I recently saw that there is an option of groups for Conv2d. But I don’t understand how to use it.
In my case, I want to take each convolution output.
In traditional convolution networks, if we assume that input channel (R,G,B) is 3 and output channel (J1, J2) is 2,
J1 = F1 * R + F2 * G + F3 * B
J1 = F4 * R + F5 * G + F6 * B
where F1,F2,… denotes filters (say 3x3 filters) and * denotes convolution operator.
In my case, I want to manipulate the each convolved result. So I want to take the followings, not output images:
F1 * R, F2 * G, F3 * B,
F4 * R, F5 * G, F6 * B
Is this possible by using group
option? Thanks.