Keras函数式(functional)API的使用

本是古典何须时尚 2022-01-27 10:21 160阅读 0赞

### 多层感知器(Multilayer Perceptron) ###

定义了用于二分类的多层感知器模型。模型输入32维特征，经过三个全连接层，每层使用relu线性激活函数，并且在输出层中使用sigmoid激活函数，最后用于二分类。

##------ Multilayer Perceptron ------##
    from keras.models import Model
    from keras.layers import Input, Dense
    from keras import backend as K
    K.clear_session() 
    
    # MLP model
    x = Input(shape=(32,))
    hidden1 = Dense(10, activation='relu')(x)
    hidden2 = Dense(20, activation='relu')(hidden1)
    hidden3 = Dense(10, activation='relu')(hidden2)
    output = Dense(1, activation='sigmoid')(hidden3)
    model = Model(inputs=x, outputs=output)
    
    # summarize layers
    model.summary()

模型的结构和参数如下：  
![Multilayer Perceptron][]

### 卷积神经网络(Convolutional Neural Network) ###

定义用于图像分类的卷积神经网络。该模型接收3通道的64×64图像作为输入，然后经过两个卷积和池化层的序列作为特征提取器，接着过一个全连接层，最后输出层过softmax激活函数进行10个类别的分类。

##------ Convolutional Neural Network ------##
    from keras.models import Model
    from keras.layers import Input
    from keras.layers import Dense, Flatten
    from keras.layers import Conv2D, MaxPooling2D
    from keras import backend as K
    K.clear_session()
    
    # CNN model
    x = Input(shape=(64,64,3))
    conv1 = Conv2D(16, (5,5), activation='relu')(x)
    pool1 = MaxPooling2D((2,2))(conv1)
    conv2 = Conv2D(32, (3,3), activation='relu')(pool1)
    pool2 = MaxPooling2D((2,2))(conv2)
    conv3 = Conv2D(32, (3,3), activation='relu')(pool2)
    pool3 = MaxPooling2D((2,2))(conv3)
    flat = Flatten()(pool3)
    hidden1 = Dense(512, activation='relu')(flat)
    output = Dense(10, activation='softmax')(hidden1)
    model = Model(inputs=x, outputs=output)
    
    # summarize layers
    model.summary()

模型的结构和参数如下：  
![CNN][]

### 循环神经网络(Recurrent Neural Network) ###

定义一个用于文本序列分类的LSTM网络。该模型需要100个时间步长作为输入，然后经过一个Embedding层，每个时间步变成128维特征表示，然后经过一个LSTM层，LSTM输出过一个全连接层，最后输出用sigmoid激活函数用于进行二分类预测。

##------ Recurrent Neural Network ------##
    from keras.models import Model
    from keras.layers import Input
    from keras.layers import Dense, LSTM, Embedding
    from keras import backend as K
    K.clear_session()
    
    VOCAB_SIZE = 10000
    EMBED_DIM = 128
    
    x = Input(shape=(100,), dtype='int32')
    embedding = Embedding(VOCAB_SIZE, EMBED_DIM, mask_zero=True)(x)
    hidden1 = LSTM(64)(embedding)
    hidden2 = Dense(32, activation='relu')(hidden1)
    output = Dense(1, activation='sigmoid')(hidden2)
    model = Model(inputs=x, outputs=output)
    
    # summarize layers
    model.summary()

模型的结构和参数如下：  
![LSTM][]

### Bidirectional recurrent neural network ###

定义一个双向循环神经网络，可以用来完成序列标注等任务，相比上面的LSTM网络，多了一个反向的LSTM，其它设置一样。

##------ Bidirectional recurrent neural network ------##
    from keras.models import Model
    from keras.layers import Input, Embedding
    from keras.layers import Dense, LSTM, Bidirectional
    from keras import backend as K
    K.clear_session()
    
    VOCAB_SIZE = 10000
    EMBED_DIM = 128
    HIDDEN_SIZE = 64
    # input layer
    x = Input(shape=(100,), dtype='int32')
    # embedding layer
    embedding = Embedding(VOCAB_SIZE, EMBED_DIM, mask_zero=True)(x)
    # BiLSTM layer
    hidden = Bidirectional(LSTM(HIDDEN_SIZE, return_sequences=True))(embedding)
    # prediction layer
    output = Dense(10, activation='softmax')(hidden)
    model = Model(inputs=x, outputs=output)
    
    model.summary()

模型的结构和参数如下：  
![BiLSTM][]

### 共享输入层模型(Shared Input Layer Model) ###

定义了具有不同大小内核的多个卷积层来解释图像输入。该模型采用尺寸为64×64像素的3通道图像。 有两个共享此输入的CNN特征提取子模型; 第一个内核大小为5x5，第二个内核大小为3x3。把提取的特征展平为向量然后拼接成一个长向量，然后过一个全连接层，最后输出层完成10分类。

##------ Shared Input Layer Model ------##
    
    from keras.models import Model
    from keras.layers import Input
    from keras.layers import Dense, Flatten
    from keras.layers import Conv2D, MaxPooling2D, Concatenate
    from keras import backend as K
    K.clear_session()
    
    # input layer
    x = Input(shape=(64,64,3))
    # first feature extractor
    conv1 = Conv2D(32, (3,3), activation='relu')(x)
    pool1 = MaxPooling2D((2,2))(conv1)
    flat1 = Flatten()(pool1)
    # second feature extractor 
    conv2 = Conv2D(16, (5,5), activation='relu')(x)
    pool2 = MaxPooling2D((2,2))(conv2)
    flat2 = Flatten()(pool2)
    # merge feature
    merge = Concatenate()([flat1, flat2])
    # interpretation layer
    hidden1 = Dense(128, activation='relu')(merge)
    # prediction layer
    output = Dense(10, activation='softmax')(merge)
    model = Model(inputs=x, outputs=output)
    
    model.summary()

模型的结构和参数如下：  
![Shared-Input-Layer][]

### Shared Feature Extraction Layer ###

定义一个共享特征抽取层的模型，这里共享的是LSTM层的输出，具体共享参见代码

##------ Shared Feature Extraction Layer ------##
    from keras.models import Model
    from keras.layers import Input, Embedding
    from keras.layers import Dense, LSTM, Concatenate
    from keras import backend as K
    K.clear_session()
    
    # input layer
    x = Input(shape=(100,32))
    # feature extraction
    extract1 = LSTM(64)(x)
    # first interpretation model
    interp1 = Dense(32, activation='relu')(extract1)
    # second interpretation model
    interp11 = Dense(64, activation='relu')(extract1)
    interp12 = Dense(32, activation='relu')(interp11)
    # merge interpretation
    merge = Concatenate()([interp1, interp12])
    # output layer
    output = Dense(10, activation='softmax')(merge)
    model = Model(inputs=x, outputs=output)
    
    model.summary()

模型的结构和参数如下：  
![Shared-Feature-Extraction-Layer][]

### 多输入模型(Multiple Input Model) ###

定义有两个输入的模型，这里测试的是输入两张图片，一个输入是单通道的64x64，另一个是3通道的32x32，两个经过卷积层、池化层后，展平拼接，最后进行二分类。

##------ Multiple Input Model ------##
    from keras.models import Model
    from keras.layers import Input
    from keras.layers import Dense, Flatten
    from keras.layers import Conv2D, MaxPooling2D, Concatenate
    from keras import backend as K
    K.clear_session()
    
    # first input model
    input1 = Input(shape=(64,64,1))
    conv11 = Conv2D(32, (5,5), activation='relu')(input1)
    pool11 = MaxPooling2D(pool_size=(2,2))(conv11)
    conv12 = Conv2D(16, (3,3), activation='relu')(pool11)
    pool12 = MaxPooling2D(pool_size=(2,2))(conv12)
    flat1 = Flatten()(pool12)
    # second input model
    input2 = Input(shape=(32,32,3))
    conv21 = Conv2D(32, (5,5), activation='relu')(input2)
    pool21 = MaxPooling2D(pool_size=(2,2))(conv21)
    conv22 = Conv2D(16, (3,3), activation='relu')(pool21)
    pool22 = MaxPooling2D(pool_size=(2,2))(conv22)
    flat2 = Flatten()(pool22)
    # merge input models
    merge = Concatenate()([flat1, flat2])
    # interpretation model
    hidden1 = Dense(20, activation='relu')(merge)
    output = Dense(1, activation='sigmoid')(hidden1)
    model = Model(inputs=[input1, input2], outputs=output)
    
    model.summary()

模型的结构和参数如下：  
![Multiple-Input-Model][]

### 多输出模型(Multiple Output Model) ###

定义有多个输出的模型，以文本序列输入LSTM网络为例，一个输出是对文本的分类，另外一个输出是对文本进行序列标注。

##------ Multiple Output Model ------ ##
    from keras.models import Model
    from keras.layers import Input
    from keras.layers import Dense, Flatten, TimeDistributed, LSTM
    from keras.layers import Conv2D, MaxPooling2D, Concatenate
    from keras import backend as K
    K.clear_session()
    
    x = Input(shape=(100,1))
    extract = LSTM(10, return_sequences=True)(x)
    
    class11 = LSTM(10)(extract)
    class12 = Dense(10, activation='relu')(class11)
    output1 = Dense(1, activation='sigmoid')(class12)
    output2 = TimeDistributed(Dense(1, activation='linear'))(extract)
    model = Model(inputs=x, outputs=[output1, output2])
    
    model.summary()

模型的结构和参数如下：  
![Multiple-Output-Model][]

### 参考 ###

\[1\] [https://machinelearningmastery.com/keras-functional-api-deep-learning/][https_machinelearningmastery.com_keras-functional-api-deep-learning]  
\[2\] [https://keras.io/getting-started/functional-api-guide/][https_keras.io_getting-started_functional-api-guide]  
\[3\] [https://tensorflow.google.cn/alpha/guide/keras/functional][https_tensorflow.google.cn_alpha_guide_keras_functional]

[Multilayer Perceptron]: /images/20220127/59c358e99634416bbdb21d2eaf5ec0ab.png
[CNN]: /images/20220127/7c094ddfee074b5a9763a34ad13ef45d.png
[LSTM]: /images/20220127/7390cf106ea844f8b44236c21c8d8d4e.png
[BiLSTM]: /images/20220127/dd8fe7ce397f4d36bf9ed79228549659.png
[Shared-Input-Layer]: /images/20220127/667256f7b8f840b4827a636ca7217ad5.png
[Shared-Feature-Extraction-Layer]: /images/20220127/65ea04ea127447b58b2df8774e8fa3a8.png
[Multiple-Input-Model]: /images/20220127/6c681b816af44be481a294d994a83330.png
[Multiple-Output-Model]: /images/20220127/e9fe3964e2ee4aad91bf26297f8142bb.png
[https_machinelearningmastery.com_keras-functional-api-deep-learning]: https://machinelearningmastery.com/keras-functional-api-deep-learning/
[https_keras.io_getting-started_functional-api-guide]: https://keras.io/getting-started/functional-api-guide/
[https_tensorflow.google.cn_alpha_guide_keras_functional]: https://tensorflow.google.cn/alpha/guide/keras/functional