TextCNN keras 实现

参考：https://zhuanlan.zhihu.com/p/77634533

from IPython.core.interactiveshell import InteractiveShell
InteractiveShell.ast_node_interactivity = "all"
import numpy as np
from sklearn import metrics
import keras
from keras.preprocessing import sequence
from keras.models import Sequential, Model
from keras.layers import Dense, Dropout, Activation, concatenate
from keras.layers import Embedding, Input, MaxPooling1D, Flatten
from keras.layers import Conv1D, GlobalMaxPooling1D
from keras.datasets import imdb
from sklearn.model_selection import train_test_split
import pandas as pd
import jieba
import re
import gensim
from keras.preprocessing.sequence import pad_sequences
from gensim.models import KeyedVectors
from gensim.models import word2vec
from keras.preprocessing.text import Tokenizer
import tensorflow as tf
import keras.backend.tensorflow_backend as KTF
from keras.models import load_model
import joblib
from sklearn.model_selection import GridSearchCV
from keras.utils import to_categorical
from keras.wrappers.scikit_learn import KerasClassifier 

KTF.set_session(tf.Session(config=tf.ConfigProto(device_count={'gpu':0})))


# 构建TextCNN模型，模型结构：词嵌入-卷积池化*3-拼接-全连接-dropout-全连接
def TextCNN_model(maxlen, max_features, embedding_dims, class_num, embedding_weights, activation):
    main_input = Input(shape=(maxlen,), dtype='float64')

    # 词嵌入（使用预训练的词向量）
    embedder = Embedding(max_features + 1, embedding_dims, input_length=maxlen, weights=[embedding_weights])
    embed = embedder(main_input)

    # 词窗大小分别为3,4,5
    cnn1 = Conv1D(256, 3, padding='same', strides=1, activation='relu')(embed)
    cnn1 = MaxPooling1D()(cnn1)
    cnn2 = Conv1D(256, 4, padding='same', strides=1, activation='relu')(embed)
    cnn2 = MaxPooling1D()(cnn2)
    cnn3 = Conv1D(256, 5, padding='same', strides=1, activation='relu')(embed)
    cnn3 = MaxPooling1D()(cnn3)

    # 合并三个模型的输出向量
    cnn = concatenate([cnn1, cnn2, cnn3], axis=-1)
    flat = Flatten()(cnn)
    drop = Dropout(0.2)(flat)
    main_output = Dense(class_num, activation=activation)(drop)

    model = Model(inputs=main_input, outputs=main_output)
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

    return model

word2vec = gensim.models.Word2Vec.load("../model/word2vec.model")
max_features = word2vec.syn1.shape[0]
embedding_dims = word2vec.syn1.shape[1]

# We add an additional row of zeros to the embeddings matrix to represent unseen words and the NULL token.
embedding_weights = np.zeros((word2vec.syn1.shape[0] + 1, word2vec.syn1.shape[1]), dtype="float32")
embedding_weights[:word2vec.syn1.shape[0]] = word2vec.syn1

label_dict = np.load('../data/label_dict.npy', allow_pickle=True).tolist()
label_dict = dict(zip(label_dict.values(), label_dict.keys()))
jieba.load_userdict('../data/OTA_words.txt')

st = re.compile('\[.*?\]')
rule = re.compile(u"[^\u4e00-\u9fa5]")


def handel_query(line):
    mytext = st.sub('', line)
    line = rule.sub('', mytext)
    line = list(jieba.cut(line))
    return line


dataset = pd.read_csv('../data/all_data_final.csv')
dataset['words'] = dataset['query'].apply(lambda x: handel_query(x))

dataset['token'] = dataset['words'].apply(lambda words: [word2vec.wv.vocab[token].index if token in word2vec.wv.vocab else max_features for token in words])


batch_size = 1000
epochs = 10
class_num = 339
activation = 'softmax'
maxlen = 256

train_df = dataset[0: int(dataset.shape[0] * 0.9)]
test_df = dataset[int(dataset.shape[0] * 0.9):]

x_train = train_df['token'].tolist()
x_train = pad_sequences(x_train, maxlen=maxlen)  # 将超过固定值的部分截掉，不足的在最前面用0填充

y_train = to_categorical(train_df['label_id'].tolist(), num_classes=class_num)  # 将标签转换为one-hot编码

model = TextCNN_model(maxlen, max_features, embedding_dims, class_num, embedding_weights, activation)

print('Train...')
model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs)

# 模型保存
model.save('./model/textcnn_model.h5')

words = test_df['token'].tolist()
result = model.predict(pad_sequences(words, maxlen=maxlen))
y_predict = np.argmax(result, axis=1)  # 获得最大概率对应的标签

print('准确率', metrics.accuracy_score(test_df['label_id'].tolist(), y_predict))

word2vec 预训练模型维度：

模型结构：

训练过程：

测试集上准确率：