tensorflow文本卷积

卷积例子：
inputs = tf.placeholder(‘float’, shape=[None, 6, 8])
out = tf.layers.conv1d(inputs, 5, 3)
说明： 对于一个样本而言，句子有6个词，词向量的维度为8，filters=5（输出维度为5）, kernel_size=3（卷积核大小为3）， 因为是一维卷积，所以卷积核的大小为3，那么输入6经过3的卷积核卷积后得到的是4的一个向量(4=6-3+1)，又因为有5个过滤器，所以是得到5个4*1的向量（一维卷积的只能在一个维度上进行卷积，第二个维度在卷积之后的维度大小为1）

import numpy as np
import matplotlib.pyplot as plt
from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras.preprocessing.sequence import pad_sequences
#定义超参
num_features = 3000
sequence_length = 300
embedding_dimension = 100
#加载数据集
(x_train, y_train), (x_test, y_test) = keras.datasets.imdb.load_data(num_words=num_features)
print(x_train.shape)
print(x_test.shape)
print(y_train.shape)
print(y_test.shape)
#文本的padding处理
x_train = pad_sequences(x_train, maxlen=sequence_length)
x_test = pad_sequences(x_test, maxlen=sequence_length)
print(x_train.shape)
print(x_test.shape)
print(y_train.shape)
print(y_test.shape)

基本的文本卷积

构建模型

def imdb_cnn():
    model = keras.Sequential([
        layers.Embedding(input_dim=num_features, output_dim=embedding_dimension,input_length=sequence_length),
        #input_dim词汇表的大小（取数据的时候取的词汇频率最高的num_features个词汇，所以词汇表的大小为num_features）
        #input_length输入句子padding之后的长度
        #输出为sequence_length*embedding_dimension（300*100）
        layers.Conv1D(filters=50, kernel_size=5, strides=1, padding='valid'),
        #卷积核大小5，50个过滤器，输出为（sequence_length-kernel_size+1）*50
        layers.MaxPool1D(pool_size=2, padding='valid'),#每两个过滤器进行一次maxpool得到（sequence_length-kernel_size+1）/2*50
        layers.Flatten(),#将结果展平，作为之后网络的输入
        layers.Dense(10, activation='relu'),
        layers.Dense(1, activation='sigmoid')
    ])
    model.compile(optimizer=keras.optimizers.Adam(1e-3),
                 loss=keras.losses.BinaryCrossentropy(),
                 metrics=['accuracy'])
    
    return model

model = imdb_cnn()
for i in range(len(model.layers)):#查看各层的输出shape
    print(model.get_layer(index=i).output)
model.summary()
输出：
Tensor("embedding_4/Identity:0", shape=(None, 300, 100), dtype=float32)
Tensor("conv1d_1/Identity:0", shape=(None, 296, 50), dtype=float32)
Tensor("max_pooling1d_1/Identity:0", shape=(None, 148, 50), dtype=float32)
Tensor("flatten_3/Identity:0", shape=(None, 7400), dtype=float32)
Tensor("dense_6/Identity:0", shape=(None, 10), dtype=float32)
Tensor("dense_7/Identity:0", shape=(None, 1), dtype=float32)
Model: "sequential_3"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_4 (Embedding)      (None, 300, 100)          300000    
_________________________________________________________________
conv1d_1 (Conv1D)            (None, 296, 50)           25050     
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 148, 50)           0         
_________________________________________________________________
flatten_3 (Flatten)          (None, 7400)              0         
_________________________________________________________________
dense_6 (Dense)              (None, 10)                74010     
_________________________________________________________________
dense_7 (Dense)              (None, 1)                 11        
=================================================================
Total params: 399,071
Trainable params: 399,071
Non-trainable params: 0
_________________________________________________________________

模型训练

%%time
history = model.fit(x_train, y_train, batch_size=64, epochs=5, validation_split=0.1)

plt.plot(history.history['accuracy'])
plt.plot(history.history['val_accuracy'])
plt.legend(['training', 'valiation'], loc='upper left')
plt.show()

多核卷积网络

filter_sizes=[3,4,5]#定义三种不同大小的卷积核
def convolution():#函数式加入多核卷积的写法
    inn = layers.Input(shape=(sequence_length, embedding_dimension, 1))#定义函数的输入层
    cnns = []
    for size in filter_sizes:#不同大小的卷积核
        conv = layers.Conv2D(filters=64, kernel_size=(size, embedding_dimension),#卷积核大小为size*embedding_dimension
                            strides=1, padding='valid', activation='relu')(inn)#64个卷积核，输出为64个（sequence_length-kernel_size+1）*11
        pool = layers.MaxPool2D(pool_size=(sequence_length-size+1, 1), padding='valid')(conv)#每个卷积核的结果只保存最大的
        cnns.append(pool)
    #所以for的每一次循环都得到64个卷积结果，三次循环就得到192个，以此作为多核卷积的输出结果
    outt = layers.concatenate(cnns)#函数的输出层
    
    model = keras.Model(inputs=inn, outputs=outt)
    return model

def cnn_mulfilter():
    model = keras.Sequential([
        layers.Embedding(input_dim=num_features, output_dim=embedding_dimension,
                        input_length=sequence_length),
        layers.Reshape((sequence_length, embedding_dimension, 1)),#二维卷积
        convolution(),
        layers.Flatten(),
        layers.Dense(10, activation='relu'),
        layers.Dropout(0.2),
        layers.Dense(1, activation='sigmoid')
    ])
    model.compile(optimizer=keras.optimizers.Adam(),
                 loss=keras.losses.BinaryCrossentropy(),
                 metrics=['accuracy'])
    return model

model = cnn_mulfilter()
for i in range(len(model.layers)):#查看各层的输出shape
    print(model.get_layer(index=i).output)
model.summary()
输出：
Tensor("embedding_3/Identity:0", shape=(None, 300, 100), dtype=float32)
Tensor("reshape_1/Identity:0", shape=(None, 300, 100, 1), dtype=float32)
Tensor("concatenate_1/Identity:0", shape=(None, 1, 1, 192), dtype=float32)
Tensor("flatten_2/Identity:0", shape=(None, 192), dtype=float32)
Tensor("dense_4/Identity:0", shape=(None, 10), dtype=float32)
Tensor("dropout_1/Identity:0", shape=(None, 10), dtype=float32)
Tensor("dense_5/Identity:0", shape=(None, 1), dtype=float32)
Model: "sequential_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_3 (Embedding)      (None, 300, 100)          300000    
_________________________________________________________________
reshape_1 (Reshape)          (None, 300, 100, 1)       0         
_________________________________________________________________
model_1 (Model)              (None, 1, 1, 192)         76992     
_________________________________________________________________
flatten_2 (Flatten)          (None, 192)               0         
_________________________________________________________________
dense_4 (Dense)              (None, 10)                1930      
_________________________________________________________________
dropout_1 (Dropout)          (None, 10)                0         
_________________________________________________________________
dense_5 (Dense)              (None, 1)                 11        
=================================================================
Total params: 378,933
Trainable params: 378,933
Non-trainable params: 0
_________________________________________________________________

模型训练

%%time
history = model.fit(x_train, y_train, batch_size=64, epochs=5, validation_split=0.1)
plt.plot(history.history['accuracy'])
plt.plot(history.history['val_accuracy'])
plt.legend(['training', 'valiation'], loc='upper left')
plt.show()