说明
本文主要是实现NLP文本分类任务，目前该系统为四系统融合模型，其中包括基于字向量char的CNN模型、RCNN模型 以及 基于词向量word2vec的CNN模型、RCNN模型。目前的融合规则采用权值平均法进行融合。
github:
环境
python3
tensorflow
数据
所使用的数据为网上的cnews数据，共10个类别，分为训练集，验证集和测试集。 类别如下：
体育, 财经, 房产, 家居, 教育, 科技, 时尚, 时政, 游戏, 娱乐
训练集每个类别5千数据。共5万。验证集每个类别500，测试集每个类别1000。 训练集、验证集和测试集的文件如下：
cnews.train.txt: 训练集(50000条)
cnews.val.txt: 验证集(5000条)
cnews.test.txt: 测试集(10000条)
数据链接：
链接：https://pan.baidu.com/s/1vEWbsIcBW_vgRJuYeLtj9Q
提取码：8czc
Train
下面我开放了训练部分的代码:
#coding: utf-8
import numpy as np
import tensorflow as tf
from tools import tools
import jieba
import os
import time
import logging
from model_net import TextMix_C_W_model as TM
import tensorflow.contrib.keras as kr
import config
os.environ['CUDA_VISIBLE_DEVICES']='1'
flages=tf.flags
FLAGS = flages.FLAGS
flages.DEFINE_float("learning_rate",0.001,"learning rate")
flages.DEFINE_string("stop_word","./data_pre/stop_word.txt","stop word data path")
flages.DEFINE_string("word2vec","./data_pre/word2voc_all.bin","word2voc data path")
flages.DEFINE_boolean("model_restore",False,"wheater to restore model")
flages.DEFINE_string("vocab_dir","./data_pre/cnews_vocab.txt","vocab data path")
flages.DEFINE_string("train_data","../data/cnews.train.txt","training data path")
flages.DEFINE_string("val_data","../data/cnews.val.txt","val data path")
flages.DEFINE_string("test_data","../data/cnews.test.txt","test data path")
def preprocess_data(vocab_data,word2vec,stop_word_dir,train_data,sequence_length):
###### 制作vocab 字典 ###################
vocab_list = tools.read_file(vocab_data)
vocabulary_word2index_char = {} ### word ：index
for i, vocab in enumerate(vocab_list):
vocabulary_word2index_char[vocab] = i
###### 加载 word2vec ###################
word2vec_list = tools.read_file(word2vec)
vocabulary_word2index_word = {}
len_vec = 0
for i, one_word2vec in enumerate(word2vec_list):
if i ==0:
pass
else:
word2vec_split = one_word2vec.split(' ')
words = word2vec_split[0]
vecs = np.array(word2vec_split[1:])
vocabulary_word2index_word[words] = vecs
len_vec = len(vecs)
######## 文本编码 ##############################
train_data_lines = tools.read_file(train_data)
categories, cat_to_id = tools.label_dict()
X_char = []
Y = []
total_data = []
for line in train_data_lines:
label, content = line.strip().split('\t')
contents = [vocabulary_word2index_char.get(e, 0) for e in content]
label = cat_to_id[label]
X_char.append(contents)
Y.append(label)
content_cut = jieba.cut(content, cut_all=False) ## 结巴分词
content_cut_smooth = tools.data_smooth(content_cut, stop_word_dir) ## 平滑
vec_line = np.zeros([sequence_length, len_vec]) ## 设置一个空矩阵
i = 0
# print(len(content_cut_smooth))
for word_content in content_cut_smooth:
if word_content in vocabulary_word2index_word.keys():
i += 1
if i > sequence_length - 1: ## 定长处理
break
vec_content = vocabulary_word2index_word[word_content] ### 对一个词编码
vec_content_np = np.array(vec_content, dtype=float)
vec_line[i - 1] = vec_content_np
total_data.append(vec_line)
X = kr.preprocessing.sequence.pad_sequences(X_char, sequence_length)
X_word = np.array(total_data)
Y_all = kr.utils.to_categorical(Y, num_classes=len(cat_to_id))
return X,X_word,Y_all
def TextMix_train():
########### load data ###################
if not os.path.exists(FLAGS.vocab_dir):
words = tools.build_vocab(train_data=FLAGS.train_data, vocab_dir=FLAGS.vocab_dir) ### 制作词汇表
else:
words = tools.read_file(FLAGS.vocab_dir)
vocab_size = len(words)
train_char_X, train_word_X,train_Y = preprocess_data(vocab_data=FLAGS.vocab_dir,word2vec=FLAGS.word2vec,
stop_word_dir=FLAGS.stop_word,train_data=FLAGS.train_data,sequence_length=config.seq_length)
val_char_X, val_word_X, val_Y = preprocess_data(vocab_data=FLAGS.vocab_dir, word2vec=FLAGS.word2vec,
stop_word_dir=FLAGS.stop_word, train_data=FLAGS.val_data,
sequence_length=config.seq_length)
print("Data deal down!")
###########################################
input_x_char = tf.placeholder(tf.int32, [None, config.seq_length], name='input_x')
input_x_word = tf.placeholder(tf.float32, [None, config.seq_length, config.word2vec_size], name='input_x')
input_y = tf.placeholder(tf.float32, [None, config.num_classes], name='input_y')
######### Char Model ####################################################
textcnn = TM.TextCNN_Char(config, vocab_size, keep_prob=config.dropout_keep_prob)
textrcnn = TM.TextRCNN_Char(config, vocab_size, keep_prob=config.dropout_keep_prob)
logits_cnn = textcnn.cnn_inference(input_x_char) ### (?,10)
logits_rcnn = textrcnn.RCNN_inference(input_x_char) ### (?,10)
######## Word Model ########################################################
textrcnn = TM.TextRCNN_word(config, keep_prob=config.dropout_keep_prob)
logits_RCNN = textrcnn.RCNN_inference(input_x_word) ### (?,10)
textcnn = TM.Text_word_CNN(config, keep_prob=config.dropout_keep_prob)
logits_CNN = textcnn.cnn(input_x_word) ### (?,10)
###### 模型融合
logits = logits_cnn + logits_rcnn + logits_RCNN + logits_CNN
############# 计算 loss 和 acc ######################################
loss = TM.Loss(logits=logits, label=input_y)
acc = TM.Acc(logits=logits, labels=input_y)
global_step = tf.Variable(0, name='global_step', trainable=False)
#learning_rate = FLAGS.learning_rate
learning_rate = tf.train.exponential_decay(
learning_rate=FLAGS.learning_rate,
global_step=global_step,
decay_steps=2500,
decay_rate=0.5,
staircase=True)
optim = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(loss=loss,global_step=global_step)
#optim = tf.train.MomentumOptimizer(learning_rate=learning_rate,momentum=0.9).minimize(loss=loss,global_step=global_step)
tensorboard_dir = 'tensorboard/textcnn_word'
tf.summary.scalar("loss", loss)
merged_summary = tf.summary.merge_all()
writer = tf.summary.FileWriter(tensorboard_dir)
saver = tf.train.Saver(max_to_keep=3) ### 保存模型
model_save_dir = 'checkpoints/'
train_start_time = time.strftime('%Y-%m-%d-%H-%M-%S', time.localtime(time.time()))
model_name = 'TextMixNet_{:s}.ckpt'.format(str(train_start_time))
model_save_path = os.path.join(model_save_dir, model_name)
logging.basicConfig(filename='./checkpoints/' + model_name + '.log',
format='%(asctime)s - %(pathname)s - %(levelname)s: %(message)s',
level=logging.DEBUG, filemode='a', datefmt='%Y-%m-%d%I:%M:%S %p')
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
sess = tf.Session(config=sess_config)
with sess.as_default():
if FLAGS.model_restore:
weights_path = './checkpoints/TextWordRCNNnet_2019-11-06-16-07-55.ckpt-12500'
saver.restore(sess=sess, save_path=weights_path)
step = sess.run(tf.train.get_global_step())
writer.add_graph(sess.graph)
infor = '##### Restore model : ' + weights_path + ' ########'
logging.info(infor)
print(infor)
else:
step = 0
init = tf.global_variables_initializer()
sess.run(init)
writer.add_graph(sess.graph)
print('First step is:', step)
num_batch = int((len(train_char_X) - 1) / config.batch_size) + 1 ### 总batch数
acc_begain = 0
for epoch in range(config.epochs):
batch_train = tools.batch_iter_Mix(train_char_X, train_word_X,train_Y, config.batch_size) ### 生成批次数据
Begain_learn_rate = FLAGS.learning_rate
for x_char_batch,x_word_batch, y_batch in batch_train:
step += 1
_, learn_rate, train_loss_value, train_pred, train_acc, merge_summary_value = sess.run(
[optim, learning_rate, loss, logits, acc, merged_summary],
feed_dict={input_x_char: x_char_batch,input_x_word:x_word_batch, input_y: y_batch})
############# 输出 learning_rate
if Begain_learn_rate != learn_rate:
information = '############ New Learning_Rate {:6f} in step {:d} ###########'.format(learn_rate,
step)
logging.info(information)
print(information)
Begain_learn_rate = learn_rate
if step % 10 == 0:
information = '## Epoch {:d} Step_Train / Total_Batch: {:d} / {:d} train_loss= {:5f} train_acc={:5f}'. \
format(epoch, step, num_batch, train_loss_value, train_acc)
print(information) ### 输出到屏幕
logging.info(information) ### 输出到log文件
if step % 500 == 0: ### 每 500 步进行一次验证，并保存最优模型
val_acc_all = 0
val_loss_all = 0
val_step = 0
batch_val = tools.batch_iter_Mix(val_char_X, val_word_X, val_Y, config.batch_size) ### 生成批次数据
for x_char_val,x_word_val, y_val in batch_val:
if x_char_val.shape[0] < config.batch_size:
pass
else:
_, val_loss_value, val_pred, val_acc, merge_summary_value = sess.run(
[optim, loss, logits, acc, merged_summary], feed_dict={input_x_char: x_char_val,
input_x_word:x_word_val,input_y: y_val})
writer.add_summary(merge_summary_value, step)
val_acc_all = val_acc_all + val_acc
val_loss_all = val_loss_all + val_loss_value
val_step += 1
ave_acc = val_acc_all / val_step
ave_loss = val_loss_all / val_step
information = "############## Val_loss = {:5f} Val_acc = {:5f} #####################". \
format(ave_loss, ave_acc)
print(information) ### 输出到屏幕
logging.info(information) ### 输出到log文件
if (ave_acc - acc_begain) > 0.0005:
saver.save(sess, model_save_path, global_step=step)
acc_begain = ave_acc
详细代码，可以去我的 github 里面参考。记得给个星星哦
代码所用到的word2vec模型，本文这里公开了自己使用的word2vec模型，也可以自己根据自己的数据集制作词向量模型。 链接：
https://pan.baidu.com/s/1dYRIPVH1N0y0gNOCw8rAuQ 提取码：g889
具体训练情况如下：
最后
本小白的四系统较单一的CNN系统或者其他系统，其效果提升大概2个点以上（具体看应用场景）。相应的，其效率较其他单一系统略低。毕竟效率和效果之间总是要平衡的。
最后的最后，github网址：
https://github.com/Tian14267/TextClassify_FourSystems
欢迎各位英雄好汉 ，前来交流学习，共同进步。
扫一扫
2775
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
