GitLab

Big refactor, fixes and multi-GPU/TPU support

* Fix do_eval default value in training_args.py

* Update PULL_REQUEST_TEMPLATE.md

* Adding `AutomaticSpeechRecognitionPipeline`.

- Because we added everything to enable this pipeline, we probably
should add it to `transformers`.
- This PR tries to limit the scope and focuses only on the pipeline part
(what should go in, and out).
- The tests are very specific for S2T and Wav2vec2 to make sure both
architectures are supported by the pipeline. We don't use the mixin for
tests right now, because that requires more work in the `pipeline`
function (will be done in a follow up PR).
- Unsure about the "helper" function `ffmpeg_read`. It makes a lot of
  sense from a user perspective, it does not add any additional
dependencies (as in hard dependency, because users can always use their
own load mechanism). Meanwhile, it feels slightly clunky to have so much
optional preprocessing.
- The pipeline is not done to support streaming audio right now.

Future work:

- Add `automatic-speech-recognition` as a `task`. And add the
FeatureExtractor.from_pretrained within `pipeline` function.
- Add small models within tests
- Add the Mixin to tests.
- Make the logic between ForCTC vs ForConditionalGeneration better.

* Update tests/test_pipelines_automatic_speech_recognition.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Adding docs + main import + type checking + LICENSE.

* Doc style !.

* Fixing TYPE_HINT.

* Specifying waveform shape in the docs.

* Adding asserts + specify in the documentation the shape of the input
np.ndarray.

* Update src/transformers/pipelines/automatic_speech_recognition.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Adding require to tests + move the `feature_extractor` doc.
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Implement gradient checkpoinging for T5Stack

* A bit more robust type checking

* Add `gradient_checkpointing` to T5Config

* Formatting

* Set requires_grad only when training

* None return value will only cause problems when training

* Change the output tuple according to `use_cache`

* Enable gradient checkpointing for the decoder

Squashed commit of the following:

commit 658bdd0bd1215353a8770f558bda2ea69a0ad0c7
Author: Ceshine Lee <shuanck@gmail.com>
Date:   Sat Apr 24 14:08:17 2021 +0800

    Only set `require_grad` for gradient checkpointing

commit acaeee6b2e675045fb28ce2176444c1d63e908bd
Author: Ceshine Lee <shuanck@gmail.com>
Date:   Sat Apr 24 13:59:35 2021 +0800

    Make gradient checkpointing work with the decoder

* Formatting

add test for pickle

simplify test

fix test code style

add missing pickle import

fix test

fix test

fix test

Co-authored-by: Frederik Bode <frederik@paperbox.ai>

* Split checkpoint from model_name_or_path in examples

* Address review comments

* Address review comments

Co-authored-by: Michael Benayoun <michael@huggingface.co>

* Reformat to make code clearer

* Reformat to make code clearer

* add attentions & hidden states

* add model outputs + docs

* finish docs

* finish tests

* finish impl

* del @

* finish

* finish

* correct test

* apply sylvains suggestions

* Update src/transformers/models/bert/modeling_flax_bert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

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

* Update tokenization_utils_base.py

* add assertion

* check batch len

* Update src/transformers/tokenization_utils_base.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

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

* fix docs for decoder_input_ids

* revert the changes for bart and mbart

* removed max_len

* removed max_length from BeamSearchScorer

* correct max length

* finish

* del vim

* finish & add test
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* adding Z-inf

* revamp config process

* up version requirement

* wip

* massive rewrite

* cleanup

* cleanup

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* consistent json commas

* act on suggestions

* leave this feature for 0.3.16

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

* fix invalid class name

* proper ref

* proper ref

* Improve documentation for is_split_into_words argument

* Change description wording

* Pass along seed to DistributedSampler

* Add seed to DistributedLengthGroupedSampler

* Add FP16 support for SageMaker MP

* Add print debugs

* Squeeze

* Remove debug statements

* Add defensive check

* Typo

TF BART models - Add `cross_attentions` to model output and fix cross-attention head masking (#10699)

* Add cross_attn_head_mask to BART

* Fix cross_attentions in TFBart-like models

* This commit enables returning of `cross_attentions`
for TFBart-like models

* It also fixes attention head masking in cross-attenion module

* Update TF model templates

* Fix missing , in TF model templates

* Fix typo: congig -> config

* removes the creation of separate config objects and uses the existing ones instead+overwrite resize_token_embeddings from parent class because it is not working for the EncoderDecoderModel

* rollback to current version of the huggingface master branch

* reworked version that ties the encoder and decoder config of the parent encoderdecoder instance

* overwrite of resize_token_embeddings throws an error now

* review comment suggestion
Co-authored-by: Suraj Patil <surajp815@gmail.com>

* implemented warning in case encoderdecoder is created with differing configs of encoderdecoderconfig and decoderconfig or encoderconfig

* added test to avoid diverging configs of wrapper class and wrapped classes

* Update src/transformers/models/encoder_decoder/modeling_encoder_decoder.py

* make style
Co-authored-by: Suraj Patil <surajp815@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Add head_mask & decoder_head_mask + some corrections

* Fix head masking for N-grams

* Enable test_headmasking for encoder and decod

* Fix one typo regarding in modeling_propgetnet.py

* Enable test_headmasking for ProphetNetStandaloneDecoderModelTest
and ProphetNetStandaloneEncoderModelTest in test_modeling_prophetnet.py

* make style

* Fix cross_head_mask

* Fix attention head mask naming

* `cross_head_mask` -> `cross_attn_head_mask`

* `cross_layer_head_mask` -> `cross_attn_layer_head_mask`

* Still need to merge #10605 to master to pass the tests

documentation linked to the parent class PreTrainedTokenizerFast but it should be the slow tokenizer (#11410)

* enable subword regularization.

* fix tokenizer storage

* fix docstring formatting

* Update src/transformers/models/xlm_roberta/tokenization_xlm_roberta.py
Co-authored-by: Stefan Schweter <stefan@schweter.it>

* fix docstring formatting

* add test for subword regularization tokenizer

* improve comments of test

* add sp_model_kwargs

* reformat docstring to match the style

* add some more documentation

* Update src/transformers/models/xlm_roberta/tokenization_xlm_roberta.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* improve docstring

* empty commit to trigger CI

* Update src/transformers/models/xlm_roberta/tokenization_xlm_roberta.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fix docstring formatting for sphinx
Co-authored-by: Stefan Schweter <stefan@schweter.it>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Fix cross-attention head mask for Torch BART models

* Fix head masking for cross-attention module for the following
models: BART, Blenderbot, Blenderbot_small, M2M_100, Marian, MBart,
Pegasus

* Enable test_headmasking for M2M_100 model

* Fix cross_head_mask for FSMT, LED and T5

* This commit fixes `head_mask` for cross-attention modules
in the following models: FSMT, LED, T5

* It also contains some smaller changes in doc so that
it is be perfectly clear the shape of `cross_head_mask`
is the same as of `decoder_head_mask`

* Update template

* Fix template for BartForCausalLM

* Fix cross_head_mask for Speech2Text models

* Fix cross_head_mask in templates

* Fix args order in BartForCausalLM template

* Fix doc in BART templates

* Make more explicit naming

* `cross_head_mask` -> `cross_attn_head_mask`

* `cross_layer_head_mask` -> `cross_attn_layer_head_mask`

* Fix doc

* make style quality

* Fix speech2text docstring

When passing `inputs_embeds` and not `input_ids=None` the generation function fails because `input_ids` is created but the function but it should not.