A simple neural network training template for board games
Put your dataset at the root dir as alldata.npz
alldata.npz should contain 4 matrices: bf,gf,vt,pt
bf: Board Features, Location-specific features. shape=[N,C1,H,W] eg. where are black stones and white stones
gf: Global Features. shape=[N,C2] eg. which color is the next player
vt: Value Target. shape=[N,3] probability of win/loss/draw for the next player
pt: Policy Target. shape=[N,H*W] probability of the next move
input_c in model.py should be C1+C2
C2 should not be zero. If you don't need global features, you can let C2=1 and fill it with zero
You should modify Board.nninput() in play.py based on the format of your dataset
Run train.py -h to see the usage
You can play with your trained model here.
You should implement your game rules and NNinput(bf and gf) in "Board" class
A very simple ResNet.
You should change input_c here based on your dataset
You can design other models in this file and add it to ModelDic
This script will concat bf and gf, and apply random symmetry.
Global variables