def start(self):
if self.model is None:
self.model = self.load_model()
self.buffer = []
idx = 1
mtcs_info = None
while True:
start_time = time()
if mtcs_info is None and self.config.play.share_mtcs_info_in_self_play:
mtcs_info = ReversiPlayer.create_mtcs_info()
env = self.start_game(idx, mtcs_info)
end_time = time()
logger.debug(
f"play game {idx} time={end_time - start_time} sec, "
f"turn={env.turn}:{env.board.number_of_black_and_white}:{env.winner}"
)
if self.config.play.use_newest_next_generation_model:
model_changed = reload_newest_next_generation_model_if_changed(
self.model, clear_session=True)
else:
model_changed = reload_best_model_weight_if_changed(
self.model, clear_session=True)
if model_changed:
mtcs_info = None
idx += 1
def start(self):
if self.model is None:
self.model = self.load_model()
self.buffer = []
idx = self.read_as_int(self.config.resource.self_play_game_idx_file) or 1
mtcs_info = None
while True:
start_time = time()
if mtcs_info is None and self.config.play.share_mtcs_info_in_self_play:
mtcs_info = ReversiPlayer.create_mtcs_info()
env = self.start_game(idx, mtcs_info)
end_time = time()
logger.debug(f"play game {idx} time={end_time - start_time} sec, "
f"turn={env.turn}:{env.board.number_of_black_and_white}:{env.winner}")
while True:
try:
if self.config.play.use_newest_next_generation_model:
reload_newest_next_generation_model_if_changed(self.model, clear_session=True)
else:
reload_best_model_weight_if_changed(self.model, clear_session=True)
break
except Exception as e:
logger.error(e)
if idx % self.config.play.reset_mtcs_info_per_game == 0:
logger.debug("reset MTCS info")
mtcs_info = None
idx += 1
with open(self.config.resource.self_play_game_idx_file, "wt") as f:
f.write(str(idx))
def _start(self):
logger.debug("SelfPlayWorker#start()")
np.random.seed(None)
worker_name = f"worker{self.worker_index:03d}"
self.tensor_board = TensorBoardLogger(
os.path.join(self.config.resource.self_play_log_dir, worker_name))
self.buffer = []
mtcs_info = None
local_idx = 0
while True:
np.random.seed(None)
local_idx += 1
game_idx = self.shared_var.game_idx
start_time = time()
if mtcs_info is None and self.config.play.share_mtcs_info_in_self_play:
mtcs_info = ReversiPlayer.create_mtcs_info()
# play game
env = self.start_game(local_idx, game_idx, mtcs_info)
game_idx = self.shared_var.incr_game_idx()
# just log
end_time = time()
time_spent = end_time - start_time
logger.debug(
f"play game {game_idx} time={time_spent} sec, "
f"turn={env.turn}:{env.board.number_of_black_and_white}:{env.winner}"
)
# log play info to tensor board
prefix = "self"
log_info = {
f"{prefix}/time": time_spent,
f"{prefix}/turn": env.turn
}
if mtcs_info:
log_info[f"{prefix}/mcts_buffer_size"] = len(mtcs_info.var_p)
self.tensor_board.log_scaler(log_info, game_idx)
# reset MCTS info per X games
if self.config.play.reset_mtcs_info_per_game and local_idx % self.config.play.reset_mtcs_info_per_game == 0:
logger.debug("reset MCTS info")
mtcs_info = None
with open(self.config.resource.self_play_game_idx_file, "wt") as f:
f.write(str(game_idx))
