Hi, is there a smart way to get the output of a convolution without summing over the kernels spatially?
i.e from input [B, 4, 6, 16, 16] to output [B, 6, 14, 14, K, K] where K=3 is the kernel size in a convolution operation.
This can be achieved trivially with unfold, but is it possible using a Conv2d/3d layer with a fancy filter?