# 创建策略网络
net = ActorCritic(input_dim=env.state_dims, output_dim=env.action_dims).to(device)
if len(glob.glob(os.path.join(ckpt_folder, '*.pt'))) > 0:
# 加载最后一个模型
last_ckpt = sorted(glob.glob(os.path.join(ckpt_folder, '*.pt')))[-1]
print("load checkpoint", last_ckpt)
checkpoint = torch.load(last_ckpt,map_location=device)
net.load_state_dict(checkpoint['model_G_state_dict'])
last_episode_id = checkpoint['episode_id']
REWARDS = checkpoint['REWARDS']
for episode_id in range(last_episode_id, max_m_episode):
# 循环一局训练
state = env.reset()
rewards, log_probs, values, masks = [], [], [], []
for step_id in range(max_steps):
# 根据策略网络计算动作和对数概率，并获取动作值
action, log_prob, value = net.get_action(state)
# 执行动作并获取下一个状态和奖励
state, reward, done = env.step(action)
rewards.append(reward)
log_probs.append(log_prob)
values.append(value)
masks.append(1-done)
if episode_id % 1000 == 1:
env.render()
if done or step_id == max_steps-1:
