张量的元素 0 不需要 grad 并且没有 grad_fn

Question

我正在尝试将强化学习机制应用于分类任务。我知道这样做是没有用的，因为深度学习在任务中的表现可能超过 RL。不管怎样，我正在做研究。

如果代理人正确为正 1 或不为负 -1，我会奖励他并用 predicted_action(predicted_class) 和奖励计算损失 FUNC。

但是我得到一个错误：

element 0 of tensors does not require grad and does not have a grad_fn

 # creating model
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()

        self.pipe = nn.Sequential(nn.Linear(9, 120),
                                 nn.ReLU(),
                                 nn.Linear(120, 64),
                                 nn.ReLU(),
                                 nn.Linear(64,2),
                                 nn.Softmax()
                                 )

    def forward(self, x):
        return self.pipe(x)


def env_step(action, label, size):
    total_reward = []

    for i in range(size):
        reward = 0

        if action[i] == label[i]:
            total_reward.append(reward+1)
            continue
        else:
            total_reward.append(reward-1)
            continue

    return total_reward




if __name__=='__main__':
    epoch_size = 100
    net = Net()
    criterion = nn.MSELoss()
    optimizer = optim.Adam(params=net.parameters(), lr=0.01)

    total_loss = deque(maxlen = 50)

    for epoch in range(epoch_size):
        batch_index = 0
        for i in range(13):
            # batch sample
            batch_xs = torch.FloatTensor(train_state[batch_index: batch_index+50])   # make tensor
            batch_ys = torch.from_numpy(train_label[batch_index: batch_index+50]).type('torch.LongTensor')  # make tensor

            # action_prob; e.g classification prob
            actions_prob = net(batch_xs)                                
            #print(actions_prob)
            action = torch.argmax(actions_prob, dim=1).unsqueeze(1)    
            #print(action)
            reward = np.array(env_step(action, batch_ys, 50))  
            #print(reward)

            reward = torch.from_numpy(reward).unsqueeze(1).type('torch.FloatTensor')
            #print(reward)
            action = action.type('torch.FloatTensor')

            optimizer.zero_grad()
            loss = criterion(action, reward)    
            loss.backward()
            optimizer.step()


            batch_index += 50

Answer 1

action 是由 argmax 函数产生的，它是不可微分的。相反，您希望在奖励和所采取行动的 责任概率 之间取损失。

通常，在强化学习中为策略选择的 "loss" 就是所谓的 score function：

这是采取的行动 a 的责任概率乘以获得的奖励的对数乘积。

张量的元素 0 不需要 grad 并且没有 grad_fn

element 0 of tensors does not require grad and does not have a grad_fn

python

classification

reinforcement-learning

pytorch