GitLab

* pad_to_multiple_of added to DataCollatorForWholeWordMask

* pad_to_multiple_of added to DataCollatorForWholeWordMask
Co-authored-by: Цвигун Аким Олегович <AOTsvigun@sberbank.ru>

* Return raw outputs in TextClassificationPipeline

* Style

* Support for problem type

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

* Apply Nicolas' comments
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Fix from_pretrained with corrupted state_dict

* Adapt test

* Use better checkpoint

* Style

* Clean up

This change enables tf.keras.mixed_precision with bf16

* fix #12970

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

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

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

* remove unnecessary issue link

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

* fix typo in example/text-classification README

* add space to align the table

* Add multilingual documentation support

* Add multilingual documentation support

* make style

* make style

* revert

Co-authored-by: Lukas Wutschitz <lukas.wutschitz@microsoft.com>

* minor change to log azureml only for rank 0

* fix typo

help for `ModelArguments.gradient_checkpointing` should be
"If True, use gradient checkpointing to save memory
at the expense of slower backward pass."
not "Whether to freeze the feature extractor layers of the model."
(which is duplicated from `freeze_feature_extractor` arg)

* Add CpmTokenizerFast

* Fix isort

* Overwrite _batch_encode_plus

* Update feature extraction pipelilne.

* Leaving 1 small model for actual values check.

* Fixes tests

- Better support for tokenizer with no pad token
- Increasing PegasusModelTesterConfig for pipelines
- Test of feature extraction are more permissive + don't test Multimodel
models + encoder-decoder.

* Fixing model loading with incorrect shape (+ model with HEAD).

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

* Revert modeling_utils modification.

* Some corrections.

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

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

* Syntax.

* Fixing text-classification tests.

* Don't modify this file.
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Raise an issue if the pytorch version is < 1.8.0

* Attempt to add a test to ensure it correctly raises.

* Missing docstring.

* Second attempt, patch with string absolute import.

* Let's do the call before checking it was called ...

* use the correct function ... 🤦

* Raise ImportError and AssertionError respectively when unable to find torch and torch version is not sufficient.

* Correct path mock patching

* relax constraint for torch_onnx_dict_inputs to ge instead of eq.

* Style.

* Split each version requirements for torch.

* Let's compare version directly.

* Import torch_version after checking pytorch is installed.

* @require_torch

Change `PreTrainedConfig` -> `PretrainedConfig` in the docstring for `AutoTokenizer.from_pretrained(...)`.

Fix `config.decoder.__class` -> `config.decoder.__class__`

Change `torch.Tensor` -> `torch.FloatTensor` in `TemperatureLogitsWarper` to be consistent with the `LogitsWarper` ABC signature annotation.

While `Iterable[Iterable[int]]` is a nicer annotation (it's covariant!), the defensive statements parsing out `bad_words_ids` in `__init__(...)` force the caller to pass in `List[List[int]]`. I've changed the annotation to make that clear.

* fix distiller

* fix style

`_BaseAutoModelClass` was missing `classmethod` decorators on the `from_config(...)` and `from_pretrained(...)` methods.

Change `score` -> `scores` because the argument is not positional-only, so you need consistently named parameters for the subclasses. The subclasses appear to favor `scores` over `score`.

* Fixed train_test_split test_size argument

* `Seq2SeqTrainer` set max_length and num_beams only when non None  (#12899)

* set max_length and num_beams only when non None

* fix instance variables

* fix code style

* [FLAX] Minor fixes in CLM example (#12914)

* readme: fix retrieval of vocab size for flax clm example

* examples: fix flax clm example when using training/evaluation files

* Fix module path for symbolic_trace example
Co-authored-by: cchen-dialpad <47165889+cchen-dialpad@users.noreply.github.com>
Co-authored-by: Stefan Schweter <stefan@schweter.it>
Co-authored-by: Sylvain Gugger <sylvain.gugger@gmail.com>

* readme: fix retrieval of vocab size for flax clm example

* examples: fix flax clm example when using training/evaluation files

* set max_length and num_beams only when non None

* fix instance variables

* fix code style

* Better heuristic for token-classification pipeline.

Relooking at the problem makes thing actually much simpler,
when we look at ids from a tokenizer, we have no way in **general**
to recover if some substring is part of a word or not.

However, within the pipeline, with offsets we still have access to the
original string, so we can simply look if previous character (if it
exists) of a token, is actually a space. This will obviously be wrong
for tokenizers that contain spaces within tokens, tokenizers where
offsets include spaces too (Don't think there are a lot).

This heuristic hopefully is fully bc and still can handle non-word based
tokenizers.

* Updating test with real values.

* We still need the older "correct" heuristic to prevent fusing
punctuation.

* Adding a real warning when important.

* Add new multiple-choice example, remove old one