GAN模型生成手写字
概述:在前期的文章中,我们用TensorFlow完成了对手写数字的识别,得到了94.09%的识别准确度,效果还算不错。在这篇文章中,笔者将带领大家用GAN模型,生成我们想要的手写数字。
GAN简介
对抗性生成网络(GenerativeAdversarial Network),由 Ian Goodfellow 首先提出,由两个网络组成,分别是generator网络(用于生成)和discriminator网络(用于判别)。GAN网络的目的就是使其自己生成一副图片,比如说经过对一系列猫的图片的学习,generator网络可以自己“绘制”出一张猫的图片,且尽量真实。discriminator网络则是用来进行判断的,将一张真实的图片和一张由generator网络生成的照片同时交给discriminator网络,不断训练discriminator网络,使其可以准确将discriminator网络生成的“假图片”找出来。就这样,generator网络不断改进使其可以骗过discriminator网络,而discriminator网络不断改进使其可以更准确找到“假图片”,这种相互促进相互对抗的关系,就叫做对抗网络。图一中展示了GAN模型的结构。
思路梳理
将MNIST数据集中标签为0的图片提取出来,然后训练discriminator网络,进行手写数字0识别,接着让generator产生一张随机图片,让训练好的discriminator去识别这张生成的图片,不断训练discriminator,直到discriminator网络将生成的图片当做数字0为止。
生成“假图片”
生成一张随机像素的28*28的图片,分别进行全连接,Leaky ReLU函数激活,dropout处理(随机丢弃一些神经元,防止过拟合),全连接,tanh函数激活,最终生成一张“假图片”,TensorFlow代码如下:
1def get_generator(noise_img, n_units, out_dim, reuse=False, alpha=0.01):
2 with tf.variable_scope("generator", reuse=reuse):
3 hidden1 = tf.layers.dense(noise_img, n_units) # 全连接层
4 hidden1 = tf.maximum(alpha * hidden1, hidden1)
5 hidden1 = tf.layers.dropout(hidden1, rate=0.2)
6 logits = tf.layers.dense(hidden1, out_dim)
7 outputs = tf.tanh(logits)
8 return logits, outputs
图像判别
将需要进行判别的图片先后经过全连接,Leaky ReLU函数激活,全连接,sigmoid函数激活处理,最终输出图片的识别结果,TensorFlow代码如下:
1def get_discriminator(img, n_units, reuse=False, alpha=0.01):
2 with tf.variable_scope("discriminator", reuse=reuse):
3 hidden1 = tf.layers.dense(img, n_units)
4 hidden1 = tf.maximum(alpha * hidden1, hidden1)
5 logits = tf.layers.dense(hidden1, 1)
6 outputs = tf.sigmoid(logits)
7 return logits, outputs
完整代码
GAN手写数字识别的完整代码如下:
1import tensorflow as tf
2from tensorflow.examples.tutorials.mnist import input_data
3import matplotlib.pyplot as plt
4import numpy as np
5
6mnist = input_data.read_data_sets("E:/Tensor/MNIST_data/")
7img = mnist.train.images[50]
8
9
10def get_inputs(real_size, noise_size):
11 real_img = tf.placeholder(tf.float32, [None, real_size], name="real_img")
12 noise_img = tf.placeholder(tf.float32, [None, noise_size], name="noise_img")
13 return real_img, noise_img
14
15
16# 生成图像
17def get_generator(noise_img, n_units, out_dim, reuse=False, alpha=0.01):
18 with tf.variable_scope("generator", reuse=reuse):
19 hidden1 = tf.layers.dense(noise_img, n_units) # 全连接层
20 hidden1 = tf.maximum(alpha * hidden1, hidden1)
21 hidden1 = tf.layers.dropout(hidden1, rate=0.2)
22 logits = tf.layers.dense(hidden1, out_dim)
23 outputs = tf.tanh(logits)
24 return logits, outputs
25
26
27# 图像判别
28def get_discriminator(img, n_units, reuse=False, alpha=0.01):
29 with tf.variable_scope("discriminator", reuse=reuse):
30 hidden1 = tf.layers.dense(img, n_units)
31 hidden1 = tf.maximum(alpha * hidden1, hidden1)
32 logits = tf.layers.dense(hidden1, 1)
33 outputs = tf.sigmoid(logits)
34 return logits, outputs
35#真实图像size
36img_size = mnist.train.images[0].shape[0]
37#传入generator的噪声size
38noise_size = 100
39#生成器隐层参数
40g_units = 128
41#判别器隐层参数
42d_units = 128
43#Leaky ReLU参数
44alpha = 0.01
45#学习率
46learning_rate = 0.001
47#label smoothing
48smooth = 0.1
49tf.reset_default_graph()
50real_img, noise_img = get_inputs(img_size, noise_size)
51g_logits, g_outputs = get_generator(noise_img, g_units, img_size)
52
53d_logits_real, d_outputs_real = get_discriminator(real_img, d_units)
54d_logits_fake, d_outputs_fake = get_discriminator(g_outputs, d_units, reuse=True)
55
56d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
57 logits=d_logits_real, labels=tf.ones_like(d_logits_real)
58) * (1 - smooth))
59d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
60 logits=d_logits_fake, labels=tf.zeros_like(d_logits_fake)
61))
62d_loss = tf.add(d_loss_real, d_loss_fake)
63g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
64 logits=d_logits_fake, labels=tf.ones_like(d_logits_fake)
65) * (1 - smooth))
66
67train_vars = tf.trainable_variables()
68g_vars = [var for var in train_vars if var.name.startswith("generator")]
69d_vars = [var for var in train_vars if var.name.startswith("discriminator")]
70
71d_train_opt = tf.train.AdamOptimizer(learning_rate).minimize(d_loss, var_list=d_vars)
72g_train_opt = tf.train.AdamOptimizer(learning_rate).minimize(g_loss, var_list=g_vars)
73
74
75epochs = 10000
76samples = []
77n_sample = 10
78losses = []
79
80i = j = 0
81while i<10000:
82 if mnist.train.labels[j] == 0:
83 samples.append(mnist.train.images[j])
84 i += 1
85 j += 1
86
87print(len(samples))
88size = samples[0].size
89
90with tf.Session() as sess:
91 tf.global_variables_initializer().run()
92 for e in range(epochs):
93 batch_images = samples[e] * 2 -1
94 batch_noise = np.random.uniform(-1, 1, size=noise_size)
95
96 _ = sess.run(d_train_opt, feed_dict={real_img:[batch_images], noise_img:[batch_noise]})
97 _ = sess.run(g_train_opt, feed_dict={noise_img:[batch_noise]})
98
99 sample_noise = np.random.uniform(-1, 1, size=noise_size)
100 g_logit, g_output = sess.run(get_generator(noise_img, g_units, img_size,
101 reuse=True), feed_dict={
102 noise_img:[sample_noise]
103 })
104 print(g_logit.size)
105 g_output = (g_output+1)/2
106 plt.imshow(g_output.reshape([28, 28]), cmap='Greys_r')
107 plt.show()
训练效果
在经过了10000次的迭代后,generator网络生成的图片已经接近手写数字零的形状。
本文是对GAN模型的初次探索,在后续GAN模型的系列文章中,笔者将层层深入的去讲解GAN模型复杂的应用。