Autograd Error on carrying out two times backpropagation

I got this error on optimizing a model, how the computational graph can be modified without getting this error?
one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [500, 1]], which is output 0 of AsStridedBackward0, is at version 1001; expected version 1000 instead.