🚀Add SuperSimpleNet model

[blaz-r]

📝 Description

• This PR adds the SuperSimpleNet model implementation.
• I also verified the performance on MVTec AD and the results are very close tho the ones reported in the paper.

✨ Changes

Select what type of change your PR is:

• 🐞 Bug fix (non-breaking change which fixes an issue)
• 🔨 Refactor (non-breaking change which refactors the code base)
• 🚀 New feature (non-breaking change which adds functionality)
• 💥 Breaking change (fix or feature that would cause existing functionality to not work as expected)
• 📚 Documentation update
• 🔒 Security update

✅ Checklist

Before you submit your pull request, please make sure you have completed the following steps:

• 📋 I have summarized my changes in the CHANGELOG and followed the guidelines for my type of change (skip for minor changes, documentation updates, and test enhancements).
• 📚 I have made the necessary updates to the documentation (if applicable).
• 🧪 I have written tests that support my changes and prove that my fix is effective or my feature works (if applicable).

For more information about code review checklists, see the Code Review Checklist.

[samet-akcay]

This is amazing @blaz-r! Thanks a lot! I'll review it shortly

[blaz-r]

I have fixed the linter error and formated the files. The tests for SuperSimpleNet are still failing on onnx and openvino export locally. I'm not entirely sure how to fix this, but I'll take a lot when I get the time.

[blaz-r]

Tests are failing on imgaug, somethign with numpy version incompatiblity it seems.

[samet-akcay]

@blaz-r, we will need to remove imgaug dependency to bump the numpy version.

[samet-akcay]

What are the missing points in Anomalib to support supervised setting

[blaz-r]

Right now I believe there are no standard supervised datasets. Another problem is the Folder dataset as it assumes that abnormal samples are always in test set:

anomalib/src/anomalib/data/image/folder.py

Lines 174 to 175 in bcc0b43

	samples.loc[(samples.label == DirType.NORMAL), "split"] = Split.TRAIN
	samples.loc[(samples.label == DirType.ABNORMAL) | (samples.label == DirType.NORMAL_TEST), "split"] = Split.TEST

Another thing for full reproduction of SuperSimpleNet results is the fixed flipping augmentation and frequency sampling. This is however not necessary, but needed for best results. It's also not SuperSimpleNet specific, so might be worth considering if other supervised model will be supported.

[samet-akcay]

yeah, we would like to diversify the model pool, and include more learning types than one-class models. Thanks for the feedback.
@abc-125, you might want to be aware of this discussion as you have recently worked on this stuff

[abc-125]

Thank you for adding me, it would be great to have supervised models and datasets in Anomalib. Recently, I looked at how to add a supervised dataset, and it certainly would require changing some base structures, such as paths to folders (I guess we will need abnormal_train_dir and maybe renaming the rest to make it easier to understand, normal_train_dir, etc.):

anomalib/src/anomalib/data/image/folder.py

Lines 195 to 202 in bcc0b43

	normal_dir (str | Path | Sequence): Path to the directory containing normal images.
	root (str | Path | None): Root folder of the dataset.
	Defaults to ``None``.
	abnormal_dir (str | Path | Sequence | None, optional): Path to the directory containing abnormal images.
	Defaults to ``None``.
	normal_test_dir (str | Path | Sequence | None, optional): Path to the directory containing
	normal images for the test dataset.
	Defaults to ``None``.

[samet-akcay]

@blaz-r, I've created #2436 to remove imgaug stuff and write our own perlin noise generation stuff. Would it be an idea to try this in supersimplenet ?

[blaz-r]

Sure. SuperSimpleNet used only _rand_perlin_2d function, which I think can now be replaced with the new generate_perlin_noise.