def play_game(agent):
logger = Logger.get_instance()
player_color = True if random.random() >= 0.5 else False
logger.debug(f"Player is white: {player_color}")
gam = Game(player_color=player_color)
agent.color = player_color
if player_color is False:
# Make the oponent move
gam.move(agent.best_move(gam, real_game=True))
# Play until finish
while gam.get_result() is None:
start = timer()
bm, am = agent.best_move(gam,
real_game=False,
ai_move=True,
max_iters=900)
gam.move(bm) # Make our move
gam.move(am) # Make oponent move
end = timer()
elapsed = round(end - start, 2)
logger.debug(f"\tMade move: {bm}, took: {elapsed} secs")
logger.debug(gam.get_history())
return gam
def main():
parser = argparse.ArgumentParser(description="Plays some chess games,"
"stores the result and trains a model.")
parser.add_argument('stockfish_bin', metavar='stockbin',
help="Stockfish binary path")
parser.add_argument('data_path', metavar='datadir',
help="Path of .JSON dataset.")
parser.add_argument('--games', metavar='games', type=int,
default=10)
parser.add_argument('--depth', metavar='depth', type=int,
default=1, help="Stockfish tree depth.")
parser.add_argument('--random_depth',
action='store_true',
default=False,
help="Use normal distribution of depths with "
"mean --depth.")
parser.add_argument('--debug',
action='store_true',
default=False,
help="Log debug messages on screen. Default false.")
args = parser.parse_args()
logger = Logger.get_instance()
logger.set_level(1)
if args.debug:
logger.set_level(0)
gen_data(args.stockfish_bin, args.data_path, args.games)
def gen_data(stockfish_bin, save_path, num_games=100, workers=2,
random_dep=False):
logger = Logger.get_instance()
d = DatasetGame()
pbar = tqdm(total=num_games)
with ThreadPoolExecutor(max_workers=workers) as executor:
for _ in range(num_games):
executor.submit(play_game,
stockfish_bin=stockfish_bin,
dataset=d,
tqbar=pbar)
pbar.close()
logger.info("Saving dataset...")
d.save(save_path)
def main():
parser = argparse.ArgumentParser(description="Plays some chess games,"
"stores the result and trains a model.")
parser.add_argument('model_dir',
metavar='modeldir',
help="where to store (and load from)"
"the trained model and the logs")
parser.add_argument('--games', metavar='games', type=int, default=1)
parser.add_argument('--threads', metavar='threads', type=int, default=6)
parser.add_argument('--debug',
action='store_true',
default=False,
help="Log debug messages on screen. Default false.")
args = parser.parse_args()
logger = Logger.get_instance()
logger.set_level(1)
if args.debug:
logger.set_level(0)
model_path = get_model_path(args.model_dir)
endpoint = ('localhost', 9999)
worker = PredictWorker(model_path=model_path, endpoint=endpoint)
manager = multiprocessing.Manager()
return_dict = manager.dict()
for i in range(args.games):
worker.start()
logger.info(f"Game {i+1} of {args.games}")
proci = multiprocessing.Process(target=play_game_job,
args=(endpoint, return_dict,
args.threads))
proci.start()
proci.join()
logger.debug("Stopping worker")
worker.stop()
logger.info(f"\tTraining {i+1} of {args.games}")
proci = multiprocessing.Process(target=train_model_job,
args=(return_dict['game'], model_path,
args.model_dir))
proci.start()
proci.join()
def benchmark(model_dir, workers=1, games=10, stockfish_depth=10, log=False):
""" Plays N games and gets stats about the results.
Parameters:
model_dir: str. Directory where the neural net weights and training
logs will be saved.
workers: number of concurrent games (workers which will play the games)
"""
multiprocessing.set_start_method('spawn', force=True)
if log:
logger = Logger.get_instance()
logger.info(f"Setting up {workers} concurrent games.")
model_path = get_model_path(model_dir)
with ProcessPoolExecutor(workers, initializer=process_initializer)\
as executor:
results = []
for i in range(games):
results.append(
executor.submit(play_game_job,
*[i, model_path, stockfish_depth]))
results = [r.result() for r in results]
if log:
logger.debug("Calculating stats.")
won = [
1 if x['color'] is True and x['result'] == 1
or x['color'] is False and x['result'] == -1 else 0 # noqa:W503
for x in results
]
if log:
print("##################### SUMMARY ###################")
print(f"Games played: {games}")
print(f"Games won: {len([x for x in won if x == 1])}")
print(f"Games drawn: {len([x for x in results if x['result'] == 0])}")
print("#################################################")
return dict(played=games,
won=len([x for x in won if x == 1]),
drawn=len([x for x in results if x['result'] == 0]))
def play_game_job(id: int, model_path, stockfish_depth, log=False):
""" Plays a game and returns the result..
Parameters:
id: Play ID (i.e. worker ID).
model_path: path to the .h5 model. If it not exists, it will play with
a fresh one.
stockfish_depth: int. Difficulty of stockfish.
"""
logger = Logger.get_instance()
agent_is_white = True if random.random() <= .5 else False
chess_agent = Agent(color=agent_is_white)
game_env = GameStockfish(player_color=agent_is_white,
stockfish='../../res/stockfish-10-64',
stockfish_depth=stockfish_depth)
try:
chess_agent.load(model_path)
except OSError:
logger.error("Model not found. Exiting.")
return None
if log:
logger.info(f"Starting game {id}")
try:
timer_start = timer()
while game_env.get_result() is None:
agent_move = chess_agent.best_move(game_env, real_game=True)
game_env.move(agent_move)
timer_end = timer()
if log:
logger.info(f"Game {id} done. Result: {game_env.get_result()}. "
f"took {round(timer_end-timer_start, 2)} secs")
except Exception:
logger.error(traceback.format_exc())
res = {'color': agent_is_white, 'result': game_env.get_result()}
game_env.tearup()
return res
