def setUp(self):
# Remove the symmetries for reproductibility
import symmetry
symmetry.SYMMETRIES = symmetry.SYMMETRIES[0:1]
from random import seed
seed(0)
init_simulation_workers()
def main():
print("Starting run (v{})".format(__version__))
init_directories()
if conf['THREAD_SIMULATION']:
init_simulation_workers()
model_name = "model_1"
model = create_initial_model(name=model_name)
while True:
model = load_latest_model()
best_model = load_best_model()
train(model, game_model_name=best_model.name)
evaluate(best_model, model)
K.clear_session()
def test_save_sgf(self):
model = DummyModel()
init_simulation_workers()
mcts_simulations = 8 # mcts batch size is 8 and we need at least one batch
game_data = play_game(model,
model,
mcts_simulations,
conf['STOP_EXPLORATION'],
self_play=True,
num_moves=10)
save_game_sgf("test_model", 0, game_data)
destroy_simulation_workers()
os.remove("games/test_model/game_000.sgf")
os.removedirs("games/test_model")
def setUp(self):
# Remove the symmetries for reproductibility
import symmetry
symmetry.SYMMETRIES = symmetry.SYMMETRIES[0:1]
size = conf['SIZE']
tree = {
'count': 0,
'mean_value': 0,
'value': 0,
'parent': None,
'virtual_loss': 0,
'subtree': {
0: {
'count': 0,
'p': 1,
'value': 0,
'mean_value': 0,
'virtual_loss': 0,
'subtree': {}
},
1: {
'count': 0,
'p': 0,
'mean_value': 0,
'value': 0,
'virtual_loss': 0,
'subtree': {}
}
}
}
tree['subtree'][0]['parent'] = tree
tree['subtree'][1]['parent'] = tree
board, player = game_init()
model = DummyModel()
self.model = model
self.board = board
self.tree = tree
init_simulation_workers()
def run(self):
# set environment
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = str(self._gpuid)
logger.info('cuda_visible_device %s',
os.environ["CUDA_VISIBLE_DEVICES"])
if conf['THREAD_SIMULATION']:
init_simulation_workers()
# load models
latest_model, best_model = self.load_model()
while True:
if latest_model.name != best_model.name:
total = 0
wins = 0
desc = "Evaluation %s vs %s" % (latest_model.name,
best_model.name)
tq = tqdm(range(EVALUATE_N_GAMES), desc=desc)
for game in tq:
if self._one_game_only >= 0 and tq != game:
continue
directory = os.path.join(EVAL_DIR, latest_model.name,
"game_%03d" % game)
if os.path.isdir(directory):
continue
try:
os.makedirs(directory)
except Exception:
continue
start = datetime.datetime.now()
game_data = play_game(best_model,
latest_model,
MCTS_SIMULATIONS,
stop_exploration=0)
stop = datetime.datetime.now()
# Some statistics
winner_model = game_data['winner_model']
if winner_model == latest_model.name:
wins += 1
total += 1
moves = len(game_data['moves'])
new_desc = desc + " (winrate:%s%% %.2fs/move)" % (int(
wins / total * 100), (stop - start).seconds / moves)
tq.set_description(new_desc)
# save_game_data(best_model.name, game, game_data)
self.save_eval_game(latest_model.name, game, winner_model)
if self._one_game_only >= 0:
break
else:
logger.info("No new trained model")
if self._forever:
logger.info("Sleep for %s seconds", conf['SLEEP_SECONDS'])
time.sleep(conf['SLEEP_SECONDS'])
if not self._forever:
break
latest_model, best_model = self.load_model()
destroy_simulation_workers()