How to use data augmentation during training process (especially PatchCore)?

[james77777778]

Is your feature request related to a problem? Please describe.
I successfully trained PatchCore with custom dataset, but I find that the threshold is too sensitive and the model is hard to generalize. Maybe some extra data augmentation (like color jittering, rotation, etc.) will help the generalization?

#119 (comment)

Describe the solution you'd like
I know anomalib uses albumentations for preprocessing, but I cannot found an easy way (without hard-coding) to configure it.

Describe alternatives you've considered

Additional context

Thanks for the great work. If there are some plans to make this possible, I can try to contribute.

[samet-akcay]

Hi @james77777778, you could use albumentations to serialize an augmentation pipeline into a yaml file. I've created a PR #156 that reads this augmentation yaml file and applies these custom augmentations to the input images. Following this, you don't need to hardcode anything.

[james77777778]

Thanks for the quick response!

But I have a question. Since PatchCore extracts features only once, how the augmentation pipeline works in this scenario?
It makes random augmentation for each image only once (if epoch=1)? It is recommended to set more epochs?

[samet-akcay]

There is already an ongoing discussion here in #119 you could refer to. Maybe we could continue from there?