GitLab

* add readme for flax clm

* use section link for tokenizer

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

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

* fix_torch_device_generate_test

* remove @

* upload

* Apply suggestions from code review

* Apply suggestions from code review

* Apply suggestions from code review

* Update examples/flax/language-modeling/README.md

* add more info

* finish

* fix
Co-authored-by: Patrick von Platen <patrick@huggingface.co>

* first draft

* max_seq_length => block_size

* fix arg names

* fix typos

* fix loss calculation

* add max examples, fix  train eval steps, metrics

* optimizer mask

* fix perpelexity, metric logging

* fix logging

* data_collator = > data_loader

* refactor loss_fn

* support single GPU

* pass distributed to write_metric

* fix jitting

* fix single device training

* fix single device metrics

* close inner progress bars once finished

* add overwrite_cache arg

* ifx dataset caching issue

* add more logs

* few small fixes,

* address nicholas suggestions

* fix docstr

* address patricks suggestions

* make flake happy

* pass new new_dropout_rng to apply_gradients

* reset train metrics after every epoc

* remove distributed logis, small fixes

* fix_torch_device_generate_test

* remove @

* finish refactor
Co-authored-by: Patrick von Platen <patrick@huggingface.co>

* Fix weight decay masking in `run_flax_glue.py`

Issues with the previous implementation:
- The `dict` from `traverse_util.flatten_dict` has keys which are tuples of strings, not one long string with the path separated by periods.
- `optax.masked` applies the transformation wherever the mask is True, so the masks are flipped.
- Flax's LayerNorm calls the scale parameter `scale` not `weight`

* Fix formatting with black

* adapt results
Co-authored-by: Patrick von Platen <patrick@huggingface.co>

* Remove redundant `nn.log_softmax` in `run_flax_glue.py`

`optax.softmax_cross_entropy` expects unnormalized logits, and so it already calls `nn.log_softmax`, so I believe it is not needed here. `nn.log_softmax` is idempotent so mathematically it shouldn't have made a difference.

* Remove unused 'flax.linen' import

* fix_torch_device_generate_test

* remove @

* change pytorch import to flax import

* fix_torch_device_generate_test

* remove @

* add flax glue link

* fix_torch_device_generate_test

* remove @

* correct best seed for flax fine-tuning
Co-authored-by: Patrick von Platen <patrick@huggingface.co>

* speed up flax glue

* remove unnecessary line

* remove folder

* remove run in loop
Co-authored-by: Patrick von Platen <patrick@huggingface.co>

* refactor

* update

* update

* update

* refactor run mlm

* finalize

* refactor more

* fix typo

* update

* finish refactor

* modify run mlm

* Apply suggestions from code review

* Apply suggestions from code review

* Apply suggestions from code review

* small fixes

* upload

* upload

* finish run mlm script
Co-authored-by: Patrick von Platen <patrick@huggingface.co>

* Add Flax Examples README

* Apply suggestions from code review

* Update examples/flax/README.md

* add nice table

* fix

* fix

* apply suggestions

* upload

* finish flax readme.md
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Add Cloud details to README

* Flax script and readme updates

* Some simplifications of Flax script

* Add Cloud details to README

* Flax script and readme updates

* Adds Flax BERT finetuning example

* fix traced jax tensor type

* Use Optax losses and learning schedulers

* Add 1GPU training results

* merge into master & make style

* fix input

* del file

* Fix bug in loss and add torch runs

* finish bert flax fine-tune

* Update examples/flax/text-classification/README.md

* Update examples/flax/text-classification/run_flax_glue.py

* add requirements

* finalize

* finalize
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Patrick von Platen <patrick@huggingface.co>

* add flax roberta

* make style

* correct initialiazation

* modify model to save weights

* fix copied from

* fix copied from

* correct some more code

* add more roberta models

* Apply suggestions from code review

* merge from master

* finish

* finish docs
Co-authored-by: Patrick von Platen <patrick@huggingface.co>

* Base move

* Examples reorganization

* Update references

* Put back test data

* Move conftest

* More fixes

* Move test data to test fixtures

* Update path

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

* Address review comments and clean
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>