There are minor changes that might be merged later
@simona-rc
Added changes for the ISIC's runner which support: (updated 5.8.23)
Loading model's backbone (ResNet18 & Transformer) prior training stage - if exists.
Saving model's backbone after training stage.
Loading model's backbone for inference stage.
NOTE
I played a little with it and didn't find it too useful for the "golden" subset case. I guess that's because I didn't load the header weights - which should also be the case for Poriya.
Also works OK for different image resolutions since the ResNet backbone extracts a feature vector of size 512 regardless the input image dimensions (see layer4)
Doesn't suppose to pass CICD
Update (5.8.23)
Now also supports checkpointing for the Transformer backbone type.
👁️
Reference only!
There are minor changes that might be merged later
@simona-rc
Added changes for the ISIC's runner which support: (updated 5.8.23)
NOTE
layer4
)Update (5.8.23)
Now also supports checkpointing for the Transformer backbone type.