GitLab

[HUGE] Refactoring tokenizers backend - padding - truncation - pre-tokenized pipeline - fast tokenizers - tests (#4510)

* Use tokenizers pre-tokenized pipeline

* failing pretrokenized test

* Fix is_pretokenized in python

* add pretokenized tests

* style and quality

* better tests for batched pretokenized inputs

* tokenizers clean up - new padding_strategy - split the files

* [HUGE] refactoring tokenizers - padding - truncation - tests

* style and quality

* bump up requied tokenizers version to 0.8.0-rc1

* switched padding/truncation API - simpler better backward compat

* updating tests for custom tokenizers

* style and quality - tests on pad

* fix QA pipeline

* fix backward compatibility for max_length only

* style and quality

* Various cleans up - add verbose

* fix tests

* update docstrings

* Fix tests

* Docs reformatted

* __call__ method documented
Co-authored-by: Thomas Wolf <thomwolf@users.noreply.github.com>
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>

* add tpu and torchscipt for benchmark

* fix name in tests

* "fix email"

* make style

* better log message for tpu

* add more print and info for tpu

* allow possibility to print tpu metrics

* correct cpu usage

* fix test for non-install

* remove bugus file

* include psutil in testing

* run a couple of times before tracing in torchscript

* do not allow tpu memory tracing for now

* make style

* add torchscript to env

* better name for torch tpu
Co-authored-by: Patrick von Platen <patrick@huggingface.co>

* make transformers-cli cross-platform

Using "scripts" is a useful option in setup.py particularly when you want to get access to non-python scripts. However, in this case we want to have an entry point into some of our own Python scripts. To do this in a concise, cross-platfom way, we can use entry_points.console_scripts. This change is necessary to provide the CLI on different platforms, which "scripts" does not ensure. Usage remains the same, but the "transformers-cli" script has to be moved (be part of the library) and renamed (underscore + extension)

* make style & quality

As discussed w/ @lysandrejik

packaging is maintained by PyPA (the Python Packaging Authority), and should be lightweight and stable

* 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.

* Fix: unpin flake8 and fix cs errors

* Ok we still need to quote those

* Allow BatchEncoding to be initialized empty.

This is required by recent changes introduced in TF 2.2.

* Attempt to unpin Tensorflow to 2.2 with the previous commit.

* pin TF to 2.1

* Pin flake8 as well

* Fix the issue to properly run the accumulator with TF 2.2

* Apply style

* Fix training_args_tf for TF 2.2

* Fix the TF training args when only one GPU is available

* Remove the fixed version of TF in setup.py

* Standard deviation can no longer be set to 0

* Remove torch pinned version

* 9th instead of 10th, silly me

* First pass on utility classes and python tokenizers

* finishing cleanup pass

* style and quality

* Fix tests

* Updating following @mfuntowicz comment

* style and quality

* Fix Roberta

* fix batch_size/seq_length inBatchEncoding

* add alignement methods + tests

* Fix OpenAI and Transfo-XL tokenizers

* adding trim_offsets=True default for GPT2 et RoBERTa

* style and quality

* fix tests

* add_prefix_space in roberta

* bump up tokenizers to rc7

* style

* unfortunately tensorfow does like these - removing shape/seq_len for now

* Update src/transformers/tokenization_utils.py
Co-Authored-By: Stefan Schweter <stefan@schweter.it>

* Adding doc and docstrings

* making flake8 happy
Co-authored-by: Stefan Schweter <stefan@schweter.it>

* Renamed num_added_tokens to num_special_tokens_to_add
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Cherry-Pick: Partially fix space only input without special tokens added to the output #3091
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Added property is_fast on PretrainedTokenizer and PretrainedTokenizerFast
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Make fast tokenizers unittests work on Windows.

* Entirely refactored unittest for tokenizers fast.

* Remove ABC class for CommonFastTokenizerTest

* Added embeded_special_tokens tests from allenai @dirkgr

* Make embeded_special_tokens tests from allenai more generic

* Uniformize vocab_size as a property for both Fast and normal tokenizers

* Move special tokens handling out of PretrainedTokenizer (SpecialTokensMixin)

* Ensure providing None input raise the same ValueError than Python tokenizer + tests.

* Fix invalid input for assert_padding when testing batch_encode_plus

* Move add_special_tokens from constructor to tokenize/encode/[batch_]encode_plus methods parameter.

* Ensure tokenize() correctly forward add_special_tokens to rust.

* Adding None checking on top on encode / encode_batch for TransfoXLTokenizerFast.
Avoid stripping on None values.

* unittests ensure tokenize() also throws a ValueError if provided None

* Added add_special_tokens unittest for all supported models.

* Style

* Make sure TransfoXL test run only if PyTorch is provided.

* Split up tokenizers tests for each model type.

* Fix invalid unittest with new tokenizers API.

* Filter out Roberta openai detector models from unittests.

* Introduce BatchEncoding on fast tokenizers path.

This new structure exposes all the mappings retrieved from Rust.
It also keeps the current behavior with model forward.

* Introduce BatchEncoding on slow tokenizers path.

Backward compatibility.

* Improve error message on BatchEncoding for slow path

* Make add_prefix_space True by default on Roberta fast to match Python in majority of cases.

* Style and format.

* Added typing on all methods for PretrainedTokenizerFast

* Style and format

* Added path for feeding pretokenized (List[str]) input to PretrainedTokenizerFast.

* Style and format

* encode_plus now supports pretokenized inputs.

* Remove user warning about add_special_tokens when working on pretokenized inputs.

* Always go through the post processor.

* Added support for pretokenized input pairs on encode_plus

* Added is_pretokenized flag on encode_plus for clarity and improved error message on input TypeError.

* Added pretokenized inputs support on batch_encode_plus

* Update BatchEncoding methods name to match Encoding.

* Bump setup.py tokenizers dependency to 0.7.0rc1

* Remove unused parameters in BertTokenizerFast

* Make sure Roberta returns token_type_ids for unittests.

* Added missing typings

* Update add_tokens prototype to match tokenizers side and allow AddedToken

* Bumping tokenizers to 0.7.0rc2

* Added documentation for BatchEncoding

* Added (unused) is_pretokenized parameter on PreTrainedTokenizer encode_plus/batch_encode_plus methods.

* Added higher-level typing for tokenize / encode_plus / batch_encode_plus.

* Fix unittests failing because add_special_tokens was defined as a constructor parameter on Rust Tokenizers.

* Fix text-classification pipeline using the wrong tokenizer

* Make pipelines works with BatchEncoding

* Turn off add_special_tokens on tokenize by default.
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Remove add_prefix_space from tokenize call in unittest.
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Style and quality
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Correct message for batch_encode_plus none input exception.
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Fix invalid list comprehension for offset_mapping overriding content every iteration.
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* TransfoXL uses Strip normalizer.
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Bump tokenizers dependency to 0.7.0rc3
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Support AddedTokens for special_tokens and use left stripping on mask for Roberta.
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* SpecilaTokenMixin can use slots to faster access to underlying attributes.
Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

* Remove update_special_tokens from fast tokenizers.

* Ensure TransfoXL unittests are run only when torch is available.

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

* Style

* Style 🙏🙏



* Remove slots on SpecialTokensMixin, need deep dive into pickle protocol.

* Remove Roberta warning on __init__.

* Move documentation to Google style.
Co-authored-by: LysandreJik <lysandre.debut@reseau.eseo.fr>

* [ci] Also run test_examples in py37

(will revert at the end of the experiment)

* InputExample: use immutable dataclass

* [deps] Install dataclasses for Py<3.7

* [skip ci] Revert "[ci] Also run test_examples in py37"

This reverts commit d29afd9959786b77759b0b8fa4e6b4335b952015.

cc @julien-c

hat/tip @bramvanroy

The CONTRIBUTING file pins to a specific version of isort, so we might as well install that in `dev` . This makes it easier for contributors so they don't have to manually install the specific commit.

* memory benchmark rss

* have both forward pass and line-by-line mem tracing

* cleaned up tracing

* refactored and cleaning up API

* no f-strings yet...

* add GPU mem logging

* fix GPU memory monitoring

* style and quality

* clean up and doc

* update with comments

* Switching to python 3.6+

* fix quality