GitLab

* Update modeling_wav2vec2.py

With very tiny sound files (less than 0.1 seconds) the num_masked_span can be too long. The issue is described in issue #15366 and discussed with @patrickvonplaten.

* correct errors with mask time indices

* remove bogus file

* make fix-copies
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Add 'with torch.no_grad()' to BEiT integration test forward pass

* Fix inconsistent use of tabs and spaces in indentation

* Fix spurious warning in TF TokenClassification models

* Fixing one last spurious warning

* Removing outdated warning altogether

* refactor roberta tokenizer

* refactor fast tokenizer

* remove old comment

* fix tf led

* fix

* fix

* Add test_pt_tf_model_equivalence_extra for TFLED

* add a (temporary) test
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* add a section about GPUs

* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* [Trainer] suppress warning for length-related columns

* improve message

* Update src/transformers/trainer.py

* Change REALM checkpoint to new ones

* Last checkpoint missing

* Fix loss calculation in TFFunnelForTokenClassification

* revert the change in TFFunnelForTokenClassification

* fix FunnelForTokenClassification loss

* fix other TokenClassification loss

* fix more

* fix more

* add num_labels to ElectraForTokenClassification

* revert the change to research projects
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* [deepspeed doc] fix import, extra notes

* typo

* Fix Swin model outputs

* Rename pooler

fix typo

* Add Luke training

* Fix true label tags

* Fix true label tags

* Fix true label tags

* Update the data collator for Luke

* Some training refactor for Luke

* Improve data collator for Luke

* Fix import

* Fix datasets concatenation

* Add the --max_entity_length argument for Luke models

* Remove unused code

* Fix style issues

* Fix style issues

* Move the Luke training into a separate folder

* Fix style

* Fix naming

* Fix filtering

* Fix filtering

* Fix filter

* Update some preprocessing

* Move luke to research_projects

* Checkstyle

* Address comments

* Fix style

(This is an editorial change only)

* Fix the inconsistency of loss calculation between PT/TF XLNetLMHeadModel

* overwrite test_loss_computation
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* add xlm roberta xl

* add convert xlm xl fairseq checkpoint to pytorch

* fix init and documents for xlm-roberta-xl

* fix indention

* add test for XLM-R xl,xxl

* fix model hub name

* fix some stuff

* up

* correct init

* fix more

* fix as suggestions

* add torch_device

* fix default values of doc strings

* fix leftovers

* merge to master

* up

* correct hub names

* fix docs

* fix model

* up

* finalize

* last fix

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* add copied from

* make style
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* clean commit of changes

* apply review feedback, make edits

* fix backticks, minor formatting

* 🖍 make fixup and minor edits

* 🖍 fix # in header

* 📝 update code sample without from_pt

* 📝 final review

* add model sharing tutorial

* 🖍 apply feedback from review

* 📝 make edits

* 🖍 fix formatting

* 📝 convert from pt checkpoint to flax

* 📝 final review

* fix missing eps

* Same fix for ConvBertGeneratorPredictions

* Same fix for AlbertMLMHead
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* [deepspeed] saving checkpoint fallback when fp16 weights aren't saved

* Bump required deepspeed version to match usage when saving checkpoints

* update version
Co-authored-by: Mihai Balint <balint.mihai@gmail.com>

* add xglm

* update vocab size

* fix model name

* style and tokenizer

* typo

* no mask token

* fix pos embed compute

* fix args

* fix tokenizer

* fix positions

* fix tokenization

* style and dic fixes

* fix imports

* add fast tokenizer

* update names

* add pt tests

* fix tokenizer

* fix typo

* fix tokenizer import

* fix fast tokenizer

* fix tokenizer

* fix converter

* add tokenizer test

* update checkpoint names

* fix tokenizer tests

* fix slow tests

* add copied from comments

* rst -> mdx

* flax model

* update flax tests

* quality

* style

* doc

* update index and readme

* fix copies

* fix doc

* update toctrr

* fix indent

* minor fixes

* fix config doc

* don't save embed_pos weights

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* address Sylvains commnets, few doc fixes

* fix check_repo

* align order of arguments

* fix copies

* fix labels

* remove unnecessary mapping

* fix saving tokenizer
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Make links explicit

* Removing reference to compute_metrics() since it's kind of PyTorch-specific

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Prepare deprecated ONNX exporter for PyTorch v1.11

* Add deprecation warning

* Fix `bad_word_ids` not working with sentencepiece-based tokenizers

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

* Fixing support `batch_size` and `num_return_Sequences` in
`text-generation` pipeline

And `text2text-generation` too.

The bug was caused by the batch_size containing both the incoming batch
**and** the generated `num_sequences`.

The fix simply consists into splitting both of these again into
different dimensions.

* TF support.

* Odd backward compatibility script in the way.

* Use syncfree AdamW for xla:gpu device by default

* Make syncfree AdamW optional

* Add init to BORT

* BORT should be in init