w_x1 = conv_x1_flat.view(N,H,W,four_directions//3,3) # N, H, W, 32, 3

[liuyi1989]

Hi! Why this tranformation?

[danieltan07]

According to the paper the guidance network will output a tensor of size (N, H, W, C x 3 x 4). The 3 corresponds to the three numbers in every row of the tridiagonal matrix (the numbers are the weights / contributions of the 3 adjacent pixels in the previous row/column), and 4 corresponds to the four directions (left to right, right to left, up to down, down to up).

[yliu1989]

Thanks a lot!