运行时错误（"grad can be implicitly created only for scalar outputs"）

Question

我有以下用于训练 A3C 的训练函数代码。我遇到了以下错误。

RuntimeError("grad can be implicitly created only for scalar outputs")

第 (policy_loss + 0.5 * value_loss).backward()

行

这是我的代码，有人可以帮忙检查一下这段代码有什么问题吗？

def train(rank, params, shared_model, optimizer,ticker):
torch.manual_seed(params.seed + rank) # shifting the seed with rank to asynchronize each training agent
print(ticker)
try:
    ohlcv = pd.read_csv(ticker + '.csv')
    data = ohlcv.copy()
    data['rsi'] = ab.RSI(data['Close'],14)
    data['adx'] = ab.ADX(data,20)
    data = ab.BollBnd(data,20)
    data['BBr'] = data['Close']/data['BB_dn']
    data['macd_signal'] = ab.MACD(data)
    data['macd_r'] = data['macd_signal']/data['Close']
    data['ema20'] = ab.EMA(np.asarray(data['Close']), 20)
    data['ema20_r'] = data['ema20']/data['Close']
    data['Close'] = data['Close']/data['Close'].max()
    data = data.iloc[:,[4,7,8,13,15,17]]
    data = data.dropna()
    data = torch.DoubleTensor(np.asarray(data))
    env = ENV(state_dim, action_dim, data)
    
    model = ActorCritic(env.observation_space, env.action_space)
    state = env.reset()
    done = True
    episode_length = 0
    while True:
        episode_length += 1
        model.load_state_dict(shared_model.state_dict())
        if done:
            cx = Variable(torch.zeros(1, state_dim)) # the cell states of the LSTM are reinitialized to zero
            hx = Variable(torch.zeros(1, state_dim)) # the hidden states of the LSTM are reinitialized to zero
        else:
            cx = Variable(cx.data)
            hx = Variable(hx.data)
        values = []
        log_probs = []
        rewards = []
        entropies = []
        for step in range(params.num_steps):
            value, action_values, (hx, cx) = model((Variable(state.unsqueeze(0)), (hx, cx)))
            prob = F.softmax(action_values,-1)
            log_prob = F.log_softmax(action_values,-1)
            entropy = -(log_prob * prob).sum(1)
            entropies.append(entropy)
            action = prob.multinomial(num_samples = action_dim).data
            log_prob = log_prob.gather(1, Variable(action))
            values.append(value)
            log_probs.append(log_prob)
            state, reward, done = env.step(action)
            done = (done or episode_length >= params.max_episode_length)
            reward = max(min(reward, 1), -1) # clamping the reward between -1 and +1
            if done:
                episode_length = 0
                state = env.reset()
            rewards.append(reward)
            if done:
                break
        R = torch.zeros(1, 1)
        if not done: # if we are not done:
            value, _, _ = model((Variable(state.unsqueeze(0)), (hx, cx)))
            R = value.data
        values.append(Variable(R))
        policy_loss = torch.zeros(1, 1)
        value_loss = torch.zeros(1, 1) 
        R = Variable(R)
        gae = torch.zeros(1, 1)
        for i in reversed(range(len(rewards))):
            R = params.gamma * R + rewards[i]
            advantage = R - values[i]
            print("advantage:",advantage)
            value_loss = value_loss + 0.5 * advantage.pow(2) # computing the value loss
            TD = rewards[i] + params.gamma * values[i + 1].data - values[i].data # computing the temporal difference
            gae = gae * params.gamma * params.tau + TD # gae = sum_i (gamma*tau)^i * TD(i) with gae_i = gae_(i+1)*gamma*tau + (r_i + gamma*V(state_i+1) - V(state_i))
            print("gae:",gae)
            policy_loss = policy_loss - log_probs[i] * Variable(gae) - 0.01 * entropies[i] # computing the policy loss
            print("policy_loss:",policy_loss)
        optimizer.zero_grad() # initializing the optimizer
        los = policy_loss + 0.5 * value_loss
        print("los",los.shape)
        (policy_loss + 0.5 * value_loss).backward()
        torch.nn.utils.clip_grad_norm(model.parameters(), 40) # clamping the values 
        ensure_shared_grads(model, shared_model) 
        optimizer.step() # running the optimization step
except Exception as e:
    print(e)
    traceback.print_exc()
    var = traceback.format_exc()

下面是输出：-

advantage: tensor([[-1.0750]], grad_fn=<ThSubBackward>)
gae: tensor([[-1.0750]])
policy_loss: tensor([[-25.8590, -26.1414, -25.9023, -25.2628]], grad_fn=<ThSubBackward>)
los torch.Size([1, 4])

RuntimeError：只能为标量输出隐式创建 grad PS E:\ML\Breakout_a3c\Code_With_Comments>

Answer 1

您得到的 pytorch 错误意味着“您只能在标量上向后调用，即 0 维张量”。在这里，根据您的打印，policy_loss 不是标量，它是一个 1x4 矩阵。因此，policy_loss + 0.5 * value_loss 也是如此。因此，您对 backward 的调用会产生错误。

您可能忘记了将损失减少到标量（使用 norm 或 MSELoss ... 等函数）。参见示例 here

它不起作用的原因是梯度传播在内部工作的方式（它基本上是一个雅可比乘法引擎）。您可以在 non-scalar 张量上向后调用，但是您必须自己提供梯度，例如：

# loss is 1x4
loss = policy_loss + 0.5 * value_loss
# explicit gradient backprop with non-scalar tensor
loss.backward(torch.ones(1,4))

如果没有很好地理解 Pytorch 的 Autograd 的工作原理及其含义，您真的不应该这样做。

PS：下次，请提供一个最小的工作示例:)

运行时错误（"grad can be implicitly created only for scalar outputs"）

RuntimeError("grad can be implicitly created only for scalar outputs")

python

reinforcement-learning

pytorch