Press "Enter" to skip to content

Keras深度学习实战（23）——DCGAN详解与实现

2022年9月11日

本站内容均来自兴趣收集,如不慎侵害的您的相关权益,请留言告知,我们将尽快删除.谢谢.

Keras深度学习实战（23）——DCGAN详解与实现

1. 使用 DCGAN 生成手写数字图像
2. 使用 DCGAN 生成面部图像

2.2 从零开始实现 DCGAN 生成面部图像

0. 前言

 

在 生成对抗网络 (Generative Adversarial Networks, GAN) 一节中，我们使用原始 GAN 生成了数字图片。从之前的相关学习中，我们已经知道，使用卷积神经网络 ( Convolutional Neural Network , CNN ) 体系结构能够更好地学习图像中的特征，因为 CNN 中的卷积核能够学习图像中的特定细节。深度卷积生成对抗网络 ( Deep Convolutional GAN , DCGAN ) 正是基于上述思想来生成图像，其将卷积神经网络引入 GAN 中，以代替原始 GAN 中的全连接网络。

 

1. 使用 DCGAN 生成手写数字图像

 

DCGAN 的原理与GAN 基本相同，主要区别在于 DCGAN 的生成器和鉴别器使用卷积神经网络体系结构：

 

def generator():
    model = Sequential()
    model.add(Dense(1024, input_dim=100))
    model.add(LeakyReLU(0.2))
    model.add(Dense(128*7*7))
    model.add(BatchNormalization())
    model.add(LeakyReLU(0.2))
    model.add(Reshape((7,7,128)))
    model.add(Conv2DTranspose(128, (5, 5), strides=(1, 1), padding='same', use_bias=False))
    model.add(BatchNormalization())
    model.add(LeakyReLU(0.2))
    model.add(Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False))
    model.add(BatchNormalization())
    model.add(LeakyReLU(0.2))
    model.add(Conv2DTranspose(1, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh'))
    return model
def discriminator():
    model = Sequential()
    model.add(Conv2D(64, (3, 3), padding='same', input_shape=(28,28,1)))
    model.add(LeakyReLU(0.2))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Conv2D(128, (3,3)))
    model.add(LeakyReLU(0.2))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Conv2D(512, (3,3)))
    model.add(LeakyReLU(0.2))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Flatten())
    model.add(Dense(512))
    model.add(LeakyReLU(0.2))
    model.add(Dense(1))
    model.add(Activation('sigmoid'))
    return model

 

在 DCGAN 中，我们对输入数据执行多个卷积和池化操作，而其它步骤与原始 GAN 中的步骤完全相同，区别仅在于使用卷积和池化架构定义的 GAN 模型，模型训练后生成的图像如下所示：

 

 

生成器和鉴别器训练损失值随 epoch 的增加而变化，如下所示：

 

 

与原始 GAN相比，可以看到，尽管其他条件保持不变，仅模型体系结构发生了变化，但是通过 DCGAN 生成的图像比原始 GAN 的图像更加真实。

 

2. 使用 DCGAN 生成面部图像

 

我们已经学习了如何使用 DCGAN 生成新手写数字图像。本节中，我们将继续学习如何从现有的面部数据集中生成一组新的面部图像。所用的面部数据集与在性别分类中所用数据集相同。

 

2.1 模型分析

 

在介绍了 DCGAN 的核心思想后，我们将在本节中使用 DCGAN 架构生成面部图像，所用策略如下：

使用包含面部图像的数据集
生成器开始时只能生成随机图像
通过向鉴别器输入真实面部图像和生成器生成图像来训练鉴别器，鉴别器应学会区分真实面部图像和生成的虚假面部图像
鉴别器模型训练完成后，将冻结其权重，并训练生成器网络，以使鉴别器以较高的概率将生成的虚假图片识别为真实图像
多次迭代以上前两个步骤，直到生成器能够生成足够逼真的图像

2.2 从零开始实现 DCGAN 生成面部图像

 

(1)下载数据集，与在性别分类中所用数据集相同，图像样本如下：

 

(2) 定义模型架构与所用超参数：

 

shape = (56, 56, 3)
epochs = 10000
batch_size = 64
save_interval = 100
def generator():
    model = Sequential()
    model.add(Dense(1024, input_dim=100))
    model.add(LeakyReLU(0.2))
    model.add(Dense(128*7*7))
    model.add(BatchNormalization())
    model.add(LeakyReLU(0.2))
    model.add(Reshape((7,7,128)))
    model.add(Conv2DTranspose(128, (3, 3), strides=(1, 1), padding='same', use_bias=False))
    model.add(BatchNormalization())
    model.add(LeakyReLU(0.2))
    model.add(Conv2DTranspose(64, (3, 3), strides=(2, 2), padding='same', use_bias=False))
    model.add(BatchNormalization())
    model.add(LeakyReLU(0.2))
    model.add(Conv2DTranspose(32, (3, 3), strides=(2, 2), padding='same', use_bias=False))
    model.add(BatchNormalization())
    model.add(LeakyReLU(0.2))
    model.add(Conv2DTranspose(3, (3, 3), strides=(2, 2), padding='same', use_bias=False, activation='tanh'))
    return model
def discriminator():
    model = Sequential()
    model.add(Conv2D(64, (3, 3), padding='same', input_shape=(56,56,3)))
    model.add(LeakyReLU(0.2))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Conv2D(128, (3,3)))
    model.add(LeakyReLU(0.2))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Conv2D(512, (3,3)))
    model.add(LeakyReLU(0.2))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Conv2D(1024, (3,3)))
    model.add(LeakyReLU(0.2))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Flatten())
    model.add(Dense(1024))
    model.add(LeakyReLU(0.2))
    model.add(Dense(1))
    model.add(Activation('sigmoid'))
    return model
def gan(discriminator, generator):
    discriminator.trainable = False
    model = Sequential()
    model.add(generator)
    model.add(discriminator)
    return model

 

(3)定义预处理函数、反处理函数以及绘制图像等实用函数：

 

import time
noise = np.random.normal(0, 1, (16, 100))
def plot_images(noise=noise, samples=16, step=0):
    images = generator.predict(noise)
    images = deprocess(images)
    images = np.clip(images, 0, 255)
    plt.figure(figsize=(10, 10))
    for i in range(images.shape[0]):
        plt.subplot(4, 4, i + 1)
        image = images[i, :, :, :]
        image = np.reshape(image, [56, 56,3])
        plt.imshow(image)
        plt.axis('off')
    plt.tight_layout()
    plt.show()

 

在此模型中，我们在将图像调整为较小的尺寸，这是由于 DCGAN 并不适合生成较大尺寸的图片，同时也可以减少模型的参数量：

 

def preprocess(x):
    return (x / 255) * 2 - 1
def deprocess(x):
return np.uint8((x + 1) / 2 * 255)

 

(4)导入数据集，创建输入数据将其转换为数组，并对其进行预处理：

 

import os
from glob import glob
root_dir = 'man_woman/b_resized/'
all_img = glob(os.path.join(root_dir, '*.jpg'))
x_train = []
for i in range(len(all_img)):
    img = cv2.imread(all_img[i])#, cv2.COLOR_BGR2RGB)
    img = cv2.resize(img, (56, 56))
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = preprocess(img)
    x_train.append(img)
x_train = np.array(x_train)

 

(5)编译生成器，鉴别器和 GCGAN 模型：

 

generator = generator()
generator.summary()
generator.compile(loss='binary_crossentropy', optimizer=Adam(lr=0.0002, beta_1=0.5, decay=8e-8))
discriminator = discriminator()
discriminator.summary()
discriminator.compile(loss='binary_crossentropy', optimizer=Adam(lr=0.0002, beta_1=0.5, decay=8e-8), metrics=['acc'])
gan = gan(discriminator, generator)
gan.compile(loss='binary_crossentropy', optimizer=Adam(lr=0.0002, beta_1=0.5, decay=8e-8))

 

生成器模型简要的架构信息如下所示：

 

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                (None, 1024)              103424    
_________________________________________________________________
leaky_re_lu (LeakyReLU)      (None, 1024)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 6272)              6428800   
_________________________________________________________________
batch_normalization (BatchNo (None, 6272)              25088     
_________________________________________________________________
leaky_re_lu_1 (LeakyReLU)    (None, 6272)              0         
_________________________________________________________________
reshape (Reshape)            (None, 7, 7, 128)         0         
_________________________________________________________________
conv2d_transpose (Conv2DTran (None, 7, 7, 128)         147456    
_________________________________________________________________
batch_normalization_1 (Batch (None, 7, 7, 128)         512       
_________________________________________________________________
leaky_re_lu_2 (LeakyReLU)    (None, 7, 7, 128)         0         
_________________________________________________________________
conv2d_transpose_1 (Conv2DTr (None, 14, 14, 64)        73728     
_________________________________________________________________
batch_normalization_2 (Batch (None, 14, 14, 64)        256       
_________________________________________________________________
leaky_re_lu_3 (LeakyReLU)    (None, 14, 14, 64)        0         
_________________________________________________________________
conv2d_transpose_2 (Conv2DTr (None, 28, 28, 32)        18432     
_________________________________________________________________
batch_normalization_3 (Batch (None, 28, 28, 32)        128       
_________________________________________________________________
leaky_re_lu_4 (LeakyReLU)    (None, 28, 28, 32)        0         
_________________________________________________________________
conv2d_transpose_3 (Conv2DTr (None, 56, 56, 3)         864       
=================================================================
Total params: 6,798,688
Trainable params: 6,785,696
Non-trainable params: 12,992
_________________________________________________________________

 

鉴别器模型简要的架构信息如下所示：

 

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 56, 56, 64)        1792      
_________________________________________________________________
batch_normalization_4 (Batch (None, 56, 56, 64)        256       
_________________________________________________________________
leaky_re_lu_5 (LeakyReLU)    (None, 56, 56, 64)        0         
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 26, 26, 128)       73856     
_________________________________________________________________
batch_normalization_5 (Batch (None, 26, 26, 128)       512       
_________________________________________________________________
leaky_re_lu_6 (LeakyReLU)    (None, 26, 26, 128)       0         
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 13, 13, 128)       0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 11, 11, 512)       590336    
_________________________________________________________________
batch_normalization_6 (Batch (None, 11, 11, 512)       2048      
_________________________________________________________________
leaky_re_lu_7 (LeakyReLU)    (None, 11, 11, 512)       0         
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 5, 5, 512)         0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 3, 3, 1024)        4719616   
_________________________________________________________________
batch_normalization_7 (Batch (None, 3, 3, 1024)        4096      
_________________________________________________________________
leaky_re_lu_8 (LeakyReLU)    (None, 3, 3, 1024)        0         
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 1, 1, 1024)        0         
_________________________________________________________________
flatten (Flatten)            (None, 1024)              0         
_________________________________________________________________
dense_2 (Dense)              (None, 1024)              1049600   
_________________________________________________________________
leaky_re_lu_9 (LeakyReLU)    (None, 1024)              0         
_________________________________________________________________
dense_3 (Dense)              (None, 1)                 1025      
_________________________________________________________________
activation (Activation)      (None, 1)                 0         
=================================================================
Total params: 6,443,137
Trainable params: 6,439,681
Non-trainable params: 3,456
_________________________________________________________________

 

(6)接下来，对模型训练多个 epoch ，具体训练过程与训练原始 GAN相同：

 

disc_loss = []
gen_loss = []
for cnt in range(epochs):
    random_index = np.random.randint(0, len(x_train) - batch_size / 2)
    legit_images = x_train[random_index: random_index + batch_size // 2].reshape(batch_size // 2, 56, 56, 3)
    gen_noise = np.random.normal(0, 1, (batch_size // 2, 100))/2
    synthetic_images = generator.predict(gen_noise)
    
    x_combined_batch = np.concatenate((legit_images, synthetic_images))
    y_combined_batch = np.concatenate((np.ones((batch_size // 2, 1)), np.zeros((batch_size // 2, 1))))
    d_loss = discriminator.train_on_batch(x_combined_batch, y_combined_batch)
    noise = np.random.normal(0, 1, (batch_size*2, 100))/2
    y_mislabled = np.ones((batch_size*2, 1))
    
    g_loss = gan.train_on_batch(noise, y_mislabled)
    disc_loss.append(d_loss[0])
    gen_loss.append(g_loss)
    print('epoch: {}, [Discriminator: {}], [Generator: {}]'.format(cnt, d_loss[0], g_loss))
    if cnt % 500 == 0:
        plot_images(step=cnt)

 

以上代码生成的图像如下所示：

 

 

如上图所示，尽管这些图像看起来仍然不够逼真，但仍表现出一定的有效性，可以通过更改模型体系结构或增加网络深度对 GAN 进行改进。鉴别器和生成器的损失值随 epoch 的增加的变化情况如下：

 

 

深度卷积生成对抗网络 ( Deep Convolutional GAN , DCGAN ) 将卷积神经网络引入 GAN 中，以代替原始 GAN 中的全连接网络更好地学习图像中的特征。本节中，首先介绍了 DCGAN 的基本思想，然后使用 Keras 从零开始实现用于生成手写数字图片的 DCGAN 模型，可以看到生成的图像效果比原始 GAN 更逼真，最后为了充分验证 DCGAN 模型性能，使用 DCGAN 生成复杂的面部图像。

 

Categories:机器学习

Be First to Comment

发表回复

您的电子邮箱地址不会被公开。 必填项已用*标注