def g_path_regularize(fake_img, latents, mean_path_length, decay=0.01):
    noise = paddle.randn(fake_img.shape) / math.sqrt(
        fake_img.shape[2] * fake_img.shape[3])
    grad = paddle.grad(outputs=(fake_img * noise).sum(),
                       inputs=latents,
                       create_graph=True)[0]
    path_lengths = paddle.sqrt((grad * grad).sum(2).mean(1))

    path_mean = mean_path_length + decay * (path_lengths.mean() -
                                            mean_path_length)

    path_penalty = ((path_lengths - path_mean) *
                    (path_lengths - path_mean)).mean()

    return path_penalty, path_lengths.detach().mean(), path_mean.detach()