vision
vision copied to clipboard
Clarify `TVTensor` shapes
📚 The doc issue
TVTensor
is documented as being a subclass of torch.Tensor
specifically for transforms. Perhaps it could be added that they only constrain the shape of torch.Tensor
?
For example, I only use the utility functions of transforms.v2
and, when to_image
referred me to Image
, I still had to check the source code to confirm it is just a torch.Tensor
of shape (C x H x W)
.
Suggest a potential alternative/fix
No response