The board has colored edges now. These are identical to the GUI edges, so there is no overlapping and transposing the edges gives the edges of the opponent.
This only changes Board.board_tensor and not Board.logical_board_tensor in logic.hexboard. The models have a padding reduced by 1 in the first layer, so they are reusable, obviously they are not strong anymore.
Comparison to the old models is not possible automated. However, I finetuned the weights of the old model, and after only 1 day 11_2w4_2000 beat the website model 2-1 as starting player.
This is the amazing puzzle loss function of the first half hour of finetuning. Furthermore, the model now employs the "walking the edge" strategy against itself, this might indicate that it is not as vulnerable against this strategy anymore. @simonant an evaluation against mohex would be much appreciated! @cleeff @simonant I hope this isn't a porting nightmare ;)
The board has colored edges now. These are identical to the GUI edges, so there is no overlapping and transposing the edges gives the edges of the opponent.
This only changes
Board.board_tensor
and notBoard.logical_board_tensor
inlogic.hexboard
. The models have a padding reduced by 1 in the first layer, so they are reusable, obviously they are not strong anymore.Comparison to the old models is not possible automated. However, I finetuned the weights of the old model, and after only 1 day
11_2w4_2000
beat the website model 2-1 as starting player.This is the amazing puzzle loss function of the first half hour of finetuning. Furthermore, the model now employs the "walking the edge" strategy against itself, this might indicate that it is not as vulnerable against this strategy anymore. @simonant an evaluation against mohex would be much appreciated! @cleeff @simonant I hope this isn't a porting nightmare ;)