def AsyncAgainst(nnet, game, args, iter_num):
os.environ["CUDA_VISIBLE_DEVICES"] = '3'
minimax = minimaxAI(game,depth=7)
local_args = dotdict({'numMCTSSims': 200, 'cpuct': 1.0})
mcts = MCTS(game, nnet, local_args, eval=True)
arena = Arena(lambda x: np.argmax(mcts.getActionProb(x, temp=0)),
minimax.get_move, game)
arena.displayBar = False
net_win, minimax_win, draws = arena.playGames(2)
return net_win, minimax_win, draws
def Async_Play(game, args, iter_num, bar):
bar.suffix = "iter:{i}/{x} | Total: {total:} | ETA: {eta:}".format(
i=iter_num + 1,
x=args.numPlayGames,
total=bar.elapsed_td,
eta=bar.eta_td)
bar.next()
# set gpu
if (args.multiGPU):
if (iter_num % 2 == 0):
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
else:
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
else:
os.environ["CUDA_VISIBLE_DEVICES"] = args.setGPU
# set gpu growth
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# create NN
model1 = NNet(game)
model2 = NNet(game)
# try load weight
try:
model1.load_checkpoint(folder=args.model1Folder,
filename=args.model1FileName)
except:
print("load model1 fail")
pass
try:
model2.load_checkpoint(folder=args.model2Folder,
filename=args.model2FileName)
except:
print("load model2 fail")
pass
# create MCTS
mcts1 = MCTS(game, model1, args)
mcts2 = MCTS(game, model2, args)
# each process play 2 games
arena = Arena(lambda x: np.argmax(mcts1.getActionProb(x, temp=0)),
lambda x: np.argmax(mcts2.getActionProb(x, temp=0)), game)
arena.displayBar = False
oneWon, twoWon, draws = arena.playGames(2)
return oneWon, twoWon, draws
def AsyncAgainst(game, args, iter_num, bar):
bar.suffix = "iter:{i}/{x} | Total: {total:} | ETA: {eta:}".format(
i=iter_num + 1,
x=args.numAgainstPlayProcess,
total=bar.elapsed_td,
eta=bar.eta_td)
bar.next()
#set gpu
if (args.multiGPU):
if (iter_num % 2 == 0):
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
else:
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
else:
os.environ["CUDA_VISIBLE_DEVICES"] = args.setGPU
#set gpu memory grow
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
#create nn and load
nnet = nn(game)
pnet = nn(game)
try:
nnet.load_checkpoint(folder=args.checkpoint, filename='train.pth.tar')
except:
print("load train model fail")
pass
try:
pnet.load_checkpoint(folder=args.checkpoint, filename='best.pth.tar')
except:
print("load old model fail")
pass
pmcts = MCTS(game, pnet, args)
nmcts = MCTS(game, nnet, args)
arena = Arena(lambda x: np.argmax(pmcts.getActionProb(x, temp=0)),
lambda x: np.argmax(nmcts.getActionProb(x, temp=0)), game)
arena.displayBar = True
# each against process play the number of numPerProcessAgainst games.
pwins, nwins, draws = arena.playGames(args.numPerProcessAgainst)
return pwins, nwins, draws
def AsyncAgainst(nnet, game, args, gameth):
logging.debug("play self test game " + str(gameth))
os.environ["CUDA_VISIBLE_DEVICES"] = args.setGPU
# create nn and load
minimax = minimaxAI(game)
local_args = dotdict({'numMCTSSims': 100, 'cpuct': 1.0})
# local_args.numMCTSSims = 100
# local_args.cpuct = 1
mcts = MCTS(game, nnet, local_args, eval=True)
arena = Arena(lambda x: np.argmax(mcts.getActionProb(x, temp=0)),
minimax.get_move, game)
arena.displayBar = False
net_win, minimax_win, draws = arena.playGames(2)
return net_win, minimax_win, draws
def async_against(game, args, iter_num):
import tensorflow as tf
#bar.suffix = "iter:{i}/{x} | Total: {total:} | ETA: {eta:}".format(i=iter_num + 1, x=args.arenaCompare,
# total=bar.elapsed_td, eta=bar.eta_td)
#bar.next()
# set gpu
if args.multiGPU:
if iter_num % 2 == 0:
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
else:
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
else:
os.environ["CUDA_VISIBLE_DEVICES"] = args.setGPU
# set gpu memory grow
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
_ = tf.Session(config=config)
# create nn and load
nnet = nn(game)
pnet = nn(game)
try:
nnet.load_checkpoint(folder=args.checkpoint, filename='train.pth.tar')
except:
print("load train model fail")
pass
try:
pnet.load_checkpoint(folder=args.checkpoint, filename='best.pth.tar')
except:
print("load old model fail")
pass
pmcts = MCTS(game, pnet, args)
nmcts = MCTS(game, nnet, args)
arena = Arena(lambda x: np.argmax(pmcts.get_action_prob(x, temp=0)),
lambda x: np.argmax(nmcts.get_action_prob(x, temp=0)), game)
arena.displayBar = False
pwins, nwins, draws = arena.playGames(2)
return pwins, nwins, draws
