Let's:

• remove the first call to F.cvcuda_to_tensor(image) so that the flip operation is done on the cvcuda tensor
• simplify with a single assert_equal with:

if isinstance(image, cvcuda.Tensor):
            expected = F.cvcuda_to_tensor(F.vertical_flip(image))
        else:
            expected = F.to_image(F.vertical_flip(F.to_pil_image(image)))
        assert_equal(actual, expected)

See this comment discussion: #9277 (comment)

Let's:

• remove the print statement
• remove the first call to F.cvcuda_to_tensor(image) so that the flip operation is done on the cvcuda tensor
• simplify with a single assert_equal with:

if isinstance(image, cvcuda.Tensor):
            expected = F.cvcuda_to_tensor(F.vertical_flip(image))
        else:
            expected = F.to_image(F.vertical_flip(F.to_pil_image(image)))
        assert_equal(actual, expected)

Hi @AntoineSimoulin,

Thanks a lot for the comment. After taking a closer look, I think my implementation actually works well.

1, The thing is for cvcuda input, the actual = fn(image) is a cvcuda tensor but the assert_equal could not handle the cvcuda tensor so I get an error using your suggestion: "E TypeError: No comparison pair was able to handle inputs of type <class 'nvcv.Tensor'> and <class 'nvcv.Tensor'>." Thus I have to convert the actual to the torch tensor and compare by using assert_equal(F.cvcuda_to_tensor(actual), expected). (Did we add the cvcuda tensor comparison into the assert _equal()? If so, I think I can make it consistent with the PIL implementation for this part.)

2, And for the expected = F.horizontal_flip(F.cvcuda_to_tensor(image)), I think the logic is to convert the cvcuda tensor to torch tensor first and then apply the flip to get the expected format just like the PIL image part F.horizontal_flip(F.to_pil_image(image)). And the reason the code uses "expected = F.to_image(F.horizontal_flip(F.to_pil_image(image)))" for PIL image is because the assert_equal can handle the PIL image comparison.

Please let me know if I understand correctly.
I will definitely remove the 'print' statement and the unnecessary comments.

I feel we should maybe add cvcuda.Tensor to _transformed_types in the base class Transform under "vision/torchvision/transforms/v2/_transform.py". This way, we don't have to add it every time. Maybe something like:

# Class attribute defining transformed types. Other types are passed-through without any transformation
    # We support both Types and callables that are able to do further checks on the type of the input.
    _transformed_types: tuple[type | Callable[[Any], bool], ...] = (torch.Tensor, PIL.Image.Image)
    if CVCUDA_AVAILABLE:
        _transformed_types += (cvcuda.Tensor,)

Regarding this, I have run into this issue on the other transform PRs. To solve this, I added a is_cvcuda_tensor function and added this to the _transformed_types and in the query_size function. What is the preferred method for handling this, the tuple setup or using a function?

you can dispatch based on the make_input function itself to avoid issues if CV-CUDA is not available