Open mjack3 opened 2 years ago
I unfortunately no longer have the particular training logs with me, but I did do a blockwise grid-search over EfficientNet layers.
What this means is that every layer that shares the same resolution of its featuremap is grouped together and then samples from, and then used to create the feature representations for PatchCore.
I.e. how standard PatchCore with WideResNet50 uses the outputs of layer2
and layer3
, for EfficientNet I drew a layer from similar resolution groups for comparability.
How are the layers called? I tried e.g. blocks.4
, but it wasn't found.
Just curious.