For group convolution, if both input and output have the shape of H W C, and the number of groups is 4, the kernel size is k, then for each location, the weight should have the shape of C/4 k k C/4 4 = C (C/4) k^2. But in paper, M[i, j] have the shape of C 4 k^2.
For group convolution, if both input and output have the shape of H W C, and the number of groups is 4, the kernel size is k, then for each location, the weight should have the shape of C/4 k k C/4 4 = C (C/4) k^2. But in paper, M[i, j] have the shape of C 4 k^2.