GAN网络生成手写体数字图片

深藏阁楼爱情的钟 2023-02-13 11:37 46阅读 0赞

Keras真香，以前都是用tensorflow来写神经网络，自从用了keras,发现这个Keras也蛮方便的。  
目前感觉keras的优点就是方便搭建基于标准网络组件的神经网络，这里的网络组件包括全连接层，卷积层，池化层等等。对于需要对网络本身做创新的实验，keas可能不是很方便，还是得用tensorflow来搭建。

这篇博客，我想用Keras写一个简单的生成对抗网络。  
生成对抗网络的目标是生成手写体数字。

先看看实验的效果：  
epoch=1000的时候：  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY_size_16_color_FFFFFF_t_70]  
epoch=10000的时候：数字1已经有点像了  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY_size_16_color_FFFFFF_t_70 1]  
epoch=60000,数字1就很清晰了 ，而且其他数字也越来越清晰了  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY_size_16_color_FFFFFF_t_70 2]  
epoch=80000: 生成了5,7 啥的了。  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY_size_16_color_FFFFFF_t_70 3]  
随着训练的加深，生成的数字会越来越真实了。  
代码已经开源，项目地址：

https://github.com/jmhIcoding/GAN\_MNIST.git

# 模型原理 #

模型原理就不说了，就是使用最基础GAN结构。  
模型由一个生成器和一个鉴别器组成。  
生成器用于输入噪声，然后生成一个手写体数字图片。  
鉴别器用于判断某个输入给它的图片是不是生成器合成的。

生成器的目标是生成让鉴别器判断为非合成的图片。  
鉴别器的目标则是以尽量高的正确率分类某种图片是否为合成的。

总的原理就是这些了。  
模型的损失函数就是围绕着这两个目标来展开的。

# 模型编写 #

## 生成器 ##

__author__ = 'dk'
    #生成器
    
    import sys
    import numpy as np
    
    import  keras
    from  keras import layers
    from keras import models
    from  keras import optimizers
    from keras import losses
    
    class Generator:
        def __init__(self,height=28,width=28,channel=1,latent_space_dimension=100):
            ''' :param height: 生成图片的高,minist为28 :param width: 生成图片的宽,minist为28 :param channel: 生成器所生成的图片的通道数目,对于mnist灰度图来说,channel为1 :param latent_space_dimension: 噪声的维度 :return: '''
    
            self.latent_space_dimension = latent_space_dimension
            self.height = height
            self.width = width
            self.channel = channel
            self.generator = self.build_model()
            self.generator.summary()
        def build_model(self,block_starting_size=128,num_blocks=4):
            model = models.Sequential(name='generator')
            for i in range(num_blocks):
                if i ==0 :
                    model.add(layers.Dense(block_starting_size,input_shape=(self.latent_space_dimension,)))
                else:
                    block_size = block_starting_size * (2**i)
                    model.add(layers.Dense(block_size))
                    model.add(layers.LeakyReLU())
                    model.add(layers.BatchNormalization(momentum=0.75))
    
            model.add(layers.Dense(self.height*self.channel*self.width,activation='tanh'))
            model.add(layers.Reshape((self.width,self.height,self.channel)))
            return  model
        def summary(self):
            self.model.summary()
    
        def save_model(self):
            self.generator.save("generator.h5")

注意，generator是和整个模型一起训练的，它可以不需要compile模型。

## 鉴别器 ##

__author__ = 'dk'
    #判别器
    import sys
    import os
    import keras
    from  keras import layers
    from keras import optimizers
    from keras import models
    from keras import losses
    class Discriminator:
        def __init__(self,height=28,width=28,channel=1):
            ''' :param height: 输入图片的高 :param width: 输入图片的宽 :param channel: 输入图片的通道数 :return: '''
            self.height = height
            self.width = width
            self.channel = channel
            self.discriminator = self.build_model()
            OPTIMIZER = optimizers.Adam()
            self.discriminator = self.build_model()
            self.discriminator.compile(optimizer=OPTIMIZER,loss=losses.binary_crossentropy,metrics =['accuracy'])
            self.discriminator.summary()
        def build_model(self):
            model = models.Sequential(name='discriminator')
            model.add(layers.Flatten(input_shape=(self.width,self.height,self.channel)))
            model.add(layers.Dense(self.height*self.width*self.channel,input_shape=(self.width,self.height,self.channel)))
            model.add(layers.LeakyReLU(0.2))
            model.add(layers.Dense(self.height*self.width*self.channel//2))
            model.add(layers.LeakyReLU(0.2))
            model.add(layers.Dense(1,activation='sigmoid'))
            return model
    
        def summary(self):
            return self.discriminator.summary()
    
        def save_model(self):
            self.discriminator.save("discriminator.h5")

## gan网络 ##

把生成器和鉴别器合并起来

__author__ = 'dk'
    #生成对抗网络
    
    import keras
    from keras import layers
    from  keras import optimizers
    from  keras import  losses
    from  keras import models
    
    import  sys
    import os
    
    from Discriminator import Discriminator
    from Generator import Generator
    class GAN:
        def __init__(self,latent_space_dimension,height,width,channel):
            self.generator  = Generator(height,width,channel,latent_space_dimension)
            self.discriminator = Discriminator(height,width,channel)
            self.discriminator.discriminator.trainable = False 
            #gan部分,只训练生成器，鉴别器通过显式discriminator.train_on_batch调用来训练
            self.gan =  self.build_model()
            OPTIMIZER = optimizers.Adamax()
            self.gan.compile(optimizer = OPTIMIZER,loss = losses.binary_crossentropy)
            self.gan.summary()
        def build_model(self):
            model  = models.Sequential(name='gan')
            model.add(self.generator.generator)
            model.add(self.discriminator.discriminator)
            return  model
        def summary(self):
            self.gan.summary()
    
        def save_model(self):
            self.gan.save("gan.h5")

## 数据准备模块 ##

__author__ = 'dk'
    #数据集采集器,主要是对mnist进行简单的封装
    from keras.datasets import mnist
    import numpy as np
    def sample_latent_space(instances_number,latent_space_dimension):
        return  np.random.normal(0,1,(instances_number,latent_space_dimension))
    
    class Dator:
        def __init__(self,batch_size=None,model_type=1):
            ''' :param batch_size: :param model_type: 当model_type为-1的时候,表示0-9个数字都选;当model_type=2,说明只选择数字2 :return: '''
            self.batch_size = batch_size
            self.model_type = model_type
            with np.load("mnist.npz", allow_pickle=True) as f:
                X_train, y_train = f['x_train'], f['y_train']
                #X_test, y_test = f['x_test'], f['y_test']
            if model_type != -1:
                X_train = X_train[np.where(y_train==model_type)[0]]
            if batch_size == None:
                self.batch_size = X_train.shape[0]
            else:
                self.batch_size = batch_size
    
            self.X_train = (np.float32(X_train)-128)/128.0
            self.X_train = np.expand_dims(self.X_train,3)
    
            self.watch_index = 0
            self.train_size = self.X_train.shape[0]
        def next_batch(self,batch_size = None):
            if batch_size == None:
                batch_size  =self.batch_size
    
            X=np.concatenate([self.X_train[self.watch_index:(self.watch_index+batch_size)], self.X_train[:batch_size]])[:batch_size]
            self.watch_index  = (self.watch_index + batch_size) % self.train_size
            return  X
    
    if __name__ == '__main__':
        print(sample_latent_space(5,4))

## 训练main脚本：train.py ##

__author__ = 'dk'
    #模型训练代码
    from  GAN import GAN
    from data_utils import Dator,sample_latent_space
    import  numpy as np
    from matplotlib import pyplot as plt
    import time
    
    epochs = 50000
    height = 28
    width = 28
    channel =1
    latent_space_dimension = 100
    batch = 128
    dator = Dator(batch_size=batch,model_type=-1)
    gan = GAN(latent_space_dimension,height,width,channel)
    image_index = 0
    for i in range(epochs):
        real_img = dator.next_batch(batch_size=batch*2)
        real_label = np.ones(shape=(real_img.shape[0],1))       #真实的样本设置为1的标签
    
        noise = sample_latent_space(real_img.shape[0],latent_space_dimension)
        fake_img = gan.generator.generator.predict(noise)
        fake_label = np.zeros(shape=(fake_img.shape[0],1))      #生成器生成的假图片标注为0
    
        ###合成给gan的鉴别器的数据
        x_batch = np.concatenate([real_img,fake_img])
        y_batch = np.concatenate([real_label,fake_label])
        #训练一次
        discriminator_loss = gan.discriminator.discriminator.train_on_batch(x_batch,y_batch)[0]
        ###注意，此时训练的是鉴别器，生成器部分不动。
        ###合成训练生成器的数据
        noise = sample_latent_space(batch*2,latent_space_dimension)
        noise_labels = np.ones((batch*2,1))           
        #生成器的目标是把图片的label越来越像1
    
        generator_loss = gan.gan.train_on_batch(noise,noise_labels)
    
        print('Epoch : {0}, [Discriminator Loss:{1} ], [Generator Loss:{2}]'.format(i,discriminator_loss,generator_loss))
    
        if i!=0 and (i%50)==0:
            print('show time')
            #每50次输入16张图片看看效果
            noise = sample_latent_space(16,latent_space_dimension)
            images = gan.generator.generator.predict(noise)
            plt.figure(figsize=(10,10))
            plt.suptitle('epoch={0}'.format(i),fontsize=16)
            for index in range(images.shape[0]):
                plt.subplot(4,4,index+1)
                image  =images[index,:,:,:]
                image = image.reshape(height,width)
                plt.imshow(image,cmap='gray')
            #plt.tight_layout()
            plt.savefig("./show_time/{0}.png".format(time.time()))
            image_index += 1
            plt.close()

# 运行脚本 #

python3 train.py

即可。  
输出：

Model: "generator"
    _________________________________________________________________
    Layer (type)                 Output Shape              Param # 
    =================================================================
    dense_1 (Dense)              (None, 128)               12928     
    _________________________________________________________________
    dense_2 (Dense)              (None, 256)               33024     
    _________________________________________________________________
    leaky_re_lu_1 (LeakyReLU)    (None, 256)               0         
    _________________________________________________________________
    batch_normalization_1 (Batch (None, 256)               1024      
    _________________________________________________________________
    dense_3 (Dense)              (None, 512)               131584    
    _________________________________________________________________
    leaky_re_lu_2 (LeakyReLU)    (None, 512)               0         
    _________________________________________________________________
    batch_normalization_2 (Batch (None, 512)               2048      
    _________________________________________________________________
    dense_4 (Dense)              (None, 1024)              525312    
    _________________________________________________________________
    leaky_re_lu_3 (LeakyReLU)    (None, 1024)              0         
    _________________________________________________________________
    batch_normalization_3 (Batch (None, 1024)              4096      
    _________________________________________________________________
    dense_5 (Dense)              (None, 784)               803600    
    _________________________________________________________________
    reshape_1 (Reshape)          (None, 28, 28, 1)         0         
    =================================================================
    Total params: 1,513,616
    Trainable params: 1,510,032
    Non-trainable params: 3,584
    _________________________________________________________________
    Model: "discriminator"
    _________________________________________________________________
    Layer (type)                 Output Shape              Param # 
    =================================================================
    flatten_2 (Flatten)          (None, 784)               0         
    _________________________________________________________________
    dense_9 (Dense)              (None, 784)               615440    
    _________________________________________________________________
    leaky_re_lu_6 (LeakyReLU)    (None, 784)               0         
    _________________________________________________________________
    dense_10 (Dense)             (None, 392)               307720    
    _________________________________________________________________
    leaky_re_lu_7 (LeakyReLU)    (None, 392)               0         
    _________________________________________________________________
    dense_11 (Dense)             (None, 1)                 393       
    =================================================================
    Total params: 923,553
    Trainable params: 923,553
    Non-trainable params: 0
    _________________________________________________________________
    Model: "gan"
    _________________________________________________________________
    Layer (type)                 Output Shape              Param # 
    =================================================================
    generator (Sequential)       (None, 28, 28, 1)         1513616   
    _________________________________________________________________
    discriminator (Sequential)   (None, 1)                 923553    
    =================================================================
    Total params: 2,437,169
    Trainable params: 1,510,032
    Non-trainable params: 927,137
    _________________________________________________________________
    ····
    ···
    ··
    
    Epoch : 117754, [Discriminator Loss:0.22975191473960876 ], [Generator Loss:2.57688570022583]
    Epoch : 117755, [Discriminator Loss:0.26782122254371643 ], [Generator Loss:3.1791584491729736]
    Epoch : 117756, [Discriminator Loss:0.2609345614910126 ], [Generator Loss:2.960988998413086]
    Epoch : 117757, [Discriminator Loss:0.2673880159854889 ], [Generator Loss:2.317220687866211]
    Epoch : 117758, [Discriminator Loss:0.24904575943946838 ], [Generator Loss:1.929720401763916]
    Epoch : 117759, [Discriminator Loss:0.25158950686454773 ], [Generator Loss:2.954155683517456]
    Epoch : 117760, [Discriminator Loss:0.20324105024337769 ], [Generator Loss:3.5244760513305664]
    Epoch : 117761, [Discriminator Loss:0.2849388122558594 ], [Generator Loss:3.195873498916626]
    Epoch : 117762, [Discriminator Loss:0.19631560146808624 ], [Generator Loss:2.328411340713501]
    Epoch : 117763, [Discriminator Loss:0.20523831248283386 ], [Generator Loss:2.402683973312378]
    Epoch : 117764, [Discriminator Loss:0.2625979781150818 ], [Generator Loss:3.2176101207733154]
    Epoch : 117765, [Discriminator Loss:0.29969191551208496 ], [Generator Loss:2.9656052589416504]
    Epoch : 117766, [Discriminator Loss:0.270328551530838 ], [Generator Loss:2.3880398273468018]
    Epoch : 117767, [Discriminator Loss:0.26741161942481995 ], [Generator Loss:2.7729406356811523]
    Epoch : 117768, [Discriminator Loss:0.28797847032546997 ], [Generator Loss:2.8959264755249023]
    Epoch : 117769, [Discriminator Loss:0.30181047320365906 ], [Generator Loss:2.791097402572632]
    Epoch : 117770, [Discriminator Loss:0.26939862966537476 ], [Generator Loss:2.3666043281555176]
    Epoch : 117771, [Discriminator Loss:0.26297527551651 ], [Generator Loss:2.895970582962036]
    Epoch : 117772, [Discriminator Loss:0.21928083896636963 ], [Generator Loss:3.4627976417541504]
    Epoch : 117773, [Discriminator Loss:0.3553962707519531 ], [Generator Loss:3.2194197177886963]
    Epoch : 117774, [Discriminator Loss:0.32673510909080505 ], [Generator Loss:2.473867893218994]
    Epoch : 117775, [Discriminator Loss:0.31245478987693787 ], [Generator Loss:2.999265193939209]
    Epoch : 117776, [Discriminator Loss:0.29536381363868713 ], [Generator Loss:3.733344554901123]
    Epoch : 117777, [Discriminator Loss:0.2955515682697296 ], [Generator Loss:3.2467658519744873]
    Epoch : 117778, [Discriminator Loss:0.3677394986152649 ], [Generator Loss:1.8517814874649048]
    Epoch : 117779, [Discriminator Loss:0.31648850440979004 ], [Generator Loss:2.6385254859924316]
    Epoch : 117780, [Discriminator Loss:0.31941041350364685 ], [Generator Loss:3.350475311279297]
    Epoch : 117781, [Discriminator Loss:0.47521263360977173 ], [Generator Loss:1.9556307792663574]
    Epoch : 117782, [Discriminator Loss:0.44070643186569214 ], [Generator Loss:1.9684114456176758]

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY_size_16_color_FFFFFF_t_70]: https://img-blog.csdnimg.cn/20200528214251280.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY_size_16_color_FFFFFF_t_70 1]: https://img-blog.csdnimg.cn/20200528214329865.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY_size_16_color_FFFFFF_t_70 2]: https://img-blog.csdnimg.cn/20200528214416895.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/2020052821452380.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ptaDE5OTY=,size_16,color_FFFFFF,t_70