正在解决“typeerror:max()接收到无效的参数组合-Get(out=nonetype,axis=nonetype),但应为?”

iqih9akk  于 2021-09-29  发布在  Java
关注(0)|答案(0)|浏览(723)

我正在为pommerman环境研究一种深度q-学习强化学习算法。
我最初得到这个 RuntimeError: mat1 and mat2 shapes cannot be multiplied (1x30 and 201x128) 运行主函数时出错,但通过更改 --batch 从…起 main.py 到201,以便匹配mat1和mat2尺寸。

Episode:  1 finished, result: Lose
Avg Episode Reward: -380.0
Traceback (most recent call last):
  File "/Users/tylerkim/Desktop/rlena/DQN/main.py", line 100, in <module>
    main()
  File "/Users/tylerkim/Desktop/rlena/DQN/main.py", line 84, in main
    agent1.update(args.gamma, args.batch)
  File "/Users/tylerkim/Desktop/rlena/DQN/agent.py", line 63, in update
    expectQ = rewards + gamma * np.max(self.forward(next_states))
  File "<__array_function__ internals>", line 6, in amax
  File "/Users/tylerkim/opt/anaconda3/envs/rlena/lib/python3.7/site-packages/numpy/core/fromnumeric.py", line 2734, in amax
    keepdims=keepdims, initial=initial, where=where)
  File "/Users/tylerkim/opt/anaconda3/envs/rlena/lib/python3.7/site-packages/numpy/core/fromnumeric.py", line 85, in _wrapreduction
    return reduction(axis=axis, out=out,**passkwargs)
TypeError: max() received an invalid combination of arguments - got (out=NoneType, axis=NoneType, ), but expected one of:
 * ()
 * (Tensor other)
 * (int dim, bool keepdim)
      didn't match because some of the keywords were incorrect: out, axis
 * (name dim, bool keepdim)
      didn't match because some of the keywords were incorrect: out, axis\
Process finished with exit code 1

但在我尝试了之后,我得到了以下无效参数组合的typeerror,在这里我似乎为两个参数都输入了nonetype。我不知道我在哪里犯了错误。如果有人能看一下我的代码,我将不胜感激!
这是agent.py的代码

class DQNAgent(BaseAgent):
    """
    DQN from scratch
    """
    def __init__(self, env, args, character=characters.Bomber):
        super(DQNAgent, self).__init__(character)
        self.obs_n = env.observation_space.shape[0]  # output_dim
        self.action_n = env.action_space.n  # input_dim

        self.epsilon = args.epsilon
        self.eps_decay = args.eps_decay
        self.min_eps = args.min_eps
        self.gamma = args.gamma
        self.lr = args.lr

        self.episodes = args.episodes
        self.maxsteps = args.maxsteps
        self.showevery = args.showevery

        self.capacity = args.capacity
        self.batch = args.batch
        self.buffer = ReplayBuffer(self.capacity, self.batch)

        self.model = nn.Sequential(
            nn.Linear(self.obs_n, 128),
            nn.ReLU(),
            nn.Linear(128, 256),
            nn.ReLU(),
            nn.Linear(256, self.action_n)
        )

        self.optim = optim.Adam(self.model.parameters(), self.lr)
        self.MSE_loss = nn.MSELoss()

    def forward(self, state):
        state = torch.FloatTensor(state)
        qvals = self.model(state)
        # qvals (20,2,6) state (20,2,201)
        return qvals

    def act(self, obs):
        return np.argmax(self.forward(obs))

    def update(self, gamma, batch_size):
        batch = self.buffer.sample(batch_size)

        states, actions, rewards, next_states, done = batch
        # actions = {list: 20}(batchsize) of list:2
        currQ = self.forward(states)
        expectQ = rewards + gamma * np.max(self.forward(next_states))
        loss = self.MSE_loss(currQ, expectQ) # TODO: try Huber Loss later too

        self.optim.zero_grad()
        loss.backward()
        self.optim.step()

    def epsdecay(self):
        self.epsilon = self.epsilon * self.eps_decay if self.epsilon > self.min_eps else self.epsilon

这是主要的

def main():
    parser = argparse.ArgumentParser(description='DQN pommerman MARL')
    parser.add_argument('--episodes', type=int, default=3000, help='episodes')
    parser.add_argument('--maxsteps', type=int, default=200, help='maximum steps')
    parser.add_argument('--showevery', type=int, default=300, help='report loss every n episodes')

    parser.add_argument('--epsilon', type=float, default=0.05, help='parameter for epsilon greedy')
    parser.add_argument('--eps_decay', type=float, default=0.995, help='epsilon decay rate')
    parser.add_argument('--min_eps', type=float, default=0.05, help='minimum epsilon for decaying')
    parser.add_argument('--gamma', type=float, default=0.95, help='gamma')
    parser.add_argument('--lr', type=float, default=0.01, help='learning rate')

    parser.add_argument('--capacity', type=int, default=100000, help='capacity for replay buffer')
    parser.add_argument('--batch', type=int, default=201, help='batch size for replay buffer')

    parser.add_argument('--gpu', type=str, default='0', help='gpu number')

    args = parser.parse_args()

    # GPU
    args.device = torch.device('cuda:{}'.format(args.gpu) if torch.cuda.is_available() else "cpu")
    print("GPU using status: ", args.device)

    agent_list = [agents.SimpleAgent(), agents.SimpleAgent()]  # placeholder
    env = pommerman.make('OneVsOne-v0', agent_list)

    agent1 = DQNAgent(env, args)  # TODO: assertionerror; not agents.BaseAgent??
    agent2 = agents.SimpleAgent()

    agent_list = [agent1, agent2]
    # env = pommerman.make('OneVsOne-v0', agent_list)

    episode_rewards = []
    action_n = env.action_space.n

    for episode in range(args.episodes):
        states = env.reset()
        state_feature = featurize(env, states)
        done = False
        episode_reward = 0
        for step in range(args.maxsteps):
            # if agent1.epsilon > random.random():
            #     action = random.randint(0, action_n - 1)
            # else:
            #     action = agent1.act(state_feature[0])

            env.render()

            actions = env.act(states)
            # TODO: env.set_training_agent(agents[-1].agent_id)으로 training agent를 명시
            # 그리고 위에 env.act(states)에서 training agent action만 따로 append하기

            next_state, reward, done, info = env.step(actions)  # n-array with action for each agent
            next_state_feature = featurize(env, next_state)
            episode_reward += (reward[0]+reward[1])
            agent1.buffer.append([state_feature, actions, reward, next_state_feature, done])
            # env.get_observation -> 찾아보기 (forward model)

            if len(agent1.buffer) > args.batch:
                agent1.update(args.gamma, args.batch)

        if done:
            episode_rewards.append(episode_reward)
        if episode % args.showevery == 0:
            print(f"Episode: {episode + 1:2d} finished, result: {'Win' if 0 in info.get('winners', []) else 'Lose'}")
            print(f"Avg Episode Reward: {np.mean(episode_rewards)}")

        agent1.epsdecay()

    env.close()

暂无答案!

目前还没有任何答案,快来回答吧!

相关问题