GitLab

Removed unused max_position_embeddings in config classes
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Fix typo

* Update README.md

* Update run_mlm_flax_stream.py

* Update README.md

fix typo output not ouput

typo was causing an error on pytest collection

* Set up overall resources documentation structure

* Update vit.mdx

* Removing irrelevant sections on text models

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* Update vit.mdx

* [clip] fix error message

* sync

Use image_transforms pad

* fix nit

- initialization on `_init_weights`
- fix copies

* add copied from

lazy import torch._softmax_backward_data
Signed-off-by: daquexian <daquexian566@gmail.com>
Signed-off-by: daquexian <daquexian566@gmail.com>

* Copy RoBERTa

* formatting

* implement RoBERTa with prelayer normalization

* update test expectations

* add documentation

* add convertion script for DinkyTrain weights

* update checkpoint repo

Unfortunately the original checkpoints assumes a hacked roberta model

* add to RoBERTa-PreLayerNorm docs to toc

* run utils/check_copies.py

* lint files

* remove unused import

* fix check_repo reporting wrongly a test is missing

* fix import error, caused by rebase

* run make fix-copies

* add RobertaPreLayerNormConfig to ROBERTA_EMBEDDING_ADJUSMENT_CONFIGS

* Fix documentation <Facebook> -> Facebook
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* fixup: Fix documentation <Facebook> -> Facebook
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Add missing Flax header
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* expected_slice -> EXPECTED_SLICE
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* update copies after rebase

* add missing copied from statements

* make fix-copies

* make prelayernorm explicit in code

* fix checkpoint path for the original implementation

* add flax integration tests

* improve docs

* update utils/documentation_tests.txt

* lint files

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

* make fix-copies

* Remove EXPECTED_SLICE calculation comments
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Install sentencepiece in DS CI image

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

* First draft

* Add more improvements

* Improve forward pass

* Fix layernorm

* Add upscaler

* More improvements

* More improvements

* More improvements

* Improve conversion script

* Add preprocessing

* Make output match original implementation

* Add additional attributes

* Add support for more models

* Support more models

* Add support for real world sr

* Add initial Swin2SRFeatureExtractor

* Add ImageSuperResolutionOutput

* Make more tests pass

* Use BaseModelOutput

* Fix one more test

* Fix more tests

* Fix another test

* Fix all tests

* Rename to Swin2SRImageProcessor

* Fix toctree

* Fix toctree

* Fix rebase

* Improve Swin2SRImageProcessor

* Remove feature extractor file

* Improve model

* Improve conversion script

* Fix integration test

* Fix init

* Fix conversion script

* Address comments

* Improve upsampler

* Add NearestConvUpsampler

* Improve pixel shuffle upsampler

* Improve auxiliary upsampler

* Improve conversion script

* Rename conv_last to final_convolution

* Fix rebase

* Improve upsample module

* Add padding to image processor

* Fix bug

* Update padding

* Remove print statement and fix integration test

* Improve docs

* Add image processor tests

* Convert all checkpoints, fix testsé

* Remove print statements

* Fix import
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>

* Add mapping

* Add mapping to pipeline

* Apply suggestions

* Fix feature extractor tests

* Use ForInstance, add model to universal mapping

* More fixes

* Remove model from deprecated objectsé
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>

* Revert "Fixing object detection with `layoutlm` (#20776)"

This reverts commit fca66abe.

* Better fix for layoutlm object detection.

* Style.

skip feature extraction test if in `IMAGE_PROCESSOR_MAPPING`

Recompile apex in DeepSpeed CI image
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Move convert_to_rgb to image_transforms module

* Fix tests

* generate from config mvp

* fix failing tests

* max_time test

* Load default gen config at model load time; Update docs

* further documentation; add tests

* adapt rag to the new structure

* handle models not instantiated with from_pretained (like in tests)

* better default generation config

* add can_generate fn

* handle legacy use case of ad hoc model config changes

* initialize gen config from config in individual methods, if gen config is none

* fix _get_decoder_start_token_id when called outside GenerationMixin

* correct model config load order (set attr > model config > decoder config)

* update rag to match latest changes

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

* load gen config from model config in model.from_pretrained

* fix can_generate fn

* handle generate calls without a previous from_pretrained (e.g. tests)

* add legacy behavior (and a warning)

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

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

* Fixing object detection with layoutlm.

* Fixup.

fix failing `pipeline` test

* Workaround for #20287: FlanT5-XXL 8bit support

* Make fix-copies

* revert unrelated change

* Dont apply to longt5 and switch transformers

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

* Even more validation.

* Fixing order.

* Add Swin backbone

* Remove line

* Add code example
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>

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

* Replaces xxx_required with requires_backends

* Fixup

* weight -> weights

* model embedding resize does not work with both v2 and noraml

* remove useless test

fix: 修复Trainer无法使用use_legacy_prediction_loop参数的问题

解决使用use_legacy_prediction_loop参数在predict阶段使用prediction_loop进行预测时，遇到AttributeError: 'PredictionOutput' object has no attribute 'num_samples'的问题
Co-authored-by: ZhouHang <zhouhang@idataway.com>

* Improve tests

* Improve TF tests

* Apply suggestion

* Fix test
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>

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

Uninstall torch_tensorrt for now
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

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

* added model resources for xlm-roberta

* added model resources for xlm-roberta

* resolve suggested changes

* add resources to xlm-roberta

* Add first draft

* Add out_features attribute to config

* Add corresponding test

* Add Dinat backbone

* Add BackboneMixin

* Add Backbone mixin, improve tests

* Fix embeddings

* Fix bug

* Improve backbones

* Fix Nat backbone tests

* Fix Dinat backbone tests

* Apply suggestions
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>

`image = to_channel_dimension_format(image, ChannelDimension.LAST)`
is redundant as this same conversion is also applied in to_pil_image().

This redundant call actually makes the training fail in rare cases.
The problem can be reproduced with the following code snippet:
```
from transformers.models.clip import CLIPFeatureExtractor
vision_processor = CLIPFeatureExtractor.from_pretrained('openai/clip-vit-large-patch14')
images = [
    torch.rand(size=(3, 2, 10), dtype=torch.float),
    torch.rand(size=(3, 10, 1), dtype=torch.float),
    torch.rand(size=(3, 1, 10), dtype=torch.float)
]
for image in images:
    processed_image = vision_processor(images=image, return_tensors="pt")['pixel_values']
    print(processed_image.shape)
    assert processed_image.shape == torch.Size([1, 3, 224, 224])
```

The last image has a height of 1 pixel.
The second call to to_channel_dimesion_format() will transpose the image, and the height
dimension is wrongly treated as the channels dimension afterwards.
Because of this, the following normalize() step will result in an
exception.

* Fix AdamWeightDecay for TF

* Fix AdamWeightDecay for TF

* make fixup

* Fix the pipeline test regarding TF

* Fix the pipeline test regarding TF

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

* add `keep_in_fp32_modules` support

* pass it as class attribute

* few modifs

- make tests `slow`
- fix logic

* better logic

* fix failing test

* `bfloat16` support

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

* fix

* simplify tests

* simplify tests

* fix test

* modify message

* more checks

* fix failing tests

* add more conditions

- add `is_accelerate_available`
- fixes pipleine tests that failed

* add suggestions

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

* fix failing `bnb` test

* add last safety checker
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>