One of the variables needed for gradient computation has been modified by an inplace operation
报错代码：
x2 = self.conv(x1)
sr = self.tail(x2) ------------sr 有去计算loss
x3 = self.relu(x2) ------------x3 后续操作结果也去计算loss
relu 默认 inplace为True ，relu 改变了x2的值 所以sr部分损失回传导致x2的值不对
...
改为：
x2 = self.conv(x1)
x3 = self.relu(x2)
sr = self.tail(x3)
...