RuntimeError: one of the variables needed for gradient computation has been modified by...

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation

我在计算loss的过程中,遇到了以上错误

原因1:浅复制搞的鬼

出错代码1:

temp = target

正确代码1:

temp = target.clone()

出错代码2:

output1[:, 0, :, :, :] = 1 - output1[:, 0, :, :, :]

temp = target
正确代码2:

output2 = output1.clone()
output2[:, 0, :, :, :] = 1 - output1[:, 0, :, :, :]

原因2:使用广播机制

出错代码:

intersection1 = 2. * (output1 * target1).sum()#output1 和 target1大小不一,在计算过程中,target1会经过广播机制扩成和output1同样大小。

正确代码:

target1 = torch.cat([temp,temp],1)
intersection1 = 2. * (output1 * target1).sum()

当然,此错误还有其他原因

你可能感兴趣的:(Debug)