Hi @ptrblck
Thank you for your explanations.
In my explanations, I assume kernels as the axes/dimension for each filter.
Based on your example, each one of 10 featuremaps refers to a filter which contains three kernels(because of the input dimension.)
In other words, during the Conv. operation, 30 featuremaps are produces, afterwards they are summarized to 10 featuremaps.
I meant, instead of accessing 10 featuremaps using hook function, accessing those 30 featuremaps.
Is groups argument that you mentioned a solution to what I want(30 featuremaps). If so, is it possible to apply such operation on pre-trained CNN such as VGG19? or shall I train it from scratch with the Conv. layers including groups argument?
Thank you very much.