深度有趣 | 07 生成式对抗网络

简介

除VAE之外，生成式对抗网络（Generative Adversarial Nets，GAN）也是一种非常流行的无监督生成式模型

GAN中主要包括两个核心网络

• 生成器（Generator）：记作G，通过对大量样本的学习，能够生成一些以假乱真的样本，和VAE类似
• 判别器（Discriminator）：记作D，接受真实样本和G生成的样本，并进行判别和区分
• G和D相互博弈，通过学习，G的生成能力和D的判别能力都逐渐增强并收敛

GAN的训练非常困难，有很多细节需要注意，才能生成质量较高的图片

• 恰当地使用Batch Normalization、LeakyReLU
• 用strides为2的卷积代替池化
• 交替地训练，避免一方过强

这里我们以MNIST为例，通过TensorFlow实现GAN，由于用到深度卷积神经网络，所以也称作DCGAN（Deep Convolutional GAN）

原理

对于一个服从随机分布的噪音z，生成器通过一个复杂的映射函数生成假的样本

判别器则使用另一个复杂的映射函数，对于真实样本或假的样本，输出一个0至1之间的值，越大表示越有可能是真实的样本

总的目标函数如下

实现

加载库

# -*- coding: utf-8 -*-

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
import os, imageio

加载数据

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data')

定义一些常量、网络输入、辅助函数

batch_size = 100
z_dim = 100

OUTPUT_DIR = 'samples'
if not os.path.exists(OUTPUT_DIR):
    os.mkdir(OUTPUT_DIR)

X = tf.placeholder(dtype=tf.float32, shape=[None, 28, 28, 1], name='X')
noise = tf.placeholder(dtype=tf.float32, shape=[None, z_dim], name='noise')
is_training = tf.placeholder(dtype=tf.bool, name='is_training')

def lrelu(x, leak=0.2):
    return tf.maximum(x, leak * x)

def sigmoid_cross_entropy_with_logits(x, y):
    return tf.nn.sigmoid_cross_entropy_with_logits(logits=x, labels=y)

判别器部分

def discriminator(image, reuse=None, is_training=is_training):
    momentum = 0.9
    with tf.variable_scope('discriminator', reuse=reuse):
        h0 = lrelu(tf.layers.conv2d(image, kernel_size=5, filters=64, strides=2, padding='same'))
        
        h1 = tf.layers.conv2d(h0, kernel_size=5, filters=128, strides=2, padding='same')
        h1 = lrelu(tf.contrib.layers.batch_norm(h1, is_training=is_training, decay=momentum))
        
        h2 = tf.layers.conv2d(h1, kernel_size=5, filters=256, strides=2, padding='same')
        h2 = lrelu(tf.contrib.layers.batch_norm(h2, is_training=is_training, decay=momentum))
        
        h3 = tf.layers.conv2d(h2, kernel_size=5, filters=512, strides=2, padding='same')
        h3 = lrelu(tf.contrib.layers.batch_norm(h3, is_training=is_training, decay=momentum))
        
        h4 = tf.contrib.layers.flatten(h3)
        h4 = tf.layers.dense(h4, units=1)
        return tf.nn.sigmoid(h4), h4

生成器部分

def generator(z, is_training=is_training):
    momentum = 0.9
    with tf.variable_scope('generator', reuse=None):
        d = 3
        h0 = tf.layers.dense(z, units=d * d * 512)
        h0 = tf.reshape(h0, shape=[-1, d, d, 512])
        h0 = tf.nn.relu(tf.contrib.layers.batch_norm(h0, is_training=is_training, decay=momentum))
        
        h1 = tf.layers.conv2d_transpose(h0, kernel_size=5, filters=256, strides=2, padding='same')
        h1 = tf.nn.relu(tf.contrib.layers.batch_norm(h1, is_training=is_training, decay=momentum))
        
        h2 = tf.layers.conv2d_transpose(h1, kernel_size=5, filters=128, strides=2, padding='same')
        h2 = tf.nn.relu(tf.contrib.layers.batch_norm(h2, is_training=is_training, decay=momentum))
        
        h3 = tf.layers.conv2d_transpose(h2, kernel_size=5, filters=64, strides=2, padding='same')
        h3 = tf.nn.relu(tf.contrib.layers.batch_norm(h3, is_training=is_training, decay=momentum))
        
        h4 = tf.layers.conv2d_transpose(h3, kernel_size=5, filters=1, strides=1, padding='valid', activation=tf.nn.tanh, name='g')
        return h4 

定义损失函数，注意这里实现了两个判别器，但参数是共享的

g = generator(noise)
d_real, d_real_logits = discriminator(X)
d_fake, d_fake_logits = discriminator(g, reuse=True)

vars_g = [var for var in tf.trainable_variables() if var.name.startswith('generator')]
vars_d = [var for var in tf.trainable_variables() if var.name.startswith('discriminator')]

loss_d_real = tf.reduce_mean(sigmoid_cross_entropy_with_logits(d_real_logits, tf.ones_like(d_real)))
loss_d_fake = tf.reduce_mean(sigmoid_cross_entropy_with_logits(d_fake_logits, tf.zeros_like(d_fake)))
loss_g = tf.reduce_mean(sigmoid_cross_entropy_with_logits(d_fake_logits, tf.ones_like(d_fake)))
loss_d = loss_d_real + loss_d_fake

定义优化函数，注意损失函数需要和可调参数对应上

update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
    optimizer_d = tf.train.AdamOptimizer(learning_rate=0.0002, beta1=0.5).minimize(loss_d, var_list=vars_d)
    optimizer_g = tf.train.AdamOptimizer(learning_rate=0.0002, beta1=0.5).minimize(loss_g, var_list=vars_g)

定义一个辅助函数，用于将多张图片以网格状拼在一起显示

def montage(images):
    if isinstance(images, list):
        images = np.array(images)
    img_h = images.shape[1]
    img_w = images.shape[2]
    n_plots = int(np.ceil(np.sqrt(images.shape[0])))
    m = np.ones((images.shape[1] * n_plots + n_plots + 1, images.shape[2] * n_plots + n_plots + 1)) * 0.5
    for i in range(n_plots):
        for j in range(n_plots):
            this_filter = i * n_plots + j
            if this_filter < images.shape[0]:
                this_img = images[this_filter]
                m[1 + i + i * img_h:1 + i + (i + 1) * img_h,
                  1 + j + j * img_w:1 + j + (j + 1) * img_w] = this_img
    return m

开始训练，每次迭代训练G两次

sess = tf.Session()
sess.run(tf.global_variables_initializer())
z_samples = np.random.uniform(-1.0, 1.0, [batch_size, z_dim]).astype(np.float32)
samples = []
loss = {'d': [], 'g': []}

for i in range(60000):
    n = np.random.uniform(-1.0, 1.0, [batch_size, z_dim]).astype(np.float32)
    batch = mnist.train.next_batch(batch_size=batch_size)[0]
    batch = np.reshape(batch, [-1, 28, 28, 1])
    batch = (batch - 0.5) * 2
    
    d_ls, g_ls = sess.run([loss_d, loss_g], feed_dict={X: batch, noise: n, is_training: True})
    loss['d'].append(d_ls)
    loss['g'].append(g_ls)
    
    sess.run(optimizer_d, feed_dict={X: batch, noise: n, is_training: True})
    sess.run(optimizer_g, feed_dict={X: batch, noise: n, is_training: True})
    sess.run(optimizer_g, feed_dict={X: batch, noise: n, is_training: True})
        
    if i % 1000 == 0:
        print(i, d_ls, g_ls)
        gen_imgs = sess.run(g, feed_dict={noise: z_samples, is_training: False})
        gen_imgs = (gen_imgs + 1) / 2
        imgs = [img[:, :, 0] for img in gen_imgs]
        gen_imgs = montage(imgs)
        plt.axis('off')
        plt.imshow(gen_imgs, cmap='gray')
        plt.savefig(os.path.join(OUTPUT_DIR, 'sample_%d.jpg' % i))
        plt.show()
        samples.append(gen_imgs)

plt.plot(loss['d'], label='Discriminator')
plt.plot(loss['g'], label='Generator')
plt.legend(loc='upper right')
plt.savefig('Loss.png')
plt.show()
imageio.mimsave(os.path.join(OUTPUT_DIR, 'samples.gif'), samples, fps=5)

生成的图片如下，由于损失函数中并未使用到逐像素比较，因此图形边缘不会出现模糊

保存模型，便于后续使用

saver = tf.train.Saver()
saver.save(sess, './mnist_dcgan', global_step=60000)

加载模型，如果需要的话，例如在单机上使用

# -*- coding: utf-8 -*-

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

batch_size = 100
z_dim = 100

def montage(images):
    if isinstance(images, list):
        images = np.array(images)
    img_h = images.shape[1]
    img_w = images.shape[2]
    n_plots = int(np.ceil(np.sqrt(images.shape[0])))
    m = np.ones((images.shape[1] * n_plots + n_plots + 1, images.shape[2] * n_plots + n_plots + 1)) * 0.5
    for i in range(n_plots):
        for j in range(n_plots):
            this_filter = i * n_plots + j
            if this_filter < images.shape[0]:
                this_img = images[this_filter]
                m[1 + i + i * img_h:1 + i + (i + 1) * img_h,
                  1 + j + j * img_w:1 + j + (j + 1) * img_w] = this_img
    return m

sess = tf.Session()
sess.run(tf.global_variables_initializer())

saver = tf.train.import_meta_graph('./mnist_dcgan-60000.meta')
saver.restore(sess, tf.train.latest_checkpoint('./'))

graph = tf.get_default_graph()
g = graph.get_tensor_by_name('generator/g/Tanh:0')
noise = graph.get_tensor_by_name('noise:0')
is_training = graph.get_tensor_by_name('is_training:0')

n = np.random.uniform(-1.0, 1.0, [batch_size, z_dim]).astype(np.float32)
gen_imgs = sess.run(g, feed_dict={noise: n, is_training: False})
gen_imgs = (gen_imgs + 1) / 2
imgs = [img[:, :, 0] for img in gen_imgs]
gen_imgs = montage(imgs)
plt.axis('off')
plt.imshow(gen_imgs, cmap='gray')
plt.show()

参考

• Generative Adversarial Networks：arxiv.org/abs/1406.26…
• Generative Adversarial Networks：www.iangoodfellow.com/slides/2017…
• Adversarial Nets Papers：github.com/zhangqianhu…
• DCGAN-tensorflow：github.com/carpedm20/D…

视频讲解课程

深度有趣（一）