torch.cuda.is_available() is redundant as apex handles that internally (#9350 )

[prophetnet] wrong import (#9349 )
``` python -c "from apex.normalization import FusedProphetNetLayerNorm" Traceback (most recent call last): File "<string>", line 1, in <module> ImportError: cannot import name 'FusedProphetNetLayerNorm' from 'apex.normalization' (/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/apex/normalization/__init__.py) ``` It looks like this code has never been tested, so it silently fails inside try/except. Discovered this by accident in https://github.com/huggingface/transformers/issues/9338#issuecomment-752217708
2020-12-30 10:09:51 +01:00 · 2020-12-29 22:32:07 +01:00 · 2020-12-29 19:35:06 +01:00 · 2020-12-29 16:48:44 +01:00 · 2020-12-28 20:52:18 +01:00 · 2020-12-28 20:51:40 +01:00
1192 changed files with 28372 additions and 48816 deletions
@@ -3,47 +3,6 @@ orbs:
    gcp-gke: circleci/gcp-gke@1.0.4
    go: circleci/go@1.3.0

-commands:
-  skip-job-on-doc-only-changes:
-    description: "Do not continue this job and exit with success for PRs with only doc changes"
-    steps:
-
-      - run:
-          name: docs-only changes skip check
-          command: |
-            if test -n "$CIRCLE_PR_NUMBER"
-            then
-                echo $CIRCLE_PR_NUMBER
-                resp=$(curl -Ls https://api.github.com/repos/huggingface/transformers/pulls/${CIRCLE_PR_NUMBER})
-                user=$(jq -r .user.login \<<< $resp)   # PR creator username
-                head_ref=$(jq -r .head.ref \<<< $resp) # PR user's branch name
-                echo head_ref=$head_ref, user=$user
-            fi
-            
-            if test -n "$user" && test -n "$head_ref"
-            then
-                git clone https://github.com/$user/transformers user-clone
-                cd user-clone
-                git checkout $head_ref
-                fork_point_sha=$(git merge-base --fork-point master)
-                cd -
-            fi
-
-            if test -n "$fork_point_sha" && test -n "$(git diff --name-only $fork_point_sha)"
-            then
-                git --no-pager diff --name-only $fork_point_sha
-                if git diff --name-only $fork_point_sha | egrep -qv '\.(md|rst)$'
-                then
-                    echo "Non-docs were modified in this PR, proceeding normally"
-                else
-                    echo "Only docs were modified in this PR, quitting this job"
-                    # enable skipping once we get this sorted out
-                    # circleci step halt
-                fi
-            else
-                echo "Not enough data to perform a skipping check - continuing the job"
-            fi
-

 # TPU REFERENCES
 references:
@@ -114,13 +73,13 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - skip-job-on-doc-only-changes
            - restore_cache:
                  keys:
                      - v0.4-torch_and_tf-{{ checksum "setup.py" }}
                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,tf-cpu,torch,testing,sentencepiece]
+            - run: pip install tapas torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cpu.html
            - save_cache:
                key: v0.4-{{ checksum "setup.py" }}
                paths:
@@ -141,13 +100,13 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - skip-job-on-doc-only-changes
            - restore_cache:
                  keys:
                      - v0.4-torch-{{ checksum "setup.py" }}
                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,torch,testing,sentencepiece]
+            - run: pip install tapas torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cpu.html
            - save_cache:
                  key: v0.4-torch-{{ checksum "setup.py" }}
                  paths:
@@ -168,7 +127,6 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - skip-job-on-doc-only-changes
            - restore_cache:
                  keys:
                      - v0.4-tf-{{ checksum "setup.py" }}
@@ -195,7 +153,6 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - skip-job-on-doc-only-changes
            - restore_cache:
                keys:
                    - v0.4-flax-{{ checksum "setup.py" }}
@@ -222,13 +179,13 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - skip-job-on-doc-only-changes
            - restore_cache:
                  keys:
                      - v0.4-torch-{{ checksum "setup.py" }}
                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,torch,testing,sentencepiece]
+            - run: pip install tapas torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cpu.html
            - save_cache:
                  key: v0.4-torch-{{ checksum "setup.py" }}
                  paths:
@@ -249,7 +206,6 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - skip-job-on-doc-only-changes
            - restore_cache:
                  keys:
                      - v0.4-tf-{{ checksum "setup.py" }}
@@ -274,7 +230,6 @@ jobs:
            RUN_CUSTOM_TOKENIZERS: yes
        steps:
            - checkout
-            - skip-job-on-doc-only-changes
            - restore_cache:
                  keys:
                      - v0.4-custom_tokenizers-{{ checksum "setup.py" }}
@@ -302,14 +257,13 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - skip-job-on-doc-only-changes
            - restore_cache:
                  keys:
                      - v0.4-torch_examples-{{ checksum "setup.py" }}
                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,torch,sentencepiece,testing]
-            - run: pip install -r examples/requirements.txt
+            - run: pip install -r examples/_tests_requirements.txt
            - save_cache:
                  key: v0.4-torch_examples-{{ checksum "setup.py" }}
                  paths:
@@ -53,4 +53,5 @@ deploy_doc "3ebb1b3" v3.2.0
 deploy_doc "0613f05" v3.3.1
 deploy_doc "eb0e0ce" v3.4.0
 deploy_doc "818878d" v3.5.1
-deploy_doc "c781171" # v4.0.0 Latest stable release
+deploy_doc "c781171" v4.0.0
+deploy_doc "bfa4ccf" # v4.1.1 Latest stable release
@@ -40,6 +40,8 @@ jobs:
          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/pt-encoder-bert-tokenizer.json --path=templates/adding_a_new_model
          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/standalone.json --path=templates/adding_a_new_model
          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/tf-encoder-bert-tokenizer.json --path=templates/adding_a_new_model
+          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/tf-seq-2-seq-bart-tokenizer.json --path=templates/adding_a_new_model
+          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/pt-seq-2-seq-bart-tokenizer.json --path=templates/adding_a_new_model
          make style
          python utils/check_table.py --fix_and_overwrite
          python utils/check_dummies.py --fix_and_overwrite
@@ -50,6 +50,7 @@ jobs:
          pip install --upgrade pip
          pip install .[torch,sklearn,testing,onnxruntime,sentencepiece]
          pip install git+https://github.com/huggingface/datasets
+          pip install pandas torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cu102.html

      - name: Are GPUs recognized by our DL frameworks
        run: |
@@ -187,6 +188,7 @@ jobs:
          pip install --upgrade pip
          pip install .[torch,sklearn,testing,onnxruntime,sentencepiece]
          pip install git+https://github.com/huggingface/datasets
+          pip install pandas torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cu102.html

      - name: Are GPUs recognized by our DL frameworks
        run: |
@@ -222,6 +222,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 ultilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
 1. **[SqueezeBert](https://huggingface.co/transformers/model_doc/squeezebert.html)** released with the paper [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://arxiv.org/abs/2006.11316) by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
 1. **[T5](https://huggingface.co/transformers/model_doc/t5.html)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
+1. **[TAPAS](https://huggingface.co/transformers/master/model_doc/tapas.html)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[XLM](https://huggingface.co/transformers/model_doc/xlm.html)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](https://huggingface.co/transformers/model_doc/xlmprophetnet.html)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
@@ -1,10 +1,11 @@
 // These two things need to be updated at each release for the version selector.
 // Last stable version
-const stableVersion = "v4.0.0"
+const stableVersion = "v4.1.1"
 // Dictionary doc folder to label. The last stable version should have an empty key.
 const versionMapping = {
    "master": "master",
-    "": "v4.0.0 (stable)",
+    "": "v4.1.1 (stable)",
+    "v4.0.1": "v4.0.0/v4.0.1",
    "v3.5.1": "v3.5.0/v3.5.1",
    "v3.4.0": "v3.4.0",
    "v3.3.1": "v3.3.0/v3.3.1",
@@ -34,5 +34,5 @@ help people access the inner representations, mainly adapted from the great work
  in https://arxiv.org/abs/1905.10650.

 To help you understand and use these features, we have added a specific example script: `bertology.py
-<https://github.com/huggingface/transformers/blob/master/examples/bertology/run_bertology.py>`_ while extract
-information and prune a model pre-trained on GLUE.
+<https://github.com/huggingface/transformers/blob/master/examples/research_projects/bertology/run_bertology.py>`_ while
+extract information and prune a model pre-trained on GLUE.
@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'4.0.0'
+release = u'4.1.1'


 # -- General configuration ---------------------------------------------------
@@ -226,7 +226,7 @@ Contrary to RNNs that have the position of each token embedded within them, tran
 each token. Therefore, the position IDs (``position_ids``) are used by the model to identify each token's position in
 the list of tokens.

-They are an optional parameter. If no ``position_ids`` is passed to the model, the IDs are automatically created as
+They are an optional parameter. If no ``position_ids`` are passed to the model, the IDs are automatically created as
 absolute positional embeddings.

 Absolute positional embeddings are selected in the range ``[0, config.max_position_embeddings - 1]``. Some models use
@@ -176,19 +176,22 @@ and conversion utilities for the following models:
 30. :doc:`T5 <model_doc/t5>` (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.
-31. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
+31. `TAPAS <https://huggingface.co/transformers/master/model_doc/tapas.html>`__ (from Google AI) released with the
+    paper `TAPAS: Weakly Supervised Table Parsing via Pre-training <https://arxiv.org/abs/2004.02349>`__ by Jonathan
+    Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
+32. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
    Attentive Language Models Beyond a Fixed-Length Context <https://arxiv.org/abs/1901.02860>`__ by Zihang Dai*,
    Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
-32. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
+33. :doc:`XLM <model_doc/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.
-33. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
+34. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
    Predicting Future N-gram for Sequence-to-Sequence Pre-training <https://arxiv.org/abs/2001.04063>`__ by Yu Yan,
    Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
-34. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
+35. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (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.
-35. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `XLNet: Generalized Autoregressive
+36. :doc:`XLNet <model_doc/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.

@@ -269,6 +272,8 @@ TensorFlow and/or Flax.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |             T5              |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
+|            TAPAS            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |       Transformer-XL        |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |             XLM             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
@@ -382,6 +387,7 @@ TensorFlow and/or Flax.
    model_doc/roberta
    model_doc/squeezebert
    model_doc/t5
+    model_doc/tapas
    model_doc/transformerxl
    model_doc/xlm
    model_doc/xlmprophetnet
@@ -32,6 +32,9 @@ generation.
 .. autoclass:: transformers.LogitsProcessorList
    :members: __call__

+.. autoclass:: transformers.LogitsWarper
+    :members: __call__
+
 .. autoclass:: transformers.MinLengthLogitsProcessor
    :members: __call__

@@ -67,3 +70,10 @@ BeamSearch

 .. autoclass:: transformers.BeamSearchScorer
    :members: process, finalize
+
+Utilities
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autofunction:: transformers.top_k_top_p_filtering
+
+.. autofunction:: transformers.tf_top_k_top_p_filtering
@@ -91,8 +91,6 @@ TensorFlow loss functions
 TensorFlow Helper Functions
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autofunction:: transformers.modeling_tf_utils.cast_bool_to_primitive
-
 .. autofunction:: transformers.modeling_tf_utils.get_initializer

 .. autofunction:: transformers.modeling_tf_utils.keras_serializable
@@ -22,6 +22,8 @@ Utilities

 .. autoclass:: transformers.EvalPrediction

+.. autoclass:: transformers.EvaluationStrategy
+
 .. autofunction:: transformers.set_seed

 .. autofunction:: transformers.torch_distributed_zero_first
@@ -32,8 +34,15 @@ Callbacks internals

 .. autoclass:: transformers.trainer_callback.CallbackHandler

+
 Distributed Evaluation
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.trainer_pt_utils.DistributedTensorGatherer
    :members:
+
+
+Distributed Evaluation
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.HfArgumentParser
@@ -13,9 +13,10 @@
 Models
 -----------------------------------------------------------------------------------------------------------------------

-The base classes :class:`~transformers.PreTrainedModel` and :class:`~transformers.TFPreTrainedModel` implement the
-common methods for loading/saving a model either from a local file or directory, or from a pretrained model
-configuration provided by the library (downloaded from HuggingFace's AWS S3 repository).
+The base classes :class:`~transformers.PreTrainedModel`, :class:`~transformers.TFPreTrainedModel`, and
+:class:`~transformers.FlaxPreTrainedModel` implement the common methods for loading/saving a model either from a local
+file or directory, or from a pretrained model configuration provided by the library (downloaded from HuggingFace's AWS
+S3 repository).

 :class:`~transformers.PreTrainedModel` and :class:`~transformers.TFPreTrainedModel` also implement a few methods which
 are common among all the models to:
@@ -57,6 +58,13 @@ TFModelUtilsMixin
    :members:


+FlaxPreTrainedModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxPreTrainedModel
+    :members:
+
+
 Generation
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -43,6 +43,10 @@ Schedules
 Learning Rate Schedules (Pytorch)
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

+.. autoclass:: transformers.SchedulerType
+
+.. autofunction:: transformers.get_scheduler
+
 .. autofunction:: transformers.get_constant_schedule


@@ -74,6 +78,10 @@ Learning Rate Schedules (Pytorch)
    :target: /imgs/warmup_linear_schedule.png
    :alt:

+
+.. autofunction:: transformers.get_polynomial_decay_schedule_with_warmup
+
+
 Warmup (TensorFlow)
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

@@ -126,13 +126,6 @@ CausalLMOutputWithCrossAttentions
    :members:


-CausalLMOutputWithPastAndCrossAttentions
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.CausalLMOutputWithPastAndCrossAttentions
-    :members:
-
-
 CausalLMOutputWithPast
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -34,6 +34,7 @@ There are two categories of pipeline abstractions to be aware about:
    - :class:`~transformers.TranslationPipeline`
    - :class:`~transformers.ZeroShotClassificationPipeline`
    - :class:`~transformers.Text2TextGenerationPipeline`
+    - :class:`~transformers.TableQuestionAnsweringPipeline`

 The pipeline abstraction
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -73,8 +74,9 @@ FillMaskPipeline
 NerPipeline
 =======================================================================================================================

-This class is an alias of the :class:`~transformers.TokenClassificationPipeline` defined below. Please refer to that
-pipeline for documentation and usage examples.
+.. autoclass:: transformers.NerPipeline
+
+See :class:`~transformers.TokenClassificationPipeline` for all details.

 QuestionAnsweringPipeline
 =======================================================================================================================
@@ -90,6 +92,13 @@ SummarizationPipeline
    :special-members: __call__
    :members:

+TableQuestionAnsweringPipeline
+=======================================================================================================================
+
+.. autoclass:: transformers.TableQuestionAnsweringPipeline
+    :special-members: __call__
+
+
 TextClassificationPipeline
 =======================================================================================================================

@@ -118,6 +127,13 @@ TokenClassificationPipeline
    :special-members: __call__
    :members:

+TranslationPipeline
+=======================================================================================================================
+
+.. autoclass:: transformers.TranslationPipeline
+    :special-members: __call__
+    :members:
+
 ZeroShotClassificationPipeline
 =======================================================================================================================

@@ -63,6 +63,13 @@ Trainer
    :members:


+Seq2SeqTrainer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.Seq2SeqTrainer
+    :members: evaluate, predict
+
+
 TFTrainer
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -77,6 +84,13 @@ TrainingArguments
    :members:


+Seq2SeqTrainingArguments
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.Seq2SeqTrainingArguments
+    :members:
+
+
 TFTrainingArguments
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -60,6 +60,13 @@ AlbertTokenizer
        create_token_type_ids_from_sequences, save_vocabulary


+AlbertTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AlbertTokenizerFast
+    :members:
+
+
 Albert specific outputs
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -114,6 +114,13 @@ AutoModelForQuestionAnswering
    :members:


+AutoModelForTableQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AutoModelForTableQuestionAnswering
+    :members:
+
+
 TFAutoModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -175,3 +182,10 @@ TFAutoModelForQuestionAnswering

 .. autoclass:: transformers.TFAutoModelForQuestionAnswering
    :members:
+
+
+FlaxAutoModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxAutoModel
+    :members:
@@ -55,9 +55,8 @@ Implementation Notes

 - Bart doesn't use :obj:`token_type_ids` for sequence classification. Use :class:`~transformers.BartTokenizer` or
  :meth:`~transformers.BartTokenizer.encode` to get the proper splitting.
- The forward pass of :class:`~transformers.BartModel` will create decoder inputs (using the helper function
-  :func:`transformers.models.bart.modeling_bart._prepare_bart_decoder_inputs`) if they are not passed. This is
-  different than some other modeling APIs.
+- The forward pass of :class:`~transformers.BartModel` will create the ``decoder_input_ids`` if they are not passed.
+  This is different than some other modeling APIs. A typical use case of this feature is mask filling.
 - Model predictions are intended to be identical to the original implementation when
  :obj:`force_bos_token_to_be_generated=True`. This only works, however, if the string you pass to
  :func:`fairseq.encode` starts with a space.
@@ -98,6 +97,12 @@ BartTokenizer
    :members:


+BartTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BartTokenizerFast
+    :members:
+

 BartModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -51,3 +51,9 @@ BarthezTokenizer
 .. autoclass:: transformers.BarthezTokenizer
    :members:

+
+BarthezTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BarthezTokenizerFast
+    :members:
@@ -207,3 +207,10 @@ FlaxBertModel

 .. autoclass:: transformers.FlaxBertModel
    :members: __call__
+
+
+FlaxBertForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBertForMaskedLM
+    :members: __call__
@@ -100,6 +100,15 @@ BlenderbotSmallTokenizer
    :members:


+BlenderbotModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+See :obj:`transformers.BartModel` for arguments to `forward` and `generate`
+
+.. autoclass:: transformers.BlenderbotModel
+    :members:
+
+
 BlenderbotForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -54,6 +54,13 @@ CamembertTokenizer
        create_token_type_ids_from_sequences, save_vocabulary


+CamembertTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CamembertTokenizerFast
+    :members:
+
+
 CamembertModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -97,3 +97,8 @@ TFCTRLLMHeadModel
 .. autoclass:: transformers.TFCTRLLMHeadModel
    :members: call

+TFCTRLForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFCTRLForSequenceClassification
+    :members: call
@@ -138,3 +138,9 @@ TFOpenAIGPTDoubleHeadsModel

 .. autoclass:: transformers.TFOpenAIGPTDoubleHeadsModel
    :members: call
+
+TFOpenAIGPTForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFOpenAIGPTForSequenceClassification
+    :members: call
@@ -57,6 +57,13 @@ LayoutLMTokenizer
    :members:


+LayoutLMTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LayoutLMTokenizerFast
+    :members:
+
+
 LayoutLMModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -34,6 +34,12 @@ contrast to most prior work, we also pretrain Longformer and finetune it on a va
 pretrained Longformer consistently outperforms RoBERTa on long document tasks and sets new state-of-the-art results on
 WikiHop and TriviaQA.*

+Tips:
+
+- Since the Longformer is based on RoBERTa, it doesn't have :obj:`token_type_ids`. You don't need to indicate which
+  token belongs to which segment. Just separate your segments with the separation token :obj:`tokenizer.sep_token` (or
+  :obj:`</s>`).
+
 The Authors' code can be found `here <https://github.com/allenai/longformer>`__.

 Longformer Self Attention
@@ -90,6 +90,20 @@ MBartTokenizer
    :members: build_inputs_with_special_tokens, prepare_seq2seq_batch


+MBartTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MBartTokenizerFast
+    :members:
+
+
+MBartModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MBartModel
+    :members:
+
+
 MBartForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -1,3 +1,15 @@
+.. 
+    Copyright 2020 The HuggingFace Team. 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.
+
 MPNet
 -----------------------------------------------------------------------------------------------------------------------

@@ -37,6 +37,22 @@ MT5Config
    :members:


+MT5Tokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MT5Tokenizer
+
+See :class:`~transformers.T5Tokenizer` for all details.
+
+
+MT5TokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MT5TokenizerFast
+
+See :class:`~transformers.T5TokenizerFast` for all details.
+
+
 MT5Model
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -112,6 +112,19 @@ warning: ``add_tokens`` does not work at the moment.
    :members: __call__, prepare_seq2seq_batch


+PegasusTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.PegasusTokenizerFast
+    :members:
+
+
+PegasusModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.PegasusModel
+
+
 PegasusForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -163,6 +163,13 @@ ReformerTokenizer
    :members: save_vocabulary


+ReformerTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ReformerTokenizerFast
+    :members:
+
+
 ReformerModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -44,9 +44,9 @@ Tips:

  For more information about which prefix to use, it is easiest to look into Appendix D of the `paper
  <https://arxiv.org/pdf/1910.10683.pdf>`__. - For sequence-to-sequence generation, it is recommended to use
-  :obj:`T5ForConditionalGeneration.generate()``. This method takes care of feeding the encoded input via
-  cross-attention layers to the decoder and auto-regressively generates the decoder output. - T5 uses relative scalar
-  embeddings. Encoder input padding can be done on the left and on the right.
+  :obj:`T5ForConditionalGeneration.generate()`. This method takes care of feeding the encoded input via cross-attention
+  layers to the decoder and auto-regressively generates the decoder output. - T5 uses relative scalar embeddings.
+  Encoder input padding can be done on the left and on the right.

 The original code can be found `here <https://github.com/google-research/text-to-text-transfer-transformer>`__.

@@ -55,7 +55,7 @@ Training

 T5 is an encoder-decoder model and converts all NLP problems into a text-to-text format. It is trained using teacher
 forcing. This means that for training we always need an input sequence and a target sequence. The input sequence is fed
-to the model using :obj:`input_ids``. The target sequence is shifted to the right, i.e., prepended by a start-sequence
+to the model using :obj:`input_ids`. The target sequence is shifted to the right, i.e., prepended by a start-sequence
 token and fed to the decoder using the :obj:`decoder_input_ids`. In teacher-forcing style, the target sequence is then
 appended by the EOS token and corresponds to the :obj:`labels`. The PAD token is hereby used as the start-sequence
 token. T5 can be trained / fine-tuned both in a supervised and unsupervised fashion.
@@ -107,6 +107,13 @@ T5Tokenizer
        create_token_type_ids_from_sequences, prepare_seq2seq_batch, save_vocabulary


+T5TokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.T5TokenizerFast
+    :members:
+
+
 T5Model
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -124,7 +131,7 @@ T5EncoderModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.T5EncoderModel
-    :members: forward
+    :members: forward, parallelize, deparallelize

 TFT5Model
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -0,0 +1,434 @@
+TAPAS
+-----------------------------------------------------------------------------------------------------------------------
+
+.. note::
+
+    This is a recently introduced model so the API hasn't been tested extensively. There may be some bugs or slight
+    breaking changes to fix them in the future.
+
+
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The TAPAS model was proposed in `TAPAS: Weakly Supervised Table Parsing via Pre-training
+<https://www.aclweb.org/anthology/2020.acl-main.398>`__ by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller,
+Francesco Piccinno and Julian Martin Eisenschlos. It's a BERT-based model specifically designed (and pre-trained) for
+answering questions about tabular data. Compared to BERT, TAPAS uses relative position embeddings and has 7 token types
+that encode tabular structure. TAPAS is pre-trained on the masked language modeling (MLM) objective on a large dataset
+comprising millions of tables from English Wikipedia and corresponding texts. For question answering, TAPAS has 2 heads
+on top: a cell selection head and an aggregation head, for (optionally) performing aggregations (such as counting or
+summing) among selected cells. TAPAS has been fine-tuned on several datasets: `SQA
+<https://www.microsoft.com/en-us/download/details.aspx?id=54253>`__ (Sequential Question Answering by Microsoft), `WTQ
+<https://github.com/ppasupat/WikiTableQuestions>`__ (Wiki Table Questions by Stanford University) and `WikiSQL
+<https://github.com/salesforce/WikiSQL>`__ (by Salesforce). It achieves state-of-the-art on both SQA and WTQ, while
+having comparable performance to SOTA on WikiSQL, with a much simpler architecture.
+
+The abstract from the paper is the following:
+
+*Answering natural language questions over tables is usually seen as a semantic parsing task. To alleviate the
+collection cost of full logical forms, one popular approach focuses on weak supervision consisting of denotations
+instead of logical forms. However, training semantic parsers from weak supervision poses difficulties, and in addition,
+the generated logical forms are only used as an intermediate step prior to retrieving the denotation. In this paper, we
+present TAPAS, an approach to question answering over tables without generating logical forms. TAPAS trains from weak
+supervision, and predicts the denotation by selecting table cells and optionally applying a corresponding aggregation
+operator to such selection. TAPAS extends BERT's architecture to encode tables as input, initializes from an effective
+joint pre-training of text segments and tables crawled from Wikipedia, and is trained end-to-end. We experiment with
+three different semantic parsing datasets, and find that TAPAS outperforms or rivals semantic parsing models by
+improving state-of-the-art accuracy on SQA from 55.1 to 67.2 and performing on par with the state-of-the-art on WIKISQL
+and WIKITQ, but with a simpler model architecture. We additionally find that transfer learning, which is trivial in our
+setting, from WIKISQL to WIKITQ, yields 48.7 accuracy, 4.2 points above the state-of-the-art.*
+
+In addition, the authors have further pre-trained TAPAS to recognize **table entailment**, by creating a balanced
+dataset of millions of automatically created training examples which are learned in an intermediate step prior to
+fine-tuning. The authors of TAPAS call this further pre-training intermediate pre-training (since TAPAS is first
+pre-trained on MLM, and then on another dataset). They found that intermediate pre-training further improves
+performance on SQA, achieving a new state-of-the-art as well as state-of-the-art on `TabFact
+<https://github.com/wenhuchen/Table-Fact-Checking>`__, a large-scale dataset with 16k Wikipedia tables for table
+entailment (a binary classification task). For more details, see their follow-up paper: `Understanding tables with
+intermediate pre-training <https://www.aclweb.org/anthology/2020.findings-emnlp.27/>`__ by Julian Martin Eisenschlos,
+Syrine Krichene and Thomas Müller.
+
+The original code can be found `here <https://github.com/google-research/tapas>`__.
+
+Tips:
+
+- TAPAS is a model that uses relative position embeddings by default (restarting the position embeddings at every cell
+  of the table). Note that this is something that was added after the publication of the original TAPAS paper.
+  According to the authors, this usually results in a slightly better performance, and allows you to encode longer
+  sequences without running out of embeddings. This is reflected in the ``reset_position_index_per_cell`` parameter of
+  :class:`~transformers.TapasConfig`, which is set to ``True`` by default. The default versions of the models available
+  in the `model hub <https://huggingface.co/models?search=tapas>`_ all use relative position embeddings. You can still
+  use the ones with absolute position embeddings by passing in an additional argument ``revision="no_reset"`` when
+  calling the ``.from_pretrained()`` method. Note that it's usually advised to pad the inputs on the right rather than
+  the left.
+- TAPAS is based on BERT, so ``TAPAS-base`` for example corresponds to a ``BERT-base`` architecture. Of course,
+  TAPAS-large will result in the best performance (the results reported in the paper are from TAPAS-large). Results of
+  the various sized models are shown on the `original Github repository <https://github.com/google-research/tapas>`_.
+- TAPAS has checkpoints fine-tuned on SQA, which are capable of answering questions related to a table in a
+  conversational set-up. This means that you can ask follow-up questions such as "what is his age?" related to the
+  previous question. Note that the forward pass of TAPAS is a bit different in case of a conversational set-up: in that
+  case, you have to feed every table-question pair one by one to the model, such that the `prev_labels` token type ids
+  can be overwritten by the predicted `labels` of the model to the previous question. See "Usage" section for more
+  info.
+- TAPAS is similar to BERT and therefore relies on the masked language modeling (MLM) objective. It is therefore
+  efficient at predicting masked tokens and at NLU in general, but is not optimal for text generation. Models trained
+  with a causal language modeling (CLM) objective are better in that regard.
+
+
+Usage: fine-tuning
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Here we explain how you can fine-tune :class:`~transformers.TapasForQuestionAnswering` on your own dataset.
+
+**STEP 1: Choose one of the 3 ways in which you can use TAPAS - or experiment**
+
+Basically, there are 3 different ways in which one can fine-tune :class:`~transformers.TapasForQuestionAnswering`,
+corresponding to the different datasets on which Tapas was fine-tuned:
+
+1. SQA: if you're interested in asking follow-up questions related to a table, in a conversational set-up. For example
+   if you first ask "what's the name of the first actor?" then you can ask a follow-up question such as "how old is
+   he?". Here, questions do not involve any aggregation (all questions are cell selection questions).
+2. WTQ: if you're not interested in asking questions in a conversational set-up, but rather just asking questions
+   related to a table, which might involve aggregation, such as counting a number of rows, summing up cell values or
+   averaging cell values. You can then for example ask "what's the total number of goals Cristiano Ronaldo made in his
+   career?". This case is also called **weak supervision**, since the model itself must learn the appropriate
+   aggregation operator (SUM/COUNT/AVERAGE/NONE) given only the answer to the question as supervision.
+3. WikiSQL-supervised: this dataset is based on WikiSQL with the model being given the ground truth aggregation
+   operator during training. This is also called **strong supervision**. Here, learning the appropriate aggregation
+   operator is much easier.
+
+To summarize:
+
+------------------------------------+----------------------+-------------------------------------------------------------------------------------------------------------------+
+| **Task**                           | **Example dataset**  | **Description**                                                                                                   |
+------------------------------------+----------------------+-------------------------------------------------------------------------------------------------------------------+
+| Conversational                     | SQA                  | Conversational, only cell selection questions                                                                     |
+------------------------------------+----------------------+-------------------------------------------------------------------------------------------------------------------+
+| Weak supervision for aggregation   | WTQ                  | Questions might involve aggregation, and the model must learn this given only the answer as supervision           |
+------------------------------------+----------------------+-------------------------------------------------------------------------------------------------------------------+
+| Strong supervision for aggregation | WikiSQL-supervised   | Questions might involve aggregation, and the model must learn this given the gold aggregation operator            |
+------------------------------------+----------------------+-------------------------------------------------------------------------------------------------------------------+
+
+Initializing a model with a pre-trained base and randomly initialized classification heads from the model hub can be
+done as follows (be sure to have installed the `torch-scatter dependency <https://github.com/rusty1s/pytorch_scatter>`_
+for your environment):
+
+.. code-block::
+
+        >>> from transformers import TapasConfig, TapasForQuestionAnswering
+
+        >>> # for example, the base sized model with default SQA configuration
+        >>> model = TapasForQuestionAnswering.from_pretrained('google/tapas-base')
+
+        >>> # or, the base sized model with WTQ configuration
+        >>> config = TapasConfig.from_pretrained('google/tapas-base-finetuned-wtq')
+        >>> model = TapasForQuestionAnswering.from_pretrained('google/tapas-base', config=config)
+
+        >>> # or, the base sized model with WikiSQL configuration
+        >>> config = TapasConfig('google-base-finetuned-wikisql-supervised')
+        >>> model = TapasForQuestionAnswering.from_pretrained('google/tapas-base', config=config)
+
+
+Of course, you don't necessarily have to follow one of these three ways in which TAPAS was fine-tuned. You can also
+experiment by defining any hyperparameters you want when initializing :class:`~transformers.TapasConfig`, and then
+create a :class:`~transformers.TapasForQuestionAnswering` based on that configuration. For example, if you have a
+dataset that has both conversational questions and questions that might involve aggregation, then you can do it this
+way. Here's an example:
+
+.. code-block::
+
+        >>> from transformers import TapasConfig, TapasForQuestionAnswering
+
+        >>> # you can initialize the classification heads any way you want (see docs of TapasConfig)
+        >>> config = TapasConfig(num_aggregation_labels=3, average_logits_per_cell=True, select_one_column=False)
+        >>> # initializing the pre-trained base sized model with our custom classification heads
+        >>> model = TapasForQuestionAnswering.from_pretrained('google/tapas-base', config=config)
+
+What you can also do is start from an already fine-tuned checkpoint. A note here is that the already fine-tuned
+checkpoint on WTQ has some issues due to the L2-loss which is somewhat brittle. See `here
+<https://github.com/google-research/tapas/issues/91#issuecomment-735719340>`__ for more info.
+
+For a list of all pre-trained and fine-tuned TAPAS checkpoints available in the HuggingFace model hub, see `here
+<https://huggingface.co/models?search=tapas>`__.
+
+**STEP 2: Prepare your data in the SQA format**
+
+Second, no matter what you picked above, you should prepare your dataset in the `SQA format
+<https://www.microsoft.com/en-us/download/details.aspx?id=54253>`__. This format is a TSV/CSV file with the following
+columns:
+
+- ``id``: optional, id of the table-question pair, for bookkeeping purposes.
+- ``annotator``: optional, id of the person who annotated the table-question pair, for bookkeeping purposes.
+- ``position``: integer indicating if the question is the first, second, third,... related to the table. Only required
+  in case of conversational setup (SQA). You don't need this column in case you're going for WTQ/WikiSQL-supervised.
+- ``question``: string
+- ``table_file``: string, name of a csv file containing the tabular data
+- ``answer_coordinates``: list of one or more tuples (each tuple being a cell coordinate, i.e. row, column pair that is
+  part of the answer)
+- ``answer_text``: list of one or more strings (each string being a cell value that is part of the answer)
+- ``aggregation_label``: index of the aggregation operator. Only required in case of strong supervision for aggregation
+  (the WikiSQL-supervised case)
+- ``float_answer``: the float answer to the question, if there is one (np.nan if there isn't). Only required in case of
+  weak supervision for aggregation (such as WTQ and WikiSQL)
+
+The tables themselves should be present in a folder, each table being a separate csv file. Note that the authors of the
+TAPAS algorithm used conversion scripts with some automated logic to convert the other datasets (WTQ, WikiSQL) into the
+SQA format. The author explains this `here
+<https://github.com/google-research/tapas/issues/50#issuecomment-705465960>`__. Interestingly, these conversion scripts
+are not perfect (the ``answer_coordinates`` and ``float_answer`` fields are populated based on the ``answer_text``),
+meaning that WTQ and WikiSQL results could actually be improved.
+
+**STEP 3: Convert your data into PyTorch tensors using TapasTokenizer**
+
+Third, given that you've prepared your data in this TSV/CSV format (and corresponding CSV files containing the tabular
+data), you can then use :class:`~transformers.TapasTokenizer` to convert table-question pairs into :obj:`input_ids`,
+:obj:`attention_mask`, :obj:`token_type_ids` and so on. Again, based on which of the three cases you picked above,
+:class:`~transformers.TapasForQuestionAnswering` requires different inputs to be fine-tuned:
+
+------------------------------------+----------------------------------------------------------------------------------------------+
+| **Task**                           | **Required inputs**                                                                          |
+------------------------------------+----------------------------------------------------------------------------------------------+
+| Conversational                     | ``input_ids``, ``attention_mask``, ``token_type_ids``, ``labels``                            |
+------------------------------------+----------------------------------------------------------------------------------------------+
+| Weak supervision for aggregation   | ``input_ids``, ``attention_mask``, ``token_type_ids``, ``labels``, ``numeric_values``,       |
+|                                    | ``numeric_values_scale``, ``float_answer``                                                   |
+------------------------------------+----------------------------------------------------------------------------------------------+
+| Strong supervision for aggregation | ``input ids``, ``attention mask``, ``token type ids``, ``labels``, ``aggregation_labels``    |
+------------------------------------+----------------------------------------------------------------------------------------------+
+
+:class:`~transformers.TapasTokenizer` creates the ``labels``, ``numeric_values`` and ``numeric_values_scale`` based on
+the ``answer_coordinates`` and ``answer_text`` columns of the TSV file. The ``float_answer`` and ``aggregation_labels``
+are already in the TSV file of step 2. Here's an example:
+
+.. code-block::
+
+        >>> from transformers import TapasTokenizer
+        >>> import pandas as pd
+
+        >>> model_name = 'google/tapas-base'
+        >>> tokenizer = TapasTokenizer.from_pretrained(model_name)
+
+        >>> data = {'Actors': ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], 'Number of movies': ["87", "53", "69"]}
+        >>> queries = ["What is the name of the first actor?", "How many movies has George Clooney played in?", "What is the total number of movies?"]
+        >>> answer_coordinates = [[(0, 0)], [(2, 1)], [(0, 1), (1, 1), (2, 1)]]
+        >>> answer_text = [["Brad Pitt"], ["69"], ["209"]]
+        >>> table = pd.DataFrame.from_dict(data)
+        >>> inputs = tokenizer(table=table, queries=queries, answer_coordinates=answer_coordinates, answer_text=answer_text, padding='max_length', return_tensors='pt')
+        >>> inputs
+        {'input_ids': tensor([[ ... ]]), 'attention_mask': tensor([[...]]), 'token_type_ids': tensor([[[...]]]),
+        'numeric_values': tensor([[ ... ]]), 'numeric_values_scale: tensor([[ ... ]]), labels: tensor([[ ... ]])}
+
+Note that :class:`~transformers.TapasTokenizer` expects the data of the table to be **text-only**. You can use
+``.astype(str)`` on a dataframe to turn it into text-only data. Of course, this only shows how to encode a single
+training example. It is advised to create a PyTorch dataset and a corresponding dataloader:
+
+.. code-block::
+
+        >>> import torch
+        >>> import pandas as pd
+
+        >>> tsv_path = "your_path_to_the_tsv_file"
+        >>> table_csv_path = "your_path_to_a_directory_containing_all_csv_files"
+
+        >>> class TableDataset(torch.utils.data.Dataset):
+        ...     def __init__(self, data, tokenizer):
+        ...         self.data = data
+        ...         self.tokenizer = tokenizer
+        ...
+        ...     def __getitem__(self, idx):
+        ...         item = data.iloc[idx]
+        ...         table = pd.read_csv(table_csv_path + item.table_file).astype(str) # be sure to make your table data text only
+        ...         encoding = self.tokenizer(table=table, 
+        ...                                   queries=item.question, 
+        ...                                   answer_coordinates=item.answer_coordinates, 
+        ...                                   answer_text=item.answer_text,
+        ...                                   truncation=True,
+        ...                                   padding="max_length",
+        ...                                   return_tensors="pt"
+        ...         )
+        ...         # remove the batch dimension which the tokenizer adds by default
+        ...         encoding = {key: val.squeeze(0) for key, val in encoding.items()}
+        ...         # add the float_answer which is also required (weak supervision for aggregation case)
+        ...         encoding["float_answer"] = torch.tensor(item.float_answer) 
+        ...         return encoding
+        ...
+        ...     def __len__(self):
+        ...        return len(self.data)
+
+        >>> data = pd.read_csv(tsv_path, sep='\t')
+        >>> train_dataset = TableDataset(data, tokenizer)
+        >>> train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=32)
+
+Note that here, we encode each table-question pair independently. This is fine as long as your dataset is **not
+conversational**. In case your dataset involves conversational questions (such as in SQA), then you should first group
+together the ``queries``, ``answer_coordinates`` and ``answer_text`` per table (in the order of their ``position``
+index) and batch encode each table with its questions. This will make sure that the ``prev_labels`` token types (see
+docs of :class:`~transformers.TapasTokenizer`) are set correctly. See `this notebook
+<https://github.com/NielsRogge/Transformers-Tutorials/blob/master/Fine_tuning_TapasForQuestionAnswering_on_SQA.ipynb>`__
+for more info.
+
+**STEP 4: Train (fine-tune) TapasForQuestionAnswering**
+
+You can then fine-tune :class:`~transformers.TapasForQuestionAnswering` using native PyTorch as follows (shown here for
+the weak supervision for aggregation case):
+
+.. code-block::
+
+        >>> from transformers import TapasConfig, TapasForQuestionAnswering, AdamW
+
+        >>> # this is the default WTQ configuration
+        >>> config = TapasConfig(
+        ...            num_aggregation_labels = 4,
+        ...            use_answer_as_supervision = True,
+        ...            answer_loss_cutoff = 0.664694,
+        ...            cell_selection_preference = 0.207951,
+        ...            huber_loss_delta = 0.121194,
+        ...            init_cell_selection_weights_to_zero = True,
+        ...            select_one_column = True,
+        ...            allow_empty_column_selection = False,
+        ...            temperature = 0.0352513,
+        ... )
+        >>> model = TapasForQuestionAnswering.from_pretrained("google/tapas-base", config=config)
+
+        >>> optimizer = AdamW(model.parameters(), lr=5e-5)
+
+        >>> for epoch in range(2):  # loop over the dataset multiple times
+        ...    for idx, batch in enumerate(train_dataloader):
+        ...         # get the inputs; 
+        ...         input_ids = batch["input_ids"]
+        ...         attention_mask = batch["attention_mask"]
+        ...         token_type_ids = batch["token_type_ids"]
+        ...         labels = batch["labels"]
+        ...         numeric_values = batch["numeric_values"]
+        ...         numeric_values_scale = batch["numeric_values_scale"]
+        ...         float_answer = batch["float_answer"]
+
+        ...         # zero the parameter gradients
+        ...         optimizer.zero_grad()
+
+        ...         # forward + backward + optimize
+        ...         outputs = model(input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, 
+        ...                        labels=labels, numeric_values=numeric_values, numeric_values_scale=numeric_values_scale, 
+        ...                        float_answer=float_answer)
+        ...         loss = outputs.loss
+        ...         loss.backward()
+        ...         optimizer.step()
+
+Usage: inference
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Here we explain how you can use :class:`~transformers.TapasForQuestionAnswering` for inference (i.e. making predictions
+on new data). For inference, only ``input_ids``, ``attention_mask`` and ``token_type_ids`` (which you can obtain using
+:class:`~transformers.TapasTokenizer`) have to be provided to the model to obtain the logits. Next, you can use the
+handy ``convert_logits_to_predictions`` method of :class:`~transformers.TapasTokenizer` to convert these into predicted
+coordinates and optional aggregation indices.
+
+However, note that inference is **different** depending on whether or not the setup is conversational. In a
+non-conversational set-up, inference can be done in parallel on all table-question pairs of a batch. Here's an example
+of that:
+
+.. code-block::
+
+        >>> from transformers import TapasTokenizer, TapasForQuestionAnswering
+        >>> import pandas as pd 
+
+        >>> model_name = 'google/tapas-base-finetuned-wtq'
+        >>> model = TapasForQuestionAnswering.from_pretrained(model_name)
+        >>> tokenizer = TapasTokenizer.from_pretrained(model_name)
+
+        >>> data = {'Actors': ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], 'Number of movies': ["87", "53", "69"]}
+        >>> queries = ["What is the name of the first actor?", "How many movies has George Clooney played in?", "What is the total number of movies?"]
+        >>> table = pd.DataFrame.from_dict(data)
+        >>> inputs = tokenizer(table=table, queries=queries, padding='max_length', return_tensors="pt") 
+        >>> outputs = model(**inputs)
+        >>> predicted_answer_coordinates, predicted_aggregation_indices = tokenizer.convert_logits_to_predictions(
+        ...         inputs, 
+        ...         outputs.logits.detach(), 
+        ...         outputs.logits_aggregation.detach()
+        ... )
+
+        >>> # let's print out the results:
+        >>> id2aggregation = {0: "NONE", 1: "SUM", 2: "AVERAGE", 3:"COUNT"}
+        >>> aggregation_predictions_string = [id2aggregation[x] for x in predicted_aggregation_indices]
+
+        >>> answers = []
+        >>> for coordinates in predicted_answer_coordinates:
+        ...   if len(coordinates) == 1:
+        ...     # only a single cell:
+        ...     answers.append(table.iat[coordinates[0]])
+        ...   else:
+        ...     # multiple cells
+        ...     cell_values = []
+        ...     for coordinate in coordinates:
+        ...        cell_values.append(table.iat[coordinate])
+        ...     answers.append(", ".join(cell_values))
+
+        >>> display(table)
+        >>> print("")
+        >>> for query, answer, predicted_agg in zip(queries, answers, aggregation_predictions_string):
+        ...   print(query)
+        ...   if predicted_agg == "NONE":
+        ...     print("Predicted answer: " + answer)
+        ...   else:
+        ...     print("Predicted answer: " + predicted_agg + " > " + answer)    
+        What is the name of the first actor?
+        Predicted answer: Brad Pitt
+        How many movies has George Clooney played in?
+        Predicted answer: COUNT > 69
+        What is the total number of movies?
+        Predicted answer: SUM > 87, 53, 69
+
+In case of a conversational set-up, then each table-question pair must be provided **sequentially** to the model, such
+that the ``prev_labels`` token types can be overwritten by the predicted ``labels`` of the previous table-question
+pair. Again, more info can be found in `this notebook
+<https://github.com/NielsRogge/Transformers-Tutorials/blob/master/Fine_tuning_TapasForQuestionAnswering_on_SQA.ipynb>`__.
+
+
+Tapas specific outputs
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.models.tapas.modeling_tapas.TableQuestionAnsweringOutput
+    :members:
+
+
+TapasConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasConfig
+    :members:
+
+
+TapasTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasTokenizer
+    :members: __call__, convert_logits_to_predictions, save_vocabulary
+
+
+TapasModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasModel
+    :members: forward
+
+
+TapasForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasForMaskedLM
+    :members: forward
+
+
+TapasForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasForSequenceClassification
+    :members: forward
+
+
+TapasForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasForQuestionAnswering
+    :members: forward
@@ -87,12 +87,14 @@ TransfoXLLMHeadModel
 .. autoclass:: transformers.TransfoXLLMHeadModel
    :members: forward

+
 TransfoXLForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.TransfoXLForSequenceClassification
    :members: forward

+
 TFTransfoXLModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -105,3 +107,18 @@ TFTransfoXLLMHeadModel

 .. autoclass:: transformers.TFTransfoXLLMHeadModel
    :members: call
+
+
+TFTransfoXLForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFTransfoXLForSequenceClassification
+    :members: call
+
+
+Internal Layers
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AdaptiveEmbedding
+
+.. autoclass:: transformers.TFAdaptiveEmbedding
@@ -62,6 +62,13 @@ XLMRobertaTokenizer
        create_token_type_ids_from_sequences, save_vocabulary


+XLMRobertaTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMRobertaTokenizerFast
+    :members:
+
+
 XLMRobertaModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -62,6 +62,13 @@ XLNetTokenizer
        create_token_type_ids_from_sequences, save_vocabulary


+XLNetTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLNetTokenizerFast
+    :members:
+
+
 XLNet specific outputs
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -60,7 +60,7 @@ Basic steps
 In order to upload a model, you'll need to first create a git repo. This repo will live on the model hub, allowing
 users to clone it and you (and your organization members) to push to it.

-You can create a model repo directly from the website, `here <https://huggingface.co/new>`.
+You can create a model repo **directly from `the /new page on the website <https://huggingface.co/new>`__.**

 Alternatively, you can use the ``transformers-cli``. The next steps describe that process:

@@ -82,12 +82,12 @@ This creates a repo on the model hub, which can be cloned.

 .. code-block:: bash

-    git clone https://huggingface.co/username/your-model-name
-
    # Make sure you have git-lfs installed
    # (https://git-lfs.github.com/)
    git lfs install

+    git clone https://huggingface.co/username/your-model-name
+
 When you have your local clone of your repo and lfs installed, you can then add/remove from that clone as you would
 with any other git repo.

@@ -98,8 +98,12 @@ with any other git repo.
    echo "hello" >> README.md
    git add . && git commit -m "Update from $USER"

-We are intentionally not wrapping git too much, so as to stay intuitive and easy-to-use.
+We are intentionally not wrapping git too much, so that you can go on with the workflow you're used to and the tools
+you already know.

+The only learning curve you might have compared to regular git is the one for git-lfs. The documentation at
+`git-lfs.github.com <https://git-lfs.github.com/>`__ is decent, but we'll work on a tutorial with some tips and tricks
+in the coming weeks!

 Make your model work on all frameworks
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -110,7 +114,7 @@ Make your model work on all frameworks
 You probably have your favorite framework, but so will other users! That's why it's best to upload your model with both
 PyTorch `and` TensorFlow checkpoints to make it easier to use (if you skip this step, users will still be able to load
 your model in another framework, but it will be slower, as it will have to be converted on the fly). Don't worry, it's
-super easy to do (and in a future version, it will all be automatic). You will need to install both PyTorch and
+super easy to do (and in a future version, it might all be automatic). You will need to install both PyTorch and
 TensorFlow for this step, but you don't need to worry about the GPU, so it should be very easy. Check the `TensorFlow
 installation page <https://www.tensorflow.org/install/pip#tensorflow-2.0-rc-is-available>`__ and/or the `PyTorch
 installation page <https://pytorch.org/get-started/locally/#start-locally>`__ to see how.
@@ -192,7 +196,7 @@ status`` command:
    git add --all
    git status

-Finally, the files should be comitted:
+Finally, the files should be committed:

 .. code-block:: bash

@@ -210,23 +214,20 @@ This will upload the folder containing the weights, tokenizer and configuration
 Add a model card
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

-To make sure everyone knows what your model can do, what its limitations and potential bias or ethetical
-considerations, please add a README.md model card to the 🤗 Transformers repo under `model_cards/`. It should then be
-placed in a subfolder with your username or organization, then another subfolder named like your model
-(`awesome-name-you-picked`). Or just click on the "Create a model card on GitHub" button on the model page, it will get
-you directly to the right location. If you need one, `here <https://github.com/huggingface/model_card>`__ is a model
-card template (meta-suggestions are welcome).
+To make sure everyone knows what your model can do, what its limitations, potential bias or ethical considerations are,
+please add a README.md model card to your model repo. You can just create it, or there's also a convenient button
+titled "Add a README.md" on your model page. A model card template can be found `here
+<https://github.com/huggingface/model_card>`__ (meta-suggestions are welcome). model card template (meta-suggestions
+are welcome).
+
+.. note::
+
+    Model cards used to live in the 🤗 Transformers repo under `model_cards/`, but for consistency and scalability we
+    migrated every model card from the repo to its corresponding huggingface.co model repo.

 If your model is fine-tuned from another model coming from the model hub (all 🤗 Transformers pretrained models do),
 don't forget to link to its model card so that people can fully trace how your model was built.

-If you have never made a pull request to the 🤗 Transformers repo, look at the :doc:`contributing guide <contributing>`
-to see the steps to follow.
-
-.. note::
-
-    You can also send your model card in the folder you uploaded with the CLI by placing it in a `README.md` file
-    inside `path/to/awesome-name-you-picked/`.

 Using your model
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -262,7 +263,8 @@ First you need to install `git-lfs` in the environment used by the notebook:

    sudo apt-get install git-lfs

-Then you can use the :obj:`transformers-cli` to create your new repo:
+Then you can use either create a repo directly from `huggingface.co <https://huggingface.co/>`__ , or use the
+:obj:`transformers-cli` to create it:


 .. code-block:: bash
@@ -274,13 +276,14 @@ Once it's created, you can clone it and configure it (replace username by your u

 .. code-block:: bash

+    git lfs install
+
    git clone https://username:password@huggingface.co/username/your-model-name
    # Alternatively if you have a token,
    # you can use it instead of your password
    git clone https://username:token@huggingface.co/username/your-model-name

    cd your-model-name
-    git lfs install
    git config --global user.email "email@example.com"
    # Tip: using the same email than for your huggingface.co account will link your commits to your profile
    git config --global user.name "Your name"
@@ -16,7 +16,7 @@ Summary of the models
 This is a summary of the models available in 🤗 Transformers. It assumes you’re familiar with the original `transformer
 model <https://arxiv.org/abs/1706.03762>`_. For a gentle introduction check the `annotated transformer
 <http://nlp.seas.harvard.edu/2018/04/03/attention.html>`_. Here we focus on the high-level differences between the
-models. You can check them more in detail in their respective documentation. Also checkout the :doc:`pretrained model
+models. You can check them more in detail in their respective documentation. Also check out the :doc:`pretrained model
 page </pretrained_models>` to see the checkpoints available for each type of model and all `the community models
 <https://huggingface.co/models>`_.

@@ -30,7 +30,7 @@ Each one of the models in the library falls into one of the following categories

 Autoregressive models are pretrained on the classic language modeling task: guess the next token having read all the
 previous ones. They correspond to the decoder of the original transformer model, and a mask is used on top of the full
-sentence so that the attention heads can only see what was before in the next, and not what’s after. Although those
+sentence so that the attention heads can only see what was before in the text, and not what’s after. Although those
 models can be fine-tuned and achieve great results on many tasks, the most natural application is text generation. A
 typical example of such models is GPT.

@@ -512,8 +512,8 @@ BART
 <https://arxiv.org/abs/1910.13461>`_, Mike Lewis et al.

 Sequence-to-sequence model with an encoder and a decoder. Encoder is fed a corrupted version of the tokens, decoder is
-fed the original tokens (but has a mask to hide the future words like a regular transformers decoder). For the encoder
-, on the pretraining tasks, a composition of the following transformations are applied:
+fed the original tokens (but has a mask to hide the future words like a regular transformers decoder). A composition of
+the following transformations are applied on the pretraining tasks for the encoder:

  * mask random tokens (like in BERT)
  * delete random tokens
@@ -78,7 +78,7 @@ The library is built around three types of classes for each model:
 All these classes can be instantiated from pretrained instances and saved locally using two methods:

 - :obj:`from_pretrained()` lets you instantiate a model/configuration/tokenizer from a pretrained version either
-  provided by the library itself (the supported models are provided in the list :doc:`here <pretrained_models>` or
+  provided by the library itself (the supported models are provided in the list :doc:`here <pretrained_models>`) or
  stored locally (or on a server) by the user,
 - :obj:`save_pretrained()` lets you save a model/configuration/tokenizer locally so that it can be reloaded using
  :obj:`from_pretrained()`.
@@ -10,17 +10,17 @@
    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.

-reprocessing data
+Preprocessing data
 =======================================================================================================================

 In this tutorial, we'll explore how to preprocess your data using 🤗 Transformers. The main tool for this is what we
 call a :doc:`tokenizer <main_classes/tokenizer>`. You can build one using the tokenizer class associated to the model
 you would like to use, or directly with the :class:`~transformers.AutoTokenizer` class.

-As we saw in the :doc:`quicktour </quicktour>`, the tokenizer will first split a given text in words (or part of words,
-punctuation symbols, etc.) usually called `tokens`. Then it will convert those `tokens` into numbers, to be able to
-build a tensor out of them and feed them to the model. It will also add any additional inputs the model might expect to
-work properly.
+As we saw in the :doc:`quick tour </quicktour>`, the tokenizer will first split a given text in words (or part of
+words, punctuation symbols, etc.) usually called `tokens`. Then it will convert those `tokens` into numbers, to be able
+to build a tensor out of them and feed them to the model. It will also add any additional inputs the model might expect
+to work properly.

 .. note::

@@ -131,7 +131,7 @@ ones it should not (because they represent padding in this case).


 Note that if your model does not have a maximum length associated to it, the command above will throw a warning. You
-can safely ignore it. You can also pass ``verbose=False`` to stop the tokenizer to throw those kinds of warnings.
+can safely ignore it. You can also pass ``verbose=False`` to stop the tokenizer from throwing those kinds of warnings.

 .. _sentence-pairs:

@@ -216,7 +216,6 @@ Everything you always wanted to know about padding and truncation
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 We have seen the commands that will work for most cases (pad your batch to the length of the maximum sentence and
-
 truncate to the maximum length the mode can accept). However, the API supports more strategies if you need them. The
 three arguments you need to know for this are :obj:`padding`, :obj:`truncation` and :obj:`max_length`.

@@ -158,7 +158,7 @@ Using the tokenizer

 We mentioned the tokenizer is responsible for the preprocessing of your texts. First, it will split a given text in
 words (or part of words, punctuation symbols, etc.) usually called `tokens`. There are multiple rules that can govern
-that process (you can learn more about them in the :doc:`tokenizer summary <tokenizer_summary>`, which is why we need
+that process (you can learn more about them in the :doc:`tokenizer summary <tokenizer_summary>`), which is why we need
 to instantiate the tokenizer using the name of the model, to make sure we use the same rules as when the model was
 pretrained.

@@ -327,7 +327,7 @@ Masked Language Modeling
 Masked language modeling is the task of masking tokens in a sequence with a masking token, and prompting the model to
 fill that mask with an appropriate token. This allows the model to attend to both the right context (tokens on the
 right of the mask) and the left context (tokens on the left of the mask). Such a training creates a strong basis for
-downstream tasks, requiring bi-directional context such as SQuAD (question answering, see `Lewis, Lui, Goyal et al.
+downstream tasks requiring bi-directional context, such as SQuAD (question answering, see `Lewis, Lui, Goyal et al.
 <https://arxiv.org/abs/1910.13461>`__, part 4.2).

 Here is an example of using pipelines to replace a mask from a sequence:
@@ -657,7 +657,7 @@ Here are the expected results:
        {'word': 'Bridge', 'score': 0.990249514579773, 'entity': 'I-LOC'}
    ]

-Note, how the tokens of the sequence "Hugging Face" have been identified as an organisation, and "New York City",
+Note how the tokens of the sequence "Hugging Face" have been identified as an organisation, and "New York City",
 "DUMBO" and "Manhattan Bridge" have been identified as locations.

 Here is an example of doing named entity recognition, using a model and a tokenizer. The process is the following:
@@ -1142,3 +1142,66 @@ To start a debugger at the point of the warning, do this:
 .. code-block:: bash

    pytest tests/test_logging.py -W error::UserWarning --pdb
+
+
+
+Testing Experimental CI Features
+-----------------------------------------------------------------------------------------------------------------------
+
+Testing CI features can be potentially problematic as it can interfere with the normal CI functioning. Therefore if a
+new CI feature is to be added, it should be done as following.
+
+1. Create a new dedicated job that tests what needs to be tested
+2. The new job must always succeed so that it gives us a green ✓ (details below).
+3. Let it run for some days to see that a variety of different PR types get to run on it (user fork branches,
+   non-forked branches, branches originating from github.com UI direct file edit, various forced pushes, etc. - there
+   are so many) while monitoring the experimental job's logs (not the overall job green as it's purposefully always
+   green)
+4. When it's clear that everything is solid, then merge the new changes into existing jobs.
+
+That way experiments on CI functionality itself won't interfere with the normal workflow.
+
+Now how can we make the job always succeed while the new CI feature is being developed?
+
+Some CIs, like TravisCI support ignore-step-failure and will report the overall job as successful, but CircleCI and
+Github Actions as of this writing don't support that.
+
+So the following workaround can be used:
+
+1. ``set +euo pipefail`` at the beginning of the run command to suppress most potential failures in the bash script.
+2. the last command must be a success: ``echo "done"`` or just ``true`` will do
+
+Here is an example:
+
+.. code-block:: yaml
+
+    - run:
+        name: run CI experiment
+        command: |
+            set +euo pipefail
+            echo "setting run-all-despite-any-errors-mode"
+            this_command_will_fail
+            echo "but bash continues to run"
+            # emulate another failure
+            false
+            # but the last command must be a success
+            echo "during experiment do not remove: reporting success to CI, even if there were failures"
+
+For simple commands you could also do:
+
+.. code-block:: bash
+
+    cmd_that_may_fail || true
+
+Of course, once satisfied with the results, integrate the experimental step or job with the rest of the normal jobs,
+while removing ``set +euo pipefail`` or any other things you may have added to ensure that the experimental job doesn't
+interfere with the normal CI functioning.
+
+This whole process would have been much easier if we only could set something like ``allow-failure`` for the
+experimental step, and let it fail without impacting the overall status of PRs. But as mentioned earlier CircleCI and
+Github Actions don't support it at the moment.
+
+You can vote for this feature and see where it is at at these CI-specific threads:
+
+* `Github Actions: <https://github.com/actions/toolkit/issues/399>`__
+* `CircleCI: <https://ideas.circleci.com/ideas/CCI-I-344>`__
@@ -18,7 +18,7 @@ On this page, we will have a closer look at tokenization. As we saw in :doc:`the
 look-up table. Converting words or subwords to ids is straightforward, so in this summary, we will focus on splitting a
 text into words or subwords (i.e. tokenizing a text). More specifically, we will look at the three main types of
 tokenizers used in 🤗 Transformers: :ref:`Byte-Pair Encoding (BPE) <byte-pair-encoding>`, :ref:`WordPiece <wordpiece>`,
-and :ref:`SentencePiece <sentencepiece>`, and show exemplary which tokenizer type is used by which model.
+and :ref:`SentencePiece <sentencepiece>`, and show examples of which tokenizer type is used by which model.

 Note that on each model page, you can look at the documentation of the associated tokenizer to know which tokenizer
 type was used by the pretrained model. For instance, if we look at :class:`~transformers.BertTokenizer`, we can see
@@ -72,7 +72,7 @@ greater than 50,000, especially if they are pretrained only on a single language
 So if simple space and punctuation tokenization is unsatisfactory, why not simply tokenize on characters? While
 character tokenization is very simple and would greatly reduce memory and time complexity it makes it much harder for
 the model to learn meaningful input representations. *E.g.* learning a meaningful context-independent representation
-for the letter ``"t"`` is much harder as learning a context-independent representation for the word ``"today"``.
+for the letter ``"t"`` is much harder than learning a context-independent representation for the word ``"today"``.
 Therefore, character tokenization is often accompanied by a loss of performance. So to get the best of both worlds,
 transformers models use a hybrid between word-level and character-level tokenization called **subword** tokenization.

@@ -202,10 +202,10 @@ WordPiece

 WordPiece is the subword tokenization algorithm used for :doc:`BERT <model_doc/bert>`, :doc:`DistilBERT
 <model_doc/distilbert>`, and :doc:`Electra <model_doc/electra>`. The algorithm was outlined in `Japanese and Korean
-Voice Seach (Schuster et al., 2012)
+Voice Search (Schuster et al., 2012)
 <https://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/37842.pdf>`__ and is very similar to
 BPE. WordPiece first initializes the vocabulary to include every character present in the training data and
-progressively learn a given number of merge rules. In contrast to BPE, WordPiece does not choose the most frequent
+progressively learns a given number of merge rules. In contrast to BPE, WordPiece does not choose the most frequent
 symbol pair, but the one that maximizes the likelihood of the training data once added to the vocabulary.

 So what does this mean exactly? Referring to the previous example, maximizing the likelihood of the training data is
@@ -14,7 +14,7 @@ Training and fine-tuning
 =======================================================================================================================

 Model classes in 🤗 Transformers are designed to be compatible with native PyTorch and TensorFlow 2 and can be used
-seemlessly with either. In this quickstart, we will show how to fine-tune (or train from scratch) a model using the
+seamlessly with either. In this quickstart, we will show how to fine-tune (or train from scratch) a model using the
 standard training tools available in either framework. We will also show how to use our included
 :func:`~transformers.Trainer` class which handles much of the complexity of training for you.

@@ -16,59 +16,95 @@ limitations under the License.

 # Examples

-Version 2.9 of 🤗 Transformers introduced a new [`Trainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py) class for PyTorch, and its equivalent [`TFTrainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer_tf.py) for TF 2.
-Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.2+.
-
-Here is the list of all our examples:
- **grouped by task** (all official examples work for multiple models)
- with information on whether they are **built on top of `Trainer`/`TFTrainer`** (if not, they still work, they might
-  just lack some features),
- whether or not they leverage the [🤗 Datasets](https://github.com/huggingface/datasets) library.
- links to **Colab notebooks** to walk through the scripts and run them easily,
- links to **Cloud deployments** to be able to deploy large-scale trainings in the Cloud with little to no setup.
-
+This folder contains actively maintained examples of use of 🤗 Transformers organized along NLP tasks. If you are looking for an example that used to
+be in this folder, it may have moved to our [research projects](https://github.com/huggingface/transformers/tree/master/examples/research_projects) subfolder (which contains frozen snapshots of research projects).

 ## Important note

 **Important**

-To make sure you can successfully run the latest versions of the example scripts, you have to **install the library from source** and install some example-specific requirements.
-Execute the following steps in a new virtual environment:
-
+To make sure you can successfully run the latest versions of the example scripts, you have to **install the library from source** and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
 ```bash
 git clone https://github.com/huggingface/transformers
 cd transformers
 pip install .
-pip install -r ./examples/requirements.txt
+```
+Then cd in the example folder of your choice and run
+```bash
+pip install -r requirements.txt
 ```

-Alternatively, you can run the version of the examples as they were for your current version of Transformers via (for instance with v3.4.0):
+Alternatively, you can run the version of the examples as they were for your current version of Transformers via (for instance with v3.5.1):
 ```bash
-git checkout tags/v3.4.0
+git checkout tags/v3.5.1
 ```

 ## The Big Table of Tasks

+Here is the list of all our examples:
+- with information on whether they are **built on top of `Trainer`/`TFTrainer`** (if not, they still work, they might
+  just lack some features),
+- whether or not they leverage the [🤗 Datasets](https://github.com/huggingface/datasets) library.
+- links to **Colab notebooks** to walk through the scripts and run them easily,
+<!--
+Coming soon!
+- links to **Cloud deployments** to be able to deploy large-scale trainings in the Cloud with little to no setup.
+-->
+
 | Task | Example datasets | Trainer support | TFTrainer support | 🤗 Datasets | Colab
 |---|---|:---:|:---:|:---:|:---:|
 | [**`language-modeling`**](https://github.com/huggingface/transformers/tree/master/examples/language-modeling)       | Raw text        | ✅ | -  | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/01_how_to_train.ipynb)
-| [**`text-classification`**](https://github.com/huggingface/transformers/tree/master/examples/text-classification)   | GLUE, XNLI      | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://github.com/huggingface/notebooks/blob/master/examples/text_classification.ipynb)
-| [**`token-classification`**](https://github.com/huggingface/transformers/tree/master/examples/token-classification) | CoNLL NER       | ✅ | ✅ | ✅ | -
-| [**`multiple-choice`**](https://github.com/huggingface/transformers/tree/master/examples/multiple-choice)           | SWAG, RACE, ARC | ✅ | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ViktorAlm/notebooks/blob/master/MPC_GPU_Demo_for_TF_and_PT.ipynb)
-| [**`question-answering`**](https://github.com/huggingface/transformers/tree/master/examples/question-answering)     | SQuAD           | ✅ | ✅ | ✅ | -
-| [**`text-generation`**](https://github.com/huggingface/transformers/tree/master/examples/text-generation)           | -               | n/a | n/a | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/02_how_to_generate.ipynb)
-| [**`distillation`**](https://github.com/huggingface/transformers/tree/master/examples/distillation)                 | All             | - | -  | - | -
+| [**`multiple-choice`**](https://github.com/huggingface/transformers/tree/master/examples/multiple-choice)           | SWAG, RACE, ARC | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ViktorAlm/notebooks/blob/master/MPC_GPU_Demo_for_TF_and_PT.ipynb)
+| [**`question-answering`**](https://github.com/huggingface/transformers/tree/master/examples/question-answering)     | SQuAD           | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://github.com/huggingface/notebooks/blob/master/examples/question_answering.ipynb)
 | [**`summarization`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)                     | CNN/Daily Mail  | ✅  | - | - | -
+| [**`text-classification`**](https://github.com/huggingface/transformers/tree/master/examples/text-classification)   | GLUE, XNLI      | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://github.com/huggingface/notebooks/blob/master/examples/text_classification.ipynb)
+| [**`text-generation`**](https://github.com/huggingface/transformers/tree/master/examples/text-generation)           | -               | n/a | n/a | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/02_how_to_generate.ipynb)
+| [**`token-classification`**](https://github.com/huggingface/transformers/tree/master/examples/token-classification) | CoNLL NER       | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://github.com/huggingface/notebooks/blob/master/examples/token_classification.ipynb)
 | [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)                       | WMT             | ✅  | - | - | -
-| [**`bertology`**](https://github.com/huggingface/transformers/tree/master/examples/bertology)                       | -               | - | - | - | -
-| [**`adversarial`**](https://github.com/huggingface/transformers/tree/master/examples/adversarial)                   | HANS            | ✅ | - | - | -


-<br>
-
+<!--
 ## One-click Deploy to Cloud (wip)

 **Coming soon!**
+-->
+
+## Distributed training and mixed precision
+
+All the PyTorch scripts mentioned above work out of the box with distributed training and mixed precision, thanks to
+the [Trainer API](https://huggingface.co/transformers/main_classes/trainer.html). To launch one of them on _n_ GPUS,
+use the following command:
+
+```bash
+python -m torch.distributed.launch \
+    --nproc_per_node number_of_gpu_you_have path_to_script.py \
+	--all_arguments_of_the_script 
+```
+
+As an example, here is how you would fine-tune the BERT large model (with whole word masking) on the text
+classification MNLI task using the `run_glue` script, with 8 GPUs:
+
+```bash
+python -m torch.distributed.launch \
+    --nproc_per_node 8 text-classification/run_glue.py \
+    --model_name_or_path bert-large-uncased-whole-word-masking \
+    --task_name mnli \
+    --do_train \
+    --do_eval \
+    --max_seq_length 128 \
+    --per_device_train_batch_size 8 \
+    --learning_rate 2e-5 \
+    --num_train_epochs 3.0 \
+    --output_dir /tmp/mnli_output/
+```
+
+If you have a GPU with mixed precision capabilities (architecture Pascal or more recent), you can use mixed precision
+training with PyTorch 1.6.0 or latest, or by installing the [Apex](https://github.com/NVIDIA/apex) library for previous
+versions. Just add the flag `--fp16` to your command launching one of the scripts mentioned above!
+
+Using mixed precision training usually results in 2x-speedup for training with the same final results (as shown in
+[this table](https://github.com/huggingface/transformers/tree/master/examples/text-classification#mixed-precision-training)
+for text classification).

 ## Running on TPUs

@@ -77,27 +113,34 @@ When using Tensorflow, TPUs are supported out of the box as a `tf.distribute.Str
 When using PyTorch, we support TPUs thanks to `pytorch/xla`. For more context and information on how to setup your TPU environment refer to Google's documentation and to the
 very detailed [pytorch/xla README](https://github.com/pytorch/xla/blob/master/README.md).

-In this repo, we provide a very simple launcher script named [xla_spawn.py](https://github.com/huggingface/transformers/tree/master/examples/xla_spawn.py) that lets you run our example scripts on multiple TPU cores without any boilerplate.
-Just pass a `--num_cores` flag to this script, then your regular training script with its arguments (this is similar to the `torch.distributed.launch` helper for torch.distributed). 
-Note that this approach does not work for examples that use `pytorch-lightning`.
-
-For example for `run_glue`:
+In this repo, we provide a very simple launcher script named
+[xla_spawn.py](https://github.com/huggingface/transformers/tree/master/examples/xla_spawn.py) that lets you run our
+example scripts on multiple TPU cores without any boilerplate. Just pass a `--num_cores` flag to this script, then your
+regular training script with its arguments (this is similar to the `torch.distributed.launch` helper for
+`torch.distributed`):

 ```bash
-python examples/xla_spawn.py --num_cores 8 \
-	examples/text-classification/run_glue.py \
-	--model_name_or_path bert-base-cased \
-	--task_name mnli \
-	--data_dir ./data/glue_data/MNLI \
-	--output_dir ./models/tpu \
-	--overwrite_output_dir \
-	--do_train \
-	--do_eval \
-	--num_train_epochs 1 \
-	--save_steps 20000
+python xla_spawn.py --num_cores num_tpu_you_have \
+    path_to_script.py \
+	--all_arguments_of_the_script 
 ```

-Feedback and more use cases and benchmarks involving TPUs are welcome, please share with the community.
+As an example, here is how you would fine-tune the BERT large model (with whole word masking) on the text
+classification MNLI task using the `run_glue` script, with 8 TPUs:
+
+```bash
+python xla_spawn.py --num_cores 8 \
+    text-classification/run_glue.py \
+    --model_name_or_path bert-large-uncased-whole-word-masking \
+    --task_name mnli \
+    --do_train \
+    --do_eval \
+    --max_seq_length 128 \
+    --per_device_train_batch_size 8 \
+    --learning_rate 2e-5 \
+    --num_train_epochs 3.0 \
+    --output_dir /tmp/mnli_output/
+```

 ## Logging & Experiment tracking

@@ -0,0 +1,20 @@
+tensorboard
+scikit-learn
+seqeval
+psutil
+sacrebleu
+rouge-score
+tensorflow_datasets
+matplotlib
+git-python==1.0.3
+faiss-cpu
+streamlit
+elasticsearch
+nltk
+pandas
+datasets >= 1.1.3
+fire
+pytest
+conllu
+sentencepiece != 0.1.92
+protobuf
@@ -1,3 +1,19 @@
+<!---
+Copyright 2020 The HuggingFace Team. 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.
+-->
+
 # 🤗 Benchmark results

 Here, you can find a list of the different benchmark results created by the community.
@@ -1,3 +1,17 @@
+# Copyright 2020 The HuggingFace Team. 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.
+
 import csv
 from collections import defaultdict
 from dataclasses import dataclass, field
@@ -1,3 +1,17 @@
+# Copyright 2020 The HuggingFace Team. 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.
+
 # tests directory-specific settings - this file is run automatically
 # by pytest before any tests are run

@@ -1,5 +0,0 @@
-# Community contributed examples
-
-This folder contains examples which are not actively maintained (mostly contributed by the community).
-
-Using these examples together with a recent version of the library usually requires to make small (sometimes big) adaptations to get the scripts working.
@@ -1,3 +1,19 @@
+<!---
+Copyright 2020 The HuggingFace Team. 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.
+-->
+
 ## Language model training

 Fine-tuning (or training from scratch) the library models for language modeling on a text dataset for GPT, GPT-2,
@@ -9,8 +25,7 @@ objectives in our [model summary](https://huggingface.co/transformers/model_summ
 These scripts leverage the 🤗 Datasets library and the Trainer API. You can easily customize them to your needs if you
 need extra processing on your datasets.

-**Note:** The old script `run_language_modeling.py` is still available
-[here](https://github.com/huggingface/transformers/blob/master/examples/contrib/legacy/run_language_modeling.py).
+**Note:** The old script `run_language_modeling.py` is still available [here](https://github.com/huggingface/transformers/blob/master/examples/legacy/run_language_modeling.py).

 The following examples, will run on a datasets hosted on our [hub](https://huggingface.co/datasets) or with your own
 text files for training and validation. We give examples of both below.
@@ -0,0 +1,3 @@
+datasets >= 1.1.3
+sentencepiece != 0.1.92
+protobuf
@@ -113,6 +113,12 @@ class DataTrainingArguments:
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
@@ -188,6 +194,17 @@ def main():
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+            )
    else:
        data_files = {}
        if data_args.train_file is not None:
@@ -324,9 +341,20 @@ def main():
            if (model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path))
            else None
        )
-        trainer.train(model_path=model_path)
+        train_result = trainer.train(model_path=model_path)
        trainer.save_model()  # Saves the tokenizer too for easy upload

+        output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
+        if trainer.is_world_process_zero():
+            with open(output_train_file, "w") as writer:
+                logger.info("***** Train results *****")
+                for key, value in sorted(train_result.metrics.items()):
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+            # Need to save the state, since Trainer.save_model saves only the tokenizer with the model
+            trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
+
    # Evaluation
    results = {}
    if training_args.do_eval:
@@ -341,7 +369,7 @@ def main():
        if trainer.is_world_process_zero():
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
-                for key, value in results.items():
+                for key, value in sorted(results.items()):
                    logger.info(f"  {key} = {value}")
                    writer.write(f"{key} = {value}\n")

@@ -103,6 +103,12 @@ class DataTrainingArguments:
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
    max_seq_length: Optional[int] = field(
        default=None,
        metadata={
@@ -199,6 +205,17 @@ def main():
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+            )
    else:
        data_files = {}
        if data_args.train_file is not None:
@@ -359,9 +376,20 @@ def main():
            if (model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path))
            else None
        )
-        trainer.train(model_path=model_path)
+        train_result = trainer.train(model_path=model_path)
        trainer.save_model()  # Saves the tokenizer too for easy upload

+        output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
+        if trainer.is_world_process_zero():
+            with open(output_train_file, "w") as writer:
+                logger.info("***** Train results *****")
+                for key, value in sorted(train_result.metrics.items()):
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+            # Need to save the state, since Trainer.save_model saves only the tokenizer with the model
+            trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
+
    # Evaluation
    results = {}
    if training_args.do_eval:
@@ -376,7 +404,7 @@ def main():
        if trainer.is_world_process_zero():
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
-                for key, value in results.items():
+                for key, value in sorted(results.items()):
                    logger.info(f"  {key} = {value}")
                    writer.write(f"{key} = {value}\n")

@@ -134,6 +134,12 @@ class DataTrainingArguments:
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
    max_seq_length: Optional[int] = field(
        default=None,
        metadata={
@@ -379,7 +385,7 @@ def training_step(optimizer, batch, dropout_rng):
        # Hide away tokens which doesn't participate in the optimization
        token_mask = jnp.where(targets > 0, 1.0, 0.0)

-        pooled, logits = model(**batch, params=params, dropout_rng=dropout_rng, train=True)
+        logits = model(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
        loss, weight_sum = cross_entropy(logits, targets, token_mask)
        return loss / weight_sum

@@ -401,7 +407,7 @@ def eval_step(params, batch):

    # Hide away tokens which doesn't participate in the optimization
    token_mask = jnp.where(targets > 0, 1.0, 0.0)
-    _, logits = model(**batch, params=params, train=False)
+    logits = model(**batch, params=params, train=False)[0]

    return compute_metrics(logits, targets, token_mask)

@@ -413,7 +419,7 @@ def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndar
    if samples_to_remove != 0:
        samples_idx = samples_idx[:-samples_to_remove]
    sections_split = nb_samples // batch_size
-    batch_idx = jnp.split(samples_idx, sections_split)
+    batch_idx = np.split(samples_idx, sections_split)
    return batch_idx


@@ -473,6 +479,17 @@ if __name__ == "__main__":
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+            )
    else:
        data_files = {}
        if data_args.train_file is not None:
@@ -525,9 +542,9 @@ if __name__ == "__main__":

    def tokenize_function(examples):
        # Remove empty lines
-        examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
+        examples = [line for line in examples if len(line) > 0 and not line.isspace()]
        return tokenizer(
-            examples["text"],
+            examples,
            return_special_tokens_mask=True,
            padding=padding,
            truncation=True,
@@ -536,9 +553,10 @@ if __name__ == "__main__":

    tokenized_datasets = datasets.map(
        tokenize_function,
+        input_columns=[text_column_name],
        batched=True,
        num_proc=data_args.preprocessing_num_workers,
-        remove_columns=[text_column_name],
+        remove_columns=column_names,
        load_from_cache_file=not data_args.overwrite_cache,
    )

@@ -554,8 +572,13 @@ if __name__ == "__main__":
    rng = jax.random.PRNGKey(training_args.seed)
    dropout_rngs = jax.random.split(rng, jax.local_device_count())

-    model = FlaxBertForMaskedLM.from_pretrained("bert-base-cased", dtype=jnp.float32, dropout_rate=0.1)
-    model.init(jax.random.PRNGKey(training_args.seed), (training_args.train_batch_size, model.config.max_length))
+    model = FlaxBertForMaskedLM.from_pretrained(
+        "bert-base-cased",
+        dtype=jnp.float32,
+        input_shape=(training_args.train_batch_size, config.max_position_embeddings),
+        seed=training_args.seed,
+        dropout_rate=0.1,
+    )

    # Setup optimizer
    optimizer = Adam(
@@ -566,8 +589,9 @@ if __name__ == "__main__":
    ).create(model.params)

    # Create learning rate scheduler
+    # warmup_steps = 0 causes the Flax optimizer to return NaNs; warmup_steps = 1 is functionally equivalent.
    lr_scheduler_fn = create_learning_rate_scheduler(
-        base_learning_rate=training_args.learning_rate, warmup_steps=training_args.warmup_steps
+        base_learning_rate=training_args.learning_rate, warmup_steps=min(training_args.warmup_steps, 1)
    )

    # Create parallel version of the training and evaluation steps
@@ -606,13 +630,13 @@ if __name__ == "__main__":
        epochs.write(f"Loss: {loss}")

        # ======================== Evaluating ==============================
-        nb_eval_samples = len(tokenized_datasets["test"])
+        nb_eval_samples = len(tokenized_datasets["validation"])
        eval_samples_idx = jnp.arange(nb_eval_samples)
        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)

        eval_metrics = []
        for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
-            samples = [tokenized_datasets["test"][int(idx)] for idx in batch_idx]
+            samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
            model_inputs = data_collator(samples, pad_to_multiple_of=16)

            # Model forward
@@ -91,6 +91,12 @@ class DataTrainingArguments:
    Arguments pertaining to what data we are going to input our model for training and eval.
    """

+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
    validation_file: Optional[str] = field(
        default=None,
@@ -107,6 +113,12 @@ class DataTrainingArguments:
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
    max_seq_length: Optional[int] = field(
        default=None,
        metadata={
@@ -203,15 +215,30 @@ def main():
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
-    data_files = {}
-    if data_args.train_file is not None:
-        data_files["train"] = data_args.train_file
-    if data_args.validation_file is not None:
-        data_files["validation"] = data_args.validation_file
-    extension = data_args.train_file.split(".")[-1]
-    if extension == "txt":
-        extension = "text"
-    datasets = load_dataset(extension, data_files=data_files)
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        datasets = load_dataset(extension, data_files=data_files)
    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

@@ -281,7 +308,7 @@ def main():
    # Add the chinese references if provided
    if data_args.train_ref_file is not None:
        tokenized_datasets["train"] = add_chinese_references(tokenized_datasets["train"], data_args.train_ref_file)
-    if data_args.valid_ref_file is not None:
+    if data_args.validation_ref_file is not None:
        tokenized_datasets["validation"] = add_chinese_references(
            tokenized_datasets["validation"], data_args.validation_ref_file
        )
@@ -307,9 +334,20 @@ def main():
            if (model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path))
            else None
        )
-        trainer.train(model_path=model_path)
+        train_result = trainer.train(model_path=model_path)
        trainer.save_model()  # Saves the tokenizer too for easy upload

+        output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
+        if trainer.is_world_process_zero():
+            with open(output_train_file, "w") as writer:
+                logger.info("***** Train results *****")
+                for key, value in sorted(train_result.metrics.items()):
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+            # Need to save the state, since Trainer.save_model saves only the tokenizer with the model
+            trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
+
    # Evaluation
    results = {}
    if training_args.do_eval:
@@ -324,7 +362,7 @@ def main():
        if trainer.is_world_process_zero():
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
-                for key, value in results.items():
+                for key, value in sorted(results.items()):
                    logger.info(f"  {key} = {value}")
                    writer.write(f"{key} = {value}\n")

@@ -93,6 +93,12 @@ class DataTrainingArguments:
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
    max_seq_length: int = field(
        default=512,
        metadata={
@@ -196,6 +202,17 @@ def main():
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+            )
    else:
        data_files = {}
        if data_args.train_file is not None:
@@ -346,9 +363,20 @@ def main():
            if (model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path))
            else None
        )
-        trainer.train(model_path=model_path)
+        train_result = trainer.train(model_path=model_path)
        trainer.save_model()  # Saves the tokenizer too for easy upload

+        output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
+        if trainer.is_world_process_zero():
+            with open(output_train_file, "w") as writer:
+                logger.info("***** Train results *****")
+                for key, value in sorted(train_result.metrics.items()):
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+            # Need to save the state, since Trainer.save_model saves only the tokenizer with the model
+            trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
+
    # Evaluation
    results = {}
    if training_args.do_eval:
@@ -363,7 +391,7 @@ def main():
        if trainer.is_world_process_zero():
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
-                for key, value in results.items():
+                for key, value in sorted(results.items()):
                    logger.info(f"  {key} = {value}")
                    writer.write(f"{key} = {value}\n")

@@ -0,0 +1,21 @@
+<!---
+Copyright 2020 The HuggingFace Team. 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.
+-->
+
+# Legacy examples
+
+This folder contains examples which are not actively maintained (mostly contributed by the community).
+
+Using these examples together with a recent version of the library usually requires to make small (sometimes big) adaptations to get the scripts working.
@@ -0,0 +1,579 @@
+# 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.
+""" Multiple choice fine-tuning: utilities to work with multiple choice tasks of reading comprehension """
+
+
+import csv
+import glob
+import json
+import logging
+import os
+from dataclasses import dataclass
+from enum import Enum
+from typing import List, Optional
+
+import tqdm
+
+from filelock import FileLock
+from transformers import PreTrainedTokenizer, is_tf_available, is_torch_available
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class InputExample:
+    """
+    A single training/test example for multiple choice
+
+    Args:
+        example_id: Unique id for the example.
+        question: string. The untokenized text of the second sequence (question).
+        contexts: list of str. The untokenized text of the first sequence (context of corresponding question).
+        endings: list of str. multiple choice's options. Its length must be equal to contexts' length.
+        label: (Optional) string. The label of the example. This should be
+        specified for train and dev examples, but not for test examples.
+    """
+
+    example_id: str
+    question: str
+    contexts: List[str]
+    endings: List[str]
+    label: Optional[str]
+
+
+@dataclass(frozen=True)
+class InputFeatures:
+    """
+    A single set of features of data.
+    Property names are the same names as the corresponding inputs to a model.
+    """
+
+    example_id: str
+    input_ids: List[List[int]]
+    attention_mask: Optional[List[List[int]]]
+    token_type_ids: Optional[List[List[int]]]
+    label: Optional[int]
+
+
+class Split(Enum):
+    train = "train"
+    dev = "dev"
+    test = "test"
+
+
+if is_torch_available():
+    import torch
+    from torch.utils.data.dataset import Dataset
+
+    class MultipleChoiceDataset(Dataset):
+        """
+        This will be superseded by a framework-agnostic approach
+        soon.
+        """
+
+        features: List[InputFeatures]
+
+        def __init__(
+            self,
+            data_dir: str,
+            tokenizer: PreTrainedTokenizer,
+            task: str,
+            max_seq_length: Optional[int] = None,
+            overwrite_cache=False,
+            mode: Split = Split.train,
+        ):
+            processor = processors[task]()
+
+            cached_features_file = os.path.join(
+                data_dir,
+                "cached_{}_{}_{}_{}".format(
+                    mode.value,
+                    tokenizer.__class__.__name__,
+                    str(max_seq_length),
+                    task,
+                ),
+            )
+
+            # Make sure only the first process in distributed training processes the dataset,
+            # and the others will use the cache.
+            lock_path = cached_features_file + ".lock"
+            with FileLock(lock_path):
+
+                if os.path.exists(cached_features_file) and not overwrite_cache:
+                    logger.info(f"Loading features from cached file {cached_features_file}")
+                    self.features = torch.load(cached_features_file)
+                else:
+                    logger.info(f"Creating features from dataset file at {data_dir}")
+                    label_list = processor.get_labels()
+                    if mode == Split.dev:
+                        examples = processor.get_dev_examples(data_dir)
+                    elif mode == Split.test:
+                        examples = processor.get_test_examples(data_dir)
+                    else:
+                        examples = processor.get_train_examples(data_dir)
+                    logger.info("Training examples: %s", len(examples))
+                    self.features = convert_examples_to_features(
+                        examples,
+                        label_list,
+                        max_seq_length,
+                        tokenizer,
+                    )
+                    logger.info("Saving features into cached file %s", cached_features_file)
+                    torch.save(self.features, cached_features_file)
+
+        def __len__(self):
+            return len(self.features)
+
+        def __getitem__(self, i) -> InputFeatures:
+            return self.features[i]
+
+
+if is_tf_available():
+    import tensorflow as tf
+
+    class TFMultipleChoiceDataset:
+        """
+        This will be superseded by a framework-agnostic approach
+        soon.
+        """
+
+        features: List[InputFeatures]
+
+        def __init__(
+            self,
+            data_dir: str,
+            tokenizer: PreTrainedTokenizer,
+            task: str,
+            max_seq_length: Optional[int] = 128,
+            overwrite_cache=False,
+            mode: Split = Split.train,
+        ):
+            processor = processors[task]()
+
+            logger.info(f"Creating features from dataset file at {data_dir}")
+            label_list = processor.get_labels()
+            if mode == Split.dev:
+                examples = processor.get_dev_examples(data_dir)
+            elif mode == Split.test:
+                examples = processor.get_test_examples(data_dir)
+            else:
+                examples = processor.get_train_examples(data_dir)
+            logger.info("Training examples: %s", len(examples))
+
+            self.features = convert_examples_to_features(
+                examples,
+                label_list,
+                max_seq_length,
+                tokenizer,
+            )
+
+            def gen():
+                for (ex_index, ex) in tqdm.tqdm(enumerate(self.features), desc="convert examples to features"):
+                    if ex_index % 10000 == 0:
+                        logger.info("Writing example %d of %d" % (ex_index, len(examples)))
+
+                    yield (
+                        {
+                            "example_id": 0,
+                            "input_ids": ex.input_ids,
+                            "attention_mask": ex.attention_mask,
+                            "token_type_ids": ex.token_type_ids,
+                        },
+                        ex.label,
+                    )
+
+            self.dataset = tf.data.Dataset.from_generator(
+                gen,
+                (
+                    {
+                        "example_id": tf.int32,
+                        "input_ids": tf.int32,
+                        "attention_mask": tf.int32,
+                        "token_type_ids": tf.int32,
+                    },
+                    tf.int64,
+                ),
+                (
+                    {
+                        "example_id": tf.TensorShape([]),
+                        "input_ids": tf.TensorShape([None, None]),
+                        "attention_mask": tf.TensorShape([None, None]),
+                        "token_type_ids": tf.TensorShape([None, None]),
+                    },
+                    tf.TensorShape([]),
+                ),
+            )
+
+        def get_dataset(self):
+            self.dataset = self.dataset.apply(tf.data.experimental.assert_cardinality(len(self.features)))
+
+            return self.dataset
+
+        def __len__(self):
+            return len(self.features)
+
+        def __getitem__(self, i) -> InputFeatures:
+            return self.features[i]
+
+
+class DataProcessor:
+    """Base class for data converters for multiple choice data sets."""
+
+    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_test_examples(self, data_dir):
+        """Gets a collection of `InputExample`s for the test set."""
+        raise NotImplementedError()
+
+    def get_labels(self):
+        """Gets the list of labels for this data set."""
+        raise NotImplementedError()
+
+
+class RaceProcessor(DataProcessor):
+    """Processor for the RACE data set."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} train".format(data_dir))
+        high = os.path.join(data_dir, "train/high")
+        middle = os.path.join(data_dir, "train/middle")
+        high = self._read_txt(high)
+        middle = self._read_txt(middle)
+        return self._create_examples(high + middle, "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} dev".format(data_dir))
+        high = os.path.join(data_dir, "dev/high")
+        middle = os.path.join(data_dir, "dev/middle")
+        high = self._read_txt(high)
+        middle = self._read_txt(middle)
+        return self._create_examples(high + middle, "dev")
+
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} test".format(data_dir))
+        high = os.path.join(data_dir, "test/high")
+        middle = os.path.join(data_dir, "test/middle")
+        high = self._read_txt(high)
+        middle = self._read_txt(middle)
+        return self._create_examples(high + middle, "test")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1", "2", "3"]
+
+    def _read_txt(self, input_dir):
+        lines = []
+        files = glob.glob(input_dir + "/*txt")
+        for file in tqdm.tqdm(files, desc="read files"):
+            with open(file, "r", encoding="utf-8") as fin:
+                data_raw = json.load(fin)
+                data_raw["race_id"] = file
+                lines.append(data_raw)
+        return lines
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (_, data_raw) in enumerate(lines):
+            race_id = "%s-%s" % (set_type, data_raw["race_id"])
+            article = data_raw["article"]
+            for i in range(len(data_raw["answers"])):
+                truth = str(ord(data_raw["answers"][i]) - ord("A"))
+                question = data_raw["questions"][i]
+                options = data_raw["options"][i]
+
+                examples.append(
+                    InputExample(
+                        example_id=race_id,
+                        question=question,
+                        contexts=[article, article, article, article],  # this is not efficient but convenient
+                        endings=[options[0], options[1], options[2], options[3]],
+                        label=truth,
+                    )
+                )
+        return examples
+
+
+class SynonymProcessor(DataProcessor):
+    """Processor for the Synonym data set."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} train".format(data_dir))
+        return self._create_examples(self._read_csv(os.path.join(data_dir, "mctrain.csv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} dev".format(data_dir))
+        return self._create_examples(self._read_csv(os.path.join(data_dir, "mchp.csv")), "dev")
+
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} dev".format(data_dir))
+
+        return self._create_examples(self._read_csv(os.path.join(data_dir, "mctest.csv")), "test")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1", "2", "3", "4"]
+
+    def _read_csv(self, input_file):
+        with open(input_file, "r", encoding="utf-8") as f:
+            return list(csv.reader(f))
+
+    def _create_examples(self, lines: List[List[str]], type: str):
+        """Creates examples for the training and dev sets."""
+
+        examples = [
+            InputExample(
+                example_id=line[0],
+                question="",  # in the swag dataset, the
+                # common beginning of each
+                # choice is stored in "sent2".
+                contexts=[line[1], line[1], line[1], line[1], line[1]],
+                endings=[line[2], line[3], line[4], line[5], line[6]],
+                label=line[7],
+            )
+            for line in lines  # we skip the line with the column names
+        ]
+
+        return examples
+
+
+class SwagProcessor(DataProcessor):
+    """Processor for the SWAG data set."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} train".format(data_dir))
+        return self._create_examples(self._read_csv(os.path.join(data_dir, "train.csv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} dev".format(data_dir))
+        return self._create_examples(self._read_csv(os.path.join(data_dir, "val.csv")), "dev")
+
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} dev".format(data_dir))
+        raise ValueError(
+            "For swag testing, the input file does not contain a label column. It can not be tested in current code"
+            "setting!"
+        )
+        return self._create_examples(self._read_csv(os.path.join(data_dir, "test.csv")), "test")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1", "2", "3"]
+
+    def _read_csv(self, input_file):
+        with open(input_file, "r", encoding="utf-8") as f:
+            return list(csv.reader(f))
+
+    def _create_examples(self, lines: List[List[str]], type: str):
+        """Creates examples for the training and dev sets."""
+        if type == "train" and lines[0][-1] != "label":
+            raise ValueError("For training, the input file must contain a label column.")
+
+        examples = [
+            InputExample(
+                example_id=line[2],
+                question=line[5],  # in the swag dataset, the
+                # common beginning of each
+                # choice is stored in "sent2".
+                contexts=[line[4], line[4], line[4], line[4]],
+                endings=[line[7], line[8], line[9], line[10]],
+                label=line[11],
+            )
+            for line in lines[1:]  # we skip the line with the column names
+        ]
+
+        return examples
+
+
+class ArcProcessor(DataProcessor):
+    """Processor for the ARC data set (request from allennlp)."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} train".format(data_dir))
+        return self._create_examples(self._read_json(os.path.join(data_dir, "train.jsonl")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {} dev".format(data_dir))
+        return self._create_examples(self._read_json(os.path.join(data_dir, "dev.jsonl")), "dev")
+
+    def get_test_examples(self, data_dir):
+        logger.info("LOOKING AT {} test".format(data_dir))
+        return self._create_examples(self._read_json(os.path.join(data_dir, "test.jsonl")), "test")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1", "2", "3"]
+
+    def _read_json(self, input_file):
+        with open(input_file, "r", encoding="utf-8") as fin:
+            lines = fin.readlines()
+            return lines
+
+    def _create_examples(self, lines, type):
+        """Creates examples for the training and dev sets."""
+
+        # There are two types of labels. They should be normalized
+        def normalize(truth):
+            if truth in "ABCD":
+                return ord(truth) - ord("A")
+            elif truth in "1234":
+                return int(truth) - 1
+            else:
+                logger.info("truth ERROR! %s", str(truth))
+                return None
+
+        examples = []
+        three_choice = 0
+        four_choice = 0
+        five_choice = 0
+        other_choices = 0
+        # we deleted example which has more than or less than four choices
+        for line in tqdm.tqdm(lines, desc="read arc data"):
+            data_raw = json.loads(line.strip("\n"))
+            if len(data_raw["question"]["choices"]) == 3:
+                three_choice += 1
+                continue
+            elif len(data_raw["question"]["choices"]) == 5:
+                five_choice += 1
+                continue
+            elif len(data_raw["question"]["choices"]) != 4:
+                other_choices += 1
+                continue
+            four_choice += 1
+            truth = str(normalize(data_raw["answerKey"]))
+            assert truth != "None"
+            question_choices = data_raw["question"]
+            question = question_choices["stem"]
+            id = data_raw["id"]
+            options = question_choices["choices"]
+            if len(options) == 4:
+                examples.append(
+                    InputExample(
+                        example_id=id,
+                        question=question,
+                        contexts=[
+                            options[0]["para"].replace("_", ""),
+                            options[1]["para"].replace("_", ""),
+                            options[2]["para"].replace("_", ""),
+                            options[3]["para"].replace("_", ""),
+                        ],
+                        endings=[options[0]["text"], options[1]["text"], options[2]["text"], options[3]["text"]],
+                        label=truth,
+                    )
+                )
+
+        if type == "train":
+            assert len(examples) > 1
+            assert examples[0].label is not None
+        logger.info("len examples: %s}", str(len(examples)))
+        logger.info("Three choices: %s", str(three_choice))
+        logger.info("Five choices: %s", str(five_choice))
+        logger.info("Other choices: %s", str(other_choices))
+        logger.info("four choices: %s", str(four_choice))
+
+        return examples
+
+
+def convert_examples_to_features(
+    examples: List[InputExample],
+    label_list: List[str],
+    max_length: int,
+    tokenizer: PreTrainedTokenizer,
+) -> List[InputFeatures]:
+    """
+    Loads a data file into a list of `InputFeatures`
+    """
+
+    label_map = {label: i for i, label in enumerate(label_list)}
+
+    features = []
+    for (ex_index, example) in tqdm.tqdm(enumerate(examples), desc="convert examples to features"):
+        if ex_index % 10000 == 0:
+            logger.info("Writing example %d of %d" % (ex_index, len(examples)))
+        choices_inputs = []
+        for ending_idx, (context, ending) in enumerate(zip(example.contexts, example.endings)):
+            text_a = context
+            if example.question.find("_") != -1:
+                # this is for cloze question
+                text_b = example.question.replace("_", ending)
+            else:
+                text_b = example.question + " " + ending
+
+            inputs = tokenizer(
+                text_a,
+                text_b,
+                add_special_tokens=True,
+                max_length=max_length,
+                padding="max_length",
+                truncation=True,
+                return_overflowing_tokens=True,
+            )
+            if "num_truncated_tokens" in inputs and inputs["num_truncated_tokens"] > 0:
+                logger.info(
+                    "Attention! you are cropping tokens (swag task is ok). "
+                    "If you are training ARC and RACE and you are poping question + options,"
+                    "you need to try to use a bigger max seq length!"
+                )
+
+            choices_inputs.append(inputs)
+
+        label = label_map[example.label]
+
+        input_ids = [x["input_ids"] for x in choices_inputs]
+        attention_mask = (
+            [x["attention_mask"] for x in choices_inputs] if "attention_mask" in choices_inputs[0] else None
+        )
+        token_type_ids = (
+            [x["token_type_ids"] for x in choices_inputs] if "token_type_ids" in choices_inputs[0] else None
+        )
+
+        features.append(
+            InputFeatures(
+                example_id=example.example_id,
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                label=label,
+            )
+        )
+
+    for f in features[:2]:
+        logger.info("*** Example ***")
+        logger.info("feature: %s" % f)
+
+    return features
+
+
+processors = {"race": RaceProcessor, "swag": SwagProcessor, "arc": ArcProcessor, "syn": SynonymProcessor}
+MULTIPLE_CHOICE_TASKS_NUM_LABELS = {"race", 4, "swag", 4, "arc", 4, "syn", 5}
@@ -19,3 +19,4 @@ pytest
 conllu
 sentencepiece != 0.1.92
 protobuf
+ray
@@ -21,7 +21,7 @@ mkdir -p $OUTPUT_DIR
 # Add parent directory to python path to access lightning_base.py
 export PYTHONPATH="../":"${PYTHONPATH}"

-python3 run_pl_glue.py --gpus 1 --data_dir $DATA_DIR \
+python3 run_glue.py --gpus 1 --data_dir $DATA_DIR \
 --task $TASK \
 --model_name_or_path $BERT_MODEL \
 --output_dir $OUTPUT_DIR \
@@ -31,7 +31,7 @@ mkdir -p $OUTPUT_DIR
 # Add parent directory to python path to access lightning_base.py
 export PYTHONPATH="../":"${PYTHONPATH}"

-python3 run_pl_ner.py --data_dir ./ \
+python3 run_ner.py --data_dir ./ \
 --labels ./labels.txt \
 --model_name_or_path $BERT_MODEL \
 --output_dir $OUTPUT_DIR \
@@ -26,7 +26,7 @@ export SEED=1
 # Add parent directory to python path to access lightning_base.py
 export PYTHONPATH="../":"${PYTHONPATH}"

-python3 run_pl_ner.py --data_dir ./ \
+python3 run_ner.py --data_dir ./ \
 --task_type POS \
 --model_name_or_path $BERT_MODEL \
 --output_dir $OUTPUT_DIR \
@@ -0,0 +1,229 @@
+## Token classification
+
+Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/contrib/legacy/token-classification/run_ner.py).
+
+The following examples are covered in this section:
+
+* NER on the GermEval 2014 (German NER) dataset
+* Emerging and Rare Entities task: WNUT’17 (English NER) dataset
+
+Details and results for the fine-tuning provided by @stefan-it.
+
+### GermEval 2014 (German NER) dataset
+
+#### Data (Download and pre-processing steps)
+
+Data can be obtained from the [GermEval 2014](https://sites.google.com/site/germeval2014ner/data) shared task page.
+
+Here are the commands for downloading and pre-processing train, dev and test datasets. The original data format has four (tab-separated) columns, in a pre-processing step only the two relevant columns (token and outer span NER annotation) are extracted:
+
+```bash
+curl -L 'https://drive.google.com/uc?export=download&id=1Jjhbal535VVz2ap4v4r_rN1UEHTdLK5P' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > train.txt.tmp
+curl -L 'https://drive.google.com/uc?export=download&id=1ZfRcQThdtAR5PPRjIDtrVP7BtXSCUBbm' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > dev.txt.tmp
+curl -L 'https://drive.google.com/uc?export=download&id=1u9mb7kNJHWQCWyweMDRMuTFoOHOfeBTH' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > test.txt.tmp
+```
+
+The GermEval 2014 dataset contains some strange "control character" tokens like `'\x96', '\u200e', '\x95', '\xad' or '\x80'`.
+One problem with these tokens is, that `BertTokenizer` returns an empty token for them, resulting in misaligned `InputExample`s.
+The `preprocess.py` script located in the `scripts` folder a) filters these tokens and b) splits longer sentences into smaller ones (once the max. subtoken length is reached).
+
+Let's define some variables that we need for further pre-processing steps and training the model:
+
+```bash
+export MAX_LENGTH=128
+export BERT_MODEL=bert-base-multilingual-cased
+```
+
+Run the pre-processing script on training, dev and test datasets:
+
+```bash
+python3 scripts/preprocess.py train.txt.tmp $BERT_MODEL $MAX_LENGTH > train.txt
+python3 scripts/preprocess.py dev.txt.tmp $BERT_MODEL $MAX_LENGTH > dev.txt
+python3 scripts/preprocess.py test.txt.tmp $BERT_MODEL $MAX_LENGTH > test.txt
+```
+
+The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so an own set of labels must be used:
+
+```bash
+cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
+```
+
+#### Prepare the run
+
+Additional environment variables must be set:
+
+```bash
+export OUTPUT_DIR=germeval-model
+export BATCH_SIZE=32
+export NUM_EPOCHS=3
+export SAVE_STEPS=750
+export SEED=1
+```
+
+#### Run the Pytorch version
+
+To start training, just run:
+
+```bash
+python3 run_ner.py --data_dir ./ \
+--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
+```
+
+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.
+
+#### JSON-based configuration file
+
+Instead of passing all parameters via commandline arguments, the `run_ner.py` script also supports reading parameters from a json-based configuration file:
+
+```json
+{
+    "data_dir": ".",
+    "labels": "./labels.txt",
+    "model_name_or_path": "bert-base-multilingual-cased",
+    "output_dir": "germeval-model",
+    "max_seq_length": 128,
+    "num_train_epochs": 3,
+    "per_device_train_batch_size": 32,
+    "save_steps": 750,
+    "seed": 1,
+    "do_train": true,
+    "do_eval": true,
+    "do_predict": true
+}
+```
+
+It must be saved with a `.json` extension and can be used by running `python3 run_ner.py config.json`.
+
+#### Evaluation
+
+Evaluation on development dataset outputs the following for our example:
+
+```bash
+10/04/2019 00:42:06 - INFO - __main__ -   ***** Eval results  *****
+10/04/2019 00:42:06 - INFO - __main__ -     f1 = 0.8623348017621146
+10/04/2019 00:42:06 - INFO - __main__ -     loss = 0.07183869666975543
+10/04/2019 00:42:06 - INFO - __main__ -     precision = 0.8467916366258111
+10/04/2019 00:42:06 - INFO - __main__ -     recall = 0.8784592370979806
+```
+
+On the test dataset the following results could be achieved:
+
+```bash
+10/04/2019 00:42:42 - INFO - __main__ -   ***** Eval results  *****
+10/04/2019 00:42:42 - INFO - __main__ -     f1 = 0.8614389652384803
+10/04/2019 00:42:42 - INFO - __main__ -     loss = 0.07064602487454782
+10/04/2019 00:42:42 - INFO - __main__ -     precision = 0.8604651162790697
+10/04/2019 00:42:42 - INFO - __main__ -     recall = 0.8624150210424085
+```
+
+### Emerging and Rare Entities task: WNUT’17 (English NER) dataset
+
+Description of the WNUT’17 task from the [shared task website](http://noisy-text.github.io/2017/index.html):
+
+> The WNUT’17 shared task focuses on identifying unusual, previously-unseen entities in the context of emerging discussions.
+> Named entities form the basis of many modern approaches to other tasks (like event clustering and summarization), but recall on
+> them is a real problem in noisy text - even among annotators. This drop tends to be due to novel entities and surface forms.
+
+Six labels are available in the dataset. An overview can be found on this [page](http://noisy-text.github.io/2017/files/).
+
+#### Data (Download and pre-processing steps)
+
+The dataset can be downloaded from the [official GitHub](https://github.com/leondz/emerging_entities_17) repository.
+
+The following commands show how to prepare the dataset for fine-tuning:
+
+```bash
+mkdir -p data_wnut_17
+
+curl -L 'https://github.com/leondz/emerging_entities_17/raw/master/wnut17train.conll'  | tr '\t' ' ' > data_wnut_17/train.txt.tmp
+curl -L 'https://github.com/leondz/emerging_entities_17/raw/master/emerging.dev.conll' | tr '\t' ' ' > data_wnut_17/dev.txt.tmp
+curl -L 'https://raw.githubusercontent.com/leondz/emerging_entities_17/master/emerging.test.annotated' | tr '\t' ' ' > data_wnut_17/test.txt.tmp
+```
+
+Let's define some variables that we need for further pre-processing steps:
+
+```bash
+export MAX_LENGTH=128
+export BERT_MODEL=bert-large-cased
+```
+
+Here we use the English BERT large model for fine-tuning.
+The `preprocess.py` scripts splits longer sentences into smaller ones (once the max. subtoken length is reached):
+
+```bash
+python3 scripts/preprocess.py data_wnut_17/train.txt.tmp $BERT_MODEL $MAX_LENGTH > data_wnut_17/train.txt
+python3 scripts/preprocess.py data_wnut_17/dev.txt.tmp $BERT_MODEL $MAX_LENGTH > data_wnut_17/dev.txt
+python3 scripts/preprocess.py data_wnut_17/test.txt.tmp $BERT_MODEL $MAX_LENGTH > data_wnut_17/test.txt
+```
+
+In the last pre-processing step, the `labels.txt` file needs to be generated. This file contains all available labels:
+
+```bash
+cat data_wnut_17/train.txt data_wnut_17/dev.txt data_wnut_17/test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > data_wnut_17/labels.txt
+```
+
+#### Run the Pytorch version
+
+Fine-tuning with the PyTorch version can be started using the `run_ner.py` script. In this example we use a JSON-based configuration file.
+
+This configuration file looks like:
+
+```json
+{
+    "data_dir": "./data_wnut_17",
+    "labels": "./data_wnut_17/labels.txt",
+    "model_name_or_path": "bert-large-cased",
+    "output_dir": "wnut-17-model-1",
+    "max_seq_length": 128,
+    "num_train_epochs": 3,
+    "per_device_train_batch_size": 32,
+    "save_steps": 425,
+    "seed": 1,
+    "do_train": true,
+    "do_eval": true,
+    "do_predict": true,
+    "fp16": false
+}
+```
+
+If your GPU supports half-precision training, please set `fp16` to `true`.
+
+Save this JSON-based configuration under `wnut_17.json`. The fine-tuning can be started with `python3 run_ner_old.py wnut_17.json`.
+
+#### Evaluation
+
+Evaluation on development dataset outputs the following:
+
+```bash
+05/29/2020 23:33:44 - INFO - __main__ -   ***** Eval results *****
+05/29/2020 23:33:44 - INFO - __main__ -     eval_loss = 0.26505235286212275
+05/29/2020 23:33:44 - INFO - __main__ -     eval_precision = 0.7008264462809918
+05/29/2020 23:33:44 - INFO - __main__ -     eval_recall = 0.507177033492823
+05/29/2020 23:33:44 - INFO - __main__ -     eval_f1 = 0.5884802220680084
+05/29/2020 23:33:44 - INFO - __main__ -     epoch = 3.0
+```
+
+On the test dataset the following results could be achieved:
+
+```bash
+05/29/2020 23:33:44 - INFO - transformers.trainer -   ***** Running Prediction *****
+05/29/2020 23:34:02 - INFO - __main__ -     eval_loss = 0.30948806500973547
+05/29/2020 23:34:02 - INFO - __main__ -     eval_precision = 0.5840108401084011
+05/29/2020 23:34:02 - INFO - __main__ -     eval_recall = 0.3994439295644115
+05/29/2020 23:34:02 - INFO - __main__ -     eval_f1 = 0.47440836543753434
+```
+
+WNUT’17 is a very difficult task. Current state-of-the-art results on this dataset can be found [here](http://nlpprogress.com/english/named_entity_recognition.html).
@@ -20,7 +20,7 @@ export NUM_EPOCHS=3
 export SAVE_STEPS=750
 export SEED=1

-python3 run_ner_old.py \
+python3 run_ner.py \
 --task_type NER \
 --data_dir . \
 --labels ./labels.txt \
@@ -21,7 +21,7 @@ export NUM_EPOCHS=3
 export SAVE_STEPS=750
 export SEED=1

-python3 run_ner_old.py \
+python3 run_ner.py \
 --task_type Chunk \
 --data_dir . \
 --model_name_or_path $BERT_MODEL \
@@ -21,7 +21,7 @@ export NUM_EPOCHS=3
 export SAVE_STEPS=750
 export SEED=1

-python3 run_ner_old.py \
+python3 run_ner.py \
 --task_type POS \
 --data_dir . \
 --model_name_or_path $BERT_MODEL \
@@ -1,26 +1,35 @@
+<!---
+Copyright 2020 The HuggingFace Team. 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.
+-->
+
 ## Multiple Choice

-Based on the script [`run_multiple_choice.py`]().
+Based on the script [`run_swag.py`]().

 #### Fine-tuning on SWAG
-Download [swag](https://github.com/rowanz/swagaf/tree/master/data) data

 ```bash
-#training on 4 tesla V100(16GB) GPUS
-export SWAG_DIR=/path/to/swag_data_dir
-python ./examples/multiple-choice/run_multiple_choice.py \
--task_name swag \
+python examples/multiple-choice/run_swag.py \
 --model_name_or_path roberta-base \
 --do_train \
 --do_eval \
--data_dir $SWAG_DIR \
 --learning_rate 5e-5 \
 --num_train_epochs 3 \
--max_seq_length 80 \
--output_dir models_bert/swag_base \
+--output_dir /tmp/swag_base \
 --per_gpu_eval_batch_size=16 \
 --per_device_train_batch_size=16 \
--gradient_accumulation_steps 2 \
 --overwrite_output
 ```
 Training with the defined hyper-parameters yields the following results:
@@ -0,0 +1,2 @@
+sentencepiece != 0.1.92
+protobuf
@@ -0,0 +1,360 @@
+# coding=utf-8
+# Copyright The HuggingFace Team and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the library models for multiple choice.
+"""
+# You can also adapt this script on your own multiple choice task. Pointers for this are left as comments.
+
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Optional, Union
+
+import numpy as np
+import torch
+from datasets import load_dataset
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoModelForMultipleChoice,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    default_data_collator,
+    set_seed,
+)
+from transformers.tokenization_utils_base import PaddingStrategy, PreTrainedTokenizerBase
+from transformers.trainer_utils import is_main_process
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_seq_length: int = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. If passed, sequences longer "
+            "than this will be truncated, sequences shorter will be padded."
+        },
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to the maximum sentence length. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch. More "
+            "efficient on GPU but very bad for TPU."
+        },
+    )
+
+    def __post_init__(self):
+        if self.train_file is not None:
+            extension = self.train_file.split(".")[-1]
+            assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
+        if self.validation_file is not None:
+            extension = self.validation_file.split(".")[-1]
+            assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."
+
+
+@dataclass
+class DataCollatorForMultipleChoice:
+    """
+    Data collator that will dynamically pad the inputs for multiple choice received.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
+            among:
+
+            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+              sequence if provided).
+            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
+              maximum acceptable input length for the model if that argument is not provided.
+            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
+              different lengths).
+        max_length (:obj:`int`, `optional`):
+            Maximum length of the returned list and optionally padding length (see above).
+        pad_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the sequence to a multiple of the provided value.
+
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    padding: Union[bool, str, PaddingStrategy] = True
+    max_length: Optional[int] = None
+    pad_to_multiple_of: Optional[int] = None
+
+    def __call__(self, features):
+        label_name = "label" if "label" in features[0].keys() else "labels"
+        labels = [feature.pop(label_name) for feature in features]
+        batch_size = len(features)
+        num_choices = len(features[0]["input_ids"])
+        flattened_features = [
+            [{k: v[i] for k, v in feature.items()} for i in range(num_choices)] for feature in features
+        ]
+        flattened_features = sum(flattened_features, [])
+
+        batch = self.tokenizer.pad(
+            flattened_features,
+            padding=self.padding,
+            max_length=self.max_length,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors="pt",
+        )
+
+        # Un-flatten
+        batch = {k: v.view(batch_size, num_choices, -1) for k, v in batch.items()}
+        # Add back labels
+        batch["labels"] = torch.tensor(labels, dtype=torch.int64)
+        return batch
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
+    )
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+
+    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
+    # download the dataset.
+    if data_args.train_file is not None or data_args.validation_file is not None:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        datasets = load_dataset(extension, data_files=data_files)
+    else:
+        # Downloading and loading the swag dataset from the hub.
+        datasets = load_dataset("swag", "regular")
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    config = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_fast=model_args.use_fast_tokenizer,
+    )
+    model = AutoModelForMultipleChoice.from_pretrained(
+        model_args.model_name_or_path,
+        from_tf=bool(".ckpt" in model_args.model_name_or_path),
+        config=config,
+        cache_dir=model_args.cache_dir,
+    )
+
+    # When using your own dataset or a different dataset from swag, you will probably need to change this.
+    ending_names = [f"ending{i}" for i in range(4)]
+    context_name = "sent1"
+    question_header_name = "sent2"
+
+    # Preprocessing the datasets.
+    def preprocess_function(examples):
+        first_sentences = [[context] * 4 for context in examples[context_name]]
+        question_headers = examples[question_header_name]
+        second_sentences = [
+            [f"{header} {examples[end][i]}" for end in ending_names] for i, header in enumerate(question_headers)
+        ]
+
+        # Flatten out
+        first_sentences = sum(first_sentences, [])
+        second_sentences = sum(second_sentences, [])
+
+        # Tokenize
+        tokenized_examples = tokenizer(
+            first_sentences,
+            second_sentences,
+            truncation=True,
+            max_length=data_args.max_seq_length,
+            padding="max_length" if data_args.pad_to_max_length else False,
+        )
+        # Un-flatten
+        return {k: [v[i : i + 4] for i in range(0, len(v), 4)] for k, v in tokenized_examples.items()}
+
+    tokenized_datasets = datasets.map(
+        preprocess_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    # Data collator
+    data_collator = (
+        default_data_collator if data_args.pad_to_max_length else DataCollatorForMultipleChoice(tokenizer=tokenizer)
+    )
+
+    # Metric
+    def compute_metrics(eval_predictions):
+        predictions, label_ids = eval_predictions
+        preds = np.argmax(predictions, axis=1)
+        return {"accuracy": (preds == label_ids).astype(np.float32).mean().item()}
+
+    # Initialize our Trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+        compute_metrics=compute_metrics,
+    )
+
+    # Training
+    if training_args.do_train:
+        train_result = trainer.train(
+            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
+        )
+        trainer.save_model()  # Saves the tokenizer too for easy upload
+
+        output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
+        if trainer.is_world_process_zero():
+            with open(output_train_file, "w") as writer:
+                logger.info("***** Train results *****")
+                for key, value in sorted(train_result.metrics.items()):
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+            # Need to save the state, since Trainer.save_model saves only the tokenizer with the model
+            trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+
+        results = trainer.evaluate()
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results_swag.txt")
+        if trainer.is_world_process_zero():
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Eval results *****")
+                for key, value in sorted(results.items()):
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()
@@ -1,9 +1,29 @@
+<!---
+Copyright 2020 The HuggingFace Team. 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.
+-->

 ## SQuAD

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

+**Note:** This script only works with models that have a fast tokenizer (backed by the 🤗 Tokenizers library) as it
+uses special features of those tokenizers. You can check if your favorite model has a fast tokenizer in
+[this table](https://huggingface.co/transformers/index.html#bigtable), if it doesn't you can still use the old version
+of the script.
+
+The old version of this script can be found [here](https://github.com/huggingface/transformers/tree/master/examples/legacy/question-answering).
 #### Fine-tuning BERT on SQuAD1.0

 This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large)
@@ -0,0 +1 @@
+datasets >= 1.1.3
@@ -438,11 +438,22 @@ def main():

    # Training
    if training_args.do_train:
-        trainer.train(
+        train_result = trainer.train(
            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
        )
        trainer.save_model()  # Saves the tokenizer too for easy upload

+        output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
+        if trainer.is_world_process_zero():
+            with open(output_train_file, "w") as writer:
+                logger.info("***** Train results *****")
+                for key, value in sorted(train_result.metrics.items()):
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+            # Need to save the state, since Trainer.save_model saves only the tokenizer with the model
+            trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
+
    # Evaluation
    results = {}
    if training_args.do_eval:
@@ -453,7 +464,7 @@ def main():
        if trainer.is_world_process_zero():
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
-                for key, value in results.items():
+                for key, value in sorted(results.items()):
                    logger.info(f"  {key} = {value}")
                    writer.write(f"{key} = {value}\n")

@@ -481,11 +481,22 @@ def main():

    # Training
    if training_args.do_train:
-        trainer.train(
+        train_result = trainer.train(
            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
        )
        trainer.save_model()  # Saves the tokenizer too for easy upload

+        output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
+        if trainer.is_world_process_zero():
+            with open(output_train_file, "w") as writer:
+                logger.info("***** Train results *****")
+                for key, value in sorted(train_result.metrics.items()):
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+            # Need to save the state, since Trainer.save_model saves only the tokenizer with the model
+            trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
+
    # Evaluation
    results = {}
    if training_args.do_eval:
@@ -496,7 +507,7 @@ def main():
        if trainer.is_world_process_zero():
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
-                for key, value in results.items():
+                for key, value in sorted(results.items()):
                    logger.info(f"  {key} = {value}")
                    writer.write(f"{key} = {value}\n")

@@ -76,9 +76,7 @@ def postprocess_qa_predictions(
    assert len(predictions) == 2, "`predictions` should be a tuple with two elements (start_logits, end_logits)."
    all_start_logits, all_end_logits = predictions

-    assert len(predictions[0]) == len(
-        features
-    ), f"Got {len(predictions[0])} predicitions and {len(features)} features."
+    assert len(predictions[0]) == len(features), f"Got {len(predictions[0])} predictions and {len(features)} features."

    # Build a map example to its corresponding features.
    example_id_to_index = {k: i for i, k in enumerate(examples["id"])}
@@ -118,7 +116,7 @@ def postprocess_qa_predictions(

            # Update minimum null prediction.
            feature_null_score = start_logits[0] + end_logits[0]
-            if min_null_prediction is None or min_null_prediction["score"] < feature_null_score:
+            if min_null_prediction is None or min_null_prediction["score"] > feature_null_score:
                min_null_prediction = {
                    "offsets": (0, 0),
                    "score": feature_null_score,
@@ -208,7 +206,7 @@ def postprocess_qa_predictions(

        # Make `predictions` JSON-serializable by casting np.float back to float.
        all_nbest_json[example["id"]] = [
-            {k: (float(v) if isinstance(v, (np.float32, np.float64)) else v) for k, v in pred.items()}
+            {k: (float(v) if isinstance(v, (np.float16, np.float32, np.float64)) else v) for k, v in pred.items()}
            for pred in predictions
        ]

@@ -396,7 +394,7 @@ def postprocess_qa_predictions_with_beam_search(

        # Make `predictions` JSON-serializable by casting np.float back to float.
        all_nbest_json[example["id"]] = [
-            {k: (float(v) if isinstance(v, (np.float32, np.float64)) else v) for k, v in pred.items()}
+            {k: (float(v) if isinstance(v, (np.float16, np.float32, np.float64)) else v) for k, v in pred.items()}
            for pred in predictions
        ]

@@ -0,0 +1,28 @@
+<!---
+Copyright 2020 The HuggingFace Team. 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.
+-->
+
+# Research projects
+
+This folder contains various research projects using 🤗 Transformers. They are not maintained and require a specific
+version of 🤗 Transformers that is indicated in the requirements file of each folder. Updating them to the most recent version of the library will require some work.
+
+To use any of them, just run the command
+```
+pip install -r requirements.txt
+```
+inside the folder of your choice.
+
+If you need help with any of those, contact the author(s), indicated at the top of the `README` of each folder.
@@ -0,0 +1 @@
+transformers == 3.5.1
--- a/Show More
+++ b/Show More