8f6a6f153781d0908fb0904349aae844494026ea,models/shiftnet_model.py,ShiftNetModel,backward_D,#ShiftNetModel#,172

Before Change

            x_hat = Variable(alpha * real_AB.data + (1 - alpha) * fake_AB.detach().data, requires_grad=True)

            pred_hat = self.netD(x_hat)
            if self.use_gpu:
                gradients = torch.autograd.grad(outputs=pred_hat, inputs=x_hat, grad_outputs=torch.ones(pred_hat.size()).cuda(),
                                create_graph=True, retain_graph=True, only_inputs=True)[0]
            else:
                gradients = torch.autograd.grad(outputs=pred_hat, inputs=x_hat, grad_outputs=torch.ones(pred_hat.size()),
                                    create_graph=True, retain_graph=True, only_inputs=True)[0]

            gradient_penalty = self.gp_lambda * ((gradients.view(gradients.size()[0], -1).norm(2, 1) - 1) ** 2).mean()

            self.loss_D = self.loss_D_fake - self.loss_D_real + gradient_penalty
        else:

After Change

            x_hat.requires_grad_(True)
            pred_hat = self.netD(x_hat)

            gradients = torch.autograd.grad(outputs=pred_hat, inputs=x_hat, grad_outputs=torch.ones(pred_hat.size()).to(self.device),
                                create_graph=True, retain_graph=True, only_inputs=True)[0]

            gradient_penalty = self.gp_lambda * ((gradients.view(gradients.size(0), -1).norm(2, 1) - 1) ** 2).mean()

Instances