Allow ToImage to handle image paths

[mantasu]

🚀 The feature

Was just wondering if it was possible to add support for handling image paths when using ToImage. Here's the desired feature:

import numpy as np
from PIL import Image
from torchvision.transforms.v2 import ToImage

img_np = np.random.randint(0, 256, (100, 100, 3), dtype=np.uint8)
Image.fromarray(img_np).save("test_img.jpg")

print(ToImage()("test_img.jpg").shape) # torch.Size([3, 100, 100])

Motivation, pitch

This simply allows us to avoid manually loading images, for example in custom Datasets:

class MyDataset(torch.utils.data.Dataset):
    def __init__(self):
        self.transform = v2.Compose([v2.ToImage(), v2.Normalize([0.5]*3, [0.3]*3)])
    def __getitem__(self, idx):
        return self.transform(f"path/to/img-{idx}.jpg")

Also, would be incredibly handful when using a standalone to_image utility.

Alternatives

No response

Additional context

No response

[NicolasHug]

Hi @mantasu , thanks for the feature request. Traditionally, the decoding utilities are kept separate from the transforms, as those tend to have fairly different parametrization.

Is read_image what you're looking for?

[mantasu]

Yeah, I was wondering if this could be added to to_image with an additional final check for str. I created a PR with an example functionality. But feel free to close it if it is better to keep those separate!

[NicolasHug]

But feel free to close it if it is better to keep those separate!

Thanks for understanding @mantasu - yes, let's keep those separate. Our converstion transforms (e.g. ToImage, ToTensor, ToPILImage, etc.) have been the source of a lot of confusion in the past, e.g. ToTensor() would silently scale the values of the input and convert a uint8 PIL image to float... We've tried to clean that a little, but backward-compatibility engagement make that hard. I'm afraid that adding another surface to those APIs (i.e. the ability to pass paths) would add a new layer of complexity to this already-messy space, so I prefer avoiding this. Thank you!