The selfplay-setting.json controls the training process. Here are the parameters.
{
"NeuralNetwork" : {
"NNType" : "Residual",
"MaxBoardSize" : 19, # The max size in the self-play game. It is
# OK if this value greater than training games,
# but will hurt the performance.
"ResidualChannels" : 128, # Channel size of residual.
"PolicyExtract" : 24, # Channel size of policy head.
"ValueExtract" : 24, # Channel size of value head.
"SeRatio" : 4, # squeeze ration of SE module
"Activation" : "mish", # should be one of relu/mish
"Stack" : [
"ResidualBlock", # The 1st residual block. It is normal block.
"ResidualBlock", # The 2nd residual block. It is normal block.
"ResidualBlock-SE", # The 3rd residual block. It is block with SE module.
"ResidualBlock", # The 4th residual block.
"ResidualBlock", # The 5th residual block.
"ResidualBlock-SE" # The 6th residual block.
]
},
"Train" : {
"UseGPU" : null,
"Optimizer" : "SGD",
"StepsPerEpoch" : 4000, # Save the weight evey this steps.
"ValidationSteps" : 100,
"VerboseSteps" : 1000,
"MaxStepsPerRunning" : 4000, # Will stop the training after this steps.
"Workers" : 4, # Number of data loader workers.
"BatchSize" : 256,
"BufferSize" : 524288, # Bigger is better but it will use more memory. If your
# compute is only 32GB, you can set it as around 256000.
"DownSampleRate" : 16, # Bigger is better but may be slow down.
"MacroFactor" : 1,
"WeightDecay" : 1e-4,
"ChunksIncreasingC" : 5000, # Slowly increase the replay buffer size when probing at least
# chunks. Will load "NumberChunks" chunks if we don't give
# any value (or null).
"NumberChunks" : 20000, # Will load last X chunks at most. Each chunk has only one game
# by default.
"PolicySurpriseFactor" : 0.5, # One factor of sample rate
"LearningRateSchedule" : [
[0, 1e-2] # The format is [X, lr]. Will use the lr rate
# after X stpes. You only need to change the lr
# part in the reinforcement learning.
],
"TrainDirectory" : "selfplay/data",
"StorePath" : "workspace"
}
}
The selfplay-config.txt controls the self-play process. Here are the parameters.
--dirichlet-noise # Enable Dirichlet noise in MCTS.
--random-moves-factor 0.08 # Do the random move if the move number is
# below the (X * intersections).
--random-opening-prob 0.75 # Play opening with high temperature policy in
# this probability.
--random-fastsearch-prob 0.75 # Play random move for reduce-playouts in this
# probability.
--komi-stddev 2.5 # Apply the random komi in the self-play
# games.
--cpuct-init 0.5
--lcb-reduction 0
--score-utility-factor 0.05
--selfplay-query bkp:9:7:0.9 # The format is boardsize-komi-prob.
# It means 90% games are 9x9 with komi 7.
--selfplay-query bhp:9:2:0.1 # The format is boardsize-handicap-prob.
# It means 10% 9x9 games is handicap. The
# max handicap size is 2.
--selfplay-query srs:area # The self-play game will use area scoring (chinese-like) and
--selfplay-query srs:territory # territory scoring (japanese-like).
--selfplay-query bkp:7:9:0.1 # It means 10% games are 7x7 with komi 9.
--playouts 150 # The main playouts.
--gumbel # Enable Sequential Halving with Gumbel.
--gumbel-playouts-threshold 50 # Do the Gumbel search if the current playouts
# below this value.
--always-completed-q-policy
--reduce-playouts 50 # 75% uses 50 playouts. Will disable any noise
# do not record the training data.
--reduce-playouts-prob 0.75 # 75% uses reduce-playouts.
--resign-playouts 75 # Use this playout if someone's winrate below
# threshold.
--resign-threshold 0.05 # If someone's winrate is below this value,
# will use the resign-playouts.
--resign-discard-prob 0.8 # Discard the training data in 80% when someome
# has already won the game.
--parallel-games 128 # Parallel games at the same time.
--batch-size 64 # Network evalutaion batch size.
--cache-memory-mib 400
--early-symm-cache
--first-pass-bonus
--root-policy-temp 1.1 # The policy softmax temperature of root node.
--num-games 5000 # Self-play games per epoch.