fix tf cuda

.circleci/config.yml

@@ -73,7 +73,7 @@ jobs:
         steps:
             - checkout
             - run: sudo pip install .[tf,torch,docs]
-            - run: cd docs && make html SPHINXOPTS="-W"
+            - run: cd docs && make html
             - store_artifacts:
                 path: ./docs/_build
     deploy_doc:

.circleci/deploy.sh

@@ -17,8 +17,7 @@ function deploy_doc(){
 	fi
 }
 
-# You can find the commit for each tag on https://github.com/huggingface/transformers/tags
-deploy_doc "master" master
+deploy_doc "master"
 deploy_doc "b33a385" v1.0.0
 deploy_doc "fe02e45" v1.1.0
 deploy_doc "89fd345" v1.2.0
@@ -28,11 +27,3 @@ deploy_doc "3616209" v2.2.0
 deploy_doc "d0f8b9a" v2.3.0
 deploy_doc "6664ea9" v2.4.0
 deploy_doc "fb560dc" v2.5.0
-deploy_doc "b90745c" v2.5.1
-deploy_doc "fbc5bf1" v2.6.0
-deploy_doc "6f5a12a" v2.7.0
-deploy_doc "11c3257" v2.8.0
-deploy_doc "e7cfc1a" v2.9.0
-deploy_doc "7cb203f" v2.9.1
-deploy_doc "10d7239" v2.10.0 
-deploy_doc "b42586e" #v2.11.0 Latest stable release

README.md

@@ -59,7 +59,7 @@ Choose the right framework for every part of a model's lifetime
 | [Quick tour: Share your models ](#Quick-tour-of-model-sharing) | Upload and share your fine-tuned models with the community |
 | [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-transformers to transformers |
 | [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
-| Documentation [(master)](https://huggingface.co/transformers/master) [(stable)](https://huggingface.co/transformers/) [(v2.10.0)](https://huggingface.co/transformers/v2.10.0) [(v2.9.0/v2.9.1)](https://huggingface.co/transformers/v2.9.1) [(v2.8.0)](https://huggingface.co/transformers/v2.8.0) [(v2.7.0)](https://huggingface.co/transformers/v2.7.0) [(v2.6.0)](https://huggingface.co/transformers/v2.6.0) [(v2.5.0/v2.5.1)](https://huggingface.co/transformers/v2.5.1) [(v2.4.0/v2.4.1)](https://huggingface.co/transformers/v2.4.0)[(v2.3.0)](https://huggingface.co/transformers/v2.3.0)[(v2.2.0/v2.2.1/v2.2.2)](https://huggingface.co/transformers/v2.2.0) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) | Full API documentation and more |
+| [Documentation][(v2.5.0)](https://huggingface.co/transformers/v2.5.0)[(v2.4.0/v2.4.1)](https://huggingface.co/transformers/v2.4.0)[(v2.3.0)](https://huggingface.co/transformers/v2.3.0)[(v2.2.0/v2.2.1/v2.2.2)](https://huggingface.co/transformers/v2.2.0) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) [(master)](https://huggingface.co/transformers) | Full API documentation and more |
 
 ## Installation
 
@@ -169,7 +169,7 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 22. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
 23. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
 
-These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Pearson R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
+These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Peason R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
 
 ## Online demo
 

docs/source/glossary.rst

@@ -1,41 +1,11 @@
 Glossary
-^^^^^^^^
-
-General terms
--------------
-
-- autoencoding models: see MLM
-- autoregressive models: see CLM
-- CLM: causal language modeling, a pretraining task where the model reads the texts in order and has to predict the
-  next word. It's usually done by reading the whole sentence but using a mask inside the model to hide the future 
-  tokens at a certain timestep.
-- MLM: masked language modeling, a pretraining task where the model sees a corrupted version of the texts, usually done
-  by masking some tokens randomly, and has to predict the original text.
-- multimodal: a task taht combines texts with another kind of inputs (for instance images).
-- NLG: natural language generation, all tasks related to generating text ( for instance talk with transformers,
-  translation)
-- NLP: natural language processing, a generic way to say "deal with texts".
-- NLU: natural language understanding, all tasks related to understanding what is in a text (for instance classifying
-  the whole text, individual words)
-- pretrained model: a model that has been pretrained on some data (for instance all of Wikipedia). Pretraining methods
-  involve a self-supervised objective, which can be reading the text and trying to predict the next word (see CLM) or 
-  masking some words and trying to predict them (see MLM).
-- RNN: recurrent neural network, a type of model that uses a loop over a layer to process texts.
-- seq2seq or sequence-to-sequence: models that generate a new sequence from an input, like translation models, or
-  summarization models (such as :doc:`Bart </model_doc/bart>` or :doc:`T5 </model_doc/t5>`).
-- token: a part of a sentence, usually a word, but can also be a subword (non-common words are often split in subwords)
-  or a punctuation symbol.
-
-Model inputs
-------------
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Every model is different yet bears similarities with the others. Therefore most models use the same inputs, which are
 detailed here alongside usage examples.
 
-.. _input-ids:
-
 Input IDs
-~~~~~~~~~
+--------------------------
 
 The input ids are often the only required parameters to be passed to the model as input. *They are token indices,
 numerical representations of tokens building the sequences that will be used as input by the model*.
@@ -54,52 +24,24 @@ The tokenizer takes care of splitting the sequence into tokens available in the
 
 ::
 
+    # Continuation of the previous script
     tokenized_sequence = tokenizer.tokenize(sequence)
-    print(tokenized_sequence)
+    assert tokenized_sequence == ['A', 'Titan', 'R', '##T', '##X', 'has', '24', '##GB', 'of', 'V', '##RA', '##M']
 
-The tokens are either words or subwords. Here for instance, "VRAM" wasn't in the model vocabulary, so it's been split
-in "V", "RA" and "M". To indicate those tokens are not separate words but parts of the same word, a double-dash is
-added for "RA" and "M":
-
-::
-
-    ['A', 'Titan', 'R', '##T', '##X', 'has', '24', '##GB', 'of', 'V', '##RA', '##M']
-
-These tokens can then be converted into IDs which are understandable by the model. This can be done by directly feeding
-the sentence to the tokenizer, which leverages the Rust implementation of
+These tokens can then be converted into IDs which are understandable by the model. Several methods are available for
+this, the recommended being `encode` or `encode_plus`, which leverage the Rust implementation of
 `huggingface/tokenizers <https://github.com/huggingface/tokenizers>`__ for peak performance.
 
 ::
 
-    encoded_sequence = tokenizer(sequence)["input_ids"]
-    print(encoded_sequence)
+    # Continuation of the previous script
+    encoded_sequence = tokenizer.encode(sequence)
+    assert encoded_sequence == [101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102]
 
-The tokenizer returns a dictionary with all the arguments necessary for its corresponding model to work properly. The
-token indices are under the key "input_ids":
-
-::
-
-    [101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102]
-
-Note that the tokenizer automatically adds "special tokens" (if the associated model rely on them) which are special
-IDs the model sometimes uses. If we decode the previous sequence of ids,
-
-::
-
-    tokenizer.decode(encoded_sequence)
-
-we will see 
-
-::
-
-    '[CLS] A Titan RTX has 24GB of VRAM [SEP]'
-
-because this is the way a :class:`~transformers.BertModel` is going to expect its inputs.
-
-.. _attention-mask:
+The `encode` and `encode_plus` methods automatically add "special tokens" which are special IDs the model uses.
 
 Attention mask
-~~~~~~~~~~~~~~
+--------------------------
 
 The attention mask is an optional argument used when batching sequences together. This argument indicates to the
 model which tokens should be attended to, and which should not.
@@ -114,107 +56,83 @@ For example, consider these two sequences:
     sequence_a = "This is a short sequence."
     sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."
 
-    encoded_sequence_a = tokenizer(sequence_a)["input_ids"]
-    encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
-    
-    len(encoded_sequence_a), len(encoded_sequence_b)
+    encoded_sequence_a = tokenizer.encode(sequence_a)
+    assert len(encoded_sequence_a) == 8
 
-The encoded versions have different lengths:
+    encoded_sequence_b = tokenizer.encode(sequence_b)
+    assert len(encoded_sequence_b) == 19
+
+These two sequences have different lengths and therefore can't be put together in a same tensor as-is. The first
+sequence needs to be padded up to the length of the second one, or the second one needs to be truncated down to
+the length of the first one.
+
+In the first case, the list of IDs will be extended by the padding indices:
 
 ::
 
-    (8, 19)
+    # Continuation of the previous script
+    padded_sequence_a = tokenizer.encode(sequence_a, max_length=19, pad_to_max_length=True)
 
-Therefore, we can't be put then together in a same tensor as-is. The first sequence needs to be padded up to the length
-of the second one, or the second one needs to be truncated down to the length of the first one.
+    assert padded_sequence_a == [101, 1188, 1110, 170, 1603, 4954,  119, 102,    0,    0,    0,    0,    0,    0,    0,    0,   0,   0,   0]
+    assert encoded_sequence_b == [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]
 
-In the first case, the list of IDs will be extended by the padding indices. We can pass a list to the tokenizer and ask
-it to pad like this:
-
-::
-
-    padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
-    padded_sequences["input_ids"]
-
-We can see that 0s have been added on the right of the first sentence to make it the same length as the second one:
-
-::
-
-    [[101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
-     [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]]
-
-This can then be converted into a tensor in PyTorch or TensorFlow. The attention mask is a binary tensor indicating
+These can then be converted into a tensor in PyTorch or TensorFlow. The attention mask is a binary tensor indicating
 the position of the padded indices so that the model does not attend to them. For the
 :class:`~transformers.BertTokenizer`, :obj:`1` indicate a value that should be attended to while :obj:`0` indicate
-a padded value. This attention mask is in the dictionary returned by the tokenizer under the key "attention_mask":
+a padded value.
+
+The method :func:`~transformers.PreTrainedTokenizer.encode_plus` may be used to obtain the attention mask directly:
 
 ::
 
-    padded_sequences["attention_mask"]
+    # Continuation of the previous script
+    sequence_a_dict = tokenizer.encode_plus(sequence_a, max_length=19, pad_to_max_length=True)
 
-will give back
+    assert sequence_a_dict['input_ids'] == [101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
+    assert sequence_a_dict['attention_mask'] == [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
 
-::
-
-    [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
-     [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]
-
-.. _token-type-ids:
 
 Token Type IDs
-~~~~~~~~~~~~~~
+--------------------------
 
 Some models' purpose is to do sequence classification or question answering. These require two different sequences to
 be encoded in the same input IDs. They are usually separated by special tokens, such as the classifier and separator
 tokens. For example, the BERT model builds its two sequence input as such:
 
-::
-
-   # [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]
-
-We can use our tokenizer to automatically generate such a sentence by passing the two sequences as two arguments (and
-not a list like before) like this:
-
 ::
 
     from transformers import BertTokenizer
     tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+
+    # [CLS] SEQ_A [SEP] SEQ_B [SEP]
+
     sequence_a = "HuggingFace is based in NYC"
     sequence_b = "Where is HuggingFace based?"
 
-    encoded_dict = tokenizer(sequence_a, sequence_b)
-    tokenizer.decode(encoded_dict["input_ids"])
-
-which will return:
-
-::
-
-    "[CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]"
+    encoded_sequence = tokenizer.encode(sequence_a, sequence_b)
+    assert tokenizer.decode(encoded_sequence) == "[CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]"
 
 This is enough for some models to understand where one sequence ends and where another begins. However, other models
-such as BERT have an additional mechanism, which are the token type IDs (also called segment IDs). They are a binary
-mask identifying the different sequences in the model.
+such as BERT have an additional mechanism, which are the segment IDs. The Token Type IDs are a binary mask identifying
+the different sequences in the model.
 
-The tokenizer returns in the dictionary under the key "token_type_ids":
+We can leverage :func:`~transformers.PreTrainedTokenizer.encode_plus` to output the Token Type IDs for us:
 
 ::
 
-    encoded_dict['token_type_ids']
+    # Continuation of the previous script
+    encoded_dict = tokenizer.encode_plus(sequence_a, sequence_b)
 
-will return
-
-::
-
-    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]
+    assert encoded_dict['input_ids'] == [101, 20164, 10932, 2271, 7954, 1110, 1359, 1107, 17520, 102, 2777, 1110, 20164, 10932, 2271, 7954, 1359, 136, 102]
+    assert encoded_dict['token_type_ids'] == [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]
 
 The first sequence, the "context" used for the question, has all its tokens represented by :obj:`0`, whereas the
 question has all its tokens represented by :obj:`1`. Some models, like :class:`~transformers.XLNetModel` use an
 additional token represented by a :obj:`2`.
 
-.. _position-ids:
 
 Position IDs
-~~~~~~~~~~~~
+--------------------------
 
 The position IDs are used by the model to identify which token is at which position. Contrary to RNNs that have the
 position of each token embedded within them, transformers are unaware of the position of each token. The position
@@ -226,24 +144,13 @@ positional embeddings.
 Absolute positional embeddings are selected in the range ``[0, config.max_position_embeddings - 1]``. Some models
 use other types of positional embeddings, such as sinusoidal position embeddings or relative position embeddings.
 
-.. _feed-forward-chunking:
 
 Feed Forward Chunking
-~~~~~~~~~~~~~~~~~~~~~
+--------------------------
 
-In transformers two feed forward layers usually follows the self attention layer in each residual attention block.
-The intermediate embedding size of the feed forward layers is often bigger than the hidden size of the model (e.g.,
-for ``bert-base-uncased``). 
+In transformers two feed forward layers usually follows the self attention layer in each residual attention block. The intermediate embedding size of the feed forward layers is often bigger than the hidden size of the model (*e.g.* for ``bert-base-uncased``). 
 
-For an input of size ``[batch_size, sequence_length]``, the memory required to store the intermediate feed forward
-embeddings ``[batch_size, sequence_length, config.intermediate_size]`` can account for a large fraction of the memory
-use. The authors of `Reformer: The Efficient Transformer <https://arxiv.org/abs/2001.04451>`_ noticed that since the
-computation is independent of the ``sequence_length`` dimension, it is mathematically equivalent to compute the output
-embeddings of both feed forward layers ``[batch_size, config.hidden_size]_0, ..., [batch_size, config.hidden_size]_n`` 
-individually and concat them afterward to ``[batch_size, sequence_length, config.hidden_size]`` with
-``n = sequence_length``, which trades increased computation time against reduced memory use, but yields a
-mathematically **equivalent** result.
+For an input of size ``[batch_size, sequence_length]``, the memory required to store the intermediate feed forward embeddings ``[batch_size, sequence_length, config.intermediate_size]`` can account for a large fraction of the memory use. The authors of `Reformer: The Efficient Transformer <https://arxiv.org/abs/2001.04451>`_ noticed that since the computation is independent of the ``sequence_length`` dimension, it is mathematically equivalent to compute the output embeddings of both feed forward layers ``[batch_size, config.hidden_size]_0, ..., [batch_size, config.hidden_size]_n``  individually and concat them afterward to ``[batch_size, sequence_length, config.hidden_size]`` with ``n = sequence_length``, which trades increased computation time against reduced memory use, but yields a mathematically **equivalent** result.
 
-For models employing the function :func:`~.transformers.apply_chunking_to_forward`, the ``chunk_size`` defines the
-number of output embeddings that are computed in parallel and thus defines the trade-off between memory and time
-complexity.  If ``chunk_size`` is set to 0, no feed forward chunking is done.
+For models employing the function :func:`~.transformers.apply_chunking_to_forward`, the ``chunk_size`` defines the number of output embeddings that are computed in parallel and thus defines the trade-off between memory and time complexity. 
+If ``chunk_size`` is set to 0, no feed forward chunking is done.

docs/source/index.rst

@@ -38,16 +38,6 @@ Choose the right framework for every part of a model's lifetime:
 Contents
 ---------------------------------
 
-The documentation is organized in five parts:
-
-- **GET STARTED** contains a quick tour, the installation instructions and some useful information about our philosophy
-  and a glossary.
-- **USING TRANSFORMERS** contains general tutorials on how to use the library.
-- **ADVANCED GUIDES** contains more advanced guides that are more specific to a given script or part of the library.
-- **RESEARCH** focuses on tutorials that have less to do with how to use the library but more about general resarch in
-  transformers model
-- **PACKAGE REFERENCE** contains the documentation of each public class and function.
-
 The library currently contains PyTorch and Tensorflow implementations, pre-trained model weights, usage scripts and
 conversion utilities for the following models:
 
@@ -128,17 +118,16 @@ conversion utilities for the following models:
     :maxdepth: 2
     :caption: Get started
 
-    quicktour
     installation
-    philosophy
+    quickstart
     glossary
 
 .. toctree::
     :maxdepth: 2
     :caption: Using Transformers
 
-    task_summary
-    model_summary
+    usage
+    summary
     serialization
     model_sharing
     multilingual
@@ -195,4 +184,3 @@ conversion utilities for the following models:
     model_doc/marian
     model_doc/longformer
     model_doc/retribert
-    model_doc/mobilebert

docs/source/main_classes/pipelines.rst

@@ -17,7 +17,7 @@ The pipeline abstraction
 The `pipeline` abstraction is a wrapper around all the other available pipelines. It is instantiated as any
 other pipeline but requires an additional argument which is the `task`.
 
-.. autofunction:: transformers.pipeline
+... autofunction:: transformers.pipeline
 
 
 The task specific pipelines

docs/source/model_doc/auto.rst

@@ -1,15 +1,11 @@
 AutoModels
 -----------
 
-In many cases, the architecture you want to use can be guessed from the name or the path of the pretrained model you
-are supplying to the ``from_pretrained`` method.
+In many cases, the architecture you want to use can be guessed from the name or the path of the pretrained model you are supplying to the ``from_pretrained`` method.
 
-AutoClasses are here to do this job for you so that you automatically retrieve the relevant model given the name/path
-to the pretrained weights/config/vocabulary:
+AutoClasses are here to do this job for you so that you automatically retrieve the relevant model given the name/path to the pretrained weights/config/vocabulary:
 
-Instantiating one of ``AutoModel``, ``AutoConfig`` and ``AutoTokenizer`` will directly create a class of the relevant
-architecture (ex: ``model = AutoModel.from_pretrained('bert-base-cased')`` will create a instance of 
-:class:`~transformers.BertModel`).
+Instantiating one of ``AutoModel``, ``AutoConfig`` and ``AutoTokenizer`` will directly create a class of the relevant architecture (ex: ``model = AutoModel.from_pretrained('bert-base-cased')`` will create a instance of ``BertModel``).
 
 
 ``AutoConfig``
@@ -67,43 +63,3 @@ architecture (ex: ``model = AutoModel.from_pretrained('bert-base-cased')`` will
 .. autoclass:: transformers.AutoModelForTokenClassification
     :members:
 
-``TFAutoModel``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFAutoModel
-    :members:
-
-
-``TFAutoModelForPreTraining``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFAutoModelForPreTraining
-    :members:
-
-
-``TFAutoModelWithLMHead``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFAutoModelWithLMHead
-    :members:
-
-
-``TFAutoModelForSequenceClassification``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFAutoModelForSequenceClassification
-    :members:
-
-
-``TFAutoModelForQuestionAnswering``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFAutoModelForQuestionAnswering
-    :members:
-
-
-``TFAutoModelForTokenClassification``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFAutoModelForTokenClassification
-    :members:

docs/source/model_doc/mobilebert.rst

@@ -1,169 +0,0 @@
-MobileBERT
-----------------------------------------------------
-
-Overview
-~~~~~~~~~~~~~~~~~~~~~
-
-The MobileBERT model was proposed in `MobileBERT: a Compact Task-Agnostic BERT
-for Resource-Limited Devices <https://arxiv.org/abs/2004.02984>`__
-by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. It's a bidirectional transformer
-based on the BERT model, which is compressed and accelerated using several approaches.
-
-The abstract from the paper is the following:
-
-*Natural Language Processing (NLP) has recently achieved great success by using huge pre-trained models with hundreds
-of millions of parameters. However, these models suffer from heavy model sizes and high latency such that they cannot
-be deployed to resource-limited mobile devices. In this paper, we propose MobileBERT for compressing and accelerating
-the popular BERT model. Like the original BERT, MobileBERT is task-agnostic, that is, it can be generically applied
-to various downstream NLP tasks via simple fine-tuning. Basically, MobileBERT is a thin version of BERT_LARGE, while
-equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward
-networks. To train MobileBERT, we first train a specially designed teacher model, an inverted-bottleneck incorporated
-BERT_LARGE model. Then, we conduct knowledge transfer from this teacher to MobileBERT. Empirical studies show that
-MobileBERT is 4.3x smaller and 5.5x faster than BERT_BASE while achieving competitive results on well-known
-benchmarks. On the natural language inference tasks of GLUE, MobileBERT achieves a GLUEscore o 77.7
-(0.6 lower than BERT_BASE), and 62 ms latency on a Pixel 4 phone. On the SQuAD v1.1/v2.0 question answering task,
-MobileBERT achieves a dev F1 score of 90.0/79.2 (1.5/2.1 higher than BERT_BASE).*
-
-Tips:
-
-- MobileBERT is a model with absolute position embeddings so it's usually advised to pad the inputs on
-  the right rather than the left.
-- MobileBERT is similar to BERT and therefore relies on the masked language modeling (MLM) objective.
-  It is therefore efficient at predicting masked tokens and at NLU in general, but is not optimal for
-  text generation. Models trained with a causal language modeling (CLM) objective are better in that regard.
-
-The original code can be found `here <https://github.com/google-research/mobilebert>`_.
-
-MobileBertConfig
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertConfig
-    :members:
-
-
-MobileBertTokenizer
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertTokenizer
-    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
-        create_token_type_ids_from_sequences, save_vocabulary
-
-
-MobileBertTokenizerFast
-~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertTokenizerFast
-    :members:
-
-
-MobileBertModel
-~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertModel
-    :members:
-
-
-MobileBertForPreTraining
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertForPreTraining
-    :members:
-
-
-MobileBertForMaskedLM
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertForMaskedLM
-    :members:
-
-
-MobileBertForNextSentencePrediction
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertForNextSentencePrediction
-    :members:
-
-
-MobileBertForSequenceClassification
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertForSequenceClassification
-    :members:
-
-
-MobileBertForMultipleChoice
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertForMultipleChoice
-    :members:
-
-
-MobileBertForTokenClassification
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertForTokenClassification
-    :members:
-
-
-MobileBertForQuestionAnswering
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MobileBertForQuestionAnswering
-    :members:
-
-
-TFMobileBertModel
-~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFMobileBertModel
-    :members:
-
-
-TFMobileBertForPreTraining
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFMobileBertForPreTraining
-    :members:
-
-
-TFMobileBertForMaskedLM
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFMobileBertForMaskedLM
-    :members:
-
-
-TFMobileBertForNextSentencePrediction
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFMobileBertForNextSentencePrediction
-    :members:
-
-
-TFMobileBertForSequenceClassification
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFMobileBertForSequenceClassification
-    :members:
-
-
-TFMobileBertForMultipleChoice
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFMobileBertForMultipleChoice
-    :members:
-
-
-TFMobileBertForTokenClassification
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFMobileBertForTokenClassification
-    :members:
-
-
-TFMobileBertForQuestionAnswering
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFMobileBertForQuestionAnswering
-    :members:
-

docs/source/philosophy.rst

@@ -1,73 +0,0 @@
-Philosophy
-==========
-
-Transformers is an opinionated library built for:
-
-- NLP researchers and educators seeking to use/study/extend large-scale transformers models
-- hands-on practitioners who want to fine-tune those models and/or serve them in production
-- engineers who just want to download a pretrained model and use it to solve a given NLP task.
-
-The library was designed with two strong goals in mind:
-
-- Be as easy and fast to use as possible:
-
-    - We strongly limited the number of user-facing abstractions to learn, in fact, there are almost no abstractions,
-      just three standard classes required to use each model: :doc:`configuration <main_classes/configuration>`, 
-      :doc:`models <main_classes/model>` and :doc:`tokenizer <main_classes/tokenizer>`.
-    - All of these classes can be initialized in a simple and unified way from pretrained instances by using a common
-      :obj:`from_pretrained()` instantiation method which will take care of downloading (if needed), caching and
-      loading the related class instance and associated data (configurations' hyper-parameters, tokenizers' vocabulary, 
-      and models' weights) from a pretrained checkpoint provided on 
-      `Hugging Face Hub <https://huggingface.co/models>`__ or your own saved checkpoint.
-    - On top of those three base classes, the library provides two APIs: :func:`~transformers.pipeline` for quickly
-      using a model (plus its associated tokenizer and configuration) on a given task and 
-      :func:`~transformers.Trainer`/:func:`~transformers.TFTrainer` to quickly train or fine-tune a given model.
-    - As a consequence, this library is NOT a modular toolbox of building blocks for neural nets. If you want to
-      extend/build-upon the library, just use regular Python/PyTorch/TensorFlow/Keras modules and inherit from the base
-      classes of the library to reuse functionalities like model loading/saving.
-
-- Provide state-of-the-art models with performances as close as possible to the original models:
-
-    - We provide at least one example for each architecture which reproduces a result provided by the official authors
-      of said architecture.
-    - The code is usually as close to the original code base as possible which means some PyTorch code may be not as
-      *pytorchic* as it could be as a result of being converted TensorFlow code and vice versa.
-
-A few other goals:
-
-- Expose the models' internals as consistently as possible:
-
-    - We give access, using a single API, to the full hidden-states and attention weights.
-    - Tokenizer and base model's API are standardized to easily switch between models.
-
-- Incorporate a subjective selection of promising tools for fine-tuning/investigating these models:
-
-    - A simple/consistent way to add new tokens to the vocabulary and embeddings for fine-tuning.
-    - Simple ways to mask and prune transformer heads.
-
-- Switch easily between PyTorch and TensorFlow 2.0, allowing training using one framwork and inference using another.
-
-Main concepts
-~~~~~~~~~~~~~
-
-The library is build around three types of classes for each model:
-
-- **Model classes**  such as :class:`~transformers.BertModel`, which are 30+ PyTorch models 
-  (`torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__) or Keras models 
-  (`tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__) that work with the pretrained
-  weights provided in the library.
-- **Configuration classes** such as :class:`~transformers.BertConfig`, which store all the parameters required to build
-  a model. You don't always need to instantiate these yourself. In particular, if you are using a pretrained model
-  without any modification, creating the model will automatically take care of instantiating the configuration (which
-  is part of the model).
-- **Tokenizer classes** such as :class:`~transformers.BertTokenizer`, which store the vocabulary for each model and
-  provide methods for encoding/decoding strings in a list of token embeddings indices to be fed to a model.
-
-All these classes can be instantiated from pretrained instances and saved locally using two methods:
-
-- :obj:`from_pretrained()` let you instantiate a model/configuration/tokenizer from a pretrained version either
-  provided by the library itself (the suported models are provided in the list :doc:`here <pretrained_models>`
-  or stored locally (or on a server) by the user,
-- :obj:`save_pretrained()` let you save a model/configuration/tokenizer locally so that it can be reloaded using
-  :obj:`from_pretrained()`.
-

docs/source/quickstart.md

@@ -0,0 +1,222 @@
+# Quickstart
+
+## Philosophy
+
+Transformers is an opinionated library built for NLP researchers seeking to use/study/extend large-scale transformers models.
+
+The library was designed with two strong goals in mind:
+
+- be as easy and fast to use as possible:
+
+  - we strongly limited the number of user-facing abstractions to learn, in fact, there are almost no abstractions, just three standard classes required to use each model: configuration, models and tokenizer,
+  - all of these classes can be initialized in a simple and unified way from pretrained instances by using a common `from_pretrained()` instantiation method which will take care of downloading (if needed), caching and loading the related class from a pretrained instance supplied in the library or your own saved instance.
+  - as a consequence, this library is NOT a modular toolbox of building blocks for neural nets. If you want to extend/build-upon the library, just use regular Python/PyTorch modules and inherit from the base classes of the library to reuse functionalities like model loading/saving.
+
+- provide state-of-the-art models with performances as close as possible to the original models:
+
+  - we provide at least one example for each architecture which reproduces a result provided by the official authors of said architecture,
+  - the code is usually as close to the original code base as possible which means some PyTorch code may be not as *pytorchic* as it could be as a result of being converted TensorFlow code.
+
+A few other goals:
+
+- expose the models' internals as consistently as possible:
+
+  - we give access, using a single API to the full hidden-states and attention weights,
+  - tokenizer and base model's API are standardized to easily switch between models.
+
+- incorporate a subjective selection of promising tools for fine-tuning/investigating these models:
+
+  - a simple/consistent way to add new tokens to the vocabulary and embeddings for fine-tuning,
+  - simple ways to mask and prune transformer heads.
+
+## Main concepts
+
+The library is build around three types of classes for each model:
+
+- **model classes**  e.g., `BertModel` which are 20+ PyTorch models (`torch.nn.Modules`) that work with the pretrained weights provided in the library. In TF2, these are `tf.keras.Model`.
+- **configuration classes** which store all the parameters required to build a model, e.g., `BertConfig`. You don't always need to instantiate these your-self. In particular, if you are using a pretrained model without any modification, creating the model will automatically take care of instantiating the configuration (which is part of the model)
+- **tokenizer classes** which store the vocabulary for each model and provide methods for encoding/decoding strings in a list of token embeddings indices to be fed to a model, e.g., `BertTokenizer`
+
+All these classes can be instantiated from pretrained instances and saved locally using two methods:
+
+- `from_pretrained()` let you instantiate a model/configuration/tokenizer from a pretrained version either provided by the library itself (currently 27 models are provided as listed [here](https://huggingface.co/transformers/pretrained_models.html)) or stored locally (or on a server) by the user,
+- `save_pretrained()` let you save a model/configuration/tokenizer locally so that it can be reloaded using `from_pretrained()`.
+
+We'll finish this quickstart tour by going through a few simple quick-start examples to see how we can instantiate and use these classes. The rest of the documentation is organized into two parts:
+
+- the **MAIN CLASSES** section details the common functionalities/method/attributes of the three main type of classes (configuration, model, tokenizer) plus some optimization related classes provided as utilities for training,
+- the **PACKAGE REFERENCE** section details all the variants of each class for each model architectures and, in particular, the input/output that you should expect when calling each of them.
+
+## Quick tour: Usage
+
+Here are two examples showcasing a few `Bert` and `GPT2` classes and pre-trained models.
+
+See the full API reference for examples of each model class.
+
+### BERT example
+
+Let's start by preparing a tokenized input (a list of token embeddings indices to be fed to Bert) from a text string using `BertTokenizer`
+
+```python
+import torch
+from transformers import BertTokenizer, BertModel, BertForMaskedLM
+
+# OPTIONAL: if you want to have more information on what's happening under the hood, activate the logger as follows
+import logging
+logging.basicConfig(level=logging.INFO)
+
+# Load pre-trained model tokenizer (vocabulary)
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+# Tokenize input
+text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
+tokenized_text = tokenizer.tokenize(text)
+
+# Mask a token that we will try to predict back with `BertForMaskedLM`
+masked_index = 8
+tokenized_text[masked_index] = '[MASK]'
+assert tokenized_text == ['[CLS]', 'who', 'was', 'jim', 'henson', '?', '[SEP]', 'jim', '[MASK]', 'was', 'a', 'puppet', '##eer', '[SEP]']
+
+# Convert token to vocabulary indices
+indexed_tokens = tokenizer.convert_tokens_to_ids(tokenized_text)
+# Define sentence A and B indices associated to 1st and 2nd sentences (see paper)
+segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
+
+# Convert inputs to PyTorch tensors
+tokens_tensor = torch.tensor([indexed_tokens])
+segments_tensors = torch.tensor([segments_ids])
+```
+
+Let's see how we can use `BertModel` to encode our inputs in hidden-states:
+
+```python
+# Load pre-trained model (weights)
+model = BertModel.from_pretrained('bert-base-uncased')
+
+# Set the model in evaluation mode to deactivate the DropOut modules
+# This is IMPORTANT to have reproducible results during evaluation!
+model.eval()
+
+# If you have a GPU, put everything on cuda
+tokens_tensor = tokens_tensor.to('cuda')
+segments_tensors = segments_tensors.to('cuda')
+model.to('cuda')
+
+# Predict hidden states features for each layer
+with torch.no_grad():
+    # See the models docstrings for the detail of the inputs
+    outputs = model(tokens_tensor, token_type_ids=segments_tensors)
+    # Transformers models always output tuples.
+    # See the models docstrings for the detail of all the outputs
+    # In our case, the first element is the hidden state of the last layer of the Bert model
+    encoded_layers = outputs[0]
+# We have encoded our input sequence in a FloatTensor of shape (batch size, sequence length, model hidden dimension)
+assert tuple(encoded_layers.shape) == (1, len(indexed_tokens), model.config.hidden_size)
+```
+
+And how to use `BertForMaskedLM` to predict a masked token:
+
+```python
+# Load pre-trained model (weights)
+model = BertForMaskedLM.from_pretrained('bert-base-uncased')
+model.eval()
+
+# If you have a GPU, put everything on cuda
+tokens_tensor = tokens_tensor.to('cuda')
+segments_tensors = segments_tensors.to('cuda')
+model.to('cuda')
+
+# Predict all tokens
+with torch.no_grad():
+    outputs = model(tokens_tensor, token_type_ids=segments_tensors)
+    predictions = outputs[0]
+
+# confirm we were able to predict 'henson'
+predicted_index = torch.argmax(predictions[0, masked_index]).item()
+predicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]
+assert predicted_token == 'henson'
+```
+
+### OpenAI GPT-2
+
+Here is a quick-start example using `GPT2Tokenizer` and `GPT2LMHeadModel` class with OpenAI's pre-trained model to predict the next token from a text prompt.
+
+First let's prepare a tokenized input from our text string using `GPT2Tokenizer`
+
+```python
+import torch
+from transformers import GPT2Tokenizer, GPT2LMHeadModel
+
+# OPTIONAL: if you want to have more information on what's happening, activate the logger as follows
+import logging
+logging.basicConfig(level=logging.INFO)
+
+# Load pre-trained model tokenizer (vocabulary)
+tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+
+# Encode a text inputs
+text = "Who was Jim Henson ? Jim Henson was a"
+indexed_tokens = tokenizer.encode(text)
+
+# Convert indexed tokens in a PyTorch tensor
+tokens_tensor = torch.tensor([indexed_tokens])
+```
+
+Let's see how to use `GPT2LMHeadModel` to generate the next token following our text:
+
+```python
+# Load pre-trained model (weights)
+model = GPT2LMHeadModel.from_pretrained('gpt2')
+
+# Set the model in evaluation mode to deactivate the DropOut modules
+# This is IMPORTANT to have reproducible results during evaluation!
+model.eval()
+
+# If you have a GPU, put everything on cuda
+tokens_tensor = tokens_tensor.to('cuda')
+model.to('cuda')
+
+# Predict all tokens
+with torch.no_grad():
+    outputs = model(tokens_tensor)
+    predictions = outputs[0]
+
+# get the predicted next sub-word (in our case, the word 'man')
+predicted_index = torch.argmax(predictions[0, -1, :]).item()
+predicted_text = tokenizer.decode(indexed_tokens + [predicted_index])
+assert predicted_text == 'Who was Jim Henson? Jim Henson was a man'
+```
+
+Examples for each model class of each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the [documentation](#documentation).
+
+#### Using the past
+
+GPT-2, as well as some other models (GPT, XLNet, Transfo-XL, CTRL), make use of a `past` or `mems` attribute which can be used to prevent re-computing the key/value pairs when using sequential decoding. It is useful when generating sequences as a big part of the attention mechanism benefits from previous computations.
+
+Here is a fully-working example using the `past` with `GPT2LMHeadModel` and argmax decoding (which should only be used as an example, as argmax decoding introduces a lot of repetition):
+
+```python
+from transformers import GPT2LMHeadModel, GPT2Tokenizer
+import torch
+
+tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+model = GPT2LMHeadModel.from_pretrained('gpt2')
+
+generated = tokenizer.encode("The Manhattan bridge")
+context = torch.tensor([generated])
+past = None
+
+for i in range(100):
+    print(i)
+    output, past = model(context, past=past)
+    token = torch.argmax(output[..., -1, :])
+
+    generated += [token.tolist()]
+    context = token.unsqueeze(0)
+
+sequence = tokenizer.decode(generated)
+
+print(sequence)
+```
+
+The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.

docs/source/quicktour.rst

@@ -1,379 +0,0 @@
-Quick tour
-==========
-
-Let's have a quick look at the 🤗 Transformers library features. The library downloads pretrained models for
-Natural Language Understanding (NLU) tasks, such as analyzing the sentiment of a text, and Natural Language Generation (NLG),
-such as completing a prompt with new text or translating in another language.
-
-First we will see how to easily leverage the pipeline API to quickly use those pretrained models at inference. Then, we
-will dig a little bit more and see how the library gives you access to those models and helps you preprocess your data.
-
-.. note::
-
-    All code examples presented in the documentation have a switch on the top left for Pytorch versus TensorFlow. If
-    not, the code is expected to work for both backends without any change needed.
-
-Getting started on a task with a pipeline
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-The easiest way to use a pretrained model on a given task is to use :func:`~transformers.pipeline`. 🤗 Transformers
-provides the following tasks out of the box:
-
-- Sentiment analysis: is a text positive or negative?
-- Text generation (in English): provide a prompt and the model will generate what follows.
-- Name entity recognition (NER): in an input sentence, label each word with the entity it represents (person, place,
-  etc.)
-- Question answering: provide the model with some context and a question, extract the answer from the context.
-- Filling masked text: given a text with masked words (e.g., replaced by ``[MASK]``), fill the blanks.
-- Summarization: generate a summary of a long text.
-- Translation: translate a text in another language.
-- Feature extraction: return a tensor representation of the text.
-
-Let's see how this work for sentiment analysis (the other tasks are all covered in the
-:doc:`task summary </task_summary>`):
-
-::
-
-    from transformers import pipeline
-    classifier = pipeline('sentiment-analysis')
-
-When typing this command for the first time, a pretrained model and its tokenizer are downloaded and cached. We will
-look at both later on, but as an introduction the tokenizer's job is to preprocess the text for the model, which is
-then responsible for making predictions. The pipeline groups all of that together, and post-process the predictions to
-make them readable. For instance
-
-::
-
-    classifier('We are very happy to show you the Transformers library.')
-
-will return something like this:
-
-::
-
-    [{'label': 'POSITIVE', 'score': 0.999799370765686}]
-
-That's encouraging! You can use it on a list of sentences, which will be preprocessed then fed to the model as a
-`batch`:
-
-::
-
-    classifier(["We are very happy to show you the Transformers library.",
-                "We hope you don't hate it."])
-
-returning a list of dictionaries like this one:
-
-::
-
-    [{'label': 'POSITIVE', 'score': 0.999799370765686},
-     {'label': 'NEGATIVE', 'score': 0.5308589935302734}]
-
-You can see the second sentence has been classified as negative (it needs to be positive or negative) but its score is
-fairly neutral.
-
-By default, the model downloaded for this pipeline is called "distilbert-base-uncased-finetuned-sst-2-english". We can
-look at its `model page <https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english>`__ to get more
-information about it. It uses the :doc:`DistilBERT architecture </model_doc/distilbert>` and has been fine-tuned on a
-dataset called SST-2 for the sentiment analysis task.
-
-Let's say we want to use another model; for instance, one that has been trained on French data. We can search through
-the `model hub <https://huggingface.co/models>`__ that gathers models pretrained on a lot of data by research labs, but
-also community models (usually fine-tuned versions of those big models on a specific dataset). Applying the tags
-"French" and "text-classification" gives back a suggestion "nlptown/bert-base-multilingual-uncased-sentiment". Let's
-see how we can use it. 
-
-You can directly pass the name of the model to use to :func:`~transformers.pipeline`:
-
-::
-
-    classifier = pipeline('sentiment-analysis', model="nlptown/bert-base-multilingual-uncased-sentiment")
-
-This classifier can now deal with texts in English, French, but also Dutch, German, Italian and Spanish! You can also
-replace that name by a local folder where you have saved a pretrained model (see below). You can also pass a model
-object and its associated tokenizer.
-
-We will need two classes for this. The first is :class:`~transformers.AutoTokenizer`, which we will use to download the
-tokenizer associated to the model we picked and instantiate it. The second is
-:class:`~transformers.AutoModelForSequenceClassification` (or
-:class:`~transformers.TFAutoModelForSequenceClassification` if you are using TensorFlow), which we will use to download
-the model itself. Note that if we were using the library on an other task, the class of the model would change. The
-:doc:`task summary </task_summary>` tutorial summarizes which class is used for which task.
-
-::
-
-    ## PYTORCH CODE
-    from transformers import AutoTokenizer, AutoModelForSequenceClassification
-    ## TENSORFLOW CODE
-    from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
-
-Now, to download the models and tokenizer we found previously, we just have to use the 
-:func:`~transformers.AutoModelForSequenceClassification.from_pretrained` method (feel free to replace ``model_name`` by
-any other model from the model hub):
-
-::
-
-    ## PYTORCH CODE
-    model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
-    model = AutoModelForSequenceClassification.from_pretrained(model_name)
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-    pipe = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
-    ## TENSORFLOW CODE
-    model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
-    model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-    classifier = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
-
-If you don't find a model that has been pretrained on some data similar to yours, you will need to fine-tune a
-pretrained model on your data. We provide :doc:`example scripts </examples>` to do so. Once you're done, don't forget
-to share your fine-tuned model on the hub with the community, using :doc:`this tutorial </model_sharing>`.
-
-.. _pretrained-model:
-
-Under the hood: pretrained models
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-Let's now see what happens beneath the hood when using those pipelines. As we saw, the model and tokenizer are created
-using the :obj:`from_pretrained` method:
-
-::
-
-    ## PYTORCH CODE
-    from transformers import AutoTokenizer, AutoModelForSequenceClassification
-    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    model = AutoModelForSequenceClassification.from_pretrained(model_name)
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-    ## TENSORFLOW CODE
-    from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
-    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-
-Using the tokenizer
-^^^^^^^^^^^^^^^^^^^
-
-We mentioned the tokenizer is responsible for the preprocessing of your texts. First, it will split a given text in
-words (or part of words, punctuation symbols, etc.) usually called `tokens`. There are multiple rules that can govern
-that process, which is why we need to instantiate the tokenizer using the name of the model, to make sure we use the
-same rules as when the model was pretrained.
-
-The second step is to convert those `tokens` into numbers, to be able to build a tensor out of them and feed them to
-the model. To do this, the tokenizer has a `vocab`, which is the part we download when we instantiate it with the
-:obj:`from_pretrained` method, since we need to use the same `vocab` as when the model was pretrained.
-
-To apply these steps on a given text, we can just feed it to our tokenizer:
-
-::
-
-    input = tokenizer("We are very happy to show you the Transformers library.")
-    print(input)
-
-This returns a dictionary string to list of ints. It contains the `ids of the tokens <glossary.html#input-ids>`__,
-as mentioned before, but also additional arguments that will be useful to the model. Here for instance, we also have an
-`attention mask <glossary.html#attention-mask>`__ that the model will use to have a better understanding of the sequence:
-
-
-::
-
-    {'input_ids': [101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012, 102],
-     'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
-
-You can pass a list of sentences directly to your tokenizer. If your goal is to send them through your model as a
-batch, you probably want to pad them all to the same length, truncate them to the maximum length the model can accept
-and get tensors back. You can specify all of that to the tokenizer:
-
-::
-
-    ## PYTORCH CODE
-    batch = tokenizer(
-        ["We are very happy to show you the Transformers library.",
-         "We hope you don't hate it."],
-        padding=True, truncation=True, return_tensors="pt")
-    print(batch)
-    ## TENSORFLOW CODE
-    batch = tokenizer(
-        ["We are very happy to show you the Transformers library.",
-         "We hope you don't hate it."],
-        padding=True, truncation=True, return_tensors="tf")
-    print(batch)
-
-The padding is automatically applied on the side the model expect it (in this case, on the right), with the
-padding token the model was pretrained with. The attention mask is also adapted to take the padding into account:
-
-::
-
-    {'input_ids': tensor([[  101,  2057,  2024,  2200,  3407,  2000,  2265,  2017,  1996, 19081, 3075,  1012,   102],
-                          [  101,  2057,  3246,  2017,  2123,  1005,  1056,  5223,  2009,  1012,  102,     0,     0]]), 
-     'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
-                               [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]])}
-
-You can learn more about tokenizers on their :doc:`doc page <main_classes/tokenizer>` (tutorial coming soon).
-
-Using the model
-^^^^^^^^^^^^^^^
-
-Once your input has been preprocessed by the tokenizer, you can directly send it to the model. As we mentioned, it will
-contain all the relevant information the model needs. If you're using a TensorFlow model, you can directly pass the
-dictionary keys to tensor, for a PyTorch model, you need to unpack the dictionary by adding :obj:`**`.
-
-::
-
-    ## PYTORCH CODE
-    outputs = model(**batch)
-    ## TENSORFLOW CODE
-    outputs = model(batch)
-
-In 🤗 Transformers, all outputs are tuples (with only one element potentially). Here, we get a tuple with just the
-final activations of the model.
-
-::
-
-    (tensor([[-4.1329,  4.3811],
-             [ 0.0818, -0.0418]]),)
-
-.. note::
-
-    All 🤗 Transformers models (PyTorch or TensorFlow) return the activations of the model *before* the final
-    activation function (like SoftMax) since this final activation function is often fused with the loss.
-
-Let's apply the SoftMax activation to get predictions.
-
-::
-
-    ## PYTORCH CODE
-    import torch.nn.functional as F
-    predictions = F.softmax(outputs[0], dim=-1)
-    print(predictions)
-    ## TENSORFLOW CODE
-    predictions = tf.nn.softmax(outputs[0], axis=-1)
-    print(predictions)
-
-We can see we get the numbers from before:
-
-::
-
-    tensor([[2.0060e-04, 9.9980e-01],
-            [5.3086e-01, 4.6914e-01]])
-
-If you have labels, you can provide them to the model, it will return a tuple with the loss and the final activations.
-
-::
-
-    ## PYTORCH CODE
-    import torch
-    outputs = model(**batch, labels = torch.tensor([1, 0])
-    ## TENSORFLOW CODE
-    import tensorflow as tf
-    outputs = model(batch, labels = tf.constant([1, 0])
-
-Models are standard `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__ or
-`tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ so you can use them in your usual
-training loop. 🤗 Transformers also provides a :class:`~transformers.Trainer` (or :class:`~transformers.TFTrainer` if
-you are using TensorFlow) class to help with your training (taking care of things such as distributed training, mixed
-precision, etc.). See the training tutorial (coming soon) for more details.
-
-Once your model is fine-tuned, you can save it with its tokenizer the following way:
-
-::
-
-    tokenizer.save_pretrained(save_directory)
-    model.save_pretrained(save_directory)
-
-You can then load this model back using the :func:`~transformers.AutoModel.from_pretrained` method by passing the
-directory name instead of the model name. One cool feature of 🤗 Transformers is that you can easily switch between
-PyTorch and TensorFlow: any model saved as before can be loaded back either in PyTorch or TensorFlow. If you are
-loading a saved PyTorch model in a TensorFlow model, use :func:`~transformers.TFAutoModel.from_pretrained` like this:
-
-::
-
-    tokenizer = AutoTokenizer.from_pretrained(save_directory)
-    model = TFAutoModel.from_pretrained(save_directory, from_pt=True)
-
-and if you are loading a saved TensorFlow model in a PyTorch model, you should use the following code:
-
-::
-
-    tokenizer = AutoTokenizer.from_pretrained(save_directory)
-    model = AutoModel.from_pretrained(save_directory, from_tf=True)
-
-Lastly, you can also ask the model to return all hidden states and all attention weights if you need them:
-
-
-::
-
-    ## PYTORCH CODE
-    outputs = model(**batch, output_hidden_states=True, output_attentions=True)
-    all_hidden_states, all_attentions = outputs[-2:]
-    ## TENSORFLOW CODE
-    outputs = model(batch, output_hidden_states=True, output_attentions=True)
-    all_hidden_states, all_attentions = outputs[-2:]
-
-Accessing the code
-^^^^^^^^^^^^^^^^^^
-
-The :obj:`AutoModel` and :obj:`AutoTokenizer` classes are just shortcuts that will automatically work with any
-pretrained model. Behind the scenes, the library has one model class per combination of architecture plus class, so the
-code is easy to access and tweak if you need to.
-
-In our previous example, the model was called "distilbert-base-uncased-finetuned-sst-2-english", which means it's
-using the :doc:`DistilBERT </model_doc/distilbert>` architecture. The model automatically created is then a
-:class:`~transformers.DistilBertForSequenceClassification`. You can look at its documentation for all details relevant
-to that specific model, or browse the source code. This is how you would directly instantiate model and tokenizer
-without the auto magic:
-
-::
-
-    ## PYTORCH CODE
-    from transformers import DistilBertTokenizer, DistilBertForSequenceClassification
-    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    model = DistilBertForSequenceClassification.from_pretrained(model_name)
-    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
-    ## TENSORFLOW CODE
-    from transformers import DistilBertTokenizer, TFDistilBertForSequenceClassification
-    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    model = TFDistilBertForSequenceClassification.from_pretrained(model_name)
-    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
-
-Customizing the model
-^^^^^^^^^^^^^^^^^^^^^
-
-If you want to change how the model itself is built, you can define your custom configuration class. Each architecture
-comes with its own relevant configuration (in the case of DistilBERT, :class:`~transformers.DistilBertConfig`) which
-allows you to specify any of the hidden dimension, dropout rate etc. If you do core modifications, like changing the
-hidden size, you won't be able to use a pretrained model anymore and will need to train from scratch. You would then
-instantiate the model directly from this configuration.
-
-Here we use the predefined vocabulary of DistilBERT (hence load the tokenizer with the
-:func:`~transformers.DistilBertTokenizer.from_pretrained` method) and initialize the model from scratch (hence
-instantiate the model from the configuration instead of using the
-:func:`~transformers.DistilBertForSequenceClassification.from_pretrained` method).
-
-::
-
-    ## PYTORCH CODE
-    from transformers import DistilBertConfig, DistilBertTokenizer, DistilBertForSequenceClassification
-    config = DistilBertConfig(n_heads=8, dim=512, hidden_dim=4*512)
-    tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
-    model = DistilBertForSequenceClassification(config)
-    ## TENSORFLOW CODE
-    from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForSequenceClassification
-    config = DistilBertConfig(n_heads=8, dim=512, hidden_dim=4*512)
-    tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
-    model = TFDistilBertForSequenceClassification(config)
-
-For something that only changes the head of the model (for instance, the number of labels), you can still use a
-pretrained model for the body. For instance, let's define a classifier for 10 different labels using a pretrained body.
-We could create a configuration with all the default values and just change the number of labels, but more easily, you
-can directly pass any argument a configuration would take to the :func:`from_pretrained` method and it will update the
-default configuration with it:
-
-::
-
-    ## PYTORCH CODE
-    from transformers import DistilBertConfig, DistilBertTokenizer, DistilBertForSequenceClassification
-    model_name = "distilbert-base-uncased"
-    model = DistilBertForSequenceClassification.from_pretrained(model_name, num_labels=10)
-    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
-    ## TENSORFLOW CODE
-    from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForSequenceClassification
-    model_name = "distilbert-base-uncased"
-    model = TFDistilBertForSequenceClassification.from_pretrained(model_name, num_labels=10)
-    tokenizer = DistilBertTokenizer.from_pretrained(model_name)

docs/source/usage.rst

@@ -1,4 +1,4 @@
-Summary of the tasks
+Usage
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 This page shows the most frequent use-cases when using the library. The models available allow for many different

examples/benchmarking/run_benchmark.py

@@ -1,5 +1,5 @@
 # coding=utf-8
-# Copyright 2020 The HuggingFace Inc. team.
+# Copyright 2018 The HuggingFace Inc. team.
 # Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");

examples/benchmarking/run_benchmark_tf.py

@@ -1,6 +1,6 @@
 # coding=utf-8
 # Copyright 2018 The HuggingFace Inc. team.
-# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.

examples/bert-loses-patience/README.md

@@ -1,89 +0,0 @@
-# Patience-based Early Exit
-
-Patience-based Early Exit (PABEE) is a plug-and-play inference method for pretrained language models.
-We have already implemented it on BERT and ALBERT. Basically, you can make your LM faster and more robust with PABEE. It can even improve the performance of ALBERT on GLUE. The only sacrifice is that the batch size can only be 1.
-Learn more in the paper ["BERT Loses Patience: Fast and Robust Inference with Early Exit"](https://arxiv.org/abs/2006.04152) and the official [GitHub repo](https://github.com/JetRunner/PABEE).
-
-![PABEE](https://github.com/JetRunner/PABEE/raw/master/bert-loses-patience.png)
-
-## Training
-
-You can fine-tune a pretrained language model (you can choose from BERT and ALBERT) and train the internal classifiers by:
-```bash
-export GLUE_DIR=/path/to/glue_data
-export TASK_NAME=MRPC
-
-python ./run_glue_with_pabee.py \
-  --model_type albert \
-  --model_name_or_path bert-base-uncased/albert-base-v2 \
-  --task_name $TASK_NAME \
-  --do_train \
-  --do_eval \
-  --do_lower_case \
-  --data_dir "$GLUE_DIR/$TASK_NAME" \
-  --max_seq_length 128 \
-  --per_gpu_train_batch_size 32 \
-  --per_gpu_eval_batch_size 32 \
-  --learning_rate 2e-5 \
-  --save_steps 50 \
-  --logging_steps 50 \
-  --num_train_epochs 5 \
-  --output_dir /path/to/save/ \
-  --evaluate_during_training
-```
-
-## Inference
-
-You can inference with different patience settings by:
-```bash
-export GLUE_DIR=/path/to/glue_data
-export TASK_NAME=MRPC
-
-python ./run_glue_with_pabee.py \
-  --model_type albert \
-  --model_name_or_path /path/to/save/ \
-  --task_name $TASK_NAME \
-  --do_eval \
-  --do_lower_case \
-  --data_dir "$GLUE_DIR/$TASK_NAME" \
-  --max_seq_length 128 \
-  --per_gpu_eval_batch_size 1 \
-  --learning_rate 2e-5 \
-  --logging_steps 50 \
-  --num_train_epochs 15 \
-  --output_dir /path/to/save/ \
-  --eval_all_checkpoints \
-  --patience 3,4,5,6,7,8
-```
-where `patience` can be a list of patience settings, separated by a comma. It will help determine which patience works best.
-
-When evaluating on a regression task (STS-B), you may add `--regression_threshold 0.1` to define the regression threshold.
-
-## Results
-On the GLUE dev set:
-
-| Model        | \#Param | Speed  | CoLA  | MNLI  | MRPC  | QNLI  | QQP   | RTE   | SST\-2 | STS\-B |
-|--------------|---------|--------|-------|-------|-------|-------|-------|-------|--------|--------|
-| ALBERT\-base | 12M     |        | 58\.9 | 84\.6 | 89\.5 | 91\.7 | 89\.6 | 78\.6 | 92\.8  | 89\.5  |
-| \+PABEE      | 12M     | 1\.57x | 61\.2 | 85\.1 | 90\.0 | 91\.8 | 89\.6 | 80\.1 | 93\.0  | 90\.1  |
-
-| Model         | \#Param | Speed\-up | MNLI  | SST\-2 | STS\-B |
-|---------------|---------|-----------|-------|--------|--------|
-| BERT\-base    | 108M    |           | 84\.5 | 92\.1  | 88\.9  |
-| \+PABEE       | 108M    | 1\.62x    | 83\.6 | 92\.0  | 88\.7  |
-| ALBERT\-large | 18M     |           | 86\.4 | 94\.9  | 90\.4  |
-| \+PABEE       | 18M     | 2\.42x    | 86\.8 | 95\.2  | 90\.6  |
-
-
-## Citation
-If you find this resource useful, please consider citing the following paper:
-```bibtex
-@misc{zhou2020bert,
-    title={BERT Loses Patience: Fast and Robust Inference with Early Exit},
-    author={Wangchunshu Zhou and Canwen Xu and Tao Ge and Julian McAuley and Ke Xu and Furu Wei},
-    year={2020},
-    eprint={2006.04152},
-    archivePrefix={arXiv},
-    primaryClass={cs.CL}
-}
-```

examples/bert-loses-patience/pabee/modeling_pabee_albert.py

@@ -1,310 +0,0 @@
-# coding=utf-8
-# Copyright 2020 Google AI, Google Brain, the HuggingFace Inc. team and Microsoft Corporation.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""PyTorch ALBERT model with Patience-based Early Exit. """
-
-import logging
-
-import torch
-import torch.nn as nn
-from torch.nn import CrossEntropyLoss, MSELoss
-
-from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
-from transformers.modeling_albert import (
-    ALBERT_INPUTS_DOCSTRING,
-    ALBERT_START_DOCSTRING,
-    AlbertModel,
-    AlbertPreTrainedModel,
-    AlbertTransformer,
-)
-
-
-logger = logging.getLogger(__name__)
-
-
-class AlbertTransformerWithPabee(AlbertTransformer):
-    def adaptive_forward(self, hidden_states, current_layer, attention_mask=None, head_mask=None):
-        if current_layer == 0:
-            hidden_states = self.embedding_hidden_mapping_in(hidden_states)
-        else:
-            hidden_states = hidden_states[0]
-
-        layers_per_group = int(self.config.num_hidden_layers / self.config.num_hidden_groups)
-
-        # Index of the hidden group
-        group_idx = int(current_layer / (self.config.num_hidden_layers / self.config.num_hidden_groups))
-
-        layer_group_output = self.albert_layer_groups[group_idx](
-            hidden_states,
-            attention_mask,
-            head_mask[group_idx * layers_per_group : (group_idx + 1) * layers_per_group],
-        )
-        hidden_states = layer_group_output[0]
-
-        return (hidden_states,)
-
-
-@add_start_docstrings(
-    "The bare ALBERT Model transformer with PABEE outputting raw hidden-states without any specific head on top.",
-    ALBERT_START_DOCSTRING,
-)
-class AlbertModelWithPabee(AlbertModel):
-    def __init__(self, config):
-        super().__init__(config)
-
-        self.encoder = AlbertTransformerWithPabee(config)
-
-        self.init_weights()
-        self.patience = 0
-        self.inference_instances_num = 0
-        self.inference_layers_num = 0
-
-        self.regression_threshold = 0
-
-    def set_regression_threshold(self, threshold):
-        self.regression_threshold = threshold
-
-    def set_patience(self, patience):
-        self.patience = patience
-
-    def reset_stats(self):
-        self.inference_instances_num = 0
-        self.inference_layers_num = 0
-
-    def log_stats(self):
-        avg_inf_layers = self.inference_layers_num / self.inference_instances_num
-        message = f"*** Patience = {self.patience} Avg. Inference Layers = {avg_inf_layers:.2f} Speed Up = {1 - avg_inf_layers / self.config.num_hidden_layers:.2f} ***"
-        print(message)
-
-    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        output_dropout=None,
-        output_layers=None,
-        regression=False,
-    ):
-        r"""
-    Return:
-        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.AlbertConfig`) and inputs:
-        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        pooler_output (:obj:`torch.FloatTensor`: of shape :obj:`(batch_size, hidden_size)`):
-            Last layer hidden-state of the first token of the sequence (classification token)
-            further processed by a Linear layer and a Tanh activation function. The Linear
-            layer weights are trained from the next sentence prediction (classification)
-            objective during pre-training.
-
-            This output is usually *not* a good summary
-            of the semantic content of the input, you're often better with averaging or pooling
-            the sequence of hidden-states for the whole input sequence.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
-
-        if input_ids is not None and inputs_embeds is not None:
-            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
-        elif input_ids is not None:
-            input_shape = input_ids.size()
-        elif inputs_embeds is not None:
-            input_shape = inputs_embeds.size()[:-1]
-        else:
-            raise ValueError("You have to specify either input_ids or inputs_embeds")
-
-        device = input_ids.device if input_ids is not None else inputs_embeds.device
-
-        if attention_mask is None:
-            attention_mask = torch.ones(input_shape, device=device)
-        if token_type_ids is None:
-            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
-
-        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
-        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
-        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
-        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
-
-        embedding_output = self.embeddings(
-            input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
-        )
-        encoder_outputs = embedding_output
-
-        if self.training:
-            res = []
-            for i in range(self.config.num_hidden_layers):
-                encoder_outputs = self.encoder.adaptive_forward(
-                    encoder_outputs, current_layer=i, attention_mask=extended_attention_mask, head_mask=head_mask,
-                )
-
-                pooled_output = self.pooler_activation(self.pooler(encoder_outputs[0][:, 0]))
-                logits = output_layers[i](output_dropout(pooled_output))
-                res.append(logits)
-        elif self.patience == 0:  # Use all layers for inference
-            encoder_outputs = self.encoder(encoder_outputs, extended_attention_mask, head_mask=head_mask)
-            pooled_output = self.pooler_activation(self.pooler(encoder_outputs[0][:, 0]))
-            res = [output_layers[self.config.num_hidden_layers - 1](pooled_output)]
-        else:
-            patient_counter = 0
-            patient_result = None
-            calculated_layer_num = 0
-            for i in range(self.config.num_hidden_layers):
-                calculated_layer_num += 1
-                encoder_outputs = self.encoder.adaptive_forward(
-                    encoder_outputs, current_layer=i, attention_mask=extended_attention_mask, head_mask=head_mask,
-                )
-
-                pooled_output = self.pooler_activation(self.pooler(encoder_outputs[0][:, 0]))
-                logits = output_layers[i](pooled_output)
-                if regression:
-                    labels = logits.detach()
-                    if patient_result is not None:
-                        patient_labels = patient_result.detach()
-                    if (patient_result is not None) and torch.abs(patient_result - labels) < self.regression_threshold:
-                        patient_counter += 1
-                    else:
-                        patient_counter = 0
-                else:
-                    labels = logits.detach().argmax(dim=1)
-                    if patient_result is not None:
-                        patient_labels = patient_result.detach().argmax(dim=1)
-                    if (patient_result is not None) and torch.all(labels.eq(patient_labels)):
-                        patient_counter += 1
-                    else:
-                        patient_counter = 0
-
-                patient_result = logits
-                if patient_counter == self.patience:
-                    break
-            res = [patient_result]
-            self.inference_layers_num += calculated_layer_num
-            self.inference_instances_num += 1
-
-        return res
-
-
-@add_start_docstrings(
-    """Albert Model transformer with PABEE and a sequence classification/regression head on top (a linear layer on top of
-    the pooled output) e.g. for GLUE tasks. """,
-    ALBERT_START_DOCSTRING,
-)
-class AlbertForSequenceClassificationWithPabee(AlbertPreTrainedModel):
-    def __init__(self, config):
-        super().__init__(config)
-        self.num_labels = config.num_labels
-
-        self.albert = AlbertModelWithPabee(config)
-        self.dropout = nn.Dropout(config.classifier_dropout_prob)
-        self.classifiers = nn.ModuleList(
-            [nn.Linear(config.hidden_size, self.config.num_labels) for _ in range(config.num_hidden_layers)]
-        )
-
-        self.init_weights()
-
-    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        labels=None,
-    ):
-        r"""
-        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
-            Labels for computing the sequence classification/regression loss.
-            Indices should be in ``[0, ..., config.num_labels - 1]``.
-            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
-            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
-
-    Returns:
-        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.AlbertConfig`) and inputs:
-        loss: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
-            Classification (or regression if config.num_labels==1) loss.
-        logits ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-
-        Examples::
-
-            from transformers import AlbertTokenizer
-            from pabee import AlbertForSequenceClassificationWithPabee
-            import torch
-
-            tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
-            model = AlbertForSequenceClassificationWithPabee.from_pretrained('albert-base-v2')
-            input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-            labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
-            outputs = model(input_ids, labels=labels)
-            loss, logits = outputs[:2]
-
-        """
-
-        logits = self.albert(
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-            output_dropout=self.dropout,
-            output_layers=self.classifiers,
-            regression=self.num_labels == 1,
-        )
-
-        outputs = (logits[-1],)
-
-        if labels is not None:
-            total_loss = None
-            total_weights = 0
-            for ix, logits_item in enumerate(logits):
-                if self.num_labels == 1:
-                    #  We are doing regression
-                    loss_fct = MSELoss()
-                    loss = loss_fct(logits_item.view(-1), labels.view(-1))
-                else:
-                    loss_fct = CrossEntropyLoss()
-                    loss = loss_fct(logits_item.view(-1, self.num_labels), labels.view(-1))
-                if total_loss is None:
-                    total_loss = loss
-                else:
-                    total_loss += loss * (ix + 1)
-                total_weights += ix + 1
-            outputs = (total_loss / total_weights,) + outputs
-
-        return outputs

examples/bert-loses-patience/pabee/modeling_pabee_bert.py

@@ -1,342 +0,0 @@
-# coding=utf-8
-# Copyright 2020 The Google AI Language Team Authors, The HuggingFace Inc. team and Microsoft Corporation.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""PyTorch BERT model with Patience-based Early Exit. """
-
-
-import logging
-
-import torch
-from torch import nn
-from torch.nn import CrossEntropyLoss, MSELoss
-
-from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
-from transformers.modeling_bert import (
-    BERT_INPUTS_DOCSTRING,
-    BERT_START_DOCSTRING,
-    BertEncoder,
-    BertModel,
-    BertPreTrainedModel,
-)
-
-
-logger = logging.getLogger(__name__)
-
-
-class BertEncoderWithPabee(BertEncoder):
-    def adaptive_forward(self, hidden_states, current_layer, attention_mask=None, head_mask=None):
-        layer_outputs = self.layer[current_layer](hidden_states, attention_mask, head_mask[current_layer])
-
-        hidden_states = layer_outputs[0]
-
-        return hidden_states
-
-
-@add_start_docstrings(
-    "The bare Bert Model transformer with PABEE outputting raw hidden-states without any specific head on top.",
-    BERT_START_DOCSTRING,
-)
-class BertModelWithPabee(BertModel):
-    """
-
-    The model can behave as an encoder (with only self-attention) as well
-    as a decoder, in which case a layer of cross-attention is added between
-    the self-attention layers, following the architecture described in `Attention is all you need`_ by Ashish Vaswani,
-    Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
-
-    To behave as an decoder the model needs to be initialized with the
-    :obj:`is_decoder` argument of the configuration set to :obj:`True`; an
-    :obj:`encoder_hidden_states` is expected as an input to the forward pass.
-
-    .. _`Attention is all you need`:
-        https://arxiv.org/abs/1706.03762
-
-    """
-
-    def __init__(self, config):
-        super().__init__(config)
-
-        self.encoder = BertEncoderWithPabee(config)
-
-        self.init_weights()
-        self.patience = 0
-        self.inference_instances_num = 0
-        self.inference_layers_num = 0
-
-        self.regression_threshold = 0
-
-    def set_regression_threshold(self, threshold):
-        self.regression_threshold = threshold
-
-    def set_patience(self, patience):
-        self.patience = patience
-
-    def reset_stats(self):
-        self.inference_instances_num = 0
-        self.inference_layers_num = 0
-
-    def log_stats(self):
-        avg_inf_layers = self.inference_layers_num / self.inference_instances_num
-        message = f"*** Patience = {self.patience} Avg. Inference Layers = {avg_inf_layers:.2f} Speed Up = {1 - avg_inf_layers / self.config.num_hidden_layers:.2f} ***"
-        print(message)
-
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        encoder_hidden_states=None,
-        encoder_attention_mask=None,
-        output_dropout=None,
-        output_layers=None,
-        regression=False,
-    ):
-        r"""
-    Return:
-        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        pooler_output (:obj:`torch.FloatTensor`: of shape :obj:`(batch_size, hidden_size)`):
-            Last layer hidden-state of the first token of the sequence (classification token)
-            further processed by a Linear layer and a Tanh activation function. The Linear
-            layer weights are trained from the next sentence prediction (classification)
-            objective during pre-training.
-
-            This output is usually *not* a good summary
-            of the semantic content of the input, you're often better with averaging or pooling
-            the sequence of hidden-states for the whole input sequence.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
-
-        if input_ids is not None and inputs_embeds is not None:
-            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
-        elif input_ids is not None:
-            input_shape = input_ids.size()
-        elif inputs_embeds is not None:
-            input_shape = inputs_embeds.size()[:-1]
-        else:
-            raise ValueError("You have to specify either input_ids or inputs_embeds")
-
-        device = input_ids.device if input_ids is not None else inputs_embeds.device
-
-        if attention_mask is None:
-            attention_mask = torch.ones(input_shape, device=device)
-        if token_type_ids is None:
-            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
-
-        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
-        # ourselves in which case we just need to make it broadcastable to all heads.
-        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
-
-        # If a 2D ou 3D attention mask is provided for the cross-attention
-        # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
-        if self.config.is_decoder and encoder_hidden_states is not None:
-            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
-            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
-            if encoder_attention_mask is None:
-                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
-            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
-        else:
-            encoder_extended_attention_mask = None
-
-        # Prepare head mask if needed
-        # 1.0 in head_mask indicate we keep the head
-        # attention_probs has shape bsz x n_heads x N x N
-        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
-        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
-        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
-
-        embedding_output = self.embeddings(
-            input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
-        )
-        encoder_outputs = embedding_output
-
-        if self.training:
-            res = []
-            for i in range(self.config.num_hidden_layers):
-                encoder_outputs = self.encoder.adaptive_forward(
-                    encoder_outputs, current_layer=i, attention_mask=extended_attention_mask, head_mask=head_mask
-                )
-
-                pooled_output = self.pooler(encoder_outputs)
-                logits = output_layers[i](output_dropout(pooled_output))
-                res.append(logits)
-        elif self.patience == 0:  # Use all layers for inference
-            encoder_outputs = self.encoder(
-                embedding_output,
-                attention_mask=extended_attention_mask,
-                head_mask=head_mask,
-                encoder_hidden_states=encoder_hidden_states,
-                encoder_attention_mask=encoder_extended_attention_mask,
-            )
-            pooled_output = self.pooler(encoder_outputs[0])
-            res = [output_layers[self.config.num_hidden_layers - 1](pooled_output)]
-        else:
-            patient_counter = 0
-            patient_result = None
-            calculated_layer_num = 0
-            for i in range(self.config.num_hidden_layers):
-                calculated_layer_num += 1
-                encoder_outputs = self.encoder.adaptive_forward(
-                    encoder_outputs, current_layer=i, attention_mask=extended_attention_mask, head_mask=head_mask
-                )
-
-                pooled_output = self.pooler(encoder_outputs)
-                logits = output_layers[i](pooled_output)
-                if regression:
-                    labels = logits.detach()
-                    if patient_result is not None:
-                        patient_labels = patient_result.detach()
-                    if (patient_result is not None) and torch.abs(patient_result - labels) < self.regression_threshold:
-                        patient_counter += 1
-                    else:
-                        patient_counter = 0
-                else:
-                    labels = logits.detach().argmax(dim=1)
-                    if patient_result is not None:
-                        patient_labels = patient_result.detach().argmax(dim=1)
-                    if (patient_result is not None) and torch.all(labels.eq(patient_labels)):
-                        patient_counter += 1
-                    else:
-                        patient_counter = 0
-
-                patient_result = logits
-                if patient_counter == self.patience:
-                    break
-            res = [patient_result]
-            self.inference_layers_num += calculated_layer_num
-            self.inference_instances_num += 1
-
-        return res
-
-
-@add_start_docstrings(
-    """Bert Model transformer with PABEE and a sequence classification/regression head on top (a linear layer on top of
-    the pooled output) e.g. for GLUE tasks. """,
-    BERT_START_DOCSTRING,
-)
-class BertForSequenceClassificationWithPabee(BertPreTrainedModel):
-    def __init__(self, config):
-        super().__init__(config)
-        self.num_labels = config.num_labels
-
-        self.bert = BertModelWithPabee(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.classifiers = nn.ModuleList(
-            [nn.Linear(config.hidden_size, self.config.num_labels) for _ in range(config.num_hidden_layers)]
-        )
-
-        self.init_weights()
-
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        labels=None,
-    ):
-        r"""
-        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
-            Labels for computing the sequence classification/regression loss.
-            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
-            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
-            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
-
-    Returns:
-        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
-            Classification (or regression if config.num_labels==1) loss.
-        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-
-    Examples::
-
-        from transformers import BertTokenizer, BertForSequenceClassification
-        from pabee import BertForSequenceClassificationWithPabee
-        import torch
-
-        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-        model = BertForSequenceClassificationWithPabee.from_pretrained('bert-base-uncased')
-
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
-        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
-        outputs = model(input_ids, labels=labels)
-
-        loss, logits = outputs[:2]
-
-        """
-
-        logits = self.bert(
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-            output_dropout=self.dropout,
-            output_layers=self.classifiers,
-            regression=self.num_labels == 1,
-        )
-
-        outputs = (logits[-1],)
-
-        if labels is not None:
-            total_loss = None
-            total_weights = 0
-            for ix, logits_item in enumerate(logits):
-                if self.num_labels == 1:
-                    #  We are doing regression
-                    loss_fct = MSELoss()
-                    loss = loss_fct(logits_item.view(-1), labels.view(-1))
-                else:
-                    loss_fct = CrossEntropyLoss()
-                    loss = loss_fct(logits_item.view(-1, self.num_labels), labels.view(-1))
-                if total_loss is None:
-                    total_loss = loss
-                else:
-                    total_loss += loss * (ix + 1)
-                total_weights += ix + 1
-            outputs = (total_loss / total_weights,) + outputs
-
-        return outputs

examples/bert-loses-patience/run_glue_with_pabee.py

@@ -1,708 +0,0 @@
-# coding=utf-8
-# Copyright 2020 The Google AI Language Team Authors, The HuggingFace Inc. team and Microsoft Corporation.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" Training and inference using the library models for sequence classification on GLUE (Bert, Albert) with PABEE."""
-
-
-import argparse
-import glob
-import json
-import logging
-import os
-import random
-
-import numpy as np
-import torch
-from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
-from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm, trange
-
-from pabee.modeling_pabee_albert import AlbertForSequenceClassificationWithPabee
-from pabee.modeling_pabee_bert import BertForSequenceClassificationWithPabee
-from transformers import (
-    WEIGHTS_NAME,
-    AdamW,
-    AlbertConfig,
-    AlbertTokenizer,
-    BertConfig,
-    BertTokenizer,
-    get_linear_schedule_with_warmup,
-)
-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
-
-
-try:
-    from torch.utils.tensorboard import SummaryWriter
-except ImportError:
-    from tensorboardX import SummaryWriter
-
-
-logger = logging.getLogger(__name__)
-
-MODEL_CLASSES = {
-    "bert": (BertConfig, BertForSequenceClassificationWithPabee, BertTokenizer),
-    "albert": (AlbertConfig, AlbertForSequenceClassificationWithPabee, AlbertTokenizer),
-}
-
-
-def set_seed(args):
-    random.seed(args.seed)
-    np.random.seed(args.seed)
-    torch.manual_seed(args.seed)
-    if args.n_gpu > 0:
-        torch.cuda.manual_seed_all(args.seed)
-
-
-def train(args, train_dataset, model, tokenizer):
-    """ Train the model """
-    if args.local_rank in [-1, 0]:
-        tb_writer = SummaryWriter()
-
-    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
-    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
-    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
-
-    if args.max_steps > 0:
-        t_total = args.max_steps
-        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
-    else:
-        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
-
-    # Prepare optimizer and schedule (linear warmup and decay)
-    no_decay = ["bias", "LayerNorm.weight"]
-    optimizer_grouped_parameters = [
-        {
-            "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
-            "weight_decay": args.weight_decay,
-        },
-        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
-    ]
-
-    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(
-        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
-    )
-
-    # Check if saved optimizer or scheduler states exist
-    if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
-        os.path.join(args.model_name_or_path, "scheduler.pt")
-    ):
-        # Load in optimizer and scheduler states
-        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
-        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
-
-    if args.fp16:
-        try:
-            from apex import amp
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
-        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
-
-    # multi-gpu training (should be after apex fp16 initialization)
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    # Distributed training (should be after apex fp16 initialization)
-    if args.local_rank != -1:
-        model = torch.nn.parallel.DistributedDataParallel(
-            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True,
-        )
-
-    # Train!
-    logger.info("***** Running training *****")
-    logger.info("  Num examples = %d", len(train_dataset))
-    logger.info("  Num Epochs = %d", args.num_train_epochs)
-    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
-    logger.info(
-        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
-        args.train_batch_size
-        * args.gradient_accumulation_steps
-        * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
-    )
-    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
-    logger.info("  Total optimization steps = %d", t_total)
-
-    global_step = 0
-    epochs_trained = 0
-    steps_trained_in_current_epoch = 0
-    # Check if continuing training from a checkpoint
-    if os.path.exists(args.model_name_or_path):
-        # set global_step to gobal_step of last saved checkpoint from model path
-        global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
-        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
-        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
-
-        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
-        logger.info("  Continuing training from epoch %d", epochs_trained)
-        logger.info("  Continuing training from global step %d", global_step)
-        logger.info(
-            "  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch,
-        )
-
-    tr_loss, logging_loss = 0.0, 0.0
-    model.zero_grad()
-    train_iterator = trange(
-        epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0],
-    )
-    set_seed(args)  # Added here for reproductibility
-    for _ in train_iterator:
-        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
-        for step, batch in enumerate(epoch_iterator):
-
-            # Skip past any already trained steps if resuming training
-            if steps_trained_in_current_epoch > 0:
-                steps_trained_in_current_epoch -= 1
-                continue
-
-            model.train()
-            batch = tuple(t.to(args.device) for t in batch)
-            inputs = {
-                "input_ids": batch[0],
-                "attention_mask": batch[1],
-                "labels": batch[3],
-            }
-            inputs["token_type_ids"] = batch[2]
-            outputs = model(**inputs)
-            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
-
-            if args.n_gpu > 1:
-                loss = loss.mean()  # mean() to average on multi-gpu parallel training
-            if args.gradient_accumulation_steps > 1:
-                loss = loss / args.gradient_accumulation_steps
-
-            if args.fp16:
-                with amp.scale_loss(loss, optimizer) as scaled_loss:
-                    scaled_loss.backward()
-            else:
-                loss.backward()
-
-            tr_loss += loss.item()
-            if (step + 1) % args.gradient_accumulation_steps == 0:
-                if args.fp16:
-                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
-                else:
-                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
-
-                optimizer.step()
-                scheduler.step()  # Update learning rate schedule
-                model.zero_grad()
-                global_step += 1
-
-                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
-                    logs = {}
-                    if (
-                        args.local_rank == -1 and args.evaluate_during_training
-                    ):  # Only evaluate when single GPU otherwise metrics may not average well
-                        results = evaluate(args, model, tokenizer)
-                        for key, value in results.items():
-                            eval_key = "eval_{}".format(key)
-                            logs[eval_key] = value
-
-                    loss_scalar = (tr_loss - logging_loss) / args.logging_steps
-                    learning_rate_scalar = scheduler.get_lr()[0]
-                    logs["learning_rate"] = learning_rate_scalar
-                    logs["loss"] = loss_scalar
-                    logging_loss = tr_loss
-
-                    for key, value in logs.items():
-                        tb_writer.add_scalar(key, value, global_step)
-                    print(json.dumps({**logs, **{"step": global_step}}))
-
-                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
-                    # Save model checkpoint
-                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
-                    if not os.path.exists(output_dir):
-                        os.makedirs(output_dir)
-                    model_to_save = (
-                        model.module if hasattr(model, "module") else model
-                    )  # Take care of distributed/parallel training
-                    model_to_save.save_pretrained(output_dir)
-                    tokenizer.save_pretrained(output_dir)
-
-                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
-                    logger.info("Saving model checkpoint to %s", output_dir)
-
-                    torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
-                    torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
-                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
-
-            if args.max_steps > 0 and global_step > args.max_steps:
-                epoch_iterator.close()
-                break
-        if args.max_steps > 0 and global_step > args.max_steps:
-            train_iterator.close()
-            break
-
-    if args.local_rank in [-1, 0]:
-        tb_writer.close()
-
-    return global_step, tr_loss / global_step
-
-
-def evaluate(args, model, tokenizer, prefix=""):
-
-    # PABEE STATS
-    if args.model_type == "albert":
-        model.albert.reset_stats()
-    elif args.model_type == "bert":
-        model.bert.reset_stats()
-    else:
-        raise NotImplementedError()
-
-    # Loop to handle MNLI double evaluation (matched, mis-matched)
-    eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
-    eval_outputs_dirs = (args.output_dir, args.output_dir + "-MM") if args.task_name == "mnli" else (args.output_dir,)
-
-    results = {}
-    for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
-        eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)
-
-        if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(eval_output_dir)
-
-        args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
-        # Note that DistributedSampler samples randomly
-        eval_sampler = SequentialSampler(eval_dataset)
-        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
-
-        # multi-gpu eval
-        if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
-            model = torch.nn.DataParallel(model)
-
-        # Eval!
-        logger.info("***** Running evaluation {} *****".format(prefix))
-        logger.info("  Num examples = %d", len(eval_dataset))
-        logger.info("  Batch size = %d", args.eval_batch_size)
-        eval_loss = 0.0
-        nb_eval_steps = 0
-        preds = None
-        out_label_ids = None
-        for batch in tqdm(eval_dataloader, desc="Evaluating"):
-            model.eval()
-            batch = tuple(t.to(args.device) for t in batch)
-
-            with torch.no_grad():
-                inputs = {
-                    "input_ids": batch[0],
-                    "attention_mask": batch[1],
-                    "labels": batch[3],
-                }
-                inputs["token_type_ids"] = batch[2]
-                outputs = model(**inputs)
-                tmp_eval_loss, logits = outputs[:2]
-
-                eval_loss += tmp_eval_loss.mean().item()
-            nb_eval_steps += 1
-            if preds is None:
-                preds = logits.detach().cpu().numpy()
-                out_label_ids = inputs["labels"].detach().cpu().numpy()
-            else:
-                preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
-                out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
-
-        eval_loss = eval_loss / nb_eval_steps
-        if args.output_mode == "classification":
-            preds = np.argmax(preds, axis=1)
-        elif args.output_mode == "regression":
-            preds = np.squeeze(preds)
-        result = compute_metrics(eval_task, preds, out_label_ids)
-        results.update(result)
-
-        output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
-        with open(output_eval_file, "w") as writer:
-            logger.info("***** Eval results {} *****".format(prefix))
-            for key in sorted(result.keys()):
-                logger.info("  %s = %s", key, str(result[key]))
-                print("  %s = %s" % (key, str(result[key])))
-                writer.write("%s = %s\n" % (key, str(result[key])))
-
-    if args.eval_all_checkpoints:
-        if args.model_type == "albert":
-            model.albert.log_stats()
-        elif args.model_type == "bert":
-            model.bert.log_stats()
-        else:
-            raise NotImplementedError()
-
-    return results
-
-
-def load_and_cache_examples(args, task, tokenizer, evaluate=False):
-    if args.local_rank not in [-1, 0] and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    processor = processors[task]()
-    output_mode = output_modes[task]
-    # Load data features from cache or dataset file
-    cached_features_file = os.path.join(
-        args.data_dir,
-        "cached_{}_{}_{}_{}".format(
-            "dev" if evaluate else "train",
-            list(filter(None, args.model_name_or_path.split("/"))).pop(),
-            str(args.max_seq_length),
-            str(task),
-        ),
-    )
-    if os.path.exists(cached_features_file) and not args.overwrite_cache:
-        logger.info("Loading features from cached file %s", cached_features_file)
-        features = torch.load(cached_features_file)
-    else:
-        logger.info("Creating features from dataset file at %s", args.data_dir)
-        label_list = processor.get_labels()
-        if task in ["mnli", "mnli-mm"] and args.model_type in ["roberta", "xlmroberta"]:
-            # HACK(label indices are swapped in RoBERTa pretrained model)
-            label_list[1], label_list[2] = label_list[2], label_list[1]
-        examples = (
-            processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
-        )
-        features = convert_examples_to_features(
-            examples, tokenizer, label_list=label_list, max_length=args.max_seq_length, output_mode=output_mode,
-        )
-        if args.local_rank in [-1, 0]:
-            logger.info("Saving features into cached file %s", cached_features_file)
-            torch.save(features, cached_features_file)
-
-    if args.local_rank == 0 and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    # Convert to Tensors and build dataset
-    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
-    all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
-    all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
-    if output_mode == "classification":
-        all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
-    elif output_mode == "regression":
-        all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
-
-    dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
-    return dataset
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    # Required parameters
-    parser.add_argument(
-        "--data_dir",
-        default=None,
-        type=str,
-        required=True,
-        help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
-    )
-    parser.add_argument(
-        "--model_type",
-        default=None,
-        type=str,
-        required=True,
-        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
-    )
-    parser.add_argument(
-        "--model_name_or_path",
-        default=None,
-        type=str,
-        required=True,
-        help="Path to pre-trained model or shortcut name.",
-    )
-    parser.add_argument(
-        "--task_name",
-        default=None,
-        type=str,
-        required=True,
-        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()),
-    )
-    parser.add_argument(
-        "--output_dir",
-        default=None,
-        type=str,
-        required=True,
-        help="The output directory where the model predictions and checkpoints will be written.",
-    )
-    parser.add_argument(
-        "--patience", default="0", type=str, required=False,
-    )
-    parser.add_argument(
-        "--regression_threshold", default=0, type=float, required=False,
-    )
-
-    # Other parameters
-    parser.add_argument(
-        "--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name",
-    )
-    parser.add_argument(
-        "--tokenizer_name",
-        default="",
-        type=str,
-        help="Pretrained tokenizer name or path if not the same as model_name",
-    )
-    parser.add_argument(
-        "--cache_dir",
-        default="",
-        type=str,
-        help="Where do you want to store the pre-trained models downloaded from s3",
-    )
-    parser.add_argument(
-        "--max_seq_length",
-        default=128,
-        type=int,
-        help="The maximum total input sequence length after tokenization. Sequences longer "
-        "than this will be truncated, sequences shorter will be padded.",
-    )
-    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
-    parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
-    parser.add_argument(
-        "--evaluate_during_training", action="store_true", help="Run evaluation during training at each logging step.",
-    )
-    parser.add_argument(
-        "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model.",
-    )
-
-    parser.add_argument(
-        "--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.",
-    )
-    parser.add_argument(
-        "--per_gpu_eval_batch_size", default=1, type=int, help="Batch size per GPU/CPU for evaluation.",
-    )
-    parser.add_argument(
-        "--gradient_accumulation_steps",
-        type=int,
-        default=1,
-        help="Number of updates steps to accumulate before performing a backward/update pass.",
-    )
-    parser.add_argument(
-        "--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.",
-    )
-    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
-    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
-    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
-    parser.add_argument(
-        "--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.",
-    )
-    parser.add_argument(
-        "--max_steps",
-        default=-1,
-        type=int,
-        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
-    )
-    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
-
-    parser.add_argument("--logging_steps", type=int, default=500, help="Log every X updates steps.")
-    parser.add_argument(
-        "--save_steps", type=int, default=500, help="Save checkpoint every X updates steps.",
-    )
-    parser.add_argument(
-        "--eval_all_checkpoints",
-        action="store_true",
-        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
-    )
-    parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
-    parser.add_argument(
-        "--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory",
-    )
-    parser.add_argument(
-        "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets",
-    )
-    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
-
-    parser.add_argument(
-        "--fp16",
-        action="store_true",
-        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
-    )
-    parser.add_argument(
-        "--fp16_opt_level",
-        type=str,
-        default="O1",
-        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
-        "See details at https://nvidia.github.io/apex/amp.html",
-    )
-    parser.add_argument(
-        "--local_rank", type=int, default=-1, help="For distributed training: local_rank",
-    )
-    parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
-    parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
-    args = parser.parse_args()
-
-    if (
-        os.path.exists(args.output_dir)
-        and os.listdir(args.output_dir)
-        and args.do_train
-        and not args.overwrite_output_dir
-    ):
-        raise ValueError(
-            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
-                args.output_dir
-            )
-        )
-
-    # Setup distant debugging if needed
-    if args.server_ip and args.server_port:
-        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
-        import ptvsd
-
-        print("Waiting for debugger attach")
-        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
-        ptvsd.wait_for_attach()
-
-    # Setup CUDA, GPU & distributed training
-    if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
-    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
-        torch.cuda.set_device(args.local_rank)
-        device = torch.device("cuda", args.local_rank)
-        torch.distributed.init_process_group(backend="nccl")
-        args.n_gpu = 1
-    args.device = device
-
-    # Setup logging
-    logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
-        datefmt="%m/%d/%Y %H:%M:%S",
-        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
-    )
-    logger.warning(
-        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
-        args.local_rank,
-        device,
-        args.n_gpu,
-        bool(args.local_rank != -1),
-        args.fp16,
-    )
-
-    # Set seed
-    set_seed(args)
-
-    # Prepare GLUE task
-    args.task_name = args.task_name.lower()
-    if args.task_name not in processors:
-        raise ValueError("Task not found: %s" % (args.task_name))
-    processor = processors[args.task_name]()
-    args.output_mode = output_modes[args.task_name]
-    label_list = processor.get_labels()
-    num_labels = len(label_list)
-
-    if args.patience != "0" and args.per_gpu_eval_batch_size != 1:
-        raise ValueError("The eval batch size must be 1 with PABEE inference on.")
-
-    # Load pretrained model and tokenizer
-    if args.local_rank not in [-1, 0]:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    args.model_type = args.model_type.lower()
-    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(
-        args.config_name if args.config_name else args.model_name_or_path,
-        num_labels=num_labels,
-        finetuning_task=args.task_name,
-        cache_dir=args.cache_dir if args.cache_dir else None,
-    )
-    tokenizer = tokenizer_class.from_pretrained(
-        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
-        do_lower_case=args.do_lower_case,
-        cache_dir=args.cache_dir if args.cache_dir else None,
-    )
-    model = model_class.from_pretrained(
-        args.model_name_or_path,
-        from_tf=bool(".ckpt" in args.model_name_or_path),
-        config=config,
-        cache_dir=args.cache_dir if args.cache_dir else None,
-    )
-
-    if args.local_rank == 0:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    model.to(args.device)
-
-    print("Total Model Parameters:", sum(param.numel() for param in model.parameters()))
-    output_layers_param_num = sum(param.numel() for param in model.classifiers.parameters())
-    print("Output Layers Parameters:", output_layers_param_num)
-    single_output_layer_param_num = sum(param.numel() for param in model.classifiers[0].parameters())
-    print(
-        "Added Output Layers Parameters:", output_layers_param_num - single_output_layer_param_num,
-    )
-
-    logger.info("Training/evaluation parameters %s", args)
-
-    # Training
-    if args.do_train:
-        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
-        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
-        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
-
-    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
-    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
-        logger.info("Saving model checkpoint to %s", args.output_dir)
-        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
-        # They can then be reloaded using `from_pretrained()`
-        model_to_save = (
-            model.module if hasattr(model, "module") else model
-        )  # Take care of distributed/parallel training
-        model_to_save.save_pretrained(args.output_dir)
-        tokenizer.save_pretrained(args.output_dir)
-
-        # Good practice: save your training arguments together with the trained model
-        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
-
-        # Load a trained model and vocabulary that you have fine-tuned
-        model = model_class.from_pretrained(args.output_dir)
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
-        model.to(args.device)
-
-    # Evaluation
-    results = {}
-    if args.do_eval and args.local_rank in [-1, 0]:
-        patience_list = [int(x) for x in args.patience.split(",")]
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
-        checkpoints = [args.output_dir]
-        if args.eval_all_checkpoints:
-            checkpoints = list(
-                os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
-            )
-            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
-        logger.info("Evaluate the following checkpoints: %s", checkpoints)
-
-        for checkpoint in checkpoints:
-
-            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
-            prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""
-
-            model = model_class.from_pretrained(checkpoint)
-            model.to(args.device)
-
-            print(f"Evaluation for checkpoint {prefix}")
-            for patience in patience_list:
-                if args.model_type == "albert":
-                    model.albert.set_regression_threshold(args.regression_threshold)
-                    model.albert.set_patience(patience)
-                elif args.model_type == "bert":
-                    model.bert.set_regression_threshold(args.regression_threshold)
-                    model.bert.set_patience(patience)
-                else:
-                    raise NotImplementedError()
-                result = evaluate(args, model, tokenizer, prefix=prefix)
-                result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
-                results.update(result)
-    return results
-
-
-if __name__ == "__main__":
-    main()

examples/bert-loses-patience/test_run_glue_with_pabee.py

@@ -1,48 +0,0 @@
-import argparse
-import logging
-import sys
-import unittest
-from unittest.mock import patch
-
-import run_glue_with_pabee
-
-
-logging.basicConfig(level=logging.DEBUG)
-
-logger = logging.getLogger()
-
-
-def get_setup_file():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("-f")
-    args = parser.parse_args()
-    return args.f
-
-
-class PabeeTests(unittest.TestCase):
-    def test_run_glue(self):
-        stream_handler = logging.StreamHandler(sys.stdout)
-        logger.addHandler(stream_handler)
-
-        testargs = """
-            run_glue_with_pabee.py
-            --model_type albert
-            --model_name_or_path albert-base-v2
-            --data_dir ./tests/fixtures/tests_samples/MRPC/
-            --task_name mrpc
-            --do_train
-            --do_eval
-            --output_dir ./tests/fixtures/tests_samples/temp_dir
-            --per_gpu_train_batch_size=2
-            --per_gpu_eval_batch_size=1
-            --learning_rate=2e-5
-            --max_steps=50
-            --warmup_steps=2
-            --overwrite_output_dir
-            --seed=42
-            --max_seq_length=128
-            """.split()
-        with patch.object(sys, "argv", testargs):
-            result = run_glue_with_pabee.main()
-            for value in result.values():
-                self.assertGreaterEqual(value, 0.75)

examples/lightning_base.py

@@ -8,7 +8,6 @@ from typing import Any, Dict
 import numpy as np
 import pytorch_lightning as pl
 import torch
-from pytorch_lightning.utilities import rank_zero_info, rank_zero_only
 
 from transformers import (
     AdamW,
@@ -61,9 +60,10 @@ class BaseTransformer(pl.LightningModule):
         model=None,
         **config_kwargs
     ):
-        """Initialize a model, tokenizer and config."""
+        "Initialize a model."
+
         super().__init__()
-        self.hparams = hparams  # TODO: move to self.save_hyperparameters()
+        self.hparams = hparams
         self.step_count = 0
         self.tfmr_ckpts = {}
         self.output_dir = Path(self.hparams.output_dir)
@@ -84,8 +84,8 @@ class BaseTransformer(pl.LightningModule):
             )
         else:
             self.tokenizer: PreTrainedTokenizer = tokenizer
-        self.model_type = MODEL_MODES[mode]
         if model is None:
+            self.model_type = MODEL_MODES[mode]
             self.model = self.model_type.from_pretrained(
                 self.hparams.model_name_or_path,
                 from_tf=bool(".ckpt" in self.hparams.model_name_or_path),
@@ -93,13 +93,18 @@ class BaseTransformer(pl.LightningModule):
                 cache_dir=cache_dir,
             )
         else:
+            self.model_type = None
             self.model = model
 
     def load_hf_checkpoint(self, *args, **kwargs):
         self.model = self.model_type.from_pretrained(*args, **kwargs)
 
+    def is_logger(self):
+        return self.trainer.proc_rank <= 0
+
     def configure_optimizers(self):
         "Prepare optimizer and schedule (linear warmup and decay)"
+
         model = self.model
         no_decay = ["bias", "LayerNorm.weight"]
         optimizer_grouped_parameters = [
@@ -132,7 +137,7 @@ class BaseTransformer(pl.LightningModule):
     def test_step(self, batch, batch_nb):
         return self.validation_step(batch, batch_nb)
 
-    def test_epoch_end(self, outputs):
+    def test_end(self, outputs):
         return self.validation_end(outputs)
 
     def train_dataloader(self):
@@ -203,7 +208,6 @@ class BaseTransformer(pl.LightningModule):
         parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
         parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
         parser.add_argument("--warmup_steps", default=500, type=int, help="Linear warmup over warmup_steps.")
-        parser.add_argument("--num_workers", default=4, type=int, help="kwarg passed to DataLoader")
         parser.add_argument(
             "--num_train_epochs", default=3, type=int, help="Total number of training epochs to perform."
         )
@@ -213,26 +217,28 @@ class BaseTransformer(pl.LightningModule):
 
 
 class LoggingCallback(pl.Callback):
-    @rank_zero_only
     def on_validation_end(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
-        rank_zero_info("***** Validation results *****")
-        metrics = trainer.callback_metrics
-        # Log results
-        for key in sorted(metrics):
-            if key not in ["log", "progress_bar"]:
-                rank_zero_info("{} = {}\n".format(key, str(metrics[key])))
-
-    @rank_zero_only
-    def on_test_end(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
-        logger.info("***** Test results *****")
-        metrics = trainer.callback_metrics
-        # Log and save results to file
-        output_test_results_file = os.path.join(pl_module.hparams.output_dir, "test_results.txt")
-        with open(output_test_results_file, "w") as writer:
+        logger.info("***** Validation results *****")
+        if pl_module.is_logger():
+            metrics = trainer.callback_metrics
+            # Log results
             for key in sorted(metrics):
                 if key not in ["log", "progress_bar"]:
                     logger.info("{} = {}\n".format(key, str(metrics[key])))
-                    writer.write("{} = {}\n".format(key, str(metrics[key])))
+
+    def on_test_end(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
+        logger.info("***** Test results *****")
+
+        if pl_module.is_logger():
+            metrics = trainer.callback_metrics
+
+            # Log and save results to file
+            output_test_results_file = os.path.join(pl_module.hparams.output_dir, "test_results.txt")
+            with open(output_test_results_file, "w") as writer:
+                for key in sorted(metrics):
+                    if key not in ["log", "progress_bar"]:
+                        logger.info("{} = {}\n".format(key, str(metrics[key])))
+                        writer.write("{} = {}\n".format(key, str(metrics[key])))
 
 
 def add_generic_args(parser, root_dir) -> None:

examples/requirements.txt

@@ -5,7 +5,7 @@ psutil
 sacrebleu
 rouge-score
 tensorflow_datasets
-pytorch-lightning==0.8.1
+pytorch-lightning==0.7.6
 matplotlib
 git-python==1.0.3
 faiss

examples/summarization/README.md

@@ -44,7 +44,7 @@ export me=`git config user.name`
 ```
 
 Tips:
-- 1 epoch at batch size 1 for bart-large takes 24 hours, requires 13GB GPU RAM with fp16 on an NVIDIA-V100.
+- 1 epoch at batch size 1 for bart-large takes 24 hours, requires 13GB GPU RAM with fp16 on an NVIDIA-V100. 
 - try `bart-base`, `--freeze_encoder` or `--freeze_embeds` for faster training/larger batch size.  (3hr/epoch with bs=8, see below)
 - `fp16_opt_level=O1` (the default works best).
 - If you are finetuning on your own dataset, start from `bart-large-cnn` if you want long summaries and `bart-large-xsum` if you want short summaries.

examples/summarization/bart/rouge_tracker.py

@@ -1,103 +0,0 @@
-from pathlib import Path
-
-import numpy as np
-import pandas as pd
-import torch
-
-from durbango import lmap, tqdm_nice
-from transformers import BartTokenizer
-
-
-try:
-    from finetune import calculate_rouge
-except ImportError:
-    from .finetune import calculate_rouge
-
-
-LOGDIR = Path("examples/summarization/dbart/logs/").absolute()
-DATA_DIR = Path("examples/summarization/dbart/cnn_dm").absolute()
-
-
-def rouge_files(src_file: Path, tgt_file: Path):
-    src = lmap(str.strip, list(src_file.open().readlines()))
-    tgt = lmap(str.strip, list(tgt_file.open().readlines()))
-    return calculate_rouge(src, tgt)
-
-
-def read_gens(exp_name, split="test", n=None):
-    if Path(exp_name).exists():
-        expdir = exp_name
-    else:
-        expdir = LOGDIR / exp_name
-    assert expdir.exists(), expdir
-    paths = list(expdir.glob(f"{split}_generations*.txt"))
-    assert paths
-    path = paths[0]
-    lns = lmap(str.strip, list(path.open().readlines()))
-    if n is not None:
-        return lns[:n]
-    return lns
-
-
-def load_cpu(p):
-    return torch.load(p, map_location="cpu")
-
-
-class RougeTracker:
-    def __init__(self, csv_path="rouge_test_df.csv", logdir=LOGDIR, data_dir=DATA_DIR):
-        try:
-            self.df = pd.read_csv(csv_path, index_col=0)
-        except FileNotFoundError:
-            self.df = pd.DataFrame()
-        self.logdir = logdir
-        test_gt = lmap(str.strip, Path(data_dir / "test.target").open().readlines())
-        test_gt = lmap(str.strip, test_gt)
-        self.gt = test_gt  # {'test': test_gt}
-        self.tokenizer = BartTokenizer.from_pretrained("facebook/bart-large-cnn")
-        self.csv_path = csv_path
-        self.new_results = pd.DataFrame()
-
-    @property
-    def finished_experiments(self):
-        return [p.parent.name for p in list(self.logdir.glob("*/test_generations*.txt"))]
-
-    def tok_len(self, strang):
-        return len(self.tokenizer.encode(strang))
-
-    def read_all_gens(self,):
-        GENS = {}
-        for f in self.finished_experiments:
-            GENS[f] = read_gens(f)
-        return GENS
-
-    def score(self, gens, k, all_stats=False):
-        rouge_raw = calculate_rouge(gens, self.gt, all_stats=all_stats)
-        lens = np.mean(lmap(self.tok_len, gens))
-        return dict(avg_len=lens, exp_name=k, **rouge_raw)
-
-    @property
-    def new_experiments(self):
-        possible = set(self.finished_experiments).difference(self.df.index)
-        to_score = set()
-        for p in possible:
-            gens = read_gens(p)
-            if len(gens) == len(self.gt):
-                to_score.add(p)
-        return to_score
-
-    def update(self):
-        records = []
-        to_score = self.new_experiments
-        for exp_name in tqdm_nice(to_score, desc="Rouge Update"):
-            gens = read_gens(exp_name)
-            if len(gens) != len(self.gt):
-                continue
-            records.append(self.score(gens, exp_name))
-        if not records:
-            return self.df
-        new_df = (
-            pd.DataFrame(records).rename(columns=lambda x: x.replace("rouge", "R")).set_index("exp_name").astype(float)
-        )
-        self.new_results = new_df
-        self.df = pd.concat([self.df, new_df]).dsort("R2")
-        return self.df

examples/summarization/callbacks.py

@@ -19,11 +19,12 @@ logger = logging.getLogger(__name__)
 
 
 class Seq2SeqLoggingCallback(pl.Callback):
-    @rank_zero_only
     def _write_logs(
         self, trainer: pl.Trainer, pl_module: pl.LightningModule, type_path: str, save_generations=True
     ) -> None:
         logger.info(f"***** {type_path} results at step {trainer.global_step:05d} *****")
+        if not pl_module.is_logger():
+            return
         metrics = trainer.callback_metrics
         trainer.logger.log_metrics({k: v for k, v in metrics.items() if k not in ["log", "progress_bar", "preds"]})
         # Log results

examples/summarization/distillation.py

@@ -25,8 +25,6 @@ try:
         any_requires_grad,
     )
     from .finetune import main as ft_main
-    from .replacement_scheduler import LinearReplacementScheduler
-
 except ImportError:
     from finetune import SummarizationModule
     from finetune import main as ft_main
@@ -39,62 +37,6 @@ except ImportError:
         assert_all_frozen,
         any_requires_grad,
     )
-    from replacement_scheduler import LinearReplacementScheduler
-
-
-class TheseusDistiller(SummarizationModule):
-    def __init__(self, hparams):
-
-        assert Path(hparams.data_dir).exists()
-        # config = BartConfig.from_pretrained(hparams.model_name_or_path, student_encoder_layers=hparams.student_encoder_layers, student_decoder_layers=hparams.student_decoder_layers, replacing_rate=hparams.theseus_replace_rate)
-        model = BartForConditionalGeneration.from_pretrained(
-            hparams.model_name_or_path,
-            student_encoder_layers=hparams.student_encoder_layers,
-            student_decoder_layers=hparams.student_decoder_layers,
-            replacing_rate=hparams.theseus_replace_rate,
-        )
-        super().__init__(hparams, model=model)
-        self.different_encoder: bool = hparams.student_encoder_layers != self.model.config.encoder_layers
-        self.different_decoder: bool = hparams.student_decoder_layers != self.model.config.decoder_layers
-
-        if hparams.theseus_init_copy:
-            hparams.d_layers_to_copy = get_layers_to_copy(
-                hparams.student_decoder_layers, self.model.config.decoder_layers, strategy=hparams.init_strategy,
-            )
-            hparams.e_layers_to_copy: List = get_layers_to_copy(
-                hparams.student_encoder_layers, self.model.config.encoder_layers, strategy=hparams.init_strategy
-            )
-            if self.different_decoder:
-                copy_layers(
-                    self.model.model.decoder.layers, self.model.model.decoder.scc_layers, hparams.d_layers_to_copy
-                )
-            if self.different_encoder:
-                copy_layers(
-                    self.model.model.encoder.layers, self.model.model.encoder.scc_layers, hparams.e_layers_to_copy
-                )
-        else:
-            hparams.e_layers_to_copy, hparams.d_layers_to_copy = None, None
-
-        self.replace_scheduler_encoder = LinearReplacementScheduler(self.model.model.encoder, 0.6)
-        self.replace_scheduler_decoder = LinearReplacementScheduler(self.model.model.decoder, 0.6)
-        freeze_params(self.model.model.encoder.layers)  # Test
-        freeze_params(self.model.model.decoder.layers)
-
-    def optimizer_step(self, *args, **kwargs) -> None:
-        self.replace_scheduler_encoder.step()
-        replace_rate = self.replace_scheduler_decoder.step()
-        self.logger.log_metrics({"replace_rate": replace_rate})
-        super().optimizer_step(*args, **kwargs)
-
-    def copy_to_student(self, d_layers_to_copy, e_layers_to_copy, hparams, student, teacher):
-        if teacher.config.model_type == "t5":
-            return self.copy_t5_to_student(d_layers_to_copy, e_layers_to_copy, hparams, student, teacher)
-        self.different_encoder: bool = hparams.student_encoder_layers != teacher.config.encoder_layers
-        self.different_decoder = hparams.student_decoder_layers != teacher.config.decoder_layers
-        if self.different_decoder:
-            copy_layers(teacher.model.decoder.layers, student.model.decoder.layers, d_layers_to_copy)
-        if self.different_encoder:
-            copy_layers(teacher.model.encoder.layers, student.model.encoder.layers, e_layers_to_copy)
 
 
 class SummarizationDistiller(SummarizationModule):
@@ -107,9 +49,6 @@ class SummarizationDistiller(SummarizationModule):
 
         super().__init__(hparams, model=student, config=student_cfg)
         self.teacher = teacher
-        if isinstance(self.teacher, BartForConditionalGeneration):
-            assert teacher.model.encoder.scc_layers is None
-            assert self.model.model.encoder.scc_layers is None
         use_task_specific_params(self.teacher, "summarization")
         freeze_params(self.teacher)
         self.sanity_check_gradients()
@@ -140,14 +79,8 @@ class SummarizationDistiller(SummarizationModule):
             "decoder_layers": hparams.student_decoder_layers,
             "encoder_layers": hparams.student_encoder_layers,
         }
-
-        d_layers_to_copy = get_layers_to_copy(
-            student_updates["decoder_layers"], teacher.config.decoder_layers, strategy=hparams.init_strategy
-        )
-        e_layers_to_copy: List = get_layers_to_copy(
-            student_updates["encoder_layers"], teacher.config.encoder_layers, strategy=hparams.init_strategy
-        )
-
+        d_layers_to_copy = get_layers_to_copy(student_updates["decoder_layers"], teacher.config.decoder_layers)
+        e_layers_to_copy: List = get_layers_to_copy(student_updates["encoder_layers"], teacher.config.encoder_layers)
         hparams.d_layer_to_copy = d_layers_to_copy
         hparams.e_layer_to_copy = e_layers_to_copy
         kw = teacher.config.to_diff_dict()
@@ -247,13 +180,18 @@ class SummarizationDistiller(SummarizationModule):
         # parser.add_argument("--alpha_cos", default=0.0, type=float)
         parser.add_argument("--alpha_encoder_loss", default=0.0, type=float)
         parser.add_argument("--alpha_hid", default=0.0, type=float, 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)
-        parser.add_argument("--theseus_replace_rate", type=float, default=0.0)
-        parser.add_argument("--theseus_init_copy", action="store_true")
-        parser.add_argument("--init_strategy", type=str, default="alternate", choices=["alternate", "top", "bottom"])
+        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,
+        )
+        parser.add_argument(  # TODO: remove
+            "--enc_only", action="store_true", default=False,
+        )
         return parser
 
     def _step(self, batch):
@@ -333,7 +271,6 @@ class SummarizationDistiller(SummarizationModule):
 
 class T5SummarizationDistiller(SummarizationDistiller):
     def pre_init(self, hparams):
-        raise NotImplementedError("T5 Distillation does not work yet")
         teacher = T5ForConditionalGeneration.from_pretrained(hparams.teacher)
         n_layer = hparams.student_decoder_layers
         assert n_layer == hparams.student_encoder_layers  # TODO(SS): relax this
@@ -440,12 +377,13 @@ class T5SummarizationDistiller(SummarizationDistiller):
 
 def create_module(args):
     t5 = "t5" in args.model_name_or_path
-    if args.no_teacher and args.theseus_replace_rate == 0:
+    if args.no_teacher:
+        assert not args.enc_only
         module_cls = SummarizationModule
-    elif args.no_teacher and args.theseus_replace_rate > 0:
-        module_cls = TheseusDistiller
     elif t5:
         module_cls = T5SummarizationDistiller
+    elif args.enc_only:
+        raise ValueError("Deleted that")
     else:
         module_cls = SummarizationDistiller
     args.setup_cls: str = module_cls.__name__
@@ -476,30 +414,21 @@ def evaluate_checkpoint(ckpt_path: Path, dest_dir=None):
     trainer.test(model)
 
 
-DISTILBERT_ALTERNATE_PATTERN = {  # maps # layers in student -> which teacher layers to copy
-    6: [0, 2, 4, 7, 9, 11],
-    1: [11],
-    3: [0, 6, 11],
-    2: [0, 11],
-    4: [0, 4, 8, 11],
-    9: [0, 1, 2, 4, 5, 7, 9, 10, 11],
-    12: list(range(12)),
-}
-
-
-def get_layers_to_copy(n_student_layers: int, n_teacher_layers: int, strategy="alternate") -> List:
-    all_layers = list(range(n_teacher_layers))
-    if strategy == "alternate":
-        if n_teacher_layers == 12:
-            return DISTILBERT_ALTERNATE_PATTERN[n_student_layers]
-        else:
-            return all_layers[::2][:n_student_layers]
-    elif strategy == "bottom":
-        return all_layers[:n_student_layers]
-    elif strategy == "top":
-        return all_layers[-n_student_layers:]
+def get_layers_to_copy(n_to_get, tot):
+    all_layers = list(range(tot))
+    if tot == 12:  # Alternating for special cases
+        layers_to_copy = {  # maps # layers in student -> which teacher layers to copy
+            6: [0, 2, 4, 7, 9, 11],
+            1: [11],
+            3: [0, 6, 11],
+            2: [0, 11],
+            4: [0, 4, 8, 11],
+            9: [0, 1, 2, 4, 5, 7, 9, 10, 11],
+            12: all_layers,
+        }
+        return layers_to_copy[n_to_get]
     else:
-        raise ValueError(f"layer copy strategy {strategy}  not supported")
+        return all_layers[:n_to_get]
 
 
 def distill_main(args):

examples/summarization/finetune.py

@@ -9,14 +9,12 @@ from typing import Dict, List, Tuple
 import numpy as np
 import pytorch_lightning as pl
 import torch
-from pytorch_lightning.loggers import WandbLogger
 from torch.utils.data import DataLoader
 
 from lightning_base import BaseTransformer, add_generic_args, generic_train
 from transformers import get_linear_schedule_with_warmup
 
 
-WANDB_PROJ_NAME = "transformers_fork-examples_summarization_bart"
 try:
     from .utils import (
         use_task_specific_params,
@@ -54,7 +52,6 @@ class SummarizationModule(BaseTransformer):
     loss_names = ["loss"]
 
     def __init__(self, hparams, **kwargs):
-        assert Path(hparams.data_dir).exists()
         super().__init__(hparams, num_labels=None, mode=self.mode, **kwargs)
         use_task_specific_params(self.model, "summarization")
         save_git_info(self.hparams.output_dir)
@@ -82,12 +79,13 @@ class SummarizationModule(BaseTransformer):
         }
         assert self.target_lens["train"] <= self.target_lens["val"], f"target_lens: {self.target_lens}"
         assert self.target_lens["train"] <= self.target_lens["test"], f"target_lens: {self.target_lens}"
-        if not self.hparams.unfreeze_embeds:
+
+        if self.hparams.freeze_embeds:
             self.freeze_embeds()
         if self.hparams.freeze_encoder:
             freeze_params(self.model.model.encoder)  # TODO: this will break for t5
         self.hparams.git_sha = get_git_info()["repo_sha"]
-        self.num_workers = hparams.num_workers
+        self.num_workers = 4 if self.hparams.gpus <= 1 else None  # passing num_workers breaks lightning for multigpu
 
     def freeze_embeds(self):
         """Freeze token embeddings and positional embeddings for bart, just token embeddings for t5."""
@@ -128,7 +126,7 @@ class SummarizationModule(BaseTransformer):
     def validation_step(self, batch, batch_idx) -> Dict:
         return self._generative_step(batch)
 
-    def validation_epoch_end(self, outputs, prefix="val") -> Dict:
+    def validation_end(self, outputs, prefix="val") -> Dict:
         self.step_count += 1
         losses = {k: torch.stack([x[k] for x in outputs]).mean() for k in self.loss_names}
         loss = losses["loss"]
@@ -146,12 +144,14 @@ class SummarizationModule(BaseTransformer):
         self.metrics[prefix].append(metrics)
         pickle_save(self.metrics, self.metrics_save_path)
 
-    def _generative_step(self, batch: dict) -> dict:
+    def _generative_step(self, batch):
         pad_token_id = self.tokenizer.pad_token_id
         source_ids, source_mask, y = SummarizationDataset.trim_seq2seq_batch(batch, pad_token_id)
+        # TODO(SS): task specific params
+
         t0 = time.time()
         generated_ids = self.model.generate(input_ids=source_ids, attention_mask=source_mask, use_cache=True,)
-        gen_time = time.time() - t0 / source_ids.shape[0]
+        gen_time = time.time() - t0
         preds = self.ids_to_clean_text(generated_ids)
         target = self.ids_to_clean_text(y)
         loss_tensors = self._step(batch)
@@ -164,8 +164,24 @@ class SummarizationModule(BaseTransformer):
     def test_step(self, batch, batch_idx):
         return self._generative_step(batch)
 
+    def test_end(self, outputs):
+        return self.validation_end(outputs, prefix="test")
+
     def test_epoch_end(self, outputs):
-        return self.validation_epoch_end(outputs, prefix="test")
+        output_test_predictions_file = os.path.join(self.hparams.output_dir, "test_predictions.txt")
+        output_test_targets_file = os.path.join(self.hparams.output_dir, "test_targets.txt")
+        # write predictions and targets for later rouge evaluation.
+        with open(output_test_predictions_file, "w+") as p_writer, open(output_test_targets_file, "w+") as t_writer:
+            for output_batch in outputs:
+                p_writer.writelines(s + "\n" for s in output_batch["preds"])
+                t_writer.writelines(s + "\n" for s in output_batch["target"])
+            p_writer.close()
+            t_writer.close()
+
+        return self.test_end(outputs)
+
+    def validation_epoch_end(self, outputs):
+        self.validation_end(outputs, "val")
 
     def get_dataset(self, type_path) -> SummarizationDataset:
         n_obs = self.n_obs[type_path]
@@ -255,9 +271,9 @@ class SummarizationModule(BaseTransformer):
             help="The input data dir. Should contain train.source, train.target, val.source, val.target, test.source, test.target",
         )
         parser.add_argument("--freeze_encoder", action="store_true")
-        parser.add_argument("--unfreeze_embeds", action="store_true")
+        parser.add_argument("--freeze_embeds", action="store_true")
         parser.add_argument("--sortish_sampler", action="store_true", default=False)
-        parser.add_argument("--logger", type=str, choices=["default", "wandb", "wandb_shared"], default="wandb")
+        parser.add_argument("--logger", type=str, choices=["default", "wandb", "wandb_shared"], default="default")
         parser.add_argument("--n_train", type=int, default=-1, required=False, help="# examples. -1 means use all.")
         parser.add_argument("--n_val", type=int, default=500, required=False, help="# examples. -1 means use all.")
         parser.add_argument("--n_test", type=int, default=-1, required=False, help="# examples. -1 means use all.")
@@ -280,10 +296,6 @@ def main(args, model=None) -> SummarizationModule:
     elif args.logger == "wandb":
         from pytorch_lightning.loggers import WandbLogger
 
-        logger = WandbLogger(name=model.output_dir.name, project=WANDB_PROJ_NAME)
-    elif args.logger == "wandb_shared":
-        from pytorch_lightning.loggers import WandbLogger
-
         logger = WandbLogger(name=model.output_dir.name)
     elif args.logger == "wandb_shared":
         from pytorch_lightning.loggers import WandbLogger
@@ -298,7 +310,6 @@ def main(args, model=None) -> SummarizationModule:
         logger=logger,
         # TODO: early stopping callback seems messed up
     )
-    pickle_save(model.hparams, model.output_dir / "hparams.pkl")
     if not args.do_predict:
         return model
 

examples/summarization/finetune.sh

@@ -1,11 +1,14 @@
+export OUTPUT_DIR=bart_cnn_finetune
+
+# Make output directory if it doesn't exist
+mkdir -p $OUTPUT_DIR
+
 # Add parent directory to python path to access lightning_base.py
 export PYTHONPATH="../":"${PYTHONPATH}"
 
 
 # --model_name_or_path=t5-base for t5
 
-# the proper usage is documented in the README
-
 python finetune.py \
     --model_name_or_path=facebook/bart-large \
     --learning_rate=3e-5 \

examples/summarization/finetune_bart_tiny.sh

@@ -16,13 +16,14 @@ mkdir -p $OUTPUT_DIR
 export PYTHONPATH="../":"${PYTHONPATH}"
 python finetune.py \
 --data_dir=cnn_tiny/ \
+--model_type=bart \
 --model_name_or_path=sshleifer/bart-tiny-random \
 --learning_rate=3e-5 \
 --train_batch_size=2 \
 --eval_batch_size=2 \
 --output_dir=$OUTPUT_DIR \
 --num_train_epochs=1  \
---gpus=0 \
+--n_gpu=0 \
 --do_train $@
 
 rm -rf cnn_tiny

examples/summarization/finetune_t5.sh

@@ -11,6 +11,7 @@ export PYTHONPATH="../":"${PYTHONPATH}"
 python finetune.py \
 --data_dir=./cnn-dailymail/cnn_dm \
 --model_name_or_path=t5-large \
+--model_type=t5
 --learning_rate=3e-5 \
 --train_batch_size=4 \
 --eval_batch_size=4 \

examples/summarization/modeling_bert_of_theseus.py

@@ -1,639 +0,0 @@
-"""PyTorch BERT-of-Theseus model. """
-
-from __future__ import absolute_import, division, print_function, unicode_literals
-
-import logging
-
-import torch
-from torch import nn
-from torch.distributions.bernoulli import Bernoulli
-from torch.nn import CrossEntropyLoss, MSELoss
-
-from transformers.configuration_bert import BertConfig
-from transformers.modeling_bert import (
-    ACT2FN,
-    BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
-    BertAttention,
-    BertEmbeddings,
-    BertIntermediate,
-    BertLayer,
-    BertLayerNorm,
-    BertLMPredictionHead,
-    BertOnlyMLMHead,
-    BertOnlyNSPHead,
-    BertOutput,
-    BertPooler,
-    BertPredictionHeadTransform,
-    BertPreTrainingHeads,
-    BertSelfAttention,
-    BertSelfOutput,
-    gelu,
-    gelu_new,
-    load_tf_weights_in_bert,
-    mish,
-    swish,
-)
-from transformers.modeling_utils import PreTrainedModel, prune_linear_layer
-
-
-logger = logging.getLogger(__name__)
-
-
-class BertEncoder(nn.Module):
-    def __init__(self, config, scc_n_layer=6):
-        super(BertEncoder, self).__init__()
-        self.prd_n_layer = config.num_hidden_layers
-        self.scc_n_layer = scc_n_layer
-        assert self.prd_n_layer % self.scc_n_layer == 0
-        self.compress_ratio = self.prd_n_layer // self.scc_n_layer
-        self.bernoulli = None
-        self.output_attentions = config.output_attentions
-        self.output_hidden_states = config.output_hidden_states
-        self.layer = nn.ModuleList([BertLayer(config) for _ in range(self.prd_n_layer)])
-        self.scc_layer = nn.ModuleList([BertLayer(config) for _ in range(self.scc_n_layer)])
-
-    def set_replacing_rate(self, replacing_rate):
-        if not 0 < replacing_rate <= 1:
-            raise Exception("Replace rate must be in the range (0, 1]!")
-        self.bernoulli = Bernoulli(torch.tensor([replacing_rate]))
-
-    def forward(
-        self,
-        hidden_states,
-        attention_mask=None,
-        head_mask=None,
-        encoder_hidden_states=None,
-        encoder_attention_mask=None,
-    ):
-        all_hidden_states = ()
-        all_attentions = ()
-        if self.training:
-            inference_layers = []
-            for i in range(self.scc_n_layer):
-                if self.bernoulli.sample() == 1:  # REPLACE
-                    inference_layers.append(self.scc_layer[i])
-                else:  # KEEP the original
-                    for offset in range(self.compress_ratio):
-                        inference_layers.append(self.layer[i * self.compress_ratio + offset])
-
-        else:  # inference with compressed model
-            inference_layers = self.scc_layer
-
-        for i, layer_module in enumerate(inference_layers):
-            if self.output_hidden_states:
-                all_hidden_states = all_hidden_states + (hidden_states,)
-
-            layer_outputs = layer_module(
-                hidden_states, attention_mask, head_mask[i], encoder_hidden_states, encoder_attention_mask
-            )
-            hidden_states = layer_outputs[0]
-
-            if self.output_attentions:
-                all_attentions = all_attentions + (layer_outputs[1],)
-
-        # Add last layer
-        if self.output_hidden_states:
-            all_hidden_states = all_hidden_states + (hidden_states,)
-
-        outputs = (hidden_states,)
-        if self.output_hidden_states:
-            outputs = outputs + (all_hidden_states,)
-        if self.output_attentions:
-            outputs = outputs + (all_attentions,)
-        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
-
-
-class BertPreTrainedModel(PreTrainedModel):
-    config_class = BertConfig
-    pretrained_model_archive_map = BERT_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_tf_weights = load_tf_weights_in_bert
-    base_model_prefix = "bert"
-
-    def _init_weights(self, module):
-        """ Initialize the weights """
-        if isinstance(module, (nn.Linear, nn.Embedding)):
-            # Slightly different from the TF version which uses truncated_normal for initialization
-            # cf https://github.com/pytorch/pytorch/pull/5617
-            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
-        elif isinstance(module, BertLayerNorm):
-            module.bias.data.zero_()
-            module.weight.data.fill_(1.0)
-        if isinstance(module, nn.Linear) and module.bias is not None:
-            module.bias.data.zero_()
-
-
-class BertModel(BertPreTrainedModel):
-    def __init__(self, config):
-        super(BertModel, self).__init__(config)
-        self.config = config
-
-        self.embeddings = BertEmbeddings(config)
-        self.encoder = BertEncoder(config)
-        self.pooler = BertPooler(config)
-
-        self.init_weights()
-
-    def get_input_embeddings(self):
-        return self.embeddings.word_embeddings
-
-    def set_input_embeddings(self, value):
-        self.embeddings.word_embeddings = value
-
-    def _prune_heads(self, heads_to_prune):
-        """ Prunes heads of the model.
-            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
-            See base class PreTrainedModel
-        """
-        for layer, heads in heads_to_prune.items():
-            self.encoder.layer[layer].attention.prune_heads(heads)
-
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        encoder_hidden_states=None,
-        encoder_attention_mask=None,
-    ):
-        if input_ids is not None and inputs_embeds is not None:
-            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
-        elif input_ids is not None:
-            input_shape = input_ids.size()
-        elif inputs_embeds is not None:
-            input_shape = inputs_embeds.size()[:-1]
-        else:
-            raise ValueError("You have to specify either input_ids or inputs_embeds")
-
-        device = input_ids.device if input_ids is not None else inputs_embeds.device
-
-        if attention_mask is None:
-            attention_mask = torch.ones(input_shape, device=device)
-        if token_type_ids is None:
-            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
-
-        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
-        # ourselves in which case we just need to make it broadcastable to all heads.
-        if attention_mask.dim() == 3:
-            extended_attention_mask = attention_mask[:, None, :, :]
-        elif attention_mask.dim() == 2:
-            # Provided a padding mask of dimensions [batch_size, seq_length]
-            # - if the model is a decoder, apply a causal mask in addition to the padding mask
-            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
-            if self.config.is_decoder:
-                batch_size, seq_length = input_shape
-                seq_ids = torch.arange(seq_length, device=device)
-                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
-                causal_mask = causal_mask.to(
-                    torch.long
-                )  # not converting to long will cause errors with pytorch version < 1.3
-                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
-            else:
-                extended_attention_mask = attention_mask[:, None, None, :]
-        else:
-            raise ValueError(
-                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
-                    input_shape, attention_mask.shape
-                )
-            )
-
-        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
-        # masked positions, this operation will create a tensor which is 0.0 for
-        # positions we want to attend and -10000.0 for masked positions.
-        # Since we are adding it to the raw scores before the softmax, this is
-        # effectively the same as removing these entirely.
-        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
-        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
-
-        # If a 2D ou 3D attention mask is provided for the cross-attention
-        # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
-        if self.config.is_decoder and encoder_hidden_states is not None:
-            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
-            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
-            if encoder_attention_mask is None:
-                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
-
-            if encoder_attention_mask.dim() == 3:
-                encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
-            elif encoder_attention_mask.dim() == 2:
-                encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
-            else:
-                raise ValueError(
-                    "Wrong shape for encoder_hidden_shape (shape {}) or encoder_attention_mask (shape {})".format(
-                        encoder_hidden_shape, encoder_attention_mask.shape
-                    )
-                )
-
-            encoder_extended_attention_mask = encoder_extended_attention_mask.to(
-                dtype=next(self.parameters()).dtype
-            )  # fp16 compatibility
-            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
-        else:
-            encoder_extended_attention_mask = None
-
-        # Prepare head mask if needed
-        # 1.0 in head_mask indicate we keep the head
-        # attention_probs has shape bsz x n_heads x N x N
-        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
-        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
-        if head_mask is not None:
-            if head_mask.dim() == 1:
-                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
-                head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
-            elif head_mask.dim() == 2:
-                head_mask = (
-                    head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)
-                )  # We can specify head_mask for each layer
-            head_mask = head_mask.to(
-                dtype=next(self.parameters()).dtype
-            )  # switch to fload if need + fp16 compatibility
-        else:
-            head_mask = [None] * self.config.num_hidden_layers
-
-        embedding_output = self.embeddings(
-            input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
-        )
-        encoder_outputs = self.encoder(
-            embedding_output,
-            attention_mask=extended_attention_mask,
-            head_mask=head_mask,
-            encoder_hidden_states=encoder_hidden_states,
-            encoder_attention_mask=encoder_extended_attention_mask,
-        )
-        sequence_output = encoder_outputs[0]
-        pooled_output = self.pooler(sequence_output)
-
-        outputs = (sequence_output, pooled_output,) + encoder_outputs[
-            1:
-        ]  # add hidden_states and attentions if they are here
-        return outputs  # sequence_output, pooled_output, (hidden_states), (attentions)
-
-
-class BertForPreTraining(BertPreTrainedModel):
-    def __init__(self, config):
-        super(BertForPreTraining, self).__init__(config)
-
-        self.bert = BertModel(config)
-        self.cls = BertPreTrainingHeads(config)
-
-        self.init_weights()
-
-    def get_output_embeddings(self):
-        return self.cls.predictions.decoder
-
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        masked_lm_labels=None,
-        next_sentence_label=None,
-    ):
-        outputs = self.bert(
-            input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-        )
-
-        sequence_output, pooled_output = outputs[:2]
-        prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)
-
-        outputs = (prediction_scores, seq_relationship_score,) + outputs[
-            2:
-        ]  # add hidden states and attention if they are here
-
-        if masked_lm_labels is not None and next_sentence_label is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
-            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
-            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
-            total_loss = masked_lm_loss + next_sentence_loss
-            outputs = (total_loss,) + outputs
-
-        return outputs  # (loss), prediction_scores, seq_relationship_score, (hidden_states), (attentions)
-
-
-class BertForMaskedLM(BertPreTrainedModel):
-    def __init__(self, config):
-        super(BertForMaskedLM, self).__init__(config)
-
-        self.bert = BertModel(config)
-        self.cls = BertOnlyMLMHead(config)
-
-        self.init_weights()
-
-    def get_output_embeddings(self):
-        return self.cls.predictions.decoder
-
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        masked_lm_labels=None,
-        encoder_hidden_states=None,
-        encoder_attention_mask=None,
-        lm_labels=None,
-    ):
-
-        outputs = self.bert(
-            input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-            encoder_hidden_states=encoder_hidden_states,
-            encoder_attention_mask=encoder_attention_mask,
-        )
-
-        sequence_output = outputs[0]
-        prediction_scores = self.cls(sequence_output)
-
-        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
-
-        # Although this may seem awkward, BertForMaskedLM supports two scenarios:
-        # 1. If a tensor that contains the indices of masked labels is provided,
-        #    the cross-entropy is the MLM cross-entropy that measures the likelihood
-        #    of predictions for masked words.
-        # 2. If `lm_labels` is provided we are in a causal scenario where we
-        #    try to predict the next token for each input in the decoder.
-        if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)  # -1 index = padding token
-            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
-            outputs = (masked_lm_loss,) + outputs
-
-        if lm_labels is not None:
-            # we are doing next-token prediction; shift prediction scores and input ids by one
-            prediction_scores = prediction_scores[:, :-1, :].contiguous()
-            lm_labels = lm_labels[:, 1:].contiguous()
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
-            ltr_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), lm_labels.view(-1))
-            outputs = (ltr_lm_loss,) + outputs
-
-        return outputs  # (masked_lm_loss), (ltr_lm_loss), prediction_scores, (hidden_states), (attentions)
-
-
-class BertForNextSentencePrediction(BertPreTrainedModel):
-    def __init__(self, config):
-        super(BertForNextSentencePrediction, self).__init__(config)
-
-        self.bert = BertModel(config)
-        self.cls = BertOnlyNSPHead(config)
-
-        self.init_weights()
-
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        next_sentence_label=None,
-    ):
-        outputs = self.bert(
-            input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-        )
-
-        pooled_output = outputs[1]
-
-        seq_relationship_score = self.cls(pooled_output)
-
-        outputs = (seq_relationship_score,) + outputs[2:]  # add hidden states and attention if they are here
-        if next_sentence_label is not None:
-            loss_fct = CrossEntropyLoss(ignore_index=-1)
-            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
-            outputs = (next_sentence_loss,) + outputs
-
-        return outputs  # (next_sentence_loss), seq_relationship_score, (hidden_states), (attentions)
-
-
-class BertForSequenceClassification(BertPreTrainedModel):
-    def __init__(self, config):
-        super(BertForSequenceClassification, self).__init__(config)
-        self.num_labels = config.num_labels
-
-        self.bert = BertModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
-
-        self.init_weights()
-
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        labels=None,
-    ):
-
-        outputs = self.bert(
-            input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-        )
-
-        pooled_output = outputs[1]
-
-        pooled_output = self.dropout(pooled_output)
-        logits = self.classifier(pooled_output)
-
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
-
-        if labels is not None:
-            if self.num_labels == 1:
-                #  We are doing regression
-                loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1), labels.view(-1))
-            else:
-                loss_fct = CrossEntropyLoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
-            outputs = (loss,) + outputs
-
-        return outputs  # (loss), logits, (hidden_states), (attentions)
-
-
-class BertForMultipleChoice(BertPreTrainedModel):
-    def __init__(self, config):
-        super(BertForMultipleChoice, self).__init__(config)
-
-        self.bert = BertModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.classifier = nn.Linear(config.hidden_size, 1)
-
-        self.init_weights()
-
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        labels=None,
-    ):
-        num_choices = input_ids.shape[1]
-
-        input_ids = input_ids.view(-1, input_ids.size(-1))
-        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
-        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
-        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
-
-        outputs = self.bert(
-            input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-        )
-
-        pooled_output = outputs[1]
-
-        pooled_output = self.dropout(pooled_output)
-        logits = self.classifier(pooled_output)
-        reshaped_logits = logits.view(-1, num_choices)
-
-        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
-
-        if labels is not None:
-            loss_fct = CrossEntropyLoss()
-            loss = loss_fct(reshaped_logits, labels)
-            outputs = (loss,) + outputs
-
-        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
-
-
-class BertForTokenClassification(BertPreTrainedModel):
-    def __init__(self, config):
-        super(BertForTokenClassification, self).__init__(config)
-        self.num_labels = config.num_labels
-
-        self.bert = BertModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
-
-        self.init_weights()
-
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        labels=None,
-    ):
-
-        outputs = self.bert(
-            input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-        )
-
-        sequence_output = outputs[0]
-
-        sequence_output = self.dropout(sequence_output)
-        logits = self.classifier(sequence_output)
-
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
-        if labels is not None:
-            loss_fct = CrossEntropyLoss()
-            # Only keep active parts of the loss
-            if attention_mask is not None:
-                active_loss = attention_mask.view(-1) == 1
-                active_logits = logits.view(-1, self.num_labels)[active_loss]
-                active_labels = labels.view(-1)[active_loss]
-                loss = loss_fct(active_logits, active_labels)
-            else:
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
-            outputs = (loss,) + outputs
-
-        return outputs  # (loss), scores, (hidden_states), (attentions)
-
-
-class BertForQuestionAnswering(BertPreTrainedModel):
-    def __init__(self, config):
-        super(BertForQuestionAnswering, self).__init__(config)
-        self.num_labels = config.num_labels
-
-        self.bert = BertModel(config)
-        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
-
-        self.init_weights()
-
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        start_positions=None,
-        end_positions=None,
-    ):
-
-        outputs = self.bert(
-            input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-        )
-
-        sequence_output = outputs[0]
-
-        logits = self.qa_outputs(sequence_output)
-        start_logits, end_logits = logits.split(1, dim=-1)
-        start_logits = start_logits.squeeze(-1)
-        end_logits = end_logits.squeeze(-1)
-
-        outputs = (start_logits, end_logits,) + outputs[2:]
-        if start_positions is not None and end_positions is not None:
-            # If we are on multi-GPU, split add a dimension
-            if len(start_positions.size()) > 1:
-                start_positions = start_positions.squeeze(-1)
-            if len(end_positions.size()) > 1:
-                end_positions = end_positions.squeeze(-1)
-            # sometimes the start/end positions are outside our model inputs, we ignore these terms
-            ignored_index = start_logits.size(1)
-            start_positions.clamp_(0, ignored_index)
-            end_positions.clamp_(0, ignored_index)
-
-            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
-            start_loss = loss_fct(start_logits, start_positions)
-            end_loss = loss_fct(end_logits, end_positions)
-            total_loss = (start_loss + end_loss) / 2
-            outputs = (total_loss,) + outputs
-
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)

examples/summarization/replacement_scheduler.py

@@ -1,32 +0,0 @@
-class ConstantReplacementScheduler:
-    def __init__(self, module, replacing_rate, replacing_steps=None):
-        self.module = module
-        self.replacing_rate = replacing_rate
-        self.replacing_steps = replacing_steps
-        self.step_counter = 0
-        self.module.set_replacing_rate(replacing_rate)
-
-    def step(self):
-        self.step_counter += 1
-        if self.replacing_steps is None or self.replacing_rate == 1.0:
-            return self.replacing_rate
-        else:
-            if self.step_counter >= self.replacing_steps:
-                self.module.set_replacing_rate(1.0)
-                self.replacing_rate = 1.0
-            return self.replacing_rate
-
-
-class LinearReplacementScheduler:
-    def __init__(self, module, base_replacing_rate, k=1e-4):
-        self.module = module
-        self.base_replacing_rate = base_replacing_rate
-        self.step_counter = 0
-        self.k = k
-        self.module.set_replacing_rate(base_replacing_rate)
-
-    def step(self):
-        self.step_counter += 1
-        current_replacing_rate = min(self.k * self.step_counter + self.base_replacing_rate, 1.0)
-        self.module.set_replacing_rate(current_replacing_rate)
-        return current_replacing_rate

examples/summarization/run_eval.py

@@ -5,7 +5,7 @@ from pathlib import Path
 import torch
 from tqdm import tqdm
 
-from transformers import AutoModelWithLMHead, AutoTokenizer
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
 
 
 try:
@@ -23,10 +23,12 @@ def chunks(lst, n):
 
 
 def generate_summaries(
-    examples: list, out_file: str, model_name: str, batch_size: int = 8, device: str = DEFAULT_DEVICE
-):
+    examples: list, out_file: str, model_name: str, batch_size: int = 8, device: str = DEFAULT_DEVICE, fp16=False,
+) -> None:
     fout = Path(out_file).open("w", encoding="utf-8")
-    model = AutoModelWithLMHead.from_pretrained(model_name).to(device)
+    model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)
+    if fp16:
+        model = model.half()
 
     tokenizer = AutoTokenizer.from_pretrained(model_name)
 
@@ -49,32 +51,20 @@ def generate_summaries(
 
 def run_generate():
     parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "input_path", type=str, help="like cnn_dm/test.source",
-    )
-    parser.add_argument(
-        "output_path", type=str, help="where to save summaries",
-    )
-    parser.add_argument(
-        "model_name",
-        type=str,
-        default="facebook/bart-large-cnn",
-        help="like facebook/bart-large-cnn,'t5-small', 't5-base', 't5-large', 't5-3b', 't5-11b",
-    )
+    parser.add_argument("input_path", type=str, help="like cnn_dm/test.source")
+    parser.add_argument("output_path", type=str, help="where to save summaries")
+    parser.add_argument("model_name", type=str, help="like facebook/bart-large-cnn,t5-base, etc.")
     parser.add_argument("--reference_path", type=str, required=False, help="like cnn_dm/test_reference_summaries.txt")
-    parser.add_argument(
-        "--score_path", type=str, required=False, help="where to save the rouge score in json format",
-    )
-    parser.add_argument(
-        "--device", type=str, required=False, default=DEFAULT_DEVICE, help="cuda, cuda:1, cpu etc.",
-    )
-    parser.add_argument(
-        "--bs", type=int, default=8, required=False, help="batch size: how many to summarize at a time",
-    )
+    parser.add_argument("--score_path", type=str, required=False, help="where to save the rouge score in json format")
+    parser.add_argument("--device", type=str, required=False, default=DEFAULT_DEVICE, help="cuda, cuda:1, cpu etc.")
+    parser.add_argument("--bs", type=int, default=8, required=False, help="batch size")
+    parser.add_argument("--fp16", action="store_true")
     args = parser.parse_args()
     examples = [" " + x.rstrip() if "t5" in args.model_name else x.rstrip() for x in open(args.input_path).readlines()]
 
-    generate_summaries(examples, args.output_path, args.model_name, batch_size=args.bs, device=args.device)
+    generate_summaries(
+        examples, args.output_path, args.model_name, batch_size=args.bs, device=args.device, fp16=args.fp16
+    )
     if args.score_path is not None:
         output_lns = [x.rstrip() for x in open(args.output_path).readlines()]
         reference_lns = [x.rstrip() for x in open(args.reference_path).readlines()]

examples/summarization/test_summarization_examples.py

@@ -23,12 +23,9 @@ logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger()
 FP16_EVER = False
 CHEAP_ARGS = {
-    "theseus_init_copy": False,
-    "theseus_replace_rate": 0,
     "logger": "default",
-    "num_workers": 2,
     "alpha_hid": 0,
-    "unfreeze_embeds": False,
+    "freeze_embeds": True,
     "enc_only": False,
     "tgt_suffix": "",
     "resume_from_checkpoint": None,
@@ -74,8 +71,6 @@ CHEAP_ARGS = {
     "alpha_loss_encoder": 0.0,
     "freeze_encoder": False,
     "auto_scale_batch_size": False,
-    "freeze_decoder": False,
-    "init_strategy": "bottom",
 }
 
 
@@ -84,7 +79,7 @@ def _dump_articles(path: Path, articles: list):
         f.write("\n".join(articles))
 
 
-T5_TINY = "patrickvonplaten/t5-tiny-random"
+BDIR = Path("~/transformers_fork/examples/summarization/bart/").absolute()
 
 
 def make_test_data_dir():
@@ -97,6 +92,7 @@ def make_test_data_dir():
     return tmp_dir
 
 
+@unittest.skip("These wont' pass until hidden_states kwarg is merged.")
 class TestSummarizationDistiller(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
@@ -115,17 +111,44 @@ class TestSummarizationDistiller(unittest.TestCase):
         )
         self._bart_distiller_cli(updates)
 
-    def test_bdc_t5_train(self):
+    @unittest.skipUnless(torch.cuda.is_available(), "skipping fp16 test")
+    def test_bdc_fp16(self):
+        updates = dict(
+            student_encoder_layers=2,
+            student_decoder_layers=1,
+            alpha_hid=3.0,
+            freeze_encoder=True,
+            gpus=1,
+            fp16=FP16_EVER,
+            fp16_opt_level="O1",
+        )
+        self._bart_distiller_cli(updates)
+
+    @unittest.skipUnless(torch.cuda.is_available(), "skipping fp16 test")
+    def test_bdc_t5_eval_fp16(self):
         updates = dict(
             fp16=FP16_EVER,
-            gpus=1 if torch.cuda.is_available() else 0,
+            gpus=1,
             model_type="t5",
-            model_name_or_path=T5_TINY,
+            model_name_or_path="patrickvonplaten/t5-tiny-random",
+            do_train=False,
+            do_predict=True,
+            tokenizer_name=None,
+            no_teacher=True,
+        )
+        self._bart_distiller_cli(updates, check_contents=False)
+
+    @unittest.skipUnless(torch.cuda.is_available(), "skipping fp16 test")
+    def test_bdc_t5_train_fp16(self):
+        updates = dict(
+            fp16=FP16_EVER,
+            gpus=1,
+            model_type="t5",
+            model_name_or_path="patrickvonplaten/t5-tiny-random",
             do_train=True,
             do_predict=True,
-            tokenizer_name=T5_TINY,
+            tokenizer_name="patrickvonplaten/t5-tiny-random",
             no_teacher=True,
-            alpha_hid=2.0,
         )
         self._bart_distiller_cli(updates)
 
@@ -133,14 +156,18 @@ class TestSummarizationDistiller(unittest.TestCase):
         updates = dict(student_encoder_layers=2, student_decoder_layers=1, no_teacher=True,)
         self._bart_distiller_cli(updates)
 
+    def test_bdc_yes_teacher(self):
+        updates = dict(student_encoder_layers=2, student_decoder_layers=1,)
+        self._bart_distiller_cli(updates)
+
     def test_bdc_checkpointing(self):
+
         updates = dict(
             student_encoder_layers=2,
             student_decoder_layers=1,
             num_train_epochs=4,
             val_check_interval=0.25,
             alpha_hid=2.0,
-            init_strategy="alternate",
         )
         model = self._bart_distiller_cli(updates, check_contents=False)
 
@@ -152,33 +179,37 @@ class TestSummarizationDistiller(unittest.TestCase):
         self.assertEqual(len(new_transformer_ckpts), 1)
         examples = lmap(str.strip, model.hparams.data_dir.joinpath("test.source").open().readlines())
         out_path = tempfile.mktemp()
-        generate_summaries(examples, out_path, model_name=str(new_transformer_ckpts[0].parent))
+        generate_summaries(examples, out_path, new_transformer_ckpts[0].parent)
         self.assertTrue(Path(out_path).exists())
+
         evaluate_checkpoint(ckpts[0], dest_dir=Path(tempfile.mkdtemp()))
 
-    def test_bdc_theseus(self):
+    def test_bdc_t5(self):
         updates = dict(
-            theseus_replace_rate=0.5,
             student_encoder_layers=1,
             student_decoder_layers=1,
-            no_teacher=True,
-            theseus_init_copy=True,
+            alpha_hid=2.0,
+            teacher="patrickvonplaten/t5-tiny-random",
+            model_type="t5",
+            model_name_or_path="patrickvonplaten/t5-tiny-random",
+            tokenizer_name="patrickvonplaten/t5-tiny-random",
         )
         self._bart_distiller_cli(updates)
 
-    def test_bdc_frozen_theseus(self):
+    def test_bdc_t5_eval(self):
         updates = dict(
-            theseus_replace_rate=0.5,
-            student_encoder_layers=2,
-            student_decoder_layers=1,
+            model_type="t5",
+            model_name_or_path="patrickvonplaten/t5-tiny-random",
+            do_train=False,
+            do_predict=True,
+            tokenizer_name="patrickvonplaten/t5-tiny-random",
             no_teacher=True,
-            freeze_encoder=True,
-            theseus_init_copy=True,
         )
-        self._bart_distiller_cli(updates)
+        self._bart_distiller_cli(updates, check_contents=False)
 
     def _bart_distiller_cli(self, updates, check_contents=True):
         default_updates = dict(
+            model_type="bart",
             train_batch_size=1,
             eval_batch_size=2,
             num_train_epochs=2,
@@ -206,14 +237,21 @@ class TestSummarizationDistiller(unittest.TestCase):
         self.assertIn(ckpt_name, contents)
         self.assertIn("metrics.pkl", contents)
         self.assertIn("test_generations.txt", contents)
-        self.assertIn("val_generations_00001.txt", contents)
-        self.assertIn("val_results_00001.txt", contents)
+        self.assertIn("val_generations_1.txt", contents)
+        self.assertIn("val_1_results.txt", contents)
         self.assertIn("test_results.txt", contents)
+        # self.assertEqual(len(contents), 15)
 
         metrics = pickle_load(Path(output_dir) / "metrics.pkl")
-        desired_n_evals = int(args_d["num_train_epochs"] * (1 / args_d["val_check_interval"]) + 1)
-        self.assertEqual(len(metrics["val"]), desired_n_evals)
-        self.assertEqual(len(metrics["train"]), 0)  # doesn't get logged here
+        import pandas as pd
+
+        val_df = pd.DataFrame(metrics["val"])
+        train_df = pd.DataFrame(metrics["train"])
+        test_df = pd.DataFrame(metrics["test"])
+        desired_n_evals = args_d["num_train_epochs"] * 2 + 1
+        self.assertEqual(val_df.shape[0], desired_n_evals)  #
+        self.assertEqual(test_df.shape[1], val_df.shape[1])
+        self.assertEqual(train_df.shape[0], 0)
         return model
 
 
@@ -243,8 +281,9 @@ class TestBartExamples(unittest.TestCase):
         output_dir = tempfile.mkdtemp(prefix="output_")
         args_d.update(
             data_dir=tmp_dir,
-            model_name_or_path=T5_TINY,
-            tokenizer_name=None,  # T5_TINY,
+            model_type="t5",
+            model_name_or_path="patrickvonplaten/t5-tiny-random",
+            tokenizer_name=None,  # "patrickvonplaten/t5-tiny-random",
             train_batch_size=2,
             eval_batch_size=2,
             gpus=0,

examples/summarization/utils.py

@@ -32,6 +32,7 @@ def encode_file(
             examples = torch.load(cache_path)
             assert isinstance(examples, list)
             return examples
+
         except Exception:
             print(f"failed to load from {cache_path}, retokenizing {data_path}")
     data_path = Path(data_path)
@@ -46,10 +47,8 @@ def encode_file(
             max_length=max_length,
             pad_to_max_length=pad_to_max_length,
             add_prefix_space=True,
-            truncation=True,
             return_tensors=return_tensors,
         )
-        assert tokenized.input_ids.shape[1] == max_length
         examples.append(tokenized)
     torch.save(lmap(dict, examples), cache_path.open("wb"))
     return examples
@@ -95,11 +94,16 @@ class SummarizationDataset(Dataset):
             prefix=prefix,
             tok_name=tok_name,
         )
-        tgt_path = os.path.join(data_dir, type_path + ".target")
+        if type_path == "train":
+            tgt_path = os.path.join(data_dir, type_path + ".target")
+        else:
+            tgt_path = os.path.join(data_dir, type_path + ".target")
+
         self.target = encode_file(
             tokenizer, tgt_path, max_target_length, overwrite_cache=overwrite_cache, tok_name=tok_name
         )
-
+        self.source = encode_file(tokenizer, os.path.join(data_dir, type_path + ".source"), max_source_length)
+        self.target = encode_file(tokenizer, os.path.join(data_dir, type_path + ".target"), max_target_length)
         if n_obs is not None:
             self.source = self.source[:n_obs]
             self.target = self.target[:n_obs]
@@ -214,7 +218,7 @@ def get_git_info():
 ROUGE_KEYS = ["rouge1", "rouge2", "rougeL"]
 
 
-def calculate_rouge(output_lns: List[str], reference_lns: List[str], all_stats=False):
+def calculate_rouge(output_lns: List[str], reference_lns: List[str]) -> Dict:
     scorer = rouge_scorer.RougeScorer(ROUGE_KEYS, use_stemmer=True)
     aggregator = scoring.BootstrapAggregator()
 
@@ -223,11 +227,7 @@ def calculate_rouge(output_lns: List[str], reference_lns: List[str], all_stats=F
         aggregator.add_scores(scores)
 
     result = aggregator.aggregate()
-
-    if all_stats:
-        return expanded_rouge_df(result)
-    else:
-        return {k: v.mid.fmeasure for k, v in result.items()}
+    return {k: v.mid.fmeasure for k, v in result.items()}
 
 
 def freeze_params(model: nn.Module):
@@ -247,36 +247,10 @@ def assert_all_frozen(model):
     model_grads: List[bool] = list(grad_status(model))
     n_require_grad = sum(lmap(int, model_grads))
     npars = len(model_grads)
-    assert not any(model_grads), f"{n_require_grad / npars:.1%} of {npars} weights require grad"
+    assert not any(model_grads), f"{n_require_grad/npars:.1%} of {npars} weights require grad"
 
 
 def assert_not_all_frozen(model):
     model_grads: List[bool] = list(grad_status(model))
     npars = len(model_grads)
     assert any(model_grads), f"none of {npars} weights require grad"
-
-
-def dictify(rouge_obj) -> List:
-    records = []
-    for k, rouge_measurement in rouge_obj.items():
-        if k == "rouge1":
-            continue
-        for k1 in ["low", "mid", "high"]:
-            if k1 != "mid":
-                continue
-            v1 = getattr(rouge_measurement, k1)
-            for k2 in ["precision", "recall", "fmeasure"]:
-                records.append([k, k1, k2, getattr(v1, k2)])
-
-    return records
-
-
-def expanded_rouge_df(rouge_all):
-    import pandas as pd
-
-    return (
-        pd.DataFrame(dictify(rouge_all), columns=["metric", "k1", "k2", "val"])
-        .set_index(["metric", "k2"])["val"]
-        .unstack("metric")
-        .rename_axis(None)
-    )

examples/text-classification/run_pl_glue.py

@@ -108,7 +108,7 @@ class GLUETransformer(BaseTransformer):
 
         return {"val_loss": tmp_eval_loss.detach().cpu(), "pred": preds, "target": out_label_ids}
 
-    def _eval_end(self, outputs) -> tuple:
+    def _eval_end(self, outputs):
         val_loss_mean = torch.stack([x["val_loss"] for x in outputs]).mean().detach().cpu().item()
         preds = np.concatenate([x["pred"] for x in outputs], axis=0)
 
@@ -132,14 +132,20 @@ class GLUETransformer(BaseTransformer):
         logs = ret["log"]
         return {"val_loss": logs["val_loss"], "log": logs, "progress_bar": logs}
 
-    def test_epoch_end(self, outputs) -> dict:
+    def test_epoch_end(self, outputs):
+        # updating to test_epoch_end instead of deprecated test_end
         ret, predictions, targets = self._eval_end(outputs)
+
+        # Converting to the dic required by pl
+        # https://github.com/PyTorchLightning/pytorch-lightning/blob/master/\
+        # pytorch_lightning/trainer/logging.py#L139
         logs = ret["log"]
         # `val_loss` is the key returned by `self._eval_end()` but actually refers to `test_loss`
         return {"avg_test_loss": logs["val_loss"], "log": logs, "progress_bar": logs}
 
     @staticmethod
     def add_model_specific_args(parser, root_dir):
+        # Add NER specific options
         BaseTransformer.add_model_specific_args(parser, root_dir)
         parser.add_argument(
             "--max_seq_length",

model_cards/a-ware/bart-squadv2/README.md

@@ -1,76 +0,0 @@
----
-datasets:
-- squad_v2
----
-
-# BART-LARGE finetuned on SQuADv2
-
-This is bart-large model finetuned on SQuADv2 dataset for question answering task
-
-## Model details
-BART was propsed in the [paper](https://arxiv.org/abs/1910.13461) **BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension**.
-BART is a seq2seq model intended for both NLG and NLU tasks. 
-
-To use BART for question answering tasks, we feed the complete document into the encoder and decoder, and use the top
-hidden state of the decoder as a representation for each
-word. This representation is used to classify the token. As given in the paper bart-large achives comparable to ROBERTa on SQuAD.
-Another notable thing about BART is that it can handle sequences with upto 1024 tokens.
-
-| Param               | #Value |
-|---------------------|--------|
-| encoder layers      | 12     |
-| decoder layers      | 12     |
-| hidden size         | 4096   |
-| num attetion heads  | 16     |
-| on disk size        | 1.63GB |
-
-
-## Model training
-This model was trained with following parameters using simpletransformers wrapper:
-```
-train_args = {
-    'learning_rate': 1e-5,
-    'max_seq_length': 512,
-    'doc_stride': 512,
-    'overwrite_output_dir': True,
-    'reprocess_input_data': False,
-    'train_batch_size': 8,
-    'num_train_epochs': 2,
-    'gradient_accumulation_steps': 2,
-    'no_cache': True,
-    'use_cached_eval_features': False,
-    'save_model_every_epoch': False,
-    'output_dir': "bart-squadv2",
-    'eval_batch_size': 32,
-    'fp16_opt_level': 'O2',
-    }
-```
-
-[You can even train your own model using this colab notebook](https://colab.research.google.com/drive/1I5cK1M_0dLaf5xoewh6swcm5nAInfwHy?usp=sharing)
-
-## Results
-```{"correct": 6832, "similar": 4409, "incorrect": 632, "eval_loss": -14.950117511952177}```
-
-## Model in Action  🚀
-```python3
-from transformers import BartTokenizer, BartForQuestionAnswering
-import torch
-
-tokenizer = BartTokenizer.from_pretrained('a-ware/bart-squadv2')
-model = BartForQuestionAnswering.from_pretrained('a-ware/bart-squadv2')
-
-question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
-encoding = tokenizer.encode_plus(question, text, return_tensors='pt')
-input_ids = encoding['input_ids']
-attention_mask = encoding['attention_mask']
-
-start_scores, end_scores = model(input_ids, attention_mask=attention_mask, output_attentions=False)[:2]
-
-all_tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
-answer = ' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1])
-answer = tokenizer.convert_tokens_to_ids(answer.split())
-answer = tokenizer.decode(answer)
-#answer => 'a nice puppet' 
-```
-
-> Created with ❤️ by A-ware UG [![Github icon](https://cdn0.iconfinder.com/data/icons/octicons/1024/mark-github-32.png)](https://github.com/aware-ai)

model_cards/a-ware/xlmroberta-squadv2/README.md

@@ -1,59 +0,0 @@
----
-datasets:
-- squad_v2
----
-
-# XLM-ROBERTA-LARGE finetuned on SQuADv2
-
-This is xlm-roberta-large model finetuned on SQuADv2 dataset for question answering task
-
-## Model details
-XLM-Roberta was propsed in the [paper](https://arxiv.org/pdf/1911.02116.pdf) **XLM-R: State-of-the-art cross-lingual understanding through self-supervision
-
-## Model training
-This model was trained with following parameters using simpletransformers wrapper:
-```
-train_args = {
-    'learning_rate': 1e-5,
-    'max_seq_length': 512,
-    'doc_stride': 512,
-    'overwrite_output_dir': True,
-    'reprocess_input_data': False,
-    'train_batch_size': 8,
-    'num_train_epochs': 2,
-    'gradient_accumulation_steps': 2,
-    'no_cache': True,
-    'use_cached_eval_features': False,
-    'save_model_every_epoch': False,
-    'output_dir': "bart-squadv2",
-    'eval_batch_size': 32,
-    'fp16_opt_level': 'O2',
-    }
-```
-
-## Results
-```{"correct": 6961, "similar": 4359, "incorrect": 553, "eval_loss": -12.177856394381962}```
-
-## Model in Action  🚀
-```python3
-from transformers import XLMRobertaTokenizer, XLMRobertaForQuestionAnswering
-import torch
-
-tokenizer = XLMRobertaTokenizer.from_pretrained('a-ware/xlmroberta-squadv2')
-model = XLMRobertaForQuestionAnswering.from_pretrained('a-ware/xlmroberta-squadv2')
-
-question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
-encoding = tokenizer.encode_plus(question, text, return_tensors='pt')
-input_ids = encoding['input_ids']
-attention_mask = encoding['attention_mask']
-
-start_scores, end_scores = model(input_ids, attention_mask=attention_mask, output_attentions=False)[:2]
-
-all_tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
-answer = ' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1])
-answer = tokenizer.convert_tokens_to_ids(answer.split())
-answer = tokenizer.decode(answer)
-#answer => 'a nice puppet' 
-```
-
-> Created with ❤️ by A-ware UG [![Github icon](https://cdn0.iconfinder.com/data/icons/octicons/1024/mark-github-32.png)](https://github.com/aware-ai)

model_cards/aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616/README.md

@@ -1,20 +0,0 @@
-# BERT L-10 H512 fine-tuned on MLM (CORD-19 2020/06/16)
-
-BERT model with [10 Transformer layers and hidden embedding of size 512](https://huggingface.co/google/bert_uncased_L-10_H-512_A-8), referenced in [Well-Read Students Learn Better: On the Importance of Pre-training Compact Models](https://arxiv.org/abs/1908.08962).
-
-## Training the model
-
-```bash
-python run_language_modeling.py
-    --model_type bert
-    --model_name_or_path google/bert_uncased_L-10_H-512_A-8
-    --do_train
-    --train_data_file {cord19-200616-dataset}
-    --mlm
-    --mlm_probability 0.2
-    --line_by_line
-    --block_size 512
-    --per_device_train_batch_size 20
-    --learning_rate 3e-5
-    --num_train_epochs 2
-    --output_dir bert_uncased_L-10_H-512_A-8_cord19-200616

model_cards/aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616_squad2/README.md

@@ -1,26 +0,0 @@
----
-datasets:
-- squad_v2
----
-
-# BERT L-10 H-512 CORD-19 (2020/06/16) fine-tuned on SQuAD v2.0
-
-BERT model with [10 Transformer layers and hidden embedding of size 512](https://huggingface.co/google/bert_uncased_L-10_H-512_A-8), referenced in [Well-Read Students Learn Better: On the Importance of Pre-training Compact Models](https://arxiv.org/abs/1908.08962), [fine-tuned for MLM](https://huggingface.co/aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616) on CORD-19 dataset (as released on 2020/06/16) and fine-tuned for QA on SQuAD v2.0.
-
-## Training the model
-
-```bash
-python run_squad.py
-    --model_type bert
-    --model_name_or_path aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616
-    --train_file 'train-v2.0.json'
-    --predict_file 'dev-v2.0.json'
-    --do_train
-    --do_eval
-    --do_lower_case
-    --version_2_with_negative
-    --max_seq_length 384
-    --per_gpu_train_batch_size 10
-    --learning_rate 3e-5
-    --num_train_epochs 2
-    --output_dir bert_uncased_L-10_H-512_A-8_cord19-200616_squad2

model_cards/aodiniz/bert_uncased_L-2_H-512_A-8_cord19-200616/README.md

@@ -1,20 +0,0 @@
-# BERT L-2 H-512 fine-tuned on MLM (CORD-19 2020/06/16)
-
-BERT model with [2 Transformer layers and hidden embedding of size 512](https://huggingface.co/google/bert_uncased_L-2_H-512_A-8), referenced in [Well-Read Students Learn Better: On the Importance of Pre-training Compact Models](https://arxiv.org/abs/1908.08962), fine-tuned for MLM on CORD-19 dataset (as released on 2020/06/16).
-
-## Training the model
-
-```bash
-python run_language_modeling.py
-    --model_type bert
-    --model_name_or_path google/bert_uncased_L-2_H-512_A-8
-    --do_train
-    --train_data_file {cord19-200616-dataset}
-    --mlm
-    --mlm_probability 0.2
-    --line_by_line
-    --block_size 512
-    --per_device_train_batch_size 20
-    --learning_rate 3e-5
-    --num_train_epochs 2
-    --output_dir bert_uncased_L-2_H-512_A-8_cord19-200616

model_cards/aodiniz/bert_uncased_L-4_H-256_A-4_cord19-200616/README.md

@@ -1,20 +0,0 @@
-# BERT L-4 H-256 fine-tuned on MLM (CORD-19 2020/06/16)
-
-BERT model with [4 Transformer layers and hidden embedding of size 256](https://huggingface.co/google/bert_uncased_L-4_H-256_A-4), referenced in [Well-Read Students Learn Better: On the Importance of Pre-training Compact Models](https://arxiv.org/abs/1908.08962), fine-tuned for MLM on CORD-19 dataset (as released on 2020/06/16).
-
-## Training the model
-
-```bash
-python run_language_modeling.py
-    --model_type bert
-    --model_name_or_path google/bert_uncased_L-4_H-256_A-4
-    --do_train
-    --train_data_file {cord19-200616-dataset}
-    --mlm
-    --mlm_probability 0.2
-    --line_by_line
-    --block_size 256
-    --per_device_train_batch_size 20
-    --learning_rate 3e-5
-    --num_train_epochs 2
-    --output_dir bert_uncased_L-4_H-256_A-4_cord19-200616

model_cards/canwenxu/BERT-of-Theseus-MNLI/README.md

@@ -20,5 +20,3 @@ We provide a 6-layer pretrained model on MNLI as a general-purpose model, which
 | BERT-base       | 83.5 | 89.5 | 91.2 | 89.8 | 71.1 | 91.5  | 88.9  |
 | DistillBERT     | 79.0 | 87.5 | 85.3 | 84.9 | 59.9 | 90.7  | 81.2  |
 | BERT-of-Theseus | 82.1 | 87.5 | 88.8 | 88.8 | 70.1 | 91.8  | 87.8  |
-
-Please Note: this checkpoint is for [Intermediate-Task Transfer Learning](https://arxiv.org/abs/2005.00628) so it does not include the classification head for MNLI! Please fine-tune it before use (like DistilBERT).

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

@@ -4,7 +4,7 @@
 **Language model:** deepset/roberta-base-squad2  
 **Language:** English  
 **Downstream-task:** Extractive QA  
-**Training data:** [SQuAD-style CORD-19 annotations from 23rd April](https://github.com/deepset-ai/COVID-QA/blob/master/data/question-answering/200423_covidQA.json)  
+**Training data:** [SQuAD-style CORD-19 annotations](https://github.com/deepset-ai/COVID-QA/tree/master/data/question-answering)  
 **Code:**  See [example](https://github.com/deepset-ai/FARM/blob/master/examples/question_answering_crossvalidation.py) in [FARM](https://github.com/deepset-ai/FARM)  
 **Infrastructure**: Tesla v100
 

model_cards/fran-martinez/scibert_scivocab_cased_ner_jnlpba/README.md

@@ -1,161 +0,0 @@
----
-language: scientific english
----
-
-# SciBERT finetuned on JNLPA for NER downstream task
-## Language Model
- [SciBERT](https://arxiv.org/pdf/1903.10676.pdf) is a pretrained language model based on BERT and trained by the 
- [Allen Institute for AI](https://allenai.org/) on papers from the corpus of 
- [Semantic Scholar](https://www.semanticscholar.org/). 
- Corpus size is 1.14M papers, 3.1B tokens. SciBERT has its own vocabulary (scivocab) that's built to best match 
- the training corpus.
- 
-## Downstream task
-[`allenai/scibert_scivocab_cased`](https://huggingface.co/allenai/scibert_scivocab_cased#) has been finetuned for Named Entity 
-Recognition (NER) dowstream task. The code to train the NER can be found [here](https://github.com/fran-martinez/bio_ner_bert).
-
-### Data
-The corpus used to fine-tune the NER is [BioNLP / JNLPBA shared task](http://www.geniaproject.org/shared-tasks/bionlp-jnlpba-shared-task-2004).
-
-- Training data consist of 2,000 PubMed abstracts with term/word annotation. This corresponds to 18,546 samples (senteces).
-- Evaluation data consist of 404 PubMed abstracts with term/word annotation. This corresponds to 3,856 samples (sentences).
-
-The classes (at word level) and its distribution (number of examples for each class) for training and evaluation datasets are shown below:
- 
-| Class Label         | # training examples| # evaluation examples|
-|:--------------|--------------:|----------------:|
-|O              |   382,963     |     81,647      |
-|B-protein      |    30,269     |      5,067      |
-|I-protein      |    24,848     |      4,774      |
-|B-cell_type    |     6,718     |      1,921      |
-|I-cell_type    |     8,748     |      2,991      |
-|B-DNA          |     9,533     |      1,056      |
-|I-DNA          |    15,774     |      1,789      |
-|B-cell_line    |     3,830     |        500      |
-|I-cell_line    |     7,387     |       9,89      |
-|B-RNA          |       951     |        118      |
-|I-RNA          |     1,530     |        187      |
-
-### Model
-An exhaustive hyperparameter search was done.
-The hyperparameters that provided the best results are:
-
-- Max length sequence: 128
-- Number of epochs: 6
-- Batch size: 32
-- Dropout: 0.3
-- Optimizer: Adam
-
-The used learning rate was 5e-5 with a decreasing linear schedule. A warmup was used at the beggining of the training
-with a ratio of steps equal to 0.1 from the total training steps.
-
-The model from the epoch with the best F1-score was selected, in this case, the model from epoch 5.
-
-
-### Evaluation
-The following table shows the evaluation metrics calculated at span/entity level:
-
-|          |   precision|    recall|  f1-score|   
-|:---------|-----------:|---------:|---------:|
-cell_line   |  0.5205   | 0.7100   | 0.6007   | 
-cell_type   |  0.7736   | 0.7422   | 0.7576   |
-protein     |  0.6953   | 0.8459   | 0.7633   |
-DNA         |  0.6997   | 0.7894   | 0.7419   | 
-RNA         |  0.6985   | 0.8051   | 0.7480   | 
-|           |          |          |
-**micro avg**   |  0.6984   | 0.8076  |  0.7490|
-**macro avg**   | 0.7032   | 0.8076   | 0.7498 |
-
-The macro F1-score is equal to 0.7498, compared to the value provided by the Allen Institute for AI in their
-[paper](https://arxiv.org/pdf/1903.10676.pdf), which is equal to 0.7728. This drop in performance could be due to 
-several reasons, but one hypothesis could be the fact that the authors used an additional conditional random field, 
-while this model uses a regular classification layer with softmax activation on top of SciBERT model.
-
-At word level, this model achieves a precision of 0.7742, a recall of 0.8536 and a F1-score of 0.8093.
-
-### Model usage in inference
-Use the pipeline:
-````python
-from transformers import pipeline
-
-text = "Mouse thymus was used as a source of glucocorticoid receptor from normal CS lymphocytes."
-
-nlp_ner = pipeline("ner",
-                   model='fran-martinez/scibert_scivocab_cased_ner_jnlpba',
-                   tokenizer='fran-martinez/scibert_scivocab_cased_ner_jnlpba')
-
-nlp_ner(text)
-
-"""
-Output:
----------------------------
-[
-{'word': 'glucocorticoid', 
-'score': 0.9894881248474121, 
-'entity': 'B-protein'}, 
- 
-{'word': 'receptor', 
-'score': 0.989505410194397, 
-'entity': 'I-protein'}, 
-
-{'word': 'normal', 
-'score': 0.7680378556251526, 
-'entity': 'B-cell_type'}, 
-
-{'word': 'cs', 
-'score': 0.5176806449890137, 
-'entity': 'I-cell_type'}, 
-
-{'word': 'lymphocytes', 
-'score': 0.9898491501808167, 
-'entity': 'I-cell_type'}
-]
-"""
-````
-Or load model and tokenizer as follows:
-````python
-import torch
-from transformers import AutoTokenizer, AutoModelForTokenClassification
-
-# Example
-text = "Mouse thymus was used as a source of glucocorticoid receptor from normal CS lymphocytes."
-
-# Load model
-tokenizer = AutoTokenizer.from_pretrained("fran-martinez/scibert_scivocab_cased_ner_jnlpba")
-model = AutoModelForTokenClassification.from_pretrained("fran-martinez/scibert_scivocab_cased_ner_jnlpba")
-
-# Get input for BERT
-input_ids = torch.tensor(tokenizer.encode(text)).unsqueeze(0)
-
-# Predict
-with torch.no_grad():
-  outputs = model(input_ids)
-
-# From the output let's take the first element of the tuple.
-# Then, let's get rid of [CLS] and [SEP] tokens (first and last)
-predictions = outputs[0].argmax(axis=-1)[0][1:-1]
-
-# Map label class indexes to string labels.
-for token, pred in zip(tokenizer.tokenize(text), predictions):
-  print(token, '->', model.config.id2label[pred.numpy().item()])
-
-"""
-Output:
----------------------------
-mouse -> O
-thymus -> O
-was -> O
-used -> O
-as -> O
-a -> O
-source -> O
-of -> O
-glucocorticoid -> B-protein
-receptor -> I-protein
-from -> O
-normal -> B-cell_type
-cs -> I-cell_type
-lymphocytes -> I-cell_type
-. -> O
-"""
-````

model_cards/giganticode/StackOBERTflow-comments-small-v1/README.md

@@ -1,39 +0,0 @@
-# StackOBERTflow-comments-small
-
-StackOBERTflow is a RoBERTa model trained on StackOverflow comments.
-A Byte-level BPE tokenizer with dropout was used (using the `tokenizers` package).
-
-The model is *small*, i.e. has only 6-layers and the maximum sequence length was restricted to 256 tokens. 
-The model was trained for 6 epochs on several GBs of comments from the StackOverflow corpus.
-
-## Quick start: masked language modeling prediction
-
-```python
-from transformers import pipeline
-from pprint import pprint
-
-COMMENT = "You really should not do it this way, I would use <mask> instead."
-
-fill_mask = pipeline(
-    "fill-mask",
-    model="giganticode/StackOBERTflow-comments-small-v1",
-    tokenizer="giganticode/StackOBERTflow-comments-small-v1"
-)
-
-pprint(fill_mask(COMMENT))
-# [{'score': 0.019997311756014824,
-#   'sequence': '<s> You really should not do it this way, I would use jQuery instead.</s>',
-#   'token': 1738},
-#  {'score': 0.01693696901202202,
-#   'sequence': '<s> You really should not do it this way, I would use arrays instead.</s>',
-#   'token': 2844},
-#  {'score': 0.013411642983555794,
-#   'sequence': '<s> You really should not do it this way, I would use CSS instead.</s>',
-#   'token': 2254},
-#  {'score': 0.013224546797573566,
-#   'sequence': '<s> You really should not do it this way, I would use it instead.</s>',
-#   'token': 300},
-#  {'score': 0.011984303593635559,
-#   'sequence': '<s> You really should not do it this way, I would use classes instead.</s>',
-#   'token': 1779}]
-```

model_cards/google/mobilebert-uncased/README.md

@@ -1,32 +0,0 @@
----
-language: english
-thumbnail: https://huggingface.co/front/thumbnails/google.png
-
-license: apache-2.0
----
-
-## MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices
-
-MobileBERT is a thin version of BERT_LARGE, while equipped with bottleneck structures and a carefully designed balance
-between self-attentions and feed-forward networks.
-
-This checkpoint is the original MobileBert Optimized Uncased English: 
-[uncased_L-24_H-128_B-512_A-4_F-4_OPT](https://storage.googleapis.com/cloud-tpu-checkpoints/mobilebert/uncased_L-24_H-128_B-512_A-4_F-4_OPT.tar.gz) 
-checkpoint.
-
-## How to use MobileBERT in `transformers`
-
-```python
-from transformers import pipeline
-
-fill_mask = pipeline(
-	"fill-mask",
-	model="google/mobilebert-uncased",
-	tokenizer="google/mobilebert-uncased"
-)
-
-print(
-	fill_mask(f"HuggingFace is creating a {fill_mask.tokenizer.mask_token} that the community uses to solve NLP tasks.")
-)
-
-```

model_cards/google/reformer-crime-and-punishment/README.md

@@ -10,8 +10,8 @@ The ReformerLM model was trained in flax using colab notebook proposed by author
 The model is a language model that operates on small sub-word units. Text can be generated as follows:
 
 ```python
-model = ReformerModelWithLMHead.from_pretrained("google/reformer-crime-and-punishment")
-tok = ReformerTokenizer.from_pretrained("google/reformer-crime-and-punishment")
+model = ReformerModelWithLMHead.from_pretrained("patrickvonplaten/reformer-crime-and-punish")
+tok = ReformerTokenizer.from_pretrained("patrickvonplaten/reformer-crime-and-punish")
 tok.decode(model.generate(tok.encode("A few months later", return_tensors="pt"), do_sample=True,temperature=0.7, max_length=100)[0])
 
 # gives:'A few months later on was more than anything in the flat. 

model_cards/microsoft/MiniLM-L12-H384-uncased/README.md

@@ -1,43 +0,0 @@
----
-thumbnail: https://huggingface.co/front/thumbnails/microsoft.png
-tags:
-- text-classification
-license: mit
----
-
-## MiniLM: Small and Fast Pre-trained Models for Language Understanding and Generation
-
-MiniLM is a distilled model from the paper "[MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers](https://arxiv.org/abs/2002.10957)".
-
-Please find the information about preprocessing, training and full details of the MiniLM in the [original MiniLM repository](https://github.com/microsoft/unilm/blob/master/minilm/).
-
-Please note: This checkpoint can be an inplace substitution for BERT and it needs to be fine-tuned before use!
-
-### English Pre-trained Models
-We release the **uncased** **12**-layer model with **384** hidden size distilled from an in-house pre-trained [UniLM v2](/unilm) model in BERT-Base size.
-
-- MiniLMv1-L12-H384-uncased: 12-layer, 384-hidden, 12-heads, 33M parameters, 2.7x faster than BERT-Base
-
-#### Fine-tuning on NLU tasks
-
-We present the dev results on SQuAD 2.0 and several GLUE benchmark tasks.
-
-| Model                                             | #Param | SQuAD 2.0 | MNLI-m | SST-2 | QNLI | CoLA | RTE  | MRPC | QQP  |
-|---------------------------------------------------|--------|-----------|--------|-------|------|------|------|------|------|
-| [BERT-Base](https://arxiv.org/pdf/1810.04805.pdf) | 109M   | 76.8      | 84.5   | 93.2  | 91.7 | 58.9 | 68.6 | 87.3 | 91.3 |
-| **MiniLM-L12xH384**                               | 33M    | 81.7      | 85.7   | 93.0  | 91.5 | 58.5 | 73.3 | 89.5 | 91.3 |
-
-### Citation
-
-If you find MiniLM useful in your research, please cite the following paper:
-
-``` latex
-@misc{wang2020minilm,
-    title={MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers},
-    author={Wenhui Wang and Furu Wei and Li Dong and Hangbo Bao and Nan Yang and Ming Zhou},
-    year={2020},
-    eprint={2002.10957},
-    archivePrefix={arXiv},
-    primaryClass={cs.CL}
-}
-```

model_cards/microsoft/Multilingual-MiniLM-L12-H384/README.md

@@ -1,89 +0,0 @@
----
-thumbnail: https://huggingface.co/front/thumbnails/microsoft.png
-tags:
-- text-classification
-license: mit
----
-
-## MiniLM: Small and Fast Pre-trained Models for Language Understanding and Generation
-
-MiniLM is a distilled model from the paper "[MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers](https://arxiv.org/abs/2002.10957)".
-
-Please find the information about preprocessing, training and full details of the MiniLM in the [original MiniLM repository](https://github.com/microsoft/unilm/blob/master/minilm/).
-
-Please note: This checkpoint uses `BertModel` with `XLMRobertaTokenizer` so `AutoTokenizer` won't work with this checkpoint!
-
-### Multilingual Pretrained Model
-- Multilingual-MiniLMv1-L12-H384: 12-layer, 384-hidden, 12-heads, 21M Transformer parameters, 96M embedding parameters
-
-Multilingual MiniLM uses the same tokenizer as XLM-R. But the Transformer architecture of our model is the same as BERT. We provide the fine-tuning code on XNLI based on [huggingface/transformers](https://github.com/huggingface/transformers). Please replace `run_xnli.py` in transformers with [ours](https://github.com/microsoft/unilm/blob/master/minilm/examples/run_xnli.py) to fine-tune multilingual MiniLM.  
-
-We evaluate the multilingual MiniLM on cross-lingual natural language inference benchmark (XNLI) and cross-lingual question answering benchmark (MLQA).
-
-#### Cross-Lingual Natural Language Inference - [XNLI](https://arxiv.org/abs/1809.05053)
-
-We evaluate our model on cross-lingual transfer from English to other languages. Following [Conneau et al. (2019)](https://arxiv.org/abs/1911.02116), we select the best single model on the joint dev set of all the languages.
-
-| Model                                                                                       | #Layers | #Hidden | #Transformer Parameters | Average | en   | fr   | es   | de   | el   | bg   | ru   | tr   | ar   | vi   | th   | zh   | hi   | sw   | ur   |
-|---------------------------------------------------------------------------------------------|---------|---------|-------------------------|---------|------|------|------|------|------|------|------|------|------|------|------|------|------|------|------|
-| [mBERT](https://github.com/google-research/bert)                                            | 12      | 768     | 85M                     | 66.3    | 82.1 | 73.8 | 74.3 | 71.1 | 66.4 | 68.9 | 69.0 | 61.6 | 64.9 | 69.5 | 55.8 | 69.3 | 60.0 | 50.4 | 58.0 |
-| [XLM-100](https://github.com/facebookresearch/XLM#pretrained-cross-lingual-language-models) | 16      | 1280    | 315M                    | 70.7    | 83.2 | 76.7 | 77.7 | 74.0 | 72.7 | 74.1 | 72.7 | 68.7 | 68.6 | 72.9 | 68.9 | 72.5 | 65.6 | 58.2 | 62.4 |
-| [XLM-R Base](https://arxiv.org/abs/1911.02116)                                              | 12      | 768     | 85M                     | 74.5    | 84.6 | 78.4 | 78.9 | 76.8 | 75.9 | 77.3 | 75.4 | 73.2 | 71.5 | 75.4 | 72.5 | 74.9 | 71.1 | 65.2 | 66.5 |
-| **mMiniLM-L12xH384**                                                                        | 12      | 384     | 21M                     | 71.1    | 81.5 | 74.8 | 75.7 | 72.9 | 73.0 | 74.5 | 71.3 | 69.7 | 68.8 | 72.1 | 67.8 | 70.0 | 66.2 | 63.3 | 64.2 |
-
-This example code fine-tunes **12**-layer multilingual MiniLM on XNLI.
-
-```bash
-# run fine-tuning on XNLI
-DATA_DIR=/{path_of_data}/
-OUTPUT_DIR=/{path_of_fine-tuned_model}/
-MODEL_PATH=/{path_of_pre-trained_model}/
-
-python ./examples/run_xnli.py --model_type minilm \
- --output_dir ${OUTPUT_DIR} --data_dir ${DATA_DIR} \
- --model_name_or_path microsoft/Multilingual-MiniLM-L12-H384 \
- --tokenizer_name xlm-roberta-base \
- --config_name ${MODEL_PATH}/multilingual-minilm-l12-h384-config.json \
- --do_train \
- --do_eval \
- --max_seq_length 128 \
- --per_gpu_train_batch_size 128 \
- --learning_rate 5e-5 \
- --num_train_epochs 5 \
- --per_gpu_eval_batch_size 32 \
- --weight_decay 0.001 \
- --warmup_steps 500 \
- --save_steps 1500 \
- --logging_steps 1500 \
- --eval_all_checkpoints \
- --language en \
- --fp16 \
- --fp16_opt_level O2
-```
-
-#### Cross-Lingual Question Answering - [MLQA](https://arxiv.org/abs/1910.07475)
-
-Following [Lewis et al. (2019b)](https://arxiv.org/abs/1910.07475), we adopt SQuAD 1.1 as training data and use MLQA English development data for early stopping.
-
-| Model F1 Score                                                                             | #Layers | #Hidden | #Transformer Parameters | Average | en   | es   | de   | ar   | hi   | vi   | zh   |
-|--------------------------------------------------------------------------------------------|---------|---------|-------------------------|---------|------|------|------|------|------|------|------|
-| [mBERT](https://github.com/google-research/bert)                                           | 12      | 768     | 85M                     | 57.7    | 77.7 | 64.3 | 57.9 | 45.7 | 43.8 | 57.1 | 57.5 |
-| [XLM-15](https://github.com/facebookresearch/XLM#pretrained-cross-lingual-language-models) | 12      | 1024    | 151M                    | 61.6    | 74.9 | 68.0 | 62.2 | 54.8 | 48.8 | 61.4 | 61.1 |
-| [XLM-R Base](https://arxiv.org/abs/1911.02116) (Reported)                                  | 12      | 768     | 85M                     | 62.9    | 77.8 | 67.2 | 60.8 | 53.0 | 57.9 | 63.1 | 60.2 |
-| [XLM-R Base](https://arxiv.org/abs/1911.02116) (Our fine-tuned)                            | 12      | 768     | 85M                     | 64.9    | 80.3 | 67.0 | 62.7 | 55.0 | 60.4 | 66.5 | 62.3 |
-| **mMiniLM-L12xH384**                                                                       | 12      | 384     | 21M                     | 63.2    | 79.4 | 66.1 | 61.2 | 54.9 | 58.5 | 63.1 | 59.0 |
-
-### Citation
-
-If you find MiniLM useful in your research, please cite the following paper:
-
-``` latex
-@misc{wang2020minilm,
-    title={MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers},
-    author={Wenhui Wang and Furu Wei and Li Dong and Hangbo Bao and Nan Yang and Ming Zhou},
-    year={2020},
-    eprint={2002.10957},
-    archivePrefix={arXiv},
-    primaryClass={cs.CL}
-}
-```

model_cards/mrm8488/t5-base-finetuned-emotion/README.md

@@ -1,81 +0,0 @@
---
-language: english
----
-
-# T5-base fine-tuned for Emotion Recognition 😂😢😡😃😯
-
-
-[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) base fine-tuned on [emotion recognition](https://github.com/dair-ai/emotion_dataset) dataset for **Emotion Recognition** downstream task.
-
-## Details of T5
-
-The **T5** model was presented in [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf) by *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu* in Here the abstract:
-
-Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.
-
-![model image](https://camo.githubusercontent.com/623b4dea0b653f2ad3f36c71ebfe749a677ac0a1/68747470733a2f2f6d69726f2e6d656469756d2e636f6d2f6d61782f343030362f312a44304a31674e51663876727255704b657944387750412e706e67)
-
-## Details of the downstream task (Sentiment Recognition) - Dataset 📚
-
-[Elvis Saravia](https://twitter.com/omarsar0) has gathered a great [dataset](https://github.com/dair-ai/emotion_dataset) for emotion recognition. It allows to classifiy the text into one of the following **6** emotions:
-
- - sadness 😢
- - joy 😃
- - love 🥰
- - anger 😡
- - fear 😱
- - surprise 😯
-
-## Model fine-tuning 🏋️‍
-
-The training script is a slightly modified version of [this Colab Notebook](https://github.com/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) created by [Suraj Patil](https://github.com/patil-suraj), so all credits to him!
-
-## Test set metrics 🧾
-
-               |precision | recall  | f1-score |support|
-    |----------|----------|---------|----------|-------|
-    |anger     |      0.93|     0.92|      0.93|    275|
-    |fear      |      0.91|     0.87|      0.89|    224|
-    |joy       |      0.97|     0.94|      0.95|    695|
-    |love      |      0.80|     0.91|      0.85|    159|
-    |sadness   |      0.97|     0.97|      0.97|    521|
-    |surpirse  |      0.73|     0.89|      0.80|     66|
-    |----------|----------|---------|----------|-------|
-    |accuracy|            |         |      0.93|   2000|
-    |macro avg|       0.89|     0.92|      0.90|   2000|
-    |weighted avg|    0.94|     0.93|      0.93|   2000|
-    
-    
-                 
-    
-
-
-
-## Model in Action 🚀
-
-```python
-from transformers import AutoTokenizer, AutoModelWithLMHead
-
-tokenizer = AutoTokenizer.from_pretrained("mrm8488/t5-base-finetuned-emotion")
-
-model = AutoModelWithLMHead.from_pretrained("mrm8488/t5-base-finetuned-emotion")
-
-def get_emotion(text):
-  input_ids = tokenizer.encode(text + '</s>', return_tensors='pt')
-
-  output = model.generate(input_ids=input_ids,
-               max_length=2)
-  
-  dec = [tokenizer.decode(ids) for ids in output]
-  label = dec[0]
-  return label
-  
- get_emotion("i feel as if i havent blogged in ages are at least truly blogged i am doing an update cute") # Output: 'joy'
- 
- get_emotion("i have a feeling i kinda lost my best friend") # Output: 'sadness'
-
-```
-
-> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488) | [LinkedIn](https://www.linkedin.com/in/manuel-romero-cs/)
-
-> Made with <span style="color: #e25555;">&hearts;</span> in Spain

model_cards/mrm8488/t5-base-finetuned-imdb-sentiment/README.md

@@ -1,68 +0,0 @@
---
-language: english
-datasets:
-- imdb
----
-
-# T5-base fine-tuned for Sentiment Anlalysis 🎞️👍👎
-
-
-[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) base fine-tuned on [IMDB](https://huggingface.co/datasets/imdb) dataset for **Sentiment Analysis** downstream task.
-
-## Details of T5
-
-The **T5** model was presented in [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf) by *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu* in Here the abstract:
-
-Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.
-
-![model image](https://camo.githubusercontent.com/623b4dea0b653f2ad3f36c71ebfe749a677ac0a1/68747470733a2f2f6d69726f2e6d656469756d2e636f6d2f6d61782f343030362f312a44304a31674e51663876727255704b657944387750412e706e67)
-
-## Details of the downstream task (Sentiment analysis) - Dataset 📚
-
-[IMDB](https://huggingface.co/datasets/imdb)
-
-This is a dataset for binary sentiment classification containing substantially more data than previous benchmark datasets. We provide a set of **25,000** highly polar movie reviews for training, and **25,000** for testing.
-
-## Model fine-tuning 🏋️‍
-
-The training script is a slightly modified version of [this Colab Notebook](https://github.com/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) created by [Suraj Patil](https://github.com/patil-suraj), so all credits to him!
-
-## Test set metrics 🧾
-
-               |precision | recall  | f1-score |support|
-    |----------|----------|---------|----------|-------|
-    |negative  |     0.95 |     0.95|      0.95|  12500|
-    |positive  |     0.95 |     0.95|      0.95|  12500|
-    |----------|----------|---------|----------|-------|
-    |accuracy|            |         |      0.95|  25000|
-    |macro avg|       0.95|     0.95|      0.95|  25000|
-    |weighted avg|    0.95|     0.95|     0.95 |  25000|
-
-
-## Model in Action 🚀
-
-```python
-from transformers import AutoTokenizer, AutoModelWithLMHead
-
-tokenizer = AutoTokenizer.from_pretrained("mrm8488/t5-base-finetuned-imdb-sentiment")
-
-model = AutoModelWithLMHead.from_pretrained("mrm8488/t5-base-finetuned-imdb-sentiment")
-
-def get_sentiment(text):
-  input_ids = tokenizer.encode(text + '</s>', return_tensors='pt')
-
-  output = model.generate(input_ids=input_ids,
-               max_length=2)
-  
-  dec = [tokenizer.decode(ids) for ids in output]
-  label = dec[0]
-  return label
-  
-get_sentiment("I dislike a lot that film")
-
-# Output: 'negative'
-```
-
-> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488) | [LinkedIn](https://www.linkedin.com/in/manuel-romero-cs/)
-
-> Made with <span style="color: #e25555;">&hearts;</span> in Spain

model_cards/nyu-mll/roberta-base-100M-1/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-base-100M-2/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-base-100M-3/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-base-10M-1/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-base-10M-2/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-base-10M-3/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-base-1B-1/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-base-1B-2/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-base-1B-3/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-med-small-1M-1/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-med-small-1M-2/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta-med-small-1M-3/README.md

@@ -1 +0,0 @@
-../roberta_1M_to_1B/README.md

model_cards/nyu-mll/roberta_1M_to_1B/README.md

@@ -1,49 +0,0 @@
-# RoBERTa Pretrained on Smaller Datasets
-
-We pretrain RoBERTa on smaller datasets (1M, 10M, 100M, 1B tokens). We release 3 models with lowest perplexities for each pretraining data size out of 25 runs (or 10 in the case of 1B tokens). The pretraining data reproduces that of BERT: We combine English Wikipedia and a reproduction of BookCorpus using texts from smashwords in a ratio of approximately 3:1.
-
-### Hyperparameters and Validation Perplexity
-
-The hyperparameters and validation perplexities corresponding to each model are as follows:
-
-| Model Name               | Training Size | Model Size | Max Steps | Batch Size | Validation Perplexity |
-|--------------------------|---------------|------------|-----------|------------|-----------------------|
-| [roberta-base-1B-1][link-roberta-base-1B-1]        | 1B            | BASE       | 100K      | 512        | 3.93                  |
-| [roberta-base-1B-2][link-roberta-base-1B-2]        | 1B            | BASE       | 31K       | 1024       | 4.25                  |
-| [roberta-base-1B-3][link-roberta-base-1B-3]        | 1B            | BASE       | 31K       | 4096       | 3.84                  |
-| [roberta-base-100M-1][link-roberta-base-100M-1]      | 100M          | BASE       | 100K      | 512        | 4.99                  |
-| [roberta-base-100M-2][link-roberta-base-100M-2]      | 100M          | BASE       | 31K       | 1024       | 4.61                  |
-| [roberta-base-100M-3][link-roberta-base-100M-3]      | 100M          | BASE       | 31K       | 512        | 5.02                  |
-| [roberta-base-10M-1][link-roberta-base-10M-1]       | 10M           | BASE       | 10K       | 1024       | 11.31                 |
-| [roberta-base-10M-2][link-roberta-base-10M-2]       | 10M           | BASE       | 10K       | 512        | 10.78                 |
-| [roberta-base-10M-3][link-roberta-base-10M-3]       | 10M           | BASE       | 31K       | 512        | 11.58                 |
-| [roberta-med-small-1M-1][link-roberta-med-small-1M-1]   | 1M            | MED-SMALL  | 100K      | 512        | 153.38                |
-| [roberta-med-small-1M-2][link-roberta-med-small-1M-2]   | 1M            | MED-SMALL  | 10K       | 512        | 134.18                |
-| [roberta-med-small-1M-3][link-roberta-med-small-1M-3]   | 1M            | MED-SMALL  | 31K       | 512        | 139.39                |
-
-The hyperparameters corresponding to model sizes mentioned above are as follows:
-
-| Model Size | L  | AH | HS  | FFN  | P    |
-|------------|----|----|-----|------|------|
-| BASE       | 12 | 12 | 768 | 3072 | 125M |
-| MED-SMALL  | 6  | 8  | 512 | 2048 | 45M  |
-
-(AH = number of attention heads; HS = hidden size; FFN = feedforward network dimension; P = number of parameters.)
-
-For other hyperparameters, we select:
-- Peak Learning rate: 5e-4
-- Warmup Steps: 6% of max steps
-- Dropout: 0.1
-
-[link-roberta-med-small-1M-1]: https://huggingface.co/nyu-mll/roberta-med-small-1M-1
-[link-roberta-med-small-1M-2]: https://huggingface.co/nyu-mll/roberta-med-small-1M-2
-[link-roberta-med-small-1M-3]: https://huggingface.co/nyu-mll/roberta-med-small-1M-3
-[link-roberta-base-10M-1]: https://huggingface.co/nyu-mll/roberta-base-10M-1
-[link-roberta-base-10M-2]: https://huggingface.co/nyu-mll/roberta-base-10M-2
-[link-roberta-base-10M-3]: https://huggingface.co/nyu-mll/roberta-base-10M-3
-[link-roberta-base-100M-1]: https://huggingface.co/nyu-mll/roberta-base-100M-1
-[link-roberta-base-100M-2]: https://huggingface.co/nyu-mll/roberta-base-100M-2
-[link-roberta-base-100M-3]: https://huggingface.co/nyu-mll/roberta-base-100M-3
-[link-roberta-base-1B-1]: https://huggingface.co/nyu-mll/roberta-base-1B-1
-[link-roberta-base-1B-2]: https://huggingface.co/nyu-mll/roberta-base-1B-2
-[link-roberta-base-1B-3]: https://huggingface.co/nyu-mll/roberta-base-1B-3

notebooks/02-transformers.ipynb

@@ -221,7 +221,7 @@
    },
    "source": [
     "The code you saw in the previous section introduced all the steps required to do simple model invocation.\n",
-    "For more day-to-day usage, transformers provides you higher-level methods which will makes your NLP journey easier.\n",
+    "For more day-to-day usage, transformers provides you higher-level methods which will makes your NLP journey easier\n",
     "Let's improve our previous example"
    ]
   },
@@ -268,7 +268,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "As you can see above, the method `encode_plus` provides a convenient way to generate all the required parameters\n",
+    "As you can see above, the methode `encode_plus` provides a convenient way to generate all the required parameters\n",
     "that will go through the model. \n",
     "\n",
     "Moreover, you might have noticed it generated some additional tensors: \n",

notebooks/README.md

@@ -37,4 +37,3 @@ Pull Request so it can be included under the Community notebooks.
 | [Fine-tune DistilBert for Multiclass Classification](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_multiclass_classification.ipynb) | How to fine-tune DistilBert for multiclass classification with PyTorch | [Abhishek Kumar Mishra](https://github.com/abhimishra91) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_multiclass_classification.ipynb)|
 |[Fine-tune BERT for Multi-label Classification](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_multi_label_classification.ipynb)|How to fine-tune BERT for multi-label classification using PyTorch|[Abhishek Kumar Mishra](https://github.com/abhimishra91) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_multi_label_classification.ipynb)|
 |[Fine-tune T5 for Summarization](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_summarization_wandb.ipynb)|How to fine-tune T5 for summarization in PyTorch and track experiments with WandB|[Abhishek Kumar Mishra](https://github.com/abhimishra91) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_summarization_wandb.ipynb)|
-|[Speed up Fine-Tuning in Transformers with Dynamic Padding / Bucketing](https://github.com/ELS-RD/transformers-notebook/blob/master/Divide_Hugging_Face_Transformers_training_time_by_2_or_more.ipynb)|How to speed up fine-tuning by a factor of 2 using dynamic padding / bucketing|[Michael Benesty](https://github.com/pommedeterresautee) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1CBfRU1zbfu7-ijiOqAAQUA-RJaxfcJoO?usp=sharing)|

rouge_test_df.csv

@@ -1,40 +0,0 @@
-exp_name,avg_len,R2_56,R1,R2,RL
-dl6_no_teacher,89.17162750217581,,0.4426298740974814,0.21214627660372445,0.30352876002780776
-cnn_f12_9_noteach,93.90400348128807,,0.44021766057145817,0.2112684144707048,0.30203109862281485
-baseline_140,83.3504,0.2012,0.4405,0.2106,0.3063
-dl6_yes_teacher,65.3845953002611,,0.4358592269823179,0.2094872879851778,0.3055267315672064
-cnn_12_9_no_teacher,65.63707571801567,,0.4349178365295365,0.20912713468933475,0.30577661504959625
-cnn_f12_9,66.25691906005223,,0.4360757701741097,0.20874355812934306,0.30534340603938714
-dl6_ckpt,64.7923,0.211,0.4344,0.2087,0.3049
-run_6l,64.7923,0.2109,0.4344,0.2086,0.3049
-dl6_4ep,66.4003,0.2112,0.4351,0.2082,0.3047
-cnn_9_9_yes_teacher,65.75987815491732,,0.43407500508924857,0.2077387062275759,0.3041112175512638
-dl6_maxlen_140_gacc2,65.2891,0.2099,0.433,0.2072,0.3038
-brewer_cnn_12_6_v2,58.29869451697128,,0.4275596591964453,0.2061398841242274,0.3043299614540603
-cnn_enc_only_6_12_brutasse,78.37284595300261,,0.4352447538697767,0.2057028264435689,0.3015751874385552
-cnn_no_teacher_f12_3,84.8461270670148,,0.4341908103753934,0.20524584699497528,0.29823737701046504
-cnn_6_6_yes_teacher,66.56605744125326,,0.4271799135757925,0.2017168386739932,0.2971576748941705
-el6_12_fix,94.3361183637946,,0.4295203853115971,0.20114288940092984,0.29082690956646995
-dl3,56.8062,0.2047,0.4197,0.2002,0.2989
-dl3_4ep,56.585,0.2057,0.4192,0.1998,0.3
-pseudo_dd_140,102.0577,0.1864,0.4254,0.1997,0.2886
-dl6_4ep_sasha_ds,60.0613,0.2082,0.4217,0.1993,0.2965
-cnn_enc_only_3_12_brutasse,75.5237597911227,,0.4278349430460213,0.19820455870913525,0.2945791235929408
-pseudo_6l_140_fix,62.5668,0.2035,0.4197,0.1968,0.2941
-staged_cnn_6_6,89.18093994778067,,0.4255412775299003,0.19658951474224035,0.2885276829133131
-el9_12,66.26144473455179,,0.42086970025135134,0.1948512909973004,0.29011441089763124
-el9_9,65.4414273281114,,0.419385405970708,0.1941718068172067,0.2900372909854251
-dl2_4ep,56.013054830287196,,0.4097263081894015,0.1919076904112108,0.29258786794511826
-dl3_4ep_sasha_ds,56.9671,0.2007,0.4111,0.1912,0.2914
-dl2_4ep_sasha_ds,55.999129677980854,,0.4057876788402908,0.1871524671665716,0.28735289627716404
-brewer_cnn_12_1_v4,56.39904264577894,,0.4031071697939543,0.18464802550618464,0.2839190569931007
-blarge_12_6_no_teacher,98.89782419495216,,0.4056381963795483,0.1826793955605352,0.2700670904734034
-2l_96,56.0208,0.1808,0.395,0.1799,0.2799
-2l_140,56.0208,0.1808,0.3951,0.1799,0.2798
-2l_56,56.0336,0.1762,0.3895,0.1758,0.2762
-2l_eval_140,56.0336,0.1762,0.3897,0.1757,0.2762
-el6_12,66.68172323759791,,0.4013734110868821,0.1752702303482339,0.2711560926634892
-el6_6,57.5147954743255,,0.38798696359999946,0.1672810503629001,0.2648421408851355
-dl1,56.0179,0.1264,0.3411,0.1285,0.226
-last_layer,55.6623,0.0596,0.2654,0.0616,0.1764
-cnn_enc_only_9_12_brutasse,122.32297650130549,,0.12110269669223532,0.011040620587715645,0.09900079897379864

rouge_test_df_xsum.csv

@@ -1,30 +0,0 @@
-exp_name,avg_len,R1,R2,RL
-brewer_xsum_12_6,27.251213270978557,0.45257522039021847,0.2195833165680984,0.3679937489447971
-xsum_baseline,29.06997264625431,0.45234885716476203,0.21848808526257507,0.3650268517677788
-xsum_f12_9,29.35074561016501,0.451213614023561,0.2163378899615816,0.363539265208816
-brewer_xsum_9_6,27.35771640342363,0.4492537548158509,0.2162358630026469,0.3641728953722555
-xsum_no_teacher_12_6,27.75319862348893,0.4486508407826425,0.2149919319149121,0.36419730252620985
-xsum_clean_baseline,29.314656313420983,0.4477747607857255,0.21439354376130093,0.3593759295843251
-brewer_xsum_12_3,24.899232330362658,0.44368280928513437,0.21344371120165062,0.3630886297392266
-xsum_f12_9_noteach,29.46845495455749,0.4472052566240151,0.21274301655819566,0.3593763343512352
-xsum_12_9_yes_teacher_mlm_low,29.737845230742078,0.4465496602405639,0.2123133092772882,0.35962822466043964
-brewer_xsum_9_9_v8.2,27.76872849201447,0.4443257538168877,0.21127297599326367,0.3582479791521699
-xsum_theseus_12_6_copy,26.884320127062562,0.4424918692803468,0.21067481121631593,0.3583553330701445
-brewer_xsum_9_6_gaccum,27.142063001853,0.4425169154001685,0.2085901108525016,0.3575200377038006
-xsum_dl6,27.928086120180005,0.4428406843456235,0.20797826365160146,0.3570522063056493
-brewer_xsum_6_6_v3,27.265684284831906,0.43971173638675576,0.20748396024553475,0.3548038868729512
-xsum_no_teacher_f12_6,27.745874878672904,0.4410138016184357,0.20718399111770808,0.3554781824669484
-xsum_9_9_encloss10,28.9296744021883,0.4375314579000953,0.2042900182424192,0.3507282669577781
-xsum_enc_only_9_12_v8,27.90267360804729,0.4317712176227767,0.1975871213673145,0.3445507930002034
-xsum_no_teacher_f12_3,25.100238242301245,0.4212459339899797,0.19363224077581093,0.3425534541928968
-xsum_enc_only_6_12_brutasse,28.53339804111885,0.4268136827870216,0.19356708317121352,0.3391381862315658
-xsum_9_12_noteacher,29.28192005647225,0.4197414451047208,0.1873359194795431,0.33264079827956505
-eval_xsum_9_9_no_teacher,28.378540545310155,0.4154520215812441,0.1838431267358389,0.3296500161194308
-xsum_teacher_f12_3,22.863937174622784,0.4099647311777451,0.1838389871897501,0.3350523158599553
-xsum_9_9_no_teacher,28.378540545310155,0.4155138730166103,0.1837896883494328,0.3296699124599892
-xsum_6_6_yes_teacher,27.355245742521838,0.4156289447797342,0.1833851442061078,0.3312280244052153
-brewer_xsum_3_3_v2,24.55272213888644,0.4066228194707484,0.1820601762610753,0.3289002584850459
-xsum_theseus_6_6_v2,27.172240360010587,0.3843975566079777,0.16290567759393396,0.3068524775579614
-brewer_xsum_6_6_fix2,26.134651019147622,0.38678850503910334,0.15887838687468492,0.3076899937354824
-staged_xsum_9_9,27.46360187064325,0.3735212492621201,0.14991821061038266,0.2934425279687253
-brewer_xsum_6_6_newcode,59.600811788582014,0.24447308274539264,0.06887251729221451,0.17840798937780264

setup.py

@@ -85,8 +85,7 @@ extras["serving"] = ["pydantic", "uvicorn", "fastapi", "starlette"]
 extras["all"] = extras["serving"] + ["tensorflow", "torch"]
 
 extras["testing"] = ["pytest", "pytest-xdist", "timeout-decorator", "psutil"]
-# sphinx-rtd-theme==0.5.0 introduced big changes in the style.
-extras["docs"] = ["recommonmark", "sphinx", "sphinx-markdown-tables", "sphinx-rtd-theme==0.4.3"]
+extras["docs"] = ["recommonmark", "sphinx", "sphinx-markdown-tables", "sphinx-rtd-theme"]
 extras["quality"] = [
     "black",
     "isort @ git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort",
@@ -109,7 +108,7 @@ setup(
     packages=find_packages("src"),
     install_requires=[
         "numpy",
-        "tokenizers == 0.8.0-rc3",
+        "tokenizers == 0.8.0-rc1",
         # dataclasses for Python versions that don't have it
         "dataclasses;python_version<'3.7'",
         # utilities from PyPA to e.g. compare versions

src/transformers/__init__.py

@@ -34,7 +34,6 @@ from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
 from .configuration_longformer import LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, LongformerConfig
 from .configuration_marian import MarianConfig
 from .configuration_mmbt import MMBTConfig
-from .configuration_mobilebert import MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MobileBertConfig
 from .configuration_openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig
 from .configuration_reformer import REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, ReformerConfig
 from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RetriBertConfig
@@ -133,7 +132,6 @@ from .tokenization_electra import ElectraTokenizer, ElectraTokenizerFast
 from .tokenization_flaubert import FlaubertTokenizer
 from .tokenization_gpt2 import GPT2Tokenizer, GPT2TokenizerFast
 from .tokenization_longformer import LongformerTokenizer, LongformerTokenizerFast
-from .tokenization_mobilebert import MobileBertTokenizer, MobileBertTokenizerFast
 from .tokenization_openai import OpenAIGPTTokenizer, OpenAIGPTTokenizerFast
 from .tokenization_reformer import ReformerTokenizer
 from .tokenization_retribert import RetriBertTokenizer, RetriBertTokenizerFast
@@ -193,21 +191,6 @@ if is_torch_available():
         MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
     )
 
-    from .modeling_mobilebert import (
-        MobileBertPreTrainedModel,
-        MobileBertModel,
-        MobileBertForPreTraining,
-        MobileBertForSequenceClassification,
-        MobileBertForQuestionAnswering,
-        MobileBertForMaskedLM,
-        MobileBertForNextSentencePrediction,
-        MobileBertForMultipleChoice,
-        MobileBertForTokenClassification,
-        load_tf_weights_in_mobilebert,
-        MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
-        MobileBertLayer,
-    )
-
     from .modeling_bert import (
         BertPreTrainedModel,
         BertModel,
@@ -353,7 +336,6 @@ if is_torch_available():
         ElectraForMaskedLM,
         ElectraForTokenClassification,
         ElectraPreTrainedModel,
-        ElectraForMultipleChoice,
         ElectraForSequenceClassification,
         ElectraForQuestionAnswering,
         ElectraModel,
@@ -516,20 +498,6 @@ if is_tf_available():
         TFGPT2PreTrainedModel,
     )
 
-    from .modeling_tf_mobilebert import (
-        TF_MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
-        TFMobileBertModel,
-        TFMobileBertPreTrainedModel,
-        TFMobileBertForPreTraining,
-        TFMobileBertForSequenceClassification,
-        TFMobileBertForQuestionAnswering,
-        TFMobileBertForMaskedLM,
-        TFMobileBertForNextSentencePrediction,
-        TFMobileBertForMultipleChoice,
-        TFMobileBertForTokenClassification,
-        TFMobileBertMainLayer,
-    )
-
     from .modeling_tf_openai import (
         TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST,
         TFOpenAIGPTDoubleHeadsModel,

src/transformers/benchmark/benchmark.py

@@ -20,7 +20,6 @@
 
 import logging
 import timeit
-from typing import Callable, Optional
 
 from transformers import (
     MODEL_MAPPING,
@@ -30,14 +29,7 @@ from transformers import (
     is_torch_available,
 )
 
-from .benchmark_utils import (
-    Benchmark,
-    Memory,
-    MemorySummary,
-    measure_peak_memory_cpu,
-    start_memory_tracing,
-    stop_memory_tracing,
-)
+from .benchmark_utils import Benchmark, Memory, measure_peak_memory_cpu, start_memory_tracing, stop_memory_tracing
 
 
 if is_torch_available():
@@ -62,27 +54,23 @@ class PyTorchBenchmark(Benchmark):
     def framework_version(self):
         return torch.__version__
 
-    def _inference_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
+    def _inference_speed(self, model_name, batch_size, sequence_length):
         _inference = self._prepare_inference_func(model_name, batch_size, sequence_length)
         return self._measure_speed(_inference)
 
-    def _inference_memory(
-        self, model_name: str, batch_size: int, sequence_length: int
-    ) -> [Memory, Optional[MemorySummary]]:
+    def _inference_memory(self, model_name, batch_size, sequence_length):
         _inference = self._prepare_inference_func(model_name, batch_size, sequence_length)
         return self._measure_memory(_inference)
 
-    def _train_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
+    def _train_speed(self, model_name, batch_size, sequence_length):
         _train = self._prepare_train_func(model_name, batch_size, sequence_length)
         return self._measure_speed(_train)
 
-    def _train_memory(
-        self, model_name: str, batch_size: int, sequence_length: int
-    ) -> [Memory, Optional[MemorySummary]]:
+    def _train_memory(self, model_name, batch_size, sequence_length):
         _train = self._prepare_train_func(model_name, batch_size, sequence_length)
         return self._measure_memory(_train)
 
-    def _prepare_inference_func(self, model_name: str, batch_size: int, sequence_length: int) -> Callable[[], None]:
+    def _prepare_inference_func(self, model_name, batch_size, sequence_length):
         config = self.config_dict[model_name]
 
         if self.args.torchscript:
@@ -125,7 +113,7 @@ class PyTorchBenchmark(Benchmark):
         _forward = encoder_decoder_forward if config.is_encoder_decoder else encoder_forward
         return _forward
 
-    def _prepare_train_func(self, model_name: str, batch_size: int, sequence_length: int) -> Callable[[], None]:
+    def _prepare_train_func(self, model_name, batch_size, sequence_length):
         config = self.config_dict[model_name]
         model = MODEL_WITH_LM_HEAD_MAPPING[config.__class__](config)
 
@@ -166,7 +154,7 @@ class PyTorchBenchmark(Benchmark):
         )
         return _train
 
-    def _measure_speed(self, func) -> float:
+    def _measure_speed(self, func):
         try:
             if self.args.is_tpu or self.args.torchscript:
                 # run additional 10 times to stabilize compilation for tpu and torchscript
@@ -188,7 +176,7 @@ class PyTorchBenchmark(Benchmark):
             self.print_fn("Doesn't fit on GPU. {}".format(e))
             return "N/A"
 
-    def _measure_memory(self, func: Callable[[], None]) -> [Memory, MemorySummary]:
+    def _measure_memory(self, func):
         try:
             if self.args.trace_memory_line_by_line:
                 trace = start_memory_tracing("transformers")
@@ -196,7 +184,7 @@ class PyTorchBenchmark(Benchmark):
             if self.args.is_tpu:
                 # tpu
                 raise NotImplementedError(
-                    "Memory Benchmarking is currently not implemented for TPU. Please disable memory benchmarking with `--no_memory` or `args.no_memory=True`"
+                    "Memory Benchmarking is currently not implemented for TPU. Please disable memory benchmarking with `args.no_memory=True`"
                 )
             elif self.args.is_gpu:
                 if not is_py3nvml_available():

src/transformers/benchmark/benchmark_args_utils.py

@@ -114,9 +114,6 @@ class BenchmarkArguments:
 
     @property
     def model_names(self):
-        assert (
-            len(self.models) > 0
-        ), "Please make sure you provide at least one model name / model identifier, *e.g.* `--models bert-base-cased` or `args.models = ['bert-base-cased']."
         return self.models
 
     @property

src/transformers/benchmark/benchmark_tf.py

@@ -22,7 +22,6 @@ import logging
 import random
 import timeit
 from functools import wraps
-from typing import Callable, Optional
 
 from transformers import (
     TF_MODEL_MAPPING,
@@ -32,14 +31,7 @@ from transformers import (
     is_tf_available,
 )
 
-from .benchmark_utils import (
-    Benchmark,
-    Memory,
-    MemorySummary,
-    measure_peak_memory_cpu,
-    start_memory_tracing,
-    stop_memory_tracing,
-)
+from .benchmark_utils import Benchmark, Memory, measure_peak_memory_cpu, start_memory_tracing, stop_memory_tracing
 
 
 if is_tf_available():
@@ -53,20 +45,20 @@ if is_py3nvml_available():
 logger = logging.getLogger(__name__)
 
 
-def run_with_tf_optimizations(do_eager_mode: bool, use_xla: bool):
+def run_with_tf_optimizations(do_eager_mode, do_xla):
     def run_func(func):
         @wraps(func)
         def run_in_eager_mode(*args, **kwargs):
             return func(*args, **kwargs)
 
         @wraps(func)
-        @tf.function(experimental_compile=use_xla)
+        @tf.function(experimental_compile=do_xla)
         def run_in_graph_mode(*args, **kwargs):
             return func(*args, **kwargs)
 
         if do_eager_mode is True:
             assert (
-                use_xla is False
+                do_xla is False
             ), "Cannot run model in XLA, if `args.eager_mode` is set to `True`. Please set `args.eager_mode=False`."
             return run_in_eager_mode
         else:
@@ -75,7 +67,7 @@ def run_with_tf_optimizations(do_eager_mode: bool, use_xla: bool):
     return run_func
 
 
-def random_input_ids(batch_size: int, sequence_length: int, vocab_size: int) -> ["tf.Tensor"]:
+def random_input_ids(batch_size, sequence_length, vocab_size):
     rng = random.Random()
     values = [rng.randint(0, vocab_size - 1) for i in range(batch_size * sequence_length)]
     return tf.constant(values, shape=(batch_size, sequence_length), dtype=tf.int32)
@@ -91,7 +83,7 @@ class TensorflowBenchmark(Benchmark):
     def framework_version(self):
         return tf.__version__
 
-    def _inference_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
+    def _inference_speed(self, model_name, batch_size, sequence_length):
         # initialize GPU on separate process
         strategy = self.args.strategy
         assert strategy is not None, "A device strategy has to be initialized before using Tensorflow."
@@ -103,9 +95,7 @@ class TensorflowBenchmark(Benchmark):
             "Training is currently not really implemented." "Wait for TFTrainer to support CLM and MLM."
         )
 
-    def _inference_memory(
-        self, model_name: str, batch_size: int, sequence_length: int
-    ) -> [Memory, Optional[MemorySummary]]:
+    def _inference_memory(self, model_name, batch_size, sequence_length):
         # initialize GPU on separate process
         if self.args.is_gpu:
             tf.config.experimental.set_memory_growth(self.args.gpu_list[self.args.device_idx], True)
@@ -119,7 +109,7 @@ class TensorflowBenchmark(Benchmark):
             "Training is currently not really implemented. Wait for TFTrainer to support CLM and MLM."
         )
 
-    def _prepare_inference_func(self, model_name: str, batch_size: int, sequence_length: int) -> Callable[[], None]:
+    def _prepare_inference_func(self, model_name, batch_size, sequence_length):
         config = self.config_dict[model_name]
 
         if self.args.fp16:
@@ -146,7 +136,7 @@ class TensorflowBenchmark(Benchmark):
 
         return _inference
 
-    def _measure_speed(self, func) -> float:
+    def _measure_speed(self, func):
         with self.args.strategy.scope():
             try:
                 if self.args.is_tpu or self.args.use_xla:
@@ -161,7 +151,7 @@ class TensorflowBenchmark(Benchmark):
             except ResourceExhaustedError as e:
                 self.print_fn("Doesn't fit on GPU. {}".format(e))
 
-    def _measure_memory(self, func: Callable[[], None]) -> [Memory, MemorySummary]:
+    def _measure_memory(self, func):
         logger.info(
             "Note that Tensorflow allocates more memory than"
             "it might need to speed up computation."

src/transformers/benchmark/benchmark_utils.py

@@ -61,23 +61,11 @@ BenchmarkOutput = namedtuple(
 )
 
 
-def separate_process_wrapper_fn(func: Callable[[], None], do_multi_processing: bool) -> Callable[[], None]:
-    """
-        This function wraps another function into its own separated process.
-        In order to ensure accurate memory measurements it is important that the function
-        is executed in a separate process
-
-        Args:
-            - `func`: (`callable`): function() -> ...
-                generic function which will be executed in its own separate process
-            - `do_multi_processing`: (`bool`)
-                Whether to run function on separate process or not
-    """
-
+def separate_process_wrapper_fn(func, do_multi_processing):
     def multi_process_func(*args, **kwargs):
         # run function in an individual
         # process to get correct memory
-        def wrapper_func(queue: Queue, *args):
+        def wrapper_func(queue, *args):
             try:
                 result = func(*args)
             except Exception as e:
@@ -94,7 +82,7 @@ def separate_process_wrapper_fn(func: Callable[[], None], do_multi_processing: b
         return result
 
     if do_multi_processing:
-        logging.info("fFunction {func} is executed in its own process...")
+        logging.info("Use multiprocessing...")
         return multi_process_func
     else:
         return func
@@ -602,35 +590,31 @@ class Benchmark(ABC):
         pass
 
     @abstractmethod
-    def _inference_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
+    def _inference_speed(self, model_name, batch_size, sequence_length):
         pass
 
     @abstractmethod
-    def _train_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
+    def _train_speed(self, model_name, batch_size, sequence_length):
         pass
 
     @abstractmethod
-    def _inference_memory(
-        self, model_name: str, batch_size: int, sequence_length: int
-    ) -> [Memory, Optional[MemorySummary]]:
+    def _inference_memory(self, model_name, batch_size, sequence_length):
         pass
 
     @abstractmethod
-    def _train_memory(
-        self, model_name: str, batch_size: int, sequence_length: int
-    ) -> [Memory, Optional[MemorySummary]]:
+    def _train_memory(self, model_name, batch_size, sequence_length):
         pass
 
-    def inference_speed(self, *args, **kwargs) -> float:
+    def inference_speed(self, *args, **kwargs):
         return separate_process_wrapper_fn(self._inference_speed, self.args.do_multi_processing)(*args, **kwargs)
 
-    def train_speed(self, *args, **kwargs) -> float:
+    def train_speed(self, *args, **kwargs):
         return separate_process_wrapper_fn(self._train_speed, self.args.do_multi_processing)(*args, **kwargs)
 
-    def inference_memory(self, *args, **kwargs) -> [Memory, Optional[MemorySummary]]:
+    def inference_memory(self, *args, **kwargs):
         return separate_process_wrapper_fn(self._inference_memory, self.args.do_multi_processing)(*args, **kwargs)
 
-    def train_memory(self, *args, **kwargs) -> [Memory, Optional[MemorySummary]]:
+    def train_memory(self, *args, **kwargs):
         return separate_process_wrapper_fn(self._train_memory, self.args.do_multi_processing)(*args, **kwargs)
 
     def run(self):

src/transformers/configuration_auto.py

@@ -19,7 +19,7 @@ import logging
 from collections import OrderedDict
 
 from .configuration_albert import ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, AlbertConfig
-from .configuration_bart import BART_PRETRAINED_CONFIG_ARCHIVE_MAP, BartConfig, MBartConfig
+from .configuration_bart import BART_PRETRAINED_CONFIG_ARCHIVE_MAP, BartConfig
 from .configuration_bert import BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, BertConfig
 from .configuration_camembert import CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, CamembertConfig
 from .configuration_ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig
@@ -30,7 +30,6 @@ from .configuration_flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, Flau
 from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
 from .configuration_longformer import LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, LongformerConfig
 from .configuration_marian import MarianConfig
-from .configuration_mobilebert import MobileBertConfig
 from .configuration_openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig
 from .configuration_reformer import ReformerConfig
 from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RetriBertConfig
@@ -76,13 +75,11 @@ CONFIG_MAPPING = OrderedDict(
     [
         ("retribert", RetriBertConfig,),
         ("t5", T5Config,),
-        ("mobilebert", MobileBertConfig,),
         ("distilbert", DistilBertConfig,),
         ("albert", AlbertConfig,),
         ("camembert", CamembertConfig,),
         ("xlm-roberta", XLMRobertaConfig,),
         ("marian", MarianConfig,),
-        ("mbart", MBartConfig,),
         ("bart", BartConfig,),
         ("reformer", ReformerConfig,),
         ("longformer", LongformerConfig,),

src/transformers/configuration_bart.py

@@ -41,7 +41,6 @@ class BartConfig(PretrainedConfig):
     def __init__(
         self,
         activation_dropout=0.0,
-        extra_pos_embeddings=2,
         activation_function="gelu",
         vocab_size=50265,
         d_model=1024,
@@ -69,9 +68,6 @@ class BartConfig(PretrainedConfig):
         normalize_embedding=True,
         static_position_embeddings=False,
         add_bias_logits=False,
-        student_decoder_layers=None,
-        student_encoder_layers=None,
-        replacing_rate=0,
         **common_kwargs
     ):
         r"""
@@ -90,7 +86,6 @@ class BartConfig(PretrainedConfig):
             is_encoder_decoder=is_encoder_decoder,
             **common_kwargs,
         )
-        self.replacing_rate = replacing_rate
         self.vocab_size = vocab_size
         self.d_model = d_model  # encoder_embed_dim and decoder_embed_dim
         self.encoder_ffn_dim = encoder_ffn_dim
@@ -123,12 +118,6 @@ class BartConfig(PretrainedConfig):
         # Classifier stuff
         self.classif_dropout = classifier_dropout
 
-        self.extra_pos_embeddings = self.pad_token_id + 1
-
-        # Theseus params
-        self.student_encoder_layers = student_encoder_layers
-        self.student_decoder_layers = student_decoder_layers
-
     @property
     def num_attention_heads(self) -> int:
         return self.encoder_attention_heads
@@ -144,7 +133,3 @@ class BartConfig(PretrainedConfig):
         if self.normalize_before or self.add_final_layer_norm or self.scale_embedding:
             logger.info("This configuration is a mixture of MBART and BART settings")
         return False
-
-
-class MBartConfig(BartConfig):
-    model_type = "mbart"

src/transformers/configuration_bert.py

@@ -90,8 +90,6 @@ class BertConfig(PretrainedConfig):
                 The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
             layer_norm_eps (:obj:`float`, optional, defaults to 1e-12):
                 The epsilon used by the layer normalization layers.
-            gradient_checkpointing (:obj:`bool`, optional, defaults to False):
-                If True, use gradient checkpointing to save memory at the expense of slower backward pass.
 
         Example::
 
@@ -123,7 +121,6 @@ class BertConfig(PretrainedConfig):
         initializer_range=0.02,
         layer_norm_eps=1e-12,
         pad_token_id=0,
-        gradient_checkpointing=False,
         **kwargs
     ):
         super().__init__(pad_token_id=pad_token_id, **kwargs)
@@ -140,4 +137,3 @@ class BertConfig(PretrainedConfig):
         self.type_vocab_size = type_vocab_size
         self.initializer_range = initializer_range
         self.layer_norm_eps = layer_norm_eps
-        self.gradient_checkpointing = gradient_checkpointing

src/transformers/configuration_electra.py

@@ -76,28 +76,6 @@ class ElectraConfig(PretrainedConfig):
                 The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
             layer_norm_eps (:obj:`float`, optional, defaults to 1e-12):
                 The epsilon used by the layer normalization layers.
-            summary_type (:obj:`string`, optional, defaults to "first"):
-                Argument used when doing sequence summary. Used in for the multiple choice head in
-                :class:`~transformers.ElectraForMultipleChoice`.
-                Is one of the following options:
-
-                    - 'last' => take the last token hidden state (like XLNet)
-                    - 'first' => take the first token hidden state (like Bert)
-                    - 'mean' => take the mean of all tokens hidden states
-                    - 'cls_index' => supply a Tensor of classification token position (GPT/GPT-2)
-                    - 'attn' => Not implemented now, use multi-head attention
-            summary_use_proj (:obj:`boolean`, optional, defaults to :obj:`True`):
-                Argument used when doing sequence summary. Used in for the multiple choice head in
-                :class:`~transformers.ElectraForMultipleChoice`.
-                Add a projection after the vector extraction
-            summary_activation (:obj:`string` or :obj:`None`, optional, defaults to :obj:`None`):
-                Argument used when doing sequence summary. Used in for the multiple choice head in
-                :class:`~transformers.ElectraForMultipleChoice`.
-                'gelu' => add a gelu activation to the output, Other => no activation.
-            summary_last_dropout (:obj:`float`, optional, defaults to 0.0):
-                Argument used when doing sequence summary. Used in for the multiple choice head in
-                :class:`~transformers.ElectraForMultipleChoice`.
-                Add a dropout after the projection and activation
 
         Example::
 
@@ -129,10 +107,6 @@ class ElectraConfig(PretrainedConfig):
         type_vocab_size=2,
         initializer_range=0.02,
         layer_norm_eps=1e-12,
-        summary_type="first",
-        summary_use_proj=True,
-        summary_activation="gelu",
-        summary_last_dropout=0.1,
         pad_token_id=0,
         **kwargs
     ):
@@ -151,8 +125,3 @@ class ElectraConfig(PretrainedConfig):
         self.type_vocab_size = type_vocab_size
         self.initializer_range = initializer_range
         self.layer_norm_eps = layer_norm_eps
-
-        self.summary_type = summary_type
-        self.summary_use_proj = summary_use_proj
-        self.summary_activation = summary_activation
-        self.summary_last_dropout = summary_last_dropout

src/transformers/configuration_encoder_decoder.py

@@ -37,7 +37,7 @@ class EncoderDecoderConfig(PretrainedConfig):
                 Remaining dictionary of keyword arguments. Notably:
                     encoder (:class:`PretrainedConfig`, optional, defaults to `None`):
                         An instance of a configuration object that defines the encoder config.
-                    decoder (:class:`PretrainedConfig`, optional, defaults to `None`):
+                    encoder (:class:`PretrainedConfig`, optional, defaults to `None`):
                         An instance of a configuration object that defines the decoder config.
 
         Example::

src/transformers/configuration_mobilebert.py

@@ -1,159 +0,0 @@
-# coding=utf-8
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" MobileBERT model configuration """
-
-import logging
-
-from .configuration_utils import PretrainedConfig
-
-
-logger = logging.getLogger(__name__)
-
-MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "mobilebert-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/google/mobilebert-uncased/config.json"
-}
-
-
-class MobileBertConfig(PretrainedConfig):
-    r"""
-        This is the configuration class to store the configuration of a :class:`~transformers.MobileBertModel`.
-        It is used to instantiate a MobileBERT model according to the specified arguments, defining the model
-        architecture.
-
-        Configuration objects inherit from  :class:`~transformers.PretrainedConfig` and can be used
-        to control the model outputs. Read the documentation from  :class:`~transformers.PretrainedConfig`
-        for more information.
-
-
-        Args:
-            vocab_size (:obj:`int`, optional, defaults to 30522):
-                Vocabulary size of the MobileBERT model. Defines the different tokens that
-                can be represented by the `inputs_ids` passed to the forward method of :class:`~transformers.MobileBertModel`.
-            hidden_size (:obj:`int`, optional, defaults to 512):
-                Dimensionality of the encoder layers and the pooler layer.
-            num_hidden_layers (:obj:`int`, optional, defaults to 24):
-                Number of hidden layers in the Transformer encoder.
-            num_attention_heads (:obj:`int`, optional, defaults to 4):
-                Number of attention heads for each attention layer in the Transformer encoder.
-            intermediate_size (:obj:`int`, optional, defaults to 512):
-                Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
-            hidden_act (:obj:`str` or :obj:`function`, optional, defaults to "relu"):
-                The non-linear activation function (function or string) in the encoder and pooler.
-                If string, "gelu", "relu", "swish" and "gelu_new" are supported.
-            hidden_dropout_prob (:obj:`float`, optional, defaults to 0.0):
-                The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
-            attention_probs_dropout_prob (:obj:`float`, optional, defaults to 0.1):
-                The dropout ratio for the attention probabilities.
-            max_position_embeddings (:obj:`int`, optional, defaults to 512):
-                The maximum sequence length that this model might ever be used with.
-                Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
-            type_vocab_size (:obj:`int`, optional, defaults to 2):
-                The vocabulary size of the `token_type_ids` passed into :class:`~transformers.MobileBertModel`.
-            initializer_range (:obj:`float`, optional, defaults to 0.02):
-                The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-            layer_norm_eps (:obj:`float`, optional, defaults to 1e-12):
-                The epsilon used by the layer normalization layers.
-
-            pad_token_id (:obj:`int`, optional, defaults to 0):
-                The ID of the token in the word embedding to use as padding.
-            embedding_size (:obj:`int`, optional, defaults to 128):
-                The dimension of the word embedding vectors.
-            trigram_input (:obj:`bool`, optional, defaults to True):
-                Use a convolution of trigram as input.
-            use_bottleneck (:obj:`bool`, optional, defaults to True):
-                Whether to use bottleneck in BERT.
-            intra_bottleneck_size (:obj:`int`, optional, defaults to 128):
-                Size of bottleneck layer output.
-            use_bottleneck_attention (:obj:`bool`, optional, defaults to False):
-                Whether to use attention inputs from the bottleneck transformation.
-            key_query_shared_bottleneck (:obj:`bool`, optional, defaults to True):
-                Whether to use the same linear transformation for query&key in the bottleneck.
-            num_feedforward_networks (:obj:`int`, optional, defaults to 4):
-                Number of FFNs in a block.
-            normalization_type (:obj:`str`, optional, defaults to "no_norm"):
-                The normalization type in BERT.
-
-        Example:
-
-            from transformers import MobileBertModel, MobileBertConfig
-
-            # Initializing a MobileBERT configuration
-            configuration = MobileBertConfig()
-
-            # Initializing a model from the configuration above
-            model = MobileBertModel(configuration)
-
-            # Accessing the model configuration
-            configuration = model.config
-
-        Attributes:
-            pretrained_config_archive_map (Dict[str, str]):
-                A dictionary containing all the available pre-trained checkpoints.
-    """
-    pretrained_config_archive_map = MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
-    model_type = "mobilebert"
-
-    def __init__(
-        self,
-        vocab_size=30522,
-        hidden_size=512,
-        num_hidden_layers=24,
-        num_attention_heads=4,
-        intermediate_size=512,
-        hidden_act="relu",
-        hidden_dropout_prob=0.0,
-        attention_probs_dropout_prob=0.1,
-        max_position_embeddings=512,
-        type_vocab_size=2,
-        initializer_range=0.02,
-        layer_norm_eps=1e-12,
-        pad_token_id=0,
-        embedding_size=128,
-        trigram_input=True,
-        use_bottleneck=True,
-        intra_bottleneck_size=128,
-        use_bottleneck_attention=False,
-        key_query_shared_bottleneck=True,
-        num_feedforward_networks=4,
-        normalization_type="no_norm",
-        classifier_activation=True,
-        **kwargs
-    ):
-        super().__init__(pad_token_id=pad_token_id, **kwargs)
-
-        self.vocab_size = vocab_size
-        self.hidden_size = hidden_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.hidden_act = hidden_act
-        self.intermediate_size = intermediate_size
-        self.hidden_dropout_prob = hidden_dropout_prob
-        self.attention_probs_dropout_prob = attention_probs_dropout_prob
-        self.max_position_embeddings = max_position_embeddings
-        self.type_vocab_size = type_vocab_size
-        self.initializer_range = initializer_range
-        self.layer_norm_eps = layer_norm_eps
-        self.embedding_size = embedding_size
-        self.trigram_input = trigram_input
-        self.use_bottleneck = use_bottleneck
-        self.intra_bottleneck_size = intra_bottleneck_size
-        self.use_bottleneck_attention = use_bottleneck_attention
-        self.key_query_shared_bottleneck = key_query_shared_bottleneck
-        self.num_feedforward_networks = num_feedforward_networks
-        self.normalization_type = normalization_type
-        self.classifier_activation = classifier_activation
-
-        if self.use_bottleneck:
-            self.true_hidden_size = intra_bottleneck_size
-        else:
-            self.true_hidden_size = hidden_size

src/transformers/configuration_reformer.py

@@ -97,7 +97,7 @@ class ReformerConfig(PretrainedConfig):
                 Number of following neighbouring chunks to attend to in LocalSelfAttention layer in addition to itself.
             local_attention_probs_dropout_prob (:obj:`float`, optional, defaults to 0.1):
                 The dropout ratio for the attention probabilities in LocalSelfAttention.
-            lsh_attn_chunk_length (:obj:`int`, optional, defaults to 64):
+            lsh_chunk_length (:obj:`int`, optional, defaults to 64):
                 Length of chunk which attends to itself in LSHSelfAttention. Chunking reduces memory complexity from sequence length x sequence length (self attention) to chunk length x chunk length x sequence length / chunk length (chunked self attention).
             lsh_num_chunks_before (:obj:`int`, optional, defaults to 1):
                 Number of previous neighbouring chunks to attend to in LSHSelfAttention layer to itself.

src/transformers/configuration_t5.py

@@ -41,8 +41,6 @@ class T5Config(PretrainedConfig):
             vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `T5Model`.
             d_model: Size of the encoder layers and the pooler layer. `d_model` can also accesed via the property `hidden_size`.
             num_layers: Number of hidden layers in the Transformer encoder. `num_layers` can also be accessed via the property `num_hidden_layers`.
-            d_kv: Size of the key, query, value projections per attention head. `d_kv` has to be equal to `d_model // num_heads`.
-            d_ff: Size of the intermediate feed forward layer in each `T5Block`.
             num_heads: Number of attention heads for each attention layer in
                 the Transformer encoder. `num_heads` can also be accessed via the property `num_attention_heads`.
             intermediate_size: The size of the "intermediate" (i.e., feed-forward)

src/transformers/convert_mobilebert_original_tf_checkpoint_to_pytorch.py

@@ -1,42 +0,0 @@
-import argparse
-import logging
-
-import torch
-
-from transformers import MobileBertConfig, MobileBertForPreTraining, load_tf_weights_in_mobilebert
-
-
-logging.basicConfig(level=logging.INFO)
-
-
-def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, mobilebert_config_file, pytorch_dump_path):
-    # Initialise PyTorch model
-    config = MobileBertConfig.from_json_file(mobilebert_config_file)
-    print("Building PyTorch model from configuration: {}".format(str(config)))
-    model = MobileBertForPreTraining(config)
-    # Load weights from tf checkpoint
-    model = load_tf_weights_in_mobilebert(model, config, tf_checkpoint_path)
-    # Save pytorch-model
-    print("Save PyTorch model to {}".format(pytorch_dump_path))
-    torch.save(model.state_dict(), pytorch_dump_path)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    # Required parameters
-    parser.add_argument(
-        "--tf_checkpoint_path", default=None, type=str, required=True, help="Path to the TensorFlow checkpoint path."
-    )
-    parser.add_argument(
-        "--mobilebert_config_file",
-        default=None,
-        type=str,
-        required=True,
-        help="The config json file corresponding to the pre-trained MobileBERT model. \n"
-        "This specifies the model architecture.",
-    )
-    parser.add_argument(
-        "--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
-    )
-    args = parser.parse_args()
-    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path, args.mobilebert_config_file, args.pytorch_dump_path)

src/transformers/data/data_collator.py

@@ -42,10 +42,10 @@ def default_data_collator(features: List[InputDataClass]) -> Dict[str, torch.Ten
     # Special handling for labels.
     # Ensure that tensor is created with the correct type
     # (it should be automatically the case, but let's make sure of it.)
-    if "label" in first and first["label"] is not None:
+    if "label" in first:
         dtype = torch.long if type(first["label"]) is int else torch.float
         batch["labels"] = torch.tensor([f["label"] for f in features], dtype=dtype)
-    elif "label_ids" in first and first["label_ids"] is not None:
+    elif "label_ids" in first:
         if isinstance(first["label_ids"], torch.Tensor):
             batch["labels"] = torch.stack([f["label_ids"] for f in features])
         else:

src/transformers/data/datasets/glue.py

@@ -9,7 +9,6 @@ import torch
 from filelock import FileLock
 from torch.utils.data.dataset import Dataset
 
-from ...tokenization_bart import BartTokenizer, BartTokenizerFast
 from ...tokenization_roberta import RobertaTokenizer, RobertaTokenizerFast
 from ...tokenization_utils import PreTrainedTokenizer
 from ...tokenization_xlm_roberta import XLMRobertaTokenizer
@@ -93,8 +92,6 @@ class GlueDataset(Dataset):
             RobertaTokenizer,
             RobertaTokenizerFast,
             XLMRobertaTokenizer,
-            BartTokenizer,
-            BartTokenizerFast,
         ):
             # HACK(label indices are swapped in RoBERTa pretrained model)
             label_list[1], label_list[2] = label_list[2], label_list[1]

src/transformers/file_utils.py

@@ -16,7 +16,7 @@ from contextlib import contextmanager
 from functools import partial, wraps
 from hashlib import sha256
 from pathlib import Path
-from typing import Dict, Optional, Union
+from typing import Optional
 from urllib.parse import urlparse
 from zipfile import ZipFile, is_zipfile
 
@@ -271,7 +271,7 @@ def cached_path(
     force_download=False,
     proxies=None,
     resume_download=False,
-    user_agent: Union[Dict, str, None] = None,
+    user_agent=None,
     extract_compressed_file=False,
     force_extract=False,
     local_files_only=False,
@@ -357,7 +357,7 @@ def cached_path(
     return output_path
 
 
-def http_get(url, temp_file, proxies=None, resume_size=0, user_agent: Union[Dict, str, None] = None):
+def http_get(url, temp_file, proxies=None, resume_size=0, user_agent=None):
     ua = "transformers/{}; python/{}".format(__version__, sys.version.split()[0])
     if is_torch_available():
         ua += "; torch/{}".format(torch.__version__)
@@ -397,7 +397,7 @@ def get_from_cache(
     proxies=None,
     etag_timeout=10,
     resume_download=False,
-    user_agent: Union[Dict, str, None] = None,
+    user_agent=None,
     local_files_only=False,
 ) -> Optional[str]:
     """

src/transformers/modeling_albert.py

@@ -269,9 +269,7 @@ class AlbertLayer(nn.Module):
         self.ffn_output = nn.Linear(config.intermediate_size, config.hidden_size)
         self.activation = ACT2FN[config.hidden_act]
 
-    def forward(
-        self, hidden_states, attention_mask=None, head_mask=None, output_attentions=False, output_hidden_states=False
-    ):
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, output_attentions=False):
         attention_output = self.attention(hidden_states, attention_mask, head_mask, output_attentions)
         ffn_output = self.ffn(attention_output[0])
         ffn_output = self.activation(ffn_output)
@@ -285,11 +283,10 @@ class AlbertLayerGroup(nn.Module):
     def __init__(self, config):
         super().__init__()
 
+        self.output_hidden_states = config.output_hidden_states
         self.albert_layers = nn.ModuleList([AlbertLayer(config) for _ in range(config.inner_group_num)])
 
-    def forward(
-        self, hidden_states, attention_mask=None, head_mask=None, output_attentions=False, output_hidden_states=False
-    ):
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, output_attentions=False):
         layer_hidden_states = ()
         layer_attentions = ()
 
@@ -300,11 +297,11 @@ class AlbertLayerGroup(nn.Module):
             if output_attentions:
                 layer_attentions = layer_attentions + (layer_output[1],)
 
-            if output_hidden_states:
+            if self.output_hidden_states:
                 layer_hidden_states = layer_hidden_states + (hidden_states,)
 
         outputs = (hidden_states,)
-        if output_hidden_states:
+        if self.output_hidden_states:
             outputs = outputs + (layer_hidden_states,)
         if output_attentions:
             outputs = outputs + (layer_attentions,)
@@ -316,17 +313,16 @@ class AlbertTransformer(nn.Module):
         super().__init__()
 
         self.config = config
+        self.output_hidden_states = config.output_hidden_states
         self.embedding_hidden_mapping_in = nn.Linear(config.embedding_size, config.hidden_size)
         self.albert_layer_groups = nn.ModuleList([AlbertLayerGroup(config) for _ in range(config.num_hidden_groups)])
 
-    def forward(
-        self, hidden_states, attention_mask=None, head_mask=None, output_attentions=False, output_hidden_states=False
-    ):
+    def forward(self, hidden_states, attention_mask=None, head_mask=None, output_attentions=False):
         hidden_states = self.embedding_hidden_mapping_in(hidden_states)
 
         all_attentions = ()
 
-        if output_hidden_states:
+        if self.output_hidden_states:
             all_hidden_states = (hidden_states,)
 
         for i in range(self.config.num_hidden_layers):
@@ -341,18 +337,17 @@ class AlbertTransformer(nn.Module):
                 attention_mask,
                 head_mask[group_idx * layers_per_group : (group_idx + 1) * layers_per_group],
                 output_attentions,
-                output_hidden_states,
             )
             hidden_states = layer_group_output[0]
 
             if output_attentions:
                 all_attentions = all_attentions + layer_group_output[-1]
 
-            if output_hidden_states:
+            if self.output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
         outputs = (hidden_states,)
-        if output_hidden_states:
+        if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if output_attentions:
             outputs = outputs + (all_attentions,)
@@ -494,7 +489,6 @@ class AlbertModel(AlbertPreTrainedModel):
         head_mask=None,
         inputs_embeds=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
     Return:
@@ -510,12 +504,12 @@ class AlbertModel(AlbertPreTrainedModel):
             This output is usually *not* a good summary
             of the semantic content of the input, you're often better with averaging or pooling
             the sequence of hidden-states for the whole input sequence.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -536,9 +530,6 @@ class AlbertModel(AlbertPreTrainedModel):
         """
 
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -565,11 +556,7 @@ class AlbertModel(AlbertPreTrainedModel):
             input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
         )
         encoder_outputs = self.encoder(
-            embedding_output,
-            extended_attention_mask,
-            head_mask=head_mask,
-            output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
+            embedding_output, extended_attention_mask, head_mask=head_mask, output_attentions=output_attentions,
         )
 
         sequence_output = encoder_outputs[0]
@@ -616,7 +603,6 @@ class AlbertForPreTraining(AlbertPreTrainedModel):
         labels=None,
         sentence_order_label=None,
         output_attentions=None,
-        output_hidden_states=None,
         **kwargs,
     ):
         r"""
@@ -642,12 +628,12 @@ class AlbertForPreTraining(AlbertPreTrainedModel):
         sop_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
             Prediction scores of the next sequence prediction (classification) head (scores of True/False
             continuation before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when :obj:`config.output_hidden_states=True`):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -686,7 +672,6 @@ class AlbertForPreTraining(AlbertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output, pooled_output = outputs[:2]
@@ -773,7 +758,6 @@ class AlbertForMaskedLM(AlbertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
         **kwargs
     ):
         r"""
@@ -791,12 +775,12 @@ class AlbertForMaskedLM(AlbertPreTrainedModel):
             Masked language modeling loss.
         prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -831,7 +815,6 @@ class AlbertForMaskedLM(AlbertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
         sequence_outputs = outputs[0]
 
@@ -873,7 +856,6 @@ class AlbertForSequenceClassification(AlbertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -888,12 +870,12 @@ class AlbertForSequenceClassification(AlbertPreTrainedModel):
             Classification (or regression if config.num_labels==1) loss.
         logits ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
             Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -922,7 +904,6 @@ class AlbertForSequenceClassification(AlbertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         pooled_output = outputs[1]
@@ -972,7 +953,6 @@ class AlbertForTokenClassification(AlbertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -985,12 +965,12 @@ class AlbertForTokenClassification(AlbertPreTrainedModel):
             Classification loss.
         scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`)
             Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1021,7 +1001,6 @@ class AlbertForTokenClassification(AlbertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = outputs[0]
@@ -1073,7 +1052,6 @@ class AlbertForQuestionAnswering(AlbertPreTrainedModel):
         start_positions=None,
         end_positions=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -1093,12 +1071,12 @@ class AlbertForQuestionAnswering(AlbertPreTrainedModel):
             Span-start scores (before SoftMax).
         end_scores: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
             Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1129,7 +1107,6 @@ class AlbertForQuestionAnswering(AlbertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = outputs[0]
@@ -1186,7 +1163,6 @@ class AlbertForMultipleChoice(AlbertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -1202,12 +1178,12 @@ class AlbertForMultipleChoice(AlbertPreTrainedModel):
             `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
 
             Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1252,7 +1228,6 @@ class AlbertForMultipleChoice(AlbertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         pooled_output = outputs[1]

src/transformers/modeling_auto.py

@@ -32,7 +32,6 @@ from .configuration_auto import (
     FlaubertConfig,
     GPT2Config,
     LongformerConfig,
-    MobileBertConfig,
     OpenAIGPTConfig,
     ReformerConfig,
     RetriBertConfig,
@@ -88,7 +87,6 @@ from .modeling_distilbert import (
 )
 from .modeling_electra import (
     ElectraForMaskedLM,
-    ElectraForMultipleChoice,
     ElectraForPreTraining,
     ElectraForQuestionAnswering,
     ElectraForSequenceClassification,
@@ -112,15 +110,6 @@ from .modeling_longformer import (
     LongformerModel,
 )
 from .modeling_marian import MarianMTModel
-from .modeling_mobilebert import (
-    MobileBertForMaskedLM,
-    MobileBertForMultipleChoice,
-    MobileBertForPreTraining,
-    MobileBertForQuestionAnswering,
-    MobileBertForSequenceClassification,
-    MobileBertForTokenClassification,
-    MobileBertModel,
-)
 from .modeling_openai import OpenAIGPTLMHeadModel, OpenAIGPTModel
 from .modeling_reformer import ReformerModel, ReformerModelWithLMHead
 from .modeling_retribert import RetriBertModel
@@ -176,7 +165,6 @@ MODEL_MAPPING = OrderedDict(
         (BertConfig, BertModel),
         (OpenAIGPTConfig, OpenAIGPTModel),
         (GPT2Config, GPT2Model),
-        (MobileBertConfig, MobileBertModel),
         (TransfoXLConfig, TransfoXLModel),
         (XLNetConfig, XLNetModel),
         (FlaubertConfig, FlaubertModel),
@@ -201,7 +189,6 @@ MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
         (BertConfig, BertForPreTraining),
         (OpenAIGPTConfig, OpenAIGPTLMHeadModel),
         (GPT2Config, GPT2LMHeadModel),
-        (MobileBertConfig, MobileBertForPreTraining),
         (TransfoXLConfig, TransfoXLLMHeadModel),
         (XLNetConfig, XLNetLMHeadModel),
         (FlaubertConfig, FlaubertWithLMHeadModel),
@@ -225,7 +212,6 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
         (BertConfig, BertForMaskedLM),
         (OpenAIGPTConfig, OpenAIGPTLMHeadModel),
         (GPT2Config, GPT2LMHeadModel),
-        (MobileBertConfig, MobileBertForMaskedLM),
         (TransfoXLConfig, TransfoXLLMHeadModel),
         (XLNetConfig, XLNetLMHeadModel),
         (FlaubertConfig, FlaubertWithLMHeadModel),
@@ -262,7 +248,6 @@ MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
         (LongformerConfig, LongformerForMaskedLM),
         (RobertaConfig, RobertaForMaskedLM),
         (BertConfig, BertForMaskedLM),
-        (MobileBertConfig, MobileBertForMaskedLM),
         (FlaubertConfig, FlaubertWithLMHeadModel),
         (XLMConfig, XLMWithLMHeadModel),
         (ElectraConfig, ElectraForMaskedLM),
@@ -289,7 +274,6 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
         (RobertaConfig, RobertaForSequenceClassification),
         (BertConfig, BertForSequenceClassification),
         (XLNetConfig, XLNetForSequenceClassification),
-        (MobileBertConfig, MobileBertForSequenceClassification),
         (FlaubertConfig, FlaubertForSequenceClassification),
         (XLMConfig, XLMForSequenceClassification),
         (ElectraConfig, ElectraForSequenceClassification),
@@ -307,7 +291,6 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
         (BertConfig, BertForQuestionAnswering),
         (XLNetConfig, XLNetForQuestionAnsweringSimple),
         (FlaubertConfig, FlaubertForQuestionAnsweringSimple),
-        (MobileBertConfig, MobileBertForQuestionAnswering),
         (XLMConfig, XLMForQuestionAnsweringSimple),
         (ElectraConfig, ElectraForQuestionAnswering),
     ]
@@ -322,7 +305,6 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
         (LongformerConfig, LongformerForTokenClassification),
         (RobertaConfig, RobertaForTokenClassification),
         (BertConfig, BertForTokenClassification),
-        (MobileBertConfig, MobileBertForTokenClassification),
         (XLNetConfig, XLNetForTokenClassification),
         (AlbertConfig, AlbertForTokenClassification),
         (ElectraConfig, ElectraForTokenClassification),
@@ -333,13 +315,11 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
 MODEL_FOR_MULTIPLE_CHOICE_MAPPING = OrderedDict(
     [
         (CamembertConfig, CamembertForMultipleChoice),
-        (ElectraConfig, ElectraForMultipleChoice),
         (XLMRobertaConfig, XLMRobertaForMultipleChoice),
         (LongformerConfig, LongformerForMultipleChoice),
         (RobertaConfig, RobertaForMultipleChoice),
         (BertConfig, BertForMultipleChoice),
         (DistilBertConfig, DistilBertForMultipleChoice),
-        (MobileBertConfig, MobileBertForMultipleChoice),
         (XLNetConfig, XLNetForMultipleChoice),
         (AlbertConfig, AlbertForMultipleChoice),
     ]

src/transformers/modeling_bart.py

@@ -23,13 +23,12 @@ import numpy as np
 import torch
 import torch.nn.functional as F
 from torch import Tensor, nn
-from torch.distributions.bernoulli import Bernoulli
 from torch.nn import CrossEntropyLoss
 
 from .activations import ACT2FN
 from .configuration_bart import BartConfig
 from .file_utils import add_start_docstrings, add_start_docstrings_to_callable
-from .modeling_utils import PreTrainedModel
+from .modeling_utils import PreTrainedModel, create_position_ids_from_input_ids
 
 
 logger = logging.getLogger(__name__)
@@ -45,23 +44,6 @@ BART_PRETRAINED_MODEL_ARCHIVE_LIST = [
 ]
 
 
-def get_layers_to_copy(n_to_get, tot):
-    all_layers = list(range(tot))
-    if tot == 12:  # Alternating for special cases
-        layers_to_copy = {  # maps # layers in student -> which teacher layers to copy
-            6: [0, 2, 4, 7, 9, 11],
-            1: [11],
-            3: [0, 6, 11],
-            2: [0, 11],
-            4: [0, 4, 8, 11],
-            9: [0, 1, 2, 4, 5, 7, 9, 10, 11],
-            12: all_layers,
-        }
-        return layers_to_copy[n_to_get]
-    else:
-        return all_layers[:n_to_get]
-
-
 BART_START_DOCSTRING = r"""
 
     This model is a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`_ sub-class. Use it as a regular PyTorch Module and
@@ -78,9 +60,8 @@ BART_GENERATION_EXAMPLE = r"""
 
         from transformers import BartTokenizer, BartForConditionalGeneration, BartConfig
         # see ``examples/summarization/bart/run_eval.py`` for a longer example
-        model = BartForConditionalGeneration.from_pretrained('facebook/bart-large-cnn')
-        tokenizer = BartTokenizer.from_pretrained('facebook/bart-large-cnn')
-
+        model = BartForConditionalGeneration.from_pretrained('bart-large-cnn')
+        tokenizer = BartTokenizer.from_pretrained('bart-large-cnn')
         ARTICLE_TO_SUMMARIZE = "My friends are cool but they eat too many carbs."
         inputs = tokenizer.batch_encode_plus([ARTICLE_TO_SUMMARIZE], max_length=1024, return_tensors='pt')
         # Generate Summary
@@ -115,7 +96,6 @@ BART_INPUTS_DOCSTRING = r"""
 
 
 def invert_mask(attention_mask):
-    """Turns 1->0, 0->1, False->True, True-> False"""
     assert attention_mask.dim() == 2
     return attention_mask.eq(0)
 
@@ -254,37 +234,7 @@ class EncoderLayer(nn.Module):
         return x, attn_weights
 
 
-class TheseusMixin:
-    compress_ratio = 2
-
-    def set_replacing_rate(self, replacing_rate):
-        if not 0 < replacing_rate <= 1:
-            raise Exception("Replace rate must be in the range (0, 1]!")
-        self.bernoulli = Bernoulli(torch.tensor([replacing_rate]))
-
-    def determine_inference_layers(self) -> nn.ModuleList:
-        if self.scc_layers is None:
-            return self.layers
-        if self.training:
-            inference_layers = []
-            for i in range(len(self.scc_layers)):
-                if self.bernoulli.sample() == 1:  # REPLACE
-                    inference_layers.append(self.scc_layers[i])
-                else:  # KEEP the original
-                    for offset in range(self.compress_ratio):
-                        inference_layers.append(self.layers[i * self.compress_ratio + offset])
-
-        else:  # inference with compressed model
-            inference_layers = self.scc_layers
-
-        return inference_layers
-
-    def init_successor_layers(self, replacing_rate):
-        if replacing_rate > 0 and replacing_rate < 1:
-            self.set_replacing_rate(replacing_rate)
-
-
-class BartEncoder(nn.Module, TheseusMixin):
+class BartEncoder(nn.Module):
     """
     Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer
     is a :class:`EncoderLayer`.
@@ -298,6 +248,7 @@ class BartEncoder(nn.Module, TheseusMixin):
 
         self.dropout = config.dropout
         self.layerdrop = config.encoder_layerdrop
+        self.output_hidden_states = config.output_hidden_states
 
         embed_dim = embed_tokens.embedding_dim
         self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
@@ -311,20 +262,14 @@ class BartEncoder(nn.Module, TheseusMixin):
             )
         else:
             self.embed_positions = LearnedPositionalEmbedding(
-                config.max_position_embeddings, embed_dim, self.padding_idx, config.extra_pos_embeddings,
+                config.max_position_embeddings, embed_dim, self.padding_idx,
             )
         self.layers = nn.ModuleList([EncoderLayer(config) for _ in range(config.encoder_layers)])
-
         self.layernorm_embedding = LayerNorm(embed_dim) if config.normalize_embedding else nn.Identity()
         # mbart has one extra layer_norm
         self.layer_norm = LayerNorm(config.d_model) if config.normalize_before else None
-        self.scc_layers = None
-        if config.student_encoder_layers is not None and config.student_encoder_layers < config.encoder_layers:
-            self.scc_layers = nn.ModuleList([EncoderLayer(config) for _ in range(config.student_encoder_layers)])
-            self.compress_ratio = len(self.scc_layers) // len(self.layers)
-        self.init_successor_layers(config.replacing_rate)
 
-    def forward(self, input_ids, attention_mask=None, output_attentions=False, output_hidden_states=False):
+    def forward(self, input_ids, attention_mask=None, output_attentions=False):
         """
         Args:
             input_ids (LongTensor): tokens in the source language of shape
@@ -336,7 +281,7 @@ class BartEncoder(nn.Module, TheseusMixin):
                   shape `(src_len, batch, embed_dim)`
                 - **encoder_states** (List[Tensor]): all intermediate
                   hidden states of shape `(src_len, batch, embed_dim)`.
-                  Only populated if *output_hidden_states:* is True.
+                  Only populated if *self.output_hidden_states:* is True.
                 - **all_attentions** (List[Tensor]): Attention weights for each layer.
                 During training might not be of length n_layers because of layer dropout.
         """
@@ -349,14 +294,13 @@ class BartEncoder(nn.Module, TheseusMixin):
         x = inputs_embeds + embed_pos
         x = self.layernorm_embedding(x)
         x = F.dropout(x, p=self.dropout, training=self.training)
-        inference_layers = self.determine_inference_layers()
 
         # B x T x C -> T x B x C
         x = x.transpose(0, 1)
 
         encoder_states, all_attentions = [], []
-        for encoder_layer in inference_layers:
-            if output_hidden_states:
+        for encoder_layer in self.layers:
+            if self.output_hidden_states:
                 encoder_states.append(x)
             # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
             dropout_probability = random.uniform(0, 1)
@@ -370,7 +314,7 @@ class BartEncoder(nn.Module, TheseusMixin):
 
         if self.layer_norm:
             x = self.layer_norm(x)
-        if output_hidden_states:
+        if self.output_hidden_states:
             encoder_states.append(x)
 
         # T x B x C -> B x T x C
@@ -469,7 +413,7 @@ class DecoderLayer(nn.Module):
         )  # just self_attn weights for now, following t5, layer_state = cache for decoding
 
 
-class BartDecoder(nn.Module, TheseusMixin):
+class BartDecoder(nn.Module):
     """
     Transformer decoder consisting of *config.decoder_layers* layers. Each layer
     is a :class:`DecoderLayer`.
@@ -480,6 +424,7 @@ class BartDecoder(nn.Module, TheseusMixin):
 
     def __init__(self, config: BartConfig, embed_tokens: nn.Embedding):
         super().__init__()
+        self.output_hidden_states = config.output_hidden_states
         self.dropout = config.dropout
         self.layerdrop = config.decoder_layerdrop
         self.padding_idx = embed_tokens.padding_idx
@@ -492,18 +437,13 @@ class BartDecoder(nn.Module, TheseusMixin):
             )
         else:
             self.embed_positions = LearnedPositionalEmbedding(
-                config.max_position_embeddings, config.d_model, self.padding_idx, config.extra_pos_embeddings,
+                config.max_position_embeddings, config.d_model, self.padding_idx,
             )
         self.layers = nn.ModuleList(
             [DecoderLayer(config) for _ in range(config.decoder_layers)]
         )  # type: List[DecoderLayer]
         self.layernorm_embedding = LayerNorm(config.d_model) if config.normalize_embedding else nn.Identity()
         self.layer_norm = LayerNorm(config.d_model) if config.add_final_layer_norm else None
-        self.scc_layers = None
-        if config.student_decoder_layers is not None and config.student_decoder_layers < config.decoder_layers:
-            self.scc_layers = nn.ModuleList([DecoderLayer(config) for _ in range(config.student_decoder_layers)])
-            self.compress_ratio = len(self.scc_layers) // len(self.layers)
-        self.init_successor_layers(config.replacing_rate)
 
     def forward(
         self,
@@ -515,7 +455,6 @@ class BartDecoder(nn.Module, TheseusMixin):
         decoder_cached_states=None,
         use_cache=False,
         output_attentions=False,
-        output_hidden_states=False,
         **unused,
     ):
         """
@@ -561,18 +500,16 @@ class BartDecoder(nn.Module, TheseusMixin):
         all_hidden_states = ()
         all_self_attns = ()
         next_decoder_cache = []
-
-        inference_layers = self.determine_inference_layers()
-        for idx, decoder_layer in enumerate(inference_layers):
+        for idx, decoder_layer in enumerate(self.layers):
             # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
-            if output_hidden_states:
+            if self.output_hidden_states:
                 all_hidden_states += (x,)
             dropout_probability = random.uniform(0, 1)
             if self.training and (dropout_probability < self.layerdrop):
                 continue
 
             layer_state = decoder_cached_states[idx] if decoder_cached_states is not None else None
-            # DecoderLayer.forward()
+
             x, layer_self_attn, layer_past = decoder_layer(
                 x,
                 encoder_hidden_states,
@@ -809,23 +746,23 @@ class LearnedPositionalEmbedding(nn.Embedding):
     position ids are passed to the forward function.
     """
 
-    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, offset):
-        # Bart is set up so that if padding_idx is specified then offset the embedding ids by 2
-        # and adjust num_embeddings appropriately. Other models dont have this hack
-        self.offset = offset
+    def __init__(
+        self, num_embeddings: int, embedding_dim: int, padding_idx: int,
+    ):
+        # if padding_idx is specified then offset the embedding ids by
+        # this index and adjust num_embeddings appropriately
         assert padding_idx is not None
-        num_embeddings += offset
+        num_embeddings += padding_idx + 1  # WHY?
         super().__init__(num_embeddings, embedding_dim, padding_idx=padding_idx)
 
-    def forward(self, input_ids, use_cache=False):
+    def forward(self, input, use_cache=False):
         """Input is expected to be of size [bsz x seqlen]."""
-        bsz, seq_len = input_ids.shape[:2]
-        if use_cache:
-            positions = input_ids.data.new(1, 1).fill_(seq_len - 1)  # called before slicing
+        if use_cache:  # the position is our current step in the decoded sequence
+            pos = int(self.padding_idx + input.size(1))
+            positions = input.data.new(1, 1).fill_(pos)
         else:
-            # starts at 0, ends at 1-seq_len
-            positions = torch.arange(seq_len, dtype=torch.long, device=self.weight.device)
-        return super().forward(positions + self.offset)
+            positions = create_position_ids_from_input_ids(input, self.padding_idx)
+        return super().forward(positions)
 
 
 def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True):
@@ -860,6 +797,8 @@ def _get_shape(t):
 class BartModel(PretrainedBartModel):
     def __init__(self, config: BartConfig):
         super().__init__(config)
+        self.output_hidden_states = config.output_hidden_states
+
         padding_idx, vocab_size = config.pad_token_id, config.vocab_size
         self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
 
@@ -879,12 +818,8 @@ class BartModel(PretrainedBartModel):
         decoder_cached_states=None,
         use_cache=False,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
 
         # make masks if user doesn't supply
         if not use_cache:
@@ -902,10 +837,7 @@ class BartModel(PretrainedBartModel):
 
         if encoder_outputs is None:
             encoder_outputs = self.encoder(
-                input_ids=input_ids,
-                attention_mask=attention_mask,
-                output_attentions=output_attentions,
-                output_hidden_states=output_hidden_states,
+                input_ids=input_ids, attention_mask=attention_mask, output_attentions=output_attentions,
             )
         assert isinstance(encoder_outputs, tuple)
         # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
@@ -917,7 +849,6 @@ class BartModel(PretrainedBartModel):
             decoder_causal_mask=causal_mask,
             decoder_cached_states=decoder_cached_states,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
             use_cache=use_cache,
         )
 
@@ -979,7 +910,6 @@ class BartForConditionalGeneration(PretrainedBartModel):
         labels=None,
         use_cache=False,
         output_attentions=None,
-        output_hidden_states=None,
         **unused,
     ):
         r"""
@@ -996,12 +926,12 @@ class BartForConditionalGeneration(PretrainedBartModel):
             Masked language modeling loss.
         prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1038,7 +968,6 @@ class BartForConditionalGeneration(PretrainedBartModel):
             decoder_attention_mask=decoder_attention_mask,
             decoder_cached_states=decoder_cached_states,
             use_cache=use_cache,
-            output_hidden_states=output_hidden_states,
             output_attentions=output_attentions,
         )
         lm_logits = F.linear(outputs[0], self.model.shared.weight, bias=self.final_logits_bias)
@@ -1053,6 +982,7 @@ class BartForConditionalGeneration(PretrainedBartModel):
 
     def prepare_inputs_for_generation(self, decoder_input_ids, past, attention_mask, use_cache, **kwargs):
         assert past is not None, "past has to be defined for encoder_outputs"
+
         encoder_outputs, decoder_cached_states = past
         return {
             "input_ids": None,  # encoder_outputs is defined. input_ids not needed
@@ -1063,7 +993,7 @@ class BartForConditionalGeneration(PretrainedBartModel):
             "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
         }
 
-    def adjust_logits_during_generation(self, logits, cur_len, max_length):
+    def prepare_logits_for_generation(self, logits, cur_len, max_length):
         if cur_len == 1:
             self._force_token_ids_generation(logits, self.config.bos_token_id)
         if cur_len == max_length - 1 and self.config.eos_token_id is not None:
@@ -1130,7 +1060,6 @@ class BartForSequenceClassification(PretrainedBartModel):
         decoder_attention_mask=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -1140,18 +1069,19 @@ class BartForSequenceClassification(PretrainedBartModel):
 
     Returns:
         :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BartConfig`) and inputs:
-        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
-            Classification loss (cross entropy)
-        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
+            loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
+                Classification loss (cross entropy)
+            logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+                Classification (or regression if config.num_labels==1) scores (before SoftMax).
+            hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+                Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+                of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+                Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+            attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
+                Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+                Attentions weights after the attention softmax, used to compute the weighted average in the
+                self-attention
+                heads.
 
     Examples::
 
@@ -1173,7 +1103,6 @@ class BartForSequenceClassification(PretrainedBartModel):
             decoder_input_ids=decoder_input_ids,
             decoder_attention_mask=decoder_attention_mask,
             encoder_outputs=encoder_outputs,
-            output_hidden_states=output_hidden_states,
             output_attentions=output_attentions,
         )
         x = outputs[0]  # last hidden state
@@ -1219,7 +1148,6 @@ class BartForQuestionAnswering(PretrainedBartModel):
         start_positions=None,
         end_positions=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -1239,12 +1167,15 @@ class BartForQuestionAnswering(PretrainedBartModel):
             Span-start scores (before SoftMax).
         end_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
             Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
             Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
             heads.
 
@@ -1275,7 +1206,6 @@ class BartForQuestionAnswering(PretrainedBartModel):
             decoder_attention_mask=decoder_attention_mask,
             encoder_outputs=encoder_outputs,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = outputs[0]

src/transformers/modeling_bert.py

@@ -22,7 +22,6 @@ import os
 import warnings
 
 import torch
-import torch.utils.checkpoint
 from torch import nn
 from torch.nn import CrossEntropyLoss, MSELoss
 
@@ -392,7 +391,7 @@ class BertLayer(nn.Module):
 class BertEncoder(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.config = config
+        self.output_hidden_states = config.output_hidden_states
         self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
 
     def forward(
@@ -403,50 +402,32 @@ class BertEncoder(nn.Module):
         encoder_hidden_states=None,
         encoder_attention_mask=None,
         output_attentions=False,
-        output_hidden_states=False,
     ):
         all_hidden_states = ()
         all_attentions = ()
         for i, layer_module in enumerate(self.layer):
-            if output_hidden_states:
+            if self.output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
-            if getattr(self.config, "gradient_checkpointing", False):
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs, output_attentions)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(layer_module),
-                    hidden_states,
-                    attention_mask,
-                    head_mask[i],
-                    encoder_hidden_states,
-                    encoder_attention_mask,
-                )
-            else:
-                layer_outputs = layer_module(
-                    hidden_states,
-                    attention_mask,
-                    head_mask[i],
-                    encoder_hidden_states,
-                    encoder_attention_mask,
-                    output_attentions,
-                )
+            layer_outputs = layer_module(
+                hidden_states,
+                attention_mask,
+                head_mask[i],
+                encoder_hidden_states,
+                encoder_attention_mask,
+                output_attentions,
+            )
             hidden_states = layer_outputs[0]
 
             if output_attentions:
                 all_attentions = all_attentions + (layer_outputs[1],)
 
         # Add last layer
-        if output_hidden_states:
+        if self.output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
         outputs = (hidden_states,)
-        if output_hidden_states:
+        if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if output_attentions:
             outputs = outputs + (all_attentions,)
@@ -675,7 +656,6 @@ class BertModel(BertPreTrainedModel):
         encoder_hidden_states=None,
         encoder_attention_mask=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
     Return:
@@ -691,12 +671,12 @@ class BertModel(BertPreTrainedModel):
             This output is usually *not* a good summary
             of the semantic content of the input, you're often better with averaging or pooling
             the sequence of hidden-states for the whole input sequence.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -718,9 +698,6 @@ class BertModel(BertPreTrainedModel):
 
         """
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -770,7 +747,6 @@ class BertModel(BertPreTrainedModel):
             encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=encoder_extended_attention_mask,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
         sequence_output = encoder_outputs[0]
         pooled_output = self.pooler(sequence_output)
@@ -810,7 +786,6 @@ class BertForPreTraining(BertPreTrainedModel):
         labels=None,
         next_sentence_label=None,
         output_attentions=None,
-        output_hidden_states=None,
         **kwargs
     ):
         r"""
@@ -836,12 +811,12 @@ class BertForPreTraining(BertPreTrainedModel):
         seq_relationship_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
             Prediction scores of the next sequence prediction (classification) head (scores of True/False
             continuation before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when :obj:`config.output_hidden_states=True`):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -879,7 +854,6 @@ class BertForPreTraining(BertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output, pooled_output = outputs[:2]
@@ -928,7 +902,6 @@ class BertLMHeadModel(BertPreTrainedModel):
         encoder_hidden_states=None,
         encoder_attention_mask=None,
         output_attentions=None,
-        output_hidden_states=None,
         **kwargs
     ):
         r"""
@@ -946,12 +919,12 @@ class BertLMHeadModel(BertPreTrainedModel):
             Next token prediction loss.
         prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -983,7 +956,6 @@ class BertLMHeadModel(BertPreTrainedModel):
             encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=encoder_attention_mask,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = outputs[0]
@@ -1040,7 +1012,6 @@ class BertForMaskedLM(BertPreTrainedModel):
         encoder_hidden_states=None,
         encoder_attention_mask=None,
         output_attentions=None,
-        output_hidden_states=None,
         **kwargs
     ):
         r"""
@@ -1058,12 +1029,12 @@ class BertForMaskedLM(BertPreTrainedModel):
             Masked language modeling loss.
         prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1103,7 +1074,6 @@ class BertForMaskedLM(BertPreTrainedModel):
             encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=encoder_attention_mask,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = outputs[0]
@@ -1156,7 +1126,6 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
         inputs_embeds=None,
         next_sentence_label=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         next_sentence_label (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -1171,12 +1140,12 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
             Next sequence prediction (classification) loss.
         seq_relationship_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
             Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1207,7 +1176,6 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         pooled_output = outputs[1]
@@ -1250,7 +1218,6 @@ class BertForSequenceClassification(BertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -1265,12 +1232,12 @@ class BertForSequenceClassification(BertPreTrainedModel):
             Classification (or regression if config.num_labels==1) loss.
         logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
             Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1301,7 +1268,6 @@ class BertForSequenceClassification(BertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         pooled_output = outputs[1]
@@ -1350,7 +1316,6 @@ class BertForMultipleChoice(BertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -1366,12 +1331,12 @@ class BertForMultipleChoice(BertPreTrainedModel):
             `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
 
             Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1417,7 +1382,6 @@ class BertForMultipleChoice(BertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         pooled_output = outputs[1]
@@ -1463,7 +1427,6 @@ class BertForTokenClassification(BertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -1476,12 +1439,12 @@ class BertForTokenClassification(BertPreTrainedModel):
             Classification loss.
         scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`)
             Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1512,7 +1475,6 @@ class BertForTokenClassification(BertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = outputs[0]
@@ -1565,7 +1527,6 @@ class BertForQuestionAnswering(BertPreTrainedModel):
         start_positions=None,
         end_positions=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -1585,12 +1546,12 @@ class BertForQuestionAnswering(BertPreTrainedModel):
             Span-start scores (before SoftMax).
         end_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
             Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1625,7 +1586,6 @@ class BertForQuestionAnswering(BertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = outputs[0]

src/transformers/modeling_ctrl.py

@@ -296,6 +296,7 @@ CTRL_INPUTS_DOCSTRING = r"""
 class CTRLModel(CTRLPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
+        self.output_hidden_states = config.output_hidden_states
 
         self.d_model_size = config.n_embd
         self.num_layers = config.n_layer
@@ -337,7 +338,6 @@ class CTRLModel(CTRLPreTrainedModel):
         inputs_embeds=None,
         use_cache=True,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
     Return:
@@ -347,12 +347,12 @@ class CTRLModel(CTRLPreTrainedModel):
         past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
             Contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -374,9 +374,6 @@ class CTRLModel(CTRLPreTrainedModel):
 
         """
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -449,7 +446,7 @@ class CTRLModel(CTRLPreTrainedModel):
         all_hidden_states = ()
         all_attentions = []
         for i, (h, layer_past) in enumerate(zip(self.h, past)):
-            if output_hidden_states:
+            if self.output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)
             outputs = h(
                 hidden_states,
@@ -469,13 +466,13 @@ class CTRLModel(CTRLPreTrainedModel):
 
         hidden_states = self.layernorm(hidden_states)
         hidden_states = hidden_states.view(*output_shape)
-        if output_hidden_states:
+        if self.output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
         outputs = (hidden_states,)
         if use_cache is True:
             outputs = outputs + (presents,)
-        if output_hidden_states:
+        if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if output_attentions:
             # let the number of heads free (-1) so we can extract attention even after head pruning
@@ -521,7 +518,6 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         labels=None,
         use_cache=True,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -540,12 +536,12 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
             Contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -575,7 +571,6 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
             inputs_embeds=inputs_embeds,
             use_cache=use_cache,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         hidden_states = transformer_outputs[0]

src/transformers/modeling_distilbert.py

@@ -259,11 +259,12 @@ class Transformer(nn.Module):
     def __init__(self, config):
         super().__init__()
         self.n_layers = config.n_layers
+        self.output_hidden_states = config.output_hidden_states
 
         layer = TransformerBlock(config)
         self.layer = nn.ModuleList([copy.deepcopy(layer) for _ in range(config.n_layers)])
 
-    def forward(self, x, attn_mask=None, head_mask=None, output_attentions=False, output_hidden_states=False):
+    def forward(self, x, attn_mask=None, head_mask=None, output_attentions=False):
         """
         Parameters
         ----------
@@ -288,7 +289,7 @@ class Transformer(nn.Module):
 
         hidden_state = x
         for i, layer_module in enumerate(self.layer):
-            if output_hidden_states:
+            if self.output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_state,)
 
             layer_outputs = layer_module(
@@ -304,11 +305,11 @@ class Transformer(nn.Module):
                 assert len(layer_outputs) == 1
 
         # Add last layer
-        if output_hidden_states:
+        if self.output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_state,)
 
         outputs = (hidden_state,)
-        if output_hidden_states:
+        if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if output_attentions:
             outputs = outputs + (all_attentions,)
@@ -410,25 +411,19 @@ class DistilBertModel(DistilBertPreTrainedModel):
 
     @add_start_docstrings_to_callable(DISTILBERT_INPUTS_DOCSTRING)
     def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        head_mask=None,
-        inputs_embeds=None,
-        output_attentions=None,
-        output_hidden_states=None,
+        self, input_ids=None, attention_mask=None, head_mask=None, inputs_embeds=None, output_attentions=None,
     ):
         r"""
     Return:
         :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.DistilBertConfig`) and inputs:
         last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
             Sequence of hidden-states at the output of the last layer of the model.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -450,9 +445,6 @@ class DistilBertModel(DistilBertPreTrainedModel):
 
         """
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -474,11 +466,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
         if inputs_embeds is None:
             inputs_embeds = self.embeddings(input_ids)  # (bs, seq_length, dim)
         tfmr_output = self.transformer(
-            x=inputs_embeds,
-            attn_mask=attention_mask,
-            head_mask=head_mask,
-            output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
+            x=inputs_embeds, attn_mask=attention_mask, head_mask=head_mask, output_attentions=output_attentions,
         )
         hidden_state = tfmr_output[0]
         output = (hidden_state,) + tfmr_output[1:]
@@ -492,6 +480,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
 class DistilBertForMaskedLM(DistilBertPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
+        self.output_hidden_states = config.output_hidden_states
 
         self.distilbert = DistilBertModel(config)
         self.vocab_transform = nn.Linear(config.dim, config.dim)
@@ -514,7 +503,6 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
         **kwargs
     ):
         r"""
@@ -532,12 +520,12 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
             Masked language modeling loss.
         prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -570,7 +558,6 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
         hidden_states = dlbrt_output[0]  # (bs, seq_length, dim)
         prediction_logits = self.vocab_transform(hidden_states)  # (bs, seq_length, dim)
@@ -612,7 +599,6 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -627,12 +613,12 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
             Classification (or regression if config.num_labels==1) loss.
         logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
             Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -658,7 +644,6 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
         hidden_state = distilbert_output[0]  # (bs, seq_len, dim)
         pooled_output = hidden_state[:, 0]  # (bs, dim)
@@ -706,7 +691,6 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
         start_positions=None,
         end_positions=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -726,12 +710,12 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
             Span-start scores (before SoftMax).
         end_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
             Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -758,7 +742,6 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
         hidden_states = distilbert_output[0]  # (bs, max_query_len, dim)
 
@@ -814,7 +797,6 @@ class DistilBertForTokenClassification(DistilBertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -827,12 +809,12 @@ class DistilBertForTokenClassification(DistilBertPreTrainedModel):
             Classification loss.
         scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`)
             Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -859,7 +841,6 @@ class DistilBertForTokenClassification(DistilBertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = outputs[0]
@@ -910,7 +891,6 @@ class DistilBertForMultipleChoice(DistilBertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -926,12 +906,12 @@ class DistilBertForMultipleChoice(DistilBertPreTrainedModel):
             `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
 
             Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -973,7 +953,6 @@ class DistilBertForMultipleChoice(DistilBertPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         hidden_state = outputs[0]  # (bs * num_choices, seq_len, dim)

src/transformers/modeling_electra.py

@@ -10,7 +10,6 @@ from .activations import get_activation
 from .configuration_electra import ElectraConfig
 from .file_utils import add_start_docstrings, add_start_docstrings_to_callable
 from .modeling_bert import BertEmbeddings, BertEncoder, BertLayerNorm, BertPreTrainedModel
-from .modeling_utils import SequenceSummary
 
 
 logger = logging.getLogger(__name__)
@@ -273,19 +272,18 @@ class ElectraModel(ElectraPreTrainedModel):
         head_mask=None,
         inputs_embeds=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
     Return:
         :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.ElectraConfig`) and inputs:
         last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
             Sequence of hidden-states at the output of the last layer of the model.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -308,9 +306,6 @@ class ElectraModel(ElectraPreTrainedModel):
         """
 
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -343,7 +338,6 @@ class ElectraModel(ElectraPreTrainedModel):
             attention_mask=extended_attention_mask,
             head_mask=head_mask,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         return hidden_states
@@ -393,7 +387,6 @@ class ElectraForSequenceClassification(ElectraPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -408,12 +401,12 @@ class ElectraForSequenceClassification(ElectraPreTrainedModel):
             Classification (or regression if config.num_labels==1) loss.
         logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
             Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -436,14 +429,7 @@ class ElectraForSequenceClassification(ElectraPreTrainedModel):
 
         """
         discriminator_hidden_states = self.electra(
-            input_ids,
-            attention_mask,
-            token_type_ids,
-            position_ids,
-            head_mask,
-            inputs_embeds,
-            output_attentions,
-            output_hidden_states,
+            input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, output_attentions
         )
 
         sequence_output = discriminator_hidden_states[0]
@@ -491,7 +477,6 @@ class ElectraForPreTraining(ElectraPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (``torch.LongTensor`` of shape ``(batch_size, sequence_length)``, `optional`, defaults to :obj:`None`):
@@ -506,12 +491,12 @@ class ElectraForPreTraining(ElectraPreTrainedModel):
             Total loss of the ELECTRA objective.
         scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`)
             Prediction scores of the head (scores for each token before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when :obj:`config.output_hidden_states=True`):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -535,14 +520,7 @@ class ElectraForPreTraining(ElectraPreTrainedModel):
         """
 
         discriminator_hidden_states = self.electra(
-            input_ids,
-            attention_mask,
-            token_type_ids,
-            position_ids,
-            head_mask,
-            inputs_embeds,
-            output_attentions,
-            output_hidden_states,
+            input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, output_attentions,
         )
         discriminator_sequence_output = discriminator_hidden_states[0]
 
@@ -599,7 +577,6 @@ class ElectraForMaskedLM(ElectraPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
         **kwargs
     ):
         r"""
@@ -617,12 +594,12 @@ class ElectraForMaskedLM(ElectraPreTrainedModel):
             Masked language modeling loss.
         prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -652,14 +629,7 @@ class ElectraForMaskedLM(ElectraPreTrainedModel):
         assert kwargs == {}, f"Unexpected keyword arguments: {list(kwargs.keys())}."
 
         generator_hidden_states = self.electra(
-            input_ids,
-            attention_mask,
-            token_type_ids,
-            position_ids,
-            head_mask,
-            inputs_embeds,
-            output_attentions,
-            output_hidden_states,
+            input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, output_attentions,
         )
         generator_sequence_output = generator_hidden_states[0]
 
@@ -706,7 +676,6 @@ class ElectraForTokenClassification(ElectraPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -719,12 +688,12 @@ class ElectraForTokenClassification(ElectraPreTrainedModel):
             Classification loss.
         scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`)
             Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -748,14 +717,7 @@ class ElectraForTokenClassification(ElectraPreTrainedModel):
         """
 
         discriminator_hidden_states = self.electra(
-            input_ids,
-            attention_mask,
-            token_type_ids,
-            position_ids,
-            head_mask,
-            inputs_embeds,
-            output_attentions,
-            output_hidden_states,
+            input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, output_attentions,
         )
         discriminator_sequence_output = discriminator_hidden_states[0]
 
@@ -813,7 +775,6 @@ class ElectraForQuestionAnswering(ElectraPreTrainedModel):
         start_positions=None,
         end_positions=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -833,12 +794,12 @@ class ElectraForQuestionAnswering(ElectraPreTrainedModel):
             Span-start scores (before SoftMax).
         end_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
             Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -871,7 +832,6 @@ class ElectraForQuestionAnswering(ElectraPreTrainedModel):
             head_mask=head_mask,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = discriminator_hidden_states[0]
@@ -900,115 +860,3 @@ class ElectraForQuestionAnswering(ElectraPreTrainedModel):
             outputs = (total_loss,) + outputs
 
         return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
-
-
-@add_start_docstrings(
-    """ELECTRA Model with a multiple choice classification head on top (a linear layer on top of
-    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
-    ELECTRA_INPUTS_DOCSTRING,
-)
-class ElectraForMultipleChoice(ElectraPreTrainedModel):
-    def __init__(self, config):
-        super().__init__(config)
-
-        self.electra = ElectraModel(config)
-        self.summary = SequenceSummary(config)
-        self.classifier = nn.Linear(config.hidden_size, 1)
-
-        self.init_weights()
-
-    @add_start_docstrings_to_callable(ELECTRA_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        labels=None,
-        output_attentions=None,
-    ):
-        r"""
-        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
-            Labels for computing the multiple choice classification loss.
-            Indices should be in ``[0, ..., num_choices-1]`` where `num_choices` is the size of the second dimension
-            of the input tensors. (see `input_ids` above)
-
-    Returns:
-        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.ElectraConfig`) and inputs:
-        loss (:obj:`torch.FloatTensor` of shape `(1,)`, `optional`, returned when :obj:`labels` is provided):
-            Classification loss.
-        classification_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
-            `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
-
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-
-    Examples::
-
-        from transformers import ElectraTokenizer, ElectraForMultipleChoice
-        import torch
-
-        tokenizer = ElectraTokenizer.from_pretrained('google/electra-base-discriminator')
-        model = ElectraForMultipleChoice.from_pretrained('google/electra-base-discriminator')
-
-        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."
-        choice1 = "It is eaten while held in the hand."
-        labels = torch.tensor(0) # choice0 is correct (according to Wikipedia ;))
-
-        encoding = tokenizer.batch_encode_plus([[prompt, choice0], [prompt, choice1]], return_tensors='pt', pad_to_max_length=True)
-        outputs = model(**{k: v.unsqueeze(0) for k,v in encoding.items()}, labels=labels) # batch size is 1
-
-        # the linear classifier still needs to be trained
-        loss, logits = outputs[:2]
-        """
-        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
-
-        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
-        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
-        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
-        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
-        inputs_embeds = (
-            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
-            if inputs_embeds is not None
-            else None
-        )
-
-        discriminator_hidden_states = self.electra(
-            input_ids,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-            output_attentions=output_attentions,
-        )
-
-        sequence_output = discriminator_hidden_states[0]
-
-        pooled_output = self.summary(sequence_output)
-        logits = self.classifier(pooled_output)
-        reshaped_logits = logits.view(-1, num_choices)
-
-        outputs = (reshaped_logits,) + discriminator_hidden_states[
-            1:
-        ]  # add hidden states and attention if they are here
-
-        if labels is not None:
-            loss_fct = CrossEntropyLoss()
-            loss = loss_fct(reshaped_logits, labels)
-            outputs = (loss,) + outputs
-
-        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)

src/transformers/modeling_flaubert.py

@@ -131,19 +131,18 @@ class FlaubertModel(XLMModel):
         head_mask=None,
         inputs_embeds=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
     Return:
         :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.XLMConfig`) and inputs:
         last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
             Sequence of hidden-states at the output of the last layer of the model.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -163,9 +162,6 @@ class FlaubertModel(XLMModel):
 
         """
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
 
         # removed: src_enc=None, src_len=None
         if input_ids is not None:
@@ -244,7 +240,7 @@ class FlaubertModel(XLMModel):
             if self.training and (dropout_probability < self.layerdrop):
                 continue
 
-            if output_hidden_states:
+            if self.output_hidden_states:
                 hidden_states = hidden_states + (tensor,)
 
             # self attention
@@ -285,7 +281,7 @@ class FlaubertModel(XLMModel):
             tensor *= mask.unsqueeze(-1).to(tensor.dtype)
 
         # Add last hidden state
-        if output_hidden_states:
+        if self.output_hidden_states:
             hidden_states = hidden_states + (tensor,)
 
         # update cache length
@@ -296,7 +292,7 @@ class FlaubertModel(XLMModel):
         # tensor = tensor.transpose(0, 1)
 
         outputs = (tensor,)
-        if output_hidden_states:
+        if self.output_hidden_states:
             outputs = outputs + (hidden_states,)
         if output_attentions:
             outputs = outputs + (attentions,)

src/transformers/modeling_gpt2.py

@@ -347,6 +347,7 @@ GPT2_INPUTS_DOCSTRING = r"""
 class GPT2Model(GPT2PreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
+        self.output_hidden_states = config.output_hidden_states
 
         self.wte = nn.Embedding(config.vocab_size, config.n_embd)
         self.wpe = nn.Embedding(config.n_positions, config.n_embd)
@@ -381,7 +382,6 @@ class GPT2Model(GPT2PreTrainedModel):
         inputs_embeds=None,
         use_cache=True,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
     Return:
@@ -392,7 +392,7 @@ class GPT2Model(GPT2PreTrainedModel):
         past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
             Contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True``) is passed or when ``config.output_hidden_states=True``:
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
@@ -417,9 +417,6 @@ class GPT2Model(GPT2PreTrainedModel):
 
         """
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -489,7 +486,7 @@ class GPT2Model(GPT2PreTrainedModel):
         all_attentions = []
         all_hidden_states = ()
         for i, (block, layer_past) in enumerate(zip(self.h, past)):
-            if output_hidden_states:
+            if self.output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)
 
             outputs = block(
@@ -512,13 +509,13 @@ class GPT2Model(GPT2PreTrainedModel):
 
         hidden_states = hidden_states.view(*output_shape)
         # Add last hidden state
-        if output_hidden_states:
+        if self.output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
         outputs = (hidden_states,)
         if use_cache is True:
             outputs = outputs + (presents,)
-        if output_hidden_states:
+        if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if output_attentions:
             # let the number of heads free (-1) so we can extract attention even after head pruning
@@ -564,7 +561,6 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         labels=None,
         use_cache=True,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -583,12 +579,12 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
             Contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -618,7 +614,6 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
             inputs_embeds=inputs_embeds,
             use_cache=use_cache,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
         hidden_states = transformer_outputs[0]
 
@@ -673,7 +668,6 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         mc_labels=None,
         use_cache=True,
         output_attentions=None,
-        output_hidden_states=None,
         **kwargs
     ):
         r"""
@@ -706,12 +700,12 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
             Contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -760,7 +754,6 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             inputs_embeds=inputs_embeds,
             use_cache=use_cache,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         hidden_states = transformer_outputs[0]

src/transformers/modeling_longformer.py

@@ -587,7 +587,6 @@ class LongformerModel(RobertaModel):
         position_ids=None,
         inputs_embeds=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
 
@@ -595,12 +594,12 @@ class LongformerModel(RobertaModel):
         :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
         prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -628,9 +627,6 @@ class LongformerModel(RobertaModel):
         """
 
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
 
         # padding
         attention_window = (
@@ -672,7 +668,6 @@ class LongformerModel(RobertaModel):
             encoder_hidden_states=None,
             encoder_attention_mask=None,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         # undo padding
@@ -711,7 +706,6 @@ class LongformerForMaskedLM(BertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
         **kwargs
     ):
         r"""
@@ -729,12 +723,12 @@ class LongformerForMaskedLM(BertPreTrainedModel):
             Masked language modeling loss.
         prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -773,7 +767,6 @@ class LongformerForMaskedLM(BertPreTrainedModel):
             position_ids=position_ids,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
         sequence_output = outputs[0]
         prediction_scores = self.lm_head(sequence_output)
@@ -817,7 +810,6 @@ class LongformerForSequenceClassification(BertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -832,12 +824,12 @@ class LongformerForSequenceClassification(BertPreTrainedModel):
             Classification (or regression if config.num_labels==1) loss.
         logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
             Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -872,7 +864,6 @@ class LongformerForSequenceClassification(BertPreTrainedModel):
             position_ids=position_ids,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
         sequence_output = outputs[0]
         logits = self.classifier(sequence_output)
@@ -940,7 +931,6 @@ class LongformerForQuestionAnswering(BertPreTrainedModel):
         start_positions=None,
         end_positions=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -959,15 +949,13 @@ class LongformerForQuestionAnswering(BertPreTrainedModel):
             Span-start scores (before SoftMax).
         end_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
             Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
             Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
             heads.
 
@@ -1009,7 +997,6 @@ class LongformerForQuestionAnswering(BertPreTrainedModel):
             position_ids=position_ids,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = outputs[0]
@@ -1070,7 +1057,6 @@ class LongformerForTokenClassification(BertPreTrainedModel):
         inputs_embeds=None,
         labels=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -1083,12 +1069,12 @@ class LongformerForTokenClassification(BertPreTrainedModel):
             Classification loss.
         scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`)
             Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1117,7 +1103,6 @@ class LongformerForTokenClassification(BertPreTrainedModel):
             position_ids=position_ids,
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = outputs[0]
@@ -1173,7 +1158,6 @@ class LongformerForMultipleChoice(BertPreTrainedModel):
         position_ids=None,
         inputs_embeds=None,
         output_attentions=None,
-        output_hidden_states=None,
     ):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
@@ -1189,12 +1173,12 @@ class LongformerForMultipleChoice(BertPreTrainedModel):
             `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
 
             Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1255,7 +1239,6 @@ class LongformerForMultipleChoice(BertPreTrainedModel):
             global_attention_mask=flat_global_attention_mask,
             inputs_embeds=flat_inputs_embeds,
             output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
         )
         pooled_output = outputs[1]
 

src/transformers/modeling_marian.py

@@ -46,7 +46,7 @@ class MarianMTModel(BartForConditionalGeneration):
 
     """
 
-    def adjust_logits_during_generation(self, logits, cur_len, max_length):
+    def prepare_logits_for_generation(self, logits, cur_len, max_length):
         logits[:, self.config.pad_token_id] = float("-inf")
         if cur_len == max_length - 1 and self.config.eos_token_id is not None:
             self._force_token_ids_generation(logits, self.config.eos_token_id)

src/transformers/modeling_mmbt.py

@@ -163,7 +163,7 @@ class MMBTModel(nn.Module, ModuleUtilsMixin):
                 objective during Bert pretraining. This output is usually *not* a good summary
                 of the semantic content of the input, you're often better with averaging or pooling
                 the sequence of hidden-states for the whole input sequence.
-            **hidden_states**: (`optional`, returned when ``output_hidden_states=True``)
+            **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
                 list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
                 of shape ``(batch_size, sequence_length, hidden_size)``:
                 Hidden-states of the model at the output of each layer plus the initial embedding outputs.
@@ -200,7 +200,6 @@ class MMBTModel(nn.Module, ModuleUtilsMixin):
         inputs_embeds=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
-        output_hidden_states=None,
     ):
 
         if input_ids is not None and inputs_embeds is not None:
@@ -258,7 +257,6 @@ class MMBTModel(nn.Module, ModuleUtilsMixin):
             head_mask=head_mask,
             encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=encoder_extended_attention_mask,
-            output_hidden_states=output_hidden_states,
         )
 
         sequence_output = encoder_outputs[0]
@@ -295,7 +293,7 @@ class MMBTForClassification(nn.Module):
                 Classification (or regression if config.num_labels==1) loss.
             **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
                 Classification (or regression if config.num_labels==1) scores (before SoftMax).
-            **hidden_states**: (`optional`, returned when ``output_hidden_states=True``)
+            **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
                 list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
                 of shape ``(batch_size, sequence_length, hidden_size)``:
                 Hidden-states of the model at the output of each layer plus the initial embedding outputs.