Different Behaviors of tranforms.Resize and transforms.v2.Resize

[donglihe-hub]

🐛 Describe the bug

Test codes

import numpy as np
from torchvision.transforms import v2
from torchvision import transforms


if __name__ == "__main__":
    t = np.ones((1023, 587))

    transform_v2 = v2.Compose(
        [
            v2.Resize((224, 224)),
            v2.ToTensor(),
        ]
    )
    print(f"v2 passed. {transform_v2(t).shape} {type(transform_v2(t))}")

    transform_v1 = transforms.Compose(
        [
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
        ]
    )
    print(transform_v1(t).shape)

The expected output of the test codes is a (1, 224, 224) tensor. When input is an numpy.ndarray, transforms.Resize raises

TypeError: Unexpected type <class 'numpy.ndarray'>

While transforms.v2.Resize silently runs the code without resizing the input

v2 passed. torch.Size([1, 1023, 587]) <class 'torch.Tensor'>

IMO, shouldn't v2 behave the same as v1. If their bahaviors are not expected to be the same, shouldn't v2 at least tell users resizing doesn't take effect?

Versions

torchvision 0.21.0

[NicolasHug]

numpy array aren't supported by torchvision's transforms, and they're treated as foreign types (just like int, str, etc.). So on v1 you pass a foreign type and you get an error. On v2, those foreign types are passed-through, which is useful e.g. to keep track of metadata associated with files, like their path, or anything else.