Issue with patchcore anomaly score for large confidence scores

[bsl546]

Describe the bug
When the confidence is very high for images treated by the patchcore model, the anomaly score can be nan.
This is due to the formula:
weights = 1 - (torch.max(torch.exp(confidence)) / torch.sum(torch.exp(confidence)))
in anomaly_map.py, torch.exp() gives nan for values too large.

Expected behavior
Two possible solutions:

• Either a threshold should be set on the tensors, like
  confidence[confidence>thres] = thres
  Possible values for thres are e.g., 50.0.

• Or, even cleaner, replace 1- (torch.max(torch.exp(confidence)) / torch.sum(torch.exp(confidence))) by 1-torch.max(torch.nn.Softmax()(confidence))

[djdameln]

Hi, thanks for spotting this and apologies for the late response. I agree that the latter solution, replacing by 1 - torch.max(torch.nn.Softmax()(confidence)), would be the best way to address this. Would you be interested in submitting a PR for this so you could become a contributor?

[bsl546]

Yes sure, I will do it today if possible.