代码实践 | 自编码器

1.  # /chapter4/4_7_5_AutoEncoder.ipynb
2.  import keras  
3.  import numpy as np  
4.  import matplotlib.pyplot as plt  
5.  from keras.datasets import mnist  
6.  from keras.models import Model  
7.  from keras.layers import Input, add  
8.  from keras.layers.core import Layer, Dense, Dropout, Activation, Flatten, Reshape  
9.  from keras import regularizers  
10.  from keras.regularizers import l2  
11.  from keras.layers.convolutional import Conv2D, MaxPooling2D, UpSampling2D, ZeroPadding2D  
12.  from keras.utils import np_utils

1) 读取手写体数据及与图像预处理

1.  (X_train, _), (X_test, _) = mnist.load_data()  
2.    
3.  #  归一化  
4.  X_train = X_train.astype("float32")/255.  
5.  X_test = X_test.astype("float32")/255.  
6.    
7.  print('X_train shape:', X_train.shape)  
8.  print(X_train.shape[0], 'train samples')  
9.  print(X_test.shape[0], 'test samples')

X_train shape: (60000, 28, 28)

60000 train samples

10000 test samples

1.  # np.prod是将28X28矩阵转化成1X784,方便全连接神经网络输入层784个神经元读取。
2.  X_train = X_train.reshape((len(X_train), np.prod(X_train.shape[1:])))  
3.  X_test = X_test.reshape((len(X_test), np.prod(X_test.shape[1:])))

2) 构建自编码器模型

1.  input_size = 784  
2.  hidden_size = 64  
3.  output_size = 784  
4.    
5.  x = Input(shape=(input_size,))  
6.  h = Dense(hidden_size, activation='relu')(x)  
7.  r = Dense(output_size, activation='sigmoid')(h)  
8.    
9.  autoencoder = Model(inputs=x, outputs=r)  
10.  autoencoder.compile(optimizer='adam', loss='mse')

3) 训练

1.  epochs = 5  
2.  batch_size = 128  
3.    
4.  history = autoencoder.fit(X_train, X_train,   
5.                            batch_size=batch_size,   
6.                            epochs=epochs, verbose=1,   
7.                            validation_data=(X_test, X_test)  
8.                           )

4) 查看自编码器的压缩效果

1.  conv_encoder = Model(x, h)  # 只取编码器做模型  
2.  encoded_imgs = conv_encoder.predict(X_test)  
3.    
4.  # 打印10张测试集手写体的压缩效果  
5.  n = 10  
6.  plt.figure(figsize=(20, 8))  
7.  for i in range(n):  
8.      ax = plt.subplot(1, n, i+1)  
9.      plt.imshow(encoded_imgs[i].reshape(4, 16).T)  
10.      plt.gray()  
11.      ax.get_xaxis().set_visible(False)  
12.      ax.get_yaxis().set_visible(False)  
13.  plt.show()

5) 查看自编码器的解码效果

1.  decoded_imgs = autoencoder.predict(X_test)  
2.  n = 10  
3.  plt.figure(figsize=(20, 6))  
4.  for i in range(n):  
5.      # 打印原图  
6.      ax = plt.subplot(3, n, i+1)  
7.      plt.imshow(X_test[i].reshape(28, 28))  
8.      plt.gray()  
9.      ax.get_xaxis().set_visible(False)  
10.      ax.get_yaxis().set_visible(False)  
11.    
12.        
13.      # 打印解码图  
14.      ax = plt.subplot(3, n, i+n+1)  
15.      plt.imshow(decoded_imgs[i].reshape(28, 28))  
16.      plt.gray()  
17.      ax.get_xaxis().set_visible(False)  
18.      ax.get_yaxis().set_visible(False)  
19.        
20.  plt.show()

6) 训练过程可视化

1.  print(history.history.keys())  
2.    
3.  plt.plot(history.history['loss'])  
4.  plt.plot(history.history['val_loss'])  
5.  plt.title('model loss')  
6.  plt.ylabel('loss')  
7.  plt.xlabel('epoch')  
8.  plt.legend(['train', 'validation'], loc='upper right')  
9.  plt.show()

dict_keys(['val_loss', 'loss'])

2. 多层自编码器

1) 多层自编码器建模

1.  input_size = 784  
2.  hidden_size = 128  
3.  code_size = 64  
4.    
5.  x = Input(shape=(input_size,))  
6.  hidden_1 = Dense(hidden_size, activation='relu')(x)  
7.  h = Dense(code_size, activation='relu')(hidden_1)  
8.  hidden_2 = Dense(hidden_size, activation='relu')(h)  
9.  r = Dense(input_size, activation='sigmoid')(hidden_2)  
10.    
11.  autoencoder = Model(inputs=x, outputs=r)  
12.  autoencoder.compile(optimizer='adam', loss='mse')

2) 训练模型

1.  epochs = 5  
2.  batch_size = 128  
3.    
4.  history = autoencoder.fit(X_train, X_train,   
5.                            batch_size=batch_size,   
6.                            epochs=epochs,   
7.                            verbose=1,   
8.                            validation_data=(X_test, X_test))

3) 查看编码效果

1.  conv_encoder = Model(x, h)  # 只取编码器做模型  
2.  encoded_imgs = conv_encoder.predict(X_test)  
3.    
4.  # 打印10张测试集手写体的压缩效果  
5.  n = 10  
6.  plt.figure(figsize=(20, 8))  
7.  for i in range(n):  
8.      ax = plt.subplot(1, n, i+1)  
9.      plt.imshow(encoded_imgs[i].reshape(4, 16).T)  
10.      plt.gray()  
11.      ax.get_xaxis().set_visible(False)  
12.      ax.get_yaxis().set_visible(False)  
13.  plt.show()

4) 查看解码效果

1.  decoded_imgs = autoencoder.predict(X_test)  
2.    
3.  n = 10  
4.  plt.figure(figsize=(20, 6))  
5.  for i in range(n):  
6.      # 原图  
7.      ax = plt.subplot(3, n, i+1)  
8.      plt.imshow(X_test[i].reshape(28, 28))  
9.      plt.gray()  
10.      ax.get_xaxis().set_visible(False)  
11.      ax.get_yaxis().set_visible(False)  
12.    
13.        
14.      # 解码效果图  
15.      ax = plt.subplot(3, n, i+n+1)  
16.      plt.imshow(decoded_imgs[i].reshape(28, 28))  
17.      plt.gray()  
18.      ax.get_xaxis().set_visible(False)  
19.      ax.get_yaxis().set_visible(False)  
20.        
21.  plt.show()

5) 训练过程可视化

1.  print(history.history.keys())  
2.    
3.  plt.plot(history.history['loss'])  
4.  plt.plot(history.history['val_loss'])  
5.  plt.title('model loss')  
6.  plt.ylabel('loss')  
7.  plt.xlabel('epoch')  
8.  plt.legend(['train', 'validation'], loc='upper right')  
9.  plt.show()

dict_keys(['val_loss', 'loss'])

3. 卷积自编码器

1) 读取数据集

1.  nb_classes = 10  # 10类  
2.    
3.  (X_train, y_train), (X_test, y_test) = mnist.load_data()  
4.    
5.  X_train = X_train.reshape(X_train.shape[0], 28, 28, 1)  
6.  X_test = X_test.reshape(X_test.shape[0], 28, 28, 1)  
7.    
8.  # 归一化  
9.  X_train = X_train.astype("float32")/255.  
10.  X_test = X_test.astype("float32")/255.  
11.  print('X_train shape:', X_train.shape)  
12.  print(X_train.shape[0], 'train samples')  
13.  print(X_test.shape[0], 'test samples')  
14.    
15.  y_train = np_utils.to_categorical(y_train, nb_classes)  
16.  y_test = np_utils.to_categorical(y_test, nb_classes)

2) 卷积自编码器建模

1.  x = Input(shape=(28, 28,1))     
2.      
3.  # 编码器    
4.  conv1_1 = Conv2D(16, (3, 3), activation='relu', padding='same')(x)    
5.  pool1 = MaxPooling2D((2, 2), padding='same')(conv1_1)    
6.  conv1_2 = Conv2D(8, (3, 3), activation='relu', padding='same')(pool1)    
7.  pool2 = MaxPooling2D((2, 2), padding='same')(conv1_2)    
8.  conv1_3 = Conv2D(8, (3, 3), activation='relu', padding='same')(pool2)    
9.  h = MaxPooling2D((2, 2), padding='same')(conv1_3)    
10.      
11.      
12.  # 解码器    
13.  conv2_1 = Conv2D(8, (3, 3), activation='relu', padding='same')(h)    
14.  up1 = UpSampling2D((2, 2))(conv2_1)    
15.  conv2_2 = Conv2D(8, (3, 3), activation='relu', padding='same')(up1)    
16.  up2 = UpSampling2D((2, 2))(conv2_2)    
17.  conv2_3 = Conv2D(16, (3, 3), activation='relu')(up2)    
18.  up3 = UpSampling2D((2, 2))(conv2_3)    
19.  r = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(up3)    
20.      
21.  autoencoder = Model(inputs=x, outputs=r)    
22.  autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')

3) 训练

1.  epochs = 3  
2.  batch_size = 128  
3.    
4.  history = autoencoder.fit(X_train, X_train,   
5.                            batch_size=batch_size,   
6.                            epochs=epochs, verbose=1,   
7.                            validation_data=(X_test, X_test)  
8.                           )

4) 查看解码效果

1.  decoded_imgs = autoencoder.predict(X_test)  
2.    
3.  n = 10  
4.  plt.figure(figsize=(20, 6))  
5.  for i in range(n):  
6.      # 原图  
7.      ax = plt.subplot(3, n, i+1)  
8.      plt.imshow(X_test[i].reshape(28, 28))  
9.      plt.gray()  
10.      ax.get_xaxis().set_visible(False)  
11.      ax.get_yaxis().set_visible(False)  
12.    
13.        
14.      # 解码效果图  
15.      ax = plt.subplot(3, n, i+n+1)  
16.      plt.imshow(decoded_imgs[i].reshape(28, 28))  
17.      plt.gray()  
18.      ax.get_xaxis().set_visible(False)  
19.      ax.get_yaxis().set_visible(False)  
20.        
21.  plt.show()

5) 训练过程可视化

1.  print(history.history.keys())  
2.    
3.  plt.plot(history.history['loss'])  
4.  plt.plot(history.history['val_loss'])  
5.  plt.title('model loss')  
6.  plt.ylabel('loss')  
7.  plt.xlabel('epoch')  
8.  plt.legend(['train', 'validation'], loc='upper right')  
9.  plt.show()

dict_keys(['val_loss', 'loss'])

4. 稀疏正则自编码器

1) 读取数据集

1.  (X_train, _), (X_test, _) = mnist.load_data()  
2.    
3.  #  归一化  
4.  X_train = X_train.astype("float32")/255.  
5.  X_test = X_test.astype("float32")/255.  
6.    
7.  print('X_train shape:', X_train.shape)  
8.  print(X_train.shape[0], 'train samples')  
9.  print(X_test.shape[0], 'test samples')

X_train shape: (60000, 28, 28)

60000 train samples

10000 test samples

1.  # np.prod是将28X28矩阵转化成1X784,方便全连接神经网络输入层784个神经元读取。  
2.  X_train = X_train.reshape((len(X_train), np.prod(X_train.shape[1:])))  
3.  X_test = X_test.reshape((len(X_test), np.prod(X_test.shape[1:])))

2) 稀疏正则自编码器建模

1.  input_size = 784  
2.  hidden_size = 32  
3.  output_size = 784  
4.    
5.  x = Input(shape=(input_size,))  
6.  h = Dense(hidden_size, activation='relu', activity_regularizer=regularizers.l1(10e-5))(x)  
7.  r = Dense(output_size, activation='sigmoid')(h)  
8.    
9.  autoencoder = Model(inputs=x, outputs=r)  
10.  autoencoder.compile(optimizer='adam', loss='mse')

3) 训练

1.  epochs = 15  
2.  batch_size = 128  
3.    
4.  history = autoencoder.fit(X_train, X_train,   
5.                            batch_size=batch_size,   
6.                            epochs=epochs,   
7.                            verbose=1,   
8.                            validation_data=(X_test, X_test)  
9.                           )

4) 查看解码效果

1.  n = 10  
2.  plt.figure(figsize=(20, 6))  
3.  for i in range(n):  
4.      # 原图  
5.      ax = plt.subplot(3, n, i+1)  
6.      plt.imshow(X_test[i].reshape(28, 28))  
7.      plt.gray()  
8.      ax.get_xaxis().set_visible(False)  
9.      ax.get_yaxis().set_visible(False)  
10.    
11.        
12.      # 解码效果图  
13.      ax = plt.subplot(3, n, i+n+1)  
14.      plt.imshow(decoded_imgs[i].reshape(28, 28))  
15.      plt.gray()  
16.      ax.get_xaxis().set_visible(False)  
17.      ax.get_yaxis().set_visible(False)  
18.        
19.  plt.show()

5) 训练过程可视化

1.  print(history.history.keys())  
2.    
3.  plt.plot(history.history['loss'])  
4.  plt.plot(history.history['val_loss'])  
5.  plt.title('model loss')  
6.  plt.ylabel('loss')  
7.  plt.xlabel('epoch')  
8.  plt.legend(['train', 'validation'], loc='upper right')  
9.  plt.show()

dict_keys(['val_loss', 'loss'])

5. 去噪自编码器

1) 读取数据集

1.  (X_train, _), (X_test, _) = mnist.load_data()  
2.    
3.  X_train = X_train.reshape(X_train.shape[0], 28, 28, 1)  
4.  X_test = X_test.reshape(X_test.shape[0], 28, 28, 1)  
5.    
6.  X_train = X_train.astype("float32")/255.  
7.  X_test = X_test.astype("float32")/255.

2) 加噪

1.  noise_factor = 0.5  
2.  X_train_noisy = X_train + noise_factor * np.random.normal(loc=0.0, scale=1.0, size=X_train.shape)   
3.  X_test_noisy = X_test + noise_factor * np.random.normal(loc=0.0, scale=1.0, size=X_test.shape)   
4.    
5.  X_train_noisy = np.clip(X_train_noisy, 0., 1.)  
6.  X_test_noisy = np.clip(X_test_noisy, 0., 1.)

3) 去噪自编码器建模

1.  x = Input(shape=(28, 28, 1))  
2.    
3.  # 编码器  
4.  conv1_1 = Conv2D(32, (3, 3), activation='relu', padding='same')(x)  
5.  pool1 = MaxPooling2D((2, 2), padding='same')(conv1_1)  
6.  conv1_2 = Conv2D(32, (3, 3), activation='relu', padding='same')(pool1)  
7.  h = MaxPooling2D((2, 2), padding='same')(conv1_2)  
8.    
9.    
10.  # 解码器  
11.  conv2_1 = Conv2D(32, (3, 3), activation='relu', padding='same')(h)  
12.  up1 = UpSampling2D((2, 2))(conv2_1)  
13.  conv2_2 = Conv2D(32, (3, 3), activation='relu', padding='same')(up1)  
14.  up2 = UpSampling2D((2, 2))(conv2_2)  
15.  r = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(up2)  
16.    
17.  autoencoder = Model(inputs=x, outputs=r)  
18.  autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')

4) 训练

1.  epochs = 3  
2.  batch_size = 128  
3.    
4.  history = autoencoder.fit(X_train_noisy, X_train,   
5.                            batch_size=batch_size,  
6.                            epochs=epochs, verbose=1,  
7.                            validation_data=(X_test_noisy, X_test))

5) 查看解码效果

1.  decoded_imgs = autoencoder.predict(X_test_noisy)  
2.    
3.  n = 10  
4.  plt.figure(figsize=(20, 6))  
5.  for i in range(n):  
6.      # 原图  
7.      ax = plt.subplot(3, n, i+1)  
8.      plt.imshow(X_test_noisy[i].reshape(28, 28))  
9.      plt.gray()  
10.      ax.get_xaxis().set_visible(False)  
11.      ax.get_yaxis().set_visible(False)  
12.    
13.        
14.      # 解码效果图  
15.      ax = plt.subplot(3, n, i+n+1)  
16.      plt.imshow(decoded_imgs[i].reshape(28, 28))  
17.      plt.gray()  
18.      ax.get_xaxis().set_visible(False)  
19.      ax.get_yaxis().set_visible(False)  
20.        
21.  plt.show()

6) 训练过程可视化

1.  print(history.history.keys())  
2.    
3.  plt.plot(history.history['loss'])  
4.  plt.plot(history.history['val_loss'])  
5.  plt.title('model loss')  
6.  plt.ylabel('loss')  
7.  plt.xlabel('epoch')  
8.  plt.legend(['train', 'validation'], loc='upper right')  
9.  plt.show()

dict_keys(['val_loss', 'loss'])