GitLab

Slightly breaking change, changes functionality for `use_cache` in XLNet: if use_cache is True and mem_len is 0 or None (which is the case in the base model config), the model behaves like GPT-2 and returns mems to be used as past in generation. At training time `use_cache` is overriden and always True.

This reverts commit 0b2da0e5.

Slightly breaking change, changes functionality for `use_cache` in XLNet: if use_cache is True and mem_len is 0 or None (which is the case in the base model config), the model behaves like GPT-2 and returns mems to be used as past in generation. At training time `use_cache` is overriden and always True.

* [WIP] Proposal for model outputs

* All Bert models

* Make CI green maybe?

* Fix ONNX test

* Isolate ModelOutput from pt and tf

* Formatting

* Add Electra models

* Auto-generate docstrings from outputs

* Add TF outputs

* Add some BERT models

* Revert TF side

* Remove last traces of TF changes

* Fail with a clear error message

* Add Albert and work through Bart

* Add CTRL and DistilBert

* Formatting

* Progress on Bart

* Renames and finish Bart

* Formatting

* Fix last test

* Add DPR

* Finish Electra and add FlauBERT

* Add GPT2

* Add Longformer

* Add MMBT

* Add MobileBert

* Add GPT

* Formatting

* Add Reformer

* Add Roberta

* Add T5

* Add Transformer XL

* Fix test

* Add XLM + fix XLMForTokenClassification

* Style + XLMRoberta

* Add XLNet

* Formatting

* Add doc of return_tuple arg

* Update PretrainedConfig doc

* Formatting

* Small fixes

* Forgotten args and more cleanup

* all save_pretrained methods mkdir if not os.path.exists

* fix doc

* add format file

* add output attentions to all docs

* add also for bart

* fix naming

* re-add doc to config

* DOC: Replace instances of ``config.output_attentions`` with function argument ``output_attentions``

* DOC: Apply Black Formatting

* Fix errors where output_attentions was undefined

* Remove output_attentions in classes per review

* Fix regressions on tests having `output_attention`

* Fix further regressions in tests relating to `output_attentions`

Ensure proper propagation of `output_attentions` as a function parameter
to all model subclasses

* Fix more regressions in `test_output_attentions`

* Fix issues with BertEncoder

* Rename related variables to `output_attentions`

* fix pytorch tests

* fix bert and gpt2 tf

* Fix most TF tests for `test_output_attentions`

* Fix linter errors and more TF tests

* fix conflicts

* DOC: Apply Black Formatting

* Fix errors where output_attentions was undefined

* Remove output_attentions in classes per review

* Fix regressions on tests having `output_attention`

* fix conflicts

* fix conflicts

* fix conflicts

* fix conflicts

* fix pytorch tests

* fix conflicts

* fix conflicts

* Fix linter errors and more TF tests

* fix tf tests

* make style

* fix isort

* improve output_attentions

* improve tensorflow
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Kill model archive maps

* Fixup

* Also kill model_archive_map for MaskedBertPreTrainedModel

* Unhook config_archive_map

* Tokenizers: align with model id changes

* make style && make quality

* Fix CI

Fixes bug reported in https://github.com/huggingface/transformers/issues/4669

See #3967 for context

* [file_utils] use_cdn + documentation

* Move to cdn. urls for weights

* [urls] Hotfix for bert-base-japanese

cc @sshleifer

* doc

* [tests] Add sample files for a regression task

* [HUGE] Trainer

* Feedback from @sshleifer

* Feedback from @thomwolf + logging tweak

* [file_utils] when downloading concurrently, get_from_cache will use the cached file for subsequent processes

* [glue] Use default max_seq_length of 128 like before

* [glue] move DataTrainingArguments around

* [ner] Change interface of InputExample, and align run_{tf,pl}

* Re-align the pl scripts a little bit

* ner

* [ner] Add integration test

* Fix language_modeling with API tweak

* [ci] Tweak loss target

* Don't break console output

* amp.initialize: model must be on right device before

* [multiple-choice] update for Trainer

* Re-align to 827d6d6e

* better config serialization

* finish configuration utils

* remove output_past from pt

* make style

* add optional input length for gpt2

* add use cache to prepare input

* save memory in gpt2

* correct gpt2 test inputs

* make past input optional for gpt2

* finish use_cache for all models

* make style

* delete modeling_gpt2 change in test file

* correct docstring

* correct is true statements for gpt2

* Initial commit to get BERT + run_glue.py on TPU

* Add README section for TPU and address comments.

* Cleanup TPU bits from run_glue.py (#3)

TPU runner is currently implemented in:
https://github.com/pytorch-tpu/transformers/blob/tpu/examples/run_glue_tpu.py.

We plan to upstream this directly into `huggingface/transformers`
(either `master` or `tpu`) branch once it's been more thoroughly tested.

* Cleanup TPU bits from run_glue.py

TPU runner is currently implemented in:
https://github.com/pytorch-tpu/transformers/blob/tpu/examples/run_glue_tpu.py

.

We plan to upstream this directly into `huggingface/transformers`
(either `master` or `tpu`) branch once it's been more thoroughly tested.

* No need to call `xm.mark_step()` explicitly (#4)

Since for gradient accumulation we're accumulating on batches from
`ParallelLoader` instance which on next() marks the step itself.

* Resolve R/W conflicts from multiprocessing (#5)

* Add XLNet in list of models for `run_glue_tpu.py` (#6)

* Add RoBERTa to list of models in TPU GLUE (#7)

* Add RoBERTa and DistilBert to list of models in TPU GLUE (#8)

* Use barriers to reduce duplicate work/resources (#9)

* Shard eval dataset and aggregate eval metrics (#10)

* Shard eval dataset and aggregate eval metrics

Also, instead of calling `eval_loss.item()` every time do summation with
tensors on device.

* Change defaultdict to float

* Reduce the pred, label tensors instead of metrics

As brought up during review some metrics like f1 cannot be aggregated
via averaging. GLUE task metrics depends largely on the dataset, so
instead we sync the prediction and label tensors so that the metrics can
be computed accurately on those instead.

* Only use tb_writer from master (#11)

* Apply huggingface black code formatting

* Style

* Remove `--do_lower_case` as example uses cased

* Add option to specify tensorboard logdir

This is needed for our testing framework which checks regressions
against key metrics writtern by the summary writer.

* Using configuration for `xla_device`

* Prefix TPU specific comments.

* num_cores clarification and namespace eval metrics

* Cache features file under `args.cache_dir`

Instead of under `args.data_dir`. This is needed as our test infra uses
data_dir with a read-only filesystem.

* Rename `run_glue_tpu` to `run_tpu_glue`
Co-authored-by: LysandreJik <lysandre.debut@reseau.eseo.fr>

* add bad words list

* make style

* add bad_words_tokens

* make style

* better naming

* make style

* fix typo

* solve conflicts

* move warnings below

* incorporate changes

* add pad_to_max_length to pipelines

* add bug fix for T5 beam search

* add prefix patterns

* make style

* fix conflicts

* adapt pipelines for task specific parameters

* improve docstring

* remove unused patterns

* add new default configs

* change prefix default to None

* make style

* fix conflicts

* better error message

* better error message

* update to model identifier instead of url

* update to model identifier instead of ur

* Pass kwargs to configuration

* Setter

* test

* Add disable_outgoing to pretrained items

Setting disable_outgoing=True disables outgonig traffic:
- etags are not looked up
- models are not downloaded

* parameter name change

* Remove forgotten print

* improving generation

* finalized special token behaviour for no_beam_search generation

* solved modeling_utils merge conflict

* solve merge conflicts in modeling_utils.py

* add run_generation improvements from PR #2749

* adapted language generation to not use hardcoded -1 if no padding token is available

* remove the -1 removal as hard coded -1`s are not necessary anymore

* add lightweight language generation testing for randomely initialized models - just checking whether no errors are thrown

* add slow language generation tests for pretrained models using hardcoded output with pytorch seed

* delete ipdb

* check that all generated tokens are valid

* renaming

* renaming Generation -> Generate

* make style

* updated so that generate_beam_search has same token behavior than generate_no_beam_search

* consistent return format for run_generation.py

* deleted pretrain lm generate tests -> will be added in another PR

* cleaning of unused if statements and renaming

* run_generate will always return an iterable

* make style

* consistent renaming

* improve naming, make sure generate function always returns the same tensor, add docstring

* add slow tests for all lmhead models

* make style and improve example comments modeling_utils

* better naming and refactoring in modeling_utils

* improving generation

* finalized special token behaviour for no_beam_search generation

* solved modeling_utils merge conflict

* solve merge conflicts in modeling_utils.py

* add run_generation improvements from PR #2749

* adapted language generation to not use hardcoded -1 if no padding token is available

* remove the -1 removal as hard coded -1`s are not necessary anymore

* add lightweight language generation testing for randomely initialized models - just checking whether no errors are thrown

* add slow language generation tests for pretrained models using hardcoded output with pytorch seed

* delete ipdb

* check that all generated tokens are valid

* renaming

* renaming Generation -> Generate

* make style

* updated so that generate_beam_search has same token behavior than generate_no_beam_search

* consistent return format for run_generation.py

* deleted pretrain lm generate tests -> will be added in another PR

* cleaning of unused if statements and renaming

* run_generate will always return an iterable

* make style

* consistent renaming

* improve naming, make sure generate function always returns the same tensor, add docstring

* add slow tests for all lmhead models

* make style and improve example comments modeling_utils

* better naming and refactoring in modeling_utils

* changed fast random lm generation testing design to more general one

* delete in old testing design in gpt2

* correct old variable name

* temporary fix for encoder_decoder lm generation tests - has to be updated when t5 is fixed

* adapted all fast random generate tests to new design

* better warning description in modeling_utils

* better comment

* better comment and error message
Co-authored-by: Thomas Wolf <thomwolf@users.noreply.github.com>