Added typing annotations to io/_video_opts

[frgfm]

Following up on #2025, this PR adds missing typing annotations in io/_video_opts.py.

Any feedback is welcome!

[pmeier]

Thanks for the PR @frgfm. Unfortunately this breaks torch.jit. Relevant test failure from test_probe_video_from_memory_script - test.test_video_reader.TestVideoReader

E           RuntimeError: 
E           Unknown type name 'np.ndarray':
E             File "/root/project/torchvision/io/_video_opt.py", line 432
E           def _probe_video_from_memory(
E               video_data: Union[torch.Tensor, np.ndarray],
E                                               ~~~~~~~~~~ <--- HERE
E           ) -> VideoMetaData:
E               """

[frgfm]

Thanks for the PR @frgfm. Unfortunately this breaks torch.jit. Relevant test failure from test_probe_video_from_memory_script - test.test_video_reader.TestVideoReader

E           RuntimeError: 
E           Unknown type name 'np.ndarray':
E             File "/root/project/torchvision/io/_video_opt.py", line 432
E           def _probe_video_from_memory(
E               video_data: Union[torch.Tensor, np.ndarray],
E                                               ~~~~~~~~~~ <--- HERE
E           ) -> VideoMetaData:
E               """

Oh, should I remove the typing for this function only then @pmeier?

[pmeier]

The problem here is that torch.jit currently does not support the Union type. There is a PR pytorch/pytorch#53180 that adds support, but it has not been landed yet.

I think the way forward is to try to eliminate the Union and try to get mypy to live with it. If that is not possible than yes, we need to remove the annotations again since in general torch.jit > mypy.

[frgfm]

Apart from the Union issue with torch.jit, I checked locally and mypy is working perfectly.
I had to ignore assignment I couldn't get to work at the end of the file, and if you have an idea how to get around the jit issue, I'm all ears @pmeier :)

[pmeier]

Hey @frgfm and sorry for the delay.

Apart from the Union issue with torch.jit

Until this is fixed upstream, there is nothing we can do about it.

I had to ignore assignment I couldn't get to work at the end of the file

I've answered inline with some more comments.

[pmeier]

Thanks a lot @frgfm!

[datumbox]

It seems that start_pts, end_pts, *pts* switch between ints and floats in different methods. It's worth checking that in all signatures where we declare them as ints are indeed always ints. That might be easier to do with a debugger.

cc @prabhat00155

[frgfm]

As mentioned in the other comment, there is no clear dtype in

vision/torchvision/csrc/io/video_reader/video_reader.h

Line 29 in 7947fc8

torch::List<torch::Tensor> read_video_from_file(

and considering it's not casted afterwards, I'm not sure how to answer this 🤷‍♂️

[datumbox]

Sorry I missed the other comment. I understand it's not straightforward to identify the types. That's why I think finding out what they are and annotating the code-base is useful. Nevertheless if we fail to annotate them correctly (say they are floats and we mark them as ints) it's going to be really confusing.

Can we confirm the types of the various pts vars by running the unit-tests and putting a debugger to observe their materialized types?

[datumbox]

Unfortunately you got a few linter and typing failures.

[prabhat00155]

Thanks @frgfm!

[pmeier]

Two minor things left from my side. LGTM, when the other open comments are resolved and CI is happy. Thanks a lot @frgfm!

[github-actions]

Hey @prabhat00155!

You merged this PR, but no labels were added. The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py