Error in functional.uniform_temporal_subsample when the input tensor has non-default device type

[Michael Hu]

🐛 Bugs / Unexpected behaviors

In the transforms.functional.uniform_temporal_subsample, when the input tensor x and the indices do not have the same device type, the code will fail to run.

Possible causes is that we set some default device torch.set_default_device(device) at global level, however our input tensor x might have a different device, thus causing the issue.

   return pytorchvideo.transforms.functional.uniform_temporal_subsample(
  File "/home/michael/.local/lib/python3.10/site-packages/pytorchvideo/transforms/functional.py", line 41, in uniform_temporal_subsample
    return torch.index_select(x, temporal_dim, indices)
  File "/home/michael/.local/lib/python3.10/site-packages/torch/utils/_device.py", line 77, in __torch_function__
    return func(*args, **kwargs)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

Instructions To Reproduce the Issue:

Please include the following (depending on what the issue is):

1. Set default global level

torch.set_default_device('cuda')

2. Use load_and_transform_video_data from ImageBind to process some video data https://github.com/facebookresearch/ImageBind/blob/c6a47d6dc2b53eced51d398c181d57049ca59286/imagebind/data.py#L287