import os import pickle os.environ['OMP_WAIT_POLICY'] = 'PASSIVE' # 确保在pytorch前设置 from copy import deepcopy import pandas as pd import torch.nn.functional as F from models.env import WgzGym from models.net import ActorPPO, CriticAdv from models.tools import get_episode_return, test_one_episode def smooth_rewards(rewards, window=10): rewards = rewards.unsqueeze(0).unsqueeze(0) # 将 rewards 转为 [1, 1, len] 的形状以适应 conv1d kernel = torch.ones(1, 1, window, device=rewards.device) / window # 创建一个均匀的滑动平均核 smoothed_rewards = F.conv1d(rewards, kernel, padding='valid') # 滑动平均 smoothed_rewards = smoothed_rewards.squeeze(0).squeeze(0) # 去掉多余的维度 # 保持与原始奖励序列相同的长度，将前 window-1 个奖励保持不变 return torch.cat((rewards[0, 0, :window - 1], smoothed_rewards)) def update_buffer(_trajectory): _trajectory = list(map(list, zip(*_trajectory))) # 2D-list transpose, here cut the trajectory into 5 parts ten_state = torch.as_tensor(_trajectory[0]) # tensor state here ten_reward = torch.as_tensor(_trajectory[1], dtype=torch.float32) # _trajectory[2] = done, 将 done 替换为掩码，节省内存 ten_mask = (1.0 - torch.as_tensor(_trajectory[2], dtype=torch.float32)) * gamma ten_action = torch.as_tensor(_trajectory[3]) ten_noise = torch.as_tensor(_trajectory[4], dtype=torch.float32) buffer[:] = (ten_state, ten_action, ten_noise, ten_reward, ten_mask) # list store tensors _steps = ten_reward.shape[0] # steps collected in all trajectories _r_exp = ten_reward.mean() # the mean reward return _steps, _r_exp class AgentPPO: def __init__(self): super().__init__() self.state = None self.device = None self.action_dim = None self.criterion = torch.nn.SmoothL1Loss() self.cri = self.cri_target = self.if_use_cri_target = self.cri_optim = self.ClassCri = None self.act = self.act_target = self.if_use_act_target = self.act_optim = self.ClassAct = None '''init modify''' self.ClassCri = CriticAdv self.ClassAct = ActorPPO self.ratio_clip = 0.2 self.lambda_entropy = 0.02 # be 0.01~0.05 self.lambda_gae_adv = 0.98 # be 0.95~0.99 self.get_reward_sum = None self.trajectory_list = None def init(self, net_dim, state_dim, action_dim, learning_rate=1e-4, gpu_id=0): self.device = torch.device(f"cuda:{gpu_id}" if (torch.cuda.is_available() and (gpu_id >= 0)) else "cpu") self.trajectory_list = list() self.get_reward_sum = self.get_reward_sum_gae self.cri = self.ClassCri(net_dim, state_dim, action_dim).to(self.device) self.act = self.ClassAct(net_dim, state_dim, action_dim).to(self.device) if self.ClassAct else self.cri self.cri_target = deepcopy(self.cri) if self.if_use_cri_target else self.cri self.act_target = deepcopy(self.act) if self.if_use_act_target else self.act self.cri_optim = torch.optim.Adam(self.cri.parameters(), learning_rate) self.act_optim = torch.optim.Adam(self.act.parameters(), learning_rate) if self.ClassAct else self.cri def select_action(self, state): states = torch.as_tensor((state,), dtype=torch.float32, device=self.device) actions, noises = self.act.get_action(states) return actions[0].detach().cpu().numpy(), noises[0].detach().cpu().numpy() def explore_env(self, env, target_step): state = self.state trajectory_temp = list() last_done = 0 for i in range(target_step): action, noise = self.select_action(state) state, next_state, reward, done = env.step(np.tanh(action)) trajectory_temp.append((state, reward, done, action, noise)) if done: state = env.reset() last_done = i else: state = next_state self.state = state '''splice list''' # store 0 trajectory information to list trajectory_list = self.trajectory_list + trajectory_temp[:last_done + 1] self.trajectory_list = trajectory_temp[last_done:] return trajectory_list def update_net(self, buffer, batch_size, repeat_times, soft_update_tau): """put data extract and update network together""" with torch.no_grad(): buf_len = buffer[0].shape[0] # decompose buffer data buf_state, buf_action, buf_noise, buf_reward, buf_mask = [ten.to(self.device) for ten in buffer] '''get buf_r_sum, buf_logprob''' bs = batch_size buf_value = [self.cri_target(buf_state[i:i + bs]) for i in range(0, buf_len, bs)] buf_value = torch.cat(buf_value, dim=0) buf_logprob = self.act.get_old_logprob(buf_action, buf_noise) buf_r_sum, buf_advantage = self.get_reward_sum(buf_len, buf_reward, buf_mask, buf_value) # detach() # normalize advantage buf_advantage = (buf_advantage - buf_advantage.mean()) / (buf_advantage.std() + 1e-5) buf_advantage = smooth_rewards(buf_advantage, window=10) del buf_noise, buffer[:] '''PPO: Surrogate objective of Trust Region''' obj_critic = obj_actor = None for _ in range(int(buf_len / batch_size * repeat_times)): indices = torch.randint(buf_len, size=(batch_size,), requires_grad=False, device=self.device) state = buf_state[indices] action = buf_action[indices] r_sum = buf_r_sum[indices] logprob = buf_logprob[indices] advantage = buf_advantage[indices] new_logprob, obj_entropy = self.act.get_logprob_entropy(state, action) # it is obj_actor ratio = (new_logprob - logprob.detach()).exp() surrogate1 = advantage * ratio surrogate2 = advantage * ratio.clamp(1 - self.ratio_clip, 1 + self.ratio_clip) obj_surrogate = -torch.min(surrogate1, surrogate2).mean() obj_actor = obj_surrogate + obj_entropy * self.lambda_entropy self.optim_update(self.act_optim, obj_actor) # update actor value = self.cri(state).squeeze(1) # critic network predicts the reward_sum (Q value) of state # use smoothloss L1 to evaluate the value loss # obj_critic = self.criterion(value, r_sum) / (r_sum.std() + 1e-6) obj_critic = self.criterion(value, r_sum) self.optim_update(self.cri_optim, obj_critic) # calculate and update the back propogation of value loss # choose to use soft update self.soft_update(self.cri_target, self.cri, soft_update_tau) if self.cri_target is not self.cri else None a_std_log = getattr(self.act, 'a_std_log', torch.zeros(1)) return obj_critic.item(), obj_actor.item(), a_std_log.mean().item() # logging_tuple def get_reward_sum_gae(self, buf_len, ten_reward, ten_mask, ten_value) -> (torch.Tensor, torch.Tensor): buf_r_sum = torch.empty(buf_len, dtype=torch.float32, device=self.device) # old policy value buf_advantage = torch.empty(buf_len, dtype=torch.float32, device=self.device) # advantage value pre_r_sum = 0.0 pre_advantage = 0.0 # advantage value of previous step for i in range(buf_len - 1, -1, -1): buf_r_sum[i] = ten_reward[i] + ten_mask[i] * pre_r_sum pre_r_sum = buf_r_sum[i] buf_advantage[i] = ten_reward[i] + ten_mask[i] * (pre_advantage - ten_value[i]) pre_advantage = ten_value[i] + buf_advantage[i] * self.lambda_gae_adv return buf_r_sum, buf_advantage @staticmethod def optim_update(optimizer, objective): optimizer.zero_grad() objective.backward() optimizer.step() @staticmethod def soft_update(target_net, current_net, tau): for tar, cur in zip(target_net.parameters(), current_net.parameters()): tar.data.copy_(cur.data.__mul__(tau) + tar.data.__mul__(1.0 - tau)) class Arguments: def __init__(self, agent=None, env=None): self.agent = agent self.env = env self.cwd = None # current work directory. None means set automatically self.if_remove = False # remove the cwd folder? (True, False, None:ask me) self.visible_gpu = '0' # os.environ['CUDA_VISIBLE_DEVICES'] = '0, 2,' self.num_threads = 32 # cpu_num for evaluate model '''Arguments for training''' self.num_episode = 1000 self.gamma = 0.995 # discount factor of reward self.learning_rate = 1e-4 # 1e-4 2 ** -14 2e-4 self.soft_update_tau = 2 ** -8 # 1e-3 2 ** -8 self.net_dim = 256 # the network width self.batch_size = 4096 # num of transitions sampled from replay buffer self.repeat_times = 2 ** 5 # collect target_step, then update network self.target_step = 4096 # repeatedly update network to keep critic's loss small self.max_memo = self.target_step # capacity of replay buffer '''Arguments for evaluate''' self.random_seed = 1234 # self.random_seed_list = [1234, 2234, 3234, 4234, 5234] self.random_seed_list = [1234] self.train = True self.save_network = True self.test_network = True self.save_test_data = True def init_before_training(self): if self.cwd is None: agent_name = self.agent.__class__.__name__ self.cwd = f'./{agent_name}' np.random.seed(self.random_seed) torch.manual_seed(self.random_seed) torch.set_num_threads(self.num_threads) torch.set_default_dtype(torch.float32) os.environ['CUDA_VISIBLE_DEVICES'] = str(self.visible_gpu) if __name__ == '__main__': args = Arguments() reward_record = {'episode': [], 'steps': [], 'mean_episode_reward': [], 'unbalance': []} loss_record = {'episode': [], 'steps': [], 'critic_loss': [], 'actor_loss': [], 'entropy_loss': []} args.visible_gpu = '0' for seed in args.random_seed_list: args.random_seed = seed args.agent = AgentPPO() agent_name = f'{args.agent.__class__.__name__}' args.agent.cri_target = True args.env = WgzGym() args.init_before_training() '''init agent and environment''' agent = args.agent env = args.env agent.init(args.net_dim, env.state_space.shape[0], env.action_space.shape[0], args.learning_rate) gamma = args.gamma batch_size = args.batch_size # data used to update net target_step = args.target_step # steps of one episode should stop repeat_times = args.repeat_times # times should update for one batch size data soft_update_tau = args.soft_update_tau num_episode = args.num_episode agent.state = env.reset() '''init buffer''' buffer = list() '''init training params''' # args.train = False # args.save_network = False if args.train: for i_episode in range(num_episode): with torch.no_grad(): trajectory_list = agent.explore_env(env, target_step) _steps, _r_exp = update_buffer(trajectory_list) critic_loss, actor_loss, entropy_loss = agent.update_net(buffer, batch_size, repeat_times, soft_update_tau) loss_record['critic_loss'].append(critic_loss) loss_record['actor_loss'].append(actor_loss) loss_record['entropy_loss'].append(entropy_loss) with torch.no_grad(): episode_reward, episode_unbalance = get_episode_return(env, agent.act, agent.device) reward_record['mean_episode_reward'].append(episode_reward) reward_record['unbalance'].append(episode_unbalance) print(f'epsiode: {i_episode}, reward: {episode_reward}, unbalance: {episode_unbalance}') act_save_path = f'{args.cwd}/actor.pth' loss_record_path = f'{args.cwd}/loss.pkl' reward_record_path = f'{args.cwd}/reward.pkl' if args.save_network: with open(loss_record_path, 'wb') as tf: pickle.dump(loss_record, tf) with open(reward_record_path, 'wb') as tf: pickle.dump(reward_record, tf) torch.save(agent.act.state_dict(), act_save_path) print('actor params have been saved')