GAN_GP_Img.py 文件源码

def cal_real_nearby_grad_penalty(self, real_data):
        # WGAN lipschitz-penalty
        epsilon = tf.random_uniform(shape=[self.batch_size, 1, 1], minval=0., maxval=1.)

        data_diff = get_perturbed_batch(real_data) - real_data
        interp_data = real_data + epsilon * data_diff
        disc_real_nearby, _ = discriminator(
            self.d_net, interp_data, self.conv_hidden_num,
            self.normalize_d
        )
        grad_real_nearby = tf.gradients(disc_real_nearby, [interp_data])[0]
        print('The shape of grad_real_nearby: {}'.format(grad_real_nearby.get_shape().as_list()))
        grad_real_nearby_flat = tf.reshape(grad_real_nearby, [self.batch_size, -1])
        slope = tf.norm(grad_real_nearby_flat, axis=1)
        print('The shape of slope: {}'.format(slope.get_shape().as_list()))

        grad_penalty = tf.reduce_mean((slope - 1.) ** 2)
        return grad_penalty