Keras【极简】GAN

1、序言

• 生成式对抗网络（Generative Adversarial Networks）至少包含2个模型：生成模型（Generative Model）和判别模型（Discriminative Model）。
• 本文用Keras实现极简的GAN，利用面向对象的思想将模型封装成3部分：图像生成器、判别者和欺诈者

2、网络结构

2.1、图像生成器

图像生成器属于欺诈者的一部分，没有编译损失函数

2.2、判别者

判别者负责判别真假，损失函数为二元交叉熵

2.3、欺诈者

欺诈者训练时，判别者不训练，损失函数为二元交叉熵

3、代码（直接复制可用）

from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten
from keras.models import Sequential, Model
import numpy as np, matplotlib.pyplot as mp, os
from keras.utils import plot_model

dir_imgs = 'imgs/'  # 图像储存文件夹
path_imgs = 'imgs/%02d.png'

img_shape = (28, 28)
noise_dim = 100  # 输入层维度

batch_size = 512  # 每次训练数量
times = 1000  # 每轮训练次数
epochs = 40  # 训练轮数

real = np.ones((batch_size, 1))  # 【正品】图像标签
fake = np.zeros((batch_size, 1))  # 【赝品】图像标签


def mkdir():
    if os.path.exists(dir_imgs):
        for fname in os.listdir(dir_imgs):
            os.remove(dir_imgs + fname)
    else:
        os.mkdir(dir_imgs)


def load_data():
    _, (x, _) = mnist.load_data()
    x = x / 127.5 - 1  # 图像预处理，映射到[-1,1]
    return x


class GAN:
    def __init__(self):
        self.generator = None  # 图像生成器
        self.discriminator = None  # 判别者
        self.welsher = None  # 欺诈者

    def init(self):
        # 建模
        self.build_generator()
        self.build_discriminator()
        self.build_welsher()
        # 模型可视化
        plot_model(self.generator, 'generator.png', show_shapes=True, show_layer_names=False)
        plot_model(self.discriminator, 'discriminator.png', show_shapes=True, show_layer_names=False)
        plot_model(self.welsher, 'welsher.png', show_shapes=True, show_layer_names=False)

    def build_generator(self):
        model = Sequential()
        model.add(Dense(250, activation='relu', input_dim=noise_dim))
        model.add(Dense(500, activation='relu'))
        model.add(Dense(np.prod(img_shape), activation='tanh'))
        model.add(Reshape(img_shape))
        self.generator = model

    def build_discriminator(self):
        model = Sequential()
        model.add(Flatten(input_shape=img_shape))
        model.add(Dense(500, activation='relu'))
        model.add(Dense(1, activation='sigmoid'))
        model.compile('adam', 'binary_crossentropy', ['acc'])
        self.discriminator = model

    def build_welsher(self):
        # 训练【欺诈者】时，不训练【判别者】
        self.discriminator.trainable = False
        # 伪造【赝品】
        noise = Input(shape=(noise_dim,))
        imgs_fake = self.generator(noise)
        # 【判别者】判别【赝品】
        discrimination = self.discriminator(imgs_fake)
        # 编译
        self.welsher = Model(noise, discrimination)
        self.welsher.compile('adam', 'binary_crossentropy')

    def train_discriminator(self, x):
        # 【正品】抽样
        idx = np.random.randint(0, x.shape[0], batch_size)
        imgs_real = x[idx]
        # 【赝品】制造
        noise = np.random.normal(0, 1, (batch_size, noise_dim))
        imgs_fake = self.generator.predict(noise)
        # 批训练
        d_loss_real = self.discriminator.train_on_batch(imgs_real, real)
        d_loss_fake = self.discriminator.train_on_batch(imgs_fake, fake)
        # 返回损失
        return np.add(d_loss_real, d_loss_fake) / 2

    def train_welsher(self):
        noise = np.random.normal(0, 1, (batch_size, noise_dim))
        return self.welsher.train_on_batch(noise, real)  # 返回损失

    def train(self, x):
        for i in range(epochs):
            loss_d, loss_w = [], []
            for _ in range(times):
                # 训练【判别者】
                loss_d.append(self.train_discriminator(x))
                # 训练【欺诈者】
                loss_w.append(self.train_welsher())
            # 训练过程展示
            print(i, '[loss_d acc_d]', np.mean(loss_d, axis=0), 'loss_w', np.mean(loss_w))
            # 保存【赝品】
            self.save_fig(i)

    def save_fig(self, epoch):
        nrows, ncols = 4, 6
        noise = np.random.normal(size=(nrows * ncols, noise_dim))
        imgs = self.generator.predict(noise)
        imgs = 0.5 * imgs + 0.5  # 预处理还原
        for i in range(nrows):
            for j in range(ncols):
                mp.subplot(nrows, ncols, i * ncols + j + 1)
                mp.imshow(imgs[i * ncols + j], cmap='gray')
                mp.axis('off')
        mp.savefig(path_imgs % epoch)
        mp.close()


if __name__ == '__main__':
    mkdir()
    gan = GAN()
    gan.init()
    x = load_data()
    gan.train(x)

4、生成的图像

本模型使用神经网络中最简单的MLP，纯粹为了方便理解GAN，不追求图像生成的效果，下面是训练3万次的效果：