生成对抗网络简介
生成对抗网络（GAN）启发自博弈论中的二人零和博弈（two-player game），类似于周伯通的绝学——“左右互搏”。GAN 模型中的两位博弈方分别由生成式模型（generative model）和判别式模型（discriminative model）充当。生成模型 G 捕捉样本数据的分布，用服从某一分布（均匀分布，高斯分布等）的噪声 z 生成一个类似真实训练数据的样本，追求效果是越像真实样本越好；判别模型 D 是一个二分类器，估计一个样本来自于训练数据（而非生成数据）的概率，如果样本来自于真实的训练数据，D 输出大概率，否则，D 输出小概率。可以做如下类比：生成网络 G 好比假币制造团伙，专门制造假币，判别网络 D 好比警察，专门检测使用的货币是真币还是假币，G 的目标是想方设法生成和真币一样的货币，使得 D 判别不出来，D 的目标是想方设法检测出来 G 生成的假币。随着训练时间的增加，判别模型与生成模型的能力都相应的提升！
具体生成网络的示意图如下所示：
[图片上传失败...(image-587d3a-1551263869078)]
Tensorflow生成对抗网络实现
from __future__ import division, print_function, absolute_import
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
导入数据集
# 导入mnist数据集
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("./data/", one_hot=True)
Extracting ./data/train-images-idx3-ubyte.gz
Extracting ./data/train-labels-idx1-ubyte.gz
Extracting ./data/t10k-images-idx3-ubyte.gz
Extracting ./data/t10k-labels-idx1-ubyte.gz
参数设置
# Training Params
num_steps = 70000 #总迭代次数
batch_size = 128 # 批量大小
learning_rate = 0.0002 #学习率
# Network Params
image_dim = 784 # 28*28 pixels，生成器的输出层节点数，也是判别器的输入
gen_hidden_dim = 256 # 生成器隐藏层节点数
disc_hidden_dim = 256 # 判别器隐藏层节点数
noise_dim = 100 # Noise data points 生成器输入节点数
# Xavier 初始化方式（更适合有ReLU的网络训练）
def glorot_init(shape):
return tf.random_normal(shape=shape, stddev=1. / tf.sqrt(shape[0] / 2.))
Xavier 初始化方式方差：
image
这里的参数是标准差。
设置每一层的权重与偏置
# 设置每一层的权重（Xavier初始化）与偏置（初始化为零）
weights = {
'gen_hidden1': tf.Variable(glorot_init([noise_dim, gen_hidden_dim])),#（100 - 256）
'gen_out': tf.Variable(glorot_init([gen_hidden_dim, image_dim])), #（256 - 784）
'disc_hidden1': tf.Variable(glorot_init([image_dim, disc_hidden_dim])),#（784 - 256）
'disc_out': tf.Variable(glorot_init([disc_hidden_dim, 1])),#（256 - 1）
}
biases = {
'gen_hidden1': tf.Variable(tf.zeros([gen_hidden_dim])),
'gen_out': tf.Variable(tf.zeros([image_dim])),
'disc_hidden1': tf.Variable(tf.zeros([disc_hidden_dim])),
'disc_out': tf.Variable(tf.zeros([1])),
}
定义生成对抗网络
# 定义生成器函数
def generator(x):
# 输入x是1x100的矩阵，weights['gen_hidden1']是100x256的矩阵，矩阵相乘结果是1x256的矩阵，生成器隐藏层含256个节点
hidden_layer = tf.matmul(x, weights['gen_hidden1'])
# biases['gen_hidden1']是1x256的矩阵，生成器隐藏层含256个节点
hidden_layer = tf.add(hidden_layer, biases['gen_hidden1'])
# 激活函数 relu
hidden_layer = tf.nn.relu(hidden_layer)
# hidden_layer是1x256的矩阵，weights['gen_out']是256x784的矩阵，矩阵相乘结果是1x784的矩阵，生成器输出层含784个节点
out_layer = tf.matmul(hidden_layer, weights['gen_out'])
# biases['gen_out']是1x784的矩阵，生成器输出层含784个节点
out_layer = tf.add(out_layer, biases['gen_out'])
# 激活函数 sigmoid
out_layer = tf.nn.sigmoid(out_layer)
return out_layer
# 定义判别器函数
def discriminator(x):
# 输入x是生成器生成的1x784的矩阵（生成的图片），weights['disc_hidden1']是784x256的矩阵，矩阵相乘结果是1x256的矩阵，判别器隐藏层含256个节点
hidden_layer = tf.matmul(x, weights['disc_hidden1'])
# biases['disc_hidden1']是1x256的矩阵，生成器隐藏层含256个节点
hidden_layer = tf.add(hidden_layer, biases['disc_hidden1'])
# 激活函数 relu
hidden_layer = tf.nn.relu(hidden_layer)
# hidden_layer是1x256的矩阵，weights['disc_out']是256x1的矩阵，矩阵相乘结果是一个数，判别器输出层含1个节点
out_layer = tf.matmul(hidden_layer, weights['disc_out'])
# biases['disc_out']是一个数，判别器输出层含1个节点
out_layer = tf.add(out_layer, biases['disc_out'])
# 激活函数 sigmoid
out_layer = tf.nn.sigmoid(out_layer)
return out_layer
# 构建网络
# 网络输入
gen_input = tf.placeholder(tf.float32, shape=[None, noise_dim], name='input_noise') # 生成器 输入噪点 batch*100，none是一个空值，后面赋值batch_size
disc_input = tf.placeholder(tf.float32, shape=[None, image_dim], name='disc_input') # 判别器 输入真实图像 batch*784
# 构建生成器（generator）
gen_sample = generator(gen_input)
# 构建两个判别器（一个是真实图像输入，一个是生成图像）
disc_real = discriminator(disc_input) # 真实图像
disc_fake = discriminator(gen_sample) # 通过生成器生成的图像
# 创建损失函数
# 关于GAN的理论推导，可参见 [^1]
gen_loss = -tf.reduce_mean(tf.log(disc_fake)) # 生成器损失函数
disc_loss = -tf.reduce_mean(tf.log(disc_real) + tf.log(1. - disc_fake)) # 判别器损失函数
# 创建优化器（采用Adam方法），可参见 [^2]
optimizer_gen = tf.train.AdamOptimizer(learning_rate=learning_rate)
optimizer_disc = tf.train.AdamOptimizer(learning_rate=learning_rate)
# Training Variables for each optimizer
# By default in TensorFlow, all variables are updated by each optimizer, so we
# need to precise for each one of them the specific variables to update.
# 生成网络的变量
gen_vars = [weights['gen_hidden1'], weights['gen_out'],
biases['gen_hidden1'], biases['gen_out']]
# 判别网络的变量
disc_vars = [weights['disc_hidden1'], weights['disc_out'],
biases['disc_hidden1'], biases['disc_out']]
# 创建训练操作
train_gen = optimizer_gen.minimize(gen_loss, var_list=gen_vars)
train_disc = optimizer_disc.minimize(disc_loss, var_list=disc_vars)
# 变量全局初始化
init = tf.global_variables_initializer()
GAN的网络结构类似于多层感知机：
image
训练生成对抗网络
# 开始训练
# 创建一个会话
sess = tf.Session()
# 初始化
sess.run(init)
# 训练
for i in range(1, num_steps+1):
# 准备数据
# 拿到下一批次的 MNIST 数据 (仅需要图像, 不需要标签)
batch_x, _ = mnist.train.next_batch(batch_size) # 判别器输入 真实图像，batch_*784
# 给生成器生成噪点数据
z = np.random.uniform(-1., 1., size=[batch_size, noise_dim]) # 生成器输入 噪声，batch*100
# 训练
feed_dict = {disc_input: batch_x, gen_input: z} #给placeholder填入值
_, _, gl, dl = sess.run([train_gen, train_disc, gen_loss, disc_loss],
feed_dict=feed_dict)
if i % 2000 == 0 or i == 1:
print('Step %i: Generator Loss: %f, Discriminator Loss: %f' % (i, gl, dl))
Step 1: Generator Loss: 0.223592, Discriminator Loss: 2.090910
Step 2000: Generator Loss: 4.678916, Discriminator Loss: 0.041115
Step 4000: Generator Loss: 3.605874, Discriminator Loss: 0.068698
Step 6000: Generator Loss: 3.845584, Discriminator Loss: 0.190420
Step 8000: Generator Loss: 4.470613, Discriminator Loss: 0.117488
Step 10000: Generator Loss: 3.813103, Discriminator Loss: 0.146255
Step 12000: Generator Loss: 2.991248, Discriminator Loss: 0.392258
Step 14000: Generator Loss: 3.769275, Discriminator Loss: 0.153639
Step 16000: Generator Loss: 4.366917, Discriminator Loss: 0.206618
Step 18000: Generator Loss: 4.052875, Discriminator Loss: 0.225112
Step 20000: Generator Loss: 3.574747, Discriminator Loss: 0.362798
Step 22000: Generator Loss: 3.760236, Discriminator Loss: 0.188211
Step 24000: Generator Loss: 3.055995, Discriminator Loss: 0.354645
Step 26000: Generator Loss: 3.619049, Discriminator Loss: 0.211489
Step 28000: Generator Loss: 3.523777, Discriminator Loss: 0.273607
Step 30000: Generator Loss: 3.889854, Discriminator Loss: 0.286803
Step 32000: Generator Loss: 3.106094, Discriminator Loss: 0.298111
Step 34000: Generator Loss: 3.548391, Discriminator Loss: 0.343262
Step 36000: Generator Loss: 3.081174, Discriminator Loss: 0.332788
Step 38000: Generator Loss: 2.946176, Discriminator Loss: 0.335102
Step 40000: Generator Loss: 3.078653, Discriminator Loss: 0.465524
Step 42000: Generator Loss: 2.601799, Discriminator Loss: 0.409574
Step 44000: Generator Loss: 3.168177, Discriminator Loss: 0.325075
Step 46000: Generator Loss: 2.601811, Discriminator Loss: 0.428143
Step 48000: Generator Loss: 2.853810, Discriminator Loss: 0.403768
Step 50000: Generator Loss: 2.690175, Discriminator Loss: 0.483180
Step 52000: Generator Loss: 3.278867, Discriminator Loss: 0.375016
Step 54000: Generator Loss: 2.869437, Discriminator Loss: 0.477840
Step 56000: Generator Loss: 2.561056, Discriminator Loss: 0.449300
Step 58000: Generator Loss: 2.814199, Discriminator Loss: 0.484522
Step 60000: Generator Loss: 2.469474, Discriminator Loss: 0.428359
Step 62000: Generator Loss: 2.721684, Discriminator Loss: 0.494090
Step 64000: Generator Loss: 2.491284, Discriminator Loss: 0.654795
Step 66000: Generator Loss: 2.725388, Discriminator Loss: 0.423149
Step 68000: Generator Loss: 2.758215, Discriminator Loss: 0.513224
Step 70000: Generator Loss: 3.072056, Discriminator Loss: 0.481437
测试
# 测试
# 通过训练出的生成网络输入噪点，生成图像
n = 6
canvas = np.empty((28 * n, 28 * n))
for i in range(n):
# 噪点输入
z = np.random.uniform(-1., 1., size=[n, noise_dim])
# 生成图像
g = sess.run(gen_sample, feed_dict={gen_input: z})
# 颜色反转便于显示
g = -1 * (g - 1)
for j in range(n):
# 绘制生成的手写体数字
canvas[i * 28:(i + 1) * 28, j * 28:(j + 1) * 28] = g[j].reshape([28, 28])
plt.figure(figsize=(n, n))
plt.imshow(canvas, origin="upper", cmap="gray")
plt.show()
image
参考
[1] 机器之心GitHub项目：GAN完整理论推导与实现，Perfect！
