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

起因：pytorch训练，之前能train的程序，一点没改，现在报错，记录改bug过程。

报错代码段：

loss_qf = L1loss(QF, qf)
loss_img = L1loss(img_E, img_H_tensor)
loss_train = loss_img + 0.1 * loss_qf
print(loss_train)
optimizer.zero_grad()
loss_train.backward()
optimizer.step()
scheduler.step()

报错：

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [16, 512, 16, 16]], which is output 0 of AddBackward0, is at version 1; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

搜索解决方法：

1、两个loss都进行backward()操作时，丢失参数。但是报错和我的有区别，原链接是TBackward且是两个loss，我只有一个loss，报错AddBackward0。方法不行。

https://discuss.pytorch.org/t/solved-pytorch1-5-runtimeerror-one-of-the-variables-needed-for-gradient-computation-has-been-modified-by-an-inplace-operation/90256

2、将网络中的nn.ReLU(inplace=True)改成False；将所有inplace操作转换为非inplace操作，如将x += 1换为y = x + 1。也无效。

3、更改torch版本，原版本1.11.0升级到1.91报错相同，降级到1.10.0，报错变成ReluBackward0。尝试步骤2，删去nn.ReLU(inplace=True)中的inplace=True，程序跑通了！！！

https://github.com/pytorch/pytorch/issues/24853

总结，pytorch版本问题，有点玄学。如果是双loss回传，第一条能解决，报错ReluBackward0用第二条，没发现问题，建议更换pytorch版本。