I have code similar to this I would like to make faster: # indices: indices of a 3d tensor # values associated to the indices result = torch.zeros((L, N, N)) for idx, (i,j,k) in enumerate(indices): mask = torch.zeros_like(result) mask[j][k] = 1.0 img = img + mask * values[idx] Now, even if I chose to make a sparse mask, I notice that each iteration runs slower. Is there a simple solution to a function that will propagate the gradient of the form img = func(indices, values) I'm looking for a solution that takes advantage of vectorization or/and sparse data structures Continue reading...
