def __init__(self, NetType='ResNet'):
self.game = game(BoardSize)
self.board = self.game.getInitBoard()
self.n = self.game.getBoardSize()[0]
self.players = [self.AlphaPlay, None, self.HumanPlay]
self.curPlayer = 1
self.gameStatus = 0
if NetType == 'ResNet':
self.AlphaNet = nn(self.game, t='RES')
self.AlphaNet.load_checkpoint(
'/home/zc1213/course/alphabackend/alphabrain/HistoryLog/Go/R_Ver2_checkpoint/{}/'
.format(BoardSize), 'best.pth.tar')
self.AlphaArgs = dotdict({'numMCTSSims': 2000, 'cpuct': 21.3})
self.AlphaMCTS = MCTS(self.game, self.AlphaNet, self.AlphaArgs)
self.Alpha = lambda x: np.argmax(
self.AlphaMCTS.getActionProb(x, temp=0))
else:
self.AlphaNet = nn(self.game, t='CNN')
self.AlphaNet.load_checkpoint(
'/home/zc1213/course/alphabackend/alphabrain/HistoryLog/Go/C_checkpoint/{}/'
.format(BoardSize), 'best.pth.tar')
self.AlphaArgs = dotdict({'numMCTSSims': 2000, 'cpuct': 17.3})
self.AlphaMCTS = MCTS(self.game, self.AlphaNet, self.AlphaArgs)
self.Alpha = lambda x: np.argmax(
self.AlphaMCTS.getActionProb(x, temp=0))
self.alphaMoveCache = {}
def tour():
pathAtt = './HistoryLog/Go/'
Rcand = {
'R1_10': [pathAtt + 'R_Ver1_checkpoint/7/', 'checkpoint_11.pth.tar'],
'R1_40': [pathAtt + 'R_Ver1_checkpoint/7/', 'checkpoint_47.pth.tar'],
'R1_B': [pathAtt + 'R_Ver1_checkpoint/7/', 'best.pth.tar'],
'R2_B': [pathAtt + 'R_Ver2_checkpoint/7/', 'best.pth.tar'],
'R3_B': [pathAtt + 'R_Ver3_checkpoint/7/', 'best.pth.tar']
}
Ccand = {
'C_10': [pathAtt + 'C_Ver1_checkpoint/7/', 'checkpoint_6.pth.tar'],
'C_40': [pathAtt + 'C_Ver1_checkpoint/7/', 'checkpoint_40.pth.tar'],
'C_B': [pathAtt + 'C_Ver1_checkpoint/7/', 'best.pth.tar']
}
compares = [('R1_10', 'C_10'), ('R1_40', 'C_40'), ('R1_B', 'C_B'),
('R2_B', 'C_B'), ('R3_B', 'C_B'), ('R1_B', 'R2_B'),
('R1_B', 'R3_B'), ('R2_B', 'R3_B')]
res = []
for c in [('R1_10', 'C_10')]:
print(c)
p1type = 'RES' if c[0][0] == 'R' else 'CNN'
p2type = 'RES' if c[1][0] == 'R' else 'CNN'
p1checkpoint = Rcand[c[0]] if c[0][0] == 'R' else Ccand[c[0]]
p2checkpoint = Rcand[c[1]] if c[1][0] == 'R' else Ccand[c[1]]
print(p1type, p2type)
print(p1checkpoint, p2checkpoint)
Net1 = nn(g, t=p1type)
Net1.load_checkpoint(p1checkpoint[0], p1checkpoint[1])
Args1 = dotdict({'numMCTSSims': 3000, 'cpuct': 17.5})
MCTS1 = MCTS(g, Net1, Args1)
Player1 = lambda x: np.argmax(MCTS1.getActionProb(x, temp=0))
Net2 = nn(g, t=p2type)
Net2.load_checkpoint(p2checkpoint[0], p2checkpoint[1])
Args2 = dotdict({
'numMCTSSims': 3000 if p2type == 'RNN' else 250,
'cpuct': 17.5 if p2type == 'RNN' else 3.0
})
MCTS2 = MCTS(g, Net2, Args2)
Player2 = lambda x: np.argmax(MCTS2.getActionProb(x, temp=0))
arena = Arena.Arena(Player1, Player2, g, display=display)
_res = arena.playGames(10, verbose=True)
res.append(_res)
result = {'1win': [], '2win': [], 'draw': []}
for r in res:
result['1win'].append(r[0])
result['2win'].append(r[1])
result['draw'].append(r[2])
pd.DataFrame(data=result).to_csv('reuslt.csv')
def main():
np.random.seed(123)
log.info('Loading %s...', Game.__name__)
g = Game(7)
log.info('Loading %s...', nn.__name__)
nnet = nn(g)
if args.load_model:
log.info('Loading checkpoint "%s/%s"...', args.load_folder_file)
nnet.load_checkpoint(args.load_folder_file[0],
args.load_folder_file[1])
else:
log.warning('Not loading a checkpoint!')
log.info('Loading the Coach...')
c = Coach(g, nnet, args)
if args.load_model:
log.info("Loading 'trainExamples' from file...")
c.loadTrainExamples()
log.info('Starting the learning process 🎉')
c.learn()
from go.pytorch.NNet import NNetWrapper as nn
args = dotdict({
'numIters': 10,
'numEps': 40,
'tempThreshold': 15,
'updateThreshold': 0.6,
'maxlenOfQueue': 200000,
'numMCTSSims': 10,
'arenaCompare': 20,
'cpuct': 1,
'checkpoint': './temp/',
'load_model': False,
'load_folder_file': ('/dev/models/8x100x50', 'best.pth.tar'),
'numItersForTrainExamplesHistory': 20,
})
if __name__ == "__main__":
g = GoGame(3)
nnet = nn(g)
if args.load_model:
nnet.load_checkpoint(args.load_folder_file[0],
args.load_folder_file[1])
c = Coach(g, nnet, args)
if args.load_model:
print("Load trainExamples from file")
c.loadTrainExamples()
c.learn()
'maxlenOfQueue': 200000,
'numMCTSSims': 200,
'arenaCompare': 50,
'cpuct': 3,
'checkpoint': './HistoryLog/Go/{}_checkpoint/{}/'.format(NetType+'_'+tag,BoardSize),
'load_model': True,
'load_folder_file': ('./HistoryLog/Go/{}_checkpoint/{}/'.format(NetType+'_'+tag,BoardSize),'best.pth.tar'),
'numItersForTrainExamplesHistory': 25,
'display':NO_DIS #True to display board, False to display progress bar
})
if __name__=="__main__":
g = Game(BoardSize)
nnet = nn(g,t='RES' if NetType=='R' else 'CNN')
logPath='./HistoryLog/Go/{}_Log/{}/'.format(NetType+'_'+tag,BoardSize)
try:
os.makedirs(logPath)
except:
pass
if args.load_model:
nnet.load_checkpoint(args.load_folder_file[0], args.load_folder_file[1])
c = Coach(g, nnet, args,log=True,logPath=logPath)
if args.load_model:
print("Load trainExamples from file")
c.loadTrainExamples()
c.learn()
"""
use this script to play any two agents against each other, or play manually with
any agent.
"""
BoardSize = 7
g = game(BoardSize)
# all players
rp = RandomPlayer(g).play
gp = GreedyGoPlayer(g).play
hp = HumanGoPlayer(g).play
# nnet players
NetType = 'CNN'
ResNet = nn(g, t='RES')
ResNet.load_checkpoint(
'./HistoryLog/Go/R_Ver2_checkpoint/{}/'.format(BoardSize),
'RVer2.best.pth.tar')
ResArgs = dotdict({'numMCTSSims': 3000, 'cpuct': 17.0})
ResMCTS = MCTS(g, ResNet, ResArgs)
ResPlayer = lambda x: np.argmax(ResMCTS.getActionProb(x, temp=0))
CNN = nn(g, t='CNN')
CNN.load_checkpoint('./HistoryLog/Go/C_checkpoint/{}/'.format(BoardSize),
'checkpoint_4.pth.tar')
CNNArgs = dotdict({'numMCTSSims': 250, 'cpuct': 3.0})
CNNMCTS = MCTS(g, CNN, CNNArgs)
CNNPlayer = lambda x: np.argmax(CNNMCTS.getActionProb(x, temp=0))
arena = Arena.Arena(ResPlayer, CNNPlayer, g, display=display)