+                if url.startswith("s3://"):
+                    if resume_download:
+                        logger.warn('Warning: resumable downloads are not implemented for "s3://" urls')
+                    s3_get(url, temp_file, proxies=proxies)
+                else:
+                    http_get(url, temp_file, proxies=proxies, resume_size=resume_size, user_agent=user_agent)
 
-            logger.info("creating metadata file for %s", cache_path)
-            meta = {'url': url, 'etag': etag}
-            meta_path = cache_path + '.json'
-            with open(meta_path, 'w') as meta_file:
-                output_string = json.dumps(meta)
-                if sys.version_info[0] == 2 and isinstance(output_string, str):
-                    output_string = unicode(output_string, 'utf-8')  # The beauty of python 2
-                meta_file.write(output_string)
+                # we are copying the file before closing it, so flush to avoid truncation
+                temp_file.flush()
 
-            logger.info("removing temp file %s", temp_file.name)
+                logger.info("storing %s in cache at %s", url, cache_path)
+                os.rename(temp_file.name, cache_path)
+
+                logger.info("creating metadata file for %s", cache_path)
+                meta = {'url': url, 'etag': etag}
+                meta_path = cache_path + '.json'
+                with open(meta_path, 'w') as meta_file:
+                    output_string = json.dumps(meta)
+                    if sys.version_info[0] == 2 and isinstance(output_string, str):
+                        output_string = unicode(output_string, 'utf-8')  # The beauty of python 2
+                    meta_file.write(output_string)
 
     return cache_path

transformers/hf_api.py

@@ -16,10 +16,11 @@ from __future__ import absolute_import, division, print_function
 
 import os
 from os.path import expanduser
-import six
 
 import requests
+import six
 from requests.exceptions import HTTPError
+from tqdm import tqdm
 
 ENDPOINT = "https://huggingface.co"
 
@@ -129,10 +130,14 @@ class HfApi:
         # Even though we presign with the correct content-type,
         # the client still has to specify it when uploading the file.
         with open(filepath, "rb") as f:
-            r = requests.put(urls.write, data=f, headers={
+            pf = TqdmProgressFileReader(f)
+            data = f if pf.total_size > 0 else ""
+
+            r = requests.put(urls.write, data=data, headers={
                 "content-type": urls.type,
             })
             r.raise_for_status()
+            pf.close()
         return urls.access
 
     def list_objs(self, token):
@@ -148,6 +153,34 @@ class HfApi:
 
 
 
+class TqdmProgressFileReader:
+    """
+    Wrap an io.BufferedReader `f` (such as the output of `open(…, "rb")`)
+    and override `f.read()` so as to display a tqdm progress bar.
+
+    see github.com/huggingface/transformers/pull/2078#discussion_r354739608
+    for implementation details.
+    """
+    def __init__(
+        self,
+        f   # type: io.BufferedReader
+    ):
+        self.f = f
+        self.total_size = os.fstat(f.fileno()).st_size # type: int
+        self.pbar = tqdm(total=self.total_size, leave=False)
+        if six.PY3:
+            # does not work unless PY3
+            # no big deal as the CLI does not currently support PY2 anyways.
+            self.read = f.read
+            f.read = self._read
+
+    def _read(self, n=-1):
+        self.pbar.update(n)
+        return self.read(n)
+
+    def close(self):
+        self.pbar.close()
+
 
 
 class HfFolder:

transformers/modelcard.py

@@ -0,0 +1,229 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Configuration base class and utilities."""
+
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import copy
+import json
+import logging
+import os
+from io import open
+
+from .configuration_auto import ALL_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+from .file_utils import CONFIG_NAME, MODEL_CARD_NAME, WEIGHTS_NAME, TF2_WEIGHTS_NAME, \
+                        cached_path, is_remote_url, hf_bucket_url
+
+
+logger = logging.getLogger(__name__)
+
+
+class ModelCard(object):
+    r""" Model Card class.
+        Store model card as well as methods for loading/downloading/saving model cards.
+
+        Please read the following paper for details and explanation on the sections:
+            "Model Cards for Model Reporting"
+                by Margaret Mitchell, Simone Wu,
+                Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer,
+                Inioluwa Deborah Raji and Timnit Gebru for the proposal behind model cards.
+            Link: https://arxiv.org/abs/1810.03993
+
+        Note:
+            A model card can be loaded and saved to disk.
+
+        Parameters:
+    """
+    def __init__(self, **kwargs):
+        # Recomended attributes from https://arxiv.org/abs/1810.03993 (see papers)
+        self.model_details = kwargs.pop('model_details', {})
+        self.intended_use = kwargs.pop('intended_use', {})
+        self.factors = kwargs.pop('factors', {})
+        self.metrics = kwargs.pop('metrics', {})
+        self.evaluation_data = kwargs.pop('evaluation_data', {})
+        self.training_data = kwargs.pop('training_data', {})
+        self.quantitative_analyses = kwargs.pop('quantitative_analyses', {})
+        self.ethical_considerations = kwargs.pop('ethical_considerations', {})
+        self.caveats_and_recommendations = kwargs.pop('caveats_and_recommendations', {})
+
+        # Open additional attributes
+        for key, value in kwargs.items():
+            try:
+                setattr(self, key, value)
+            except AttributeError as err:
+                logger.error("Can't set {} with value {} for {}".format(key, value, self))
+                raise err
+
+    def save_pretrained(self, save_directory_or_file):
+        """ Save a model card object to the directory or file `save_directory_or_file`.
+        """
+        if os.path.isdir(save_directory_or_file):
+            # If we save using the predefined names, we can load using `from_pretrained`
+            output_model_card_file = os.path.join(save_directory_or_file, MODEL_CARD_NAME)
+        else:
+            output_model_card_file = save_directory_or_file
+
+        self.to_json_file(output_model_card_file)
+        logger.info("Model card saved in {}".format(output_model_card_file))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        r""" Instantiate a :class:`~transformers.ModelCard` from a pre-trained model model card.
+
+        Parameters:
+            pretrained_model_name_or_path: either:
+
+                - a string with the `shortcut name` of a pre-trained model card to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model card that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
+                - a path to a `directory` containing a mode card file saved using the :func:`~transformers.ModelCard.save_pretrained` method, e.g.: ``./my_model_directory/``.
+                - a path or url to a saved model card JSON `file`, e.g.: ``./my_model_directory/modelcard.json``.
+
+            cache_dir: (`optional`) string:
+                Path to a directory in which a downloaded pre-trained model
+                card should be cached if the standard cache should not be used.
+
+            kwargs: (`optional`) dict: key/value pairs with which to update the ModelCard object after loading.
+
+                - The values in kwargs of any keys which are model card attributes will be used to override the loaded values.
+                - Behavior concerning key/value pairs whose keys are *not* model card attributes is controlled by the `return_unused_kwargs` keyword parameter.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            find_from_standard_name: (`optional`) boolean, default True:
+                If the pretrained_model_name_or_path ends with our standard model or config filenames, replace them with our standard modelcard filename.
+                Can be used to directly feed a model/config url and access the colocated modelcard.
+
+            return_unused_kwargs: (`optional`) bool:
+
+                - If False, then this function returns just the final model card object.
+                - If True, then this functions returns a tuple `(model card, unused_kwargs)` where `unused_kwargs` is a dictionary consisting of the key/value pairs whose keys are not model card attributes: ie the part of kwargs which has not been used to update `ModelCard` and is otherwise ignored.
+
+        Examples::
+
+            modelcard = ModelCard.from_pretrained('bert-base-uncased')    # Download model card from S3 and cache.
+            modelcard = ModelCard.from_pretrained('./test/saved_model/')  # E.g. model card was saved using `save_pretrained('./test/saved_model/')`
+            modelcard = ModelCard.from_pretrained('./test/saved_model/modelcard.json')
+            modelcard = ModelCard.from_pretrained('bert-base-uncased', output_attention=True, foo=False)
+
+        """
+        cache_dir = kwargs.pop('cache_dir', None)
+        proxies = kwargs.pop('proxies', None)
+        find_from_standard_name = kwargs.pop('find_from_standard_name', True)
+        return_unused_kwargs = kwargs.pop('return_unused_kwargs', False)
+
+        if pretrained_model_name_or_path in ALL_PRETRAINED_CONFIG_ARCHIVE_MAP:
+            # For simplicity we use the same pretrained url than the configuration files
+            # but with a different suffix (modelcard.json). This suffix is replaced below.
+            model_card_file = ALL_PRETRAINED_CONFIG_ARCHIVE_MAP[pretrained_model_name_or_path]
+        elif os.path.isdir(pretrained_model_name_or_path):
+            model_card_file = os.path.join(pretrained_model_name_or_path, MODEL_CARD_NAME)
+        elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
+            model_card_file = pretrained_model_name_or_path
+        else:
+            model_card_file = hf_bucket_url(pretrained_model_name_or_path, postfix=MODEL_CARD_NAME)
+
+        if find_from_standard_name or pretrained_model_name_or_path in ALL_PRETRAINED_CONFIG_ARCHIVE_MAP:
+            model_card_file = model_card_file.replace(CONFIG_NAME, MODEL_CARD_NAME)
+            model_card_file = model_card_file.replace(WEIGHTS_NAME, MODEL_CARD_NAME)
+            model_card_file = model_card_file.replace(TF2_WEIGHTS_NAME, MODEL_CARD_NAME)
+
+        try:
+            # Load from URL or cache if already cached
+            resolved_model_card_file = cached_path(model_card_file, cache_dir=cache_dir, force_download=True,
+                                               proxies=proxies, resume_download=False)
+            if resolved_model_card_file == model_card_file:
+                logger.info("loading model card file {}".format(model_card_file))
+            else:
+                logger.info("loading model card file {} from cache at {}".format(
+                    model_card_file, resolved_model_card_file))
+            # Load model card
+            modelcard = cls.from_json_file(resolved_model_card_file)
+
+        except EnvironmentError:
+            if pretrained_model_name_or_path in ALL_PRETRAINED_CONFIG_ARCHIVE_MAP:
+                logger.warning("Couldn't reach server at '{}' to download model card file.".format(
+                        model_card_file))
+            else:
+                logger.warning("Model name '{}' was not found in model name list ({}). " \
+                      "We assumed '{}' was a path or url to a model card file named {} or " \
+                      "a directory containing such a file but couldn't find any such file at this path or url.".format(
+                        pretrained_model_name_or_path,
+                        ', '.join(ALL_PRETRAINED_CONFIG_ARCHIVE_MAP.keys()),
+                        model_card_file, MODEL_CARD_NAME))
+            logger.warning("Creating an empty model card.")
+
+            # We fall back on creating an empty model card
+            modelcard = cls()
+
+        except json.JSONDecodeError:
+            logger.warning("Couldn't reach server at '{}' to download model card file or "
+                           "model card file is not a valid JSON file. "
+                           "Please check network or file content here: {}.".format(model_card_file, resolved_model_card_file))
+            logger.warning("Creating an empty model card.")
+
+            # We fall back on creating an empty model card
+            modelcard = cls()
+
+        # Update model card with kwargs if needed
+        to_remove = []
+        for key, value in kwargs.items():
+            if hasattr(modelcard, key):
+                setattr(modelcard, key, value)
+                to_remove.append(key)
+        for key in to_remove:
+            kwargs.pop(key, None)
+
+        logger.info("Model card: %s", str(modelcard))
+        if return_unused_kwargs:
+            return modelcard, kwargs
+        else:
+            return modelcard
+
+    @classmethod
+    def from_dict(cls, json_object):
+        """Constructs a `ModelCard` from a Python dictionary of parameters."""
+        return cls(**json_object)
+
+    @classmethod
+    def from_json_file(cls, json_file):
+        """Constructs a `ModelCard` from a json file of parameters."""
+        with open(json_file, "r", encoding='utf-8') as reader:
+            text = reader.read()
+        dict_obj = json.loads(text)
+        return cls(**dict_obj)
+
+    def __eq__(self, other):
+        return self.__dict__ == other.__dict__
+
+    def __repr__(self):
+        return str(self.to_json_string())
+
+    def to_dict(self):
+        """Serializes this instance to a Python dictionary."""
+        output = copy.deepcopy(self.__dict__)
+        return output
+
+    def to_json_string(self):
+        """Serializes this instance to a JSON string."""
+        return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
+
+    def to_json_file(self, json_file_path):
+        """ Save this instance to a json file."""
+        with open(json_file_path, "w", encoding='utf-8') as writer:
+            writer.write(self.to_json_string())

transformers/modeling_albert.py

@@ -572,7 +572,7 @@ class AlbertForMaskedLM(AlbertPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -624,7 +624,7 @@ class AlbertForMaskedLM(AlbertPreTrainedModel):
 
         outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
         if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             outputs = (masked_lm_loss,) + outputs
 

transformers/modeling_auto.py

@@ -18,18 +18,31 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import logging
 
-from .modeling_bert import BertModel, BertForMaskedLM, BertForSequenceClassification, BertForQuestionAnswering
-from .modeling_openai import OpenAIGPTModel, OpenAIGPTLMHeadModel
-from .modeling_gpt2 import GPT2Model, GPT2LMHeadModel
-from .modeling_ctrl import CTRLModel, CTRLLMHeadModel
-from .modeling_transfo_xl import TransfoXLModel, TransfoXLLMHeadModel
-from .modeling_xlnet import XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification, XLNetForQuestionAnswering
-from .modeling_xlm import XLMModel, XLMWithLMHeadModel, XLMForSequenceClassification, XLMForQuestionAnswering
-from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification
-from .modeling_distilbert import DistilBertModel, DistilBertForQuestionAnswering, DistilBertForMaskedLM, DistilBertForSequenceClassification
-from .modeling_camembert import CamembertModel, CamembertForMaskedLM, CamembertForSequenceClassification, CamembertForMultipleChoice
-from .modeling_camembert import CamembertModel, CamembertForMaskedLM, CamembertForSequenceClassification, CamembertForMultipleChoice
-from .modeling_albert import AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification, AlbertForQuestionAnswering
+from .configuration_auto import (AlbertConfig, BertConfig, CamembertConfig, CTRLConfig,
+                                 DistilBertConfig, GPT2Config, OpenAIGPTConfig, RobertaConfig,
+                                 TransfoXLConfig, XLMConfig, XLNetConfig, XLMRobertaConfig)
+
+from .modeling_bert import BertModel, BertForMaskedLM, BertForSequenceClassification, BertForQuestionAnswering, \
+    BertForTokenClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_openai import OpenAIGPTModel, OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_gpt2 import GPT2Model, GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_ctrl import CTRLModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_transfo_xl import TransfoXLModel, TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_xlnet import XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification, XLNetForQuestionAnswering, \
+    XLNetForTokenClassification, XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_xlm import XLMModel, XLMWithLMHeadModel, XLMForSequenceClassification, XLMForQuestionAnswering, \
+    XLM_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification, \
+    RobertaForTokenClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_distilbert import DistilBertModel, DistilBertForQuestionAnswering, DistilBertForMaskedLM, \
+    DistilBertForSequenceClassification, DistilBertForTokenClassification, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_camembert import CamembertModel, CamembertForMaskedLM, CamembertForSequenceClassification, \
+    CamembertForMultipleChoice, CamembertForTokenClassification, CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_albert import AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification, \
+    AlbertForQuestionAnswering, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_t5 import T5Model, T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_xlm_roberta import XLMRobertaModel, XLMRobertaForMaskedLM, XLMRobertaForSequenceClassification, \
+    XLMRobertaForMultipleChoice, XLMRobertaForTokenClassification, XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
 
 from .modeling_utils import PreTrainedModel, SequenceSummary
 
@@ -38,21 +51,42 @@ from .file_utils import add_start_docstrings
 logger = logging.getLogger(__name__)
 
 
+ALL_PRETRAINED_MODEL_ARCHIVE_MAP = dict((key, value)
+    for pretrained_map in [
+        BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
+        CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        T5_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ]
+    for key, value, in pretrained_map.items())
+
+
 class AutoModel(object):
     r"""
         :class:`~transformers.AutoModel` is a generic model class
         that will be instantiated as one of the base model classes of the library
         when created with the `AutoModel.from_pretrained(pretrained_model_name_or_path)`
-        class method.
+        or the `AutoModel.from_config(config)` class methods.
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         using pattern matching on the `pretrained_model_name_or_path` string.
 
         The base model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5Model (T5 model)
             - contains `distilbert`: DistilBertModel (DistilBERT model)
             - contains `albert`: AlbertModel (ALBERT model)
             - contains `camembert`: CamembertModel (CamemBERT model)
+            - contains `xlm-roberta`: XLMRobertaModel (XLM-RoBERTa model)
             - contains `roberta`: RobertaModel (RoBERTa model)
             - contains `bert`: BertModel (Bert model)
             - contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
@@ -66,7 +100,56 @@ class AutoModel(object):
     """
     def __init__(self):
         raise EnvironmentError("AutoModel is designed to be instantiated "
-            "using the `AutoModel.from_pretrained(pretrained_model_name_or_path)` method.")
+            "using the `AutoModel.from_pretrained(pretrained_model_name_or_path)` or "
+            "`AutoModel.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `openai-gpt` configuration class: OpenAIGPTModel (OpenAI GPT model)
+                    - isInstance of `gpt2` configuration class: GPT2Model (OpenAI GPT-2 model)
+                    - isInstance of `ctrl` configuration class: CTRLModel (Salesforce CTRL  model)
+                    - isInstance of `transfo-xl` configuration class: TransfoXLModel (Transformer-XL model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: XLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModel.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, DistilBertConfig):
+            return DistilBertModel(config)
+        elif isinstance(config, RobertaConfig):
+            return RobertaModel(config)
+        elif isinstance(config, BertConfig):
+            return BertModel(config)
+        elif isinstance(config, OpenAIGPTConfig):
+            return OpenAIGPTModel(config)
+        elif isinstance(config, GPT2Config):
+            return GPT2Model(config)
+        elif isinstance(config, TransfoXLConfig):
+            return TransfoXLModel(config)
+        elif isinstance(config, XLNetConfig):
+            return XLNetModel(config)
+        elif isinstance(config, XLMConfig):
+            return XLMModel(config)
+        elif isinstance(config, CTRLConfig):
+            return CTRLModel(config)
+        elif isinstance(config, AlbertConfig):
+            return AlbertModel(config)
+        elif isinstance(config, CamembertConfig):
+            return CamembertModel(config)
+        elif isinstance(config, XLMRobertaConfig):
+            return XLMRobertaModel(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -75,9 +158,11 @@ class AutoModel(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5Model (T5 model)
             - contains `distilbert`: DistilBertModel (DistilBERT model)
             - contains `albert`: AlbertModel (ALBERT model)
             - contains `camembert`: CamembertModel (CamemBERT model)
+            - contains `xlm-roberta`: XLMRobertaModel (XLM-RoBERTa model)
             - contains `roberta`: RobertaModel (RoBERTa model)
             - contains `bert`: BertModel (Bert model)
             - contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
@@ -94,6 +179,7 @@ class AutoModel(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -146,12 +232,16 @@ class AutoModel(object):
             model = AutoModel.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return T5Model.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return DistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'albert' in pretrained_model_name_or_path:
             return AlbertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'camembert' in pretrained_model_name_or_path:
             return CamembertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'xlm-roberta' in pretrained_model_name_or_path:
+            return XLMRobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -170,7 +260,7 @@ class AutoModel(object):
             return CTRLModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                          "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
-                         "'xlm', 'roberta, 'ctrl', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
+                         "'xlm-roberta', 'xlm', 'roberta, 'ctrl', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
 
 
 class AutoModelWithLMHead(object):
@@ -185,9 +275,11 @@ class AutoModelWithLMHead(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5ModelWithLMHead (T5 model)
             - contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
             - contains `albert`: AlbertForMaskedLM (ALBERT model)
             - contains `camembert`: CamembertForMaskedLM (CamemBERT model)
+            - contains `xlm-roberta`: XLMRobertaForMaskedLM (XLM-RoBERTa model)
             - contains `roberta`: RobertaForMaskedLM (RoBERTa model)
             - contains `bert`: BertForMaskedLM (Bert model)
             - contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -201,7 +293,52 @@ class AutoModelWithLMHead(object):
     """
     def __init__(self):
         raise EnvironmentError("AutoModelWithLMHead is designed to be instantiated "
-            "using the `AutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
+            "using the `AutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` or "
+            "`AutoModelWithLMHead.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `openai-gpt` configuration class: OpenAIGPTModel (OpenAI GPT model)
+                    - isInstance of `gpt2` configuration class: GPT2Model (OpenAI GPT-2 model)
+                    - isInstance of `ctrl` configuration class: CTRLModel (Salesforce CTRL  model)
+                    - isInstance of `transfo-xl` configuration class: TransfoXLModel (Transformer-XL model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: XLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelWithLMHead.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, DistilBertConfig):
+            return DistilBertForMaskedLM(config)
+        elif isinstance(config, RobertaConfig):
+            return RobertaForMaskedLM(config)
+        elif isinstance(config, BertConfig):
+            return BertForMaskedLM(config)
+        elif isinstance(config, OpenAIGPTConfig):
+            return OpenAIGPTLMHeadModel(config)
+        elif isinstance(config, GPT2Config):
+            return GPT2LMHeadModel(config)
+        elif isinstance(config, TransfoXLConfig):
+            return TransfoXLLMHeadModel(config)
+        elif isinstance(config, XLNetConfig):
+            return XLNetLMHeadModel(config)
+        elif isinstance(config, XLMConfig):
+            return XLMWithLMHeadModel(config)
+        elif isinstance(config, CTRLConfig):
+            return CTRLLMHeadModel(config)
+        elif isinstance(config, XLMRobertaConfig):
+            return XLMRobertaForMaskedLM(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -213,9 +350,11 @@ class AutoModelWithLMHead(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5ModelWithLMHead (T5 model)
             - contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
             - contains `albert`: AlbertForMaskedLM (ALBERT model)
             - contains `camembert`: CamembertForMaskedLM (CamemBERT model)
+            - contains `xlm-roberta`: XLMRobertaForMaskedLM (XLM-RoBERTa model)
             - contains `roberta`: RobertaForMaskedLM (RoBERTa model)
             - contains `bert`: BertForMaskedLM (Bert model)
             - contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -232,6 +371,7 @@ class AutoModelWithLMHead(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -283,12 +423,16 @@ class AutoModelWithLMHead(object):
             model = AutoModelWithLMHead.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return T5WithLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return DistilBertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'albert' in pretrained_model_name_or_path:
             return AlbertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'camembert' in pretrained_model_name_or_path:
             return CamembertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'xlm-roberta' in pretrained_model_name_or_path:
+            return XLMRobertaForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -307,7 +451,7 @@ class AutoModelWithLMHead(object):
             return CTRLLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                          "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
-                         "'xlm', 'roberta','ctrl', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
+                         "'xlm-roberta', 'xlm', 'roberta','ctrl', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
 
 
 class AutoModelForSequenceClassification(object):
@@ -325,6 +469,7 @@ class AutoModelForSequenceClassification(object):
             - contains `distilbert`: DistilBertForSequenceClassification (DistilBERT model)
             - contains `albert`: AlbertForSequenceClassification (ALBERT model)
             - contains `camembert`: CamembertForSequenceClassification (CamemBERT model)
+            - contains `xlm-roberta`: XLMRobertaForSequenceClassification (XLM-RoBERTa model)
             - contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
             - contains `bert`: BertForSequenceClassification (Bert model)
             - contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -333,8 +478,45 @@ class AutoModelForSequenceClassification(object):
         This class cannot be instantiated using `__init__()` (throws an error).
     """
     def __init__(self):
-        raise EnvironmentError("AutoModelWithLMHead is designed to be instantiated "
-            "using the `AutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
+        raise EnvironmentError("AutoModelForSequenceClassification is designed to be instantiated "
+            "using the `AutoModelForSequenceClassification.from_pretrained(pretrained_model_name_or_path)` or "
+            "`AutoModelForSequenceClassification.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: XLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelForSequenceClassification.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, AlbertConfig):
+            return AlbertForSequenceClassification(config)
+        elif isinstance(config, CamembertConfig):
+            return CamembertForSequenceClassification(config)
+        elif isinstance(config, DistilBertConfig):
+            return DistilBertForSequenceClassification(config)
+        elif isinstance(config, RobertaConfig):
+            return RobertaForSequenceClassification(config)
+        elif isinstance(config, BertConfig):
+            return BertForSequenceClassification(config)
+        elif isinstance(config, XLNetConfig):
+            return XLNetForSequenceClassification(config)
+        elif isinstance(config, XLMConfig):
+            return XLMForSequenceClassification(config)
+        elif isinstance(config, XLMRobertaConfig):
+            return XLMRobertaForSequenceClassification(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -349,6 +531,7 @@ class AutoModelForSequenceClassification(object):
             - contains `distilbert`: DistilBertForSequenceClassification (DistilBERT model)
             - contains `albert`: AlbertForSequenceClassification (ALBERT model)
             - contains `camembert`: CamembertForSequenceClassification (CamemBERT model)
+            - contains `xlm-roberta`: XLMRobertaForSequenceClassification (XLM-RoBERTa model)
             - contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
             - contains `bert`: BertForSequenceClassification (Bert model)
             - contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -361,6 +544,7 @@ class AutoModelForSequenceClassification(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -419,6 +603,8 @@ class AutoModelForSequenceClassification(object):
             return AlbertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'camembert' in pretrained_model_name_or_path:
             return CamembertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'xlm-roberta' in pretrained_model_name_or_path:
+            return XLMRobertaForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -429,7 +615,7 @@ class AutoModelForSequenceClassification(object):
             return XLMForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'xlnet', 'xlm', 'roberta', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
+                         "'bert', 'xlnet', 'xlm-roberta', 'xlm', 'roberta', 'distilbert', 'camembert', 'albert'".format(pretrained_model_name_or_path))
 
 
 class AutoModelForQuestionAnswering(object):
@@ -453,8 +639,38 @@ class AutoModelForQuestionAnswering(object):
         This class cannot be instantiated using `__init__()` (throws an error).
     """
     def __init__(self):
-        raise EnvironmentError("AutoModelWithLMHead is designed to be instantiated "
-            "using the `AutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
+        raise EnvironmentError("AutoModelForQuestionAnswering is designed to be instantiated "
+            "using the `AutoModelForQuestionAnswering.from_pretrained(pretrained_model_name_or_path)` or "
+            "`AutoModelForQuestionAnswering.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: XLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelForSequenceClassification.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, AlbertConfig):
+            return AlbertForQuestionAnswering(config)
+        elif isinstance(config, DistilBertConfig):
+            return DistilBertForQuestionAnswering(config)
+        elif isinstance(config, BertConfig):
+            return BertForQuestionAnswering(config)
+        elif isinstance(config, XLNetConfig):
+            return XLNetForQuestionAnswering(config)
+        elif isinstance(config, XLMConfig):
+            return XLMForQuestionAnswering(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -479,6 +695,7 @@ class AutoModelForQuestionAnswering(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -541,3 +758,130 @@ class AutoModelForQuestionAnswering(object):
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                          "'bert', 'xlnet', 'xlm', 'distilbert', 'albert'".format(pretrained_model_name_or_path))
+
+
+class AutoModelForTokenClassification:
+    def __init__(self):
+        raise EnvironmentError("AutoModelForTokenClassification is designed to be instantiated "
+                               "using the `AutoModelForTokenClassification.from_pretrained(pretrained_model_name_or_path)` or "
+                               "`AutoModelForTokenClassification.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+    
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `camembert` configuration class: CamembertModel (Camembert model)
+                    - isInstance of `roberta` configuration class: RobertaModel (Roberta model)
+
+        Examples::
+    
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelForTokenClassification.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, CamembertConfig):
+            return CamembertForTokenClassification(config)
+        elif isinstance(config, DistilBertConfig):
+            return DistilBertForTokenClassification(config)
+        elif isinstance(config, BertConfig):
+            return BertForTokenClassification(config)
+        elif isinstance(config, XLNetConfig):
+            return XLNetForTokenClassification(config)
+        elif isinstance(config, RobertaConfig):
+            return RobertaForTokenClassification(config)
+        elif isinstance(config, XLMRobertaConfig):
+            return XLMRobertaForTokenClassification(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
+    
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        r""" Instantiates one of the question answering model classes of the library
+        from a pre-trained model configuration.
+
+        The `from_pretrained()` method takes care of returning the correct model class instance
+        using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertForTokenClassification (DistilBERT model)
+            - contains `camembert`: CamembertForTokenClassification (Camembert model)
+            - contains `bert`: BertForTokenClassification (Bert model)
+            - contains `xlnet`: XLNetForTokenClassification (XLNet model)
+            - contains `roberta`: RobertaForTokenClassification (Roberta model)
+
+        The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
+        To train the model, you should first set it back in training mode with `model.train()`
+
+        Params:
+            pretrained_model_name_or_path: either:
+
+                - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
+                - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
+
+            model_args: (`optional`) Sequence of positional arguments:
+                All remaning positional arguments will be passed to the underlying model's ``__init__`` method
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:
+
+                - the model is a model provided by the library (loaded with the ``shortcut-name`` string of a pretrained model), or
+                - the model was saved using :func:`~transformers.PreTrainedModel.save_pretrained` and is reloaded by suppling the save directory.
+                - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
+
+            state_dict: (`optional`) dict:
+                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                This option can be used if you want to create a model from a pretrained configuration but load your own weights.
+                In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
+
+            cache_dir: (`optional`) string:
+                Path to a directory in which a downloaded pre-trained model
+                configuration should be cached if the standard cache should not be used.
+
+            force_download: (`optional`) boolean, default False:
+                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            output_loading_info: (`optional`) boolean:
+                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+
+            kwargs: (`optional`) Remaining dictionary of keyword arguments:
+                Can be used to update the configuration object (after it being loaded) and initiate the model. (e.g. ``output_attention=True``). Behave differently depending on whether a `config` is provided or automatically loaded:
+
+                - If a configuration is provided with ``config``, ``**kwargs`` will be directly passed to the underlying model's ``__init__`` method (we assume all relevant updates to the configuration have already been done)
+                - If a configuration is not provided, ``kwargs`` will be first passed to the configuration class initialization function (:func:`~transformers.PretrainedConfig.from_pretrained`). Each key of ``kwargs`` that corresponds to a configuration attribute will be used to override said attribute with the supplied ``kwargs`` value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model's ``__init__`` function.
+
+        Examples::
+
+            model = AutoModelForTokenClassification.from_pretrained('bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = AutoModelForTokenClassification.from_pretrained('./test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = AutoModelForTokenClassification.from_pretrained('bert-base-uncased', output_attention=True)  # Update configuration during loading
+            assert model.config.output_attention == True
+            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+            config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
+            model = AutoModelForTokenClassification.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+
+        """
+        if 'camembert' in pretrained_model_name_or_path:
+            return CamembertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
+            return DistilBertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'xlm-roberta' in pretrained_model_name_or_path:
+            return XLMRobertaForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'roberta' in pretrained_model_name_or_path:
+            return RobertaForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'bert' in pretrained_model_name_or_path:
+            return BertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'xlnet' in pretrained_model_name_or_path:
+            return XLNetForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+
+        raise ValueError("Unrecognized model identifier in {}. Should contains one of "
+                         "'bert', 'xlnet', 'camembert', 'distilbert', 'xlm-roberta', 'roberta'".format(pretrained_model_name_or_path))

transformers/modeling_beam_search.py

@@ -1,271 +0,0 @@
-# coding=utf-8
-# Copyright (c) 2019 Yang Liu
-
-# Permission is hereby granted, free of charge, to any person obtaining a copy
-# of this software and associated documentation files (the "Software"), to deal
-# in the Software without restriction, including without limitation the rights
-# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-# copies of the Software, and to permit persons to whom the Software is
-# furnished to do so, subject to the following conditions:
-
-# The above copyright notice and this permission notice shall be included in all
-# copies or substantial portions of the Software.
-
-# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-# SOFTWARE.
-"""
-A general wrapper around models with LM heads to generate sequences
-using beam search.
-"""
-import torch
-from torch import nn
-
-
-class TransformerBeamSearch(nn.Module):
-    def __init__(
-        self,
-        model,
-        tokenizer,
-        batch_size,
-        beam_size,
-        min_length,
-        max_length,
-        alpha=0,
-        block_repeating_trigram=True,
-    ):
-        """
-        Attributes:
-            mask_word_id: token id that corresponds to the mask
-        """
-        super(TransformerBeamSearch, self).__init__()
-        self.model = model
-        self.tokenizer = tokenizer
-
-        self.start_token_id = tokenizer.start_token_id
-        self.end_token_id = tokenizer.end_token_id
-        self.pad_token_id = tokenizer.pad_token_id
-
-        self.beam_size = beam_size
-        self.min_length = min_length
-        self.max_length = max_length
-
-        self.block_repeating_trigram = block_repeating_trigram
-        self.apply_length_penalty = False if alpha == 0 else True
-        self.alpha = alpha
-
-        # State of the beam
-        self.hypotheses = [[] for _ in range(batch_size)]
-        self.batch_offset = torch.arange(batch_size, dtype=torch.long)
-        self.beam_offset = torch.arange(
-            0, batch_size * self.beam_size, step=self.beam_size, dtype=torch.long
-        )
-        self.growing_beam = torch.full(
-            (batch_size * self.beam_size, 1), self.start_token_id, dtype=torch.long
-        )
-        self.topk_log_probabilities = torch.tensor(
-            [0.0] + [float("-inf")] * (self.beam_size - 1), dtype=torch.float
-        ).repeat(batch_size)
-        self.results = {
-            "prediction": [[] for _ in batch_size],
-            "scores": [[] for _ in batch_size],
-        }
-        self._step = 0
-        self.is_done = False
-
-    def step(self, log_probabilities):
-        """ Grows the beam by one step. """
-        self._step += 1
-
-        # The batch size changes as some beams finish so we define _B
-        vocab_size = log_probabilities.size(-1)
-        _B = log_probabilities.size(0) // self.beam_size
-
-        # Multiply each beam probability with the probability of the
-        # next token (conditioned on the words in the beam).
-        log_probabilities += self.topk_log_probabilities.view(-1, 1)
-
-        self.enforce_min_length(log_probabilities)
-        if self.block_repeating_trigram:
-            self.remove_repeating_trigrams(log_probabilities, _B)
-
-        # Find the `beam_size` (previous_beam + token) combinations with
-        # the highest score
-        topk_log_probabilities, topk_ids = log_probabilities.topk(
-            log_probabilities.view(_B, self.beam_size * vocab_size),
-            self.beam_size,
-            dim=1,
-        )
-
-        # Apply the length penalty. The +1 accounts for the [EOS] token
-        # that will be added if the beam ends.
-        topk_scores = topk_log_probabilities / self.length_penalty()
-
-        # Retrieve the corresponding respective beam and token id
-        # topk_token_ids[i] will be added to topk_beam_ids[i]
-        topk_beam_ids = topk_ids.div(vocab_size)
-        topk_token_ids = topk_ids.fmod(vocab_size)
-
-        # Retrieve the row index of the surviving beams in the original
-        # view of the log_probabilities tensor
-        surviving_beams_rows = (topk_beam_ids + self.beam_offset[:_B].view(-1, 1)).view(
-            -1
-        )
-
-        # Append the last predictions
-        self.growing_beam = torch.cat(
-            [
-                self.growing_beam.index_select(0, surviving_beams_rows),
-                topk_token_ids.view(-1, 1),
-            ],
-            1,
-        )
-
-        # Check if any of the beam searches has ended during this
-        # growth step. Also if top beam (most probable) has ended
-        # for one element of the batch.
-        is_finished = topk_token_ids.eq(self.end_token_id)
-        self.enforce_max_length()
-        is_top_beam_finished = is_finished[:, 0].eq(1)
-
-        # Save the finished searches
-        if is_finished.any():
-            predictions = self.growing_beam.view(
-                -1, self.beam_size, self.growing_beam.size(1)
-            )
-            for i in range(is_finished.size(0)):
-                if is_top_beam_finished[i]:
-                    is_finished[i].fill_(1)
-                finished_hyp = is_finished[i].nonzero().view(-1)
-
-                # Store finished hypotheses for this batch.
-                b = self.batch_offset[i]
-                for j in finished_hyp:
-                    self.hypotheses[b].append((topk_scores[i, j], predictions[i, j, :]))
-
-                # If the batch reached the end, save the best hypotheses
-                # in terms of length-penalized score.
-                if is_top_beam_finished[i]:
-                    best_hyp = sorted(
-                        self.hypotheses[b], key=lambda x: x[0], reverse=True
-                    )
-                    best_score, best_prediction = best_hyp[0]
-                    self.results["scores"][b].append(best_score)
-                    self.results["predictions"][b].append(best_prediction)
-
-            non_finished = is_top_beam_finished.eq(0).nonzero().view(-1)
-            if len(non_finished) == 0:
-                self.is_done = True
-
-            # Remove finished batches for the next step.
-            topk_log_probabilities = topk_log_probabilities.index_select(
-                0, non_finished
-            )
-            self.batch_offset = self.batch_offset.index_select(0, non_finished)
-            self.growing_beam = predictions.index_select(0, non_finished).view(
-                -1, self.growing_beam.size(-1)
-            )
-
-            surviving_beams_rows = surviving_beams_rows.index_select(0, non_finished)
-
-        return surviving_beams_rows
-
-    def forward(self, encoder_input_ids, **kwargs):
-        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
-        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
-        # that apply to the model as whole.
-        # We let the specific kwargs override the common ones in case of conflict.
-        kwargs_encoder = {
-            argument[len("encoder_"):]: value
-            for argument, value in kwargs.items()
-            if argument.startswith("encoder_")
-        }
-        kwargs_decoder = {
-            argument[len("decoder_"):]: value
-            for argument, value in kwargs.items()
-            if argument.startswith("decoder_")
-        }
-        kwargs_common = {
-            argument: value
-            for argument, value in kwargs.items()
-            if not (argument.startswith("encoder_") or argument.startswith("decoder_"))
-        }
-        kwargs_decoder = dict(kwargs_common, **kwargs_decoder)
-        kwargs_encoder = dict(kwargs_common, **kwargs_encoder)
-
-        # forward pass on the encoder
-        encoder_outputs = self.model.encoder.forward(encoder_input_ids, kwargs_encoder)
-        kwargs_decoder["encoder_hidden_states"] = tile(
-            encoder_outputs, self.beam_size, dim=0
-        )
-
-        # grow the beam by generating sequences in an autoregressive way
-        self.growing_beam = torch.full(
-            (self.batch_size * self.beam_size, 1), self.start_token_id, dtype=torch.long
-        )
-        for step in range(self.max_length):
-            decoder_input = self.growing_beam[:, -1]
-            outputs = self.model.decoder(decoder_input, kwargs_decoder)
-            log_probabilities = torch.nn.functional.log_softmax(outputs[1])
-            surviving_beams_rows = self.step(log_probabilities)
-            if self.is_done:
-                break
-
-            kwargs_decoder["encoder_hidden_states"] = kwargs_decoder[
-                "encoder_hidden_states"
-            ].index_select(0, surviving_beams_rows)
-
-        return self.results
-
-    def remove_repeating_trigrams(self, log_probabilities, _B):
-        if(self._step + 1 > 3):
-            for i in range(_B * self.beam_size):
-                tokens = [t for t in self.growing_beam[i]]
-                trigrams = [(tokens[i-1], tokens[i], tokens[i+1]) for i in range(1, len(words) - 1)]
-                last_trigram = tuple(trigrams[-1])
-                if last_trigram in trigrams[:-1]:
-                    log_probabilities[i] = -1e20
-
-    def enforce_min_length(self):
-        if self._step < self.min_length:
-            self.log_probabilities[self.end_token_id] = -1e20
-
-    def enforce_max_length(self):
-        if self._step + 1 == self.max_length:
-            self.is_finished.fill_(1)
-
-    def length_penalty(self):
-        return ((5.0 + (self._step + 1)) / 6.0) ** self.alpha
-
-
-def tile(x, count, dim=0):
-    """
-    Tiles `x` along dimension `dim` `count` times.
-
-    Example:
-        >> ex = torch.tensor([1,2],[3,4])
-        >> tile(ex, 2, 0)
-        torch.Tensor([[1,2],[1,2],[3,4],[3,4]])
-    """
-    perm = list(range(len(x.size())))
-    if dim != 0:
-        perm[0], perm[dim] = perm[dim], perm[0]
-        x = x.permute(perm).contiguous()
-    out_size = list(x.size())
-    out_size[0] *= count
-    batch = x.size(0)
-    x = (
-        x.view(batch, -1)
-        .transpose(0, 1)
-        .repeat(count, 1)
-        .transpose(0, 1)
-        .contiguous()
-        .view(*out_size)
-    )
-    if dim != 0:
-        x = x.permute(perm).contiguous()
-    return x

transformers/modeling_bert.py

@@ -48,6 +48,12 @@ BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-pytorch_model.bin",
     'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-pytorch_model.bin",
     'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-pytorch_model.bin",
+    'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-pytorch_model.bin",
+    'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-pytorch_model.bin",
+    'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-pytorch_model.bin",
+    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-pytorch_model.bin",
+    'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/pytorch_model.bin",
+    'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/pytorch_model.bin",
 }
 
 
@@ -478,58 +484,6 @@ class BertPreTrainingHeads(nn.Module):
         return prediction_scores, seq_relationship_score
 
 
-class RBertClassificationHead(nn.Module):
-    def __init__(self, config):
-        """
-        :param config: a BertConfig or aRobertaConfig.
-        """
-        super(RBertClassificationHead, self).__init__()
-        self.ent_2_index_id = config.entity_2_token_id
-        self.ent_1_index_id = config.entity_1_token_id
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.sentence_layer = torch.nn.Sequential(
-            self.dropout,
-            torch.nn.Linear(config.hidden_size, config.hidden_size),
-            torch.nn.modules.activation.Tanh(),
-        )
-        self.entity_layer = torch.nn.Sequential(
-            self.dropout,
-            torch.nn.Linear(config.hidden_size, config.hidden_size),
-            torch.nn.modules.activation.Tanh()
-        )
-
-        self.classifier = nn.Linear(config.hidden_size * 3, config.num_labels)
-
-    def get_indices_tensor(self, instance_input_ids, entity_char_id):
-        index_tensors = (instance_input_ids == entity_char_id).nonzero()
-        start_index = index_tensors[0].item() + 1  # do not include the symbol itself in the calculation
-        end_index = index_tensors[1].item()
-        return torch.arange(start_index, end_index, device=instance_input_ids.device.type)
-
-    def average_entity_vectors(self, last_hidden_states, input_ids, entity_char_id):
-        batch_return_list = []
-        for instance_hidden_state, instance_input_ids in zip(last_hidden_states, input_ids):
-            lookup_tensor = self.get_indices_tensor(instance_input_ids, entity_char_id)
-            average_of_entity_vectors = torch.index_select(instance_hidden_state, 0, lookup_tensor).unsqueeze(0).mean(1)
-            batch_return_list.append(average_of_entity_vectors)
-        return_tensor = torch.cat(batch_return_list)
-        return_tensor[return_tensor.ne(return_tensor)] = 0 #set any nan's created by mean operation to 0
-        return return_tensor
-
-    def forward(self, input_ids, last_hidden_states):
-        cls_tensor = self.sentence_layer(
-            last_hidden_states[:, 0])  # get the sentence embedding and pass through FC layer
-        ent1_tensor = self.entity_layer(self.average_entity_vectors(last_hidden_states, input_ids,
-                                                                    self.ent_1_index_id))  # average wordpieces of ent1 and pass through FC entity layer
-        ent2_tensor = self.entity_layer(self.average_entity_vectors(last_hidden_states, input_ids,
-                                                                    self.ent_2_index_id))  # ditto - ent2 shared parameters with ent1
-
-        entities_and_cls_tensor = torch.cat((cls_tensor, ent1_tensor, ent2_tensor),
-                                            dim=1)  # concat all and pass through classifier layer
-        logits = self.classifier(entities_and_cls_tensor)
-        return logits
-
-
 class BertPreTrainedModel(PreTrainedModel):
     """ An abstract class to handle weights initialization and
         a simple interface for dowloading and loading pretrained models.
@@ -719,18 +673,20 @@ class BertModel(BertPreTrainedModel):
         # ourselves in which case we just need to make it broadcastable to all heads.
         if attention_mask.dim() == 3:
             extended_attention_mask = attention_mask[:, None, :, :]
-
-        # Provided a padding mask of dimensions [batch_size, seq_length]
-        # - if the model is a decoder, apply a causal mask in addition to the padding mask
-        # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
-        if attention_mask.dim() == 2:
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
             if self.config.is_decoder:
                 batch_size, seq_length = input_shape
                 seq_ids = torch.arange(seq_length, device=device)
                 causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                causal_mask = causal_mask.to(torch.long)  # not converting to long will cause errors with pytorch version < 1.3
                 extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
             else:
                 extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError("Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(input_shape, attention_mask.shape))
 
         # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
         # masked positions, this operation will create a tensor which is 0.0 for
@@ -742,14 +698,19 @@ class BertModel(BertPreTrainedModel):
 
         # If a 2D ou 3D attention mask is provided for the cross-attention
         # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
-        if self.config.is_decoder:
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
             if encoder_attention_mask is None:
-                encoder_attention_mask = torch.ones(input_shape, device=device)
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
 
             if encoder_attention_mask.dim() == 3:
                 encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
-            if encoder_attention_mask.dim() == 2:
+            elif encoder_attention_mask.dim() == 2:
                 encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+            else:
+                raise ValueError("Wrong shape for encoder_hidden_shape (shape {}) or encoder_attention_mask (shape {})".format(encoder_hidden_shape,
+                                                                                                                               encoder_attention_mask.shape))
 
             encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
             encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
@@ -793,7 +754,7 @@ class BertForPreTraining(BertPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
         **next_sentence_label**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
             Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair (see ``input_ids`` docstring)
@@ -852,7 +813,7 @@ class BertForPreTraining(BertPreTrainedModel):
         outputs = (prediction_scores, seq_relationship_score,) + outputs[2:]  # add hidden states and attention if they are here
 
         if masked_lm_labels is not None and next_sentence_label is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
             total_loss = masked_lm_loss + next_sentence_loss
@@ -869,12 +830,12 @@ class BertForMaskedLM(BertPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
         **lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the left-to-right language modeling loss (next word prediction).
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -936,7 +897,7 @@ class BertForMaskedLM(BertPreTrainedModel):
         # 2. If `lm_labels` is provided we are in a causal scenario where we
         #    try to predict the next token for each input in the decoder.
         if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)  # -1 index = padding token
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             outputs = (masked_lm_loss,) + outputs
 
@@ -944,7 +905,7 @@ class BertForMaskedLM(BertPreTrainedModel):
             # we are doing next-token prediction; shift prediction scores and input ids by one
             prediction_scores = prediction_scores[:, :-1, :].contiguous()
             lm_labels = lm_labels[:, 1:].contiguous()
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             ltr_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), lm_labels.view(-1))
             outputs = (ltr_lm_loss,) + outputs
 
@@ -1008,87 +969,13 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
         outputs = (seq_relationship_score,) + outputs[2:]  # add hidden states and attention if they are here
         if next_sentence_label is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
             outputs = (next_sentence_loss,) + outputs
 
         return outputs  # (next_sentence_loss), seq_relationship_score, (hidden_states), (attentions)
 
 
-@add_start_docstrings(r"""Bert Model transformer with an R-BERT head on top, making use of
-        entity offsets for relationship classification. See the details in the R-Bert paper https://arxiv.org/pdf/1905.08284.pdf""",
-        BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
-class BertForRelationshipClassification(BertPreTrainedModel):
-    r"""
-        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-        Labels for computing the sequence classification/regression loss.
-        Indices should be in ``[0, ..., config.num_labels - 1]``.
-        If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
-        If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
-
-    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
-            Classification (or regression if config.num_labels==1) loss.
-        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
-            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
-            of shape ``(batch_size, sequence_length, hidden_size)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
-
-    Examples::
-
-        test_string = "The $ cat $ sat on the # mat # ."
-        ent1_sep_token = '$'
-        ent2_sep_token = '#'
-
-        model_name = 'bert-base-uncased'
-
-        tokenizer = BertTokenizer.from_pretrained(model_name)
-        ent1_sep_token_id = tokenizer.encode(ent1_sep_token)[0]
-        ent2_sep_token_id = tokenizer.encode(ent2_sep_token)[0]
-        config = BertConfig.from_pretrained(model_name, entity_1_token_id=ent1_sep_token_id, entity_2_token_id=ent2_sep_token_id)
-        model = BertForRelationshipClassification.from_pretrained(model_name, config=config)
-        input_ids = torch.tensor(tokenizer.encode(test_string,add_special_tokens=True)).unsqueeze(0)  # Batch size 1
-        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
-        outputs = model(input_ids, labels=labels)
-        loss, logits = outputs[:2]
-
-    """
-
-    def __init__(self, config):
-        super(BertForRelationshipClassification, self).__init__(config)
-        self.num_labels = config.num_labels
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.bert = BertModel(config)
-        self.rbert = RBertClassificationHead(config)
-        self.init_weights()
-
-    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
-                position_ids=None, head_mask=None, labels=None):
-
-        outputs = self.bert(input_ids, position_ids=position_ids, token_type_ids=token_type_ids,
-                            attention_mask=attention_mask, head_mask=head_mask)
-        last_hidden_states = outputs[0]
-        logits = self.rbert(input_ids, last_hidden_states)
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
-
-        if labels is not None:
-            if self.num_labels == 1:
-                #  We are doing regression
-                loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1), labels.view(-1))
-            else:
-                loss_fct = CrossEntropyLoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
-            outputs = (loss,) + outputs
-
-        return outputs  # (loss), logits, (hidden_states), (attentions)
-
-
 @add_start_docstrings("""Bert Model transformer with a sequence classification/regression head on top (a linear layer on top of
                       the pooled output) e.g. for GLUE tasks. """,
                       BERT_START_DOCSTRING,
@@ -1352,9 +1239,9 @@ class BertForQuestionAnswering(BertPreTrainedModel):
         question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
         input_text = "[CLS] " + question + " [SEP] " + text + " [SEP]"
         input_ids = tokenizer.encode(input_text)
-        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))] 
+        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))]
         start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))
-        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)  
+        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)
         print(' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]))
         # a nice puppet
 

transformers/modeling_camembert.py

@@ -156,7 +156,7 @@ class CamembertForMaskedLM(RobertaForMaskedLM):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:

transformers/modeling_ctrl.py

@@ -252,7 +252,7 @@ class CTRLModel(CTRLPreTrainedModel):
         **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -268,7 +268,7 @@ class CTRLModel(CTRLPreTrainedModel):
 
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = CTRLModel.from_pretrained('ctrl')
-        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -429,7 +429,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -438,7 +438,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -458,7 +458,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = CTRLLMHeadModel.from_pretrained('ctrl')
 
-        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 
@@ -494,7 +494,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()
             # Flatten the tokens
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                             shift_labels.view(-1))
             outputs = (loss,) + outputs

transformers/modeling_distilbert.py

@@ -415,7 +415,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertModel.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -491,7 +491,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -511,7 +511,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForMaskedLM.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, masked_lm_labels=input_ids)
         loss, prediction_scores = outputs[:2]
 
@@ -528,7 +528,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
 
         self.init_weights()
 
-        self.mlm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
+        self.mlm_loss_fct = nn.CrossEntropyLoss()
 
     def get_output_embeddings(self):
         return self.vocab_projector
@@ -581,7 +581,7 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, logits = outputs[:2]
@@ -656,7 +656,7 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForQuestionAnswering.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         start_positions = torch.tensor([1])
         end_positions = torch.tensor([3])
         outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)

transformers/modeling_encoder_decoder.py

@@ -18,9 +18,11 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import logging
 import os
+import warnings
 
 import torch
 from torch import nn
+from tqdm import trange
 
 from .modeling_auto import AutoModel, AutoModelWithLMHead
 
@@ -59,12 +61,14 @@ class PreTrainedEncoderDecoder(nn.Module):
             encoder_pretrained_model_name_or_path: information necessary to initiate the encoder. Either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/encoder``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
             decoder_pretrained_model_name_or_path: information necessary to initiate the decoder. Either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/decoder``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -117,8 +121,7 @@ class PreTrainedEncoderDecoder(nn.Module):
         kwargs_common = {
             argument: value
             for argument, value in kwargs.items()
-            if not argument.startswith("encoder_")
-            and not argument.startswith("decoder_")
+            if not argument.startswith("encoder_") and not argument.startswith("decoder_")
         }
         kwargs_decoder = kwargs_common.copy()
         kwargs_encoder = kwargs_common.copy()
@@ -164,7 +167,39 @@ class PreTrainedEncoderDecoder(nn.Module):
 
         We save the encoder' and decoder's parameters in two separate directories.
         """
+
+        # If the root output directory does not exist, create it 
+        if not os.path.exists(save_directory):
+            os.mkdir(save_directory)
+
+        # Check whether the output directory is empty or not
+        sub_directories = [directory for directory in os.listdir(save_directory)
+            if os.path.isdir(os.path.join(save_directory, directory))]
+
+        if len(sub_directories) > 0:
+            if "encoder" in sub_directories and "decoder" in sub_directories:
+                print("WARNING: there is an older version of encoder-decoder saved in" +\
+                    " the output directory. The default behaviour is to overwrite them.")
+
+            # Empty the output directory
+            for directory_to_remove in sub_directories:
+                # Remove all files into the subdirectory
+                files_to_remove = os.listdir(os.path.join(save_directory, directory_to_remove))
+                for file_to_remove in files_to_remove:
+                    os.remove(os.path.join(save_directory, directory_to_remove, file_to_remove))
+                # Remove the subdirectory itself
+                os.rmdir(os.path.join(save_directory, directory_to_remove))
+
+            assert(len(os.listdir(save_directory)) == 0) # sanity check
+
+        # Create the "encoder" directory inside the output directory and save the encoder into it
+        if not os.path.exists(os.path.join(save_directory, "encoder")):
+            os.mkdir(os.path.join(save_directory, "encoder"))
         self.encoder.save_pretrained(os.path.join(save_directory, "encoder"))
+
+        # Create the "encoder" directory inside the output directory and save the decoder into it
+        if not os.path.exists(os.path.join(save_directory, "decoder")):
+            os.mkdir(os.path.join(save_directory, "decoder"))
         self.decoder.save_pretrained(os.path.join(save_directory, "decoder"))
 
     def forward(self, encoder_input_ids, decoder_input_ids, **kwargs):
@@ -186,51 +221,56 @@ class PreTrainedEncoderDecoder(nn.Module):
                 Indices of decoder input sequence tokens in the vocabulary.
             kwargs: (`optional`) Remaining dictionary of keyword arguments.
         """
-        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
-        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
-        # that apply to the model as whole.
-        # We let the specific kwargs override the common ones in case of conflict.
+        kwargs_encoder, kwargs_decoder = self.prepare_model_kwargs(**kwargs)
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        if encoder_hidden_states is None:
+            encoder_outputs = self.encoder(encoder_input_ids, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        decoder_outputs = self.decoder(decoder_input_ids, encoder_hidden_states, **kwargs_decoder)
+
+        return decoder_outputs + encoder_outputs
+
+    @staticmethod
+    def prepare_model_kwargs(**kwargs):
+        """ Prepare the encoder and decoder's keyword arguments.
+
+        Keyword arguments come in 3 flavors:
+        - encoder-specific (prefixed by `encoder_`)
+        - decoder-specific (prefixed by `decoder_`)
+        - those that apply to the model as whole.
+
+        We let the specific kwargs override the common ones in case of
+        conflict.
+        """
         kwargs_common = {
             argument: value
             for argument, value in kwargs.items()
-            if not argument.startswith("encoder_")
-            and not argument.startswith("decoder_")
+            if not argument.startswith("encoder_") and not argument.startswith("decoder_")
         }
-        kwargs_decoder = kwargs_common.copy()
-        kwargs_encoder = kwargs_common.copy()
-        kwargs_encoder.update(
+        decoder_kwargs = kwargs_common.copy()
+        encoder_kwargs = kwargs_common.copy()
+        encoder_kwargs.update(
             {
                 argument[len("encoder_") :]: value
                 for argument, value in kwargs.items()
                 if argument.startswith("encoder_")
             }
         )
-        kwargs_decoder.update(
+        decoder_kwargs.update(
             {
                 argument[len("decoder_") :]: value
                 for argument, value in kwargs.items()
                 if argument.startswith("decoder_")
             }
         )
-
-        # Encode if needed (training, first prediction pass)
-        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
-        if encoder_hidden_states is None:
-            encoder_outputs = self.encoder(encoder_input_ids, **kwargs_encoder)
-            encoder_hidden_states = encoder_outputs[
-                0
-            ]  # output the last layer hidden state
-        else:
-            encoder_outputs = ()
-
-        # Decode
-        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
-        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get(
-            "attention_mask", None
-        )
-        decoder_outputs = self.decoder(decoder_input_ids, **kwargs_decoder)
-
-        return decoder_outputs + encoder_outputs
+        decoder_kwargs["encoder_attention_mask"] = encoder_kwargs.get("attention_mask", None)
+        return encoder_kwargs, decoder_kwargs
 
 
 class Model2Model(PreTrainedEncoderDecoder):

transformers/modeling_gpt2.py

@@ -329,7 +329,7 @@ class GPT2Model(GPT2PreTrainedModel):
         **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -345,7 +345,7 @@ class GPT2Model(GPT2PreTrainedModel):
 
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = GPT2Model.from_pretrained('gpt2')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -494,7 +494,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -503,7 +503,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -523,7 +523,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = GPT2LMHeadModel.from_pretrained('gpt2')
 
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 
@@ -557,7 +557,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()
             # Flatten the tokens
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                             shift_labels.view(-1))
             outputs = (loss,) + outputs
@@ -579,7 +579,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
         **mc_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size)``:
             Labels for computing the multiple choice classification loss.
@@ -596,7 +596,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         **mc_prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)``
             Prediction scores of the multiplechoice classification head (scores for each choice before SoftMax).
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -634,6 +634,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
     """
     def __init__(self, config):
         super(GPT2DoubleHeadsModel, self).__init__(config)
+        config.num_labels = 1
         self.transformer = GPT2Model(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
         self.multiple_choice_head = SequenceSummary(config)
@@ -667,7 +668,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         if lm_labels is not None:
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = lm_labels[..., 1:].contiguous()
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                             shift_labels.view(-1))
             outputs = (loss,) + outputs

transformers/modeling_openai.py

@@ -50,8 +50,10 @@ def load_tf_weights_in_openai_gpt(model, config, openai_checkpoint_folder_path):
 
     logger.info("Loading weights from {}".format(openai_checkpoint_folder_path))
 
-    names = json.load(open(openai_checkpoint_folder_path + '/parameters_names.json', "r", encoding='utf-8'))
-    shapes = json.load(open(openai_checkpoint_folder_path + '/params_shapes.json', "r", encoding='utf-8'))
+    with open(openai_checkpoint_folder_path + '/parameters_names.json', "r", encoding='utf-8') as names_handle:
+        names = json.load(names_handle)
+    with open(openai_checkpoint_folder_path + '/params_shapes.json', "r", encoding='utf-8') as shapes_handle:
+        shapes = json.load(shapes_handle)
     offsets = np.cumsum([np.prod(shape) for shape in shapes])
     init_params = [np.load(openai_checkpoint_folder_path + '/params_{}.npy'.format(n)) for n in range(10)]
     init_params = np.split(np.concatenate(init_params, 0), offsets)[:-1]
@@ -347,7 +349,7 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = OpenAIGPTModel.from_pretrained('openai-gpt')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -469,7 +471,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -489,7 +491,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = OpenAIGPTLMHeadModel.from_pretrained('openai-gpt')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 
@@ -521,7 +523,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()
             # Flatten the tokens
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                             shift_labels.view(-1))
             outputs = (loss,) + outputs
@@ -543,7 +545,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
         **mc_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size)``:
             Labels for computing the multiple choice classification loss.
@@ -588,6 +590,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
     def __init__(self, config):
         super(OpenAIGPTDoubleHeadsModel, self).__init__(config)
 
+        config.num_labels = 1
         self.transformer = OpenAIGPTModel(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
         self.multiple_choice_head = SequenceSummary(config)
@@ -619,7 +622,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
         if lm_labels is not None:
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = lm_labels[..., 1:].contiguous()
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                             shift_labels.view(-1))
             outputs = (loss,) + outputs

transformers/modeling_roberta.py

@@ -24,7 +24,7 @@ import torch
 import torch.nn as nn
 from torch.nn import CrossEntropyLoss, MSELoss
 
-from .modeling_bert import BertEmbeddings, BertLayerNorm, BertModel, BertPreTrainedModel, gelu, RBertClassificationHead
+from .modeling_bert import BertEmbeddings, BertLayerNorm, BertModel, BertPreTrainedModel, gelu
 from .configuration_roberta import RobertaConfig
 from .file_utils import add_start_docstrings
 
@@ -51,24 +51,44 @@ class RobertaEmbeddings(BertEmbeddings):
                                                 padding_idx=self.padding_idx)
 
     def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
-        if input_ids is not None:
-            input_shape = input_ids.size()
-        else:
-            input_shape = inputs_embeds.size()[:-1]
-
-        seq_length = input_shape[1]
-        device = input_ids.device if input_ids is not None else inputs_embeds.device
-
         if position_ids is None:
-            # Position numbers begin at padding_idx+1. Padding symbols are ignored.
-            # cf. fairseq's `utils.make_positions`
-            position_ids = torch.arange(self.padding_idx+1, seq_length+self.padding_idx+1, dtype=torch.long, device=device)
-            position_ids = position_ids.unsqueeze(0).expand(input_shape)
+            if input_ids is not None:
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = self.create_position_ids_from_input_ids(input_ids).to(input_ids.device)
+            else:
+                position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
+
         return super(RobertaEmbeddings, self).forward(input_ids,
                                                       token_type_ids=token_type_ids,
                                                       position_ids=position_ids,
                                                       inputs_embeds=inputs_embeds)
 
+    def create_position_ids_from_input_ids(self, x):
+        """ Replace non-padding symbols with their position numbers. Position numbers begin at
+        padding_idx+1. Padding symbols are ignored. This is modified from fairseq's
+        `utils.make_positions`.
+
+        :param torch.Tensor x:
+        :return torch.Tensor:
+        """
+        mask = x.ne(self.padding_idx).long()
+        incremental_indicies = torch.cumsum(mask, dim=1) * mask
+        return incremental_indicies + self.padding_idx
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """ We are provided embeddings directly. We cannot infer which are padded so just generate
+        sequential position ids.
+
+        :param torch.Tensor inputs_embeds:
+        :return torch.Tensor:
+        """
+        input_shape = inputs_embeds.size()[:-1]
+        sequence_length = input_shape[1]
+
+        position_ids = torch.arange(self.padding_idx+1, sequence_length+self.padding_idx+1, dtype=torch.long,
+                                    device=inputs_embeds.device)
+        return position_ids.unsqueeze(0).expand(input_shape)
+
 
 ROBERTA_START_DOCSTRING = r"""    The RoBERTa model was proposed in
     `RoBERTa: A Robustly Optimized BERT Pretraining Approach`_
@@ -168,7 +188,7 @@ class RobertaModel(BertModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = RobertaModel.from_pretrained('roberta-base')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -196,7 +216,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
         **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -216,7 +236,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = RobertaForMaskedLM.from_pretrained('roberta-base')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, masked_lm_labels=input_ids)
         loss, prediction_scores = outputs[:2]
 
@@ -250,7 +270,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
         outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
 
         if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss_fct = CrossEntropyLoss()
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             outputs = (masked_lm_loss,) + outputs
 
@@ -307,7 +327,7 @@ class RobertaForSequenceClassification(BertPreTrainedModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = RobertaForSequenceClassification.from_pretrained('roberta-base')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, logits = outputs[:2]
@@ -435,8 +455,8 @@ class RobertaForMultipleChoice(BertPreTrainedModel):
 
         self.init_weights()
 
-    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
-                labels=None):
+    def forward(self, input_ids=None, token_type_ids=None, attention_mask=None, labels=None,
+                position_ids=None, head_mask=None, inputs_embeds=None):
         num_choices = input_ids.shape[1]
 
         flat_input_ids = input_ids.view(-1, input_ids.size(-1))
@@ -460,86 +480,6 @@ class RobertaForMultipleChoice(BertPreTrainedModel):
 
         return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
 
-@add_start_docstrings("""Roberta Model transformer with an R-BERT head on top, making use of
-        entity offsets for relationship classification. See the details in the R-Bert paper https://arxiv.org/pdf/1905.08284.pdf""",
-    ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
-class RobertaForRelationshipClassification(BertPreTrainedModel):
-    r"""
-        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-        Labels for computing the sequence classification/regression loss.
-        Indices should be in ``[0, ..., config.num_labels - 1]``.
-        If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
-        If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
-
-    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
-            Classification (or regression if config.num_labels==1) loss.
-        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
-            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
-            of shape ``(batch_size, sequence_length, hidden_size)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
-
-    Examples::
-
-        test_string = "The $ cat $ sat on the # mat # ."
-        ent1_sep_token = '$'
-        ent2_sep_token = '#'
-
-        model_name = 'roberta-large'
-
-        tokenizer = RobertaTokenizer.from_pretrained(model_name)
-        ent1_sep_token_id = tokenizer.encode(ent1_sep_token, text_pair=None, add_special_tokens=True, add_prefix_space=True)[1]
-        ent2_sep_token_id = tokenizer.encode(ent2_sep_token, text_pair=None, add_special_tokens=True, add_prefix_space=True)[1]
-        config = RobertaConfig.from_pretrained(model_name,entity_1_token_id=ent1_sep_token_id, entity_2_token_id=ent2_sep_token_id)
-        model = RobertaForRelationshipClassification.from_pretrained(model_name, config=config)
-        input_ids = torch.tensor(tokenizer.encode(test_string,add_special_tokens=True)).unsqueeze(0)  # Batch size 1
-        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
-        outputs = model(input_ids, labels=labels)
-        loss, logits = outputs[:2]
-    """
-    config_class = RobertaConfig
-    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
-    base_model_prefix = "roberta"
-
-    def __init__(self, config):
-        super(RobertaForRelationshipClassification, self).__init__(config)
-
-        self.num_labels = config.num_labels
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.roberta = RobertaModel(config)
-        self.rbert = RBertClassificationHead(config)
-
-        self.init_weights()
-
-    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
-                labels=None):
-
-        outputs = self.roberta(input_ids,
-                               attention_mask=attention_mask,
-                               token_type_ids=token_type_ids,
-                               position_ids=position_ids,
-                               head_mask=head_mask,
-                               inputs_embeds=inputs_embeds)
-        last_hidden_states = outputs[0]
-        logits = self.rbert(input_ids, last_hidden_states)
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
-
-        if labels is not None:
-            if self.num_labels == 1:
-                #  We are doing regression
-                loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1), labels.view(-1))
-            else:
-                loss_fct = CrossEntropyLoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
-            outputs = (loss,) + outputs
-
-        return outputs  # (loss), logits, (hidden_states), (attentions)
 
 @add_start_docstrings("""Roberta Model with a token classification head on top (a linear layer on top of
     the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
@@ -635,3 +575,89 @@ class RobertaClassificationHead(nn.Module):
         x = self.dropout(x)
         x = self.out_proj(x)
         return x
+
+
+@add_start_docstrings("""Roberta Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
+    the hidden-states output to compute `span start logits` and `span end logits`). """,
+    ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
+class RobertaForQuestionAnswering(BertPreTrainedModel):
+    r"""
+        **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+        **end_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        **start_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-start scores (before SoftMax).
+        **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-end scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+    Examples::
+        tokenizer = RobertaTokenizer.from_pretrained('roberta-large')
+        model = RobertaForQuestionAnswering.from_pretrained('roberta-large')
+        question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        input_ids = tokenizer.encode(question, text)
+        start_scores, end_scores = model(torch.tensor([input_ids]))
+        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)
+        answer = ' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1])
+    """
+    config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "roberta"
+
+    def __init__(self, config):
+        super(RobertaForQuestionAnswering, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.roberta = RobertaModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                start_positions=None, end_positions=None):
+
+        outputs = self.roberta(input_ids,
+                               attention_mask=attention_mask,
+                               token_type_ids=token_type_ids,
+                               position_ids=position_ids,
+                               head_mask=head_mask)
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        outputs = (start_logits, end_logits,) + outputs[2:]
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions.clamp_(0, ignored_index)
+            end_positions.clamp_(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+            outputs = (total_loss,) + outputs
+
+        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)

transformers/modeling_t5.py

@@ -0,0 +1,886 @@
+# coding=utf-8
+# Copyright 2018 Mesh TensorFlow authors, T5 Authors and HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch T5 model. """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import math
+import os
+import sys
+import copy
+import itertools
+from io import open
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from .modeling_utils import PreTrainedModel, prune_linear_layer
+from .configuration_t5 import T5Config
+from .file_utils import add_start_docstrings, DUMMY_INPUTS, DUMMY_MASK
+
+logger = logging.getLogger(__name__)
+
+####################################################
+# This dict contrains shortcut names and associated url
+# for the pretrained weights provided with the models
+####################################################
+T5_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-small-pytorch_model.bin",
+    't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-base-pytorch_model.bin",
+    't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-large-pytorch_model.bin",
+    't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-3b-pytorch_model.bin",
+    't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-11b-pytorch_model.bin",
+}
+
+####################################################
+# This is a conversion method from TF 1.0 to PyTorch
+# More details: https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28
+####################################################
+def load_tf_weights_in_t5(model, config, tf_checkpoint_path):
+    """ Load tf checkpoints in a pytorch model.
+    """
+    try:
+        import re
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions.")
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    tf_weights = {}
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        tf_weights[name] = array
+
+    for txt_name in names:
+        name = txt_name.split('/')
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(n in ["adam_v", "adam_m", "global_step"] for n in name):
+            logger.info("Skipping {}".format("/".join(name)))
+            tf_weights.pop(txt_name, None)
+            continue
+        if '_slot_' in name[-1]:
+            logger.info("Skipping {}".format("/".join(name)))
+            tf_weights.pop(txt_name, None)
+            continue
+        pointer = model
+        array = tf_weights[txt_name]
+        for m_name in name:
+            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
+                l = re.split(r'_(\d+)', m_name)
+            else:
+                l = [m_name]
+            if l[0] in ['kernel', 'scale', 'embedding']:
+                pointer = getattr(pointer, 'weight')
+            # elif l[0] == 'scale':
+            #     pointer = getattr(pointer, 'weight')
+            # elif l[0] == 'output_bias' or l[0] == 'beta':
+            #     pointer = getattr(pointer, 'bias')
+            # elif l[0] == 'squad':
+            #     pointer = getattr(pointer, 'classifier')
+            else:
+                try:
+                    pointer = getattr(pointer, l[0])
+                except AttributeError:
+                    logger.info("Skipping {}".format("/".join(name)))
+                    continue
+            if len(l) >= 2:
+                num = int(l[1])
+                pointer = pointer[num]
+        if l[0] not in ['kernel', 'scale', 'embedding']:
+            pointer = getattr(pointer, 'weight')
+        if l[0] != 'embedding':
+            logger.info("Transposing numpy weight of shape {} for {}".format(array.shape, name))
+            array = np.transpose(array)
+        try:
+            assert pointer.shape == array.shape
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info("Initialize PyTorch weight {}".format(name))
+        pointer.data = torch.from_numpy(array.astype(np.float32))
+        tf_weights.pop(txt_name, None)
+
+    logger.info("Weights not copied to PyTorch model: {}".format(', '.join(tf_weights.keys())))
+    # logger.info("Weights not copied to PyTorch model: {}".format(', '.join(tf_weights.keys())))
+    return model
+
+
+####################################################
+# PyTorch Models are constructed by sub-classing
+# - torch.nn.Module for the layers and
+# - PreTrainedModel for the models (it-self a sub-class of torch.nn.Module)
+####################################################
+
+class T5LayerNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """ Construct a layernorm module in the T5 style
+            No bias and no substraction of mean.
+        """
+        super(T5LayerNorm, self).__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, x):
+        variance = x.pow(2).mean(-1, keepdim=True)
+        x = x / torch.sqrt(variance + self.variance_epsilon)
+        return self.weight * x
+
+
+class T5DenseReluDense(nn.Module):
+    def __init__(self, config):
+        super(T5DenseReluDense, self).__init__()
+        self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)
+        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states):
+        h = self.wi(hidden_states)
+        h = F.relu(h)
+        h = self.dropout(h)
+        h = self.wo(h)
+        return h
+
+
+class T5LayerFF(nn.Module):
+    def __init__(self, config):
+        super(T5LayerFF, self).__init__()
+        self.DenseReluDense = T5DenseReluDense(config)
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states):
+        norm_x = self.layer_norm(hidden_states)
+        y = self.DenseReluDense(norm_x)
+        layer_output = hidden_states + self.dropout(y)
+        return layer_output
+
+
+class T5Attention(nn.Module):
+    NEW_ID = itertools.count()
+
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5Attention, self).__init__()
+        self.layer_id = next(T5Attention.NEW_ID)
+        self.is_decoder = config.is_decoder
+        self.has_relative_attention_bias = has_relative_attention_bias
+
+        self.output_attentions = config.output_attentions
+        self.relative_attention_num_buckets = config.relative_attention_num_buckets
+        self.d_model = config.d_model
+        self.d_kv = config.d_kv
+        self.n_heads = config.num_heads
+        self.dropout = config.dropout_rate
+        self.inner_dim = self.n_heads * self.d_kv
+
+        # Mesh TensorFlow initialization to avoid scaling before softmax
+        self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)
+
+        if self.has_relative_attention_bias:
+            self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        mask = torch.ones(self.n_heads, self.d_kv)
+        heads = set(heads) - self.pruned_heads
+        for head in heads:
+            head -= sum(1 if h < head else 0 for h in self.pruned_heads)
+            mask[head] = 0
+        mask = mask.view(-1).contiguous().eq(1)
+        index = torch.arange(len(mask))[mask].long()
+        # Prune linear layers
+        self.q = prune_linear_layer(self.q, index)
+        self.k = prune_linear_layer(self.k, index)
+        self.v = prune_linear_layer(self.v, index)
+        self.o = prune_linear_layer(self.o, index, dim=1)
+        # Update hyper params
+        self.n_heads = self.n_heads - len(heads)
+        self.inner_dim = self.d_kv * self.n_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    @staticmethod
+    def _relative_position_bucket(relative_position,
+                                  bidirectional=True,
+                                  num_buckets=32,
+                                  max_distance=128):
+        """
+        Adapted from Mesh Tensorflow:
+        https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
+
+        Translate relative position to a bucket number for relative attention.
+        The relative position is defined as memory_position - query_position, i.e.
+        the distance in tokens from the attending position to the attended-to
+        position.  If bidirectional=False, then positive relative positions are
+        invalid.
+        We use smaller buckets for small absolute relative_position and larger buckets
+        for larger absolute relative_positions.  All relative positions >=max_distance
+        map to the same bucket.  All relative positions <=-max_distance map to the
+        same bucket.  This should allow for more graceful generalization to longer
+        sequences than the model has been trained on.
+        Args:
+            relative_position: an int32 Tensor
+            bidirectional: a boolean - whether the attention is bidirectional
+            num_buckets: an integer
+            max_distance: an integer
+        Returns:
+            a Tensor with the same shape as relative_position, containing int32
+            values in the range [0, num_buckets)
+        """
+        ret = 0
+        n = -relative_position
+        if bidirectional:
+            num_buckets //= 2
+            ret += (n < 0).to(torch.long) * num_buckets  # mtf.to_int32(mtf.less(n, 0)) * num_buckets
+            n = torch.abs(n)
+        else:
+            n = torch.max(n, torch.zeros_like(n))
+        # now n is in the range [0, inf)
+
+        # half of the buckets are for exact increments in positions
+        max_exact = num_buckets // 2
+        is_small = (n < max_exact)
+
+        # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
+        val_if_large = max_exact + (
+            torch.log(n.float() / max_exact)
+            / math.log(max_distance / max_exact) * (num_buckets - max_exact)).to(torch.long)
+        val_if_large = torch.min(val_if_large, torch.full_like(val_if_large, num_buckets - 1))
+
+        ret += torch.where(is_small, n, val_if_large)
+        return ret
+
+    def compute_bias(self, qlen, klen):
+        """ Compute binned relative position bias """
+        context_position = torch.arange(qlen, dtype=torch.long)[:, None]
+        memory_position = torch.arange(klen, dtype=torch.long)[None, :]
+        relative_position = memory_position - context_position  # shape (qlen, klen)
+        rp_bucket = self._relative_position_bucket(relative_position,  # shape (qlen, klen)
+                                                   bidirectional=not self.is_decoder,
+                                                   num_buckets=self.relative_attention_num_buckets)
+        values = self.relative_attention_bias(rp_bucket)  # shape (qlen, klen, num_heads)
+        values = values.permute([2, 0, 1]).unsqueeze(0) # shape (1, num_heads, qlen, klen)
+        return values
+
+    def forward(self, input, mask=None, kv=None, position_bias=None, cache=None, head_mask=None):
+        """
+        Self-attention (if kv is None) or attention over source sentence (provided by kv).
+        """
+        # Input is (bs, qlen, dim)
+        # Mask is (bs, klen) (non-causal) or (bs, klen, klen)
+        bs, qlen, dim = input.size()
+        if kv is None:
+            klen = qlen if cache is None else cache['slen'] + qlen
+        else:
+            klen = kv.size(1)
+
+        def shape(x):
+            """  projection """
+            return x.view(bs, -1, self.n_heads, self.d_kv).transpose(1, 2)
+
+        def unshape(x):
+            """  compute context """
+            return x.transpose(1, 2).contiguous().view(bs, -1, self.inner_dim)
+
+        q = shape(self.q(input))                                          # (bs, n_heads, qlen, dim_per_head)
+        if kv is None:
+            k = shape(self.k(input))                                      # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v(input))                                      # (bs, n_heads, qlen, dim_per_head)
+        elif cache is None or self.layer_id not in cache:
+            k = v = kv
+            k = shape(self.k(k))                                          # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v(v))                                          # (bs, n_heads, qlen, dim_per_head)
+
+        if cache is not None:
+            if self.layer_id in cache:
+                if kv is None:
+                    k_, v_ = cache[self.layer_id]
+                    k = torch.cat([k_, k], dim=2)                             # (bs, n_heads, klen, dim_per_head)
+                    v = torch.cat([v_, v], dim=2)                             # (bs, n_heads, klen, dim_per_head)
+                else:
+                    k, v = cache[self.layer_id]
+            cache[self.layer_id] = (k, v)
+
+        # q = q / math.sqrt(dim_per_head)                                     # No scaling in T5
+        scores = torch.einsum('bnqd,bnkd->bnqk', q, k)                        # (bs, n_heads, qlen, klen)
+
+        if position_bias is None:
+            if not self.has_relative_attention_bias:
+                raise ValueError("No position_bias provided and no weights to compute position_bias")
+            position_bias = self.compute_bias(qlen, klen)
+            if mask is not None:
+                position_bias = position_bias + mask                          # (bs, n_heads, qlen, klen)
+
+        scores += position_bias
+        weights = F.softmax(scores.float(), dim=-1).type_as(scores)           # (bs, n_heads, qlen, klen)
+        weights = F.dropout(weights, p=self.dropout, training=self.training)  # (bs, n_heads, qlen, klen)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            weights = weights * head_mask
+
+        context = torch.matmul(weights, v)                                    # (bs, n_heads, qlen, dim_per_head)
+        context = unshape(context)                                            # (bs, qlen, dim)
+
+        context = self.o(context)
+
+        outputs = (context,)
+        if self.output_attentions:
+            outputs = outputs + (weights,)
+        if self.has_relative_attention_bias:
+            outputs = outputs + (position_bias,)
+        return outputs
+
+
+class T5LayerSelfAttention(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5LayerSelfAttention, self).__init__()
+        self.SelfAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states, attention_mask=None, position_bias=None, head_mask=None):
+        norm_x = self.layer_norm(hidden_states)
+        attention_output = self.SelfAttention(norm_x,
+                                              mask=attention_mask,
+                                              position_bias=position_bias,
+                                              head_mask=head_mask)
+        y = attention_output[0]
+        layer_output = hidden_states + self.dropout(y)
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class T5LayerCrossAttention(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5LayerCrossAttention, self).__init__()
+        self.EncDecAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states, kv, attention_mask=None, position_bias=None, head_mask=None):
+        norm_x = self.layer_norm(hidden_states)
+        attention_output = self.EncDecAttention(norm_x,
+                                                mask=attention_mask,
+                                                kv=kv,
+                                                position_bias=position_bias,
+                                                head_mask=head_mask)
+        y = attention_output[0]
+        layer_output = hidden_states + self.dropout(y)
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class T5Block(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5Block, self).__init__()
+        self.is_decoder = config.is_decoder
+        self.layer = nn.ModuleList()
+        self.layer.append(T5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias))
+        if self.is_decoder:
+            self.layer.append(T5LayerCrossAttention(config, has_relative_attention_bias=has_relative_attention_bias))
+            self.layer.append(T5LayerFF(config))
+        else:
+            self.layer.append(T5LayerFF(config))
+
+    def forward(self, hidden_states, attention_mask=None, position_bias=None,
+                encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None,
+                head_mask=None):
+        self_attention_outputs = self.layer[0](hidden_states,
+                                                attention_mask=attention_mask,
+                                                position_bias=position_bias,
+                                                head_mask=head_mask)
+        hidden_states = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # Keep self-attention outputs and relative position weights
+
+        if not self.is_decoder:
+            hidden_states = self.layer[1](hidden_states)
+        else:
+            cross_attention_outputs = self.layer[1](hidden_states,
+                                                    kv=encoder_hidden_states,
+                                                    attention_mask=encoder_attention_mask,
+                                                    position_bias=encoder_decoder_position_bias,
+                                                    head_mask=head_mask)
+            hidden_states = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:]  # Keep cross-attention outputs and relative position weights
+            hidden_states = self.layer[2](hidden_states)
+
+        outputs = (hidden_states,) + outputs  # add attentions if we output them
+        return outputs  # hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+
+
+class T5PreTrainedModel(PreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = T5Config
+    pretrained_model_archive_map = T5_PRETRAINED_MODEL_ARCHIVE_MAP
+    load_tf_weights = load_tf_weights_in_t5
+    base_model_prefix = "transformer"
+
+    @property
+    def dummy_inputs(self):
+        input_ids = torch.tensor(DUMMY_INPUTS)
+        input_mask = torch.tensor(DUMMY_MASK)
+        dummy_inputs = {'decoder_input_ids': input_ids,
+                        'encoder_input_ids': input_ids,
+                        'decoder_attention_mask': input_mask}
+        return dummy_inputs
+
+    def _init_weights(self, module):
+        """ Initialize the weights """
+        factor = self.config.initializer_factor  # Used for testing weights initialization
+        if isinstance(module, T5LayerNorm):
+            module.weight.data.fill_(factor*1.0)
+        elif isinstance(module, (T5Model, T5WithLMHeadModel)):
+            # Mesh TensorFlow embeddings initialization
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624
+            module.shared.weight.data.normal_(mean=0.0, std=factor*1.0)
+        elif isinstance(module, T5DenseReluDense):
+            # Mesh TensorFlow FF initialization
+            # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56
+            # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89
+            module.wi.weight.data.normal_(mean=0.0, std=factor*((self.config.d_model) ** -0.5))
+            if hasattr(module.wi, 'bias') and module.wi.bias is not None:
+                module.wi.bias.data.zero_()
+            module.wo.weight.data.normal_(mean=0.0, std=factor*((self.config.d_ff) ** -0.5))
+            if hasattr(module.wo, 'bias') and module.wo.bias is not None:
+                module.wo.bias.data.zero_()
+        elif isinstance(module, T5Attention):
+            # Mesh TensorFlow attention initialization to avoid scaling before softmax
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136
+            d_model = self.config.d_model
+            d_kv = self.config.d_kv
+            n_heads = self.config.num_heads
+            module.q.weight.data.normal_(mean=0.0, std=factor*((d_model * d_kv) ** -0.5))
+            module.k.weight.data.normal_(mean=0.0, std=factor*(d_model ** -0.5))
+            module.v.weight.data.normal_(mean=0.0, std=factor*(d_model ** -0.5))
+            module.o.weight.data.normal_(mean=0.0, std=factor*((n_heads * d_kv) ** -0.5))
+            if module.has_relative_attention_bias:
+                module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor*((d_model) ** -0.5))
+
+
+class T5Stack(T5PreTrainedModel):
+    def __init__(self, config):
+        super(T5Stack, self).__init__(config)
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.is_decoder = config.is_decoder
+
+        self.block = nn.ModuleList([T5Block(config, has_relative_attention_bias=bool(i == 0))
+                                    for i in range(config.num_layers)])
+        self.final_layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+        self.init_weights()
+
+    def forward(self,
+                hidden_states,
+                attention_mask=None,
+                encoder_hidden_states=None,
+                encoder_attention_mask=None,
+                head_mask=None):
+
+        batch_size, seq_length = hidden_states.shape[0], hidden_states.shape[1]
+        if attention_mask is None:
+            attention_mask = torch.ones(batch_size, seq_length).to(hidden_states.device)
+        if self.is_decoder and encoder_attention_mask is None:
+            encoder_seq_length = encoder_hidden_states.shape[1]
+            encoder_attention_mask = torch.ones(batch_size, encoder_seq_length).to(hidden_states.device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+        # Provided a padding mask of dimensions [batch_size, seq_length]
+        # - if the model is a decoder, apply a causal mask in addition to the padding mask
+        # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if self.config.is_decoder:
+                seq_ids = torch.arange(seq_length, device=hidden_states.device)
+                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                causal_mask = causal_mask.to(attention_mask)
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -1e9 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+
+        # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
+        # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
+        # extended_attention_mask = (extended_attention_mask == extended_attention_mask.transpose(-1, -2))
+
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -1e9
+
+        if self.is_decoder:
+            # If a 2D ou 3D attention mask is provided for the cross-attention
+            # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+            if encoder_attention_mask.dim() == 3:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+            if encoder_attention_mask.dim() == 2:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+
+            # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
+            # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
+            # encoder_extended_attention_mask = (encoder_extended_attention_mask == encoder_extended_attention_mask.transpose(-1, -2))
+
+            encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e9
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.num_layers, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.num_layers
+
+        all_hidden_states = ()
+        all_attentions = ()
+        position_bias = None
+        encoder_decoder_position_bias = None
+
+        hidden_states = self.dropout(hidden_states)
+        for i, layer_module in enumerate(self.block):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_outputs = layer_module(hidden_states,
+                                         attention_mask=extended_attention_mask,
+                                         position_bias=position_bias,
+                                         encoder_hidden_states=encoder_hidden_states,
+                                         encoder_attention_mask=encoder_extended_attention_mask,
+                                         encoder_decoder_position_bias=encoder_decoder_position_bias,
+                                         head_mask=head_mask[i])
+            # layer_outputs is a tuple with:
+            # hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+            hidden_states = layer_outputs[0]
+            if i == 0:
+                # We share the position biases between the layers - the first layer store them
+                # layer_outputs = hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+                position_bias = layer_outputs[2 if self.output_attentions else 1]
+                if self.is_decoder:
+                    encoder_decoder_position_bias = layer_outputs[4 if self.output_attentions else 2]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)  # We keep only self-attention weights for now
+
+        hidden_states = self.final_layer_norm(hidden_states)
+        layer_output = self.dropout(hidden_states)
+
+        # Add last layer
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
+
+
+T5_START_DOCSTRING = r"""    The T5 model was proposed in
+    `Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`_
+    by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.
+    It's an encoder decoder transformer pre-trained in a text-to-text denoising generative setting.
+
+    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+    refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+    .. _`Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`:
+        https://arxiv.org/abs/1910.10683
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~transformers.T5Config`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+T5_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, T5 input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+
+            T5 is a model with relative position embeddings so you should be able to pad the inputs on
+            the right or the left.
+
+            Indices can be obtained using :class:`transformers.T5Tokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+"""
+
+@add_start_docstrings("The bare T5 Model transformer outputting raw hidden-states"
+                      "without any specific head on top.",
+                      T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+class T5Model(T5PreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = T5Model.from_pretrained('t5-small')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids=input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config):
+        super(T5Model, self).__init__(config)
+        self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+        encoder_config = copy.deepcopy(config)
+        self.encoder = T5Stack(encoder_config)
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        self.decoder = T5Stack(decoder_config)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, new_embeddings):
+        self.shared = new_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def forward(self, **kwargs):
+        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
+        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
+        # that apply to the model as whole.
+        # We let the specific kwargs override the common ones in case of conflict.
+        kwargs_common = dict((k, v) for k, v in kwargs.items()
+                             if not k.startswith("encoder_") and not k.startswith("decoder_"))
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
+        kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        encoder_attention_mask = kwargs_encoder.get("attention_mask", None)
+        if encoder_hidden_states is None:
+            # Convert encoder inputs in embeddings if needed
+            hidden_states = kwargs_encoder.pop("inputs_embeds", None)
+            if hidden_states is None:
+                encoder_inputs_ids = kwargs_encoder.pop("input_ids")
+                hidden_states = self.shared(encoder_inputs_ids)  # Convert inputs in embeddings
+
+            if encoder_attention_mask is not None:
+                # Apply masking
+                encoder_attention_mask = (encoder_attention_mask != 0).to(hidden_states)
+                hidden_states = hidden_states * encoder_attention_mask.unsqueeze(-1)
+
+            encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        # Convert decoder inputs in embeddings if needed
+        hidden_states = kwargs_decoder.pop("inputs_embeds", None)
+        if hidden_states is None:
+            decoder_inputs_ids = kwargs_decoder.pop("input_ids")
+            hidden_states = self.shared(decoder_inputs_ids)
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = encoder_attention_mask
+        decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
+
+        return decoder_outputs + encoder_outputs
+
+
+@add_start_docstrings("""T5 Model with a `language modeling` head on top. """,
+    T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+class T5WithLMHeadModel(T5PreTrainedModel):
+    r"""
+        **lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = T5WithLMHeadModel.from_pretrained('t5-small')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids=input_ids, lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(T5WithLMHeadModel, self).__init__(config)
+        self.model_dim = config.d_model
+
+        self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+        encoder_config = copy.deepcopy(config)
+        self.encoder = T5Stack(encoder_config)
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        self.decoder = T5Stack(decoder_config)
+
+        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, new_embeddings):
+        self.shared = new_embeddings
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def forward(self, **kwargs):
+        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
+        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
+        # that apply to the model as whole.
+        # We let the specific kwargs override the common ones in case of conflict.
+
+        lm_labels = kwargs.pop('decoder_lm_labels', None)
+
+        kwargs_common = dict((k, v) for k, v in kwargs.items()
+                             if not k.startswith("encoder_") and not k.startswith("decoder_"))
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
+        kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        if encoder_hidden_states is None:
+            # Convert encoder inputs in embeddings if needed
+            hidden_states = kwargs_encoder.pop("inputs_embeds", None)
+            if hidden_states is None:
+                encoder_inputs_ids = kwargs_encoder.pop("input_ids")
+                hidden_states = self.shared(encoder_inputs_ids)  # Convert inputs in embeddings
+
+            encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        # Convert decoder inputs in embeddings if needed
+        hidden_states = kwargs_decoder.pop("inputs_embeds", None)
+        if hidden_states is None:
+            decoder_inputs_ids = kwargs_decoder.pop("input_ids")
+            hidden_states = self.shared(decoder_inputs_ids)
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get("attention_mask", None)
+        decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
+
+        sequence_output = decoder_outputs[0]
+        # Rescale output before projecting on vocab
+        # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
+        sequence_output = sequence_output * (self.model_dim ** -0.5)
+        lm_logits = self.lm_head(sequence_output)
+
+        decoder_outputs = (lm_logits,) + decoder_outputs[1:]  # Add hidden states and attention if they are here
+        if lm_labels is not None:
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = lm_labels[..., 1:].contiguous()
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
+                            shift_labels.view(-1))
+            decoder_outputs = (loss,) + decoder_outputs  # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666
+
+        return decoder_outputs + encoder_outputs

transformers/modeling_tf_albert.py

@@ -587,8 +587,8 @@ class TFAlbertModel(TFAlbertPreTrainedModel):
         import tensorflow as tf
         from transformers import AlbertTokenizer, TFAlbertModel
 
-        tokenizer = AlbertTokenizer.from_pretrained('bert-base-uncased')
-        model = TFAlbertModel.from_pretrained('bert-base-uncased')
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v1')
+        model = TFAlbertModel.from_pretrained('albert-base-v1')
         input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple

transformers/modeling_tf_auto.py

@@ -18,21 +18,48 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import logging
 
-from .modeling_tf_bert import TFBertModel, TFBertForMaskedLM, TFBertForSequenceClassification, TFBertForQuestionAnswering
-from .modeling_tf_openai import TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel
-from .modeling_tf_gpt2 import TFGPT2Model, TFGPT2LMHeadModel
-from .modeling_tf_transfo_xl import TFTransfoXLModel, TFTransfoXLLMHeadModel
-from .modeling_tf_xlnet import TFXLNetModel, TFXLNetLMHeadModel, TFXLNetForSequenceClassification, TFXLNetForQuestionAnsweringSimple
-from .modeling_tf_xlm import TFXLMModel, TFXLMWithLMHeadModel, TFXLMForSequenceClassification, TFXLMForQuestionAnsweringSimple
-from .modeling_tf_roberta import TFRobertaModel, TFRobertaForMaskedLM, TFRobertaForSequenceClassification
-from .modeling_tf_distilbert import TFDistilBertModel, TFDistilBertForQuestionAnswering, TFDistilBertForMaskedLM, TFDistilBertForSequenceClassification
-from .modeling_tf_ctrl import TFCTRLModel, TFCTRLLMHeadModel
+from .configuration_auto import (BertConfig, CTRLConfig, DistilBertConfig,
+                                 GPT2Config, OpenAIGPTConfig, RobertaConfig,
+                                 TransfoXLConfig, XLMConfig, XLNetConfig)
+
+from .modeling_tf_bert import TFBertModel, TFBertForMaskedLM, TFBertForSequenceClassification, \
+    TFBertForQuestionAnswering, TFBertForTokenClassification, TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_openai import TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel, TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_gpt2 import TFGPT2Model, TFGPT2LMHeadModel, TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_transfo_xl import TFTransfoXLModel, TFTransfoXLLMHeadModel, TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_xlnet import TFXLNetModel, TFXLNetLMHeadModel, TFXLNetForSequenceClassification, \
+    TFXLNetForQuestionAnsweringSimple, TFXLNetForTokenClassification, TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_xlm import TFXLMModel, TFXLMWithLMHeadModel, TFXLMForSequenceClassification, \
+    TFXLMForQuestionAnsweringSimple, TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_roberta import TFRobertaModel, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, \
+    TFRobertaForTokenClassification, TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_distilbert import TFDistilBertModel, TFDistilBertForQuestionAnswering, TFDistilBertForMaskedLM, TFDistilBertForSequenceClassification, TFDistilBertForTokenClassification, TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_ctrl import TFCTRLModel, TFCTRLLMHeadModel, TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_albert import TFAlbertModel, TFAlbertForMaskedLM, TFAlbertForSequenceClassification, TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_t5 import TFT5Model, TFT5WithLMHeadModel, TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP
 
 from .file_utils import add_start_docstrings
 
 logger = logging.getLogger(__name__)
 
 
+TF_ALL_PRETRAINED_MODEL_ARCHIVE_MAP = dict((key, value)
+    for pretrained_map in [
+        TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ]
+    for key, value, in pretrained_map.items())
+
+
 class TFAutoModel(object):
     r"""
         :class:`~transformers.TFAutoModel` is a generic model class
@@ -45,6 +72,7 @@ class TFAutoModel(object):
 
         The base model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5Model (T5 model)
             - contains `distilbert`: TFDistilBertModel (DistilBERT model)
             - contains `roberta`: TFRobertaModel (RoBERTa model)
             - contains `bert`: TFBertModel (Bert model)
@@ -59,7 +87,50 @@ class TFAutoModel(object):
     """
     def __init__(self):
         raise EnvironmentError("TFAutoModel is designed to be instantiated "
-            "using the `TFAutoModel.from_pretrained(pretrained_model_name_or_path)` method.")
+            "using the `TFAutoModel.from_pretrained(pretrained_model_name_or_path)` or "
+            "`TFAutoModel.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: TFDistilBertModel (DistilBERT model)
+                    - isInstance of `roberta` configuration class: TFRobertaModel (RoBERTa model)
+                    - isInstance of `bert` configuration class: TFBertModel (Bert model)
+                    - isInstance of `openai-gpt` configuration class: TFOpenAIGPTModel (OpenAI GPT model)
+                    - isInstance of `gpt2` configuration class: TFGPT2Model (OpenAI GPT-2 model)
+                    - isInstance of `ctrl` configuration class: TFCTRLModel (Salesforce CTRL  model)
+                    - isInstance of `transfo-xl` configuration class: TFTransfoXLModel (Transformer-XL model)
+                    - isInstance of `xlnet` configuration class: TFXLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: TFXLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = TFAutoModel.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, DistilBertConfig):
+            return TFDistilBertModel(config)
+        elif isinstance(config, RobertaConfig):
+            return TFRobertaModel(config)
+        elif isinstance(config, BertConfig):
+            return TFBertModel(config)
+        elif isinstance(config, OpenAIGPTConfig):
+            return TFOpenAIGPTModel(config)
+        elif isinstance(config, GPT2Config):
+            return TFGPT2Model(config)
+        elif isinstance(config, TransfoXLConfig):
+            return TFTransfoXLModel(config)
+        elif isinstance(config, XLNetConfig):
+            return TFXLNetModel(config)
+        elif isinstance(config, XLMConfig):
+            return TFXLMModel(config)
+        elif isinstance(config, CTRLConfig):
+            return TFCTRLModel(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -68,6 +139,7 @@ class TFAutoModel(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5Model (T5 model)
             - contains `distilbert`: TFDistilBertModel (DistilBERT model)
             - contains `roberta`: TFRobertaModel (RoBERTa model)
             - contains `bert`: TFTFBertModel (Bert model)
@@ -81,6 +153,7 @@ class TFAutoModel(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
 
@@ -136,8 +209,12 @@ class TFAutoModel(object):
             model = TFAutoModel.from_pretrained('./pt_model/bert_pytorch_model.bin', from_pt=True, config=config)
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return TFT5Model.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return TFDistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return TFAlbertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return TFRobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -156,7 +233,7 @@ class TFAutoModel(object):
             return TFCTRLModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
+                         "'distilbert', 'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
                          "'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))
 
 
@@ -172,6 +249,7 @@ class TFAutoModelWithLMHead(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5WithLMHeadModel (T5 model)
             - contains `distilbert`: TFDistilBertForMaskedLM (DistilBERT model)
             - contains `roberta`: TFRobertaForMaskedLM (RoBERTa model)
             - contains `bert`: TFBertForMaskedLM (Bert model)
@@ -186,7 +264,50 @@ class TFAutoModelWithLMHead(object):
     """
     def __init__(self):
         raise EnvironmentError("TFAutoModelWithLMHead is designed to be instantiated "
-            "using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
+            "using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` or "
+            "`TFAutoModelWithLMHead.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `openai-gpt` configuration class: OpenAIGPTModel (OpenAI GPT model)
+                    - isInstance of `gpt2` configuration class: GPT2Model (OpenAI GPT-2 model)
+                    - isInstance of `ctrl` configuration class: CTRLModel (Salesforce CTRL  model)
+                    - isInstance of `transfo-xl` configuration class: TransfoXLModel (Transformer-XL model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: XLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelWithLMHead.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, DistilBertConfig):
+            return TFDistilBertForMaskedLM(config)
+        elif isinstance(config, RobertaConfig):
+            return TFRobertaForMaskedLM(config)
+        elif isinstance(config, BertConfig):
+            return TFBertForMaskedLM(config)
+        elif isinstance(config, OpenAIGPTConfig):
+            return TFOpenAIGPTLMHeadModel(config)
+        elif isinstance(config, GPT2Config):
+            return TFGPT2LMHeadModel(config)
+        elif isinstance(config, TransfoXLConfig):
+            return TFTransfoXLLMHeadModel(config)
+        elif isinstance(config, XLNetConfig):
+            return TFXLNetLMHeadModel(config)
+        elif isinstance(config, XLMConfig):
+            return TFXLMWithLMHeadModel(config)
+        elif isinstance(config, CTRLConfig):
+            return TFCTRLLMHeadModel(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -198,6 +319,7 @@ class TFAutoModelWithLMHead(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5WithLMHeadModel (T5 model)
             - contains `distilbert`: TFDistilBertForMaskedLM (DistilBERT model)
             - contains `roberta`: TFRobertaForMaskedLM (RoBERTa model)
             - contains `bert`: TFBertForMaskedLM (Bert model)
@@ -212,6 +334,7 @@ class TFAutoModelWithLMHead(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
 
@@ -267,8 +390,12 @@ class TFAutoModelWithLMHead(object):
             model = TFAutoModelWithLMHead.from_pretrained('./pt_model/bert_pytorch_model.bin', from_pt=True, config=config)
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return TFT5WithLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return TFDistilBertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return TFAlbertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return TFRobertaForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -287,7 +414,7 @@ class TFAutoModelWithLMHead(object):
             return TFCTRLLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
+                         "'distilbert', 'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
                          "'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))
 
 
@@ -312,8 +439,39 @@ class TFAutoModelForSequenceClassification(object):
         This class cannot be instantiated using `__init__()` (throws an error).
     """
     def __init__(self):
-        raise EnvironmentError("TFAutoModelWithLMHead is designed to be instantiated "
-            "using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
+        raise EnvironmentError("TFAutoModelForSequenceClassification is designed to be instantiated "
+            "using the `TFAutoModelForSequenceClassification.from_pretrained(pretrained_model_name_or_path)` or "
+            "`TFAutoModelForSequenceClassification.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: XLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelForSequenceClassification.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, DistilBertConfig):
+            return TFDistilBertForSequenceClassification(config)
+        elif isinstance(config, RobertaConfig):
+            return TFRobertaForSequenceClassification(config)
+        elif isinstance(config, BertConfig):
+            return TFBertForSequenceClassification(config)
+        elif isinstance(config, XLNetConfig):
+            return TFXLNetForSequenceClassification(config)
+        elif isinstance(config, XLMConfig):
+            return TFXLMForSequenceClassification(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -338,6 +496,7 @@ class TFAutoModelForSequenceClassification(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
 
@@ -395,6 +554,8 @@ class TFAutoModelForSequenceClassification(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return TFDistilBertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return TFAlbertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return TFRobertaForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -405,7 +566,7 @@ class TFAutoModelForSequenceClassification(object):
             return TFXLMForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'xlnet', 'xlm', 'roberta'".format(pretrained_model_name_or_path))
+                         "'distilbert', 'bert', 'xlnet', 'xlm', 'roberta'".format(pretrained_model_name_or_path))
 
 
 class TFAutoModelForQuestionAnswering(object):
@@ -428,8 +589,36 @@ class TFAutoModelForQuestionAnswering(object):
         This class cannot be instantiated using `__init__()` (throws an error).
     """
     def __init__(self):
-        raise EnvironmentError("TFAutoModelWithLMHead is designed to be instantiated "
-            "using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
+        raise EnvironmentError("TFAutoModelForQuestionAnswering is designed to be instantiated "
+            "using the `TFAutoModelForQuestionAnswering.from_pretrained(pretrained_model_name_or_path)` or "
+            "`TFAutoModelForQuestionAnswering.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: XLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelForSequenceClassification.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, DistilBertConfig):
+            return TFDistilBertForQuestionAnswering(config)
+        elif isinstance(config, BertConfig):
+            return TFBertForQuestionAnswering(config)
+        elif isinstance(config, XLNetConfig):
+            return TFXLNetForQuestionAnswering(config)
+        elif isinstance(config, XLMConfig):
+            return TFXLMForQuestionAnswering(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -453,6 +642,7 @@ class TFAutoModelForQuestionAnswering(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
 
@@ -518,4 +708,121 @@ class TFAutoModelForQuestionAnswering(object):
             return TFXLMForQuestionAnsweringSimple.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'xlnet', 'xlm'".format(pretrained_model_name_or_path))
+                         "'distilbert', 'bert', 'xlnet', 'xlm'".format(pretrained_model_name_or_path))
+
+
+class TFAutoModelForTokenClassification:
+    def __init__(self):
+        raise EnvironmentError("TFAutoModelForTokenClassification is designed to be instantiated "
+                               "using the `TFAutoModelForTokenClassification.from_pretrained(pretrained_model_name_or_path)` or "
+                               "`AutoModelForTokenClassification.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBert model)
+                    - isInstance of `roberta` configuration class: RobteraModel (Roberta model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = TFAutoModelForTokenClassification.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, BertConfig):
+            return TFBertForTokenClassification(config)
+        elif isinstance(config, XLNetConfig):
+            return TFXLNetForTokenClassification(config)
+        elif isinstance(config, DistilBertConfig):
+            return TFDistilBertForTokenClassification(config)
+        elif isinstance(config, RobertaConfig):
+            return TFRobertaForTokenClassification(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        r""" Instantiates one of the question answering model classes of the library
+        from a pre-trained model configuration.
+
+        The `from_pretrained()` method takes care of returning the correct model class instance
+        using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `bert`: BertForTokenClassification (Bert model)
+            - contains `xlnet`: XLNetForTokenClassification (XLNet model)
+            - contains `distilbert`: DistilBertForTokenClassification (DistilBert model)
+            - contains `roberta`: RobertaForTokenClassification (Roberta model)
+
+        The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
+        To train the model, you should first set it back in training mode with `model.train()`
+
+        Params:
+            pretrained_model_name_or_path: either:
+
+                - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
+                - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
+
+            model_args: (`optional`) Sequence of positional arguments:
+                All remaning positional arguments will be passed to the underlying model's ``__init__`` method
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:
+
+                - the model is a model provided by the library (loaded with the ``shortcut-name`` string of a pretrained model), or
+                - the model was saved using :func:`~transformers.PreTrainedModel.save_pretrained` and is reloaded by suppling the save directory.
+                - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
+
+            state_dict: (`optional`) dict:
+                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                This option can be used if you want to create a model from a pretrained configuration but load your own weights.
+                In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
+
+            cache_dir: (`optional`) string:
+                Path to a directory in which a downloaded pre-trained model
+                configuration should be cached if the standard cache should not be used.
+
+            force_download: (`optional`) boolean, default False:
+                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            output_loading_info: (`optional`) boolean:
+                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+
+            kwargs: (`optional`) Remaining dictionary of keyword arguments:
+                Can be used to update the configuration object (after it being loaded) and initiate the model. (e.g. ``output_attention=True``). Behave differently depending on whether a `config` is provided or automatically loaded:
+
+                - If a configuration is provided with ``config``, ``**kwargs`` will be directly passed to the underlying model's ``__init__`` method (we assume all relevant updates to the configuration have already been done)
+                - If a configuration is not provided, ``kwargs`` will be first passed to the configuration class initialization function (:func:`~transformers.PretrainedConfig.from_pretrained`). Each key of ``kwargs`` that corresponds to a configuration attribute will be used to override said attribute with the supplied ``kwargs`` value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model's ``__init__`` function.
+
+        Examples::
+
+            model = TFAutoModelForTokenClassification.from_pretrained('bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = TFAutoModelForTokenClassification.from_pretrained('./test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = TFAutoModelForTokenClassification.from_pretrained('bert-base-uncased', output_attention=True)  # Update configuration during loading
+            assert model.config.output_attention == True
+            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+            config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
+            model = TFAutoModelForTokenClassification.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+
+        """
+        if 'bert' in pretrained_model_name_or_path:
+            return TFBertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'xlnet' in pretrained_model_name_or_path:
+            return TFXLNetForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
+            return TFDistilBertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'roberta' in pretrained_model_name_or_path:
+            return TFRobertaForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+
+        raise ValueError("Unrecognized model identifier in {}. Should contains one of "
+                         "'bert', 'xlnet', 'distilbert', 'roberta'".format(pretrained_model_name_or_path))