def load_envs_and_config(model_file):
save_dict = torch.load(model_file)
config = save_dict['config']
config['device'] = 'cuda'
env_factory = EnvFactory(config=config)
reward_env = env_factory.generate_reward_env()
reward_env.load_state_dict(save_dict['model'])
real_env = env_factory.generate_real_env()
return reward_env, real_env, config
def load_envs_and_config(model_file):
save_dict = torch.load(model_file)
config = save_dict['config']
config['device'] = 'cpu'
config['envs']['CartPole-v0']['solved_reward'] = 100000 # something big enough to prevent early out triggering
env_factory = EnvFactory(config=config)
reward_env = env_factory.generate_reward_env()
reward_env.load_state_dict(save_dict['model'])
real_env = env_factory.generate_real_env()
return reward_env, real_env, config
def load_envs_and_config(model_file):
save_dict = torch.load(model_file)
config = save_dict['config']
if BREAK == 'solved':
config['envs']['Cliff']['solved_reward'] = -20 # something big enough to prevent early out triggering
else:
config['envs']['Cliff']['solved_reward'] = 100000 # something big enough to prevent early out triggering
env_factory = EnvFactory(config=config)
reward_env = env_factory.generate_reward_env()
reward_env.load_state_dict(save_dict['model'])
real_env = env_factory.generate_real_env()
return reward_env, real_env, config
def compute(self, working_dir, bohb_id, config_id, cso, budget, *args,
**kwargs):
with open(CONFIG_FILE, 'r') as stream:
default_config = yaml.safe_load(stream)
config = self.get_specific_config(cso, default_config, budget)
print('----------------------------')
print("START BOHB ITERATION")
print('CONFIG: ' + str(config))
print('CSO: ' + str(cso))
print('BUDGET: ' + str(budget))
print('----------------------------')
info = {}
# generate environment
env_fac = EnvFactory(config)
real_env = env_fac.generate_real_env()
reward_env = env_fac.generate_reward_env()
save_dict = torch.load(SAVE_FILE)
#config = save_dict['config']
reward_env.load_state_dict(save_dict['model'])
score = 0
for i in range(NUM_EVALS):
td3 = TD3(env=reward_env,
max_action=reward_env.get_max_action(),
config=config)
reward_list_train, _, _ = td3.train(reward_env, test_env=real_env)
reward_list_test, _, _ = td3.test(real_env)
avg_reward_test = statistics.mean(reward_list_test)
unsolved_weight = config["agents"]["gtn"]["unsolved_weight"]
score += len(reward_list_train) + max(
0, (real_env.get_solved_reward() -
avg_reward_test)) * unsolved_weight
score = score / NUM_EVALS
info['config'] = str(config)
print('----------------------------')
print('FINAL SCORE: ' + str(score))
print("END BOHB ITERATION")
print('----------------------------')
return {"loss": score, "info": info}
else:
self.eps *= self.eps_decay
self.eps = max(self.eps, self.eps_min)
if __name__ == "__main__":
with open("../default_config_gridworld.yaml", "r") as stream:
config = yaml.safe_load(stream)
print(config)
torch.set_num_threads(1)
# generate environment
env_fac = EnvFactory(config)
real_env = env_fac.generate_real_env()
# virtual_env = env_fac.generate_virtual_env()
reward_env = env_fac.generate_reward_env()
reward_list_len = []
for i in range(20):
ql = QL(env=real_env, config=config, count_based=False)
reward_list_train, episode_length_list_train, _ = ql.train(
env=real_env, test_env=real_env, time_remaining=5000)
reward_list_test, episode_length_list_test, _ = ql.test(
env=real_env, time_remaining=500)
reward_list_len.append(len(reward_list_train))
print(len(reward_list_train))
print(sum(episode_length_list_train))
import statistics
print(statistics.mean(reward_list_len))
