def setup_dqn(self):
print_ts("Setup DQN of Move2BeaconAgent")
self.DQN = DQN_module(self.batch_size, self.gamma, self.history_length,
self.size_replaybuffer, self.optim_learning_rate,
self.dim_actions)
self.device = self.DQN.device
print_ts("DQN module has been initalized")
def setup_dqn(self):
# TODO(vloet): what to do with history length? unecessary imho.
self.history_length = 1
self.DQN = DQN_module(self.batch_size, self.gamma, self.history_length,
self.size_replaybuffer, self.optim_learning_rate,
self.dim_actions)
self.device = self.DQN.device
print_ts("DQN module has been initalized")
def update_target_network(self):
"""
Transferring the estimator weights to the target weights and resetting
the agent.
"""
if self.episodes % self.target_update_period == 0:
print_ts("About to update")
self.DQN.update_target_net()
def _setup_torch(self):
"""
Setting GPU if available. Else, use the CPU.
"""
# Initalizing
device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
torch.set_printoptions(linewidth=750, profile="full")
print_ts("Performing calculations on {}".format(device))
return device
def setup_dqn(self):
self.dim_actions = 2
self.smart_actions = SMART_ACTIONS
print_ts("Setup DQN of CartPoleAgent")
self.DQN = DQN_module(self.batch_size, self.gamma, self.history_length,
self.size_replaybuffer, self.optim_learning_rate,
self.dim_actions)
self.device = self.DQN.device
print_ts("DQN module has been initalized")
def save_model(self, exp_path, emergency=False):
if emergency:
print("KeyboardInterrupt detected, saving last model!")
save_path = exp_path + "/model/emergency_model.pt"
else:
save_path = exp_path + "/model/model.pt"
# Path((self.exp_path + "/model")).mkdir(parents=True, exist_ok=True)
self.DQN.save(save_path)
print_ts("Model has been saved at {}".format(save_path))
def setup_agent(agent_specs):
"""
Initializing right agent and agent_interface for the environment.
"""
agent_type = agent_specs["AGENT_TYPE"]
if agent_type == 'pendulum':
return PendulumAgent(agent_specs)
else:
print_ts("The agent type {} is not known".format(agent_type))
exit()
def change_cwd(self, change_path):
"""
Change cwd if directory exists.
"""
if not path.exists(change_path):
print_ts("Directory is unknown")
return
self.cwd = change_path
self.current_test_dir = change_path + "/test_reports"
self.current_train_dir = change_path + "/training_reports"
print_ts("Changed cwd to {}".format(self.cwd))
def setup_dqn(self):
"""
Setting up the DQN module with the grid action space.
"""
self._xy_pairs, self.dim_actions, self.smart_actions = \
self.discretize_xy_grid()
self.DQN = DQN_module(self.batch_size, self.gamma, self.history_length,
self.size_replaybuffer, self.optim_learning_rate,
self.dim_actions)
self.device = self.DQN.device
print_ts("DQN module has been initalized")
def setup_agent(agent_specs):
"""
Initializing right agent and agent_interface for the environment.
"""
agent_type = agent_specs["AGENT_TYPE"]
if agent_type == 'defeatroaches':
return DefeatRoachesAgent(agent_specs)
if agent_type == 'move2beacon':
return Move2BeaconAgent(agent_specs)
else:
print_ts("The agent type {} is not known".format(agent_type))
exit()
def create_experiment(self, exp_name):
"""
Creates the folder structure which is necessary to save files.
"""
# Create free path
free_exp_path, _ = self.find_free_path(self.exp_dir, exp_name)
# print(free_exp_path, exp_name)
# Create the experiment folders
creation_success = self.create_folders(free_exp_path)
if creation_success:
print_ts("Created experiment at path {}".format(free_exp_path))
self.change_cwd(free_exp_path)
else:
pass
def setup_dqn(self):
"""
Setting up the DQN module with the compass action space.
"""
self.smart_actions = SMART_ACTIONS
self.dim_actions = len(self.smart_actions)
self.DQN = DQN_module(self.batch_size,
self.gamma,
self.history_length,
self.size_replaybuffer,
self.optim_learning_rate,
self.dim_actions)
self.device = self.DQN.device
print_ts("DQN module has been initalized")
def setup_agent(agent_specs):
"""
Initializing right agent and agent_interface for the environment.
"""
agent_type = agent_specs["AGENT_TYPE"]
if agent_type == 'compass':
return CompassAgent(agent_specs)
if agent_type == 'grid':
return GridAgent(agent_specs)
if agent_type == 'original':
raise("This agent type has not been implemeted yet.")
exit()
else:
print_ts("The agent type {} is not known".format(agent_type))
exit()
def print_cwd(self):
"""
"""
print_ts(self.cwd)
def print_architecture(self):
print_ts("Network: \n{}".format(self.net))
print_ts("Optimizer: \n{}".format(self.optimizer))
print_ts("Target Network: \n{}".format(self.target_net))
