Pytorch新手，新写了个loss函数，运行报错，有高手知道怎么...

lovelylucky

RuntimeError: there are no graph nodes that require computing gradients


我把阈值＞0.5去掉就没有问题
代码如下：

1. class DiceLoss(nn.Module):
   
2. def __init__(self):
   
3. super(DiceLoss, self).__init__()
   
4. self.sigmoid = nn.Sigmoid()
   
5. 
   
6. def forward(self, output, labels):
   
7. 
   
8. batch_size = labels.size(0)
   
9. 
   
10. Sigmoidout = self.sigmoid(output) > 0.5
    
11. 
    
12. loss = 0
    
13. 
    
14. for i in range(batch_size):
    
15. 
    
16. sampleloss = -(2.0*torch.sum(torch.mul(Sigmoidout[i,:],labels[i,:])) + 1e-5)/(torch.sum(labels[i,:]) + torch.sum(Sigmoidout[i,:]) + 1e-5)
    
17. 
    
18. loss += sampleloss
    
19. 
    
20. meanloss = loss/batch_size
    
21. 
    
22. return meanloss

复制代码

2919005896

虽然没有贴完整的代码和关键的代码，但是，损失函数的第一项是网络的输出，而不是：

1. criterion = nn.SmoothL1Loss()
   
2. loss = criterion(input, target)

复制代码

以上是错的，以下是对的：

1. criterion = nn.SmoothL1Loss()
   
2. loss = criterion(output, target)

复制代码