I would like to translate and rotate a torch.autograd.Variable() using the output of a ConvNet and calculate the loss for back propagation from translated+rotated images and input fixed image. If I convert it to numpy and translate, the loss.backward() does not work. How do I solve this problem?