Reorder AUC and Python version 3.7

[alexriedel1]

anomalib/anomalib/core/metrics/auroc.py

Line 17 in f1c8a6b

return auc(fpr, tpr)

throws a ValueError: The 'x' tensor is neither increasing or decreasing. Try setting the reorder argument to 'True'.
It happens for large datasets I guess with only little variance in some validation datapoints?

I set reorder=True to resolve the problem.

Also as the development branch is compatible to Python 3.7, it should also be possible to install the package und 3.7
#72

[samet-akcay]

Hi @alexriedel1, thank you for your contribution to the repo! I've just triggered the Pre-merge checks, which is currently running.

[djdameln]

@alexriedel1 Thanks for your PR! I agree that ideally the AUC should be computed with reorder=True. However, calling auc with this parameter may currently lead to an underestimation of the model's performance. The reason for this is that pytorch does not use stable sorting (see https://stackoverflow.com/questions/60366033/torch-sort-and-argsort-sorting-randomly-in-case-of-same-element and pytorch/pytorch#28871). This means that equal elements in the fpr array will lose their original ordering, which can lead to unpredictable behavior in the AUC computation.

It seems that stable sorting has been added in pytorch 1.9.0 as an option to sort, but for technical reasons we are currently bound to 1.8.1. Besides, even in torch 1.9.0, stable sorting is disabled by default, so updating pytorch would not solve the problem by itself.

As a temporary solution we could apply the sorting ourselves, using a stable sorting algorithm, inside the compute method (before computing the AUC).

[djdameln]

@alexriedel1 I created a new PR that should address your problem, but maintain consistent results when the fpr is already ordered correctly. Will be closing this one.