def train(env): if env == "MountainCar-v0": agent = DQN(2, 3) play(env, agent, until=195, ckpt=True) else: agent = DQN(state_dim=(210, 160, 3), n_actions=14) play(env, agent) agent.save_model(env)
def train(cfg): print('Start to train !') device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu env = gym.make('CartPole-v0').unwrapped # 可google为什么unwrapped gym,此处一般不需要 env.seed(1) # 设置env随机种子 n_states = env.observation_space.shape[0] n_actions = env.action_space.n agent = DQN(n_states=n_states, n_actions=n_actions, device=device, gamma=cfg.gamma, epsilon_start=cfg.epsilon_start, epsilon_end=cfg.epsilon_end, epsilon_decay=cfg.epsilon_decay, policy_lr=cfg.policy_lr, memory_capacity=cfg.memory_capacity, batch_size=cfg.batch_size) rewards = [] moving_average_rewards = [] ep_steps = [] log_dir=os.path.split(os.path.abspath(__file__))[0]+"/logs/train/" + SEQUENCE writer = SummaryWriter(log_dir) for i_episode in range(1, cfg.train_eps+1): state = env.reset() # reset环境状态 ep_reward = 0 for i_step in range(1, cfg.train_steps+1): action = agent.choose_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 agent.memory.push(state, action, reward, next_state, done) # 将state等这些transition存入memory state = next_state # 跳转到下一个状态 agent.update() # 每步更新网络 if done: break # 更新target network,复制DQN中的所有weights and biases if i_episode % cfg.target_update == 0: agent.target_net.load_state_dict(agent.policy_net.state_dict()) print('Episode:', i_episode, ' Reward: %i' % int(ep_reward), 'n_steps:', i_step, 'done: ', done,' Explore: %.2f' % agent.epsilon) 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() print('Complete training!') ''' 保存模型 ''' if not os.path.exists(SAVED_MODEL_PATH): # 检测是否存在文件夹 os.mkdir(SAVED_MODEL_PATH) agent.save_model(SAVED_MODEL_PATH+'checkpoint.pth') print('model saved!') '''存储reward等相关结果''' save_results(rewards,moving_average_rewards,ep_steps,tag='train',result_path=RESULT_PATH)
def main(): env = retro.make(game='Frogger-Genesis', use_restricted_actions=retro.Actions.DISCRETE) gamma = 0.99 copy_step = 25 num_actions = env.action_space.n num_states = len(env.observation_space.sample()) hidden_units = [200, 200] max_experiences = 10000 min_experiences = 100 batch_size = 32 lr = 1e-2 current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") log_dir = 'logs/dqn/' + current_time summary_writer = tf.summary.create_file_writer(log_dir) # For stable weights, use one net to train, and copy their weights over to the TargetNet every copy_steps TrainNet = DQN(num_actions=num_actions, gamma=gamma, max_experiences=max_experiences, min_experiences=min_experiences, batch_size=batch_size, lr=lr, hidden_units=hidden_units, num_states=num_states) TargetNet = DQN(num_actions=num_actions, gamma=gamma, max_experiences=max_experiences, min_experiences=min_experiences, batch_size=batch_size, lr=lr, hidden_units=hidden_units, num_states=num_states) # Loading check while True: if os.path.exists(save_dir): if input("\n\nWould you like to load the previous network weights? (y/n) ") == 'y': # load weights and copy to train net TargetNet.load_model(save_path) TrainNet.copy_weights(TargetNet) print("Loaded model weights...") break elif input("\nWould you like to delete the old checkpoints and start again? (y/n)") == 'y': shutil.rmtree(save_dir) print("Removed old checkpoint...") break else: break N = 50000 total_rewards = np.empty(N) epsilon = 0.99 decay = 0.9999 min_epsilon = 0.1 # play N games for n in range(N): epsilon = max(min_epsilon, epsilon * decay) total_reward = play_game(env, TrainNet, TargetNet, epsilon, copy_step) total_rewards[n] = total_reward avg_rewards = total_rewards[max(0, n - 100):(n + 1)].mean() with summary_writer.as_default(): tf.summary.scalar("episode reward", total_reward, step=n) tf.summary.scalar("running avg reward(100)", avg_rewards, step=n) if n % 100 == 0: print("episode:", n, "episode reward:", total_reward, "eps:", epsilon, "avg reward (last 100):", avg_rewards) # save the model weights TargetNet.save_model(save_path) print("avg reward for last 100 episodes:", avg_rewards) if create_video: make_video(env, TrainNet) env.close()