conv for seperable kernels
It's efficient to do convolution with larger kernels if they are seperable, the sum of seperalable kernels. Suppose each kernel was the outer product of two matrices: R, S. We would like an Op that computes something like conv_seperable(images, R, S, mode='valid').
Leave a comment
on 2010-11-08 12:10 *
By Ian Goodfellow
Is there a reason to do this as one op instead of just composing 2 ops? (I am doing separable convolution for the DARPA video project by composing 3 ops)