生成对抗网络 | 实验

上期我们介绍了

同样地，我们依旧通过实验来巩固我们刚刚所学的知识点。本次实验是基于Jupyer Notebook、Anaconda Python3.7与Keras环境。数据集是利用Minst手写体图像数据集。

5.3.1 代码

1.  # chapter5/5_3_GAN.ipynb
2.  import random  
3.  import numpy as np  
4.  from keras.layers import Input  
5.  from keras.layers.core import Reshape,Dense,Dropout,Activation,Flatten  
6.  from keras.layers.advanced_activations import LeakyReLU  
7.  from keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D, Deconv2D, UpSampling2D  
8.  from keras.regularizers import *  
9.  from keras.layers.normalization import *  
10.  from keras.optimizers import *  
11.  from keras.datasets import mnist  
12.  import matplotlib.pyplot as plt  
13.  from keras.models import Model  
14.  from tqdm import tqdm  
15.  from IPython import display

1. 读取数据集

1.  img_rows, img_cols = 28, 28  
2.    
3.  # 数据集的切分与混洗（shuffle） 
4.  (X_train, y_train), (X_test, y_test) = mnist.load_data()  
5.    
6.  X_train = X_train.reshape(X_train.shape[0], 1, img_rows, img_cols)  
7.  X_test = X_test.reshape(X_test.shape[0], 1, img_rows, img_cols)  
8.  X_train = X_train.astype('float32')  
9.  X_test = X_test.astype('float32')  
10.  X_train /= 255  
11.  X_test /= 255  
12.    
13.  print(np.min(X_train), np.max(X_train))  
14.  print('X_train shape:', X_train.shape)  
15.  print(X_train.shape[0], 'train samples')  
16.  print(X_test.shape[0], 'test samples')

0.0 1.0

X_train shape: (60000, 1, 28,28)

60000 train samples

10000 test samples

2. 超参数设置

1.  shp = X_train.shape[1:]  
2.  dropout_rate = 0.25  
3.    
4.  # 优化器  
5.  opt = Adam(lr=1e-4)  
6.  dopt = Adam(lr=1e-5)

3. 定义生成器

1.  K.set_image_dim_ordering('th')  # 用theano的图片输入顺序  
2.  # 生成1 * 28 * 28的图片  
3.  nch = 200  
4.  g_input = Input(shape=[100])  
5.  H = Dense(nch*14*14, kernel_initializer='glorot_normal')(g_input)  
6.  H = BatchNormalization()(H)  
7.  H = Activation('relu')(H)  
8.  H = Reshape( [nch, 14, 14] )(H)  # 转成200 * 14 * 14  
9.  H = UpSampling2D(size=(2, 2))(H)  
10.  H = Convolution2D(100, (3, 3), padding="same", kernel_initializer='glorot_normal')(H)  
11.  H = BatchNormalization()(H)  
12.  H = Activation('relu')(H)  
13.  H = Convolution2D(50, (3, 3), padding="same", kernel_initializer='glorot_normal')(H)  
14.  H = BatchNormalization()(H)  
15.  H = Activation('relu')(H)  
16.  H = Convolution2D(1, (1, 1), padding="same", kernel_initializer='glorot_normal')(H)  
17.  g_V = Activation('sigmoid')(H)  
18.  generator = Model(g_input,g_V)  
19.  generator.compile(loss='binary_crossentropy', optimizer=opt)  
20.  generator.summary()

4. 定义辨别器

1.  # 辨别是否来自真实训练集  
2.  d_input = Input(shape=shp)  
3.  H = Convolution2D(256, (5, 5), activation="relu", strides=(2, 2), padding="same")(d_input)  
4.  H = LeakyReLU(0.2)(H)  
5.  H = Dropout(dropout_rate)(H)  
6.  H = Convolution2D(512, (5, 5), activation="relu", strides=(2, 2), padding="same")(H)  
7.  H = LeakyReLU(0.2)(H)  
8.  H = Dropout(dropout_rate)(H)  
9.  H = Flatten()(H)  
10.  H = Dense(256)(H)  
11.  H = LeakyReLU(0.2)(H)  
12.  H = Dropout(dropout_rate)(H)  
13.  d_V = Dense(2,activation='softmax')(H)  
14.  discriminator = Model(d_input,d_V)  
15.  discriminator.compile(loss='categorical_crossentropy', optimizer=dopt)  
16.  discriminator.summary()

5. 构造生成对抗网络

1.  # 冷冻训练层  
2.  def make_trainable(net, val):  
3.      net.trainable = val  
4.      for l in net.layers:  
5.          l.trainable = val  
6.  make_trainable(discriminator, False)  
7.    
8.  # 构造GAN  
9.  gan_input = Input(shape=[100])  
10.  H = generator(gan_input)  
11.  gan_V = discriminator(H)  
12.  GAN = Model(gan_input, gan_V)  
13.  GAN.compile(loss='categorical_crossentropy', optimizer=opt)  
14.  GAN.summary()

6. 训练

1.  # 描绘损失收敛过程  
2.  def plot_loss(losses):  
3.          display.clear_output(wait=True)  
4.          display.display(plt.gcf())  
5.          plt.figure(figsize=(10,8))  
6.          plt.plot(losses["d"], label='discriminitive loss')  
7.          plt.plot(losses["g"], label='generative loss')  
8.          plt.legend()  
9.          plt.show()  
10.            
11.            
12.  #  描绘生成器生成图像          
13.  def plot_gen(n_ex=16,dim=(4,4), figsize=(10,10) ):  
14.      noise = np.random.uniform(0,1,size=[n_ex,100])  
15.      generated_images = generator.predict(noise)  
16.    
17.      plt.figure(figsize=figsize)  
18.      for i in range(generated_images.shape[0]):  
19.          plt.subplot(dim[0],dim[1],i+1)  
20.          img = generated_images[i,0,:,:]  
21.          plt.imshow(img)  
22.          plt.axis('off')  
23.      plt.tight_layout()  
24.      plt.show()  
25.    
26.  # 抽取训练集样本  
27.  ntrain = 10000  
28.  trainidx = random.sample(range(0,X_train.shape[0]), ntrain)  
29.  XT = X_train[trainidx,:,:,:]    
30.    
31.  # 预训练辨别器  
32.  noise_gen = np.random.uniform(0,1,size=[XT.shape[0],100])  
33.  generated_images = generator.predict(noise_gen)  # 生成器产生样本  
34.  X = np.concatenate((XT, generated_images))    
35.  n = XT.shape[0]  
36.  y = np.zeros([2*n,2])  # 构造辨别器标签 one-hot encode  
37.  y[:n,1] = 1  
38.  y[n:,0] = 1  
39.    
40.  make_trainable(discriminator,True)  
41.  discriminator.fit(X,y, epochs=1, batch_size=32)  
42.  y_hat = discriminator.predict(X)

1.  #  计算辨别器的准确率  
2.  y_hat_idx = np.argmax(y_hat,axis=1)  
3.  y_idx = np.argmax(y,axis=1)  
4.  diff = y_idx-y_hat_idx  
5.  n_total = y.shape[0]  
6.  n_right = (diff==0).sum()  
7.    
8.  print( "(%d of %d) right"  % (n_right, n_total))

1.  def train_for_n(nb_epoch=5000, plt_frq=25,BATCH_SIZE=32):  
2.      for e in tqdm(range(nb_epoch)):    
3.            
4.          # 生成器生成样本  
5.          image_batch = X_train[np.random.randint(0,X_train.shape[0],size=BATCH_SIZE),:,:,:]      
6.          noise_gen = np.random.uniform(0,1,size=[BATCH_SIZE,100])  
7.          generated_images = generator.predict(noise_gen)  
8.            
9.          # 训练辨别器  
10.          X = np.concatenate((image_batch, generated_images))  
11.          y = np.zeros([2*BATCH_SIZE,2])  
12.          y[0:BATCH_SIZE,1] = 1  
13.          y[BATCH_SIZE:,0] = 1  
14.            
15.          # 存储辨别器损失loss  
16.          make_trainable(discriminator,True)  
17.          d_loss  = discriminator.train_on_batch(X,y)  
18.          losses["d"].append(d_loss)    
19.        
20.          # 生成器生成样本  
21.          noise_tr = np.random.uniform(0,1,size=[BATCH_SIZE,100])  
22.          y2 = np.zeros([BATCH_SIZE,2])  
23.          y2[:,1] = 1  
24.            
25.          # 存储生成器损失loss  
26.          make_trainable(discriminator,False)  # 辨别器的训练关掉  
27.          g_loss = GAN.train_on_batch(noise_tr, y2)  
28.          losses["g"].append(g_loss)  
29.            
30.          # 更新损失loss图  
31.          if e%plt_frq == plt_frq-1:  
32.              plot_loss(losses)  
33.              plot_gen()  
34.  train_for_n(nb_epoch=1000, plt_frq=10,BATCH_SIZE=128)

5.3.2 结果分析

从模型输出的loss我们可以知道生成器与辨别器两者拟合的loss并不是特别地好，因此我们可以通过调参来解决。主要调参方向有以下四点：

1. batch size

2. adam优化器的learning rate

3. 迭代次数nb_epoch

4. 生成器generator和辨别器discriminator的网络结构

5.4 小结

好了，到这里，我们就已经将生成对抗网络（GAN）的知识点讲完了。大家在掌握了整个流程之后，就可以将笔者的代码修改成自己所需要的场景，进而训练自己的GAN模型了。

最后，笔者在本章介绍的GAN只是2014年的开山之作，后面有很多人基于GAN提出了许多有趣的实验，但是所用的网络原理都差不多，这里就不一一赘述了。而且GAN的应用范围非常广阔，比如市面上很火的“换脸”软件，大多都是基于GAN的原理去做的。甚至我们也可以利用GAN去做数据增强，比如在我们缺少训练集的时候，可以考虑用GAN去生成一些数据，扩充我们的训练样本。

下一期，我们将讲授

李宏毅老师的无监督学习讲座PPT总结

敬请期待~