Syntax error in type annotations

[pmeier]

In the light of first inline type annotations, a static type check via mypy was requested in #2034 (review).
Running mypy on the current torchvision results in syntax errors in several type annotations. For example:

torchvision/io/_video_opt.py:86: error: syntax error in type comment '(List[int])'

Since syntax errors are fatal they cannot be silenced even with ignore_errors set. Furthermore, mypy stops the checking after it encountered a module with syntax errors.

Thus, if we want to include mypy checks in CI and test if they work we need to either fix them or disable them temporarily and fix them later. Since they seem to be related to torch.jit, I have no idea how much work fixing them would be or how much damage temporarily disabling them will do.

[fmassa]

I've noticed similar issues as well in the past, and I tried to fix some of those in #1696

We need those type annotations as those functions are being used with TorchScript, so we need to find a way to make them compliant with TorchScript and mypy.

cc @eellison for thoughts.

[pmeier]

You never merged the PR although @eellison approved it. Does that mean the way you did it does not work with TorchScript? If adding the return type as you did does indeed work with TorchScript could you rebase and merge your PR. I would take up the rest.

[fmassa]

That PR had failing tests, and I never had the chance to get back to it and fix it so that it is compatible with mypy and torchscript.

@zhangguanheng66 are you still looking at fixing type annotations in #1696?

[zhangguanheng66]

I think the the output annotation Tuple[Tensor, Tensor] is not supported by torchscript. @eellison

It seems that a tuple of tensor is supported. For example, the multi_head_attention_forward function here

[pmeier]

This is fixed by #1696 and #2195.