gpu训练神经网络时报错

Terence888

1. #plus 1 to avoid log(0)
   
2. def MSLE_loss(pred, target):
   
3.     log_pred = torch.log(pred + 1)  
   
4.     log_target = torch.log(target + 1)  
   
5.     loss = nn.MSELoss()(log_pred, log_target)  
   
6.     return loss
   
7. 
   
8. model = Net().cuda()
   
9. optimizer = optim.AdamW(model.parameters(),lr=1e-3,weight_decay=1e-3)
   
10. for epoch in range(3500):
    
11.     pred = model(X_train)
    
12.     pred = pred.squeeze()
    
13.     loss = MSLE_loss(pred,y_train)
    
14.    
    
15.     if epoch % 500 == 0 :
    
16.         print(loss)
    
17.         print(pred)
    
18.         print(y_train)
    
19.     loss.backward()
    
20.     optimizer.step()
    
21.     optimizer.zero_grad()

复制代码

报错如下

1. ---------------------------------------------------------------------------
   
2. RuntimeError                              Traceback (most recent call last)
   
3. Cell In[54], line 11
   
4.       9 optimizer = optim.AdamW(model.parameters(),lr=1e-3,weight_decay=1e-3)
   
5.      10 for epoch in range(3500):
   
6. ---> 11     pred = model(X_train)
   
7.      12     pred = pred.squeeze()
   
8.      13     loss = MSLE_loss(pred,y_train)
   
9. 
   
10. File /opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py:1553, in Module._wrapped_call_impl(self, *args, **kwargs)
    
11.    1551     return self._compiled_call_impl(*args, **kwargs)  # type: ignore[misc]
    
12.    1552 else:
    
13. -> 1553     return self._call_impl(*args, **kwargs)
    
14. 
    
15. File /opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py:1562, in Module._call_impl(self, *args, **kwargs)
    
16.    1557 # If we don't have any hooks, we want to skip the rest of the logic in
    
17.    1558 # this function, and just call forward.
    
18.    1559 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
    
19.    1560         or _global_backward_pre_hooks or _global_backward_hooks
    
20.    1561         or _global_forward_hooks or _global_forward_pre_hooks):
    
21. -> 1562     return forward_call(*args, **kwargs)
    
22.    1564 try:
    
23.    1565     result = None
    
24. 
    
25. Cell In[45], line 23, in Net.forward(self, x)
    
26.      21 def forward(self,x):
    
27.      22     inputs = x
    
28. ---> 23     layer1_out = torch.nn.functional.gelu(self.hidden_layer1(inputs))
    
29.      24     layer1_out = self.dropout(layer1_out)
    
30.      25     layer2_out = torch.nn.functional.gelu(self.hidden_layer2(layer1_out))
    
31. 
    
32. File /opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py:1553, in Module._wrapped_call_impl(self, *args, **kwargs)
    
33.    1551     return self._compiled_call_impl(*args, **kwargs)  # type: ignore[misc]
    
34.    1552 else:
    
35. -> 1553     return self._call_impl(*args, **kwargs)
    
36. 
    
37. File /opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py:1562, in Module._call_impl(self, *args, **kwargs)
    
38.    1557 # If we don't have any hooks, we want to skip the rest of the logic in
    
39.    1558 # this function, and just call forward.
    
40.    1559 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
    
41.    1560         or _global_backward_pre_hooks or _global_backward_hooks
    
42.    1561         or _global_forward_hooks or _global_forward_pre_hooks):
    
43. -> 1562     return forward_call(*args, **kwargs)
    
44.    1564 try:
    
45.    1565     result = None
    
46. 
    
47. File /opt/conda/lib/python3.10/site-packages/torch/nn/modules/linear.py:117, in Linear.forward(self, input)
    
48.     116 def forward(self, input: Tensor) -> Tensor:
    
49. --> 117     return F.linear(input, self.weight, self.bias)
    
50. 
    
51. RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument mat1 in method wrapper_CUDA_addmm)

复制代码

FishC

报错显示您的张量在不同设备上：有的在GPU上（cuda:0），有的在CPU上。解决办法如下：

1. **将数据移动到GPU**：确保`X_train`和`y_train`在使用时都在GPU上。使用`.cuda()`方法将它们移到GPU上。例如：

  

1. X_train = X_train.cuda()
   
2.    y_train = y_train.cuda()

复制代码

2. **检查模型参数和其他张量**：确保模型的所有参数和输入张量都在同一设备上。可以在定义模型和数据加载时直接使用`.to(device)`，其中`device`是定义的计算设备，如：

  

1. device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
   
2.    model = Net().to(device)
   
3.    X_train = X_train.to(device)
   
4.    y_train = y_train.to(device)

复制代码

确保所有的计算都在同一设备上，将解决这个错误。

以上回复来自 FishC AI 小助理，如未能正确解答您的问题，请继续追问。