Fix bug with limit version of torchvision because of basicsr and gfpgan.archs

[wladradchenko]

Hi @xinntao,

I've submitted a pull request addressing the compatibility issue mentioned in issue #768. My changes can be reviewed in the following branch: wladradchenko/Real-ESRGAN.

I’ve updated the files inference_realesrgan_video.py and inference_realesrgan.py, which are key entry points, to dynamically handle missing functions without restricting the torchvision version. With these updates, it’s possible to remove the existing limitation on torchvision (and by extension, torch), allowing users to leverage the latest versions of both torch and CUDA.

This updated code:

import sys
import types

try:
    # Check if `torchvision.transforms.functional_tensor` and `rgb_to_grayscale` are missing
    from torchvision.transforms.functional_tensor import rgb_to_grayscale
except ImportError:
    # Import `rgb_to_grayscale` from `functional` if it’s missing in `functional_tensor`
    from torchvision.transforms.functional import rgb_to_grayscale

    # Create a module for `torchvision.transforms.functional_tensor`
    functional_tensor = types.ModuleType("torchvision.transforms.functional_tensor")
    functional_tensor.rgb_to_grayscale = rgb_to_grayscale

    # Add this module to `sys.modules` so other imports can access it
    sys.modules["torchvision.transforms.functional_tensor"] = functional_tensor

Thanks for considering these improvements!