tagging

.circleci/config.yml

@@ -281,7 +281,7 @@ jobs:
                       - v0.4-build_doc-{{ checksum "setup.py" }}
                       - v0.4-{{ checksum "setup.py" }}
             - run: pip install --upgrade pip
-            - run: pip install ."[all, docs]"
+            - run: pip install .[tf,torch,sentencepiece,docs]
             - save_cache:
                   key: v0.4-build_doc-{{ checksum "setup.py" }}
                   paths:
@@ -303,7 +303,7 @@ jobs:
                   keys:
                       - v0.4-deploy_doc-{{ checksum "setup.py" }}
                       - v0.4-{{ checksum "setup.py" }}
-            - run: pip install ."[all,docs]"
+            - run: pip install .[tf,torch,sentencepiece,docs]
             - save_cache:
                   key: v0.4-deploy_doc-{{ checksum "setup.py" }}
                   paths:

.circleci/deploy.sh

@@ -52,4 +52,4 @@ deploy_doc "4b3ee9c" v3.1.0
 deploy_doc "3ebb1b3" v3.2.0
 deploy_doc "0613f05" v3.3.1
 deploy_doc "eb0e0ce" v3.4.0
-deploy_doc "818878d" # v3.5.1 Latest stable release
+deploy_doc "818878d" # v3.5.0 Latest stable release

.coveragerc

@@ -1,12 +0,0 @@
-[run]
-source=transformers
-omit =
-    # skip convertion scripts from testing for now
-    */convert_*
-    */__main__.py
-[report]
-exclude_lines =
-    pragma: no cover
-    raise
-    except
-    register_parameter

.github/workflows/self-push.yml

@@ -4,12 +4,10 @@ on:
   push:
     branches:
       - master
-      - model-templates
     paths:
       - "src/**"
       - "tests/**"
       - ".github/**"
-      - "templates/**"
   # pull_request:
   repository_dispatch:
 
@@ -57,14 +55,6 @@ jobs:
           python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
           python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
 
-      - name: Create model files
-        run: |
-          source .env/bin/activate
-          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/encoder-bert-tokenizer.json --path=templates/adding_a_new_model
-          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
-
       - name: Run all non-slow tests on GPU
         env:
           OMP_NUM_THREADS: 1
@@ -126,14 +116,6 @@ jobs:
           TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('TF GPUs available:', bool(tf.config.list_physical_devices('GPU')))"
           TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('Number of TF GPUs available:', len(tf.config.list_physical_devices('GPU')))"
 
-      - name: Create model files
-        run: |
-          source .env/bin/activate
-          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/encoder-bert-tokenizer.json --path=templates/adding_a_new_model
-          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
-
       - name: Run all non-slow tests on GPU
         env:
           OMP_NUM_THREADS: 1

docs/source/_static/js/custom.js

@@ -4,7 +4,7 @@ const stableVersion = "v3.5.0"
 // Dictionary doc folder to label
 const versionMapping = {
     "master": "master",
-    "": "v3.5.0/v3.5.1",
+    "": "v3.5.0",
     "v3.4.0": "v3.4.0",
     "v3.3.1": "v3.3.0/v3.3.1",
     "v3.2.0": "v3.2.0",

docs/source/main_classes/trainer.rst

@@ -35,7 +35,7 @@ Here is an example of how to customize :class:`~transformers.Trainer` using a cu
     class MyTrainer(Trainer):
         def compute_loss(self, model, inputs):
             labels = inputs.pop("labels")
-            outputs = model(**inputs)
+            outputs = models(**inputs)
             logits = outputs[0]
             return my_custom_loss(logits, labels)
 

docs/source/model_doc/bert.rst

@@ -188,10 +188,3 @@ TFBertForQuestionAnswering
 
 .. autoclass:: transformers.TFBertForQuestionAnswering
     :members: call
-
-
-FlaxBertModel
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.FlaxBertModel
-    :members: __call__

docs/source/model_doc/bertgeneration.rst

@@ -40,7 +40,7 @@ Usage:
   labels = tokenizer('This is a short summary', return_tensors="pt").input_ids
 
   # train...
-  loss = bert2bert(input_ids=input_ids, decoder_input_ids=labels, labels=labels).loss
+  loss = bert2bert(input_ids=input_ids, decoder_input_ids=labels, labels=labels, return_dict=True).loss
   loss.backward()
 
 

docs/source/model_doc/dpr.rst

@@ -99,22 +99,3 @@ DPRReader
 
 .. autoclass:: transformers.DPRReader
     :members: forward
-
-TFDPRContextEncoder
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFDPRContextEncoder
-    :members: call
-
-TFDPRQuestionEncoder
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFDPRQuestionEncoder
-    :members: call
-
-
-TFDPRReader
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFDPRReader
-    :members: call

docs/source/model_doc/marian.rst

@@ -139,19 +139,11 @@ Example of translating english to many romance languages, using old-style 2 char
 .. code-block::python
 
     from transformers import MarianMTModel, MarianTokenizer
-    src_text = [
-        '>>fr<< this is a sentence in english that we want to translate to french',
-        '>>pt<< This should go to portuguese',
-        '>>es<< And this to Spanish'
-    ]
+     src_text = [ '>>fr<< this is a sentence in english that we want to translate to french', '>>pt<< This should go to portuguese', '>>es<< And this to Spanish']
 
-    model_name = 'Helsinki-NLP/opus-mt-en-ROMANCE'
-    tokenizer = MarianTokenizer.from_pretrained(model_name)
-    print(tokenizer.supported_language_codes)
-
-    model = MarianMTModel.from_pretrained(model_name)
-    translated = model.generate(**tokenizer.prepare_seq2seq_batch(src_text))
-    tgt_text = [tokenizer.decode(t, skip_special_tokens=True) for t in translated]
+    model_name = 'Helsinki-NLP/opus-mt-en-ROMANCE' tokenizer = MarianTokenizer.from_pretrained(model_name)
+    print(tokenizer.supported_language_codes) model = MarianMTModel.from_pretrained(model_name) translated =
+    model.generate(**tokenizer.prepare_seq2seq_batch(src_text)) tgt_text = [tokenizer.decode(t, skip_special_tokens=True) for t in translated]
     # ["c'est une phrase en anglais que nous voulons traduire en français", 'Isto deve ir para o português.',  'Y esto al español']
 
 

docs/source/model_doc/roberta.rst

@@ -146,10 +146,3 @@ TFRobertaForQuestionAnswering
 
 .. autoclass:: transformers.TFRobertaForQuestionAnswering
     :members: call
-
-
-FlaxRobertaModel
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.FlaxRobertaModel
-    :members: __call__

docs/source/model_doc/t5.rst

@@ -64,7 +64,7 @@ token. T5 can be trained / fine-tuned both in a supervised and unsupervised fash
   input_ids = tokenizer('The <extra_id_0> walks in <extra_id_1> park', return_tensors='pt').input_ids
   labels = tokenizer('<extra_id_0> cute dog <extra_id_1> the <extra_id_2>', return_tensors='pt').input_ids
   # the forward function automatically creates the correct decoder_input_ids
-  loss = model(input_ids=input_ids, labels=labels).loss
+  loss = model(input_ids=input_ids, labels=labels, return_dict=True).loss
 
 - Supervised training
 
@@ -77,7 +77,7 @@ token. T5 can be trained / fine-tuned both in a supervised and unsupervised fash
   input_ids = tokenizer('translate English to German: The house is wonderful.', return_tensors='pt').input_ids
   labels = tokenizer('Das Haus ist wunderbar.', return_tensors='pt').input_ids
   # the forward function automatically creates the correct decoder_input_ids
-  loss = model(input_ids=input_ids, labels=labels).loss
+  loss = model(input_ids=input_ids, labels=labels, return_dict=True).loss
 
 
 T5Config

docs/source/model_sharing.rst

@@ -37,7 +37,7 @@ For instance:
 
 .. code-block::
 
-    >>> model = AutoModel.from_pretrained(
+    >>> tokenizer = AutoTokenizer.from_pretrained(
     >>>   "julien-c/EsperBERTo-small",
     >>>   revision="v2.0.1" # tag name, or branch name, or commit hash
     >>> )
@@ -46,42 +46,31 @@ 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.
+users to clone it and you (and your organization members) to push to it. First, you should ensure you are logged in the
+``transformers-cli``:
 
-You can create a model repo directly from the website, `here <https://huggingface.co/new>`.
-
-Alternatively, you can use the ``transformers-cli``. The next steps describe that process:
-
-Go to a terminal and run the following command. It should be in the virtual environment where you installed 🤗
+Go in a terminal and run the following command. It should be in the virtual environment where you installed 🤗
 Transformers, since that command :obj:`transformers-cli` comes from the library.
 
-.. code-block:: bash
+.. code-block::
 
     transformers-cli login
 
 
 Once you are logged in with your model hub credentials, you can start building your repositories. To create a repo:
 
-.. code-block:: bash
+.. code-block::
 
     transformers-cli repo create your-model-name
 
-This creates a repo on the model hub, which can be cloned.
+This creates a repo on the model hub, which can be cloned. You can then add/remove from that repo as you would with any
+other git repo.
 
-.. code-block:: bash
+.. code-block::
 
     git clone https://huggingface.co/username/your-model-name
 
-    # Make sure you have git-lfs installed
-    # (https://git-lfs.github.com/)
-    git lfs install
-
-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.
-
-.. code-block:: bash
-
-    # Commit as usual
+    # Then commit as usual
     cd your-model-name
     echo "hello" >> README.md
     git add . && git commit -m "Update from $USER"
@@ -170,25 +159,24 @@ Or, if you're using the Trainer API
 .. code-block::
 
     >>> trainer.save_model("path/to/awesome-name-you-picked")
-    >>> tokenizer.save_pretrained("path/to/repo/clone/your-model-name")
 
 You can then add these files to the staging environment and verify that they have been correctly staged with the ``git
 status`` command:
 
-.. code-block:: bash
+.. code-block::
 
     git add --all
     git status
 
 Finally, the files should be comitted:
 
-.. code-block:: bash
+.. code-block::
 
     git commit -m "First version of the your-model-name model and tokenizer."
 
 And pushed to the remote:
 
-.. code-block:: bash
+.. code-block::
 
     git push
 
@@ -211,7 +199,7 @@ don't forget to link to its model card so that people can fully trace how your m
 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::
+.. 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/`.
@@ -237,47 +225,3 @@ You may specify a revision by using the ``revision`` flag in the ``from_pretrain
     >>>   "julien-c/EsperBERTo-small",
     >>>   revision="v2.0.1" # tag name, or branch name, or commit hash
     >>> )
-
-Workflow in a Colab notebook
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-If you're in a Colab notebook (or similar) with no direct access to a terminal, here is the workflow you can use to
-upload your model. You can execute each one of them in a cell by adding a ! at the beginning.
-
-First you need to install `git-lfs` in the environment used by the notebook:
-
-.. code-block:: bash
-
-    sudo apt-get install git-lfs
-
-Then you can use the :obj:`transformers-cli` to create your new repo:
-
-
-.. code-block:: bash
-
-    transformers-cli login
-    transformers-cli repo create your-model-name
-
-Once it's created, you can clone it and configure it (replace username by your username on huggingface.co):
-
-.. code-block:: bash
-
-    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"
-
-Once you've saved your model inside, and your clone is setup with the right remote URL, you can add it and push it with
-usual git commands.
-
-.. code-block:: bash
-
-    git add .
-    git commit -m "Initial commit"
-    git push

docs/source/task_summary.rst

@@ -89,7 +89,7 @@ each other. The process is the following:
     >>> import torch
 
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased-finetuned-mrpc")
-    >>> model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased-finetuned-mrpc")
+    >>> model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased-finetuned-mrpc", return_dict=True)
 
     >>> classes = ["not paraphrase", "is paraphrase"]
 
@@ -122,7 +122,7 @@ each other. The process is the following:
     >>> import tensorflow as tf
 
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased-finetuned-mrpc")
-    >>> model = TFAutoModelForSequenceClassification.from_pretrained("bert-base-cased-finetuned-mrpc")
+    >>> model = TFAutoModelForSequenceClassification.from_pretrained("bert-base-cased-finetuned-mrpc", return_dict=True)
 
     >>> classes = ["not paraphrase", "is paraphrase"]
 
@@ -211,7 +211,7 @@ Here is an example of question answering using a model and a tokenizer. The proc
     >>> import torch
 
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
-    >>> model = AutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
+    >>> model = AutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad", return_dict=True)
 
     >>> text = r"""
     ... 🤗 Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides general-purpose
@@ -253,7 +253,7 @@ Here is an example of question answering using a model and a tokenizer. The proc
     >>> import tensorflow as tf
 
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
-    >>> model = TFAutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
+    >>> model = TFAutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad", return_dict=True)
 
     >>> text = r"""
     ... 🤗 Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides general-purpose
@@ -373,7 +373,7 @@ Here is an example of doing masked language modeling using a model and a tokeniz
     >>> import torch
 
     >>> tokenizer = AutoTokenizer.from_pretrained("distilbert-base-cased")
-    >>> model = AutoModelWithLMHead.from_pretrained("distilbert-base-cased")
+    >>> model = AutoModelWithLMHead.from_pretrained("distilbert-base-cased", return_dict=True)
 
     >>> sequence = f"Distilled models are smaller than the models they mimic. Using them instead of the large versions would help {tokenizer.mask_token} our carbon footprint."
 
@@ -389,7 +389,7 @@ Here is an example of doing masked language modeling using a model and a tokeniz
     >>> import tensorflow as tf
 
     >>> tokenizer = AutoTokenizer.from_pretrained("distilbert-base-cased")
-    >>> model = TFAutoModelWithLMHead.from_pretrained("distilbert-base-cased")
+    >>> model = TFAutoModelWithLMHead.from_pretrained("distilbert-base-cased", return_dict=True)
 
     >>> sequence = f"Distilled models are smaller than the models they mimic. Using them instead of the large versions would help {tokenizer.mask_token} our carbon footprint."
 
@@ -437,7 +437,7 @@ of tokens.
     >>> from torch.nn import functional as F
 
     >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
-    >>> model = AutoModelWithLMHead.from_pretrained("gpt2")
+    >>> model = AutoModelWithLMHead.from_pretrained("gpt2", return_dict=True)
 
     >>> sequence = f"Hugging Face is based in DUMBO, New York City, and "
 
@@ -461,7 +461,7 @@ of tokens.
     >>> import tensorflow as tf
 
     >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
-    >>> model = TFAutoModelWithLMHead.from_pretrained("gpt2")
+    >>> model = TFAutoModelWithLMHead.from_pretrained("gpt2", return_dict=True)
 
     >>> sequence = f"Hugging Face is based in DUMBO, New York City, and "
 
@@ -520,7 +520,7 @@ Here is an example of text generation using ``XLNet`` and its tokenizer.
     >>> ## PYTORCH CODE
     >>> from transformers import AutoModelWithLMHead, AutoTokenizer
 
-    >>> model = AutoModelWithLMHead.from_pretrained("xlnet-base-cased")
+    >>> model = AutoModelWithLMHead.from_pretrained("xlnet-base-cased", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("xlnet-base-cased")
 
     >>> # Padding text helps XLNet with short prompts - proposed by Aman Rusia in https://github.com/rusiaaman/XLNet-gen#methodology
@@ -545,7 +545,7 @@ Here is an example of text generation using ``XLNet`` and its tokenizer.
     >>> ## TENSORFLOW CODE
     >>> from transformers import TFAutoModelWithLMHead, AutoTokenizer
 
-    >>> model = TFAutoModelWithLMHead.from_pretrained("xlnet-base-cased")
+    >>> model = TFAutoModelWithLMHead.from_pretrained("xlnet-base-cased", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("xlnet-base-cased")
 
     >>> # Padding text helps XLNet with short prompts - proposed by Aman Rusia in https://github.com/rusiaaman/XLNet-gen#methodology
@@ -664,7 +664,7 @@ Here is an example of doing named entity recognition, using a model and a tokeni
     >>> from transformers import AutoModelForTokenClassification, AutoTokenizer
     >>> import torch
 
-    >>> model = AutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english")
+    >>> model = AutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
 
     >>> label_list = [
@@ -692,7 +692,7 @@ Here is an example of doing named entity recognition, using a model and a tokeni
     >>> from transformers import TFAutoModelForTokenClassification, AutoTokenizer
     >>> import tensorflow as tf
 
-    >>> model = TFAutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english")
+    >>> model = TFAutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
 
     >>> label_list = [
@@ -790,7 +790,7 @@ CNN / Daily Mail), it yields very good results.
     >>> ## PYTORCH CODE
     >>> from transformers import AutoModelWithLMHead, AutoTokenizer
 
-    >>> model = AutoModelWithLMHead.from_pretrained("t5-base")
+    >>> model = AutoModelWithLMHead.from_pretrained("t5-base", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
 
     >>> # T5 uses a max_length of 512 so we cut the article to 512 tokens.
@@ -799,7 +799,7 @@ CNN / Daily Mail), it yields very good results.
     >>> ## TENSORFLOW CODE
     >>> from transformers import TFAutoModelWithLMHead, AutoTokenizer
 
-    >>> model = TFAutoModelWithLMHead.from_pretrained("t5-base")
+    >>> model = TFAutoModelWithLMHead.from_pretrained("t5-base", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
 
     >>> # T5 uses a max_length of 512 so we cut the article to 512 tokens.
@@ -843,7 +843,7 @@ Here is an example of doing translation using a model and a tokenizer. The proce
     >>> ## PYTORCH CODE
     >>> from transformers import AutoModelWithLMHead, AutoTokenizer
 
-    >>> model = AutoModelWithLMHead.from_pretrained("t5-base")
+    >>> model = AutoModelWithLMHead.from_pretrained("t5-base", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
 
     >>> inputs = tokenizer.encode("translate English to German: Hugging Face is a technology company based in New York and Paris", return_tensors="pt")
@@ -851,7 +851,7 @@ Here is an example of doing translation using a model and a tokenizer. The proce
     >>> ## TENSORFLOW CODE
     >>> from transformers import TFAutoModelWithLMHead, AutoTokenizer
 
-    >>> model = TFAutoModelWithLMHead.from_pretrained("t5-base")
+    >>> model = TFAutoModelWithLMHead.from_pretrained("t5-base", return_dict=True)
     >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
 
     >>> inputs = tokenizer.encode("translate English to German: Hugging Face is a technology company based in New York and Paris", return_tensors="tf")

docs/source/training.rst

@@ -39,7 +39,7 @@ head on top of the encoder with an output size of 2. Models are initialized in `
 .. code-block:: python
 
     from transformers import BertForSequenceClassification
-    model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+    model = BertForSequenceClassification.from_pretrained('bert-base-uncased', return_dict=True)
     model.train()
 
 This is useful because it allows us to make use of the pre-trained BERT encoder and easily train it on whatever

examples/adversarial/run_hans.py

@@ -23,7 +23,6 @@ from typing import Dict, List, Optional
 import numpy as np
 import torch
 
-import transformers
 from transformers import (
     AutoConfig,
     AutoModelForSequenceClassification,
@@ -34,7 +33,6 @@ from transformers import (
     default_data_collator,
     set_seed,
 )
-from transformers.trainer_utils import is_main_process
 from utils_hans import HansDataset, InputFeatures, hans_processors, hans_tasks_num_labels
 
 
@@ -126,11 +124,6 @@ def main():
         bool(training_args.local_rank != -1),
         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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Set seed

examples/bert-loses-patience/run_glue_with_pabee.py

@@ -29,7 +29,6 @@ from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, Tenso
 from torch.utils.data.distributed import DistributedSampler
 from tqdm import tqdm, trange
 
-import transformers
 from pabee.modeling_pabee_albert import AlbertForSequenceClassificationWithPabee
 from pabee.modeling_pabee_bert import BertForSequenceClassificationWithPabee
 from transformers import (
@@ -45,7 +44,6 @@ from transformers import glue_compute_metrics as compute_metrics
 from transformers import glue_convert_examples_to_features as convert_examples_to_features
 from transformers import glue_output_modes as output_modes
 from transformers import glue_processors as processors
-from transformers.trainer_utils import is_main_process
 
 
 try:
@@ -632,11 +630,7 @@ def main():
         bool(args.local_rank != -1),
         args.fp16,
     )
-    # Set the verbosity to info of the Transformers logger (on main process only):
-    if is_main_process(args.local_rank):
-        transformers.utils.logging.set_verbosity_info()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
+
     # Set seed
     set_seed(args)
 

examples/bertology/run_bertology.py

@@ -30,7 +30,6 @@ from torch.utils.data import DataLoader, SequentialSampler, Subset
 from torch.utils.data.distributed import DistributedSampler
 from tqdm import tqdm
 
-import transformers
 from transformers import (
     AutoConfig,
     AutoModelForSequenceClassification,
@@ -42,7 +41,6 @@ from transformers import (
     glue_processors,
     set_seed,
 )
-from transformers.trainer_utils import is_main_process
 
 
 logger = logging.getLogger(__name__)
@@ -370,11 +368,6 @@ def main():
     # Setup logging
     logging.basicConfig(level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
     logger.info("device: {} n_gpu: {}, distributed: {}".format(args.device, args.n_gpu, bool(args.local_rank != -1)))
-    # Set the verbosity to info of the Transformers logger (on main process only):
-    if is_main_process(args.local_rank):
-        transformers.utils.logging.set_verbosity_info()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
 
     # Set seeds
     set_seed(args.seed)

examples/contrib/legacy/run_language_modeling.py

@@ -29,7 +29,6 @@ from typing import Optional
 
 from torch.utils.data import ConcatDataset
 
-import transformers
 from transformers import (
     CONFIG_MAPPING,
     MODEL_WITH_LM_HEAD_MAPPING,
@@ -48,7 +47,6 @@ from transformers import (
     TrainingArguments,
     set_seed,
 )
-from transformers.trainer_utils import is_main_process
 
 
 logger = logging.getLogger(__name__)
@@ -221,11 +219,6 @@ def main():
         bool(training_args.local_rank != -1),
         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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Set seed

examples/contrib/mm-imdb/run_mmimdb.py

@@ -31,7 +31,6 @@ from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
 from torch.utils.data.distributed import DistributedSampler
 from tqdm import tqdm, trange
 
-import transformers
 from transformers import (
     WEIGHTS_NAME,
     AdamW,
@@ -42,7 +41,6 @@ from transformers import (
     MMBTForClassification,
     get_linear_schedule_with_warmup,
 )
-from transformers.trainer_utils import is_main_process
 from utils_mmimdb import ImageEncoder, JsonlDataset, collate_fn, get_image_transforms, get_mmimdb_labels
 
 
@@ -478,11 +476,7 @@ def main():
         bool(args.local_rank != -1),
         args.fp16,
     )
-    # Set the verbosity to info of the Transformers logger (on main process only):
-    if is_main_process(args.local_rank):
-        transformers.utils.logging.set_verbosity_info()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
+
     # Set seed
     set_seed(args)
 

examples/contrib/run_swag.py

@@ -31,10 +31,8 @@ from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, Tenso
 from torch.utils.data.distributed import DistributedSampler
 from tqdm import tqdm, trange
 
-import transformers
 from transformers import WEIGHTS_NAME, AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup
 from transformers.modeling_auto import AutoModelForMultipleChoice
-from transformers.trainer_utils import is_main_process
 
 
 try:
@@ -622,11 +620,6 @@ def main():
         bool(args.local_rank != -1),
         args.fp16,
     )
-    # Set the verbosity to info of the Transformers logger (on main process only):
-    if is_main_process(args.local_rank):
-        transformers.utils.logging.set_verbosity_info()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
 
     # Set seed
     set_seed(args)

examples/deebert/run_glue_deebert.py

@@ -13,7 +13,6 @@ from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, Tenso
 from torch.utils.data.distributed import DistributedSampler
 from tqdm import tqdm, trange
 
-import transformers
 from src.modeling_highway_bert import DeeBertForSequenceClassification
 from src.modeling_highway_roberta import DeeRobertaForSequenceClassification
 from transformers import (
@@ -29,7 +28,6 @@ from transformers import glue_compute_metrics as compute_metrics
 from transformers import glue_convert_examples_to_features as convert_examples_to_features
 from transformers import glue_output_modes as output_modes
 from transformers import glue_processors as processors
-from transformers.trainer_utils import is_main_process
 
 
 try:
@@ -582,11 +580,7 @@ def main():
         bool(args.local_rank != -1),
         args.fp16,
     )
-    # Set the verbosity to info of the Transformers logger (on main process only):
-    if is_main_process(args.local_rank):
-        transformers.utils.logging.set_verbosity_info()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
+
     # Set seed
     set_seed(args)
 

examples/distillation/run_squad_w_distillation.py

@@ -30,7 +30,6 @@ from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
 from torch.utils.data.distributed import DistributedSampler
 from tqdm import tqdm, trange
 
-import transformers
 from transformers import (
     WEIGHTS_NAME,
     AdamW,
@@ -58,7 +57,6 @@ from transformers.data.metrics.squad_metrics import (
     squad_evaluate,
 )
 from transformers.data.processors.squad import SquadResult, SquadV1Processor, SquadV2Processor
-from transformers.trainer_utils import is_main_process
 
 
 try:
@@ -747,11 +745,7 @@ def main():
         bool(args.local_rank != -1),
         args.fp16,
     )
-    # Set the verbosity to info of the Transformers logger (on main process only):
-    if is_main_process(args.local_rank):
-        transformers.utils.logging.set_verbosity_info()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
+
     # Set seed
     set_seed(args)
 

examples/language-modeling/run_clm.py

@@ -168,8 +168,6 @@ def main():
     # 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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Set seed before initializing model.
@@ -256,7 +254,7 @@ def main():
         tokenize_function,
         batched=True,
         num_proc=data_args.preprocessing_num_workers,
-        remove_columns=column_names,
+        remove_columns=[text_column_name],
         load_from_cache_file=not data_args.overwrite_cache,
     )
 

examples/language-modeling/run_mlm.py

@@ -179,8 +179,6 @@ def main():
     # 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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Set seed before initializing model.
@@ -294,7 +292,7 @@ def main():
             tokenize_function,
             batched=True,
             num_proc=data_args.preprocessing_num_workers,
-            remove_columns=column_names,
+            remove_columns=[text_column_name],
             load_from_cache_file=not data_args.overwrite_cache,
         )
 

examples/language-modeling/run_mlm_wwm.py

@@ -186,8 +186,6 @@ def main():
     # 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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Set seed before initializing model.

examples/language-modeling/run_plm.py

@@ -176,8 +176,6 @@ def main():
     # 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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Set seed before initializing model.
@@ -281,7 +279,7 @@ def main():
             tokenize_function,
             batched=True,
             num_proc=data_args.preprocessing_num_workers,
-            remove_columns=column_names,
+            remove_columns=[text_column_name],
             load_from_cache_file=not data_args.overwrite_cache,
         )
 

examples/lxmert/demo.ipynb

@@ -210,6 +210,7 @@
     "        visual_feats=features,\n",
     "        visual_pos=normalized_boxes,\n",
     "        token_type_ids=inputs.token_type_ids,\n",
+    "        return_dict=True,\n",
     "        output_attentions=False,\n",
     "    )\n",
     "    output_vqa = lxmert_vqa(\n",
@@ -218,6 +219,7 @@
     "        visual_feats=features,\n",
     "        visual_pos=normalized_boxes,\n",
     "        token_type_ids=inputs.token_type_ids,\n",
+    "        return_dict=True,\n",
     "        output_attentions=False,\n",
     "    )\n",
     "    # get prediction\n",
@@ -264,4 +266,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 4
-}
+}

examples/multiple-choice/run_multiple_choice.py

@@ -23,7 +23,6 @@ from typing import Dict, Optional
 
 import numpy as np
 
-import transformers
 from transformers import (
     AutoConfig,
     AutoModelForMultipleChoice,
@@ -34,7 +33,6 @@ from transformers import (
     TrainingArguments,
     set_seed,
 )
-from transformers.trainer_utils import is_main_process
 from utils_multiple_choice import MultipleChoiceDataset, Split, processors
 
 
@@ -117,11 +115,6 @@ def main():
         bool(training_args.local_rank != -1),
         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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Set seed

examples/multiple-choice/run_tf_multiple_choice.py

@@ -33,15 +33,9 @@ from transformers import (
     TFTrainingArguments,
     set_seed,
 )
-from transformers.utils import logging as hf_logging
 from utils_multiple_choice import Split, TFMultipleChoiceDataset, processors
 
 
-hf_logging.set_verbosity_info()
-hf_logging.enable_default_handler()
-hf_logging.enable_explicit_format()
-
-
 logger = logging.getLogger(__name__)
 
 

examples/question-answering/run_squad.py

@@ -29,7 +29,6 @@ from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
 from torch.utils.data.distributed import DistributedSampler
 from tqdm import tqdm, trange
 
-import transformers
 from transformers import (
     MODEL_FOR_QUESTION_ANSWERING_MAPPING,
     WEIGHTS_NAME,
@@ -46,7 +45,6 @@ from transformers.data.metrics.squad_metrics import (
     squad_evaluate,
 )
 from transformers.data.processors.squad import SquadResult, SquadV1Processor, SquadV2Processor
-from transformers.trainer_utils import is_main_process
 
 
 try:
@@ -321,7 +319,7 @@ def evaluate(args, model, tokenizer, prefix=""):
             eval_feature = features[feature_index.item()]
             unique_id = int(eval_feature.unique_id)
 
-            output = [to_list(output[i]) for output in outputs.to_tuple()]
+            output = [to_list(output[i]) for output in outputs]
 
             # Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
             # models only use two.
@@ -714,11 +712,7 @@ def main():
         bool(args.local_rank != -1),
         args.fp16,
     )
-    # Set the verbosity to info of the Transformers logger (on main process only):
-    if is_main_process(args.local_rank):
-        transformers.utils.logging.set_verbosity_info()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
+
     # Set seed
     set_seed(args)
 
@@ -736,7 +730,6 @@ def main():
         args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
         do_lower_case=args.do_lower_case,
         cache_dir=args.cache_dir if args.cache_dir else None,
-        use_fast=False,  # SquadDataset is not compatible with Fast tokenizers which have a smarter overflow handeling
     )
     model = AutoModelForQuestionAnswering.from_pretrained(
         args.model_name_or_path,
@@ -785,10 +778,7 @@ def main():
 
         # Load a trained model and vocabulary that you have fine-tuned
         model = AutoModelForQuestionAnswering.from_pretrained(args.output_dir)  # , force_download=True)
-
-        # SquadDataset is not compatible with Fast tokenizers which have a smarter overflow handeling
-        # So we use use_fast=False here for now until Fast-tokenizer-compatible-examples are out
-        tokenizer = AutoTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case, use_fast=False)
+        tokenizer = AutoTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
         model.to(args.device)
 
     # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory

examples/question-answering/run_squad_trainer.py

@@ -22,11 +22,9 @@ import sys
 from dataclasses import dataclass, field
 from typing import Optional
 
-import transformers
 from transformers import AutoConfig, AutoModelForQuestionAnswering, AutoTokenizer, HfArgumentParser, SquadDataset
 from transformers import SquadDataTrainingArguments as DataTrainingArguments
 from transformers import Trainer, TrainingArguments
-from transformers.trainer_utils import is_main_process
 
 
 logger = logging.getLogger(__name__)
@@ -93,11 +91,6 @@ def main():
         bool(training_args.local_rank != -1),
         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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Prepare Question-Answering task
@@ -114,7 +107,6 @@ def main():
     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=False,  # SquadDataset is not compatible with Fast tokenizers which have a smarter overflow handeling
     )
     model = AutoModelForQuestionAnswering.from_pretrained(
         model_args.model_name_or_path,

examples/question-answering/run_tf_squad.py

@@ -33,12 +33,6 @@ from transformers import (
     squad_convert_examples_to_features,
 )
 from transformers.data.processors.squad import SquadV1Processor, SquadV2Processor
-from transformers.utils import logging as hf_logging
-
-
-hf_logging.set_verbosity_info()
-hf_logging.enable_default_handler()
-hf_logging.enable_explicit_format()
 
 
 logger = logging.getLogger(__name__)

examples/rag/eval_rag.py

@@ -95,7 +95,7 @@ def evaluate_batch_retrieval(args, rag_model, questions):
         truncation=True,
     )["input_ids"].to(args.device)
 
-    question_enc_outputs = rag_model.rag.question_encoder(retriever_input_ids)
+    question_enc_outputs = rag_model.rag.question_encoder(retriever_input_ids, return_dict=True)
     question_enc_pool_output = question_enc_outputs.pooler_output
 
     result = rag_model.retriever(

examples/rag/finetune.py

@@ -204,6 +204,7 @@ class GenerativeQAModule(BaseTransformer):
             decoder_input_ids=decoder_input_ids,
             use_cache=False,
             labels=lm_labels,
+            return_dict=True,
             **rag_kwargs,
         )
 

examples/rag/use_own_knowledge_dataset.py

@@ -47,7 +47,7 @@ def embed(documents: dict, ctx_encoder: DPRContextEncoder, ctx_tokenizer: DPRCon
     input_ids = ctx_tokenizer(
         documents["title"], documents["text"], truncation=True, padding="longest", return_tensors="pt"
     )["input_ids"]
-    embeddings = ctx_encoder(input_ids.to(device=device)).pooler_output
+    embeddings = ctx_encoder(input_ids.to(device=device), return_dict=True).pooler_output
     return {"embeddings": embeddings.detach().cpu().numpy()}
 
 

examples/requirements.txt

@@ -18,4 +18,3 @@ fire
 pytest
 conllu
 sentencepiece != 0.1.92
-protobuf

examples/seq2seq/README.md

@@ -380,7 +380,7 @@ cp xsum/test* all_pl
 then use `all_pl` as DATA in the command above.
 
 #### Direct Knowledge Distillation (KD)
-+ In this method, we use try to enforce that the student and teacher produce similar encoder_outputs, logits, and hidden_states using `SummarizationDistiller`.
++ In this method, we use try to enforce that the student and teacher produce similar encoder_outputs, logits, and hidden_states using `BartSummarizationDistiller`.
 + This method was used for `sshleifer/distilbart-xsum-12-6`, `6-6`, and `9-6` checkpoints were produced.
 + You must use [`distillation.py`](./distillation.py). Note that this command initializes the student for you.
 

examples/seq2seq/distillation.py

@@ -25,8 +25,8 @@ sys.path.insert(2, str(Path(__file__).resolve().parents[1]))
 from lightning_base import generic_train  # noqa
 
 
-class SummarizationDistiller(SummarizationModule):
-    """Supports T5, Bart, Pegasus and other models that inherit from Bart."""
+class BartSummarizationDistiller(SummarizationModule):
+    """Supports Bart, Pegasus and other models that inherit from Bart."""
 
     loss_names = ["loss", "ce_loss", "mlm_loss", "hid_loss_enc", "hid_loss_dec"]
 
@@ -40,38 +40,26 @@ class SummarizationDistiller(SummarizationModule):
         hparams.model_name_or_path = str(save_dir)  # Tell lightning we are training the student
         teacher = AutoModelForSeq2SeqLM.from_pretrained(hparams.teacher).eval()
         use_task_specific_params(teacher, hparams.task)  # We copy good generation parameters to student by default
-        if hparams.student is not None:
-            student = AutoModelForSeq2SeqLM.from_pretrained(hparams.student)
-            use_task_specific_params(student, hparams.task)
-            e_layer_ids, d_layer_ids = None, None
-        else:
-            student, e_layer_ids, d_layer_ids = create_student_by_copying_alternating_layers(
-                teacher, e=hparams.student_encoder_layers, d=hparams.student_decoder_layers, save_path=save_dir
-            )
-
+        student, e_layer_ids, d_layer_ids = create_student_by_copying_alternating_layers(
+            teacher, e=hparams.student_encoder_layers, d=hparams.student_decoder_layers, save_path=save_dir
+        )
         if hparams.length_penalty != -1:
             student.config.length_penalty = hparams.length_penalty
         hparams.tokenizer_name = hparams.teacher  # Use teacher's tokenizer
         super().__init__(hparams, model=student, config=student.config)
-        assert (
-            student.config.model_type == teacher.config.model_type
-        ), f"teacher, student model types should be the same, got {student.config.model_type} != {teacher.config.model_type}"
+        model_type = student.config.model_type
+        self.e_layer_ids, self.d_layer_ids = e_layer_ids, d_layer_ids  # type: List[int], List[int]
 
-        if student.config.model_type == "t5":
-            student_encoder_layers = len(student.get_encoder().block)
-            student_decoder_layers = len(student.get_decoder().block)
+        if model_type == "t5":
             teacher_encoder_layers = len(teacher.get_encoder().block)
             teacher_decoder_layers = len(teacher.get_decoder().block)
         else:
-            student_encoder_layers = student.config.encoder_layers
-            student_decoder_layers = student.config.decoder_layers
             teacher_encoder_layers = teacher.config.encoder_layers
             teacher_decoder_layers = teacher.config.decoder_layers
 
-        self.different_base_models = not (hparams.student is None or hparams.teacher == hparams.student)
-        self.do_calc_hidden_loss = (not self.different_base_models) and hparams.alpha_hid > 0
-        self.different_encoder = self.different_base_models or (student_encoder_layers != teacher_encoder_layers)
-        # self.different_encoder determines whether we need to run the teacher encoder
+        self.different_encoder = hparams.student_encoder_layers != teacher_encoder_layers
+        self.different_decoder = hparams.student_decoder_layers != teacher_decoder_layers
+
         self.teacher = teacher
         freeze_params(self.teacher)
 
@@ -80,28 +68,13 @@ class SummarizationDistiller(SummarizationModule):
                 del self.teacher.model.encoder
             except AttributeError:  # T5
                 del self.teacher.encoder
-
-        if e_layer_ids is None:
-            e_layer_ids = list(range(student_encoder_layers))
-        if d_layer_ids is None:
-            d_layer_ids = list(range(student_decoder_layers))
-
-        self.e_layer_ids, self.d_layer_ids = e_layer_ids, d_layer_ids  # type: List[int], List[int]
-
-        if self.do_calc_hidden_loss:  # Intermediate supervision: Decide which layers to supervise
-            if hparams.supervise_forward:
-                self.e_matches = get_layers_to_supervise(
-                    n_student=len(self.e_layer_ids), n_teacher=teacher_encoder_layers
-                )
-                self.d_matches = get_layers_to_supervise(
-                    n_student=len(self.d_layer_ids), n_teacher=teacher_decoder_layers
-                )
-            else:  # student layer should emulate hidden states of the teacher layer it was copied from
-                self.e_matches = self.e_layer_ids
-                self.d_matches = self.d_layer_ids
-        else:
-            self.e_matches = None
-            self.d_matches = None
+        # Intermediate supervision: Decide which layers to supervise
+        if hparams.supervise_forward:
+            self.e_matches = get_layers_to_supervise(n_student=len(self.e_layer_ids), n_teacher=teacher_encoder_layers)
+            self.d_matches = get_layers_to_supervise(n_student=len(self.d_layer_ids), n_teacher=teacher_decoder_layers)
+        else:  # student layer should emulate hidden states of the teacher layer it was copied from
+            self.e_matches = self.e_layer_ids
+            self.d_matches = self.d_layer_ids
 
         self.ce_loss_fct = nn.KLDivLoss(reduction="batchmean")
         self.temperature = 2.0
@@ -111,8 +84,22 @@ class SummarizationDistiller(SummarizationModule):
         gc.collect()
         torch.cuda.empty_cache()
 
+    def calc_mse_loss(self, teacher_outputs: torch.Tensor, student_outputs: torch.Tensor, mask) -> torch.FloatTensor:
+        """Supervise MSE(teacher.encoder_outputs, student.encoder_outputs)."""
+        # raise NotImplementedError()
+        if mask is not None:
+            # mask has False at padding_idx
+            sel_mask = mask[:, :, None].expand_as(student_outputs).bool()
+            s_logits_slct = torch.masked_select(student_outputs, sel_mask)
+            t_logits_slct = torch.masked_select(teacher_outputs, sel_mask)
+        else:
+            t_logits_slct = teacher_outputs
+            s_logits_slct = student_outputs
+        return F.mse_loss(s_logits_slct, t_logits_slct)
+
     def calc_ce_loss(self, mask, s_logits, t_logits):
         """Copy pasted from distillbert (transformers/examples/distillation/)"""
+
         # mask has False at padding_idx
         sel_mask = mask[:, :, None].expand_as(s_logits)
         vocab_size = s_logits.size(-1)
@@ -136,8 +123,8 @@ class SummarizationDistiller(SummarizationModule):
         add_distill_args(parser)
         return parser
 
-    def _step(self, batch: dict) -> tuple:
-        """Compute the loss for a batch"""
+    def _step(self, batch):
+        # assert is_frozen(self.teacher) copied_decoder_layers
         pad_token_id = self.tokenizer.pad_token_id
         input_ids, src_mask, labels = batch["input_ids"], batch["attention_mask"], batch["labels"]
         if isinstance(self.model, T5ForConditionalGeneration):
@@ -146,15 +133,14 @@ class SummarizationDistiller(SummarizationModule):
             decoder_input_ids = shift_tokens_right(labels, pad_token_id)
 
         # noinspection PyCallingNonCallable
-        student_outputs = self(
+        lm_logits, dec_hidden, enc_outputs, enc_hidden_state = self(
             input_ids,
             attention_mask=src_mask,
             decoder_input_ids=decoder_input_ids,
-            output_hidden_states=self.do_calc_hidden_loss,
+            output_hidden_states=True,
             output_attentions=False,
             use_cache=False,
         )
-        lm_logits = student_outputs.logits
 
         # Same cross entropy vs. label smoothing logic as finetune.py
         assert lm_logits.shape[-1] == self.model.config.vocab_size
@@ -163,7 +149,7 @@ class SummarizationDistiller(SummarizationModule):
             loss_fct = torch.nn.CrossEntropyLoss(ignore_index=pad_token_id)
             student_lm_loss = loss_fct(lm_logits.view(-1, lm_logits.shape[-1]), labels.view(-1))
         else:
-            lprobs = F.log_softmax(lm_logits, dim=-1)
+            lprobs = torch.nn.functional.log_softmax(lm_logits, dim=-1)
             student_lm_loss, _ = label_smoothed_nll_loss(
                 lprobs, labels, self.hparams.label_smoothing, ignore_index=pad_token_id
             )
@@ -171,42 +157,37 @@ class SummarizationDistiller(SummarizationModule):
         def zero_tensor():
             return torch.tensor(0.0).type_as(student_lm_loss)
 
-        teacher_enc_outputs = student_outputs.encoder_last_hidden_state  # use this unless self.different_base_models
         hid_loss_enc, hid_loss_dec = zero_tensor(), zero_tensor()
         if self.different_encoder:  # compute encoder hidden state loss
-            all_teacher_encoder_outputs = self.teacher.get_encoder()(
+            with torch.no_grad():
+                teacher_enc_hid = self.teacher.get_encoder()(
+                    input_ids, attention_mask=src_mask, output_hidden_states=True, return_dict=True
+                ).hidden_states
+
+            hid_loss_enc = self.calc_hidden_loss(
+                src_mask,
+                enc_hidden_state,
+                teacher_enc_hid,
+                self.e_matches,
+                normalize_hidden=self.hparams.normalize_hidden,
+            )
+
+        with torch.no_grad():
+            outputs = self.teacher(
                 input_ids,
                 attention_mask=src_mask,
-                output_hidden_states=self.do_calc_hidden_loss,
+                encoder_outputs=(enc_outputs,),
+                decoder_input_ids=decoder_input_ids,
+                lm_labels=labels,
+                output_hidden_states=True,
+                return_dict=True,
             )
-            if self.different_base_models:
-                teacher_enc_outputs = all_teacher_encoder_outputs.last_hidden_state
-            elif self.do_calc_hidden_loss:
-                hid_loss_enc = self.calc_hidden_loss(
-                    src_mask,
-                    student_outputs.encoder_hidden_states,
-                    all_teacher_encoder_outputs.hidden_states,
-                    self.e_matches,
-                    normalize_hidden=self.hparams.normalize_hidden,
-                )
-
-        teacher_outputs = self.teacher(
-            input_ids,
-            attention_mask=src_mask,
-            encoder_outputs=(teacher_enc_outputs,),
-            decoder_input_ids=decoder_input_ids,
-            output_hidden_states=self.do_calc_hidden_loss,
-            use_cache=False,  # since we are not passing labels, never let this default to True
-        )
+            tlogits, tdec_hidden = outputs.logits, outputs.decoder_hidden_states
         dec_mask = decoder_input_ids.ne(pad_token_id)
-        loss_ce = self.calc_ce_loss(dec_mask, lm_logits, teacher_outputs.logits)
-        if self.do_calc_hidden_loss:  # Intermediate supervision of decoder hidden states
+        loss_ce = self.calc_ce_loss(dec_mask, lm_logits, tlogits)
+        if self.alpha_hid > 0:  # Intermediate supervision of decoder hidden states
             hid_loss_dec = self.calc_hidden_loss(
-                dec_mask,
-                student_outputs.decoder_hidden_states,
-                teacher_outputs.decoder_hidden_states,
-                self.d_matches,
-                normalize_hidden=self.hparams.normalize_hidden,
+                dec_mask, dec_hidden, tdec_hidden, self.d_matches, normalize_hidden=self.hparams.normalize_hidden
             )
 
         blended_loss = (
@@ -226,7 +207,6 @@ class SummarizationDistiller(SummarizationModule):
         valid_count = mask.sum() * hidden_states[0].size(-1)
         student_states = torch.stack([hidden_states[i] for i in range(len(matches))])
         teacher_states = torch.stack([hidden_states_T[j] for j in matches])
-        assert student_states.shape == teacher_states.shape, f"{student_states.shape} != {teacher_states.shape}"
         if normalize_hidden:
             student_states = F.layer_norm(student_states, student_states.shape[1:])
             teacher_states = F.layer_norm(teacher_states, teacher_states.shape[1:])
@@ -236,16 +216,10 @@ class SummarizationDistiller(SummarizationModule):
 
 
 def add_distill_args(parser):
-    # NOTE: if --student argument was specified and the teacher and student base models
-    # are different, the models still have to have the same tokenizer, specified by
-    # --tokenizer_name. So, for example, you can distill from t5_large to t5_small but not
-    # from bart to t5. This s because if the tokenizers are different, the output space
-    # for the two models is also different and their logits are not comparable.
     parser.add_argument("--teacher", type=str)
     parser.add_argument("--alpha_ce", default=0.8, type=float)
     parser.add_argument("--alpha_mlm", default=0.2, type=float)
     parser.add_argument("--alpha_hid", default=0.0, type=float, required=False)
-    parser.add_argument("--student", type=str, required=False)
     parser.add_argument("--student_decoder_layers", default=12, type=int, required=False)
     parser.add_argument("--student_encoder_layers", default=12, type=int, required=False)
     parser.add_argument("--no_teacher", action="store_true", default=False)
@@ -254,8 +228,8 @@ def add_distill_args(parser):
     parser.add_argument("--normalize_hidden", action="store_true", default=False)
 
 
-class TranslationDistiller(SummarizationDistiller):
-    """Supports T5, mBART, Marian, other models that inherit from Bart."""
+class BartTranslationDistiller(BartSummarizationDistiller):
+    """Supports Mbart, Marian, other models that inherit from Bart."""
 
     mode = "translation"
     metric_names = ["bleu"]
@@ -284,7 +258,7 @@ def create_module(args):
     if args.no_teacher:
         module_cls = TranslationModule if "translation" in args.task else SummarizationModule
     else:  # DISTILL WITH TEACHER
-        module_cls = TranslationDistiller if "translation" in args.task else SummarizationDistiller
+        module_cls = BartTranslationDistiller if "translation" in args.task else BartSummarizationDistiller
     args.setup_cls: str = module_cls.__name__
     print(f"using module {args.setup_cls}")
     model = module_cls(args)
@@ -302,7 +276,7 @@ def distill_main(args):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser = pl.Trainer.add_argparse_args(parser)
-    parser = SummarizationDistiller.add_model_specific_args(parser, os.getcwd())
+    parser = BartSummarizationDistiller.add_model_specific_args(parser, os.getcwd())
     args = parser.parse_args()
 
     distill_main(args)

examples/seq2seq/finetune_trainer.py

@@ -4,11 +4,10 @@ import sys
 from dataclasses import dataclass, field
 from typing import Optional
 
-import transformers
 from seq2seq_trainer import Seq2SeqTrainer
 from seq2seq_training_args import Seq2SeqTrainingArguments
 from transformers import AutoConfig, AutoModelForSeq2SeqLM, AutoTokenizer, HfArgumentParser, MBartTokenizer, set_seed
-from transformers.trainer_utils import EvaluationStrategy, is_main_process
+from transformers.trainer_utils import EvaluationStrategy
 from utils import (
     Seq2SeqDataCollator,
     Seq2SeqDataset,
@@ -132,11 +131,6 @@ def main():
         bool(training_args.local_rank != -1),
         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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Set seed

examples/seq2seq/test_bash_script.py

@@ -9,7 +9,7 @@ import pytorch_lightning as pl
 import timeout_decorator
 import torch
 
-from distillation import SummarizationDistiller, distill_main
+from distillation import BartSummarizationDistiller, distill_main
 from finetune import SummarizationModule, main
 from transformers import MarianMTModel
 from transformers.file_utils import cached_path
@@ -170,7 +170,7 @@ class TestDistilMarianNoTeacher(TestCasePlus):
         with patch.object(sys, "argv", testargs):
             parser = argparse.ArgumentParser()
             parser = pl.Trainer.add_argparse_args(parser)
-            parser = SummarizationDistiller.add_model_specific_args(parser, os.getcwd())
+            parser = BartSummarizationDistiller.add_model_specific_args(parser, os.getcwd())
             args = parser.parse_args()
             # assert args.gpus == gpus THIS BREAKS for multi_gpu
 

examples/seq2seq/test_datasets.py

@@ -197,7 +197,7 @@ class TestAll(TestCasePlus):
     )
     @require_torch_non_multi_gpu_but_fix_me
     def test_dataset_kwargs(self, tok_name):
-        tokenizer = AutoTokenizer.from_pretrained(tok_name, use_fast=False)
+        tokenizer = AutoTokenizer.from_pretrained(tok_name)
         if tok_name == MBART_TINY:
             train_dataset = Seq2SeqDataset(
                 tokenizer,

examples/seq2seq/test_seq2seq_examples.py

@@ -96,7 +96,6 @@ CHEAP_ARGS = {
     "freeze_encoder": False,
     "auto_scale_batch_size": False,
     "overwrite_output_dir": False,
-    "student": None,
 }
 
 
@@ -108,7 +107,6 @@ def _dump_articles(path: Path, articles: list):
 ARTICLES = [" Sam ate lunch today.", "Sams lunch ingredients."]
 SUMMARIES = ["A very interesting story about what I ate for lunch.", "Avocado, celery, turkey, coffee"]
 T5_TINY = "patrickvonplaten/t5-tiny-random"
-T5_TINIER = "sshleifer/t5-tinier-random"
 BART_TINY = "sshleifer/bart-tiny-random"
 MBART_TINY = "sshleifer/tiny-mbart"
 MARIAN_TINY = "sshleifer/tiny-marian-en-de"
@@ -185,7 +183,7 @@ class TestSummarizationDistiller(TestCasePlus):
 
     @require_torch_non_multi_gpu_but_fix_me
     def test_loss_fn(self):
-        model = AutoModelForSeq2SeqLM.from_pretrained(BART_TINY)
+        model = AutoModelForSeq2SeqLM.from_pretrained(BART_TINY, return_dict=True)
         input_ids, mask = model.dummy_inputs["input_ids"], model.dummy_inputs["attention_mask"]
         target_ids = torch.tensor([[0, 4, 8, 2], [0, 8, 2, 1]], dtype=torch.long, device=model.device)
         decoder_input_ids = target_ids[:, :-1].contiguous()  # Why this line?
@@ -241,16 +239,6 @@ class TestSummarizationDistiller(TestCasePlus):
         )
         self._test_distiller_cli(updates)
 
-    @require_torch_non_multi_gpu_but_fix_me
-    def test_distill_different_base_models(self):
-        updates = dict(
-            teacher=T5_TINY,
-            student=T5_TINIER,
-            model_name_or_path=T5_TINIER,
-            tokenizer_name=T5_TINIER,
-        )
-        self._test_distiller_cli(updates)
-
     def _test_distiller_cli(self, updates, check_contents=True):
         default_updates = dict(
             label_smoothing=0.0,

examples/text-classification/run_glue.py

@@ -171,8 +171,6 @@ def main():
     # 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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info(f"Training/evaluation parameters {training_args}")
 
     # Set seed before initializing model.

examples/text-classification/run_tf_glue.py

@@ -27,12 +27,6 @@ from transformers import (
     glue_processors,
     glue_tasks_num_labels,
 )
-from transformers.utils import logging as hf_logging
-
-
-hf_logging.set_verbosity_info()
-hf_logging.enable_default_handler()
-hf_logging.enable_explicit_format()
 
 
 class Split(Enum):

examples/text-classification/run_tf_text_classification.py

@@ -21,12 +21,6 @@ from transformers import (
     TFTrainer,
     TFTrainingArguments,
 )
-from transformers.utils import logging as hf_logging
-
-
-hf_logging.set_verbosity_info()
-hf_logging.enable_default_handler()
-hf_logging.enable_explicit_format()
 
 
 def get_tfds(

examples/text-classification/run_xnli.py

@@ -29,7 +29,6 @@ from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, Tenso
 from torch.utils.data.distributed import DistributedSampler
 from tqdm import tqdm, trange
 
-import transformers
 from transformers import (
     WEIGHTS_NAME,
     AdamW,
@@ -42,7 +41,6 @@ from transformers import glue_convert_examples_to_features as convert_examples_t
 from transformers import xnli_compute_metrics as compute_metrics
 from transformers import xnli_output_modes as output_modes
 from transformers import xnli_processors as processors
-from transformers.trainer_utils import is_main_process
 
 
 try:
@@ -528,11 +526,7 @@ def main():
         bool(args.local_rank != -1),
         args.fp16,
     )
-    # Set the verbosity to info of the Transformers logger (on main process only):
-    if is_main_process(args.local_rank):
-        transformers.utils.logging.set_verbosity_info()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
+
     # Set seed
     set_seed(args)
 

examples/token-classification/run_ner.py

@@ -163,8 +163,6 @@ def main():
     # 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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Set seed before initializing model.
@@ -346,7 +344,7 @@ def main():
     if training_args.do_predict:
         logger.info("*** Predict ***")
 
-        test_dataset = tokenized_datasets["test"]
+        test_dataset = datasets["test"]
         predictions, labels, metrics = trainer.predict(test_dataset)
         predictions = np.argmax(predictions, axis=2)
 

examples/token-classification/run_ner_old.py

@@ -25,7 +25,6 @@ import numpy as np
 from seqeval.metrics import accuracy_score, f1_score, precision_score, recall_score
 from torch import nn
 
-import transformers
 from transformers import (
     AutoConfig,
     AutoModelForTokenClassification,
@@ -36,7 +35,6 @@ from transformers import (
     TrainingArguments,
     set_seed,
 )
-from transformers.trainer_utils import is_main_process
 from utils_ner import Split, TokenClassificationDataset, TokenClassificationTask
 
 
@@ -141,11 +139,6 @@ def main():
         bool(training_args.local_rank != -1),
         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()
-        transformers.utils.logging.enable_default_handler()
-        transformers.utils.logging.enable_explicit_format()
     logger.info("Training/evaluation parameters %s", training_args)
 
     # Set seed

examples/token-classification/run_tf_ner.py

@@ -33,15 +33,9 @@ from transformers import (
     TFTrainer,
     TFTrainingArguments,
 )
-from transformers.utils import logging as hf_logging
 from utils_ner import Split, TFTokenClassificationDataset, TokenClassificationTask
 
 
-hf_logging.set_verbosity_info()
-hf_logging.enable_default_handler()
-hf_logging.enable_explicit_format()
-
-
 logger = logging.getLogger(__name__)
 
 

model_architectures.yaml

@@ -0,0 +1,200 @@
+- ALBERT:
+  paper: ALBERT: A Lite BERT for Self-supervised Learning of Language Representations, by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut
+  tags:
+    - encoder
+    - memory-efficient
+    - weight-sharing
+
+- BART:
+  paper: BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
+  tags:
+    - encoder-decoder
+    - denoising
+
+- BERT:
+  paper: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
+  tags:
+    - encoder
+    - bert
+
+- BERT For Sequence Generation:
+  paper: Leveraging Pre-trained Checkpoints for Sequence Generation Tasks by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
+  tags:
+    - encoder-decoder
+    - bert
+
+- Blenderbot:
+  paper:  Recipes for building an open-domain chatbot by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
+  tags:
+    - encoder-decoder
+    - chatbot
+    - conditional-generation
+
+- CamemBERT:
+  paper:  CamemBERT: a Tasty French Language Model by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
+  tags:
+    - encoder
+    - french
+    - bert
+
+- CTRL:
+  paper:  CTRL: A Conditional Transformer Language Model for Controllable Generation by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
+  tags:
+    - decoder
+    - conditional-generation
+
+- DeBERTa:
+  paper:  DeBERTa: Decoding-enhanced BERT with Disentangled Attention by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
+  tags:
+
+- DialoGPT:
+  paper:  DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
+  tags:
+    - decoder
+    - chatbot
+    - conditional-generation
+
+- DistilBERT:
+  paper:  DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into DistilGPT2, RoBERTa into DistilRoBERTa, Multilingual BERT into DistilmBERT and a German version of DistilBERT.
+  tags:
+    - encoder
+    - efficient
+    - distillation
+
+- DPR:
+  paper:  Dense Passage Retrieval for Open-Domain Question Answering by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
+  tags:
+    - encoder
+    - retrieval
+
+- ELECTRA:
+  paper:  ELECTRA: Pre-training text encoders as discriminators rather than generators by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+  tags:
+    - encoder
+    - efficient-training
+
+- FlauBERT:
+  paper: FlauBERT: Unsupervised Language Model Pre-training for French by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
+  tags:
+    - encoder
+    - french
+
+- Funnel Transformer:
+  paper: Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
+  tags:
+    - efficient
+    - bottleneck
+
+- GPT:
+  paper: Improving Language Understanding by Generative Pre-Training by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
+  tags:
+    - decoder
+    - unconditional-generation
+
+- GPT-2:
+  paper: Language Models are Unsupervised Multitask Learners by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
+  tags:
+    - decoder
+    - unconditional-generation
+    - over-1-billion-parameters
+
+- LayoutLM:
+  paper: LayoutLM: Pre-training of Text and Layout for Document Image Understanding by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
+  tags:
+
+- Longformer:
+  paper: Longformer: The Long-Document Transformer by Iz Beltagy, Matthew E. Peters, Arman Cohan.
+  tags:
+    - encoder
+    - efficient-attention
+    - data-independant-pattern
+    - long-inputs
+
+- LXMERT:
+  paper: LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering by Hao Tan and Mohit Bansal.
+  tags:
+    - encoder
+    - multi-modal
+    - image
+    - text
+
+- MarianMT:
+  paper: Machine translation models trained using OPUS data by Jörg Tiedemann. The Marian Framework is being developed by the Microsoft Translator Team.
+  tags:
+    - encoder-decoder``
+    - translation
+
+- MBart:
+  paper: Multilingual Denoising Pre-training for Neural Machine Translation by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
+  tags:
+    - encoder-decoder
+    - multi-lingual
+    - denoising
+
+- Pegasus:
+  paper: PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization> by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
+  tags:
+    - efficient-pretraining
+
+- ProphetNet:
+  paper: ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
+  tags:
+    - encoder-decoder
+    - future-prediction
+
+- Reformer:
+  paper: Reformer: The Efficient Transformer by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
+  tags:
+    - encoder
+    - efficient-attention
+    - data-dependant-pattern
+    - long-inputs
+
+- RoBERTa:
+  paper: a Robustly Optimized BERT Pretraining Approach by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov. ultilingual BERT into DistilmBERT and a German version of DistilBERT.
+  tags:
+    - encoder
+    - large-dataset-training
+
+- SqueezeBert:
+  paper SqueezeBERT: What can computer vision teach NLP about efficient neural networks? by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
+  tags:
+
+- T5:
+  paper: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer 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.
+  tags:
+    - encoder-decoder
+    - large-dataset-training
+    - over-1-billion-parameters
+
+- Transformer-XL:
+  paper: Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
+  tags:
+    - decoder
+    - long-inputs
+
+- XLM:
+  paper: Cross-lingual Language Model Pretraining by Guillaume Lample and Alexis Conneau.
+  tags:
+    - encoder
+    - multilingual
+
+- XLM-ProphetNet:
+  paper: ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
+  tags:
+    - encoder-decoder
+    - future-prediction
+    - mutlilingual
+
+- XLM-RoBERTa:
+  paper: Unsupervised Cross-lingual Representation Learning at Scale by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
+  tags:
+    - encoder
+    - mutlilingual
+
+- XLNet:
+  paper: XLNet: Generalized Autoregressive Pretraining for Language Understanding by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
+  tags:
+    - decoder
+    - bidirectional
+    - auto-regressive

model_cards/deepset/roberta-base-squad2-v2/README.md

@@ -90,9 +90,9 @@ tokenizer = Tokenizer.load(model_name)
 ### In haystack
 For doing QA at scale (i.e. many docs instead of single paragraph), you can load the model also in [haystack](https://github.com/deepset-ai/haystack/):
 ```python
-reader = FARMReader(model_name_or_path="deepset/roberta-base-squad2-v2")
+reader = FARMReader(model_name_or_path="deepset/roberta-base-squad2")
 # or 
-reader = TransformersReader(model_name_or_path="deepset/roberta-base-squad2-v2",tokenizer="deepset/roberta-base-squad2-v2")
+reader = TransformersReader(model="deepset/roberta-base-squad2",tokenizer="deepset/roberta-base-squad2")
 ```
 
 

model_cards/deepset/roberta-base-squad2/README.md

@@ -5,11 +5,7 @@ datasets:
 
 # roberta-base for QA 
 
-NOTE: This is version 2 of the model. See [this github issue](https://github.com/deepset-ai/FARM/issues/552) from the FARM repository for an explanation of why we updated. If you'd like to use version 1, specify `revision="v1.0"` when loading the model in Transformers 3.5. For exmaple:
-```
-model_name = "deepset/roberta-base-squad2"
-pipeline(model=model_name, tokenizer=model_name, revision="v1.0", task="question-answering")
-```
+NOTE: This model has been superseded by deepset/roberta-base-squad2-v2. For an explanation of why, see [this github issue](https://github.com/deepset-ai/FARM/issues/552) from the FARM repository.
 
 ## Overview
 **Language model:** roberta-base  
@@ -23,10 +19,10 @@ pipeline(model=model_name, tokenizer=model_name, revision="v1.0", task="question
 ## Hyperparameters
 
 ```
-batch_size = 96
-n_epochs = 2
+batch_size = 50
+n_epochs = 3
 base_LM_model = "roberta-base"
-max_seq_len = 386
+max_seq_len = 384
 learning_rate = 3e-5
 lr_schedule = LinearWarmup
 warmup_proportion = 0.2
@@ -36,18 +32,9 @@ max_query_length=64
 
 ## Performance
 Evaluated on the SQuAD 2.0 dev set with the [official eval script](https://worksheets.codalab.org/rest/bundles/0x6b567e1cf2e041ec80d7098f031c5c9e/contents/blob/).
-
 ```
-"exact": 79.97136359807968
-"f1": 83.00449234495325
-
-"total": 11873
-"HasAns_exact": 78.03643724696356
-"HasAns_f1": 84.11139298441825
-"HasAns_total": 5928
-"NoAns_exact": 81.90075693860386
-"NoAns_f1": 81.90075693860386
-"NoAns_total": 5945
+"exact": 78.49743114629833,
+"f1": 81.73092721240889
 ```
 
 ## Usage
@@ -98,7 +85,7 @@ For doing QA at scale (i.e. many docs instead of single paragraph), you can load
 ```python
 reader = FARMReader(model_name_or_path="deepset/roberta-base-squad2")
 # or 
-reader = TransformersReader(model_name_or_path="deepset/roberta-base-squad2",tokenizer="deepset/roberta-base-squad2")
+reader = TransformersReader(model="deepset/roberta-base-squad2",tokenizer="deepset/roberta-base-squad2")
 ```
 
 

model_cards/facebook/bart-large-mnli/README.md

@@ -2,80 +2,7 @@
 license: mit
 thumbnail: https://huggingface.co/front/thumbnails/facebook.png
 pipeline_tag: zero-shot-classification
-datasets:
-- multi_nli
+widget:
+- text: "Last week I upgraded my iOS version and ever since then my phone has been overheating whenever I use your app."
+  labels: "mobile, website, billing, account access"
 ---
-
-# bart-large-mnli
-
-This is the checkpoint for [bart-large](https://huggingface.co/facebook/bart-large) after being trained on the [MultiNLI (MNLI)](https://huggingface.co/datasets/multi_nli) dataset.
-
-Additional information about this model:
-- The [bart-large](https://huggingface.co/facebook/bart-large) model page
-- [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
-](https://arxiv.org/abs/1910.13461)
-- [BART fairseq implementation](https://github.com/pytorch/fairseq/tree/master/fairseq/models/bart)
-
-## NLI-based Zero Shot Text Classification
-
-[Yin et al.](https://arxiv.org/abs/1909.00161) proposed a method for using pre-trained NLI models as a ready-made zero-shot sequence classifiers. The method works by posing the sequence to be classified as the NLI premise and to construct a hypothesis from each candidate label. For example, if we want to evaluate whether a sequence belongs to the class "politics", we could construct a hypothesis of `This text is about politics.`. The probabilities for entailment and contradiction are then converted to label probabilities.
-
-This method is surprisingly effective in many cases, particularly when used with larger pre-trained models like BART and Roberta. See [this blog post](https://joeddav.github.io/blog/2020/05/29/ZSL.html) for a more expansive introduction to this and other zero shot methods, and see the code snippets below for examples of using this model for zero-shot classification both with Hugging Face's built-in pipeline and with native Transformers/PyTorch code.
-
-#### With the zero-shot classification pipeline
-
-The model can be loaded with the `zero-shot-classification` pipeline like so:
-
-```python
-from transformers import pipeline
-classifier = pipeline("zero-shot-classification",
-                      model="facebook/bart-large-mnli")
-```
-
-You can then use this pipeline to classify sequences into any of the class names you specify.
-
-```python
-sequence_to_classify = "one day I will see the world"
-candidate_labels = ['travel', 'cooking', 'dancing']
-classifier(sequence_to_classify, candidate_labels)
-#{'labels': ['travel', 'dancing', 'cooking'],
-# 'scores': [0.9938651323318481, 0.0032737774308770895, 0.002861034357920289],
-# 'sequence': 'one day I will see the world'}
-```
-
-If more than one candidate label can be correct, pass `multi_class=True` to calculate each class independently:
-
-```python
-candidate_labels = ['travel', 'cooking', 'dancing', 'exploration']
-classifier(sequence_to_classify, candidate_labels, multi_class=True)
-#{'labels': ['travel', 'exploration', 'dancing', 'cooking'],
-# 'scores': [0.9945111274719238,
-#  0.9383890628814697,
-#  0.0057061901316046715,
-#  0.0018193122232332826],
-# 'sequence': 'one day I will see the world'}
-```
-
-
-#### With manual PyTorch
-
-```python
-# pose sequence as a NLI premise and label as a hypothesis
-from transformers import AutoModelForSequenceClassification, AutoTokenizer
-nli_model = AutoModelForSequenceClassification.from_pretrained('joeddav/xlm-roberta-large-xnli')
-tokenizer = AutoTokenizer.from_pretrained('joeddav/xlm-roberta-large-xnli')
-
-premise = sequence
-hypothesis = f'This example is {label}.'
-
-# run through model pre-trained on MNLI
-x = tokenizer.encode(premise, hypothesis, return_tensors='pt',
-                     truncation_strategy='only_first')
-logits = nli_model(x.to(device))[0]
-
-# we throw away "neutral" (dim 1) and take the probability of
-# "entailment" (2) as the probability of the label being true 
-entail_contradiction_logits = logits[:,[0,2]]
-probs = entail_contradiction_logits.softmax(dim=1)
-prob_label_is_true = probs[:,1]
-```

model_cards/m3hrdadfi/bert2bert-fa-news-headline/README.md

@@ -1,24 +0,0 @@
----
-language: fa
-license: apache-2.0
-tags:
-- summarization
----
-
-A Bert2Bert model on VoA Persian Corpus (a medium-sized corpus of 7.9 million words, 2003-2008) generates headlines. The model achieved a 25.30 ROUGE-2 score. 
- 
-For more detail, please follow the [News Headline Generation](https://github.com/m3hrdadfi/news-headline-generation) repo. 
-
-
-## Eval results
-The following table summarizes the ROUGE scores obtained by the Bert2Bert model.
-
-|    %    | Precision | Recall | FMeasure |
-|:-------:|:---------:|:------:|:--------:|
-| ROUGE-1 |   43.78   |  45.52 |   43.54  |
-| ROUGE-2 |   24.50   | 25.30* |   24.24  |
-| ROUGE-L |   41.20   |  42.22 |   40.76  |
-
-
-## Questions?
-Post a Github issue on the [News Headline Generation](https://github.com/hooshvare/news-headline-generation/issues) repo.

model_cards/m3hrdadfi/bert2bert-fa-wiki-summary/README.md

@@ -1,24 +0,0 @@
----
-language: fa
-license: apache-2.0
-tags:
-- summarization
----
-
-A Bert2Bert model on the Wiki Summary dataset to summarize articles. The model achieved an 8.47 ROUGE-2 score. 
- 
-For more detail, please follow the [Wiki Summary](https://github.com/m3hrdadfi/wiki-summary) repo. 
-
-
-## Eval results
-The following table summarizes the ROUGE scores obtained by the Bert2Bert model.
-
-|    %    | Precision | Recall | FMeasure |
-|:-------:|:---------:|:------:|:--------:|
-| ROUGE-1 |   28.14   |  30.86 |   27.34  |
-| ROUGE-2 |   07.12   | 08.47* |   07.10  |
-| ROUGE-L |   28.49   |  25.87 |   25.50  |
-
-
-## Questions?
-Post a Github issue on the [Wiki Summary](https://github.com/m3hrdadfi/wiki-summary/issues) repo.

model_cards/microsoft/prophetnet-large-uncased/README.md

@@ -23,7 +23,7 @@ target_str = "us rejects charges against its ambassador in bolivia"
 input_ids = tokenizer(input_str, return_tensors="pt").input_ids
 labels = tokenizer(target_str, return_tensors="pt").input_ids
 
-loss = model(input_ids, labels=labels).loss
+loss = model(input_ids, labels=labels, return_dict=True).loss
 ```
 
 ### Citation

model_cards/microsoft/xprophetnet-large-wiki100-cased/README.md

@@ -26,7 +26,7 @@ target_str = "us rejects charges against its ambassador in bolivia"
 input_ids = tokenizer(input_str, return_tensors="pt").input_ids
 labels = tokenizer(target_str, return_tensors="pt").input_ids
 
-loss = model(input_ids, labels=labels).loss
+loss = model(input_ids, labels=labels, return_dict=True).loss
 ```
 
 Note that since this model is a multi-lingual model it can be fine-tuned on all kinds of other languages.

model_cards/mrm8488/codebert-base-finetuned-detect-insecure-code/README.md

@@ -45,7 +45,7 @@ from transformers import AutoTokenizer, AutoModelForSequenceClassification
 import torch
 import numpy as np
 tokenizer = AutoTokenizer.from_pretrained('mrm8488/codebert-base-finetuned-detect-insecure-code')
-model = AutoModelForSequenceClassification.from_pretrained('mrm8488/codebert-base-finetuned-detect-insecure-code')
+model = AutoModelForSequenceClassification.from_pretrained('mrm8488/codebert-base-finetuned-detect-insecure-code', return_dict=True)
 
 inputs = tokenizer("your code here", return_tensors="pt", truncation=True, padding='max_length')
 labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1

model_cards/sentence-transformers/LaBSE/README.md

@@ -13,7 +13,7 @@ sentences = ["Hello World", "Hallo Welt"]
 encoded_input = tokenizer(sentences, padding=True, truncation=True, max_length=64, return_tensors='pt')
 
 with torch.no_grad():
-    model_output = model(**encoded_input)
+    model_output = model(**encoded_input, return_dict=True)
 
 embeddings = model_output.pooler_output
 embeddings = torch.nn.functional.normalize(embeddings)

model_cards/squeezebert/squeezebert-uncased/README.md

@@ -22,7 +22,7 @@ The authors found that SqueezeBERT is 4.3x faster than `bert-base-uncased` on a
 The model is pretrained using the Masked Language Model (MLM) and Sentence Order Prediction (SOP) tasks.
 (Author's note: If you decide to pretrain your own model, and you prefer to train with MLM only, that should work too.)
 
-From the SqueezeBERT paper:
+The SqueezeBERT paper presents 2 approaches to finetuning the model:
 > We pretrain SqueezeBERT from scratch (without distillation) using the [LAMB](https://arxiv.org/abs/1904.00962) optimizer, and we employ the hyperparameters recommended by the LAMB authors: a global batch size of 8192, a learning rate of 2.5e-3, and a warmup proportion of 0.28. Following the LAMB paper's recommendations, we pretrain for 56k steps with a maximum sequence length of 128 and then for 6k steps with a maximum sequence length of 512.
 
 ## Finetuning

model_cards/uer/chinese_roberta_L-2_H-128/README.md

@@ -1,44 +0,0 @@
----
-language: zh
-datasets: 
-- CLUECorpus
----
-
-# Chinese RoBERTa Miniatures
-
-## Model description
-
-This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://www.aclweb.org/anthology/D19-3041.pdf).
-
-You can download the 24 Chinese RoBERTa miniatures either from the [UER-py Github page](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
-
-|   |H=128|H=256|H=512|H=768|
-|---|:---:|:---:|:---:|:---:|
-| **L=2**  |[**2/128 (BERT-Tiny)**][2_128]|[2/256]|[2/512]|[2/768]|
-| **L=4**  |[4/128]|[**4/256 (BERT-Mini)**]|[**4/512 (BERT-Small)**]|[4/768]|
-| **L=6**  |[6/128]|[6/256]|[6/512]|[6/768]|
-| **L=8**  |[8/128]|[8/256]|[**8/512 (BERT-Medium)**]|[8/768]|
-| **L=10** |[10/128]|[10/256]|[10/512]|[10/768]|
-| **L=12** |[12/128]|[12/256]|[12/512]|[**12/768 (BERT-Base)**]|
-
-## Training data
-
-CLUECorpus2020 and CLUECorpusSmall are used as training corpus.
-
-## Training procedure
-
-Training details can be found in [UER-py](https://github.com/dbiir/UER-py/).
-
-### BibTeX entry and citation info
-
-```
-@article{zhao2019uer,
-  title={UER: An Open-Source Toolkit for Pre-training Models},
-  author={Zhao, Zhe and Chen, Hui and Zhang, Jinbin and Zhao, Xin and Liu, Tao and Lu, Wei and Chen, Xi and Deng, Haotang and Ju, Qi and Du, Xiaoyong},
-  journal={EMNLP-IJCNLP 2019},
-  pages={241},
-  year={2019}
-}
-```
-
-[2_128]: https://huggingface.co/uer/chinese_roberta_L-2_H-128

scripts/fsmt/fsmt-make-super-tiny-model.py

@@ -59,7 +59,7 @@ print(f"num of params {tiny_model.num_parameters()}")
 
 # Test
 batch = tokenizer.prepare_seq2seq_batch(["Making tiny model"])
-outputs = tiny_model(**batch)
+outputs = tiny_model(**batch, return_dict=True)
 
 print("test output:", len(outputs.logits[0]))
 

scripts/fsmt/fsmt-make-tiny-model.py

@@ -30,7 +30,7 @@ print(f"num of params {tiny_model.num_parameters()}")
 
 # Test
 batch = tokenizer.prepare_seq2seq_batch(["Making tiny model"])
-outputs = tiny_model(**batch)
+outputs = tiny_model(**batch, return_dict=True)
 
 print("test output:", len(outputs.logits[0]))
 

setup.py

@@ -42,8 +42,6 @@ To create the package for pypi.
 8. Add the release version to docs/source/_static/js/custom.js and .circleci/deploy.sh
 
 9. Update README.md to redirect to correct documentation.
-
-10. Update the version in __init__.py, setup.py to the new version "-dev" and push to master.
 """
 
 import os
@@ -98,15 +96,14 @@ else:
     extras["retrieval"] = ["faiss-cpu", "datasets"]
     extras["flax"] = ["jaxlib==0.1.55", "jax>=0.2.0", "flax==0.2.2"]
 
-extras["tokenizers"] = ["tokenizers==0.9.4"]
+extras["tokenizers"] = ["tokenizers==0.9.2"]
 extras["onnxruntime"] = ["onnxruntime>=1.4.0", "onnxruntime-tools>=1.4.2"]
-extras["modelcreation"] = ["cookiecutter==1.7.2"]
 
 extras["serving"] = ["pydantic", "uvicorn", "fastapi", "starlette"]
 
-extras["sentencepiece"] = ["sentencepiece==0.1.91", "protobuf"]
+extras["sentencepiece"] = ["sentencepiece==0.1.91"]
 extras["retrieval"] = ["faiss-cpu", "datasets"]
-extras["testing"] = ["pytest", "pytest-xdist", "timeout-decorator", "parameterized", "psutil"] + extras["retrieval"] + extras["modelcreation"]
+extras["testing"] = ["pytest", "pytest-xdist", "timeout-decorator", "parameterized", "psutil"] + extras["retrieval"]
 # sphinx-rtd-theme==0.5.0 introduced big changes in the style.
 extras["docs"] = ["recommonmark", "sphinx==3.2.1", "sphinx-markdown-tables", "sphinx-rtd-theme==0.4.3", "sphinx-copybutton"]
 extras["quality"] = ["black >= 20.8b1", "isort >= 5.5.4", "flake8 >= 3.8.3"]
@@ -114,12 +111,12 @@ extras["quality"] = ["black >= 20.8b1", "isort >= 5.5.4", "flake8 >= 3.8.3"]
 
 extras["all"] = extras["tf"] + extras["torch"] + extras["flax"] + extras["sentencepiece"] + extras["tokenizers"]
 
-extras["dev"] = extras["all"] + extras["testing"] + extras["quality"] + extras["ja"] + extras["docs"] + extras["sklearn"] + extras["modelcreation"]
+extras["dev"] = extras["all"] + extras["testing"] + extras["quality"] + extras["ja"] + extras["docs"] + extras["sklearn"]
 
 
 setup(
     name="transformers",
-    version="4.0.0-dev",
+    version="3.5.0",
     author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Sam Shleifer, Patrick von Platen, Sylvain Gugger, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
     author_email="thomas@huggingface.co",
     description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch",
@@ -132,7 +129,7 @@ setup(
     packages=find_packages("src"),
     install_requires=[
         "numpy",
-        "tokenizers == 0.9.4",
+        "tokenizers == 0.9.3",
         # dataclasses for Python versions that don't have it
         "dataclasses;python_version<'3.7'",
         # utilities from PyPA to e.g. compare versions
@@ -145,6 +142,9 @@ setup(
         "tqdm >= 4.27",
         # for OpenAI GPT
         "regex != 2019.12.17",
+        # for SentencePiece models
+        "sentencepiece == 0.1.91",
+        "protobuf",
         # for XLM
         "sacremoses",
     ],

src/transformers/__init__.py

@@ -2,7 +2,7 @@
 # There's no way to ignore "F401 '...' imported but unused" warnings in this
 # module, but to preserve other warnings. So, don't check this module at all.
 
-__version__ = "4.0.0-dev"
+__version__ = "3.5.0"
 
 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -406,9 +406,6 @@ if is_torch_available():
         DistilBertPreTrainedModel,
     )
     from .modeling_dpr import (
-        DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
-        DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
-        DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST,
         DPRContextEncoder,
         DPRPretrainedContextEncoder,
         DPRPretrainedQuestionEncoder,
@@ -716,17 +713,6 @@ if is_tf_available():
         TFDistilBertModel,
         TFDistilBertPreTrainedModel,
     )
-    from .modeling_tf_dpr import (
-        TF_DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
-        TF_DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
-        TF_DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST,
-        TFDPRContextEncoder,
-        TFDPRPretrainedContextEncoder,
-        TFDPRPretrainedQuestionEncoder,
-        TFDPRPretrainedReader,
-        TFDPRQuestionEncoder,
-        TFDPRReader,
-    )
     from .modeling_tf_electra import (
         TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST,
         TFElectraForMaskedLM,

src/transformers/commands/add_new_model.py

@@ -1,192 +0,0 @@
-import json
-import os
-import shutil
-from argparse import ArgumentParser, Namespace
-from pathlib import Path
-from typing import List
-
-from cookiecutter.main import cookiecutter
-from transformers.commands import BaseTransformersCLICommand
-
-from ..utils import logging
-
-
-logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
-
-
-def add_new_model_command_factory(args: Namespace):
-    return AddNewModelCommand(args.testing, args.testing_file, path=args.path)
-
-
-class AddNewModelCommand(BaseTransformersCLICommand):
-    @staticmethod
-    def register_subcommand(parser: ArgumentParser):
-        add_new_model_parser = parser.add_parser("add-new-model")
-        add_new_model_parser.add_argument("--testing", action="store_true", help="If in testing mode.")
-        add_new_model_parser.add_argument("--testing_file", type=str, help="Configuration file on which to run.")
-        add_new_model_parser.add_argument(
-            "--path", type=str, help="Path to cookiecutter. Should only be used for testing purposes."
-        )
-        add_new_model_parser.set_defaults(func=add_new_model_command_factory)
-
-    def __init__(self, testing: bool, testing_file: str, path=None, *args):
-        self._testing = testing
-        self._testing_file = testing_file
-        self._path = path
-
-    def run(self):
-        # Ensure that there is no other `cookiecutter-template-xxx` directory in the current working directory
-        directories = [directory for directory in os.listdir() if "cookiecutter-template-" == directory[:22]]
-        if len(directories) > 0:
-            raise ValueError(
-                "Several directories starting with `cookiecutter-template-` in current working directory. "
-                "Please clean your directory by removing all folders startign with `cookiecutter-template-` or "
-                "change your working directory."
-            )
-
-        path_to_transformer_root = (
-            Path(__file__).parent.parent.parent.parent if self._path is None else Path(self._path).parent.parent
-        )
-        path_to_cookiecutter = path_to_transformer_root / "templates" / "cookiecutter"
-
-        # Execute cookiecutter
-        if not self._testing:
-            cookiecutter(str(path_to_cookiecutter))
-        else:
-            with open(self._testing_file, "r") as configuration_file:
-                testing_configuration = json.load(configuration_file)
-
-            cookiecutter(
-                str(path_to_cookiecutter if self._path is None else self._path),
-                no_input=True,
-                extra_context=testing_configuration,
-            )
-
-        directory = [directory for directory in os.listdir() if "cookiecutter-template-" in directory[:22]][0]
-
-        # Retrieve configuration
-        with open(directory + "/configuration.json", "r") as configuration_file:
-            configuration = json.load(configuration_file)
-
-        lowercase_model_name = configuration["lowercase_modelname"]
-        pytorch_or_tensorflow = configuration["generate_tensorflow_and_pytorch"]
-        os.remove(f"{directory}/configuration.json")
-
-        output_pytorch = "PyTorch" in pytorch_or_tensorflow
-        output_tensorflow = "TensorFlow" in pytorch_or_tensorflow
-
-        shutil.move(
-            f"{directory}/configuration_{lowercase_model_name}.py",
-            f"{path_to_transformer_root}/src/transformers/configuration_{lowercase_model_name}.py",
-        )
-
-        def remove_copy_lines(path):
-            with open(path, "r") as f:
-                lines = f.readlines()
-            with open(path, "w") as f:
-                for line in lines:
-                    if "# Copied from transformers." not in line:
-                        f.write(line)
-
-        if output_pytorch:
-            if not self._testing:
-                remove_copy_lines(f"{directory}/modeling_{lowercase_model_name}.py")
-
-            shutil.move(
-                f"{directory}/modeling_{lowercase_model_name}.py",
-                f"{path_to_transformer_root}/src/transformers/modeling_{lowercase_model_name}.py",
-            )
-
-            shutil.move(
-                f"{directory}/test_modeling_{lowercase_model_name}.py",
-                f"{path_to_transformer_root}/tests/test_modeling_{lowercase_model_name}.py",
-            )
-        else:
-            os.remove(f"{directory}/modeling_{lowercase_model_name}.py")
-            os.remove(f"{directory}/test_modeling_{lowercase_model_name}.py")
-
-        if output_tensorflow:
-            if not self._testing:
-                remove_copy_lines(f"{directory}/modeling_tf_{lowercase_model_name}.py")
-
-            shutil.move(
-                f"{directory}/modeling_tf_{lowercase_model_name}.py",
-                f"{path_to_transformer_root}/src/transformers/modeling_tf_{lowercase_model_name}.py",
-            )
-
-            shutil.move(
-                f"{directory}/test_modeling_tf_{lowercase_model_name}.py",
-                f"{path_to_transformer_root}/tests/test_modeling_tf_{lowercase_model_name}.py",
-            )
-        else:
-            os.remove(f"{directory}/modeling_tf_{lowercase_model_name}.py")
-            os.remove(f"{directory}/test_modeling_tf_{lowercase_model_name}.py")
-
-        shutil.move(
-            f"{directory}/{lowercase_model_name}.rst",
-            f"{path_to_transformer_root}/docs/source/model_doc/{lowercase_model_name}.rst",
-        )
-
-        shutil.move(
-            f"{directory}/tokenization_{lowercase_model_name}.py",
-            f"{path_to_transformer_root}/src/transformers/tokenization_{lowercase_model_name}.py",
-        )
-
-        from os import fdopen, remove
-        from shutil import copymode, move
-        from tempfile import mkstemp
-
-        def replace(original_file: str, line_to_copy_below: str, lines_to_copy: List[str]):
-            # Create temp file
-            fh, abs_path = mkstemp()
-            line_found = False
-            with fdopen(fh, "w") as new_file:
-                with open(original_file) as old_file:
-                    for line in old_file:
-                        new_file.write(line)
-                        if line_to_copy_below in line:
-                            line_found = True
-                            for line_to_copy in lines_to_copy:
-                                new_file.write(line_to_copy)
-
-            if not line_found:
-                raise ValueError(f"Line {line_to_copy_below} was not found in file.")
-
-            # Copy the file permissions from the old file to the new file
-            copymode(original_file, abs_path)
-            # Remove original file
-            remove(original_file)
-            # Move new file
-            move(abs_path, original_file)
-
-        def skip_units(line):
-            return ("generating PyTorch" in line and not output_pytorch) or (
-                "generating TensorFlow" in line and not output_tensorflow
-            )
-
-        def replace_in_files(path_to_datafile):
-            with open(path_to_datafile) as datafile:
-                lines_to_copy = []
-                skip_file = False
-                skip_snippet = False
-                for line in datafile:
-                    if "# To replace in: " in line and "##" not in line:
-                        file_to_replace_in = line.split('"')[1]
-                        skip_file = skip_units(line)
-                    elif "# Below: " in line and "##" not in line:
-                        line_to_copy_below = line.split('"')[1]
-                        skip_snippet = skip_units(line)
-                    elif "# End." in line and "##" not in line:
-                        if not skip_file and not skip_snippet:
-                            replace(file_to_replace_in, line_to_copy_below, lines_to_copy)
-
-                        lines_to_copy = []
-                    elif "# Replace with" in line and "##" not in line:
-                        lines_to_copy = []
-                    elif "##" not in line:
-                        lines_to_copy.append(line)
-
-            remove(path_to_datafile)
-
-        replace_in_files(f"{directory}/to_replace_{lowercase_model_name}.py")
-        os.rmdir(directory)

src/transformers/commands/transformers_cli.py

@@ -1,7 +1,6 @@
 #!/usr/bin/env python
 from argparse import ArgumentParser
 
-from transformers.commands.add_new_model import AddNewModelCommand
 from transformers.commands.convert import ConvertCommand
 from transformers.commands.download import DownloadCommand
 from transformers.commands.env import EnvironmentCommand
@@ -21,7 +20,6 @@ def main():
     RunCommand.register_subcommand(commands_parser)
     ServeCommand.register_subcommand(commands_parser)
     UserCommands.register_subcommand(commands_parser)
-    AddNewModelCommand.register_subcommand(commands_parser)
 
     # Let's go
     args = parser.parse_args()

src/transformers/configuration_auto.py

@@ -59,7 +59,6 @@ from .configuration_xlnet import XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP, XLNetConfi
 ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
     (key, value)
     for pretrained_map in [
-        # Add archive maps here
         BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
         BART_PRETRAINED_CONFIG_ARCHIVE_MAP,
         BLENDERBOT_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -96,7 +95,6 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
 
 CONFIG_MAPPING = OrderedDict(
     [
-        # Add configs here
         ("retribert", RetriBertConfig),
         ("t5", T5Config),
         ("mobilebert", MobileBertConfig),
@@ -138,7 +136,6 @@ CONFIG_MAPPING = OrderedDict(
 
 MODEL_NAMES_MAPPING = OrderedDict(
     [
-        # Add full (and cased) model names here
         ("retribert", "RetriBERT"),
         ("t5", "T5"),
         ("mobilebert", "MobileBERT"),

src/transformers/configuration_t5.py

@@ -75,6 +75,7 @@ class T5Config(PretrainedConfig):
     def __init__(
         self,
         vocab_size=32128,
+        n_positions=512,
         d_model=512,
         d_kv=64,
         d_ff=2048,
@@ -97,6 +98,7 @@ class T5Config(PretrainedConfig):
             **kwargs,
         )
         self.vocab_size = vocab_size
+        self.n_positions = n_positions
         self.d_model = d_model
         self.d_kv = d_kv
         self.d_ff = d_ff
@@ -110,6 +112,10 @@ class T5Config(PretrainedConfig):
         self.layer_norm_epsilon = layer_norm_epsilon
         self.initializer_factor = initializer_factor
 
+    @property
+    def max_position_embeddings(self):
+        return self.n_positions
+
     @property
     def hidden_size(self):
         return self.d_model

src/transformers/configuration_utils.py

@@ -55,7 +55,7 @@ class PretrainedConfig(object):
             Whether or not the model should returns all attentions.
         use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
             Whether or not the model should return the last key/values attentions (not used by all models).
-        return_dict (:obj:`bool`, `optional`, defaults to :obj:`True`):
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`False`):
             Whether or not the model should return a :class:`~transformers.file_utils.ModelOutput` instead of a plain
             tuple.
         is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`False`):
@@ -163,7 +163,7 @@ class PretrainedConfig(object):
 
     def __init__(self, **kwargs):
         # Attributes with defaults
-        self.return_dict = kwargs.pop("return_dict", True)
+        self.return_dict = kwargs.pop("return_dict", False)
         self.output_hidden_states = kwargs.pop("output_hidden_states", False)
         self.output_attentions = kwargs.pop("output_attentions", False)
         self.use_cache = kwargs.pop("use_cache", True)  # Not used by all models

src/transformers/convert_pytorch_checkpoint_to_tf2.py

@@ -25,9 +25,6 @@ from transformers import (
     CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,
     DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
-    DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
-    DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
-    DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST,
     ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP,
     FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -46,7 +43,6 @@ from transformers import (
     CamembertConfig,
     CTRLConfig,
     DistilBertConfig,
-    DPRConfig,
     ElectraConfig,
     FlaubertConfig,
     GPT2Config,
@@ -63,9 +59,6 @@ from transformers import (
     TFCTRLLMHeadModel,
     TFDistilBertForMaskedLM,
     TFDistilBertForQuestionAnswering,
-    TFDPRContextEncoder,
-    TFDPRQuestionEncoder,
-    TFDPRReader,
     TFElectraForPreTraining,
     TFFlaubertWithLMHeadModel,
     TFGPT2LMHeadModel,
@@ -105,9 +98,6 @@ if is_torch_available():
         CTRLLMHeadModel,
         DistilBertForMaskedLM,
         DistilBertForQuestionAnswering,
-        DPRContextEncoder,
-        DPRQuestionEncoder,
-        DPRReader,
         ElectraForPreTraining,
         FlaubertWithLMHeadModel,
         GPT2LMHeadModel,
@@ -157,18 +147,6 @@ MODEL_CLASSES = {
         BertForSequenceClassification,
         BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
-    "dpr": (
-        DPRConfig,
-        TFDPRQuestionEncoder,
-        TFDPRContextEncoder,
-        TFDPRReader,
-        DPRQuestionEncoder,
-        DPRContextEncoder,
-        DPRReader,
-        DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
-        DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
-        DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST,
-    ),
     "gpt2": (
         GPT2Config,
         TFGPT2LMHeadModel,

src/transformers/convert_slow_tokenizer.py

@@ -24,7 +24,10 @@ from typing import Dict, List, Tuple
 from tokenizers import Tokenizer, decoders, normalizers, pre_tokenizers, processors
 from tokenizers.models import BPE, Unigram, WordPiece
 
-from .file_utils import requires_protobuf, requires_sentencepiece
+# from transformers.tokenization_openai import OpenAIGPTTokenizer
+from transformers.utils import sentencepiece_model_pb2 as model
+
+from .file_utils import requires_sentencepiece
 
 
 class SentencePieceExtractor:
@@ -61,6 +64,12 @@ def check_number_comma(piece: str) -> bool:
     return len(piece) < 2 or piece[-1] != "," or not piece[-2].isdigit()
 
 
+def get_proto(filename: str):
+    m = model.ModelProto()
+    m.ParseFromString(open(filename, "rb").read())
+    return m
+
+
 class Converter:
     def __init__(self, original_tokenizer):
         self.original_tokenizer = original_tokenizer
@@ -283,15 +292,8 @@ class RobertaConverter(Converter):
 
 class SpmConverter(Converter):
     def __init__(self, *args):
-        requires_protobuf(self)
-
         super().__init__(*args)
-
-        from .utils import sentencepiece_model_pb2 as model_pb2
-
-        m = model_pb2.ModelProto()
-        m.ParseFromString(open(self.original_tokenizer.vocab_file, "rb").read())
-        self.proto = m
+        self.proto = get_proto(self.original_tokenizer.vocab_file)
 
     def vocab(self, proto):
         return [(piece.piece, piece.score) for piece in proto.pieces]

src/transformers/convert_t5_original_tf_checkpoint_to_pytorch.py

@@ -17,6 +17,8 @@
 
 import argparse
 
+import torch
+
 from transformers import T5Config, T5Model, load_tf_weights_in_t5
 from transformers.utils import logging
 
@@ -35,7 +37,7 @@ def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file, pytorch_du
 
     # Save pytorch-model
     print("Save PyTorch model to {}".format(pytorch_dump_path))
-    model.save_pretrained(pytorch_dump_path)
+    torch.save(model.state_dict(), pytorch_dump_path)
 
 
 if __name__ == "__main__":

src/transformers/data/metrics/__init__.py

@@ -28,7 +28,7 @@ if is_sklearn_available():
 DEPRECATION_WARNING = (
     "This metric will be removed from the library soon, metrics should be handled with the 🤗 Datasets "
     "library. You can have a look at this example script for pointers: "
-    "https://github.com/huggingface/transformers/blob/master/examples/text-classification/run_glue.py"
+    "https://github.com/huggingface/transformers/blob/master/examples/text-classification/run_glue.py",
 )
 
 

src/transformers/data/processors/squad.py

@@ -8,7 +8,7 @@ from tqdm import tqdm
 
 from ...file_utils import is_tf_available, is_torch_available
 from ...tokenization_bert import whitespace_tokenize
-from ...tokenization_utils_base import BatchEncoding, PreTrainedTokenizerBase, TruncationStrategy
+from ...tokenization_utils_base import PreTrainedTokenizerBase, TruncationStrategy
 from ...utils import logging
 from .utils import DataProcessor
 
@@ -765,7 +765,6 @@ class SquadFeatures:
         token_to_orig_map: mapping between the tokens and the original text, needed in order to identify the answer.
         start_position: start of the answer token index
         end_position: end of the answer token index
-        encoding: optionally store the BatchEncoding with the fast-tokenizer alignement methods.
     """
 
     def __init__(
@@ -785,7 +784,6 @@ class SquadFeatures:
         end_position,
         is_impossible,
         qas_id: str = None,
-        encoding: BatchEncoding = None,
     ):
         self.input_ids = input_ids
         self.attention_mask = attention_mask
@@ -805,8 +803,6 @@ class SquadFeatures:
         self.is_impossible = is_impossible
         self.qas_id = qas_id
 
-        self.encoding = encoding
-
 
 class SquadResult:
     """

src/transformers/file_utils.py

@@ -185,15 +185,6 @@ except ImportError:
     _sentencepiece_available = False
 
 
-try:
-    import google.protobuf  # noqa: F401
-
-    _protobuf_available = True
-
-except ImportError:
-    _protobuf_available = False
-
-
 try:
     import tokenizers  # noqa: F401
 
@@ -279,10 +270,6 @@ def is_sentencepiece_available():
     return _sentencepiece_available
 
 
-def is_protobuf_available():
-    return _protobuf_available
-
-
 def is_tokenizers_available():
     return _tokenizers_available
 
@@ -343,14 +330,6 @@ that match your environment.
 """
 
 
-# docstyle-ignore
-PROTOBUF_IMPORT_ERROR = """
-{0} requires the protobuf library but it was not found in your environment. Checkout the instructions on the
-installation page of its repo: https://github.com/protocolbuffers/protobuf/tree/master/python#installation and follow the ones
-that match your environment.
-"""
-
-
 # docstyle-ignore
 FAISS_IMPORT_ERROR = """
 {0} requires the faiss library but it was not found in your environment. Checkout the instructions on the
@@ -441,12 +420,6 @@ def requires_sentencepiece(obj):
         raise ImportError(SENTENCEPIECE_IMPORT_ERROR.format(name))
 
 
-def requires_protobuf(obj):
-    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
-    if not is_protobuf_available():
-        raise ImportError(PROTOBUF_IMPORT_ERROR.format(name))
-
-
 def add_start_docstrings(*docstr):
     def docstring_decorator(fn):
         fn.__doc__ = "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
@@ -559,7 +532,7 @@ PT_TOKEN_CLASSIFICATION_SAMPLE = r"""
         >>> import torch
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
         >>> labels = torch.tensor([1] * inputs["input_ids"].size(1)).unsqueeze(0)  # Batch size 1
@@ -576,7 +549,7 @@ PT_QUESTION_ANSWERING_SAMPLE = r"""
         >>> import torch
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
         >>> inputs = tokenizer(question, text, return_tensors='pt')
@@ -596,7 +569,7 @@ PT_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
         >>> import torch
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
         >>> labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
@@ -612,7 +585,7 @@ PT_MASKED_LM_SAMPLE = r"""
         >>> import torch
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> inputs = tokenizer("The capital of France is {mask}.", return_tensors="pt")
         >>> labels = tokenizer("The capital of France is Paris.", return_tensors="pt")["input_ids"]
@@ -629,7 +602,7 @@ PT_BASE_MODEL_SAMPLE = r"""
         >>> import torch
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
         >>> outputs = model(**inputs)
@@ -644,7 +617,7 @@ PT_MULTIPLE_CHOICE_SAMPLE = r"""
         >>> import torch
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
         >>> choice0 = "It is eaten with a fork and a knife."
@@ -666,7 +639,7 @@ PT_CAUSAL_LM_SAMPLE = r"""
         >>> from transformers import {tokenizer_class}, {model_class}
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint})
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
         >>> outputs = model(**inputs, labels=inputs["input_ids"])
@@ -681,7 +654,7 @@ TF_TOKEN_CLASSIFICATION_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
         >>> input_ids = inputs["input_ids"]
@@ -699,7 +672,7 @@ TF_QUESTION_ANSWERING_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
         >>> input_dict = tokenizer(question, text, return_tensors='tf')
@@ -718,7 +691,7 @@ TF_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
         >>> inputs["labels"] = tf.reshape(tf.constant(1), (-1, 1)) # Batch size 1
@@ -735,7 +708,7 @@ TF_MASKED_LM_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> inputs = tokenizer("The capital of France is {mask}.", return_tensors="tf")
         >>> inputs["labels"] = tokenizer("The capital of France is Paris.", return_tensors="tf")["input_ids"]
@@ -752,7 +725,7 @@ TF_BASE_MODEL_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
         >>> outputs = model(inputs)
@@ -767,7 +740,7 @@ TF_MULTIPLE_CHOICE_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
         >>> choice0 = "It is eaten with a fork and a knife."
@@ -788,7 +761,7 @@ TF_CAUSAL_LM_SAMPLE = r"""
         >>> import tensorflow as tf
 
         >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
-        >>> model = {model_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}', return_dict=True)
 
         >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
         >>> outputs = model(inputs)

src/transformers/generation_utils.py

@@ -144,10 +144,6 @@ class GenerationMixin:
         )
         input_ids = input_ids.index_select(0, expanded_return_idx)
 
-        if "token_type_ids" in model_kwargs:
-            token_type_ids = model_kwargs["token_type_ids"]
-            model_kwargs["token_type_ids"] = token_type_ids.index_select(0, expanded_return_idx)
-
         if attention_mask is not None:
             model_kwargs["attention_mask"] = attention_mask.index_select(0, expanded_return_idx)
 
@@ -198,11 +194,6 @@ class GenerationMixin:
         else:
             model_kwargs["past"] = None
 
-        # update token_type_ids with last value
-        if "token_type_ids" in model_kwargs:
-            token_type_ids = model_kwargs["token_type_ids"]
-            model_kwargs["token_type_ids"] = torch.cat([token_type_ids, token_type_ids[:, -1].unsqueeze(-1)], dim=-1)
-
         # update attention mask
         if not is_encoder_decoder:
             if "attention_mask" in model_kwargs:

src/transformers/modeling_albert.py

@@ -416,7 +416,7 @@ class AlbertTransformer(nn.Module):
         head_mask=None,
         output_attentions=False,
         output_hidden_states=False,
-        return_dict=True,
+        return_dict=False,
     ):
         hidden_states = self.embedding_hidden_mapping_in(hidden_states)
 
@@ -764,7 +764,7 @@ class AlbertForPreTraining(AlbertPreTrainedModel):
             >>> import torch
 
             >>> tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
-            >>> model = AlbertForPreTraining.from_pretrained('albert-base-v2')
+            >>> model = AlbertForPreTraining.from_pretrained('albert-base-v2', return_dict=True)
 
             >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
             >>> outputs = model(input_ids)

src/transformers/modeling_auto.py

@@ -226,14 +226,11 @@ from .modeling_xlnet import (
 from .utils import logging
 
 
-# Add modeling imports here
-
 logger = logging.get_logger(__name__)
 
 
 MODEL_MAPPING = OrderedDict(
     [
-        # Base model mapping
         (RetriBertConfig, RetriBertModel),
         (T5Config, T5Model),
         (DistilBertConfig, DistilBertModel),
@@ -269,7 +266,6 @@ MODEL_MAPPING = OrderedDict(
 
 MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
     [
-        # Model for pre-training mapping
         (LayoutLMConfig, LayoutLMForMaskedLM),
         (RetriBertConfig, RetriBertModel),
         (T5Config, T5ForConditionalGeneration),
@@ -299,7 +295,6 @@ MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
 
 MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
     [
-        # Model with LM heads mapping
         (LayoutLMConfig, LayoutLMForMaskedLM),
         (T5Config, T5ForConditionalGeneration),
         (DistilBertConfig, DistilBertForMaskedLM),
@@ -330,7 +325,6 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
 
 MODEL_FOR_CAUSAL_LM_MAPPING = OrderedDict(
     [
-        # Model for Causal LM mapping
         (CamembertConfig, CamembertForCausalLM),
         (XLMRobertaConfig, XLMRobertaForCausalLM),
         (RobertaConfig, RobertaForCausalLM),
@@ -353,7 +347,6 @@ MODEL_FOR_CAUSAL_LM_MAPPING = OrderedDict(
 
 MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
     [
-        # Model for Masked LM mapping
         (LayoutLMConfig, LayoutLMForMaskedLM),
         (DistilBertConfig, DistilBertForMaskedLM),
         (AlbertConfig, AlbertForMaskedLM),
@@ -375,7 +368,6 @@ MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
 
 MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING = OrderedDict(
     [
-        # Model for Seq2Seq Causal LM mapping
         (T5Config, T5ForConditionalGeneration),
         (PegasusConfig, PegasusForConditionalGeneration),
         (MarianConfig, MarianMTModel),
@@ -391,7 +383,6 @@ MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING = OrderedDict(
 
 MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
     [
-        # Model for Sequence Classification mapping
         (DistilBertConfig, DistilBertForSequenceClassification),
         (AlbertConfig, AlbertForSequenceClassification),
         (CamembertConfig, CamembertForSequenceClassification),
@@ -416,7 +407,6 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
 
 MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
     [
-        # Model for Question Answering mapping
         (DistilBertConfig, DistilBertForQuestionAnswering),
         (AlbertConfig, AlbertForQuestionAnswering),
         (CamembertConfig, CamembertForQuestionAnswering),
@@ -439,7 +429,6 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
 
 MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
     [
-        # Model for Token Classification mapping
         (LayoutLMConfig, LayoutLMForTokenClassification),
         (DistilBertConfig, DistilBertForTokenClassification),
         (CamembertConfig, CamembertForTokenClassification),
@@ -461,7 +450,6 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
 
 MODEL_FOR_MULTIPLE_CHOICE_MAPPING = OrderedDict(
     [
-        # Model for Multiple Choice mapping
         (CamembertConfig, CamembertForMultipleChoice),
         (ElectraConfig, ElectraForMultipleChoice),
         (XLMRobertaConfig, XLMRobertaForMultipleChoice),

src/transformers/modeling_bart.py

@@ -329,7 +329,7 @@ class BartEncoder(nn.Module):
         self.layer_norm = LayerNorm(config.d_model) if config.add_final_layer_norm else None
 
     def forward(
-        self, input_ids, attention_mask=None, output_attentions=False, output_hidden_states=False, return_dict=True
+        self, input_ids, attention_mask=None, output_attentions=False, output_hidden_states=False, return_dict=False
     ):
         """
         Args:
@@ -528,7 +528,7 @@ class BartDecoder(nn.Module):
         use_cache=False,
         output_attentions=False,
         output_hidden_states=False,
-        return_dict=True,
+        return_dict=False,
         **unused,
     ):
         """

src/transformers/modeling_bert.py

@@ -446,7 +446,7 @@ class BertEncoder(nn.Module):
         encoder_attention_mask=None,
         output_attentions=False,
         output_hidden_states=False,
-        return_dict=True,
+        return_dict=False,
     ):
         all_hidden_states = () if output_hidden_states else None
         all_self_attentions = () if output_attentions else None
@@ -920,7 +920,7 @@ class BertForPreTraining(BertPreTrainedModel):
             >>> import torch
 
             >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-            >>> model = BertForPreTraining.from_pretrained('bert-base-uncased')
+            >>> model = BertForPreTraining.from_pretrained('bert-base-uncased', return_dict=True)
 
             >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
             >>> outputs = model(**inputs)
@@ -1036,7 +1036,7 @@ class BertLMHeadModel(BertPreTrainedModel):
             >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
             >>> config = BertConfig.from_pretrained("bert-base-cased")
             >>> config.is_decoder = True
-            >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config)
+            >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config, return_dict=True)
 
             >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
             >>> outputs = model(**inputs)
@@ -1250,7 +1250,7 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
             >>> import torch
 
             >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-            >>> model = BertForNextSentencePrediction.from_pretrained('bert-base-uncased')
+            >>> model = BertForNextSentencePrediction.from_pretrained('bert-base-uncased', return_dict=True)
 
             >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
             >>> next_sentence = "The sky is blue due to the shorter wavelength of blue light."

src/transformers/modeling_bert_generation.py

@@ -463,7 +463,7 @@ class BertGenerationDecoder(BertGenerationPreTrainedModel):
             >>> tokenizer = BertGenerationTokenizer.from_pretrained('google/bert_for_seq_generation_L-24_bbc_encoder')
             >>> config = BertGenerationConfig.from_pretrained("google/bert_for_seq_generation_L-24_bbc_encoder")
             >>> config.is_decoder = True
-            >>> model = BertGenerationDecoder.from_pretrained('google/bert_for_seq_generation_L-24_bbc_encoder', config=config)
+            >>> model = BertGenerationDecoder.from_pretrained('google/bert_for_seq_generation_L-24_bbc_encoder', config=config, return_dict=True)
 
             >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
             >>> outputs = model(**inputs)

src/transformers/modeling_deberta.py

@@ -384,7 +384,7 @@ class DebertaEncoder(nn.Module):
         output_attentions=False,
         query_states=None,
         relative_pos=None,
-        return_dict=True,
+        return_dict=False,
     ):
         attention_mask = self.get_attention_mask(attention_mask)
         relative_pos = self.get_rel_pos(hidden_states, query_states, relative_pos)

src/transformers/modeling_distilbert.py

@@ -885,7 +885,7 @@ class DistilBertForMultipleChoice(DistilBertPreTrainedModel):
             >>> import torch
 
             >>> tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-cased')
-            >>> model = DistilBertForMultipleChoice.from_pretrained('distilbert-base-cased')
+            >>> model = DistilBertForMultipleChoice.from_pretrained('distilbert-base-cased', return_dict=True)
 
             >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
             >>> choice0 = "It is eaten with a fork and a knife."

src/transformers/modeling_dpr.py

@@ -455,7 +455,7 @@ class DPRContextEncoder(DPRPretrainedContextEncoder):
 
             >>> from transformers import DPRContextEncoder, DPRContextEncoderTokenizer
             >>> tokenizer = DPRContextEncoderTokenizer.from_pretrained('facebook/dpr-ctx_encoder-single-nq-base')
-            >>> model = DPRContextEncoder.from_pretrained('facebook/dpr-ctx_encoder-single-nq-base')
+            >>> model = DPRContextEncoder.from_pretrained('facebook/dpr-ctx_encoder-single-nq-base', return_dict=True)
             >>> input_ids = tokenizer("Hello, is my dog cute ?", return_tensors='pt')["input_ids"]
             >>> embeddings = model(input_ids).pooler_output
         """
@@ -533,7 +533,7 @@ class DPRQuestionEncoder(DPRPretrainedQuestionEncoder):
 
             >>> from transformers import DPRQuestionEncoder, DPRQuestionEncoderTokenizer
             >>> tokenizer = DPRQuestionEncoderTokenizer.from_pretrained('facebook/dpr-question_encoder-single-nq-base')
-            >>> model = DPRQuestionEncoder.from_pretrained('facebook/dpr-question_encoder-single-nq-base')
+            >>> model = DPRQuestionEncoder.from_pretrained('facebook/dpr-question_encoder-single-nq-base', return_dict=True)
             >>> input_ids = tokenizer("Hello, is my dog cute ?", return_tensors='pt')["input_ids"]
             >>> embeddings = model(input_ids).pooler_output
         """
@@ -609,7 +609,7 @@ class DPRReader(DPRPretrainedReader):
 
             >>> from transformers import DPRReader, DPRReaderTokenizer
             >>> tokenizer = DPRReaderTokenizer.from_pretrained('facebook/dpr-reader-single-nq-base')
-            >>> model = DPRReader.from_pretrained('facebook/dpr-reader-single-nq-base')
+            >>> model = DPRReader.from_pretrained('facebook/dpr-reader-single-nq-base', return_dict=True)
             >>> encoded_inputs = tokenizer(
             ...         questions=["What is love ?"],
             ...         titles=["Haddaway"],

src/transformers/modeling_electra.py

@@ -442,7 +442,7 @@ class ElectraEncoder(nn.Module):
         encoder_attention_mask=None,
         output_attentions=False,
         output_hidden_states=False,
-        return_dict=True,
+        return_dict=False,
     ):
         all_hidden_states = () if output_hidden_states else None
         all_self_attentions = () if output_attentions else None

src/transformers/modeling_encoder_decoder.py

@@ -370,7 +370,7 @@ class EncoderDecoderModel(PreTrainedModel):
             >>> outputs = model(input_ids=input_ids, decoder_input_ids=input_ids)
 
             >>> # training
-            >>> outputs = model(input_ids=input_ids, decoder_input_ids=input_ids, labels=input_ids)
+            >>> outputs = model(input_ids=input_ids, decoder_input_ids=input_ids, labels=input_ids, return_dict=True)
             >>> loss, logits = outputs.loss, outputs.logits
 
             >>> # save and load from pretrained

src/transformers/modeling_flax_bert.py

@@ -22,7 +22,7 @@ import jax
 import jax.numpy as jnp
 
 from .configuration_bert import BertConfig
-from .file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
+from .file_utils import add_start_docstrings
 from .modeling_flax_utils import FlaxPreTrainedModel, gelu
 from .utils import logging
 
@@ -35,20 +35,13 @@ _TOKENIZER_FOR_DOC = "BertTokenizer"
 
 BERT_START_DOCSTRING = r"""
 
-    This model inherits from :class:`~transformers.FlaxPreTrainedModel`. Check the superclass documentation for the
-    generic methods the library implements for all its model (such as downloading, saving and converting weights from
-    PyTorch models)
+    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
+    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
+    pruning heads etc.)
 
-    This model is also a Flax Linen `flax.nn.Module
-    <https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html>`__ subclass. Use it as a regular Flax
-    Module and refer to the Flax documentation for all matter related to general usage and behavior.
-
-    Finally, this model supports inherent JAX features such as:
-
-    - `Just-In-Time (JIT) compilation <https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit>`__
-    - `Automatic Differentiation <https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation>`__
-    - `Vectorization <https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap>`__
-    - `Parallelization <https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap>`__
+    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
+    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
+    general usage and behavior.
 
     Parameters:
         config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model.
@@ -59,32 +52,50 @@ BERT_START_DOCSTRING = r"""
 
 BERT_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`numpy.ndarray` of shape :obj:`({0})`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`):
             Indices of input sequence tokens in the vocabulary.
 
             Indices can be obtained using :class:`~transformers.BertTokenizer`. See
-            :meth:`transformers.PreTrainedTokenizer.encode` and :func:`transformers.PreTrainedTokenizer.__call__` for
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
             details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
-        attention_mask (:obj:`numpy.ndarray` of shape :obj:`({0})`, `optional`):
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({0})`, `optional`):
             Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
 
             - 1 for tokens that are **not masked**,
             - 0 for tokens that are **masked**.
 
             `What are attention masks? <../glossary.html#attention-mask>`__
-        token_type_ids (:obj:`numpy.ndarray` of shape :obj:`({0})`, `optional`):
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
             Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
             1]``:
 
             - 0 corresponds to a `sentence A` token,
             - 1 corresponds to a `sentence B` token.
 
-            `What are token type IDs? <../glossary.html#token-type-ids>`__
-        position_ids (:obj:`numpy.ndarray` of shape :obj:`({0})`, `optional`):
+            `What are token type IDs? <../glossary.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
             Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
             config.max_position_embeddings - 1]``.
+
+            `What are position IDs? <../glossary.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`({0}, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
         return_dict (:obj:`bool`, `optional`):
             Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
 """
@@ -280,7 +291,7 @@ class FlaxBertModule(nn.Module):
     intermediate_size: int
 
     @nn.compact
-    def __call__(self, input_ids, attention_mask, token_type_ids, position_ids):
+    def __call__(self, input_ids, token_type_ids, position_ids, attention_mask):
 
         # Embedding
         embeddings = FlaxBertEmbeddings(
@@ -399,8 +410,7 @@ class FlaxBertModel(FlaxPreTrainedModel):
     def module(self) -> nn.Module:
         return self._module
 
-    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
-    def __call__(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None):
+    def __call__(self, input_ids, token_type_ids=None, position_ids=None, attention_mask=None):
         if token_type_ids is None:
             token_type_ids = jnp.ones_like(input_ids)
 
@@ -413,7 +423,7 @@ class FlaxBertModel(FlaxPreTrainedModel):
         return self.model.apply(
             {"params": self.params},
             jnp.array(input_ids, dtype="i4"),
-            jnp.array(attention_mask, dtype="i4"),
             jnp.array(token_type_ids, dtype="i4"),
             jnp.array(position_ids, dtype="i4"),
+            jnp.array(attention_mask, dtype="i4"),
         )

src/transformers/modeling_flax_roberta.py

@@ -21,7 +21,7 @@ import jax
 import jax.numpy as jnp
 
 from .configuration_roberta import RobertaConfig
-from .file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
+from .file_utils import add_start_docstrings
 from .modeling_flax_utils import FlaxPreTrainedModel, gelu
 from .utils import logging
 
@@ -34,20 +34,13 @@ _TOKENIZER_FOR_DOC = "RobertaTokenizer"
 
 ROBERTA_START_DOCSTRING = r"""
 
-    This model inherits from :class:`~transformers.FlaxPreTrainedModel`. Check the superclass documentation for the
-    generic methods the library implements for all its model (such as downloading, saving and converting weights from
-    PyTorch models)
+    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
+    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
+    pruning heads etc.)
 
-    This model is also a Flax Linen `flax.nn.Module
-    <https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html>`__ subclass. Use it as a regular Flax
-    Module and refer to the Flax documentation for all matter related to general usage and behavior.
-
-    Finally, this model supports inherent JAX features such as:
-
-    - `Just-In-Time (JIT) compilation <https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit>`__
-    - `Automatic Differentiation <https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation>`__
-    - `Vectorization <https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap>`__
-    - `Parallelization <https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap>`__
+    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
+    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
+    general usage and behavior.
 
     Parameters:
         config (:class:`~transformers.RobertaConfig`): Model configuration class with all the parameters of the
@@ -58,32 +51,50 @@ ROBERTA_START_DOCSTRING = r"""
 
 ROBERTA_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`numpy.ndarray` of shape :obj:`({0})`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`):
             Indices of input sequence tokens in the vocabulary.
 
-            Indices can be obtained using :class:`~transformers.BertTokenizer`. See
-            :func:`transformers.PreTrainedTokenizer.encode` and :func:`transformers.PreTrainedTokenizer.__call__` for
+            Indices can be obtained using :class:`~transformers.RobertaTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
             details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
-        attention_mask (:obj:`numpy.ndarray` of shape :obj:`({0})`, `optional`):
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({0})`, `optional`):
             Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
 
             - 1 for tokens that are **not masked**,
             - 0 for tokens that are **masked**.
 
             `What are attention masks? <../glossary.html#attention-mask>`__
-        token_type_ids (:obj:`numpy.ndarray` of shape :obj:`({0})`, `optional`):
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
             Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
             1]``:
 
             - 0 corresponds to a `sentence A` token,
             - 1 corresponds to a `sentence B` token.
 
-            `What are token type IDs? <../glossary.html#token-type-ids>`__
-        position_ids (:obj:`numpy.ndarray` of shape :obj:`({0})`, `optional`):
+            `What are token type IDs? <../glossary.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
             Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
             config.max_position_embeddings - 1]``.
+
+            `What are position IDs? <../glossary.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`({0}, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
         return_dict (:obj:`bool`, `optional`):
             Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
 """
@@ -291,7 +302,7 @@ class FlaxRobertaModule(nn.Module):
     intermediate_size: int
 
     @nn.compact
-    def __call__(self, input_ids, attention_mask, token_type_ids, position_ids):
+    def __call__(self, input_ids, token_type_ids, position_ids, attention_mask):
 
         # Embedding
         embeddings = FlaxRobertaEmbeddings(
@@ -410,8 +421,7 @@ class FlaxRobertaModel(FlaxPreTrainedModel):
     def module(self) -> nn.Module:
         return self._module
 
-    @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
-    def __call__(self, input_ids, token_type_ids=None, attention_mask=None, position_ids=None):
+    def __call__(self, input_ids, token_type_ids=None, position_ids=None, attention_mask=None):
         if token_type_ids is None:
             token_type_ids = jnp.ones_like(input_ids)
 
@@ -426,7 +436,7 @@ class FlaxRobertaModel(FlaxPreTrainedModel):
         return self.model.apply(
             {"params": self.params},
             jnp.array(input_ids, dtype="i4"),
-            jnp.array(attention_mask, dtype="i4"),
             jnp.array(token_type_ids, dtype="i4"),
             jnp.array(position_ids, dtype="i4"),
+            jnp.array(attention_mask, dtype="i4"),
         )

src/transformers/modeling_fsmt.py

@@ -434,7 +434,7 @@ class FSMTEncoder(nn.Module):
         )  # type: List[EncoderLayer]
 
     def forward(
-        self, input_ids, attention_mask=None, output_attentions=False, output_hidden_states=False, return_dict=True
+        self, input_ids, attention_mask=None, output_attentions=False, output_hidden_states=False, return_dict=False
     ):
         """
         Args:
@@ -617,7 +617,7 @@ class FSMTDecoder(nn.Module):
         use_cache=False,
         output_attentions=False,
         output_hidden_states=False,
-        return_dict=True,
+        return_dict=False,
         **unused,
     ):
         """

src/transformers/modeling_funnel.py

@@ -619,7 +619,7 @@ class FunnelEncoder(nn.Module):
         token_type_ids=None,
         output_attentions=False,
         output_hidden_states=False,
-        return_dict=True,
+        return_dict=False,
     ):
         # The pooling is not implemented on long tensors, so we convert this mask.
         attention_mask = attention_mask.type_as(inputs_embeds)
@@ -698,7 +698,7 @@ class FunnelDecoder(nn.Module):
         token_type_ids=None,
         output_attentions=False,
         output_hidden_states=False,
-        return_dict=True,
+        return_dict=False,
     ):
         upsampled_hidden = upsample(
             final_hidden,
@@ -1111,7 +1111,7 @@ class FunnelForPreTraining(FunnelPreTrainedModel):
             >>> import torch
 
             >>> tokenizer = FunnelTokenizer.from_pretrained('funnel-transformer/small')
-            >>> model = FunnelForPreTraining.from_pretrained('funnel-transformer/small')
+            >>> model = FunnelForPreTraining.from_pretrained('funnel-transformer/small', return_dict=True)
 
             >>> inputs = tokenizer("Hello, my dog is cute", return_tensors= "pt")
             >>> logits = model(**inputs).logits

src/transformers/modeling_gpt2.py

@@ -708,12 +708,9 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         return self.lm_head
 
     def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
-        token_type_ids = kwargs.get("token_type_ids", None)
         # only last token for inputs_ids if past is defined in kwargs
         if past:
             input_ids = input_ids[:, -1].unsqueeze(-1)
-            if token_type_ids is not None:
-                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
 
         attention_mask = kwargs.get("attention_mask", None)
         position_ids = kwargs.get("position_ids", None)
@@ -732,7 +729,6 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
             "use_cache": kwargs.get("use_cache"),
             "position_ids": position_ids,
             "attention_mask": attention_mask,
-            "token_type_ids": token_type_ids,
         }
 
     @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
@@ -840,32 +836,14 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         return self.lm_head
 
     def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
-        token_type_ids = kwargs.get("token_type_ids", None)
         # only last token for inputs_ids if past is defined in kwargs
         if past:
             input_ids = input_ids[:, -1].unsqueeze(-1)
-            if token_type_ids is not None:
-                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
-
-        attention_mask = kwargs.get("attention_mask", None)
-        position_ids = kwargs.get("position_ids", None)
-
-        if attention_mask is not None and position_ids is None:
-            # create position_ids on the fly for batch generation
-            position_ids = attention_mask.long().cumsum(-1) - 1
-            position_ids.masked_fill_(attention_mask == 0, 1)
-            if past:
-                position_ids = position_ids[:, -1].unsqueeze(-1)
-        else:
-            position_ids = None
 
         return {
             "input_ids": input_ids,
             "past_key_values": past,
             "use_cache": kwargs.get("use_cache"),
-            "position_ids": position_ids,
-            "attention_mask": attention_mask,
-            "token_type_ids": token_type_ids,
         }
 
     @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
@@ -911,7 +889,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             >>> from transformers import GPT2Tokenizer, GPT2DoubleHeadsModel
 
             >>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
-            >>> model = GPT2DoubleHeadsModel.from_pretrained('gpt2')
+            >>> model = GPT2DoubleHeadsModel.from_pretrained('gpt2, return_dict=True)
 
             >>> # Add a [CLS] to the vocabulary (we should train it also!)
             >>> num_added_tokens = tokenizer.add_special_tokens({'cls_token': '[CLS]'})

src/transformers/modeling_layoutlm.py

@@ -376,7 +376,7 @@ class LayoutLMEncoder(nn.Module):
         encoder_attention_mask=None,
         output_attentions=False,
         output_hidden_states=False,
-        return_dict=True,
+        return_dict=False,
     ):
         all_hidden_states = () if output_hidden_states else None
         all_self_attentions = () if output_attentions else None