removes decord

[vrnvu]

The only usage that was left of decord is removed in this PR

• Context:, project abandoned for 2 years: https://github.com/dmlc/decord
• Adds toil to local dev (Error while installing dev dependencies for Apple Silicon #22617)
  • https://github.com/huggingface/transformers/issues?q=is%3Aissue+decord
• Already removed from a couple of places in favor of av: https://github.com/huggingface/transformers/pulls?q=is%3Apr+decord

The helper functions read_video_pyav and sample_frame_indices are the used in all the documentation https://github.com/search?q=repo%3Ahuggingface%2Ftransformers%20read_video_pyav&type=code and video_classification https://github.com/huggingface/transformers/blob/main/src/transformers/pipelines/video_classification.py#L132

[ArthurZucker]

Very welcome for this! Thanks for the cleanup 🤗
Could you confirm that the test_pipeline_video_classification tests work well locally?

[ArthurZucker]

Regarding the failing test:

VideoClassificationPipeline requires the PyAv library but it was not found in your environment. You can install it with:

pip install av

Please note that you may need to restart your runtime after installation.
FAILED tests/models/timesformer/test_modeling_timesformer.py::TimesformerModelTest::test_pipeline_video_classification_fp16 - ImportError: 
VideoClassificationPipeline requires the PyAv library but it was not found in your environment. You can install it with:

pip install av

Please note that you may need to restart your runtime after installation.

the workflow for pipelines needs a small fix!

[vrnvu]

Fixed the requirement, I wrongly confused "vision" with "video" and thought I didn't need the av requirement. 🤕

How can I troubleshoot them? Are they flaky?

I dont see how they are related with this PR.

[ArthurZucker]

Thanks, yep the other tests are unrelated!

[ArthurZucker]

But the tests you edited are skipped by the CIs because they need av (which was not installed)

[ArthurZucker]

Small nit, LGTM otherwise!

[ArthurZucker]

Suggested change

	frames.append(frame)
	frames.append(frame.to_rgb())

any reason we don't use this?
or directly convert before stacking?

[vrnvu]

Good question, we return the transformation at the end to directly transform the numpy data structure

return np.stack([x.to_ndarray(format="rgb24") for x in frames])

It's more efficient than converting the frames to rgb:
https://github.com/PyAV-Org/PyAV/blob/main/av/video/frame.pyx#L252
https://github.com/PyAV-Org/PyAV/blob/main/av/audio/frame.pyx#L168

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

Okay merging then! 🤗