#26566 swin2 sr allow in out channels

[marvingabler]

What does this PR do?

This PR adds the feature of accepting arbitary number of input and output channels when using the Swin2SR model. This allows to perform super resolution from greyscale (1 channel) to color (rgb), or from low resolution multi band satellite to high resolution rgb satellite.

All examples and pretrained models are running as expected based on my tests. No new dependencies have been added.

Just use it like

from transformers import Swin2SRForImageSuperResolution, Swin2SRConfig
import torch

Swin2SRConfig = (
     num_channels_in=1,
     num_channels_out=3
)
model = Swin2SRForImageSuperResolution(Swin2SRConfig)

with torch.no_grad():
    # or use the image preprocessor per default
    out = model({"pixel_values":torch.randn((1,1,264,264))})

Fixes #26566.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Yes here
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests? No, test where there already.

Tagging the reviewers

• text models: @ArthurZucker and @younesbelkada
• vision models: @amyeroberts

[younesbelkada]

Hi @marvingabler
Thanks for your contribution! as you are removing the num_channels attribute I think that this is a breaking change, what about keeping num_channels and make it behave as num_channels_in and use num_channels_out as an optional argument that is initialized as the same value as num_channels in case it is set to None. That way I believe changes will be backward compatible. What do you think?

[marvingabler]

Good point, yes lets do that! Let me update the PR soon :)

[younesbelkada]

Looking great to me! thanks!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[ArthurZucker]

Thanks! Can you add a small test as well? Making sure that a dummy model with a this can still perform as expected! 😉

[marvingabler]

Just realized that there are a couple of more changes required, as the Swin2SRForImageSuperResolution denormalizes based on the input images, while for the case of mapping from multiband images to single band, the mean&stds of inputs and outputs differ. Will add the changes soon.

[ArthurZucker]

Looks good to me thanks for adding this.

[LysandreJik]

Great, thanks!