Closed PascalCremer closed 5 years ago
or you could just rotate the input and output which is simply taking each board vector reversed.
harbecke notifications@github.com schrieb am Mo., 11. März 2019, 13:03:
we could apply all filters rotated by 180°, i've never done sth like this though
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/harbecke/hex/issues/6#issuecomment-471512247, or mute the thread https://github.com/notifications/unsubscribe-auth/ADGqn4mMiHKsntM6kSWmQAapDmCO87Kpks5vVkYXgaJpZM4balt5 .
so you want to use board symmetry for training, not enforcing evaluation symmetry?
alphago zero uses both. they enhace training data by a factor of 8 using the 8 symmetric go boards and use a randomly selected board during network evaluation. I would fiest implement it for data generation as this is the bottleneck for MCTS models.
harbecke notifications@github.com schrieb am Mo., 11. März 2019, 15:01:
so you want to use board symmetry for training, not enforcing evaluation symmetry?
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/harbecke/hex/issues/6#issuecomment-471550394, or mute the thread https://github.com/notifications/unsubscribe-auth/ADGqnwg2EUhmYC0AvAjzubPAkIysCTP5ks5vVmHSgaJpZM4balt5 .
Implemented in SelfPlayGenerator
for MCTS self play data creation
we could apply all filters rotated by 180°, i've never done sth like this though