GitLab

* Override save() to use the serving signature as the default

* Replace int32 with int64 in all our serving signatures

* Remember one very important line so as not to break every test at once

* Dtype fix for TFLED

* dtype fix for shift_tokens_right in general

* Dtype fixes in mBART and RAG

* Fix dtypes for test_unpack_inputs

* More dtype fixes

* Yet more mBART + RAG dtype fixes

* Yet more mBART + RAG dtype fixes

* Add a check that the model actually has a serving method

* Fix train_step and test_step, correctly enable CLIP fit test

* Stop using get_args on older Python versions

* Don't use get_origin either

* UnionType is actually even newer, don't use that either

* Apply the same fix to test_loss_computation

* Just realized I was accidentally skipping a bunch of tests!

* Fix test_loss_computation for models without separable labels

* Fix scalar losses in test_step and train_step

* Stop committing your breakpoints

* Fix Swin loss shape

* Fix Tapas loss shape

* Shape fixes for TAPAS, DeIT, HuBERT and ViTMAE

* Add loss computation to TFMobileBertForPreTraining

* make fixup and move copied from statement

* make fixup and move copied from statement

* Correct copied from

* Add labels and next_sentence_label inputs to TFMobileBERT

* Make sure total_loss is always defined

* Update tests/test_modeling_tf_common.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Fix copied from

* Ensure CTC models get labels in tests

* Ensure CTC models get labels in tests

* Fix tests for vit_mae

* Fix tests for vit_mae

* Fix tests for vit_mae

* Reduce batch size for wav2vec2 testing because it was causing OOM

* Skip some TAPAS tests that are failing

* Skip a failing HuBERT test

* make style

* Fix mobilebertforpretraining test

* Skip Wav2Vec2 tests that use huge amounts of mem

* Skip keras_fit for Wav2Vec2 as well
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* [WIP] Skeleton of VisualQuestionAnweringPipeline extended to support LayoutLM-like models

* Fixup

* Use the full encoding

* Basic refactoring to DocumentQuestionAnsweringPipeline

* Cleanup

* Improve args, docs, and implement preprocessing

* Integrate OCR

* Refactor question_answering pipeline

* Use refactored QA code in the document qa pipeline

* Fix tests

* Some small cleanups

* Use a string type annotation for Image.Image

* Update encoding with image features

* Wire through the basic docs

* Handle invalid response

* Handle empty word_boxes properly

* Docstring fix

* Integrate Donut model

* Fixup

* Incorporate comments

* Address comments

* Initial incorporation of tests

* Address Comments

* Change assert to ValueError

* Comments

* Wrap `score` in float to make it JSON serializable

* Incorporate AutoModeLForDocumentQuestionAnswering changes

* Fixup

* Rename postprocess function

* Fix auto import

* Applying comments

* Improve docs

* Remove extra assets and add copyright

* Address comments
Co-authored-by: Ankur Goyal <ankur@impira.com>

* Draft new cached_file

* Initial draft for config and model

* Small fixes

* Fix first batch of tests

* Look in cache when internet is down

* Fix last tests

* Bad black, not fixing all quality errors

* Make diff less

* Implement change for TF and Flax models

* Add tokenizer and feature extractor

* For compatibility with main

* Add utils to move the cache and auto-do it at first use.

* Quality

* Deal with empty commit shas

* Deal with empty etag

* Address review comments

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

* change shape to support dynamic batch input in tf.generate

* add tests
Co-authored-by: nlpcatcode <nlpcodecat@gmail.com>

* Fix failing test_xla_generate_slow tests

* Fix failing speech-to-text xla_generate tests

* Rewrite push_to_hub to use upload_files

* Adapt the doc a bit

* Address review comments and clean doc

* Add serving_output and serving methods to some vision models

* Add serving outputs for DeiT

* Don't convert hidden states - differing shapes

* Make saveable

* Fix up

* Make swin saveable

* Add in tests

* Fix funnel tests (can't convert to tensor)

* Fix numpy call

* Tidy up a bit

* Add in hidden states - resnet

* Remove numpy

* Fix failing tests - tensor shape and skipping tests

* Remove duplicated function

* PR comments - formatting and var names

* PR comments
Add suggestions made by Joao Gante:
* Use tf.shape instead of shape_list
* Use @tooslow decorator on tests
* Simplify some of the logic

* PR comments
Address Yih-Dar Sheih comments - making tensor names consistent and make types float

* Types consistent with docs; disable test on swin (slow)

* CI trigger

* Change input_features to float32

* Add serving_output for segformer

* Fixup
Co-authored-by: Amy Roberts <amyeroberts@users.noreply.github.com>

Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Return scalar losses instead of per-sample means

* Make loss shape (1,) instead of scalar

* Allow scalar losses in test_loss_computation

* Allow scalar losses in test_loss_computation

* Allow scalar losses in test_loss_computation

* Remove XLA loss function for RAG

* Copy inputs to train and test step before modifying them, as this breaks things

* Add XLA tests, fix our loss functions to be XLA-compatible

* make fixup

* Update loss computation test to expect vector of per-sample losses

* Patch loss for TFLED

* Patch loss for TFAlbert

* Add a tf_legacy_loss config flag that enables old loss functions

* Stop using config.get() because it's not a dict

* Skip loss computation test for RAG because its loss is very strange and I'm afraid to rewrite it

* make fixup

* Add XLA-compatible RAG loss

* Fix dtype of loss mask for TFAlbert

* Fix test for XLNet too because it overrides the default one

* make fixup

* Fix config test

* No more depending on GPU NaN behaviour

* Add test, avoid potential zero division

* Fix test item assignment

* Fix loss computation masking test

* make fixup

* Fix dtype bugs

* sharded conversion; add flag to control max hidden error

* better hidden name matching

* Add test: load TF from PT shards

* fix test (PT data must be local)

TF: XLA beam search + most generation-compatible models are now also XLA-generate-compatible (#17857)

* working beam search 🎉

* XLA generation compatible with ALL classes

* add xla generation slow test

* Fix tests that broke when models used batchnorm

* Initializing the model twice does not actually...
...give you the same weights each time.
I am good at machine learning.

* Fix speed regression

* initial commit

* update modeeling tf utils

* quality

* clean and update args

* update

* remove potential bug

* code quality

* update

* update max shard

* update tests for sharding from pretrained

* fix remaining test

* make style

* h5py if tf available

* update and fix test

* fix test

* style

* modified push to hub to support shard for TF

* quick fix

* update code

* merge branch main and style

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

* update based on reviews

* update doc

* update and style

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

* Update based on reviews

* fix typo

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

* Prepare CI for v0.8.0

* pin hfh (revert before merge)

* Revert "pin hfh (revert before merge)"

This reverts commit a0103140e1c77b810ffcb735192968bc03be3e1f.

* Test rc3

* Test latest rc

* Unpin to the RC
Co-authored-by: Sylvain Gugger <Sylvain.gugger@gmail.com>

* Also propagate changes to blenderbot, blenderbot_small, marian, mbart, and pegasus

* Use shape_list to safely get shapes

* Add relevant test

* Tidy and add metrics

* Resolve dynamic shaping issues and move test

* Tidy up and all samples in batch

* Formatting

* Add method to call to_tf_dataset() with column inference

* Add test for dataset creation

* Add a default arg for data collator

* Fix test

* Fix call with non-dev version of datasets

* Test correct column removal too

* make fixup

* More tests to make sure we remove unwanted columns

* Fix test to avoid predicting on unbuilt models

* Fix test to avoid predicting on unbuilt models

* Fix test to remove unwanted head mask columns from inputs

* Stop pushing your debug breakpoints to the main repo of the $2bn company you work for

* Skip the test in convnext because no grouped conv support

* Drop bools from the dataset dict

* Make style

* Skip the training test for models whose input dicts don't give us labels

* Skip transformerXL in the test because it doesn't return a simple loss

* Skip TFTapas because of some odd NaN losses

* make style

* make fixup

* Add docstring

* fixup

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

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

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

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

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

* Remove breakpoint from tests

* Fix assert, add requires_backends

* Protect tokenizer import with if TYPE_CHECKING

* make fixup

* Add noqa, more fixup

* More rearranging for ~* aesthetics *~

* Adding defaults for shuffle and batch_size to match to_tf_dataset()

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

* Initial commit

* Better label renaming

* Remove breakpoint before pushing (this is your job)

* Test a lot more in the Keras fit() test

* make fixup

* Clarify the case where we flatten y dicts into tensors

* Clarify the case where we flatten y dicts into tensors

* Extract label name remapping to a method

* Black preview

* Fixup too!

* Fix check copies

* Use the same version as the CI

* Bump black

* Add test to ensure models can take int64 inputs

* is_integer is an attribute, not a method

* Fix test when some inputs aren't tensors

* Add casts to blenderbot and blenderbot-small

* Add casts to the other failing models

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

* Update test_pt_tf_model_equivalence on PT side
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* add error message

* Use names in the error message

* allow ModelOutput

* rename to check_pt_tf_outputs and move outside

* fix style

* skip past_key_values in a better way

* Add comments

* improve code for label/loss

* make the logic clear by moving the ignore keys out

* fix _postprocessing_to_ignore

* fix _postprocessing_to_ignore: create new outputs from the remaining fields

* ignore past_key_values in TFGPT2 models for now

* make check_pt_tf_outputs better regarding names

* move check_pt_tf_models outside

* rename methods

* remove test_pt_tf_model_equivalence in TFCLIPModelTest

* Reduce TFViTMAEModelTest.test_pt_tf_model_equivalence

* move prepare_pt_inputs_from_tf_inputs outside check_pt_tf_models

* Fix quality

* Clean-up TFLxmertModelTester.test_pt_tf_model_equivalence

* Fix quality

* fix

* fix style

* Clean-up TFLEDModelTest.test_pt_tf_model_equivalence

* Fix quality

* add docstring

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

* refactor TF beam search

* refactored generate can now properly use attention masks

* add force bos/eos logit processors

* Adding new train_step logic to make things less confusing for users

* DO NOT ASK WHY WE NEED THAT SUBCLASS

* Metrics now working, at least for single-output models with type annotations!

* Updates and TODOs for the new train_step

* Make fixup

* Temporary test workaround until T5 has types

* Temporary test workaround until T5 has types

* I think this actually works! Needs a lot of tests though

* MAke style/quality

* Revert changes to T5 tests

* Deleting the aforementioned unmentionable subclass

* Deleting the aforementioned unmentionable subclass

* Adding a Keras API test

* Style fixes

* Removing unneeded TODO and comments

* Update test_step too

* Stop trying to compute metrics with the dummy_loss, patch up test

* Make style

* make fixup

* Docstring cleanup

* make fixup

* make fixup

* Stop expanding 1D input tensors when using dummy loss

* Adjust T5 test given the new compile()

* make fixup

* Skipping test for convnext

* Removing old T5-specific Keras test now that we have a common one

* make fixup

* make fixup

* Only skip convnext test on CPU

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

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

* Avoiding TF import issues

* make fixup

* Update compile() to support TF 2.3

* Skipping model.fit() on template classes for now

* Skipping model.fit() on template class tests for now

* Replace ad-hoc solution with find_labels

* make fixup
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Add unpack_inputs to remaining models

* removed kwargs to `call()` in TF models

* fix TF T5 tests

* use random_attention_mask for TF tests

* Fix for TFCLIP test (for now).
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Make Transformers use cache files when hf.co is down

* Fix tests

* Was there a random circleCI failure?

* Isolate patches

* Style

* Comment out the failure since it doesn't fail anymore

* Better comment

* revert tf gpt2

* add test for unpack_inputs and fix test case

* add changes to vision encoder decoder

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

* Skip test for TransfoXL

* Single list