Potential GPU Memory Leak in LabelStats when label is on GPU but image is on CPU in auto3dseg/analyzer.py

[benediktjohannes]

Current code only checks image tensor for CUDA device

If label tensor is on GPU (even if image is on CPU), torch.cuda.empty_cache() is not called

This can cause GPU memory leaks during connected component analysis

(in auto3dseg/analyzer.py)

[ericspod]

I believe this has now been addressed.

[benediktjohannes]

@ericspod thanks for the answer! Unforetunately, I don't think that this has been addressed since there were only 2 PRs (and one was closed because I reopened it lateron with some changes) and the other is still open? Or is there any other commit fixing this?

[ericspod]

Yes that's true, the fixing PR is another one which isn't merged yet. I think there's code formatting at least on #8708 which needs resolving then it can be merged.

[benediktjohannes]

Thanks for the quick answer! That's true! I'll have a look on this as soon as possible. 👍