torchvision.transforms.functional.adjust_hue does not respect alpha channel

[jamespltan]

🐛 Bug

when working with RGBA images, the adjust_hue function does not seem to respect the alpha channel and imposes a max_value of 255 after the transform. Taking a quick look at the source code, it seems that the image is immediately converted to HSV without retaining the alpha channel. It should be a quick fix to retain the alpha channel and include it when merging back into RGBA.

To Reproduce

Steps to reproduce the behavior:

img = Image.open('xyz.png')
img_ = adjust_hue(img, 0.1)
print(np.array(img.split()[-1]).mean())
print(np.array(img_.split()[-1]).mean())

cc @vfdev-5

[datumbox]

@jamespltan Thanks for reporting this.

You are right to say that the adjust_hue() method does not currently support transparency (neither on the PIL nor on the Tensor backend). The same applies for other transformations and unfortunately the documentation does not mention those limitations (cc @voldemortX).

Extending support and providing a proper and consistent solution across both backends is not trivial and will require storing additional meta-data about the images (for example their mode in case of tensors). This definitely requires additional discussion, so before tackling that "beast" I think it would be best to improve the current situation by:

1. Aligning the behavior of the two backends. For instance currently PIL.adjust_hue() returns the image if monochrome while Tensor.adjust_hue() throws an exception.
2. Update the documentation to clearly mention the limitations.
3. Open an issue and list all current limitations so that we can discuss potential solutions.

Any contribution that addresses the above 3 points would be highly appreciated.

[voldemortX]

Yeah I forgot to mention that in the doc...

Maybe I can do the listed 3 points tomorrow to make up for it.

[datumbox]

That would be awesome @voldemortX, thank you. There might be more limitations on the adjust_* methods that we don't clearly mention.

[jamespltan]

Many thanks my dudes 🙏

[Iteravse]

i want to fix this problem,any proposal?i just fix the same question in numpy,but,because it dosn't using gpu,seems like inefficient using for pyqt animation.so i need more efficiency solution,like use pytorch, think it is are not big issues,just copy alpha channel to result, @datumbox any good idea?, forget me dosn't really skilled the grammar with torchvision,and i ask a same questoin in here