def start(self):
"""
Do self play and write the data to the appropriate file.
"""
self.buffer = []
futures = deque()
with ProcessPoolExecutor(
max_workers=self.config.play.max_processes) as executor:
for game_idx in range(self.config.play.max_processes * 2):
futures.append(
executor.submit(self_play_buffer,
self.config,
cur=self.cur_pipes))
game_idx = 0
while True:
game_idx += 1
start_time = time()
env, data = futures.popleft().result()
print(f"game {game_idx:3} time={time() - start_time:5.1f}s "
f"halfmoves={env.num_halfmoves:3} {env.winner:12} "
f"{'by resign ' if env.resigned else ' '}")
pretty_print(env, ("current_model", "current_model"))
self.buffer += data
if (game_idx % self.config.play_data.nb_game_in_file) == 0:
self.flush_buffer()
reload_best_model_weight_if_changed(self.current_model)
futures.append(
executor.submit(self_play_buffer,
self.config,
cur=self.cur_pipes)) # Keep it going
if len(data) > 0:
self.flush_buffer()
def evaluate_model(self, ng_model):
"""
Given a model, evaluates it by playing a bunch of games against the current model.
:param ChessModel ng_model: model to evaluate
:return: true iff this model is better than the current_model
"""
ng_pipes = self.m.list([
ng_model.get_pipes(self.play_config.search_threads)
for _ in range(self.play_config.max_processes)
])
futures = []
with ProcessPoolExecutor(
max_workers=self.play_config.max_processes) as executor:
for game_idx in range(self.config.eval.game_num):
fut = executor.submit(play_game,
self.config,
cur=self.cur_pipes,
ng=ng_pipes,
current_white=(game_idx % 2 == 0))
futures.append(fut)
results = []
for fut in as_completed(futures):
# ng_score := if ng_model win -> 1, lose -> 0, draw -> 0.5
ng_score, env, current_white = fut.result()
results.append(ng_score)
win_rate = sum(results) / len(results)
game_idx = len(results)
logger.debug(
f"game {game_idx:3}: ng_score={ng_score:.1f} as {'black' if current_white else 'white'} "
f"{'by resign ' if env.resigned else ' '}"
f"win_rate={win_rate*100:5.1f}% "
f"{env.board.fen().split(' ')[0]}")
colors = ("current_model", "ng_model")
if not current_white:
colors = reversed(colors)
pretty_print(env, colors)
if len(results) - sum(results) >= self.config.eval.game_num * (
1 - self.config.eval.replace_rate):
logger.debug(
f"lose count reach {results.count(0)} so give up challenge"
)
return False
if sum(
results
) >= self.config.eval.game_num * self.config.eval.replace_rate:
logger.debug(
f"win count reach {results.count(1)} so change best model"
)
return True
win_rate = sum(results) / len(results)
logger.debug(f"winning rate {win_rate*100:.1f}%")
return win_rate >= self.config.eval.replace_rate
def evaluate_model(self, ng_model):
"""
Given a model, evaluates it by playing a bunch of games against the current model.
:param ChessModel ng_model: model to evaluate
:return: true iff this model is better than the current_model
"""
ng_pipes = self.m.list([ng_model.get_pipes(self.play_config.search_threads) for _ in range(self.play_config.max_processes)])
futures = []
with ProcessPoolExecutor(max_workers=self.play_config.max_processes) as executor:
for game_idx in range(self.config.eval.game_num):
fut = executor.submit(play_game, self.config, cur=self.cur_pipes, ng=ng_pipes, current_white=(game_idx % 2 == 0))
futures.append(fut)
results = []
for fut in as_completed(futures):
# ng_score := if ng_model win -> 1, lose -> 0, draw -> 0.5
ng_score, env, current_white = fut.result()
results.append(ng_score)
win_rate = sum(results) / len(results)
game_idx = len(results)
logger.debug(f"game {game_idx:3}: ng_score={ng_score:.1f} as {'black' if current_white else 'white'} "
f"{'by resign ' if env.resigned else ' '}"
f"win_rate={win_rate*100:5.1f}% "
f"{env.board.fen().split(' ')[0]}")
colors = ("current_model", "ng_model")
if not current_white:
colors = reversed(colors)
pretty_print(env, colors)
if len(results)-sum(results) >= self.config.eval.game_num * (1-self.config.eval.replace_rate):
logger.debug(f"lose count reach {results.count(0)} so give up challenge")
return False
if sum(results) >= self.config.eval.game_num * self.config.eval.replace_rate:
logger.debug(f"win count reach {results.count(1)} so change best model")
return True
win_rate = sum(results) / len(results)
logger.debug(f"winning rate {win_rate*100:.1f}%")
return win_rate >= self.config.eval.replace_rate
