def records_sample(model_num, verificate=False):
db_path = utils.PAI_DB_PATH if utils.USE_PAI else utils.DB_PATH
if not verificate:
new_records_pattern = os.path.join(db_path,
'game-{}*'.format(model_num))
# all_records_pattern = os.path.join(db_path, 'game*')
while True:
new_records = utils.pai_find_path(new_records_pattern)
if len(new_records) >= utils.TRAIN_EPOCH_GAME_NUM * 2 + 10:
# break
return new_records
else:
time.sleep(60)
# all_records = utils.pai_find_path(all_records_pattern)
# old_records = list(set(all_records) - set(new_records))
# try:
# old_records = list(np.random.choice(old_records, utils.TRAIN_SAMPLE_NUM))
# raise Exception('eee')
# return old_records + new_records
# except:
# return new_records
else:
records_pattern = os.path.join(
db_path, 'game-verification-{}*'.format(model_num))
while True:
records = utils.pai_find_path(records_pattern)
if len(records) >= utils.VERIFICATION_GAME_NUM * 2:
return records
else:
time.sleep(60)
def save_history_to_tfrecord(self, reward):
if utils.SAVE_RECORD:
net_model_num = self.mct.net.get_model_num()
verification_pattern = os.path.join(
utils.PAI_DB_PATH if utils.USE_PAI else utils.DB_PATH,
'game-verification-{}*'.format(net_model_num))
if len(utils.pai_find_path(
verification_pattern)) < 2 * utils.VERIFICATION_GAME_NUM:
tfr_name = 'game-verification-{}-{}-{}.tfrecord'.format(
net_model_num, time.time(), self.color_str)
else:
tfr_name = 'game-{}-{}-{}.tfrecord'.format(
net_model_num, time.time(), self.color_str)
tfr_path = os.path.join(
utils.PAI_DB_PATH if utils.USE_PAI else utils.DB_PATH,
tfr_name)
tfr_writer = generate_writer(tfr_path)
for i, expect in enumerate(self.prob_history):
feature = self.game.board.get_feature(self.color, i)
example = generate_example(feature, expect, reward)
tfr_writer.write(example.SerializeToString())
# for i, base_expect in enumerate(self.prob_history):
# base_feature = self.game.board.get_feature(self.color, i)
# base_expect = np.reshape(base_expect, (self.size, self.size))
# for rot_num in range(8):
# rot = self.mct.net.rot[rot_num]
# feature = rot(base_feature, (1, 2))
# expect = rot(base_expect).reshape(self.game.board.full_size)
# example = generate_example(feature, expect, reward)
# tfr_writer.write(example.SerializeToString())
tfr_writer.close()
def compare(compare_model_num=None, default_model_num=None):
if utils.USE_PAI:
if compare_model_num is None:
compare_model_num = utils.pai_read_best()
if default_model_num is None:
new_compare = utils.pai_find_path(
os.path.join(utils.PAI_RECORD_PATH, 'compare-{}'.format(compare_model_num))
)
if new_compare:
default_model_num = int(new_compare[0].split('-')[-1])
else:
default_model_num = utils.pai_read_best('compare')
if default_model_num != compare_model_num:
win, total = utils.pai_read_compare_record(default_model_num, compare_model_num)
else:
total = 10000
if total > utils.COMPARE_TIME:
time.sleep(120)
return
game = Game(utils.MCTS, utils.MCTS, utils.SIZE, compare_model_num, default_model_num)
game.logger.info('Compare model {} with default model {}'.format(compare_model_num, default_model_num))
while True:
win, total = utils.pai_read_compare_record(default_model_num, compare_model_num)
game.logger.info('Now compare result: {}-{}'.format(win, total))
if total > utils.COMPARE_TIME:
break
black_as_best = np.random.choice([True, False])
if black_as_best:
game.logger.info('Black as best')
game.black_player.mct.reset_net(default_model_num)
game.white_player.mct.reset_net(compare_model_num)
else:
game.logger.info('White as best')
game.black_player.mct.reset_net(compare_model_num)
game.white_player.mct.reset_net(default_model_num)
game.start()
winner = game.board.winner
if winner is utils.EMPTY:
pass
elif (winner is utils.BLACK and black_as_best) or (winner is utils.WHITE and not black_as_best):
utils.pai_write_compare_record(default_model_num, compare_model_num, False)
else:
utils.pai_write_compare_record(default_model_num, compare_model_num, True)
game.reset()
if win / total > utils.COMPARE_WIN_RATE:
utils.pai_change_best(compare_model_num, 'compare')
game.logger.info('Change best model to {}'.format(compare_model_num))
else:
utils.pai_change_best(default_model_num, 'compare')
game.logger.info('Best model does not change')
def main(_):
pai_constant_init()
records = utils.pai_find_path(
os.path.join(utils.PAI_RECORD_PATH, '*x*-*.psq'))
game = Game(utils.TRANS, utils.TRANS)
for record in records:
game.black_player.add_record(record)
game.white_player.add_record(record)
game.start()
game.reset()
def main():
if utils.USE_PAI:
try:
model_num = utils.pai_read_best()
except:
model_num = 0
utils.pai_change_best(model_num)
db_pattern = os.path.join(utils.PAI_DB_PATH, 'game-{}-*'.format(model_num))
if len(utils.pai_find_path(db_pattern)) / 2 >= utils.TRAIN_EPOCH_GAME_NUM:
pass
else:
game = Game(utils.MCTS, utils.MCTS, black_net_model_num=model_num, white_net_model_num=model_num)
while len(utils.pai_find_path(db_pattern)) / 2 < utils.TRAIN_EPOCH_GAME_NUM:
game.logger.info('There are {} records now'.format(len(utils.pai_find_path(db_pattern))/ 2))
game.start()
utils.pai_win_rate_record(model_num, game.board.winner)
game.reset()
del game
time.sleep(120)
else:
game = Game()
game.start()
def save_history_to_tfrecord(self, reward):
if utils.SAVE_RECORD:
net_model_num = 0
verification_pattern = os.path.join(
utils.PAI_DB_PATH if utils.USE_PAI else utils.DB_PATH,
'game-verification-{}*'.format(net_model_num))
if len(utils.pai_find_path(
verification_pattern)) < 2 * utils.VERIFICATION_GAME_NUM:
tfr_name = 'game-verification-{}-{}-{}.tfrecord'.format(
net_model_num, time.time(), self.color_str)
else:
tfr_name = 'game-{}-{}-{}.tfrecord'.format(
net_model_num, time.time(), self.color_str)
tfr_path = os.path.join(
utils.PAI_DB_PATH if utils.USE_PAI else utils.DB_PATH,
tfr_name)
tfr_writer = generate_writer(tfr_path)
for i, expect in enumerate(self.prob_history):
feature = self.game.board.get_feature(self.color, i)
example = generate_example(feature, expect, reward)
tfr_writer.write(example.SerializeToString())
tfr_writer.close()