def main():
import argparse
parser = argparse.ArgumentParser(
"Play the Gomoku Game!",
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-g',
'--n_games',
type=int,
default=100,
help='Number of games to play.')
args = parser.parse_args()
game = Gomoku(board_size=15, first_center=False)
import tarfile
with tarfile.open('input.tar.gz') as tar:
tar.extractall()
import construct_dnn
model = construct_dnn.construct_dnn()
model.load('tf_model/tf_model')
import player_A, player_B
player_A.tf_predict_u.model = player_B.tf_predict_u.model = model
player_A.initialize()
player_B.initialize()
p1 = Player('Black')
p1.strategy = player_A.strategy
p2 = Player('White')
p2.strategy = player_B.strategy
p1.strategy.learndata = pickle.load(open('black.learndata', 'rb'))
p2.strategy.learndata = pickle.load(open('white.learndata', 'rb'))
game.players = [p1, p2]
if args.n_games > 1:
game.fastmode = 2
def playone(i, game_output, winner_board):
game.reset()
winner = game.play()
winner_board[winner] += 1
game_output.write('Game %-4d: Winner is %s\n' % (i + 1, winner))
game_output.flush()
winner_board = collections.OrderedDict([(p.name, 0) for p in game.players])
winner_board["Draw"] = 0
with open('game_results.txt', 'w') as game_output:
for i_game in xrange(args.n_games):
playone(i_game, game_output, winner_board)
print("Name | Games Won")
for name, nwin in winner_board.items():
print("%-7s | %7d" % (name, nwin))
# save the learndata to disk
pickle.dump(p1.strategy.learndata, open('newblack.learndata', 'wb'))
pickle.dump(p2.strategy.learndata, open('newwhite.learndata', 'wb'))
print('%d black learndata and %d white learndata saved!' %
(len(p1.strategy.learndata), len(p2.strategy.learndata)))
with tarfile.open('output.tar.gz', 'w:gz') as tar:
tar.add('newblack.learndata')
tar.add('newwhite.learndata')
def main():
import argparse
parser = argparse.ArgumentParser("Play the Gomoku Game!", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-n', '--n_game', type=int, default=10, help='Play a number of games to gather statistics.')
args = parser.parse_args()
game = Gomoku(board_size=15, first_center=False)
import construct_dnn
model = construct_dnn.construct_dnn()
model.load('bench_model/tf_model')
#model.load('trained_model_001/trained_model_001')
#model.load('trained_model_002/trained_model_002')
#model.load('trained_model_003/trained_model_003')
#model.load('trained_model_007/trained_model_007')
import player_A
player_A.tf_predict_u.model = model
player_A.initialize()
p1 = Player('TF')
p1.strategy = player_A.strategy
player_A.estimate_level = 1
player_A.t_random = 0.01
import ai_bench
ai_bench.initialize()
p2 = Player('BenchAI')
p2.strategy = ai_bench.strategy
ai_bench.estimate_level = 4
ai_bench.t_random = 0.001
game.players = [p2, p1]
if args.n_game > 1:
game.fastmode = 2
else:
player_A.show_q = ai_bench.show_q = True
def playone(i, game_output, winner_board):
game.reset()
winner = game.play()
winner_board[winner] += 1
game_output.write('Game %-4d: Winner is %s\n'%(i+1, winner))
game_output.flush()
winner_board = collections.OrderedDict([(p.name, 0) for p in game.players])
winner_board["Draw"] = 0
with open('game_results.txt','w') as game_output:
for i_game in xrange(args.n_game):
playone(i_game, game_output, winner_board)
p1.strategy.learndata = dict()
print("Name | Games Won")
for name, nwin in winner_board.items():
print("%-7s | %7d"%(name, nwin))
def main():
import argparse
parser = argparse.ArgumentParser(
"Play the Gomoku Game!",
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-n',
'--n_train',
type=int,
default=10,
help='Play a number of games to gather statistics.')
parser.add_argument('-t',
'--train_step',
type=int,
default=100,
help='Train a new model after this number of games.')
parser.add_argument('-e',
'--n_epoch',
type=int,
default=100,
help="Number of epochs for each training model")
args = parser.parse_args()
game = Gomoku(board_size=15, first_center=False)
import construct_dnn
model = construct_dnn.construct_dnn()
model.load('initial_model/tf_model')
import player_A, player_B
player_A.tf_predict_u.model = player_B.tf_predict_u.model = model
player_A.initialize()
player_B.initialize()
p1 = Player('Black')
p1.strategy = player_A.strategy
p2 = Player('White')
p2.strategy = player_B.strategy
# set up linked learndata and cache (allow AI to look into opponent's data)
p1.strategy.learndata = p2.strategy.opponent_learndata = dict()
p2.strategy.learndata = p1.strategy.opponent_learndata = dict()
player_A.tf_predict_u.cache = player_B.tf_predict_u.cache = dict()
game.players = [p1, p2]
if args.train_step > 1:
game.fastmode = 2
else:
player_A.show_q = player_B.show_q = True
def playone(i, game_output, winner_board):
game.reset()
winner = game.play()
winner_board[winner] += 1
game_output.write('Game %-4d: Winner is %s\n' % (i + 1, winner))
game_output.flush()
print("Training the model for %d iterations." % args.n_train)
for i_train in xrange(args.n_train):
# check if the current model exists
model_name = "trained_model_%03d" % i_train
assert not os.path.exists(
model_name), "Current model %s already exists!" % model_name
# create and enter the model folder
os.mkdir(model_name)
os.chdir(model_name)
# play the games
print("Training model %s" % model_name)
winner_board = collections.OrderedDict([(p.name, 0)
for p in game.players])
winner_board["Draw"] = 0
with open('game_results.txt', 'w') as game_output:
for i_game in xrange(args.train_step):
playone(i_game, game_output, winner_board)
print("Name | Games Won")
for name, nwin in winner_board.items():
print("%-7s | %7d" % (name, nwin))
# collect training data
train_X, train_Y = prepare_train_data(p1.strategy.learndata,
p2.strategy.learndata)
# reset the learndata and cache, release memory
p1.strategy.learndata = p2.strategy.opponent_learndata = dict()
p2.strategy.learndata = p1.strategy.opponent_learndata = dict()
player_A.tf_predict_u.cache = player_B.tf_predict_u.cache = dict()
# fit the tf model
model.fit(train_X,
train_Y,
n_epoch=args.n_epoch,
validation_set=0.1,
shuffle=True,
show_metric=True)
model.save('tf_model')
print("Model %s saved!" % model_name)
os.chdir('..')
def main():
import argparse
parser = argparse.ArgumentParser(
"Play the Gomoku Game!",
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-n',
'--n_train',
type=int,
default=10,
help='Play a number of games to gather statistics.')
parser.add_argument('-t',
'--train_step',
type=int,
default=100,
help='Train a new model after this number of games.')
parser.add_argument('-e',
'--n_epoch',
type=int,
default=100,
help="Number of epochs for each training model")
parser.add_argument('-l', '--begin_lib', help='Begin board library file')
parser.add_argument('-p',
'--begin_lib_p',
type=float,
default=1.0,
help='Possibility of begin lib to be used')
parser.add_argument('--new',
action='store_true',
help='Start from new model')
args = parser.parse_args()
game = Gomoku(board_size=15, first_center=False)
import construct_dnn
model = construct_dnn.construct_dnn()
if not args.new:
model.load('initial_model/tf_model')
import player_A, player_B
player_A.tf_predict_u.model = player_B.tf_predict_u.model = model
player_A.initialize()
player_B.initialize()
p1 = Player('Black')
p1.strategy = player_A.strategy
p2 = Player('White')
p2.strategy = player_B.strategy
# set up linked learndata and cache (allow AI to look into opponent's data)
p1.strategy.learndata = p2.strategy.opponent_learndata = dict()
p2.strategy.learndata = p1.strategy.opponent_learndata = dict()
player_A.tf_predict_u.cache = player_B.tf_predict_u.cache = dict()
game.players = [p1, p2]
if args.train_step > 1:
game.fastmode = 2
else:
player_A.show_q = player_B.show_q = True
allstones = set([(r, c) for r in range(1, 16) for c in range(1, 16)])
if args.begin_lib != None:
begin_lib = __import__(args.begin_lib).begin_lib
else:
begin_lib = None
def playone(i, game_output, winner_board):
game.reset()
player_A.reset()
player_B.reset()
if random.random() < args.begin_lib_p:
game.board = gen_begin_board(allstones, begin_lib)
else:
game.board = gen_begin_board(allstones, None)
# randomly assign a black stone to be the last move
game.last_move = next(iter(game.board[0]))
winner = game.play()
winner_board[winner] += 1
game_output.write('Game %-4d: Winner is %s\n' % (i + 1, winner))
game_output.flush()
print("Training the model for %d iterations." % args.n_train)
for i_train in range(args.n_train):
# check if the current model exists
model_name = "trained_model_%03d" % i_train
assert not os.path.exists(
model_name), "Current model %s already exists!" % model_name
# create and enter the model folder
os.mkdir(model_name)
os.chdir(model_name)
# play the games
print("Training model %s" % model_name)
winner_board = collections.OrderedDict([(p.name, 0)
for p in game.players])
winner_board["Draw"] = 0
with open('game_results.txt', 'w') as game_output:
for i_game in range(args.train_step):
playone(i_game, game_output, winner_board)
print("Name | Games Won")
for name, nwin in winner_board.items():
print("%-7s | %7d" % (name, nwin))
# collect training data
train_X, train_Y = prepare_train_data(p1.strategy.learndata,
p2.strategy.learndata)
# reset the learndata and cache, release memory
p1.strategy.learndata = p2.strategy.opponent_learndata = dict()
p2.strategy.learndata = p1.strategy.opponent_learndata = dict()
player_A.tf_predict_u.cache = player_B.tf_predict_u.cache = dict()
# fit the tf model
#model.trainer.training_state.step = 0 # reset the training step so lr_decay is reset
model.fit(train_X,
train_Y,
n_epoch=args.n_epoch,
validation_set=0.1,
shuffle=True,
show_metric=True)
model.save('tf_model')
print("Model %s saved!" % model_name)
os.chdir('..')
def main():
import argparse
parser = argparse.ArgumentParser("Play the Gomoku Game!", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-n', '--n_train', type=int, default=10, help='Play a number of games to gather statistics.')
parser.add_argument('-t', '--train_step', type=int, default=100, help='Train a new model after this number of games.')
parser.add_argument('-e', '--n_epoch', type=int, default=100, help="Number of epochs for each training model")
parser.add_argument('-l', '--begin_lib', help='Begin board library file')
parser.add_argument('-p', '--begin_lib_p', type=float, default=1.0, help='Possibility of begin lib to be used')
parser.add_argument('--new', action='store_true', help='Start from new model')
args = parser.parse_args()
game = Gomoku(board_size=15, first_center=False)
from tf_model import get_new_model, load_existing_model, save_model
if args.new:
model = get_new_model()
else:
model = load_existing_model('initial_model/tf_model.h5')
import player_A, player_B
player_A.tf_predict_u.model = player_B.tf_predict_u.model = model
player_A.initialize()
player_B.initialize()
p1 = Player('Black')
p1.strategy = player_A.strategy
p2 = Player('White')
p2.strategy = player_B.strategy
# set up linked learndata and cache (allow AI to look into opponent's data)
p1.strategy.learndata = p2.strategy.opponent_learndata = dict()
p2.strategy.learndata = p1.strategy.opponent_learndata = dict()
player_A.tf_predict_u.cache = player_B.tf_predict_u.cache = dict()
game.players = [p1, p2]
if args.train_step > 1:
game.fastmode = 2
else:
player_A.show_q = player_B.show_q = True
allstones = set([(r,c) for r in range(1,16) for c in range(1,16)])
if args.begin_lib != None:
begin_lib = __import__(args.begin_lib).begin_lib
else:
begin_lib = None
def playone(i, game_output, winner_board, replay=False):
game.reset()
player_A.reset()
player_B.reset()
if replay:
game.board = copy.deepcopy(game.last_begin_board)
else:
if random.random() < args.begin_lib_p:
game.board = gen_begin_board(allstones, begin_lib)
else:
game.board = gen_begin_board(allstones, None)
# store the begin board
game.last_begin_board = copy.deepcopy(game.board)
# randomly assign a black stone to be the last move
game.last_move = next(iter(game.board[0]))
winner = game.play()
winner_board[winner] += 1
game_output.write('Game %-4d: Winner is %s\n'%(i+1, winner))
game_output.flush()
print("Training the model for %d iterations."%args.n_train)
last_i_train = -1
for i_train in range(args.n_train):
# check if the current model exists
model_name = "trained_model_%03d" % i_train
if os.path.exists(model_name):
print(f"Folder {model_name} exists")
last_i_train = i_train
else:
break
if last_i_train >= 0:
# try to load the last trained model
model_name = "trained_model_%03d" % last_i_train
model_fnm = os.path.join(model_name, 'tf_model.h5')
if os.path.exists(model_fnm):
model = load_existing_model(model_fnm)
print(f"Loaded trained model from {model_fnm}")
else:
# if last trained model not exist, load the previous model
if last_i_train > 0:
prev_model_name = f"trained_model_{last_i_train-1:03d}"
prev_model_fnm = os.path.join(prev_model_name, 'tf_model.h5')
model = load_existing_model(prev_model_fnm)
print(f"Loaded lastest model from {prev_model_fnm}")
# try to reuse data and start training
train_data_fnm = os.path.join(model_name, 'data.h5')
if os.path.exists(train_data_fnm):
train_X, train_Y, train_W = load_data_h5(train_data_fnm)
print(f"Training data loaded from {train_data_fnm}, start training")
model.fit(train_X, train_Y, epochs=args.n_epoch, validation_split=0.2, shuffle=True, sample_weight=train_W)
save_model(model, model_fnm)
print("Model %s saved!" % model_name)
else:
# delete this folder and start again
shutil.rmtree(model_name)
print(f"Deleting folder {model_name}")
last_i_train -= 1
for i_train in range(last_i_train+1, args.n_train):
model_name = "trained_model_%03d" % i_train
# create and enter the model folder
os.mkdir(model_name)
os.chdir(model_name)
# play the games
print("Training model %s" % model_name)
winner_board = collections.OrderedDict([(p.name, 0) for p in game.players])
winner_board["Draw"] = 0
with open('game_results.txt','w') as game_output:
replay_last_game = False
for i_game in range(args.train_step):
if replay_last_game:
print("Repeating the starting board of last game")
playone(i_game, game_output, winner_board, replay=replay_last_game)
replay_last_game = any(player.strategy.surprised for player in game.players)
print("Name | Games Won")
for name, nwin in winner_board.items():
print("%-7s | %7d"%(name, nwin))
# collect training data
train_X, train_Y, train_W = prepare_train_data(p1.strategy.learndata, p2.strategy.learndata)
# reset the learndata and cache, release memory
p1.strategy.learndata = p2.strategy.opponent_learndata = dict()
p2.strategy.learndata = p1.strategy.opponent_learndata = dict()
player_A.tf_predict_u.cache = player_B.tf_predict_u.cache = dict()
# fit the tf model
model.fit(train_X, train_Y, epochs=args.n_epoch, validation_split=0.2, shuffle=True, sample_weight=train_W)
save_model(model, 'tf_model.h5')
print("Model %s saved!" % model_name)
os.chdir('..')
s = line[4:].strip()
s = s.replace('-', '0')
s = s.replace('x', '1')
s = s.replace('o', '-1')
state.append(s.split())
print(state)
return np.array(state, dtype=np.int8)
state = read_state('state_debug.txt')
player_A.print_state(state)
last_move = (4, 11)
player_A.print_state(state)
import construct_dnn
model = construct_dnn.construct_dnn()
player_A.initialize()
player_A.strategy((({(1, 1)}, {}), (1, 1), 1, 15))
player_A.tf_predict_u.model = model
model.load('../../auto_playok_com/tf_model')
player_A.best_action_q(state, 12342, 180, last_move, -1, 1, 1, -1)
import IPython
IPython.embed()
#player_A.find_interesting_moves(state, 120, np.zeros((15,15)), 1, 50, True)
