深度有趣 | 07 生成式對抗網路
摘要：
除VAE之外，生成式對抗網路（Generative Adversarial Nets，GAN）也是一種非常流行的無監督生成式模型
GAN中主要包括兩個核心網路
生成器（Generator）：記作G，通過對大量樣本的學習，能夠生成一些以假亂真的樣本，和VAE類似
判別器（D...
除VAE之外，生成式對抗網路（Generative Adversarial Nets，GAN）也是一種非常流行的無監督生成式模型
GAN中主要包括兩個核心網路
- 生成器（Generator）：記作G，通過對大量樣本的學習，能夠生成一些以假亂真的樣本，和VAE類似
- 判別器（Discriminator）：記作D，接受真實樣本和G生成的樣本，並進行判別和區分
- G和D相互博弈，通過學習，G的生成能力和D的判別能力都逐漸增強並收斂
GAN的訓練非常困難，有很多細節需要注意，才能生成質量較高的圖片
strides
這裡我們以 MNIST 為例，通過 TensorFlow 實現GAN，由於用到深度卷積神經網路，所以也稱作DCGAN（Deep Convolutional GAN）
原理
對於一個服從隨機分佈的噪音z，生成器通過一個複雜的對映函式生成假的樣本
\hat{x}=G(z;\theta_g)
複製程式碼
判別器則使用另一個複雜的對映函式，對於真實樣本或假的樣本，輸出一個0至1之間的值，越大表示越有可能是真實的樣本
s=D(x;\theta_d) 複製程式碼
總的目標函式如下
\min_{G}\max_{D} V(D,G)=\mathbb{E}_{x\sim p_{data}}[\log D(x)] + \mathbb{E}_{z\sim p_z}[\log(1-D(G(z)))]
複製程式碼
實現
載入庫
# -*- coding: utf-8 -*- import tensorflow as tf import numpy as np import matplotlib.pyplot as plt %matplotlib inline import os, imageio 複製程式碼
載入資料
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data')
複製程式碼
定義一些常量、網路輸入、輔助函式
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) 複製程式碼
判別器部分
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
複製程式碼
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
複製程式碼
定義損失函式，注意這裡實現了兩個判別器，但引數是共享的
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
複製程式碼
定義優化函式，注意損失函式需要和可調引數對應上
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) 複製程式碼
定義一個輔助函式，用於將多張圖片以網格狀拼在一起顯示
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 複製程式碼
開始訓練，每次迭代訓練G兩次
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)
複製程式碼
生成的圖片如下，由於損失函式中並未使用到逐畫素比較，因此圖形邊緣不會出現模糊
儲存模型，便於後續使用
saver = tf.train.Saver() saver.save(sess, './mnist_dcgan', global_step=60000) 複製程式碼
載入模型，如果需要的話，例如在單機上使用
# -*- 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()
複製程式碼