Keras函数式(functional)API的使用

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：

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：

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：

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：

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 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：

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：

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：

参考

[1] https://machinelearningmastery.com/keras-functional-api-deep-learning/ [2] https://keras.io/getting-started/functional-api-guide/ [3] https://tensorflow.google.cn/alpha/guide/keras/functional