Building Autoencoders in Keras

目的：利用autoencoder进行降维，并尽量保留主要信息（仅针对特征），mnist数据集

数据处理

读取数据、归一化（仅针对特征）

import matplotlib.pyplot as plt
import numpy as np
from keras import regularizers
from keras.datasets import mnist
from keras.layers import Dense, Input
from keras.models import Model
from keras.utils import np_utils
#读取数据
file = mnist.load_data()
# tuple类型
# 包含训练集：特征(60000,28,28),标签(60000,)
# 测试集：特征(10000,28,28),标签(10000,)
x_train = file[0][0]
x_test = file[1][0]
picture_size = x_train.shape[1] * x_train.shape[2]  # 图片大小


def preprocess(data, process_x=True):
    # x转换成二维数组，并进行归一化（0-1）
    # y进行one-hot 编码
    if process_x == True:
        data = np.array(data).reshape((data.shape[0], picture_size)).astype('float32')  # 保留小数位
        data = data / 255
    else:
        data = np_utils.to_categorical(data)  # one-hot编码
    return data


x_train = preprocess(x_train, process_x=True)  # (10000, 784)
x_test = preprocess(x_test, process_x=True)  # (60000, 784)

1.简单autoencoder

784-32-784

encoding_dim = 32#降维维度
input_img = Input(shape=(784,))
encoded = Dense(encoding_dim, activation='relu')(input_img)
decoded = Dense(784, activation='sigmoid')(encoded)
autoencoder = Model(input_img, decoded)

提取出编码器、解码器

encoder = Model(input_img, encoded)  # 编码器
encoded_input = Input(shape=(encoding_dim,))
decoder_layer = autoencoder.layers[-1](encoded_input)
decoder = Model(encoded_input, decoder_layer)  # 解码器

编译、训练

autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')

autoencoder.fit(x_train, x_train,
                epochs=100,
                batch_size=500,
                shuffle=True,
                validation_data=(x_test, x_test))

预测、画图（看原始图像与decoder之后图像的差异）

encoded_imgs = encoder.predict(x_test)#编码之后图像
decoded_imgs = decoder.predict(encoded_imgs)#解码后图像

n = 10  # how many digits we will display
plt.figure(figsize=(20, 4))
for i in range(n):
    # display original
    ax = plt.subplot(2, n, i + 1)
    plt.imshow(x_test[i].reshape(28, 28))
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)

    # display reconstruction
    ax = plt.subplot(2, n, i + 1 + n)
    plt.imshow(decoded_imgs[i].reshape(28, 28))
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

loss: 0.1063 - val_loss: 0.1045

2.稀疏编码器

实质上是在编码层（encoder）加入正则化项（l1 or l2），以牺牲训练集精度的代价，提高测试集精度，以提高泛化能力。其中惩罚系数越大，稀疏性越强，则不容易过拟合，此时应提高epoch次数

encoded = Dense(encoding_dim, activation='relu', activity_regularizer=regularizers.l1(10e-8))(input_img)

loss: 0.1057 - val_loss: 0.1042

3.深度自编码器（堆叠自编码StackedAE/Deep AE）

1.多层编码/解码

784-128-64-32-64-128-784

input_img = Input(shape=(784,))
encoded = Dense(128, activation='relu')(input_img)
encoded = Dense(64, activation='relu')(encoded)
encoded = Dense(32, activation='relu')(encoded)

decoded = Dense(64, activation='relu')(encoded)
decoded = Dense(128, activation='relu')(decoded)
decoded = Dense(784, activation='sigmoid')(decoded)
autoencoder = Model(input_img, decoded)
autoencoder.summary()
# Model: "model_1"
#_________________________________________________________________
# Layer (type)                 Output Shape              Param #   
# =================================================================
# input_1 (InputLayer)         (None, 784)               0         
# _________________________________________________________________
# dense_1 (Dense)              (None, 128)               100480    
# _________________________________________________________________
# dense_2 (Dense)              (None, 64)                8256      
# _________________________________________________________________
# dense_3 (Dense)              (None, 32)                2080      
# _________________________________________________________________
# dense_4 (Dense)              (None, 64)                2112      
# _________________________________________________________________
# dense_5 (Dense)              (None, 128)               8320      
# _________________________________________________________________
# dense_6 (Dense)              (None, 784)               101136    
# =================================================================
# Total params: 222,384
# Trainable params: 222,384
# Non-trainable params: 0

之后对autoencoder进行compile、fit

2.编码器：784-128-64-32

encoder = Model(input_img, encoded)
encoder.summary()
# Model: "model_2"
# _________________________________________________________________
# Layer (type)                 Output Shape              Param #   
# =================================================================
# input_1 (InputLayer)         (None, 784)               0         
# _________________________________________________________________
# dense_1 (Dense)              (None, 128)               100480    
# _________________________________________________________________
# dense_2 (Dense)              (None, 64)                8256      
# _________________________________________________________________
# dense_3 (Dense)              (None, 32)                2080      
# =================================================================
# Total params: 110,816
# Trainable params: 110,816
# Non-trainable params: 0

3.解码器32-64-128-784

# 解码器
encoded_input = Input(shape=(32,))
decoder_layer1 = autoencoder.layers[-3](encoded_input)
decoder_layer2 = autoencoder.layers[-2](decoder_layer1)
decoder_layer3 = autoencoder.layers[-1](decoder_layer2)
decoder = Model(encoded_input, decoder_layer3)
decoder.summary()
# _________________________________________________________________
# Layer (type)                 Output Shape              Param #   
# =================================================================
# input_2 (InputLayer)         (None, 32)                0         
# _________________________________________________________________
# dense_4 (Dense)              (None, 64)                2112      
# _________________________________________________________________
# dense_5 (Dense)              (None, 128)               8320      
# _________________________________________________________________
# dense_6 (Dense)              (None, 784)               101136    
# =================================================================
# Total params: 111,568
# Trainable params: 111,568
# Non-trainable params: 0

4.预测

此时由于是调用autoencoder 已经compile、fit完的层，不用再作操作，直接预测

encoded_imgs = encoder.predict(x_test)
decoded_imgs = decoder.predict(encoded_imgs)
loss: 0.0996 - val_loss: 0.1003

这里过拟合了，可以加入正则化项

参考
Building Autoencoders in Keras