def check_result_and_save_network(pwins, nwins, draws, game, args, iter_num):
# set gpu
os.environ["CUDA_VISIBLE_DEVICES"] = args.setGPU
if pwins + nwins > 0 and float(nwins + (0.5 * draws)) / (
pwins + nwins + draws) < args.updateThreshold:
print('REJECTING NEW MODEL')
else:
print('ACCEPTING NEW MODEL')
net = nn(game)
net.load_checkpoint(folder=args.checkpoint, filename='train.pth.tar')
net.save_checkpoint(folder=args.checkpoint, filename='best.pth.tar')
net.save_checkpoint(folder=args.checkpoint,
filename='checkpoint_' + str(iter_num) +
'.pth.tar')
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
def async_self_play(game, args, iter_num, iterr):
import tensorflow as tf
#bar.suffix = "iter:{i}/{x} | Total: {total:} | ETA: {eta:}".format(i=iter_num + 1, x=iterr,
# 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
net = nn(game)
mcts = MCTS(game, net, args)
try:
net.load_checkpoint(folder=args.checkpoint, filename='best.pth.tar')
except:
pass
train_examples = []
board = game.getInitBoard()
cur_player = 1
episode_step = 0
while True:
episode_step += 1
canonical_board = game.getCanonicalForm(board, cur_player)
temp = int(episode_step < args.tempThreshold)
pi = mcts.get_action_prob(canonical_board, temp=temp)
sym = game.getSymmetries(canonical_board, pi)
for b, p in sym:
train_examples.append([b, cur_player, p, None])
action = np.random.choice(len(pi), p=pi)
board, cur_player = game.getNextState(board, cur_player, action)
r = game.getGameEnded(board, cur_player)
if r != 0:
return [(x[0], x[2], r * ((-1)**(x[1] != cur_player)))
for x in train_examples]
def async_train_network(game, args, trainhistory):
# set gpu
os.environ["CUDA_VISIBLE_DEVICES"] = args.setGPU
# create network for training
nnet = nn(game)
try:
nnet.load_checkpoint(folder=args.checkpoint, filename='best.pth.tar')
except:
pass
# ---load history file---
model_file = os.path.join(args.checkpoint, "trainhistory.pth.tar")
examples_file = model_file + ".examples"
if not os.path.isfile(examples_file):
print(examples_file)
else:
print("File with trainExamples found. Read it.")
with open(examples_file, "rb") as f:
for i in Unpickler(f).load():
trainhistory.append(i)
# ---delete if over limit---
if len(trainhistory) > args.numItersForTrainExamplesHistory:
print("len(trainExamplesHistory) =", len(trainhistory),
" => remove the oldest trainExamples")
del trainhistory[len(trainhistory) - 1]
# ---extend history---
train_examples = []
for e in trainhistory:
train_examples.extend(e)
# ---save history---
folder = args.checkpoint
if not os.path.exists(folder):
os.makedirs(folder)
filename = os.path.join(folder, 'trainhistory.pth.tar' + ".examples")
with open(filename, "wb+") as f:
Pickler(f).dump(trainhistory)
print('Train with {} examples'.format(len(train_examples)))
nnet.train(train_examples)
nnet.save_checkpoint(folder=args.checkpoint, filename='train.pth.tar')
'numIters': 1000,
'numEps': 100,
'tempThreshold': 15,
'updateThreshold': 0.6,
'maxlenOfQueue': 200000,
'numMCTSSims': 25,
'arenaCompare': 40,
'cpuct': 1,
'checkpoint': './temp/',
'load_model': True,
'load_folder_file': ('./temp/', 'best.pth.tar'),
'numItersForTrainExamplesHistory': 20,
})
g = Game(6)
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)
board = g.getInitBoard()
player = 1
while g.getGameEnded(board, player) == 0:
mcts = MCTS(g, nnet, args)
action = mcts.get_action_prob(g.getCanonicalForm(board, 1))
print(action)
board, player = g.getNextState(board, 1, np.argmax(action))
g.display(board)