def __init__(self,
print_gametick=False,
print_on_receive=False,
create_server=True):
self.network = Network(create_server)
self.print_gametick = print_gametick
self.print_on_receive = print_on_receive
self.game_states = []
self.player_states = {}
def __init__(self,
model_path: Optional[str] = None,
min_wait_time: float = 0.0,
max_wait_time: float = 0.0,
verbose: bool = False,
gpus: Union[int, List[int]] = 0,
max_gpu_mem_fraction: float = 1.0):
super().__init__("LazyBug-Policy")
self.max_wait_time = max_wait_time
self.min_wait_time = min_wait_time
self.__network = Network(model_path=model_path,
gpus=gpus,
max_gpu_mem_fraction=max_gpu_mem_fraction)
# Perform first prediction to force keras to load the model
self.__network.evaluate_positions(Positions.create_empty(1))
self.verbose = verbose
def __init__(self,
model_path: str,
max_gpu_mem_fraction: Optional[float] = None):
self.__network = Network(model_path=model_path,
max_gpu_mem_fraction=max_gpu_mem_fraction)
handler.setFormatter(formatter)
LOGGER = logging.getLogger(__name__)
is_atty = os.isatty(sys.stdout.fileno())
if not is_atty:
root = logging.getLogger()
root.setLevel(logging.DEBUG)
root.handlers = []
root.addHandler(handler)
else:
LOGGER.addHandler(handler)
network = Network(nr_of_boards=2,
num_residual_blocks=args.residual_blocks,
optimizer=SGD(lr=0.001),
gpus=args.gpus,
verbose=is_atty,
value_weight=0.1,
policy_weight=0.9)
network.compile_model()
print(network.model.summary())
name_prefix = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
rel_model_dir = name_prefix
model_dir = os.path.join(args.model_path, rel_model_dir)
base_name = os.path.join(name_prefix + "_{:02d}_{:03d}.h5py")
os.makedirs(model_dir, exist_ok=False)
for e in range(10):
LOGGER.info("Main epoch {}".format(e))
if args.database_dir is not None:
data_generator = read_database(PositionDatabase(args.database_dir))
root.addHandler(handler)
else:
LOGGER.addHandler(handler)
name_prefix, model_dir, epoch, config = args.setup(args)
base_name = os.path.join(name_prefix + "_{:04d}.h5py")
tensorboard_dir = os.path.join(model_dir, "tensorboard")
os.makedirs(tensorboard_dir, exist_ok=True)
create_link(name_prefix, os.path.join(args.output_path, "latest"))
if epoch >= 1:
model_name = base_name.format(epoch - 1)
model_path = os.path.join(model_dir, model_name)
network = Network(model_path=model_path)
else:
network = Network(nr_of_boards=2,
num_residual_blocks=config["residual_blocks"],
optimizer=SGD(lr=0.001),
gpus=args.gpus,
verbose=is_atty,
value_weight=0.01,
policy_weight=0.99,
use_batch_normalization=True)
network.compile_model()
print(network.model.summary())
training_database = PositionDatabase(args.training_database_dir)
validation_database = PositionDatabase(args.validation_database_dir)
extra_databases = {
import numpy as np
from chess.variant import BughouseBoards
from engine.data import read_batches, transform_move_id_to_chessmove, PositionDatabase
from engine.data.position_sequence import PositionSequence
from engine.network import Network
MODEL_PATH = "../models/current"
network = Network(model_path=MODEL_PATH)
network.compile_model()
# TODO: this is inaccurate
dummy_boards = BughouseBoards()
database = PositionDatabase(
"/mnt/linux_data_slow/bughouse_data/databases/validation_full")
position_sequence = PositionSequence(database, batch_size=256 * 4)
out_list = []
for i, (x_data, y_data) in enumerate(position_sequence):
print("")
print("Batch {}".format(i))
num = len(x_data)
output = network.model.test_on_batch(x_data, y_data)
out_list.append(output)
for n, o in zip(network.model.metrics_names, output):
print("{}: {}".format(n, o))
probs, value = network.model.predict([x[0:1] for x in x_data])
probs = probs[0]
value = value[0][0]
print("Example: ")