Open ckrsls opened 11 months ago
Hi @ckrsls ! Sorry for the late reply.
for the multigrid representation, We created multigrid representation by moving sampling points diagonally by adding the 'offset'.
you can find the code in utils/grid_sample : if offset: offset = torch.linspace(0,1-(1/N),N).reshape(N,1,1).to('cuda') iy = ((iy+1)/2)(IH-2) + offset ix = ((ix+1)/2)(IW-2) + offset
so an offset is added to the input sampled data position (iy, ix).
and finally, we calculated the gradients of sampling points to update the weight by using the autograd.
because the gradient will be calculated through autograd, we haven't dealt with the point of going over the edge.
Hi, Your article uses an interesting approach to multigrid representation. I want to know whether all sampling points must be located in the overlapping area of the grid. How to deal with the points at the edge?