GitLab

* added LongformerForQuestionAnswering

* add LongformerForQuestionAnswering

* fix import for LongformerForMaskedLM

* add LongformerForQuestionAnswering

* hardcoded sep_token_id

* compute attention_mask if not provided

* combine global_attention_mask with attention_mask when provided

* update example in  docstring

* add assert error messages, better attention combine

* add test for longformerForQuestionAnswering

* typo

* cast gloabl_attention_mask to long

* make style

* Update src/transformers/configuration_longformer.py

* Update src/transformers/configuration_longformer.py

* fix the code quality

* Merge branch 'longformer-for-question-answering' of https://github.com/patil-suraj/transformers

 into longformer-for-question-answering
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* added functionality for electra classification head

* unneeded dropout

* Test ELECTRA for sequence classification

* Style
Co-authored-by: Frankie <frankie@frase.io>
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>

Adds predict stage for glue tasks, and generate result files which can be submitted to gluebenchmark.com (#4463)

* Adds predict stage for glue tasks, and generate result files which could be submitted to gluebenchmark.com website.

* Use Split enum + always output the label name
Co-authored-by: Julien Chaumond <chaumond@gmail.com>

* There is one missing key in BERT

* Correct device for CamemBERT model

* RoBERTa tokenization adding prefix space

* Style

* fix gpu slow tests in pytorch

* change model to device syntax

* first commit

* bug fixes

* better examples

* undo padding

* remove wrong VOCAB_FILES_NAMES

* License

* make style

* make isort happy

* unit tests

* integration test

* make `black` happy by undoing `isort` changes!!

* lint

* no need for the padding value

* batch_size not bsz

* remove unused type casting

* seqlen not seq_len

* staticmethod

* `bert` selfattention instead of `n2`

* uint8 instead of bool + lints

* pad inputs_embeds using embeddings not a constant

* black

* unit test with padding

* fix unit tests

* remove redundant unit test

* upload model weights

* resolve todo

* simpler _mask_invalid_locations without lru_cache + backward compatible masked_fill_

* increase unittest coverage

* Distributed eval: SequentialDistributedSampler + gather all results

* For consistency only write to disk from world_master

Close https://github.com/huggingface/transformers/issues/4272

* Working distributed eval

* Hook into scripts

* Fix #3721 again

* TPU.mesh_reduce: stay in tensor space

Thanks @jysohn23

* Just a small comment

* whitespace

* torch.hub: pip install packaging

* Add test scenarii

* Test case for #3936

* multigpu tests pass on pytorch 1.4.0

* Fixup

* multigpu tests pass on pytorch 1.5.0

* Update src/transformers/modeling_utils.py

* Update src/transformers/modeling_utils.py

* rename multigpu to require_multigpu

* mode doc

* fix fp16 in t5

* make style

* refactor invert_attention_mask fn

* fix typo

* Add index to be returned by NerPipeline to allow for the creation of

* Add entity groups

* Convert entity list to dict

* Add entity to entity_group_disagg atfter updating entity gorups

* Change 'group' parameter to 'grouped_entities'

* Add unit tests for grouped NER pipeline case

* Correct variable name typo for NER_FINETUNED_MODELS

* Sync grouped tests to recent test updates

* Added generic ONNX conversion script for PyTorch model.

* WIP initial TF support.

* TensorFlow/Keras ONNX export working.

* Print framework version info

* Add possibility to check the model is correctly loading on ONNX runtime.

* Remove quantization option.

* Specify ONNX opset version when exporting.

* Formatting.

* Remove unused imports.

* Make functions more generally reusable from other part of the code.

* isort happy.

* flake happy

* Export only feature-extraction for now

* Correctly check inputs order / filter before export.

* Removed task variable

* Fix invalid args call in load_graph_from_args.

* Fix invalid args call in convert.

* Fix invalid args call in infer_shapes.

* Raise exception and catch in caller function instead of exit.

* Add 04-onnx-export.ipynb notebook

* More WIP on the notebook

* Remove unused imports

* Simplify & remove unused constants.

* Export with constant_folding in PyTorch

* Let's try to put function args in the right order this time ...

* Disable external_data_format temporary

* ONNX notebook draft ready.

* Updated notebooks charts + wording

* Correct error while exporting last chart in notebook.

* Adressing @LysandreJik comment.

* Set ONNX opset to 11 as default value.

* Set opset param mandatory

* Added ONNX export unittests

* Quality.

* flake8 happy

* Add keras2onnx dependency on extras["tf"]

* Pin keras2onnx on github master to v1.6.5

* Second attempt.

* Third attempt.

* Use the right repo URL this time ...

* Do the same for onnxconverter-common

* Added keras2onnx and onnxconveter-common to 1.7.0 to supports TF2.2

* Correct commit hash.

* Addressing PR review: Optimization are enabled by default.

* Addressing PR review: small changes in the notebook

* setup.py comment about keras2onnx versioning.

covers torch and tf. Also fixes a failing @slow test

* Fix: unpin flake8 and fix cs errors

* Ok we still need to quote those

[Marian Fixes] prevent predicting pad_token_id before softmax, support language codes, name multilingual models (#4290)

* Improvements to the wandb integration

* small reorg + no global necessary

* feat(trainer): log epoch and final metrics

* Simplify logging a bit

* Fixup

* Fix crash when just running eval
Co-authored-by: Chris Van Pelt <vanpelt@gmail.com>
Co-authored-by: Boris Dayma <boris.dayma@gmail.com>

- MarianSentencepieceTokenizer - > MarianTokenizer
- Start using unk token.
- add docs page
- add better generation params to MarianConfig
- more conversion utilities

* fix PR

* move tests to correct place

* Add AlbertForPreTraining and TFAlbertForPreTraining models.

* PyTorch conversion

* TensorFlow conversion

* style
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>

* Created using Colaboratory

* [examples] reorganize files

* remove run_tpu_glue.py as superseded by TPU support in Trainer

* Bugfix: int, not tuple

* move files around

* Rewritten batch support in pipelines.
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Fix imports sorting 🔧

Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Set pad_to_max_length=True by default on Pipeline.

* Set pad_to_max_length=False for generation pipelines.

Most of generation models doesn't have padding token.

* Address @joeddav review comment: Uniformized *args.
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Address @joeddav review comment: Uniformized *args (second).
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* first copy & past commit from Bert and morgans LSH code

* add easy way to compare to trax original code

* translate most of function

* make trax lsh self attention deterministic with numpy seed + copy paste code

* add same config

* add same config

* make layer init work

* implemented hash_vectors function for lsh attention

* continue reformer translation

* hf LSHSelfAttentionLayer gives same output as trax layer

* refactor code

* refactor code

* refactor code

* refactor

* refactor + add reformer config

* delete bogus file

* split reformer attention layer into two layers

* save intermediate step

* save intermediate step

* make test work

* add complete reformer block layer

* finish reformer layer

* implement causal and self mask

* clean reformer test and refactor code

* fix merge conflicts

* fix merge conflicts

* update init

* fix device for GPU

* fix chunk length init for tests

* include morgans optimization

* improve memory a bit

* improve comment

* factorize num_buckets

* better testing parameters

* make whole model work

* make lm model work

* add t5 copy paste tokenizer

* add chunking feed forward

* clean config

* add improved assert statements

* make tokenizer work

* improve test

* correct typo

* extend config

* add complexer test

* add new axial position embeddings

* add local block attention layer

* clean tests

* refactor

* better testing

* save intermediate progress

* clean test file

* make shorter input length work for model

* allow variable input length

* refactor

* make forward pass for pretrained model work

* add generation possibility

* finish dropout and init

* make style

* refactor

* add first version of RevNet Layers

* make forward pass work and add convert file

* make uploaded model forward pass work

* make uploaded model forward pass work

* refactor code

* add namedtuples and cache buckets

* correct head masks

* refactor

* made reformer more flexible

* make style

* remove set max length

* add attention masks

* fix up tests

* fix lsh attention mask

* make random seed optional for the moment

* improve memory in reformer

* add tests

* make style

* make sure masks work correctly

* detach gradients

* save intermediate

* correct backprob through gather

* make style

* change back num hashes

* rename to labels

* fix rotation shape

* fix detach

* update

* fix trainer

* fix backward dropout

* make reformer more flexible

* fix conflict

* fix

* fix

* add tests for fixed seed in reformer layer

* fix trainer typo

* fix typo in activations

* add fp16 tests

* add fp16 training

* support fp16

* correct gradient bug in reformer

* add fast gelu

* re-add dropout for embedding dropout

* better naming

* better naming

* renaming

* finalize test branch

* finalize tests

* add more tests

* finish tests

* fix

* fix type trainer

* fix fp16 tests

* fix tests

* fix tests

* fix tests

* fix issue with dropout

* fix dropout seeds

* correct random seed on gpu

* finalize random seed for dropout

* finalize random seed for dropout

* remove duplicate line

* correct half precision bug

* make style

* refactor

* refactor

* docstring

* remove sinusoidal position encodings for reformer

* move chunking to modeling_utils

* make style

* clean config

* make style

* fix tests

* fix auto tests

* pretrained models

* fix docstring

* update conversion file

* Update pretrained_models.rst

* fix rst

* fix rst

* update copyright

* fix test path

* fix test path

* fix small issue in test

* include reformer in generation tests

* add docs for axial position encoding

* finish docs

* Update convert_reformer_trax_checkpoint_to_pytorch.py

* remove isort

* include sams comments

* remove wrong comment in utils

* correct typos

* fix typo

* Update reformer.rst

* applied morgans optimization

* make style

* make gpu compatible

* remove bogus file

* big test refactor

* add example for chunking

* fix typo

* add to README

* First commit to add a TF version of the trainer.

* Make the TF trainer closer to what looks the PT trainer

* Refactoring common code between the PT and TF trainer into an util file.

* Some bugfix + better similarity with the PT trainer

* Add missing class in transformers init

* Bugfix over prediction + use classification report instead of simple metrics

* Fix name error

* Fix optimization tests + style

* Apply style

* Several bugfix for multi-gpu training

* Apply style

* Apply style

* Add glue example for the TF trainer

* Several bugix + address the reviews

* Fix on the TF training args file

* Add a debug mode

* Bugfix in utils_ner.py when segment_ids is None

* Apply style

* Apply style

* Add TPU strategy

* Fix selection strategy

* Standard deviation can no longer be set to 0

* Remove torch pinned version

* 9th instead of 10th, silly me

* Hoist bert model tester for patric

* indent

* make tests work

* Update tests/test_modeling_bert.py
Co-authored-by: Julien Chaumond <chaumond@gmail.com>
Co-authored-by: sshleifer <sshleifer@gmail.com>
Co-authored-by: Julien Chaumond <chaumond@gmail.com>

There's an inconsistency right now where:
- we load some models into CACHE_DIR
- and some models in the default cache
- and often, in both for the same models

When running the RUN_SLOW tests, this takes a lot of disk space, time, and bandwidth.

I'd rather always use the default cache

Co-Authored-By: Stefan Schweter <stefan@schweter.it>