GitLab

* Create README

* Update README.md
Co-authored-by: Kevin Canwen Xu <canwenxu@126.com>

* Default decoder inputs to encoder ones for T5 if neither are specified.

* Fixing typo, now all tests are passing.

* Changing einsum to operations supported by onnx

* Adding a test to ensure T5 can be exported to onnx op>9

* Modified test for onnx export to make it faster

* Styling changes.

* Styling changes.

* Changing notation for matrix multiplication
Co-authored-by: Abel Riboulot <tkai@protomail.com>

* fix attention mask

* fix slow test

* refactor attn masks

* fix fp16 generate test

* Added data collator for XLNet language modeling and related calls

Added DataCollatorForXLNetLanguageModeling in data/data_collator.py
to generate necessary inputs for language modeling training with
XLNetLMHeadModel. Also added related arguments, logic and calls in
examples/language-modeling/run_language_modeling.py.

Resolves: #4739, #2008 (partially)

* Changed name to `DataCollatorForPermutationLanguageModeling`

Changed the name of `DataCollatorForXLNetLanguageModeling` to the more general `DataCollatorForPermutationLanguageModelling`.
Removed the `--mlm` flag requirement for the new collator and defined a separate `--plm_probability` flag for its use.
CTRL uses a CLM loss just like GPT and GPT-2, so should work out of the box with this script (provided `past` is taken care of
similar to `mems` for XLNet).
Changed calls and imports appropriately.

* Added detailed comments, changed variable names

Added more detailed comments to `DataCollatorForPermutationLanguageModeling` in `data/data_collator.py` to explain working. Also cleaned up variable names and made them more informative.

* Added tests for new data collator

Added tests in `tests/test_trainer.py` for DataCollatorForPermutationLanguageModeling based on those in DataCollatorForLanguageModeling. A specific test has been added to check for odd-length sequences.

* Fixed styling issues

* BertTokenizerFast - Do not specify strip_accents by default

* Bump tokenizers to new version

* Add test for AddedToken serialization

* Fix #5507

* Fix formatting

* GPT2 tokenizer should not output token type IDs

* Same for OpenAIGPT

* fix warning

* style and quality

fixed ImportError: cannot import name 'hf_bucket_url'

* Exposing prepare_for_model for both slow & fast tokenizers

* Update method signature

* The traditional style commit

* Hide the warnings behind the verbose flag

* update default truncation strategy and prepare_for_model

* fix tests and prepare_for_models methods
Co-authored-by: Thomas Wolf <thomwolf@users.noreply.github.com>

Create model card for electicidad-small (Spanish Electra) fine-tuned on SQUAD-esv1

- fixed grammar and spelling
- added an intro
- updated Training data references

Create model card for electra-small-discriminator fine-tuned on SQUAD v2.0

* Make QA pipeline supports models with more than 2 outputs such as BART assuming start/end are the two first outputs.
Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* When using the new padding/truncation paradigm setting padding="max_length" + max_length=X actually pads the input up to max_length.

This result in every sample going through QA pipelines to be of size 384 whatever the actual input size is making the overall pipeline very slow.
Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Mask padding & question before applying softmax. Softmax has been refactored to operate in log space for speed and stability.
Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Format.
Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Use PaddingStrategy.LONGEST instead of DO_NOT_PAD
Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Revert "When using the new padding/truncation paradigm setting padding="max_length" + max_length=X actually pads the input up to max_length."

This reverts commit 1b00a9a2
Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com>

* Trigger CI after unattended failure

* Trigger CI

* Work on tokenizer summary

* Finish tutorial

* Link to it

* Apply suggestions from code review
Co-authored-by: Anthony MOI <xn1t0x@gmail.com>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Add vocab definition
Co-authored-by: Anthony MOI <xn1t0x@gmail.com>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

`ElectraDiscriminatorPredictions.forward` should not need `attention_mask`.

Create model card for electra-base-discriminator fine-tuned on SQUAD v1.1

* overriding _parse_and_tokenize in `TextGenerationPipeine` to allow for TransfoXl tokenizer arguments

* Changed expected_output_ids in TransfoXL generation test to match #4826 generation PR.

* making black happy

* making isort happy

Fixed duplicated memory use in Transformer-XL generation leading to bad predictions and performance.