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

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

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

PyTorch默认会跟踪张量的操作历史，以便计算梯度，但是原地操作会破坏操作历史，导致无法计算梯度。在计算梯度的时候检查出某个Variable有被一个 inplace operation 修改。报错信息会更加具体提示是网络那部分梯度计算出现问题。

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation 原有： pytorch 换了一个网络跑模型 整个项目是模块化编程 对应的网络都是有自己的模块的，因为替换了一下网络，结果报了这个错，很尴尬，明明什么都没动，就是换了一下网络，...

问题描述： RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation 原因分析： 跑过一个GAN的Super-Resolution

最近在用Pytorch训练网络的时候出现了一个奇怪的错误，找了很久在偶然间试出了解决方法，因此记录一下，以作备忘。但具体原理还不是很清楚，也请各位大佬指教。RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [19, 24, 30]], which is output 0 of SoftmaxBackwa

今天跑网络在进行loss.backward()的时候，出现了如下错误： one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [2, 64]], which is output 0 of ViewBackward, is at version 21; expected version 20 instead. Hint: enabl

修改网络时 pytroch报错： RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [1600, 16, 256]], which is output 0 of CudnnConvolutionBackward, is at version 1; expected version 0 ins

处理 Runtime Error: one of the variables needed for gradient computation has been

这样一来，在次运行，这个问题就解决了。根据这里面的描述我们可以得到初步结论是这个错误表明在计算梯度时，某个变量已经通过inplace操作进行了修改，但通常情况下，PyTorch会要求你不要对需要梯度计算的变量进行inplace操作，因为这样可能导致梯度计算错误。回想我们之前的代码问题，是在进行梯度传播的时候我们在保存计算图的时候出现了问题：因为梯度计算要迭代计算，所以需要在原来的基础上继续进行计算，所以需要保持计算图。所以这个的解决办法是，按照代码的执行逻辑去查看前面的计算图，然后我们把对应的代码改了。

【就看这一篇就行】RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [256]] is at version 4; expected version 3 instead. Hint: enable anomaly detection to find the operation that fai