原因是把创建loss的语句loss_aux = torch.tensor(0.)放在循环体外了，可能的解释是第一次backward后把计算图删除，第二次backward就会找不到父节点，也就无法反向传播。参考：https://stackoverflow.com/questions/55268726/pytorch-why-does-preallocating-memory-cause-trying-to-backward-through-the-gr
标签：loss,Trying,graph,second,through,backward From： https://www.cnblogs.com/lipoicyclic/p/17762890.html