qwerklkjhgfd

alpha = torch.rand(real_x.size(0), 1, 1, 1).cuda().expand_as(real_x)
beta = torch.rand(real_x.size())
y = real_x.data + 0.5 * real_x.data.std() * beta
interpolated = real_x.data + alpha * (y - real_x.data)
interpolated = Variable(interpolated, requires_grad=True)
out, out_cls = self.D(interpolated)
grad = torch.autograd.grad(outputs=[out, out_cls], inputs=interpolated,
                           grad_outputs=[torch.ones(out.size()).cuda(), torch.ones(out_cls.size()).cuda()],
                           retain_graph=True, create_graph=True, only_inputs=True)[0]
grad = grad.view(grad.size(0), -1)
d_loss_gp = ((grad.norm(p=2, dim=1) - 1) ** 2).mean()