深度有趣 | 17 CGAN和ACGAN

簡介

介紹CGAN和ACGAN的原理，經過引入額外的Condition來控制生成的圖片，並在DCGAN和WGAN的基礎上進行實現

CGAN原理

樣本x能夠包含一些屬性，或者說條件，記做y

例如MNIST中每張圖片對應的數字能夠是0至9

從一張圖來了解CGAN（Conditional GAN）的思想

生成器G從隨機噪音z和條件y生成假樣本，判別器D接受真假樣本和條件y，判斷樣本是否爲知足條件y的真實樣本

總的目標函數以下

$$ \min_{G}\max_{D} V(D,G)=\mathbb{E}{x\sim p{data}}[\log D(x|y)] + \mathbb{E}_{z\sim p_z}[\log(1-D(G(z|y)))] $$

實現

先用MNIST，在DCGAN的基礎上稍做改動以實現CGAN

加載庫

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

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

加載數據，指定one_hot=True

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

定義一些常量、網絡輸入、輔助函數，這裏加上了y_label和y_noise

batch_size = 100
z_dim = 100
WIDTH = 28
HEIGHT = 28
LABEL = 10

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

X = tf.placeholder(dtype=tf.float32, shape=[None, HEIGHT, WIDTH, 1], name='X')
y_label = tf.placeholder(dtype=tf.float32, shape=[None, HEIGHT, WIDTH, LABEL], name='y_label')
noise = tf.placeholder(dtype=tf.float32, shape=[None, z_dim], name='noise')
y_noise = tf.placeholder(dtype=tf.float32, shape=[None, LABEL], name='y_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)

判別器部分

def discriminator(image, label, reuse=None, is_training=is_training):
    momentum = 0.9
    with tf.variable_scope('discriminator', reuse=reuse):
        h0 = tf.concat([image, label], axis=3)
        h0 = lrelu(tf.layers.conv2d(h0, 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, label, is_training=is_training):
    momentum = 0.9
    with tf.variable_scope('generator', reuse=None):
        d = 3
        z = tf.concat([z, label], axis=1)
        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

損失函數

g = generator(noise, y_noise)
d_real, d_real_logits = discriminator(X, y_label)
d_fake, d_fake_logits = discriminator(g, y_label, 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

優化函數

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)

拼接圖片的函數

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

訓練模型，加入條件信息

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)
y_samples = np.zeros([batch_size, LABEL])
for i in range(LABEL):
    for j in range(LABEL):
        y_samples[i * LABEL + j, i] = 1
samples = []
loss = {'d': [], 'g': []}

for i in tqdm(range(60000)):
    n = np.random.uniform(-1.0, 1.0, [batch_size, z_dim]).astype(np.float32)
    batch, label = mnist.train.next_batch(batch_size=batch_size)
    batch = np.reshape(batch, [batch_size, HEIGHT, WIDTH, 1])
    batch = (batch - 0.5) * 2
    yn = np.copy(label)
    yl = np.reshape(label, [batch_size, 1, 1, LABEL])
    yl = yl * np.ones([batch_size, HEIGHT, WIDTH, LABEL])
    
    d_ls, g_ls = sess.run([loss_d, loss_g], feed_dict={X: batch, noise: n, y_label: yl, y_noise: yn, is_training: True})
    loss['d'].append(d_ls)
    loss['g'].append(g_ls)
    
    sess.run(optimizer_d, feed_dict={X: batch, noise: n, y_label: yl, y_noise: yn, is_training: True})
    sess.run(optimizer_g, feed_dict={X: batch, noise: n, y_label: yl, y_noise: yn, is_training: True})
    sess.run(optimizer_g, feed_dict={X: batch, noise: n, y_label: yl, y_noise: yn, is_training: True})
        
    if i % 1000 == 0:
        print(i, d_ls, g_ls)
        gen_imgs = sess.run(g, feed_dict={noise: z_samples, y_noise: y_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')
        imageio.imsave(os.path.join(OUTPUT_DIR, 'sample_%d.jpg' % i), gen_imgs)
        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)

生成的手寫數字圖片以下，每一行對應的數字相同

保存模型，便於後續使用

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

在單機上使用模型生成手寫數字圖片

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

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

batch_size = 100
z_dim = 100
LABEL = 10

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_cgan-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')
y_noise = graph.get_tensor_by_name('y_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)
y_samples = np.zeros([batch_size, LABEL])
for i in range(LABEL):
    for j in range(LABEL):
        y_samples[i * LABEL + j, i] = 1
gen_imgs = sess.run(g, feed_dict={noise: n, y_noise: y_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.show()

講條件的CelebA

瞭解CGAN的原理和實現以後，再嘗試下別的數據集，好比以前用過的CelebA

CelebA提供了每張圖片40個屬性的01標註，這裏將Male（是否爲男性）做爲條件，在WGAN的基礎上實現CGAN

加載庫

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

import tensorflow as tf
import numpy as np
import os
import matplotlib.pyplot as plt
%matplotlib inline
from imageio import imread, imsave, mimsave
import cv2
import glob
from tqdm import tqdm

加載圖片

images = glob.glob('celeba/*.jpg') 
print(len(images))

讀取圖片的Male標籤

tags = {}
target = 'Male'
with open('list_attr_celeba.txt', 'r') as fr:
    lines = fr.readlines()
    all_tags = lines[0].strip('\n').split()
    for i in range(1, len(lines)):
        line = lines[i].strip('\n').split()
        if int(line[all_tags.index(target) + 1]) == 1:
            tags[line[0]] = [1, 0] # 男
        else:
            tags[line[0]] = [0, 1] # 女
print(len(tags))
print(all_tags)

定義一些常量、網絡輸入、輔助函數

batch_size = 100
z_dim = 100
WIDTH = 64
HEIGHT = 64
LABEL = 2
LAMBDA = 10
DIS_ITERS = 3 # 5

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

X = tf.placeholder(dtype=tf.float32, shape=[batch_size, HEIGHT, WIDTH, 3], name='X')
y_label = tf.placeholder(dtype=tf.float32, shape=[batch_size, HEIGHT, WIDTH, LABEL], name='y_label')
noise = tf.placeholder(dtype=tf.float32, shape=[batch_size, z_dim], name='noise')
y_noise = tf.placeholder(dtype=tf.float32, shape=[batch_size, LABEL], name='y_noise')
is_training = tf.placeholder(dtype=tf.bool, name='is_training')

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

判別器部分

def discriminator(image, label, reuse=None, is_training=is_training):
    momentum = 0.9
    with tf.variable_scope('discriminator', reuse=reuse):
        h0 = tf.concat([image, label], axis=3)
        h0 = lrelu(tf.layers.conv2d(h0, kernel_size=5, filters=64, strides=2, padding='same'))
        
        h1 = lrelu(tf.layers.conv2d(h0, kernel_size=5, filters=128, strides=2, padding='same'))
        
        h2 = lrelu(tf.layers.conv2d(h1, kernel_size=5, filters=256, strides=2, padding='same'))
        
        h3 = lrelu(tf.layers.conv2d(h2, kernel_size=5, filters=512, strides=2, padding='same'))
        
        h4 = tf.contrib.layers.flatten(h3)
        h4 = tf.layers.dense(h4, units=1)
        return h4

生成器部分

def generator(z, label, is_training=is_training):
    momentum = 0.9
    with tf.variable_scope('generator', reuse=None):
        d = 4
        z = tf.concat([z, label], axis=1)
        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=3, strides=2, padding='same', activation=tf.nn.tanh, name='g')
        return h4

定義損失函數

g = generator(noise, y_noise)
d_real = discriminator(X, y_label)
d_fake = discriminator(g, y_label, reuse=True)

loss_d_real = -tf.reduce_mean(d_real)
loss_d_fake = tf.reduce_mean(d_fake)
loss_g = -tf.reduce_mean(d_fake)
loss_d = loss_d_real + loss_d_fake

alpha = tf.random_uniform(shape=[batch_size, 1, 1, 1], minval=0., maxval=1.)
interpolates = alpha * X + (1 - alpha) * g
grad = tf.gradients(discriminator(interpolates, y_label, reuse=True), [interpolates])[0]
slop = tf.sqrt(tf.reduce_sum(tf.square(grad), axis=[1]))
gp = tf.reduce_mean((slop - 1.) ** 2)
loss_d += LAMBDA * gp

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')]

定義優化器

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)

拼接圖片的函數

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])))
    if len(images.shape) == 4 and images.shape[3] == 3:
        m = np.ones(
            (images.shape[1] * n_plots + n_plots + 1,
             images.shape[2] * n_plots + n_plots + 1, 3)) * 0.5
    elif len(images.shape) == 4 and images.shape[3] == 1:
        m = np.ones(
            (images.shape[1] * n_plots + n_plots + 1,
             images.shape[2] * n_plots + n_plots + 1, 1)) * 0.5
    elif len(images.shape) == 3:
        m = np.ones(
            (images.shape[1] * n_plots + n_plots + 1,
             images.shape[2] * n_plots + n_plots + 1)) * 0.5
    else:
        raise ValueError('Could not parse image shape of {}'.format(images.shape))
    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

整理數據

X_all = []
Y_all = []
for i in tqdm(range(len(images))):
    image = imread(images[i])
    h = image.shape[0]
    w = image.shape[1]
    if h > w:
        image = image[h // 2 - w // 2: h // 2 + w // 2, :, :]
    else:
        image = image[:, w // 2 - h // 2: w // 2 + h // 2, :]    
    image = cv2.resize(image, (WIDTH, HEIGHT))
    image = (image / 255. - 0.5) * 2
    X_all.append(image)
    
    image_name = images[i][images[i].find('/') + 1:]
    Y_all.append(tags[image_name])

X_all = np.array(X_all)
Y_all = np.array(Y_all)
print(X_all.shape, Y_all.shape)

查看數據樣例

for i in range(10):
    plt.imshow((X_all[i, :, :, :] + 1) / 2)
    plt.show()
print(Y_all[i, :])

定義隨機產生批數據的函數

def get_random_batch():
    data_index = np.arange(X_all.shape[0])
    np.random.shuffle(data_index)
    data_index = data_index[:batch_size]
    X_batch = X_all[data_index, :, :, :]
    Y_batch = Y_all[data_index, :]
    yn = np.copy(Y_batch)
    yl = np.reshape(Y_batch, [batch_size, 1, 1, LABEL])
    yl = yl * np.ones([batch_size, HEIGHT, WIDTH, LABEL])
    
    return X_batch, yn, yl

訓練模型

sess = tf.Session()
sess.run(tf.global_variables_initializer())
zs = np.random.uniform(-1.0, 1.0, [batch_size // 2, z_dim]).astype(np.float32)
z_samples = []
y_samples = []
for i in range(batch_size // 2):
    z_samples.append(zs[i, :])
    y_samples.append([1, 0])
    z_samples.append(zs[i, :])
    y_samples.append([0, 1])
samples = []
loss = {'d': [], 'g': []}

for i in tqdm(range(60000)):
    for j in range(DIS_ITERS):
        n = np.random.uniform(-1.0, 1.0, [batch_size, z_dim]).astype(np.float32)
        X_batch, yn, yl = get_random_batch()
        _, d_ls = sess.run([optimizer_d, loss_d], feed_dict={X: X_batch, noise: n, y_label: yl, y_noise: yn, is_training: True})
    
    _, g_ls = sess.run([optimizer_g, loss_g], feed_dict={X: X_batch, noise: n, y_label: yl, y_noise: yn, is_training: True})
    
    loss['d'].append(d_ls)
    loss['g'].append(g_ls)
    
    if i % 500 == 0:
        print(i, d_ls, g_ls)
        gen_imgs = sess.run(g, feed_dict={noise: z_samples, y_noise: y_samples, is_training: False})
        gen_imgs = (gen_imgs + 1) / 2
        imgs = [img[:, :, :] for img in gen_imgs]
        gen_imgs = montage(imgs)
        plt.axis('off')
        plt.imshow(gen_imgs)
        imsave(os.path.join(OUTPUT_DIR, 'sample_%d.jpg' % i), gen_imgs)
        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()
mimsave(os.path.join(OUTPUT_DIR, 'samples.gif'), samples, fps=10)

結果以下，對於每一組圖片，噪音部分相同但條件不一樣，男左女右

保存模型

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

ACGAN

再經過一張圖瞭解ACGAN（Auxiliary Classifier GAN）的原理

和CGAN不一樣的是，C不直接輸入D。D不只須要判斷每一個樣本的真假，還須要完成一個分類任務即預測C，經過增長一個輔助分類器實現

對D而言，損失函數以下

$$ L_{adv}(D)=-\mathbb{E}{x\sim p{data}}[\log D(x)]-\mathbb{E}{z\sim p_z,c\sim p{c}}[\log(1-D(G(z,c)))] $$

$$ L_{cls}(D)=\mathbb{E}{x\sim p{data}}[L_D(c_x|x)] $$

對G而言，損失函數以下

$$ L_{adv}(G)=\mathbb{E}{z\sim p_z,c\sim p{c}}[\log(1-D(G(z,c)))] $$

$$ L_{cls}(G)=\mathbb{E}{z\sim p_z,c\sim p{c}}[L_D(c|G(z,c))] $$

仍是以CelebA的Male做爲條件，在WGAN的基礎上實現ACGAN

加載庫

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

import tensorflow as tf
import numpy as np
import os
import matplotlib.pyplot as plt
%matplotlib inline
from imageio import imread, imsave, mimsave
import cv2
import glob
from tqdm import tqdm

加載圖片

images = glob.glob('celeba/*.jpg') 
print(len(images))

讀取圖片的Male標籤

tags = {}
target = 'Male'
with open('list_attr_celeba.txt', 'r') as fr:
    lines = fr.readlines()
    all_tags = lines[0].strip('\n').split()
    for i in range(1, len(lines)):
        line = lines[i].strip('\n').split()
        if int(line[all_tags.index(target) + 1]) == 1:
            tags[line[0]] = [1, 0] # 男
        else:
            tags[line[0]] = [0, 1] # 女
print(len(tags))
print(all_tags)

定義一些常量、網絡輸入、輔助函數

batch_size = 100
z_dim = 100
WIDTH = 64
HEIGHT = 64
LABEL = 2
LAMBDA = 10
DIS_ITERS = 3 # 5

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

X = tf.placeholder(dtype=tf.float32, shape=[batch_size, HEIGHT, WIDTH, 3], name='X')
Y = tf.placeholder(dtype=tf.float32, shape=[batch_size, LABEL], name='Y')
noise = tf.placeholder(dtype=tf.float32, shape=[batch_size, 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)

判別器部分，去掉條件輸入，增長分類輸出

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 = lrelu(tf.layers.conv2d(h0, kernel_size=5, filters=128, strides=2, padding='same'))
        
        h2 = lrelu(tf.layers.conv2d(h1, kernel_size=5, filters=256, strides=2, padding='same'))
        
        h3 = lrelu(tf.layers.conv2d(h2, kernel_size=5, filters=512, strides=2, padding='same'))
        
        h4 = tf.contrib.layers.flatten(h3)
        Y_ = tf.layers.dense(h4, units=LABEL)
        h4 = tf.layers.dense(h4, units=1)
        return h4, Y_

生成器部分，沒有任何改動

def generator(z, label, is_training=is_training):
    momentum = 0.9
    with tf.variable_scope('generator', reuse=None):
        d = 4
        z = tf.concat([z, label], axis=1)
        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=3, strides=2, padding='same', activation=tf.nn.tanh, name='g')
        return h4

定義損失函數，加上分類部分對應的損失。理論上分類問題應該用交叉熵做爲損失函數，這裏使用MSE效果也不錯

g = generator(noise, Y)
d_real, y_real = discriminator(X)
d_fake, y_fake = discriminator(g, reuse=True)

loss_d_real = -tf.reduce_mean(d_real)
loss_d_fake = tf.reduce_mean(d_fake)

loss_cls_real = tf.losses.mean_squared_error(Y, y_real)
loss_cls_fake = tf.losses.mean_squared_error(Y, y_fake)

loss_d = loss_d_real + loss_d_fake + loss_cls_real
loss_g = -tf.reduce_mean(d_fake) + loss_cls_fake

alpha = tf.random_uniform(shape=[batch_size, 1, 1, 1], minval=0., maxval=1.)
interpolates = alpha * X + (1 - alpha) * g
grad = tf.gradients(discriminator(interpolates, reuse=True), [interpolates])[0]
slop = tf.sqrt(tf.reduce_sum(tf.square(grad), axis=[1]))
gp = tf.reduce_mean((slop - 1.) ** 2)
loss_d += LAMBDA * gp

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')]

定義優化器

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)

拼接圖片的函數

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])))
    if len(images.shape) == 4 and images.shape[3] == 3:
        m = np.ones(
            (images.shape[1] * n_plots + n_plots + 1,
             images.shape[2] * n_plots + n_plots + 1, 3)) * 0.5
    elif len(images.shape) == 4 and images.shape[3] == 1:
        m = np.ones(
            (images.shape[1] * n_plots + n_plots + 1,
             images.shape[2] * n_plots + n_plots + 1, 1)) * 0.5
    elif len(images.shape) == 3:
        m = np.ones(
            (images.shape[1] * n_plots + n_plots + 1,
             images.shape[2] * n_plots + n_plots + 1)) * 0.5
    else:
        raise ValueError('Could not parse image shape of {}'.format(images.shape))
    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

整理數據

X_all = []
Y_all = []
for i in tqdm(range(len(images))):
    image = imread(images[i])
    h = image.shape[0]
    w = image.shape[1]
    if h > w:
        image = image[h // 2 - w // 2: h // 2 + w // 2, :, :]
    else:
        image = image[:, w // 2 - h // 2: w // 2 + h // 2, :]    
    image = cv2.resize(image, (WIDTH, HEIGHT))
    image = (image / 255. - 0.5) * 2
    X_all.append(image)
    
    image_name = images[i][images[i].find('/') + 1:]
    Y_all.append(tags[image_name])

X_all = np.array(X_all)
Y_all = np.array(Y_all)
print(X_all.shape, Y_all.shape)

查看數據樣例

for i in range(10):
    plt.imshow((X_all[i, :, :, :] + 1) / 2)
    plt.show()
print(Y_all[i, :])

定義隨機產生批數據的函數

def get_random_batch():
    data_index = np.arange(X_all.shape[0])
    np.random.shuffle(data_index)
    data_index = data_index[:batch_size]
    X_batch = X_all[data_index, :, :, :]
    Y_batch = Y_all[data_index, :]
    
    return X_batch, Y_batch

訓練模型，根據ACGAN做相應調整

sess = tf.Session()
sess.run(tf.global_variables_initializer())
zs = np.random.uniform(-1.0, 1.0, [batch_size // 2, z_dim]).astype(np.float32)
z_samples = []
y_samples = []
for i in range(batch_size // 2):
    z_samples.append(zs[i, :])
    y_samples.append([1, 0])
    z_samples.append(zs[i, :])
    y_samples.append([0, 1])
samples = []
loss = {'d': [], 'g': []}

for i in tqdm(range(60000)):
    for j in range(DIS_ITERS):
        n = np.random.uniform(-1.0, 1.0, [batch_size, z_dim]).astype(np.float32)
        X_batch, Y_batch = get_random_batch()
        _, d_ls = sess.run([optimizer_d, loss_d], feed_dict={X: X_batch, Y: Y_batch, noise: n, is_training: True})
    
    _, g_ls = sess.run([optimizer_g, loss_g], feed_dict={X: X_batch, Y: Y_batch, noise: n, is_training: True})
    
    loss['d'].append(d_ls)
    loss['g'].append(g_ls)
    
    if i % 500 == 0:
        print(i, d_ls, g_ls)
        gen_imgs = sess.run(g, feed_dict={noise: z_samples, Y: y_samples, is_training: False})
        gen_imgs = (gen_imgs + 1) / 2
        imgs = [img[:, :, :] for img in gen_imgs]
        gen_imgs = montage(imgs)
        plt.axis('off')
        plt.imshow(gen_imgs)
        imsave(os.path.join(OUTPUT_DIR, 'sample_%d.jpg' % i), gen_imgs)
        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()
mimsave(os.path.join(OUTPUT_DIR, 'samples.gif'), samples, fps=10)

結果以下，比CGAN的結果好一些，崩掉的狀況比較少，並且人臉更真實更清晰

保存模型

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

在單機上加載模型，進行如下兩個嘗試：

• 固定噪音，漸變條件;
• 固定條件，漸變噪音。

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

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

batch_size = 100
z_dim = 100
LABEL = 2

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])))
    if len(images.shape) == 4 and images.shape[3] == 3:
        m = np.ones(
            (images.shape[1] * n_plots + n_plots + 1,
             images.shape[2] * n_plots + n_plots + 1, 3)) * 0.5
    elif len(images.shape) == 4 and images.shape[3] == 1:
        m = np.ones(
            (images.shape[1] * n_plots + n_plots + 1,
             images.shape[2] * n_plots + n_plots + 1, 1)) * 0.5
    elif len(images.shape) == 3:
        m = np.ones(
            (images.shape[1] * n_plots + n_plots + 1,
             images.shape[2] * n_plots + n_plots + 1)) * 0.5
    else:
        raise ValueError('Could not parse image shape of {}'.format(images.shape))
    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('./celeba_acgan-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')
Y = graph.get_tensor_by_name('Y:0')
is_training = graph.get_tensor_by_name('is_training:0')

# 固定噪音，漸變條件
n = np.random.uniform(-1.0, 1.0, [10, z_dim]).astype(np.float32)
ns = []
y_samples = []
for i in range(100):
    ns.append(n[i // 10, :])
    y_samples.append([i % 10 / 9, 1 - i % 10 / 9])
gen_imgs = sess.run(g, feed_dict={noise: ns, Y: y_samples, is_training: False})
gen_imgs = (gen_imgs + 1) / 2
imgs = [img[:, :, :] for img in gen_imgs]
gen_imgs = montage(imgs)
gen_imgs = np.clip(gen_imgs, 0, 1)
plt.figure(figsize=(8, 8))
plt.axis('off')
plt.imshow(gen_imgs)
plt.show()

# 固定條件，漸變噪音
n = np.random.uniform(-1.0, 1.0, [5, 2, z_dim]).astype(np.float32)
ns = []
y_samples = []
for i in range(5):
    for k in range(2):
        for j in range(10):
            start = n[i, 0, :]
            end = n[i, 1, :]
            ns.append(start + j * (end - start) / 9)
            if k == 0:
                y_samples.append([0, 1])
            else:
                y_samples.append([1, 0])
gen_imgs = sess.run(g, feed_dict={noise: ns, Y: y_samples, is_training: False})
gen_imgs = (gen_imgs + 1) / 2
imgs = [img[:, :, :] for img in gen_imgs]
gen_imgs = montage(imgs)
gen_imgs = np.clip(gen_imgs, 0, 1)
plt.figure(figsize=(8, 8))
plt.axis('off')
plt.imshow(gen_imgs)
plt.show()

由女變男的過程

人臉兩兩之間的漸變

參考

• Conditional Generative Adversarial Nets：https://arxiv.org/abs/1411.1784
• Conditional Image Synthesis With Auxiliary Classifier GANs：https://arxiv.org/abs/1610.09585

視頻講解課程

深度有趣（一）