def get_params(args):
params_manager = ParamsManager()
agent_params = params_manager.get_agent_params()
env_params = params_manager.get_env_params(args.env_params.lower())
env_params['env_name'] = args.env_name
custom_region_available = False
for key, value in env_params['useful_region'].items():
if key in env_params['env_name']:
env_params['useful_region'] = value
custom_region_available = True
break
if custom_region_available is not True:
env_params['useful_region'] = env_params['useful_region']['Default']
return agent_params, env_params
obs = next_obs
self.global_step_num += 1
if args.render:
self.env.render()
#print(self.actor_name + ":Episode#:", episode, "step#:", step_num, "\t rew=", reward, end="\r")
writer.add_scalar(self.actor_name + "/reward", reward,
self.global_step_num)
print(
"{}:Episode#:{} \t ep_reward:{} \t mean_ep_rew:{}\t best_ep_reward:{}"
.format(self.actor_name, episode, ep_reward,
np.mean(episode_rewards), self.best_reward))
writer.add_scalar(self.actor_name + "/ep_reward", ep_reward,
self.global_step_num)
if __name__ == "__main__":
agent_params = params_manager.get_agent_params()
agent_params["model_dir"] = args.model_dir
env_params = params_manager.get_env_params(
) # Used with Atari environments
env_params["env_name"] = args.env
mp.set_start_method('spawn') # Prevents RuntimeError during cuda init
agent_procs = [
DeepActorCriticAgent(id, args.env, agent_params, env_params)
for id in range(agent_params["num_agents"])
]
[p.start() for p in agent_procs]
[p.join() for p in agent_procs]
def load(self, env_name):
file_name = self.params['load_dir'] + "DQL_" + env_name + ".ptm"
agent_state = torch.load(file_name,
map_location=lambda storage, loc: storage)
self.Q.load_state_dict(agent_state["Q"])
self.Q.to(device)
self.best_mean_reward = agent_state["best_mean_reward"]
self.best_reward = agent_state["best_reward"]
print("Loaded Q model state from", file_name,
" which fetched a best mean reward of:", self.best_mean_reward,
" and an all time best reward of:", self.best_reward)
if __name__ == "__main__":
env_conf = params_manager.get_env_params()
env_conf["env_name"] = args.env_name
# If a custom useful_region configuration for this environment ID is available, use it if not use the Default
custom_region_available = False
for key, value in env_conf['useful_region'].items():
if key in args.env_name:
env_conf['useful_region'] = value
custom_region_available = True
break
if custom_region_available is not True:
env_conf['useful_region'] = env_conf['useful_region']['Default']
print("Using env_conf:", env_conf)
atari_env = False
for game in Atari.get_games_list():
if game in args.env_name.lower():
def load(self, env_name):
file_name = self.params['load_dir'] + "DQL_" + env_name + ".ptn"
agent_state = torch.load(file_name,
map_location=lambda storage, loc: storage)
self.Q.load_state.dict(agent_state["Q"])
self.Q.to(device)
self.best_reward_mean = agent_state["best_reward_mean"]
self.best_reward = agent_state["best_reward"]
print("El modelo se ha cargado desde: ", file_name,
". Que tiene una mejor recompensa media: ",
self.best_reward_mean, ". Recompensa máxima: ", self.best_reward)
if __name__ == "__main__":
env_conf = manager.get_env_params()
env_conf["env_game"] = args.env
if args.test:
env_conf["episodic_life"] = False
reward_type = "LIFE" if env_conf["episodic_life"] else "GAME"
custom_region_available = False
for key, value in env_conf["useful_region"].items():
if key in args.env:
env_conf["useful_region"] = value
custome_region_available = True
break
if custom_region_available is not True:
