def __init__(self, args, exp_model, logging_funcs):
self.args = args
# Exploration Model
self.exp_model = exp_model
self.log = logging_funcs["log"]
self.env = logging_funcs["env"]
self.replay = ExpReplay(args.exp_replay_size,
args.stale_limit,
exp_model,
args,
priority=self.args.prioritized)
self.bonus_replay = ExpReplay(args.bonus_replay_size,
args.stale_limit,
exp_model,
args,
priority=self.args.prioritized)
self.bonus_replay_stuff = 0
# DQN and Target DQN
model = get_models(args.model)
self.dqn = model(actions=args.actions)
self.target_dqn = model(actions=args.actions)
dqn_params = 0
for weight in self.dqn.parameters():
weight_params = 1
for s in weight.size():
weight_params *= s
dqn_params += weight_params
print("DQN has {:,} parameters.".format(dqn_params))
self.target_dqn.eval()
if args.gpu:
print("Moving models to GPU.")
self.dqn.cuda()
self.target_dqn.cuda()
# Optimizer
self.optimizer = RMSprop(self.dqn.parameters(), lr=args.lr)
# self.optimizer = Adam(self.dqn.parameters(), lr=args.lr)
self.T = 0
self.target_sync_T = -self.args.t_max
self.max_bonus = 0
self.player_bonus_positions = []
def __init__(self, args, exp_model, logging_func):
self.args = args
# Exploration Model
self.exp_model = exp_model
self.log = logging_func["log"]
# Experience Replay
self.replay = ExpReplay(args.exp_replay_size, args)
self.dnds = [
DND(kernel=kernel,
num_neighbors=args.nec_neighbours,
max_memory=args.dnd_size,
embedding_size=args.nec_embedding)
for _ in range(self.args.actions)
]
# DQN and Target DQN
model = get_models(args.model)
self.embedding = model(embedding=args.nec_embedding)
embedding_params = 0
for weight in self.embedding.parameters():
weight_params = 1
for s in weight.size():
weight_params *= s
embedding_params += weight_params
print(
"Embedding Network has {:,} parameters.".format(embedding_params))
if args.gpu:
print("Moving models to GPU.")
self.embedding.cuda()
# Optimizer
self.optimizer = RMSprop(self.embedding.parameters(), lr=args.lr)
# self.optimizer = Adam(self.embedding.parameters(), lr=args.lr)
self.T = 0
self.target_sync_T = -self.args.t_max
self.experiences = []
self.keys = []
self.q_val_estimates = []
self.table_updates = 0
def __init__(self, args, exp_model, logging_func):
self.args = args
# Exploration Model
self.exp_model = exp_model
self.log = logging_func["log"]
self.log_image = logging_func["image"]
os.makedirs("{}/transition_model".format(args.log_path))
# Experience Replay
self.replay = ExpReplay(args.exp_replay_size, args.stale_limit, exp_model, args, priority=self.args.prioritized)
# DQN and Target DQN
model = get_models(args.model)
print("\n\nDQN")
self.dqn = model(actions=args.actions)
print("Target DQN")
self.target_dqn = model(actions=args.actions)
dqn_params = 0
for weight in self.dqn.parameters():
weight_params = 1
for s in weight.size():
weight_params *= s
dqn_params += weight_params
print("Model DQN has {:,} parameters.".format(dqn_params))
self.target_dqn.eval()
if args.gpu:
print("Moving models to GPU.")
self.dqn.cuda()
self.target_dqn.cuda()
# Optimizer
# self.optimizer = Adam(self.dqn.parameters(), lr=args.lr)
self.optimizer = RMSprop(self.dqn.parameters(), lr=args.lr)
self.T = 0
self.target_sync_T = -self.args.t_max
# Action sequences
self.actions_to_take = []
def __init__(self, args, exp_model, logging_func):
self.args = args
# Exploration Model
self.exp_model = exp_model
self.log = logging_func["log"]
# Experience Replay
if self.args.set_replay:
self.replay = ExpReplaySet(10, 10, exp_model, args, priority=False)
else:
self.replay = ExpReplay(args.exp_replay_size,
args.stale_limit,
exp_model,
args,
priority=self.args.prioritized)
# DQN and Target DQN
model = get_models(args.model)
self.dqn = model(actions=args.actions)
self.target_dqn = model(actions=args.actions)
dqn_params = 0
for weight in self.dqn.parameters():
weight_params = 1
for s in weight.size():
weight_params *= s
dqn_params += weight_params
print("DQN has {:,} parameters.".format(dqn_params))
self.target_dqn.eval()
if args.gpu:
print("Moving models to GPU.")
self.dqn.cuda()
self.target_dqn.cuda()
# Optimizer
# self.optimizer = Adam(self.dqn.parameters(), lr=args.lr)
self.optimizer = RMSprop(self.dqn.parameters(), lr=args.lr)
self.T = 0
self.target_sync_T = -self.args.t_max
def __init__(self, args, exp_model, logging_func):
self.args = args
# Exploration Model
self.exp_model = exp_model
self.log = logging_func["log"]
# Experience Replay
if self.args.set_replay:
self.replay = ExpReplaySet(10, 10, exp_model, args, priority=False)
else:
self.replay = ExpReplay(args.exp_replay_size, args.stale_limit, exp_model, args, priority=self.args.prioritized)
# DQN and Target DQN
model = get_models(args.model)
self.dqn = model(actions=args.actions)
self.target_dqn = model(actions=args.actions)
dqn_params = 0
for weight in self.dqn.parameters():
weight_params = 1
for s in weight.size():
weight_params *= s
dqn_params += weight_params
print("DQN has {:,} parameters.".format(dqn_params))
self.target_dqn.eval()
if args.gpu:
print("Moving models to GPU.")
self.dqn.cuda()
self.target_dqn.cuda()
# Optimizer
# self.optimizer = Adam(self.dqn.parameters(), lr=args.lr)
self.optimizer = RMSprop(self.dqn.parameters(), lr=args.lr)
self.T = 0
self.target_sync_T = -self.args.t_max
def __init__(self, args, exp_model, logging_func):
self.args = args
# Exploration Model
self.exp_model = exp_model
self.log = logging_func["log"]
# Experience Replay
self.replay = ExpReplay(args.exp_replay_size, args)
self.dnds = [DND(kernel=kernel, num_neighbors=args.nec_neighbours, max_memory=args.dnd_size, embedding_size=args.nec_embedding) for _ in range(self.args.actions)]
# DQN and Target DQN
model = get_models(args.model)
self.embedding = model(embedding=args.nec_embedding)
embedding_params = 0
for weight in self.embedding.parameters():
weight_params = 1
for s in weight.size():
weight_params *= s
embedding_params += weight_params
print("Embedding Network has {:,} parameters.".format(embedding_params))
if args.gpu:
print("Moving models to GPU.")
self.embedding.cuda()
# Optimizer
self.optimizer = RMSprop(self.embedding.parameters(), lr=args.lr)
# self.optimizer = Adam(self.embedding.parameters(), lr=args.lr)
self.T = 0
self.target_sync_T = -self.args.t_max
self.experiences = []
self.keys = []
self.q_val_estimates = []
self.table_updates = 0
print(" {}: {}".format(arg, getattr(args, arg)))
print("=" * 40)
print()
# Experience Replay
replay = None
if args.prioritized:
replay = Prioritised_ExpReplay_Options(args.exp_replay_size)
else:
replay = ExperienceReplay_Options(args.exp_replay_size)
# DQN
print("\n" + "=" * 40)
print(16 * " " + "Models:" + " " * 16)
print("=" * 40)
dqn = get_models(args.model)(name="DQN")
print()
target_dqn = get_models(args.model)(name="Target_DQN")
print("=" * 40)
# Optimizer
optimiser = tf.train.AdamOptimizer(args.lr)
# Tensorflow Operations
with tf.name_scope("Sync_Target_DQN"):
dqn_vars = dqn.variables
target_dqn_vars = target_dqn.variables
sync_vars_list = []
for (ref, val) in zip(target_dqn_vars, dqn_vars):
sync_vars_list.append(tf.assign(ref, val))
sync_vars = tf.group(*sync_vars_list)
def __init__(self, args, exp_model, logging_funcs):
self.args = args
self.log = logging_funcs["log"]
self.log_image = logging_funcs["image"]
os.makedirs("{}/goal_states".format(args.log_path))
self.env = logging_funcs["env"]
# Exploration Model
self.exp_model = exp_model
# Experience Replay
self.replay = ExpReplay(args.exp_replay_size,
args.stale_limit,
exp_model,
args,
priority=self.args.prioritized)
# DQN and Target DQN
model = get_models(args.model)
self.model = model
self.dqn = model(actions=args.actions)
self.target_dqn = model(actions=args.actions)
dqn_params = 0
for weight in self.dqn.parameters():
weight_params = 1
for s in weight.size():
weight_params *= s
dqn_params += weight_params
print("DQN has {:,} parameters.".format(dqn_params))
self.target_dqn.eval()
if args.gpu:
print("Moving models to GPU.")
self.dqn.cuda()
self.target_dqn.cuda()
# Optimizer
self.optimizer = RMSprop(self.dqn.parameters(), lr=args.lr)
# self.optimizer = Adam(self.dqn.parameters(), lr=args.lr)
self.T = 0
self.target_sync_T = -self.args.t_max
# Hierarhcical stuff
self.goal_state_T = 0
self.goal_state = None
self.max_bonus = 0
self.goal_optimizer = None
self.goal_dqn = None
self.goal_dqns = []
self.goal_dqn_states = []
self.executing_option = False
self.option_num = 0
self.option_steps = 0
self.training_goal_T = 0