optimize compute_anomaly_score by using torch native funcrtions

[alexriedel1]

Hey,
I realized that the compute_anomy_score method

anomalib/anomalib/models/patchcore/model.py

Line 69 in ee8807b

def compute_anomaly_score(patch_scores: torch.Tensor) -> torch.Tensor:

can be optimized by using torch native functions that might vectorize better.

Using a GPU, the execution time reduces from 0.17330350000000116s to 0.00022549999999910142s (~700x faster)

[samet-akcay]

Thanks @alexriedel1! That's a good catch! We'd be able to merge it after @djdameln approves the PR

[alexriedel1]

Thanks @alexriedel1! That's a good catch! We'd be able to merge it after @djdameln approves the PR

Theres one more thing I'd like to adress: torch.cdist() is slow (especially for large tensors) (pytorch/pytorch#15253)

anomalib/anomalib/models/patchcore/model.py

Line 239 in 5a5e552

distances = torch.cdist(embedding, self.memory_bank, p=2.0) # euclidean norm

This cdist implementation is fast: pytorch/pytorch#15253 (comment)

Maybe you can think about using the version mentioned above. I can create another pull request for this.

[djdameln]

Thanks for addressing this @alexriedel1!

[samet-akcay]

@alexriedel1, would you like to address that in a separate PR, or in this one? If the former, I could merge this one.

[alexriedel1]

@alexriedel1, would you like to address that in a separate PR, or in this one? If the former, I could merge this one.

Let me do another PR so you can merge this one