`VideoDecoder` API accepts `device` argument as `str` or `torch.device`, but underlying C++ API `add_video_stream` only takes in a `str`.

[divyanshk]

🐛 Describe the bug

Repro:

device = torch.device('cuda:0')
VideoDecoder(video, device=device)

fails with complaints that

trainer/0 [0]: Declaration: torchcodec_ns::add_video_stream(Tensor(a!) decoder, *, int? width=None, int? height=None, int? num_threads=None, str? dimension_order=None, int? stream_index=None, str? device=None) -> ()
trainer/0 [0]:Cast error details: Unable to cast Python instance of type <class 'torch.device'> to C++ type '?' (#define PYBIND11_DETAILED_ERROR_MESSAGES or compile in debug mode for details)
trainer/0 [0]:Skipping video tensor([ 0,  0,  0,  ..., 49, 48, 48], dtype=torch.uint8) due to decoding error: torchcodec_ns::add_video_stream() Expected a value of type 'Optional[str]' for argument 'device' but instead found type 'device'.

VideoDecoder(video, device=str(self.device)) fixes the issue.

Versions

N/A

[scotts]

On the C++ side, we only accept the device as a string because that's the only thing the custom ops can accept. On the Python side, we actually accept Union[str, torch.device]. We need to make sure on the Python side that we convert the torch.device to a string before passing it to the C++ side.

[soheil-star01]

Hi, I can work on this issue.