Add interpolation of positional embedding to swin2sr #31024
Conversation
| model = Swin2SRForImageSuperResolution.from_pretrained("caidas/swin2SR-classical-sr-x2-64").to(torch_device) | ||
|
|
||
| image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png") | ||
| image = image.resize([680, 680]) # size is an unrecognized kwargs in processor |
There was a problem hiding this comment.
Swin2SRImageProcessor doesn’t have size as kwargs, tbh it is not really needed since the model can take any input resolution.
However, there might be a need to change input size for some reason like having limited hardware resource and want to train on lower resolution images, also to keep consistent code with other image processors.
If that's the case I'm willing to add it as a PR if approved.
|
Hi @MightyStud, thanks for opening this PR and apologies for the delay in reviewing. As Swin2SR doesn't require this, then I don't think it makes sense to add to the model and we can close this PR. My bad for having it on the list of models to add on the issue - I'll update this now. |
|
@amyeroberts |
What does this PR do?
This PR add the
interpolate_pos_encodingfunction to the Swin2SR transformer model, the purpose of it is to allow input images with resolution different than the pretrained resolution.This is part of the community contribution thread.
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Usefulness
Swin2SR transformer by nature supports different input resolution that is due to the case of log-spaced continuous relative position bias which allows it to generalize to higher input resolution at inference time as stated in the paper.
I'd say having
interpolate_pos_encodingfor Swin2SR is good for code consistency between it and other Visual transformers of the same nature, however its effectiveness is still to be reviewed.Discussion about Swin family ability to support different resolution was made in this PR #30656.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.
@amyeroberts