quicksign / tfgad Goto Github PK

thanks for the code; does this repo contain the code for iterative training procedure, described as algorithm1 in the paper.

Thanks for sharing the code.
I have some issues about calculating the diffused image.
A discrete numerical solution can be derived for the anisotropic case using the FTCS method as follows:

where {N, S, W, E} correspond to the pixel above, below, left and right of the pixel under consideration (i, j). And c is based on the gradient magnitude, thus it does not matter if we take forward or backward gradients.
How to determine the sign of diffusion coefficient item, i.e.

diffused = image[:, :, 1:-1, 1:-1] + lambda_ * (
        cv[:, 1:, :] * dv[:, :, 1:, :]
        - cv[:, :-1, :] * dv[:, :, :-1, :]
        + ch[:, :, 1:] * dh[:, :, :, 1:]
        - ch[:, :, :-1] * dh[:, :, :, :-1]
    )

Looking forward to your reply.