def eval(cfg, saved_model_path=SAVED_MODEL_PATH):
print('start to eval ! \n')
env = NormalizedActions(gym.make("Pendulum-v0"))
n_states = env.observation_space.shape[0]
n_actions = env.action_space.shape[0]
agent = DDPG(n_states,
n_actions,
critic_lr=1e-3,
actor_lr=1e-4,
gamma=0.99,
soft_tau=1e-2,
memory_capacity=100000,
batch_size=128)
agent.load_model(saved_model_path + 'checkpoint.pth')
rewards = []
moving_average_rewards = []
ep_steps = []
log_dir = os.path.split(
os.path.abspath(__file__))[0] + "/logs/eval/" + SEQUENCE
writer = SummaryWriter(log_dir)
for i_episode in range(1, cfg.eval_eps + 1):
state = env.reset() # reset环境状态
ep_reward = 0
for i_step in range(1, cfg.eval_steps + 1):
action = agent.select_action(state) # 根据当前环境state选择action
next_state, reward, done, _ = env.step(action) # 更新环境参数
ep_reward += reward
state = next_state # 跳转到下一个状态
if done:
break
print('Episode:', i_episode, ' Reward: %i' % int(ep_reward),
'n_steps:', i_step, 'done: ', done)
ep_steps.append(i_step)
rewards.append(ep_reward)
# 计算滑动窗口的reward
if i_episode == 1:
moving_average_rewards.append(ep_reward)
else:
moving_average_rewards.append(0.9 * moving_average_rewards[-1] +
0.1 * ep_reward)
writer.add_scalars('rewards', {
'raw': rewards[-1],
'moving_average': moving_average_rewards[-1]
}, i_episode)
writer.add_scalar('steps_of_each_episode', ep_steps[-1], i_episode)
writer.close()
'''存储reward等相关结果'''
if not os.path.exists(RESULT_PATH): # 检测是否存在文件夹
os.mkdir(RESULT_PATH)
np.save(RESULT_PATH + 'rewards_eval.npy', rewards)
np.save(RESULT_PATH + 'moving_average_rewards_eval.npy',
moving_average_rewards)
np.save(RESULT_PATH + 'steps_eval.npy', ep_steps)
import gym
from agent import DDPG
env = gym.make('Pendulum-v0')
agent = DDPG(env)
agent.load_model()
state = env.reset()
cumulative_reward = 0
for i in range(200):
action = agent.get_action(state)
env.render()
state, reward, _, _ = env.step(action * 2)
cumulative_reward += reward
print('Cumulative Reward: {}'.format(cumulative_reward))