I am trying to implement the following architecture which utilizes similarity matrix of size
n^2 x n^2 and it is difficult to fit such a huge matrix in GPU. So we are calculating matrix n^2 x M, where is M is some radius of neighborhood pixels then perform the convolution.
Image taken from Convolution RWN
- How to compute pairwise L1 distance from first Convolution layer efficiently (n^2 x M)?
- How to multiply sparse matrix with activations output final step of Random Walk Layer i.e. (k x n^2) * (n^2 x M)?
I would appreciate any pointers.