3D extension for SSIM

[weningerleon]

What does this PR do?

Fixes #812
Changes StructuralSimilarityIndexMeasure to be able to handle 3D images. Deciding whether the 2D or 3D version is used is determined automatically based on kernel_size. Basic sanity checks are carried out. Produces sensible results on my dataset.

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure to update the docs?
• Did you write any new necessary tests?

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

Did you have fun?

Make sure you had fun coding 🙃

[justusschock]

Some first comments.

Usually there are 2 ways for testing:

1.) Use random data and compare the results with a gold-standard implementation (like sklearn)

2.) Use predefined-data, calculate result manually once and hardcode those results as targets.

Ideally we would prefer Option 1 since that way you can always test it with additional data if required. I think the sklearn implementation here should also work with 3D data :)

Edit: have a look at the 2d tests here

[codecov]

Codecov Report

Merging #818 (b057212) into master (25c4b90) will decrease coverage by 0%.
The diff coverage is 88%.

@@          Coverage Diff          @@
##           master   #818   +/-   ##
=====================================
- Coverage      95%    95%   -0%     
=====================================
  Files         173    173           
  Lines        7257   7328   +71     
=====================================
+ Hits         6900   6958   +58     
- Misses        357    370   +13     

[weningerleon]

I think your current tests that compare the results from torchmetrics with the results from skimage are faulty.
Have a look at the 2d tests mentioned by @justusschock, function _sk_ssim:
You permute the 4D tensor (BxCxHxW) to BxHxWxC, with B=batch_size, C=image channels. Skimage treats the BxHxWxC tensor as a 3D image with C channels. However, it needs to be treated as B 2D images (skimage does not support batching). This also explains why you need to set atol to 6e-3, which is really high.

[justusschock]

@weningerleon If you could fix those while working on it, this would also be highly appreciated :D

[weningerleon]

@weningerleon If you could fix those while working on it, this would also be highly appreciated :D

Yes, have a look at the updates files ;)

[justusschock]

@Borda why are the tests not running?

[Borda]

@Borda why are the tests not running?

I guess it is due to GH outage...

[Borda]

@justusschock mind check the failing tests, pls? 🐰

talked to him and he is fine with passing tests :)

[Borda]

@weningerleon thank you for all your effort; very appreciated!

[weningerleon]

Glad I could help to improve this awesome project!
Thanks a lot for your support!