reverse rotate after nn forwarding ? about leela-zero HOT 4 CLOSED

No, we are mapping the rotated output vtx back into the regular orientation. vtx is the regular point we are considering and rot_idx is where we look for it in the rotated board output.

from leela-zero.

https://github.com/gcp/leela-zero/blob/8507cb47528f43729d6bd723b3890f637ce9e576/src/Network.cpp#L425
For consistency vtx should probably be named rot_idx there.

The neural network board at x contains planes[rotate_nn_idx(x)].

The output board at x will look for net_board[rev_rotate_nn_idx(x)].

x = rev_rotate_nn_idx(rotate_nn_idx(x)) so output and planes have the same orientation.

from leela-zero.

I guess this code would be easier to understand and simpler if we change it so the regular board is always on the left-hand side and the network board is on the right-hand side. Then one can probably get rid of rev_rotate_nn_idx as well.

(In image rotation one typically has to "sit at the output" and look for the corresponding input to get non-90° angles right, but that's not necessary here)

from leela-zero.

vtx is the regular point we are considering and rot_idx is where we look for it in the rotated board output.

my mistake

from leela-zero.