Samet Akcay
@JeffCHEN2017 , @longcw, can you please elaborate how you managed to train with multiple batch size? As you guys suggested, I re-ran roi_pooling_cuda.c by commenting out the relevant lines. Then,...
@jpcbertoldo, `torchvision` was previously the base package for the CNN backbones and feature extraction. Recently we made this switch to timm as it has a feature extractor, which even works...
ah yes, of course.
Looks like it is also used in reverse distillation `from torchvision.models.resnet import conv1x1, conv3x3` and `from torchvision.models.resnet import BasicBlock, Bottleneck`
> It might also be an idea to deprecate the FeatureExtractor My bad, forget about it. I just forgot that we already refactored `FeatureExtractor` to use `timm`'s feature extraction capabilities....
Addressed with #576
@Vvdinosaur, until we optimize the code, you could perhaps reduce the number of layers you are extracting. For example, you could only extract the features from 'layer3'. You could potentially...
@manuelblancovalentin, as @alexriedel1 pointed out, padim and patchcore are not memory efficient. If you get OOM even in a single epoch, you could try @alexriedel1's suggestion, or alternatively try to...
I'll convert this to a discussion, feel free to continue from there.
@ashwinvaidya17, I'm wondering if this is the case even when we use `timm`'s feature extractor