def arena_process(i):
g = Game(8)
nnet = nn(g)
nnet.load_model(filename=("model_auto_" + str(i + 1)))
nmcts = MCTS(g, nnet, args)
pnet = nn(g)
if i != 0:
pnet.load_model(filename=("model_auto_" + str(i)))
pmcts = MCTS(g, pnet, args)
def player1(x):
pi = pmcts.get_action_prob(x)
# display_pi(np.array(pi[:-1]).reshape((len(x), len(x))))
return np.random.choice(len(pi), p=pi)
def player2(x):
pi = nmcts.get_action_prob(x)
return np.random.choice(len(pi), p=pi)
arena = Arena(player1=lambda x: player1(x),
player2=lambda x: player2(x),
game=g,
display=display)
return arena.play_games(8)
def arena_process(self, r, old_model_file, new_model_file, verbose=False):
old_net = nn(self.game)
if len(old_model_file) > 1:
old_net.load_model(filename=old_model_file)
else:
print('random state')
# old_net.load_model(filename=old_model_file)
old_mcts = MCTS(self.game, old_net, self.args)
new_net = nn(self.game)
new_net.load_model(filename=new_model_file)
new_mcts = MCTS(self.game, new_net, self.args)
def old_player(x):
pi = old_mcts.get_action_prob(x, self.args['numMCTSSims'])
# display_pi(np.array(pi[:-1]).reshape((len(x), len(x))))
return np.random.choice(len(pi), p=pi)
def new_player(x):
pi = new_mcts.get_action_prob(x, int(self.args['numMCTSSims'] * 1))
return np.random.choice(len(pi), p=pi)
arena = Arena(player1=lambda x: old_player(x),
player2=lambda x: new_player(x),
game=self.game,
display=display)
return arena.play_games(r, verbose=verbose)
def mcts_test():
g = Game(15)
b = Board(15)
nnet = nn(g)
mcts = MCTS(g, nnet, args)
b.execute_move((4, 1), 1)
b.execute_move((3, 2), 1)
b.execute_move((2, 3), 1)
b.execute_move((5, 0), 1)
# b.execute_move((1, 4), 1)
b.execute_move((3, 3), -1)
b.execute_move((3, 4), -1)
b.execute_move((3, 5), -1)
b.execute_move((3, 6), -1)
# b.execute_move((3, 5), -1)
curPlayer = 1
canonicalBoard = g.get_canonical_form(np.array(b.pieces), curPlayer)
pi = mcts.get_action_prob(canonicalBoard)
display(canonicalBoard)
display_pi(np.array(pi[:-1]).reshape((len(canonicalBoard), len(canonicalBoard))))
def main():
# log.info('Loading %s...', Game.__name__)
print('Loading Inception TicTacToe')
# g = Game(6)
# log.info('Loading %s...', nn.__name__)
nnet = nn()
if args['load_model']:
# log.info('Loading checkpoint "%s/%s"...', args.load_folder_file)
print('Loading checkpoint: ', args['load_folder_file'])
nnet.load_checkpoint(args['load_folder_file'][0],
args['load_folder_file'][1])
else:
# log.warning('Not loading a checkpoint!')
print('Not loading a checkpoint')
# log.info('Loading the Coach...')
print('Loading the Coach')
c = Coach(g, nnet, args)
if args['load_model']:
# log.info("Loading 'trainExamples' from file...")
print('Loading \'trainExamples\' from file...')
c.loadTrainExamples()
# log.info('Starting the learning process 🎉')
print('Starting the learning process 🎉')
c.learn()
def __init__(self, player_num):
super(NNetPlayer, self).__init__(Player.NNET, player_num)
# TODO: edit this to load the best weights
self.nnet = nn()
self.move_to_index = pickle.load(open("AllPossibleActionDict.p", "rb"))
self.move_array = pickle.load(open("AllPossibleActionVector.p", "rb"))
def generate_data_debug(self, model_file):
nnet = nn(self.game)
if len(model_file) > 1:
nnet.load_model(filename=model_file)
else:
print('random state')
c = Coach(self.game, nnet, self.args)
train_example = c.execute_episode()
l_sum_up = [(np.sum(i[0]), i[2]) for i in train_example]
print(sum(i == (0, -1) or i == (-1, 1)
for i in l_sum_up)) # second hand win
print(sum(i == (0, 1) or i == (-1, -1)
for i in l_sum_up)) # first hand win
def generate_data(l):
g = Game(args.goBang_n)
nnet = nn(g)
c = Coach(g, nnet, args)
train_example = c.execute_episode()
l.acquire()
try:
folder = args.checkpoint
if not os.path.exists(folder):
os.makedirs(folder)
filename = os.path.join(folder + "train_examples_4")
with open(filename, "ab+") as f:
pickle.dump(train_example, f)
finally:
l.release()
def parallelEpisode(argsIteration):
iteration = argsIteration[0]
args = argsIteration[1]
move_to_index = argsIteration[2]
print("EXECUTING EPISODE" + str(iteration))
trainExamples = []
deck = Deck()
players = [MCTSNNPlayer(1, args['num_simulations']), \
MCTSNNPlayer(2, args['num_simulations']), MCTSNNPlayer(3, args['num_simulations'])]
board = Board(players, True)
winners = None
num_winners = 0
nnet = nn()
nnet.load_checkpoint(args['load_folder_file'][0], args['load_folder_file'][1])
while True:
AI = MCTSNN(board, args['num_simulations'], deck, \
board.active_player.player_num, nnet, move_to_index)
pi = AI.getActionProb(temp=1)
action = np.random.choice(len(pi), p=pi)
canonicalBoard = AI.canonicalBoard
trainExamples.append([canonicalBoard, \
board.active_player.player_num, pi])
move = StateToFeatures.action_to_move(action, board.active_player.move_array, \
board.active_player, len(board.players), deck)
board.active_player.make_move(move, board, deck, players)
if board.active_player.calculate_vp() >= settings.POINTS_TO_WIN:
winners = set()
winners.add(board.active_player.player_num)
num_winners = 1
if board.round_num >= args['round_threshold']:
vps = [player.calculate_vp() for player in players]
most = max(vps)
winners = set()
for i in range(len(vps)):
if vps[i] == most:
winners.add(i + 1)
num_winners += 1
if winners:
train = [[x[0], x[2], \
(-1) ** int(x[1] not in winners) / (1 if x[1] not in winners else num_winners)] \
for x in trainExamples]
saveEpisodeTrainExamples(args, iteration, train)
return train
def generate_data(l, model_iter):
g = Game(8)
nnet = nn(g)
nnet.load_model(filename=("model_auto_" + str(model_iter + 1)))
c = Coach(g, nnet, args)
train_example = c.execute_episode()
l.acquire()
try:
folder = args.checkpoint
if not os.path.exists(folder):
os.makedirs(folder)
filename = os.path.join(folder +
("train_examples_auto_" + str(model_iter + 1)))
with open(filename, "ab+") as f:
pickle.dump(train_example, f)
finally:
l.release()
def generate_data(self, l, model_file, train_example_filename):
nnet = nn(self.game)
if len(model_file) > 2:
nnet.load_model(filename=model_file)
else:
print('random nn model')
c = Coach(self.game, nnet, self.args)
train_example = c.execute_episode()
l.acquire()
try:
folder = self.args['checkpoint']
if not os.path.exists(folder):
os.makedirs(folder)
filename = os.path.join(folder + train_example_filename)
with open(filename, "ab+") as f:
pickle.dump(train_example, f)
finally:
l.release()
def main():
log.info('Loading %s...', Game.__name__)
g = Game()
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()
for j in range(len(mat[0])):
if mat[i][j] == 0:
matnew[i][j][0] = 0
matnew[i][j][1] = -1
else:
matnew[i][j][0] = np.abs(mat[i][j])
matnew[i][j][1] = np.sign(mat[i][j])
g = Game(5)
# all players
rp = RandomPlayer(g).play
# nnet players
n1 = nn(g)
checkpoint = torch.load("best.pth.tar")
n1.nnet.load_state_dict(checkpoint['state_dict'])
args1 = dotdict({'numMCTSSims': 50, 'cpuct': 1.0})
mcts1 = MCTS(g, n1, args1)
n1p = lambda x: np.argmax(mcts1.getActionProb(x, temp=0))
n2 = nn(g)
checkpoint = torch.load("checkpoint_2.pth.tar")
n2.nnet.load_state_dict(checkpoint['state_dict'])
args2 = dotdict({'numMCTSSims': 50, 'cpuct': 1.0})
mcts2 = MCTSnn(g, n2, args2)
n2p = lambda x: np.argmax(mcts2.getActionProb(x, temp=0))
#n2p=lambda x: alpha.alphaZeroSearch()
def __init__(self, args):
#self.nnet = nnet
self.nnet = nn()
self.args = args
self.trainExamplesHistory = [] # history of examples from args.numItersForTrainExamplesHistory latest iterations
self.move_to_index = pickle.load(open("AllPossibleActionDict.p", "rb"))
if __name__ == "__main__":
#freeze_support()
# Start processes with lower priority to prevent system overload/hangs/freezes. Also set multiprocessing start method to spawn for Linux, since forking makes trouble
p = psutil.Process(os.getpid())
if sys.platform.startswith('win32'):
p.nice(psutil.BELOW_NORMAL_PRIORITY_CLASS)
elif sys.platform.startswith('linux'):
p.nice(5)
mp.set_start_method('spawn')
# Set number of threads for OpenMP in PyTorch
os.environ["OMP_NUM_THREADS"] = "1"
g = Game(is_basic=True)
# Suppress logging from fireplace
logger = logging.getLogger("fireplace")
logger.setLevel(logging.WARNING)
nnet = nn()
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()
with open(filename) as f:
data = json.load(f)
current_nn = None
for player in data["players"]:
if player["type"] == "human":
player_list.append(Human(index))
elif player["type"] == "random":
player_list.append(RandomPlayer(index))
elif player["type"] == "MCTS":
MCTS_player = MCTSPlayer(index, int(player["num_simulations"]))
player_list.append(MCTS_player)
elif player["type"] == "MCTSNN":
player_list.append(MCTSNNPlayer(index, int(player["num_simulations"])))
modelFile = os.path.join("trainExamplesMCTS/", "temp.pth.tar")
current_nn = nn()
if os.path.isfile(modelFile):
print("using saved weights!")
current_nn.nnet.model.load_weights(modelFile)
elif player["type"] == "NN":
player_list.append(NNetPlayer(index, int(player["num_simulations"])))
modelFile = os.path.join("trainExamplesMCTS/", "temp.pth.tar")
current_nn = nn()
if os.path.isfile(modelFile):
current_nn.nnet.model.load_weights(modelFile)
index += 1
record_file = int(data["record_data"])
if record_file:
fields=['Winner Type','Winner Num','First Player Type',
'First Player VP', 'Second Player Type', 'Second Player VP',
'tempThreshold': 15,
'updateThreshold': 0.6,
'maxlenOfQueue': 200000,
'numMCTSSims': 25,
'arenaCompare': 40, #
'cpuct': 1,
'checkpoint': './log',
'load_model': False,
'load_folder_file': ('./temp', 'best.pth.tar'),
'load_folder_examples': ('./temp', 'checkpoint_1.pth.tar'), ##
'save_log_dir': None,
'numItersForTrainExamplesHistory': 20,
'start_iter': 0, ##
'prevEloScore': [0, 0], ##
'print': True,
})
if __name__ == "__main__":
g = Game(6)
nnet = nn(g, args)
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()
'numEps': 10,
'tempThreshold': 100,
'updateThreshold': 0.6,
'maxlenOfQueue': 2000000,
'numMCTSSims': 1000,
'arenaCompare': 64,
'cpuct': 3,
'checkpoint': './temp/',
'load_model': False,
'load_folder_file': ('/dev/models/8x100x50', 'best.pth.tar'),
'numItersForTrainExamplesHistory': 20,
}
if __name__ == "x":
g = Game(15)
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.load_train_examples()
c.learn()
if __name__ == '__main__':
g = Game(15)
auto_run = AutoRun(g, args)
def __init__(self):
# self.trainExamplesHistory = [] # history of examples from args.numItersForTrainExamplesHistory latest iterations
self.nnet = nn()
