def __init__(self, config):
self.config = config
# Create session
self.session = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True)))
# Create networks
self.prior_network = PolicyNetwork(
scope=config.prior_network,
temperature=config.prior_temperature,
use_symmetry=config.use_symmetry)
self.rollout_network = PolicyNetwork(
scope=config.rollout_network,
temperature=config.rollout_temperature,
reuse=config.prior_network == config.rollout_network,
use_symmetry=config.use_symmetry)
self.value_network = ValueNetwork(
scope=config.value_network, use_symmetry=config.use_symmetry)
# Load networks from checkpoints
run_dir = util.run_directory(config)
util.restore_network_or_fail(self.session, run_dir, self.prior_network)
util.restore_network_or_fail(self.session, run_dir, self.rollout_network)
util.restore_network_or_fail(self.session, run_dir, self.value_network)
# Create queues
self.prior_queue = AllQueue()
self.rollout_queue = AllQueue(maxsize=16)
self.value_queue = AllQueue(maxsize=16)
self.new_game()
def create_config():
config = flags.FLAGS
config.game = '_'.join(
[g.lower() for g in re.findall('[A-Z]?[a-z]+', config.game)])
config.num_actions = Atari.num_actions(config)
config.frameskip = eval(str(config.frameskip))
config.input_shape = eval(str(config.input_shape))
config.exploration_frame_shape = eval(str(config.exploration_frame_shape))
config.reward_clipping = config.reward_clipping and not config.reward_scaling
config.run_dir = util.run_directory(config)
if not config.bootstrapped: config.num_bootstrap_heads = 1
if config. async is None:
config.num_threads = 1
def __init__(self, config):
self.config = config
self.run_dir = util.run_directory(config)
self.position_targets = PositionTargets(config, self.run_dir)
self.session = tf.Session(config=tf.ConfigProto(
gpu_options=tf.GPUOptions(allow_growth=True)))
self.value_network = ValueNetwork('value')
util.restore_or_initialize_network(self.session, self.run_dir,
self.value_network)
# Train ops
self.create_train_op(self.value_network)
self.writer = tf.summary.FileWriter(self.run_dir)
util.restore_or_initialize_scope(self.session, self.run_dir,
self.training_scope.name)
def __init__(self, config):
self.config = config
session = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True)))
self.random_player = RandomPlayer()
self.exploratory_network = PolicyNetwork(config.exploratory_network)
self.exploratory_player = PolicyPlayer(self.exploratory_network,
session)
self.playout_network = PolicyNetwork(
config.playout_network,
reuse=config.exploratory_network == config.playout_network)
self.playout_player = PolicyPlayer(self.playout_network, session)
self.run_dir = util.run_directory(config)
util.restore_network_or_fail(session, self.run_dir,
self.exploratory_network)
util.restore_network_or_fail(session, self.run_dir,
self.playout_network)
def __init__(self, config):
self.config = config
self.run_dir = util.run_directory(config)
self.session = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True)))
self.policy_network = PolicyNetwork('policy')
self.policy_player = PolicyPlayer(self.policy_network, self.session)
util.restore_or_initialize_network(self.session, self.run_dir,
self.policy_network)
# Train ops
self.create_train_op(self.policy_network)
self.writer = tf.summary.FileWriter(self.run_dir)
util.restore_or_initialize_scope(self.session, self.run_dir,
self.training_scope.name)
self.opponents = Opponents(
[RandomPlayer(),
RandomThreatPlayer(),
MaxThreatPlayer()])
self.opponents.restore_networks(self.session, self.run_dir)
