GAN生成对抗网络的Keras实现

原创文章，转载请注明： 转载自慢慢的回味

本文链接地址: GAN生成对抗网络的Keras实现

网络组成

GAN生成对抗网络包括2部分：
生成网络部分：它是一个多层神经网络结构，把一个高斯噪声通过网络生成一个图片数据，像是一个回归问题。
鉴别网络部分：它也是一个多层神经网络结构，把一个图片数据通过网络输出1或0，分辨出是不是真实图片，像是一个分类问题。
对抗的含义在于，生成网络总是想生成一个和真的一样的图片，鉴别网络部分总是想区分出谁是真实图片，谁是生成网络生成的。

理论推导网上很多，可以参考比如：GAN论文阅读——原始GAN（基本概念及理论推导）

程序分析

生成网络模型：
输入1X100的高斯分布向量 ->
256输出的全连接 -> 用ReLU进行分类 -> 正规化 ->
512输出的全连接 -> 用ReLU进行分类 -> 正规化 ->
1024输出的全连接 -> 用ReLU进行分类 -> 正规化 ->
图像大小（28X28X1）输出的全连接 -> 用tanh进行激活输出 -> 生成28X28X1的图片
返回的模型为：输入噪声，输出图片

    def build_generator(self):
 
        model = Sequential()
 
        model.add(Dense(256, input_dim=self.latent_dim))
        model.add(LeakyReLU(alpha=0.2))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Dense(512))
        model.add(LeakyReLU(alpha=0.2))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Dense(1024))
        model.add(LeakyReLU(alpha=0.2))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Dense(np.prod(self.img_shape), activation='tanh'))
        model.add(Reshape(self.img_shape))
 
        model.summary()
 
        noise = Input(shape=(self.latent_dim,))
        img = model(noise)
 
        return Model(noise, img)

鉴别网络模型：
输入28X28X1的图片 -> Flatten矩阵为1×784的向量 ->
512输出的全连接 -> 用ReLU进行分类 ->
256输出的全连接 -> 用ReLU进行分类 ->
1输出的全连接 -> 用sigmoid进行激活输出，返回1或0（1是真实的，0是伪造的）
返回的模型为：输入图片，输出真实或伪造结果

    def build_discriminator(self):
 
        model = Sequential()
 
        model.add(Flatten(input_shape=self.img_shape))
        model.add(Dense(512))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dense(256))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dense(1, activation='sigmoid'))
        model.summary()
 
        img = Input(shape=self.img_shape)
        validity = model(img)
 
        return Model(img, validity)

整个网络构建：
构建discriminator，使用交叉熵作为loss计算，Adam优化器进行网络优化。
交叉熵可以理解为：输出要拟合成目标，目标要拟合成输出之和所减少的信息量。如果信息量一样，如1和1则熵减少为0，如1和0则熵减少最大。

构建generator，然后把generator的输出传入discriminator后构建combined模型，即此模型的输入时z噪声，输出为伪造图片鉴定结果。
在这个模型中，discriminator的trainable=False，即训练过程中不改变discriminator的权值。
要知道，构建模型的时候不会生成Tensor网络，包括权值更新的Tensor节点，只有在编译compile阶段才会生成Tensor节点，即生成计算图。

    def __init__(self):
        self.img_rows = 28
        self.img_cols = 28
        self.channels = 1
        self.img_shape = (self.img_rows, self.img_cols, self.channels)
        self.latent_dim = 100
 
        optimizer = Adam(0.0002, 0.5)
 
        # Build and compile the discriminator
        self.discriminator = self.build_discriminator()
        self.discriminator.compile(loss='binary_crossentropy',
            optimizer=optimizer,
            metrics=['accuracy'])
 
        # Build the generator
        self.generator = self.build_generator()
 
        # The generator takes noise as input and generates imgs
        z = Input(shape=(self.latent_dim,))
        img = self.generator(z)
 
        # For the combined model we will only train the generator
        self.discriminator.trainable = False
 
        # The discriminator takes generated images as input and determines validity
        validity = self.discriminator(img)
 
        # The combined model  (stacked generator and discriminator)
        # Trains the generator to fool the discriminator
        self.combined = Model(z, validity)
        self.combined.compile(loss='binary_crossentropy', optimizer=optimizer)

训练逻辑：
用真实的图片imgs和真的目标valid训练discriminator，得到把真图片判断为假图片的损失量d_loss_real，即真图片应该输出1，和1的差异即为损失；
用伪造的图片gen_imgs和假的目标fake训练discriminator，得到把假图片判断为真图片的损失量d_loss_fake，即假图片应该输出0，和0的差异即为损失；
合并上面的损失为当前鉴别器的损失。

用噪音输入和真的目标valid训练combined，得到把假图片判断为假图片的损失量g_loss。
即对于生成器模型来说，它要的结果是要把假图片判断为真的结果，这样才能提高生成的伪造能力，所以输出越靠近0，和1的差异越大，损失则越大。

            # ---------------------
            #  Train Discriminator
            # ---------------------
 
            # Select a random batch of images
            idx = np.random.randint(0, X_train.shape[0], batch_size)
            imgs = X_train[idx]
 
            noise = np.random.normal(0, 1, (batch_size, self.latent_dim))
 
            # Generate a batch of new images
            gen_imgs = self.generator.predict(noise)
 
            # Train the discriminator
            d_loss_real = self.discriminator.train_on_batch(imgs, valid)
            d_loss_fake = self.discriminator.train_on_batch(gen_imgs, fake)
            d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
 
            # ---------------------
            #  Train Generator
            # ---------------------
 
            noise = np.random.normal(0, 1, (batch_size, self.latent_dim))
 
            # Train the generator (to have the discriminator label samples as valid)
            g_loss = self.combined.train_on_batch(noise, valid)

完整程序

程序需要的数据下载：
mnist
Keras GAN完整程序：

from __future__ import print_function, division
 
from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten, Dropout
from keras.layers import BatchNormalization, Activation, ZeroPadding2D
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import UpSampling2D, Conv2D
from keras.models import Sequential, Model
from keras.optimizers import Adam
 
import matplotlib.pyplot as plt
 
import sys
 
import numpy as np
 
class GAN():
    def __init__(self):
        self.img_rows = 28
        self.img_cols = 28
        self.channels = 1
        self.img_shape = (self.img_rows, self.img_cols, self.channels)
        self.latent_dim = 100
 
        optimizer = Adam(0.0002, 0.5)
 
        # Build and compile the discriminator
        self.discriminator = self.build_discriminator()
        self.discriminator.compile(loss='binary_crossentropy',
            optimizer=optimizer,
            metrics=['accuracy'])
 
        # Build the generator
        self.generator = self.build_generator()
 
        # The generator takes noise as input and generates imgs
        z = Input(shape=(self.latent_dim,))
        img = self.generator(z)
 
        # For the combined model we will only train the generator
        self.discriminator.trainable = False
 
        # The discriminator takes generated images as input and determines validity
        validity = self.discriminator(img)
 
        # The combined model  (stacked generator and discriminator)
        # Trains the generator to fool the discriminator
        self.combined = Model(z, validity)
        self.combined.compile(loss='binary_crossentropy', optimizer=optimizer)
 
 
    def build_generator(self):
 
        model = Sequential()
 
        model.add(Dense(256, input_dim=self.latent_dim))
        model.add(LeakyReLU(alpha=0.2))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Dense(512))
        model.add(LeakyReLU(alpha=0.2))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Dense(1024))
        model.add(LeakyReLU(alpha=0.2))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Dense(np.prod(self.img_shape), activation='tanh'))
        model.add(Reshape(self.img_shape))
 
        model.summary()
 
        noise = Input(shape=(self.latent_dim,))
        img = model(noise)
 
        return Model(noise, img)
 
    def build_discriminator(self):
 
        model = Sequential()
 
        model.add(Flatten(input_shape=self.img_shape))
        model.add(Dense(512))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dense(256))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dense(1, activation='sigmoid'))
        model.summary()
 
        img = Input(shape=self.img_shape)
        validity = model(img)
 
        return Model(img, validity)
 
    def train(self, epochs, batch_size=128, sample_interval=50):
 
        # Load the dataset
        (X_train, _), (_, _) = mnist.load_data()
 
        # Rescale -1 to 1
        X_train = X_train / 127.5 - 1.
        X_train = np.expand_dims(X_train, axis=3)
 
        # Adversarial ground truths
        valid = np.ones((batch_size, 1))
        fake = np.zeros((batch_size, 1))
 
        for epoch in range(epochs):
 
            # ---------------------
            #  Train Discriminator
            # ---------------------
 
            # Select a random batch of images
            idx = np.random.randint(0, X_train.shape[0], batch_size)
            imgs = X_train[idx]
 
            noise = np.random.normal(0, 1, (batch_size, self.latent_dim))
 
            # Generate a batch of new images
            gen_imgs = self.generator.predict(noise)
 
            # Train the discriminator
            d_loss_real = self.discriminator.train_on_batch(imgs, valid)
            d_loss_fake = self.discriminator.train_on_batch(gen_imgs, fake)
            d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
 
            # ---------------------
            #  Train Generator
            # ---------------------
 
            noise = np.random.normal(0, 1, (batch_size, self.latent_dim))
 
            # Train the generator (to have the discriminator label samples as valid)
            g_loss = self.combined.train_on_batch(noise, valid)
 
            # Plot the progress
            print ("%d [D loss: %f, acc.: %.2f%%] [G loss: %f]" % (epoch, d_loss[0], 100*d_loss[1], g_loss))
 
            # If at save interval => save generated image samples
            if epoch % sample_interval == 0:
                self.sample_images(epoch)
 
    def sample_images(self, epoch):
        r, c = 5, 5
        noise = np.random.normal(0, 1, (r * c, self.latent_dim))
        gen_imgs = self.generator.predict(noise)
 
        # Rescale images 0 - 1
        gen_imgs = 0.5 * gen_imgs + 0.5
 
        fig, axs = plt.subplots(r, c)
        cnt = 0
        for i in range(r):
            for j in range(c):
                axs[i,j].imshow(gen_imgs[cnt, :,:,0], cmap='gray')
                axs[i,j].axis('off')
                cnt += 1
        fig.savefig("images/%d.png" % epoch)
        plt.close()
 
 
if __name__ == '__main__':
    gan = GAN()
    gan.train(epochs=30000, batch_size=32, sample_interval=200)

附Keras支持Tensorflow 2.0的办法：
所有import tensorflow as tf的地方修改为：
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
本作品采用知识共享署名 4.0 国际许可协议进行许可。

发表回复