pytorch自定义loss,如何进行后向传播loss.backward()?

Question

之前loss用自带的MSE,这样写的

criterion = nn.MSELoss(size_average=False).cuda()
...
loss = criterion(output, target)
loss.backward()

这样是没有问题的

---

现在需要用自定义loss函数newLoss,因为要逐个像素进行loss运算(算法需要)

#this is in model.py
class newLoss(nn.Module):
    def __init__(self):
        super(newLoss, self).__init__()
    def forward(self, output, gt):
        loss = 0
        for row_out, row_gt :
            for pixel_out, pixel_gt :
                loss += something pixelwise
        return loss
# this is in train.py
newloss = newLoss()
loss = newloss(output,gt)

这样计算出来的loss是float类型的,下面的代码会报
''AttributeError: 'float' object has no attribute 'backward''
的错

---

我现在的做法是:把newloss数值加到原来的MSE类型loss上:

criterion = nn.MSELoss(size_average=False).cuda()
...
loss = criterion(output, target)
newloss= newLoss()
loss += newloss(output,gt)
loss.backward()

这样写我新加的newloss在后向传播时能生效吗?

Answer 1

你这是数据类型的问题吧，把newLoss装载进cuda里，还有类定义的变量用torch类型，能够反向传播是要求torch.tensor