Support grayscale image-like tensors for adjust_contrast and adjust_saturation functions

[VinhLoiIT]

🚀 The feature

The following functions in functional_tensor.py should be updated to work with single-channel image-like tensor, i.e. (1, H, W).

1. adjust_contrast. Similar implementation in Pillow is:

    enhancer = ImageEnhance.Contrast(img)
    img = enhancer.enhance(contrast_factor)

2. adjust_saturation. Similar implementation in Pillow is:

    enhancer = ImageEnhance.Color(img)
    img = enhancer.enhance(saturation_factor)

Motivation, pitch

When updating the documentation, I discovered some transform functions that work well on one channel PIL image (image with mode='L'), yet they only support 3-channel tensors as in #4457.

Therefore, I think it is necessary to re-check and implement those functions to provide a similar result as in transforming PIL images.

Also, after solving this issue, we can further enable users to use AutoAugment and similar classes to work with grayscale image tensors. Currently, grayscale pillow images can use these augmentations without errors, but not with single-channel image tensors, i.e. (1,H,W)

Alternatives

No response

Additional context

No response

cc @vfdev-5 @datumbox

[VinhLoiIT]

Currently, while scanning through the source code for test cases, I find that the test cases for adjust_contrast is defined as follows:

@pytest.mark.parametrize('device', cpu_and_gpu())
@pytest.mark.parametrize('dtype', (None, torch.float32, torch.float64))
@pytest.mark.parametrize('config', [{"contrast_factor": f} for f in [0.2, 0.5, 1.0, 1.5, 2.0]])
def test_adjust_contrast(device, dtype, config):
    check_functional_vs_PIL_vs_scripted(
        F.adjust_contrast,
        F_pil.adjust_contrast,
        F_t.adjust_contrast,
        config,
        device,
        dtype
    )

which is essentially linked to the check_functional_vs_PIL_vs_scripted function that it tests only on 3 channels image-like tensors:

def check_functional_vs_PIL_vs_scripted(fn, fn_pil, fn_t, config, device, dtype, tol=2.0 + 1e-10, agg_method="max"):

    script_fn = torch.jit.script(fn)
    torch.manual_seed(15)
    tensor, pil_img = _create_data(26, 34, device=device)  # this creates 3 channels only
    batch_tensors = _create_data_batch(16, 18, num_samples=4, device=device)
    ...

So, I wonder that some functions, which are documented as they support 1 or 3 channels, are tested on 1 channel tensors or not? For example, adjust_brightness is one of those functions.

Please correct me if I'm wrong so that I can further create a PR for this.

cc @datumbox

[datumbox]

@VinhLoiIT Thanks for digging into this. Yes you are right, we test only for 3 channels and obviously that's not great.

A PR that improves the situation in this direction is very welcome.

[vfdev-5]

@datumbox @VinhLoiIT I can help or guide also with this issue/feature request if needed.

[vfdev-5]

@VinhLoiIT I submitted a draft PR on how to enhance the gray images support for adjust_contrast. I think something similar can be done for adjust_saturation. Once, we figure out all hidden issues (if there are) with adjust_contrast and merged my PR, if you would like to submit another one for adjust_contrast it would be awesome !

[VinhLoiIT]

@VinhLoiIT I submitted a draft PR on how to enhance the gray images support for adjust_contrast. I think something similar can be done for adjust_saturation. Once, we figure out all hidden issues (if there are) with adjust_contrast and merged my PR, if you would like to submit another one for adjust_contrast it would be awesome !

Thank you, with a glance through the functional_tensor.py file, I have found these two related functions as below:

• adjust_contrast
• adjust_saturation

I guess that this issue is because of the rgb_to_grayscale method calls, which could be done without converting, i.e. broadcasting. However, I think we have to write some test cases for this. I've had some struggles with the tensor-pillow comparison and your PR is really helpful. Do I have to create a PR for this repo or your fork?

[vfdev-5]

@VinhLoiIT yes, I agree with you about rgb_to_grayscale and that we can bypass it if input image is gray.
I think you can checkout this branch: https://github.com/vfdev-5/vision/tree/add-gray-support-adjust_contrast and submit a PR to this repo: pytorch/vision. After the first PR has been merged you can update your branch to main and I think we will be good.

[VinhLoiIT]

Sure I'll do. I'll create a PR soon.

[vfdev-5]

@VinhLoiIT the PR was merged (thanks to Vasilis for a quick review) you can just fork from pytorch/vision:main and not bother with my fork

[VinhLoiIT]

@vfdev-5 I've created a PR, please kindly have a look.