e9e215c2f7ca0bbbded08d34c834c2a3f43bb8d0,pro_gan_pytorch/Losses.py,WGAN_GP,__gradient_penalty,#WGAN_GP#Any#Any#Any#Any#Any#,109

Before Change

        // create the merge of both real and fake samples
        merged = epsilon * real_samps + ((1 - epsilon) * fake_samps)
        merged = merged.cuda(fake_samps.device)
        merged = th.autograd.Variable(merged, requires_grad=True)

        // forward pass
        op = self.dis(merged, height, alpha)

After Change

        // perform backward pass from op to merged for obtaining the gradients
        gradient = th.autograd.grad(outputs=op, inputs=merged,
                                    grad_outputs=th.ones_like(op), create_graph=True,
                                    retain_graph=True, only_inputs=True)[0]

        // calculate the penalty using these gradients

Instances