Add pathlib.Path support to read_image

[Yura52]

When I try to pass an object of pathlib.Path to torchvision.io.read_image, I get:

RuntimeError: image::read_file() Expected a value of type 'str' for argument '_0' but instead found type 'PosixPath'.
Position: 0
Value: PosixPath('<SOME PATH>')
Declaration: image::read_file(str _0) -> (Tensor _0)
Cast error details: Unable to cast Python instance to C++ type (compile in debug mode for details)

Casting the object to str fixes the problem. It would be nice to support the usual set of "path-like" types as input (str, bytes, pathlib.Path, ...)

python==3.8.3
torch==1.7.0
torchvision==0.8.1

[vfdev-5]

@Yura52 thanks for the feature request. This seems reasonable, however there could be a blocker if we would like to use it with torchscript. We have to check.

[fmassa]

As @vfdev-5 mentioned, we would like to support pathlib.Path, but it wouldn't support torchscript for pathlib.Path.

As you mentioned, this could be probably easily done by casting internally to str in

vision/torchvision/io/image.py

Line 61 in 32e5700

data = torch.ops.image.read_file(path)

and

vision/torchvision/io/image.py

Line 74 in 32e5700

torch.ops.image.write_file(filename, data)

and adding something like

path = str(path)

before the function calls.

[vict0rsch]

Any news on this front?

[LifeGains]

+1 support for this please

[eldar]

would be nice to have.

[NicolasHug]

I believe this was addressed by #8314