【训练结果很糟糕】SGAN

2020年10月16号更新：

应该是由于判别器本身的网络结构导致的(应证了猜想1)。

在训练ACGAN时，梯度也出现了问题（warning），我将判别器由卷积层换成了全连接层，再次运行就能正常训练了，详见这篇文章：

什么是SGAN

在原始GAN架构的基础上，将判别器的二分类（真实样本or生成样本）改为多分类(共N+1，N classes+fake)，便得到了SGAN( Semi-Supervised GAN) 。

此时，分类器（同时也是判别器）的表现更好，并且生成器生成的图片的质量也更高。

如何实现SGAN

生成器模型无需改动，需要改动的是判别模型以及后续的训练过程

导入相关函数

import tensorflow as tf
import matplotlib.pyplot as plt
import numpy as np
from tensorflow.keras import layers
from tensorflow.keras.utils import to_categorical

准备数据

(train_images,train_labels),(_,_)=tf.keras.datasets.mnist.load_data()

train_images.shape

(60000, 28, 28)

train_images.dtype

dtype('uint8')

train_images=train_images.reshape(train_images.shape[0],28,28,1).astype('float32')

train_images.shape

(60000, 28, 28, 1)

train_images.dtype

dtype('float32')

train_images=(train_images-127.5)/127.1#归一化

train_labels.shape

(60000,)

BATCH_SIZE=256
BUFFER_SIZE=60000

datasets=tf.data.Dataset.from_tensor_slices((train_images,train_labels))

datasets

<TensorSliceDataset shapes: ((28, 28, 1), ()), types: (tf.float32, tf.uint8)>

datasets=datasets.shuffle(BUFFER_SIZE).batch(BATCH_SIZE)

datasets

<BatchDataset shapes: ((None, 28, 28, 1), (None,)), types: (tf.float32, tf.uint8)>

搭建生成器和判别器网络

class Generator_model(tf.keras.Model):
    def __init__(self):
        super().__init__()
        
        self.dense=tf.keras.layers.Dense(7*7*256,use_bias=False)
        self.bn1=tf.keras.layers.BatchNormalization()
        self.leakyrelu1=tf.keras.layers.LeakyReLU()
        
        self.reshape=tf.keras.layers.Reshape((7,7,256))
        
        self.convT1=tf.keras.layers.Conv2DTranspose(128,(5,5),strides=(1,1),padding='same',use_bias=False)
        self.bn2=tf.keras.layers.BatchNormalization()
        self.leakyrelu2=tf.keras.layers.LeakyReLU()
        
        self.convT2=tf.keras.layers.Conv2DTranspose(64,(5,5),strides=(2,2),padding='same',use_bias=False)
        self.bn3=tf.keras.layers.BatchNormalization()
        self.leakyrelu3=tf.keras.layers.LeakyReLU()
        
        self.convT3=tf.keras.layers.Conv2DTranspose(1,(5,5),strides=(2,2),padding='same',use_bias=False,activation='tanh')
        
    def call(self,inputs,training=True):
        x=self.dense(inputs)
        x=self.bn1(x,training)
        x=self.leakyrelu1(x)
        
        x=self.reshape(x)
        
        x=self.convT1(x)
        x=self.bn2(x,training)
        x=self.leakyrelu2(x)
        
        x=self.convT2(x)
        x=self.bn3(x,training)
        x=self.leakyrelu3(x)
        
        x=self.convT3(x)
        
        return x

class Discriminator_model(tf.keras.Model):
    def __init__(self,num_classes):
        self.num_classes=num_classes
        super().__init__()
        
        self.conv1=tf.keras.layers.Conv2D(64,(5,5),strides=(2,2),padding='same')
        self.leakyrelu1=tf.keras.layers.LeakyReLU()
        self.dropout1=tf.keras.layers.Dropout(0.3)
        
        self.conv2=tf.keras.layers.Conv2D(128,(5,5),strides=(2,2),padding='same')
        self.leakyrelu2=tf.keras.layers.LeakyReLU()
        self.dropout2=tf.keras.layers.Dropout(0.3)
        
        self.flatten=tf.keras.layers.Flatten()
        
        #真实图片还是生成图片，二分类
        self.dense_valid=tf.keras.layers.Dense(1,activation='sigmoid')
        
        #分类器
        self.dense_label=tf.keras.layers.Dense(self.num_classes+1)
        self.softmax=tf.keras.layers.Softmax()
        
    def call(self,inputs,training=True):
        x=self.conv1(inputs)
        x=self.leakyrelu1(x)
        x=self.dropout1(x,training)
        
        x=self.conv2(inputs)
        x=self.leakyrelu2(x)
        x=self.dropout2(x,training)  
        
        features=self.flatten(x)#特征
        
        valid=self.dense_valid(features)#真or假
        
        label=self.dense_label(features)
        label=self.softmax(label)#属于哪一类
        
        return valid,label#真实图片还是生成图片：valid；属于哪一类（n_classes,fake）:label

定义损失函数

cross_entropy=tf.keras.losses.BinaryCrossentropy(from_logits=True)
categorical_cross_entropy=tf.keras.losses.CategoricalCrossentropy()#接受one_hot形式

def discriminator_loss(real_out,fake_out,real_pred_labels,fake_pred_labels,to_categorical_real_labels,to_categorical_fake_labels):
    #判别损失+分类损失
    real_loss=cross_entropy(tf.ones_like(real_out),real_out)+categorical_cross_entropy(real_pred_labels,to_categorical_real_labels)
    fake_loss=cross_entropy(tf.zeros_like(fake_out),fake_out)+categorical_cross_entropy(fake_pred_labels,to_categorical_fake_labels)
    return real_loss+fake_loss

def generator_loss(fake_out):
    fake_loss=cross_entropy(tf.ones_like(fake_out),fake_out)
    return fake_loss

定义优化器

generator_opt=tf.keras.optimizers.Adam(1e-4)
discriminator_opt=tf.keras.optimizers.Adam(1e-4)

设置超参数，实例化生成器和判别器

EPOCHS=100
noise_dim=100
num_examples_to_generate=16
seed=tf.random.normal([num_examples_to_generate,noise_dim])

generator=Generator_model()
discriminator=Discriminator_model(num_classes=10)

定义每个batch的训练过程

def train_step(images_one_batch,one_batch_labels):
    noise=tf.random.normal([images_one_batch.shape[0],noise_dim])#noise=seed
    with tf.GradientTape() as gen_tape,tf.GradientTape() as disc_tape:
        real_out,real_pred_labels=discriminator(images_one_batch,training=True)#真实图片送入判别器之后得到的预测标签，预测类别
        
        gen_image=generator(noise,training=True)
        fake_out,fake_pred_labels=discriminator(gen_image,training=True)#生成的假图片送入判别器之后得到的预测标签，预测类别
        
        to_categorical_real_labels=to_categorical(one_batch_labels,num_classes=10+1)
        to_categorical_fake_labels=to_categorical(np.full((images_one_batch.shape[0],1),10),num_classes=10+1)#类别10代表假，真实样本的类别为0-9
        
        #分别计算两者的损失
        gen_loss=generator_loss(fake_out)
        disc_loss=discriminator_loss(real_out,fake_out,real_pred_labels,fake_pred_labels,to_categorical_real_labels,to_categorical_fake_labels)
    
    #求可训练参数的梯度
    gradient_gen=gen_tape.gradient(gen_loss,generator.trainable_variables)
    gradient_disc=disc_tape.gradient(disc_loss,discriminator.trainable_variables)
    
    #使用优化器更新可训练参数的权值
    generator_opt.apply_gradients(zip(gradient_gen,generator.trainable_variables))
    discriminator_opt.apply_gradients(zip(gradient_disc,discriminator.trainable_variables))

定义生成图片的展示函数

#将test_noise送入gen_model，以产生假图片
def generate_plot_image(gen_model,test_noise):
    pre_images=gen_model(test_noise,training=False)#此时无需训练生成器网络
    fig=plt.figure(figsize=(4,4))
    for i in range(pre_images.shape[0]):
        plt.subplot(4,4,i+1)
        plt.imshow((pre_images[i,:,:,0]+1)/2,cmap='gray')
        plt.axis('off')
    plt.show()

定义训练函数

def train(dataset,epochs):
    for epoch in range(epochs):
        for image_batch,one_batch_labels in dataset:
            train_step(image_batch,one_batch_labels)
        generate_plot_image(generator,seed)

开始训练

train(datasets,EPOCHS)

训练结果

一直报warning：

WARNING:tensorflow:Gradients do not exist for variables ['discriminator_model/conv2d/kernel:0', 'discriminator_model/conv2d/bias:0'] when minimizing the loss.

之前使用继承自Model类的方式实现的DCGAN的代码时，如果判别器discriminator结构中的Dropout不设置training=True，也会报这个warning

问题是现在的SGAN设置了Dropout的traing=True也会报错

网上说只对某次迭代中梯度不为None的参数做梯度更新可以避免这个warning，试了下确实是这样，只需将discriminator_opt.apply_gradients(zip(gradient_disc,discriminator.trainable_variables))改为

discriminator_opt.apply_gradients([
    (grad,var)
    for (grad,var) in zip(gradient_disc,discriminator.trainable_variables)
    if grad is not None                             
    ])

但是生成的图片是惨不忍睹的

不理想原因瞎猜

• 猜想1：是不是出现了梯度爆炸或梯度消失(有文章说是这个 https://www.jiqizhixin.com/articles/2018-11-27-24 )的问题？如果是的话，考虑下改动判别器的网络结构？

  

• 猜想2：看别人用纯keras写的代码（ https://github.com/eriklindernoren/Keras-GAN/blob/master/sgan/sgan.py ）中，在train_on_batch的时候加入了class_weight来平衡真实样本和生成样本，难道是因为我没有做这一步？

这个问题有待后续再了解了，先这样。