add torch.script tests to convert_image_dtype

[pmeier]

Addresses #2078 (comment)

[pmeier]

Test failure seems related. @fmassa Could you have a look if I misunderstood the usage of torch.jit.script?

[fmassa]

I think the test failures are because torch.jit.script expects a function, but you are passing the output of the function (which is a tensor)

[pmeier]

I think the test failures are because torch.jit.script expects a function, but you are passing the output of the function (which is a tensor)

Oh dang, that should have been obvious. Sorry, I was tired when I wrote it.

[fmassa]

I think test failures might be either due to a breakage in PyTorch (cc @eellison does it ring a bell for the Faster R-CNN and Mask R-CNN tests?) or you need to rebase on top of master (which I think it's the case, just confirming)

[pmeier]

Will try the rebase.

I've seen errors such as this:

RuntimeError: 
Tried to access nonexistent attribute or method 'is_floating_point' of type 'int'.:

I don't think this will be fixed by a rebase. Is it possible that I'm not allowed to use that attribute in torch.script?

Edit: This branch is based on c2e8a00, which is the HEAD of master at the time of writing this.

[fmassa]

@pmeier I see, it's an issue with torchscript (it currently doesn't have dtype mapped to torchscript internally).

For now, the workaround would probably be to do

img.dtype in [torch.float16, torch.float32, torch.float64]

until torchscript supports it.

But then, we will have another issue afterwards, which is that torchscript doesn't support finfo and iinfo

I think for now it's better to wait until torchscript adds better support for dtype, so let's hold on with this PR for now. :-/

cc @eellison

[pmeier]

I've pushed the last commit before I saw your comment. You can ignore it for noe and I will revert it when we move forward with this PR.

[pmeier]

Superseded by #2485.