GitLab

* Adding support for multiple mask tokens.

- Original implem: https://github.com/huggingface/transformers/pull/10222

Co-authored-by: njafer <naveen.jafer@oracle.com>

* In order to accomodate optionally multimodal models like Perceiver

we add information to the tasks to specify tasks where we know for sure
if we need the tokenizer/feature_extractor or not.

* Adding info in the documentation about multi masks.

+ marked as experimental.

* Add a copy() to prevent overriding the same tensor over and over.

* Fixup.

* Adding small test for multi mask with real values..
Co-authored-by: njafer <naveen.jafer@oracle.com>

* Adding some slow test to check for perceiver at least from a high level.

* Re-enabling fast tests for Perceiver ImageClassification.

* Perceiver might try to run without Tokenizer (Fast doesn't exist) and
with FeatureExtractor some text only pipelines.

* Oops.

* Adding a comment for `update_config_with_model_class`.

* Remove `model_architecture` to get `tiny_config`.

* Finalize rebase.

* Smarter way to handle undefined FastTokenizer.

* Remove old code.

* Addressing some nits.

* Don't instantiate `None`.

* First draft

* Improve docstring + clean up tests

* Remove unused code

* Add check in case one doesn't provide a preprocessor

* avoid tf.tile in embeddings

* remove more tf.tile in embeddings

* clean
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Fix doc examples: cannot import name

* remove copy because of some necessary minor changes (maybe add copy to the individual methods instead)

* Keep copy with some modifications
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

- Do not run image-classification pipeline (_CHECKPOINT_FOR_DOC uses the checkpoint for
langage, which cannot load a FeatureExtractor so current logic fails).
- Add a safeguard to not run tests when `tokenizer_class` or
`feature_extractor_class` **are** defined, but cannot be loaded
This happens for Perceiver for the "FastTokenizer" (which doesn't exist
so None) and FeatureExtractor (which does exist but cannot be loaded
because the checkpoint doesn't define one which is reasonable for the
said checkpoint)
- Added `get_vocab` function to `PerceiverTokenizer` since it is used by
`fill-mask` pipeline when the argument `targets` is used to narrow a
subset of possible values.
Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>

* Migrate docs to mdx

* Update TAPAS docs

* Remove lines

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

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

* Add pt/tf switch to code examples

* More improvements

* Improve docstrings

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

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* correct changes

* add comment

* Fixing tests for perceiver (texts)

* For MaskedLM

* Fix doc examples: unexpected keyword argument

* Don't delete token_type_ids from inputs
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

Fix examples: 'CausalLMOutputWithCrossAttentions' object has no attribute 'last_hidden_state' (#14678)
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* add str hub token to repository when provided else fallback to default True

* make style

* Fix doc examples: name '...' is not defined

* remove >>> and ... in some docstrings in visual_bert
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Move pyctcdecode dep

* Fix doc and last objects

* Quality

* Style

* Ignore this black

* [bf16 support] tweaks

* corrections
Co-authored-by: Manuel R. Ciosici <manuelrciosici@gmail.com>

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Fixes in init

* Style

* change args to address overwriting issue

* remove project name from args

* remove passing args as kwargs to experiment object

* remove passing args as kwargs to offline experiment

* fix offline directory assignment in experiment kwargs

* log checkpoint folder on training end

* log entire output_dir as asset folder

* log asset folder  recursively

* end experiment at the end of training

* clean up

* clean up

* Default to always log training assets to Comet when using CometCallback

* change logging training assets to be true when running callback setup

* fix so that experiment always ends when training ends

* styling and quality fixes

* update docstring for COMET_LOG_ASSETS environment variable

* run styling and quality checks

* clean up to docstring

* remove merge markers

* change asset logging to false to avoid hitting max assets per experiment limit

* update training asset description

* fix styling

This reverts commit 0c70f145.

* Added support for other features for already supported models

* Partial support for causal and seq2seq models

* Partial support for causal and seq2seq models

* OnnxSeq2SeqConfigWithPast to support seq2seq models

* Parameterized the onnx tests

* Restored run_mlm.py

* Restored run_mlm.py

* [WIP] BART update

* BART and MBART

* Added comments

* Another sequence length of the past_key_values

* [AutoProcessor] Add Wav2Vec2WithLM & small fix

* revert line removal

* Update src/transformers/__init__.py

* add test

* up

* up

* small fix

tf.matrix_band_part -> tf.linalg.band_part

* First draft

* Style and remove mlm

* Make forward pass work

* More improvements

* More improvements

* Fix bug

* More improvements

* More improvements

* Add PerceiverTokenizer first draft

* Improve conversion script

* More improvements

* Make conversion script work for the encoder

* Make conversion script work with local pickle files

* Style & quality, fix-copies

* Add dummy input to conversion script

* Add absolute position embeddings to TextPreProcessor

* Make forward pass of encoder work

* More improvements

* Move text preprocessor to separate script

* More improvements

* More improvements

* Add post processor

* Make MLM model work

* Style

* Add PerceiverForMaskedLM

* Add PerceiverImagePreprocessor

* Make style

* Make PerceiverForImageClassification work

* More improvements

* More improvements

* Use tokenizer in conversion script

* Use PerceiverForMaskedLM in conversion script

* Define custom PerceiverModelOutput

* Improve PerceiverAttention to make it work for both MLM and image classification

* More improvements

* More improvements

* More improvements to the conversion script

* Make conversion script work for both MLM and image classification

* Add PerceiverFeatureExtractor

* More improvements

* Style and quality

* Add center cropping

* Fix bug

* Small fix

* Add print statement

* Fix bug in image preprocessor

* Fix bug with conversion script

* Make output position embeddings an nn.Parameter layer instead of nn.Embedding

* Comment out print statements

* Add position encoding classes

* More improvements

* Use position_encoding_kwargs

* Add PerceiverForImageClassificationFourier

* Make style & quality

* Add PerceiverForImageClassificationConvProcessing

* Style & quality

* Add flow model

* Move processors to modeling file

* Make position encodings modular

* Make basic decoder use modular position encodings

* Add PerceiverForOpticalFlow to conversion script

* Add AudioPreprocessor

* Make it possible for the basic decoder to use Fourier position embeddings

* Add PerceiverForMultimodalAutoencoding

* Improve model for optical flow

* Improve _build_network_inputs method

* Add print statement

* Fix device issue

* Fix device of Fourier embeddings

* Add print statements for debugging

* Add another print statement

* Add another print statement

* Add another print statement

* Add another print statement

* Improve PerceiverAudioPreprocessor

* Improve conversion script for multimodal modal

* More improvements

* More improvements

* Improve multimodal model

* Make forward pass multimodal model work

* More improvements

* Improve tests

* Fix some more tests

* Add output dataclasses

* Make more tests pass

* Add print statements for debuggin

* Add tests for image classification

* Add PerceiverClassifierOutput

* More improvements

* Make more tests pass for the optical flow model

* Make style & quality

* Small improvements

* Don't support training for optical flow model for now

* Fix _prepare_for_class for tests

* Make more tests pass, add some docs

* Add multimodal model to tests

* Minor fixes

* Fix tests

* Improve conversion script

* Make fixup

* Remove pos_dim argument

* Fix device issue

* Potential fix for OOM

* Revert previous commit

* Fix test_initialization

* Add print statements for debugging

* Fix print statement

* Add print statement

* Add print statement

* Add print statement

* Add print statement

* Add print statement

* Add print statement

* Remove need for output_shape

* Comment out output_shape

* Remove unnecessary code

* Improve docs

* Fix make fixup

* Remove PerceiverTextProcessor from init

* Improve docs

* Small improvement

* Apply first batch of suggestions from code review

* Apply more suggestions from code review

* Update docstrings

* Define dicts beforehand for readability

* Rename task to architecture in conversion script, include PerceiverModel in tests

* Add print statements for debugging

* Fix tests on GPU

* Remove preprocessors, postprocessors and decoders from main init

* Add integration test

* Fix docs

* Replace einops by torch

* Update for new docs frontend

* Rename PerceiverForImageClassification

* Improve docs

* Improve docs

* Improve docs of PerceiverModel

* Fix some more tests

* Improve center_crop

* Add PerceiverForSequenceClassification

* Small improvements

* Fix tests

* Add integration test for optical flow model

* Clean up

* Add tests for tokenizer

* Fix tokenizer by adding special tokens properly

* Fix CI

* up

* up

* up

* make it cleaner

* correct

* make styhahalal

* add more tests

* finish

* small fix

* make style

* up

* tryout to solve cicrle ci

* up

* fix more tests

* fix more tests

* apply sylvains suggestions

* fix import

* correct docs

* add pyctcdecode only to speech tests

* fix more tests

* add tf, flax and pt tests

* add pt

* fix last tests

* fix more tests

* Apply suggestions from code review

* change lines

* Apply suggestions from code review
Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com>

* correct tests

* correct tests

* add doc string
Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com>

* Fixing Dataset for TQA + token-classification.

* Fixing the tests.

* Making sure `offset_mappings` is a valid argument.

* [trainer] conditional ctx managers into one wrapper

* workaround for contextlib.nullcontext for py<3.7

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

* one more autocast

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

Fix a Bug, trainer_seq2seq.py, in the else branch at Line 172, generation_inputs should be a dict (#14546)

* fix bug, trainer_seq2seq.py, Line 172, generation_inputs must be a dict before feeding into self.model.generation()

* fix bug, trainer_seq2seq.py, Line 172, generation_inputs must be a dict before feeding into self.model.generation()

* quick fix SummarizationPipeline error messages

Fix error messages to avoid spam errors, and errors of type:
`Your max_length is set to 50, but you input_length is only 46. You might consider decreasing max_length manually, e.g. summarizer('...', max_length=50)`

* correcto SummarizationPipeline error messages fixes

* [deepspeed] fix load_best_model_at_end

* try with pull_request_target

* revert: try with pull_request_target

* style

* add test

* cleanup

* implement MLukeTokenizer and LukeForMaskedLM

* update tests

* update docs

* add LukeForMaskedLM to check_repo.py

* update README

* fix test and specify the entity pad id in tokenization_(m)luke

* fix EntityPredictionHeadTransform

* add cross_attention_hidden_size to text-2-text encoder-decoder models (PT/Flax)

* for TFEncoderDecoderModel

* add equivalence test for TFEncoderDecoderModel

* fix

* fix failed equivalence tests

* remove unused import

* add detailed comment

* Fix check_equivalence_tf_to_pt by using encoder/decoder

* cleaning

* Use cross_attention_hidden_size in speech-to-text

* clean fast init logging msg in encoder decoder models

* increase tol from 1e-5 to 1e-3 for tf test

* style

* style

* make sure projection layer can run

* remove type conversion + add check

* fix conflict (config.output_hidden_size)

* Remove TF -> PT in check_pt_tf_equivalence for TFEncoderDecoderModel
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Add AutoProcessor class
Init and tests
Add doc
Fix init
Update src/transformers/models/auto/processing_auto.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
Reverts to tokenizer or feature extractor when available
Adapt test

* Revert "Adapt test"

This reverts commit bbdde5fab02465f24b54b227390073082cb32093.

* Revert "Reverts to tokenizer or feature extractor when available"

This reverts commit 77659ff5d21b6cc0baf6f443017e35e056a525bb.

* Don't revert everything Lysandre!
Co-authored-by: Sylvain Gugger <sylvain.gugger@gmail.com>