retain_graph=True 相关问题

运行406_conditional_GAN时，出现以下错误：
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [128, 1]], which is output 0 of TBackward, is at version 2; expected version 1 instead.
请问是什么原因？

原因：在计算生成器网络梯度之前先对判别器进行更新，修改了某些值，导致Generator网络的梯度计算失败。
解决方法：按图修改该部分代码。