Python 如何调试索引0处的无效渐变？-皮托克

我试图训练一个演员-评论家模型，但当我到达评论家的后支柱时，我得到了以下错误：

运行时错误：索引0处的梯度无效-预期类型为torch.cuda.FloatTensor，但获得torch.FloatTensor

我无法识别错误所指的梯度。有人能帮忙吗

以下是堆栈跟踪：

Traceback (most recent call last):
  File "train.py", line 338, in <module>
    main()
  File "train.py", line 327, in main
    reinforce_trainer.train(opt.start_reinforce, opt.start_reinforce + opt.critic_pretrain_epochs - 1, True, start_time)
  File "/home/fbommfim/init-tests/treeLSTM/lib/train/reinforce_trainer.py", line 56, in train
    train_reward, critic_loss = self.train_epoch(epoch, pretrain_critic, no_update)
  File "/home/fbommfim/init-tests/treeLSTM/lib/train/reinforce_trainer.py", line 153, in train_epoch
    critic_loss = self.critic.backward(baselines.cuda(), rewards, critic_weights.cuda(), num_words, self.critic_loss_func, regression=True)
  File "/home/fbommfim/init-tests/treeLSTM/lib/model/encoder_decoder/hybrid2seq_model.py", line 67, in backward
    outputs.backward(grad_output)
  File "/home/linuxbrew/.linuxbrew/Cellar/python/3.7.6_1/lib/python3.7/site-packages/torch/tensor.py", line 195, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)
  File "/home/linuxbrew/.linuxbrew/Cellar/python/3.7.6_1/lib/python3.7/site-packages/torch/autograd/__init__.py", line 99, in backward
    allow_unreachable=True)  # allow_unreachable flag
RuntimeError: invalid gradient at index 0 - expected type torch.cuda.FloatTensor but got torch.FloatTensor

对于

hybrid2seq\u model.py

，和

backward

：

def backward(self, outputs, targets, weights, normalizer, criterion, regression=False):
        grad_output, loss = self.generator.backward(outputs, targets, weights, normalizer, criterion, regression)
        outputs.cuda()
        grad_output.cuda()
        outputs.backward(grad_output)
        return loss

---

错误表明它期望一个cuda张量，而得到一个非cuda张量，所以这就是我要寻找的

像

grad\u output.cuda（）这样的调用将返回一个cuda张量。这不是就地操作。您可能想要
grad\u output=grad\u output.cuda（）
，所以我首先要修复这样的调用

def backward(self, outputs, targets, weights, normalizer, criterion, regression=False):
        grad_output, loss = self.generator.backward(outputs, targets, weights, normalizer, criterion, regression)
        outputs.cuda()
        grad_output.cuda()
        outputs.backward(grad_output)
        return loss