首先说明使用的工具和环境:python3.6.8 tensorflow1.14.0 centos7.0(最好用Ubuntu)
关于环境的搭建只做简单说明,我这边是使用pip搭建了python的虚拟环境(virtualenv),并在虚拟环境中安装tensorflow。详细步骤可以查看tensorflow的官网。
注:本文参考于 基于tensorflow、CNN、清华数据集THUCNews的新浪新闻文本分类
训练数据
训练(train.txt)和测试(test.txt)数据,两个文件的分类相同为100个分类,其中test.txt每个类下有200条测试数据,train.txt每个类下有1800条训练数据;数据共有两列,第一列为标签信息 第二列为标题,见下图
百度云链接:https://pan.baidu.com/s/1MZX8SOJ7lerov_UqZhSWeQ
提取码:9nvj
训练代码
闲话少说直接上代码,支持训练模型固化,代码粘贴前都经过了测试,均可复现,并且在相应位置给出了注释,有任何疑问欢迎留言,不忙就会回复。
###################cnn训练代码################# #coding=utf8 import os import codecs import random import heapq import numpy as np import pandas as pd import tensorflow as tf import tensorflow.contrib.keras as kr from collections import Counter #简单的计数器 用于统计字符出现的个数 from sklearn.preprocessing import LabelEncoder #标签编码 from sklearn.metrics import confusion_matrix #混淆矩阵 labelEncoder = LabelEncoder() #标签编码 #输出格式显示设置 #pd.set_option('display.max_columns', None) #显示所有列 pd.set_option('display.max_rows', None) #显示所有行 #pd.set_option('max_colwidth', 500) #设置value的显示长度为100,默认为50 np.set_printoptions(threshold=np.inf) #设置print输出完整性 os.environ["CUDA_VISIBLE_DEVICES"] = "1" #设置只有一个gpu可见 #cnn参数设置 可根据实际情况自行修改 vocab_size = 5000 #词汇表大小 seq_length = 100 #标题序列长度 embedding_dim = 64 #词向量维度 num_classes = 100 #类别数(初始值,后边会根据具体训练数据类目数量修改) num_filters = 128 #卷积核数目 kernel_size = 5 #卷积核尺寸 hidden_dim = 128 #全连接层神经元 dropout_keep_prob = 0.5 #dropout保留比例 learning_rate = 1e-3 #学习率 batch_size = 100 #每批训练大小 with open('train.txt', encoding='utf8') as file: #加载训练数据 line_list = [k.strip() for k in file.readlines()] #提取训练数据中的一行 train_label_list = [k.split()[0] for k in line_list] #提取标签数据 train_content_list = [k.split(maxsplit=1)[1] for k in line_list] #提取标题数据 def getVocabularyList(content_list, vocabulary_size): allContent_str = ''.join(content_list) counter = Counter(allContent_str) vocabulary_list = [k[0] for k in counter.most_common(vocabulary_size)] return vocabulary_list def makeVocabularyFile(content_list, vocabulary_size): vocabulary_list = getVocabularyList(content_list, vocabulary_size) with open('vocab_last.txt', 'w', encoding='utf8') as file: for vocabulary in vocabulary_list: file.write(vocabulary + '\n') #makeVocabularyFile(train_content_list, 5000) #根据训练数据集创建新的 单字表 with open('vocab.txt', encoding='utf8') as file: #加载 词汇表(单字表) 数据 vocabulary_list = [k.strip() for k in file.readlines()] word2id_dict = dict([(b ,a) for a,b in enumerate(vocabulary_list)]) #单字 与 id对照表 content2idList = lambda content : [word2id_dict[word] for word in content if word in word2id_dict] train_idlist_list = [content2idList(content) for content in train_content_list] #标题训练id列表 train_X = kr.preprocessing.sequence.pad_sequences(train_idlist_list, seq_length) #按照seq_length补全和截断 训练数据 train_y = labelEncoder.fit_transform(train_label_list) #train_y 所有的训练标签数据 做编码 num_classes = len(labelEncoder.classes_) #根据训练数据重新定义类目数量 #保存类目预测标签 y_lable = codecs.open('y_lable_last.txt', 'w', 'utf-8') for label in labelEncoder.classes_: str1 = label + '\n' y_lable.write(str1) y_lable.close() print('参与训练的类目数量', num_classes) train_Y = kr.utils.to_categorical(train_y, num_classes) #标签的one-hot #搭建神经网络 这里使用的是默认的W、b tf.reset_default_graph() X_holder = tf.placeholder(tf.int32, [None, seq_length]) Y_holder = tf.placeholder(tf.float32, [None, num_classes]) embedding = tf.get_variable('embedding', [vocab_size, embedding_dim]) #vocab_size*embedding_dim 矩阵形状 5000*64 embedding_inputs = tf.nn.embedding_lookup(embedding, X_holder) #batch_size*seq_length*embedding_dim 100*100*64 conv = tf.layers.conv1d(embedding_inputs, num_filters, kernel_size) #形状为batch_size*(100-5+1)*num_filter 64*96*128 这里的100指标题序列长度(seq_length) max_pooling = tf.reduce_max(conv, reduction_indices=[1]) #最大值池化 形状为batch_size*num_filter 100*128 full_connect = tf.layers.dense(max_pooling, hidden_dim) #添加全连接层 形状为batch_size*hidden_dim 100*128 full_connect_dropout = tf.contrib.layers.dropout(full_connect, dropout_keep_prob) #防止全连接过拟合 full_connect_activate = tf.nn.relu(full_connect_dropout) #全连接激活函数 softmax_before = tf.layers.dense(full_connect_activate, num_classes) #添加全连接层形状为batch_size*num_classes 100*100 predict_Y = tf.nn.softmax(softmax_before) #softmax方法给出预测概率值 cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(labels=Y_holder, logits=softmax_before) #交叉熵作为损失函数 loss = tf.reduce_mean(cross_entropy) #反向传播计算损失值 optimizer = tf.train.AdamOptimizer(learning_rate) #优化器 train = optimizer.minimize(loss) #最小化损失 isCorrect = tf.equal(tf.argmax(Y_holder, 1), tf.argmax(predict_Y, 1)) #预算准确率 accuracy = tf.reduce_mean(tf.cast(isCorrect, tf.float32)) #参数初始化 对于神经网络模型,重要是其中的W、b这两个参数 init = tf.global_variables_initializer() session = tf.Session() session.run(init) #测试数据加载、转换 with open('test.txt', encoding='utf8') as file: line_list = [k.strip() for k in file.readlines()] test_label_list = [k.split()[0] for k in line_list] test_content_list = [k.split(maxsplit=1)[1] for k in line_list] test_idlist_list = [content2idList(content) for content in test_content_list] test_X = kr.preprocessing.sequence.pad_sequences(test_idlist_list, seq_length) test_y = labelEncoder.transform(test_label_list) test_Y = kr.utils.to_categorical(test_y, num_classes) #做one-hot for i in range(10000): #表示模型迭代的次数 selected_index = random.sample(list(range(len(train_y))), k=batch_size) batch_X = train_X[selected_index] batch_Y = train_Y[selected_index] session.run(train, {X_holder:batch_X, Y_holder:batch_Y}) #模型训练 step = i+1 if step % 100 == 0: selected_index = random.sample(list(range(len(test_y))), k=100) #在测试集中随机抽取100行进行测试 batch_X = test_X[selected_index] batch_Y = test_Y[selected_index] loss_value, accuracy_value = session.run([loss, accuracy], {X_holder:batch_X, Y_holder:batch_Y}) print('step:%d loss:%.4f accuracy:%.4f' %(step, loss_value, accuracy_value)) #保存模型 saver = tf.train.Saver() save_path = saver.save(session, 'train_model/fenlei_cnn.ckpt') print('Save to path:', save_path) def predictAll(test_X, data_size=100): predict_value_list = [] for i in range(0, len(test_X), data_size): selected_X = test_X[i:i+data_size] predict_value = session.run(predict_Y, {X_holder:selected_X}) predict_value_list.extend(predict_value) return np.array(predict_value_list) #预测全部的测试数据 Y = predictAll(test_X) #可修改Y的提取数量 y = np.argmax(Y, axis=1) predict_label_list = labelEncoder.inverse_transform(y) df=pd.DataFrame(confusion_matrix(test_label_list, predict_label_list), columns = labelEncoder.classes_, index = labelEncoder.classes_ ) #print(df) ###报告表 import numpy as np from sklearn.metrics import precision_recall_fscore_support def eval_model(y_true, y_pred, labels): #计算每个分类的Precision, Recall, f1, support p,r,f1,s = precision_recall_fscore_support(y_true, y_pred) #计算总体的平局Precision,recall,f1,support tot_p = np.average(p, weights=s) tot_r = np.average(r, weights=s) tot_f1 = np.average(f1, weights=s) tot_s = np.sum(s) res1 = pd.DataFrame({ u'Label':labels, u'Precision':p, u'Recall':r, u'F1':f1, u'Support':s }) res2 = pd.DataFrame({ u'Label':['总体'], u'Precision':[tot_p], u'Recall':[tot_r], u'F1':[tot_f1], u'Support':[tot_s] }) res2.index = [999] res = pd.concat([res1, res2]) return res[['Label', 'Precision', 'Recall', 'F1', 'Support']] tables = eval_model(test_label_list, predict_label_list, labelEncoder.classes_) print(tables)
预测代码
python预测代码,代码支持批量预测和单条测试
执行方式:python cnn_predict.py debug或python cnn_predict.py batch
#coding=utf8 import tensorflow as tf import os import sys import time import codecs import random import heapq import numpy as np import tensorflow.contrib.keras as kr from sklearn.preprocessing import LabelEncoder #标签编码 labelEncoder = LabelEncoder() #标签编码 os.environ["CUDA_VISIBLE_DEVICES"] = "1" #设置只有一个gpu可见
#可根据实际情况自行修改 vocab_size = 5000 #词汇表大小 seq_length = 100 #序列长度 embedding_dim = 64 #词向量维度 num_classes = 100 #类别数 num_filters = 128 #卷积核数目 kernel_size = 5 #卷积核尺寸 hidden_dim = 128 #全连接层神经元 dropout_keep_prob = 1 #dropout保留比例 注意:这里要保留为1 与训练参数的差距 learning_rate = 1e-3 #学习率 batch_size = 100 #每批训练大小 np.set_printoptions(threshold=np.inf) #设置print输出完整性 with open('vocab.txt', encoding='utf8') as file: #加载 词汇表(单字表) 数据 vocabulary_list = [k.strip() for k in file.readlines()] word2id_dict = dict([(b ,a) for a,b in enumerate(vocabulary_list)]) #单字 与 id对照表 content2idList = lambda content : [word2id_dict[word] for word in content if word in word2id_dict] with open('y_lable.txt', encoding='utf8') as file: train_label_list = [k.strip() for k in file.readlines()] labelEncoder.fit_transform(train_label_list) #所有的训练标签数据 做编码 #搭建神经网络 tf.reset_default_graph() X_holder = tf.placeholder(tf.int32, [None, seq_length]) Y_holder = tf.placeholder(tf.float32, [None, num_classes]) embedding = tf.get_variable('embedding', [vocab_size, embedding_dim]) #vocab_size*embedding_dim 矩阵形状 5000*64 embedding_inputs = tf.nn.embedding_lookup(embedding, X_holder) #batch_size*seq_length*embedding_dim 100*100*64 conv = tf.layers.conv1d(embedding_inputs, num_filters, kernel_size) #形状为batch_size*(100-5+1)*num_filter 64*96*128 max_pooling = tf.reduce_max(conv, reduction_indices=[1]) #最大值池化 形状为batch_size*num_filter 64*128 full_connect = tf.layers.dense(max_pooling, hidden_dim) #添加全连接层 形状为batch_size*hidden_dim 64*128 full_connect_dropout = tf.contrib.layers.dropout(full_connect, dropout_keep_prob) #防止全连接过拟合 full_connect_activate = tf.nn.relu(full_connect_dropout) #全连接激活函数 softmax_before = tf.layers.dense(full_connect_activate, num_classes) #添加全连接层形状为batch_size*num_classes 100*100 predict_Y = tf.nn.softmax(softmax_before) #softmax方法给出预测概率值 cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(labels=Y_holder, logits=softmax_before) #交叉熵作为损失函数 loss = tf.reduce_mean(cross_entropy) #反向传播计算损失值 optimizer = tf.train.AdamOptimizer(learning_rate) #优化器 train = optimizer.minimize(loss) #最小化损失 isCorrect = tf.equal(tf.argmax(Y_holder, 1), tf.argmax(predict_Y, 1)) #预算准确率 accuracy = tf.reduce_mean(tf.cast(isCorrect, tf.float32)) session = tf.Session() #加载预测模型 saver = tf.train.Saver() saver.restore(session, 'train_model/fenlei_cnn.ckpt') print('load model succesful') def predictAll(test_X, data_size=100): predict_value_list = [] for i in range(0, len(test_X), data_size): selected_X = test_X[i:i+data_size] predict_value = session.run(predict_Y, {X_holder:selected_X}) predict_value_list.extend(predict_value) return np.array(predict_value_list) #给出五个预测结果,及预测分数,测试准确率 def format_predict5(Y): y_index = [] y_value = [] for Y_l in Y: index_l = heapq.nlargest(5, range(len(Y_l)), Y_l.take) #获取前五个下标 value_l = heapq.nlargest(5, Y_l) #获取前五个类目数值 y_index.append(index_l) y_value.append(value_l) #print('前五个类目得分:',value_l) for i in range(0,len(test_id_list)): num=0 flag=0 #判断预测类目阈值 for n in range(5): if test_label_list[i] in train_label_list and y_value[i][n] > 0.1: num+=1 elif y_value[i][n] > 0.9: #如果原类目没有训练 则阈值要大于0.9 num+=1 flag=1 if num == 0: continue #把名称index转换成name pre_Yname = [] for ii in range(num): pre_Yname.append(labelEncoder.classes_[y_index[i][ii]]) #判断源目录与预测类目相同 if test_label_list[i] in pre_Yname: continue str1 = test_id_list[i]+'\t'+test_content_list[i]+'\t'+test_label_list[i]+'\t'+ str(flag) for j in range(num): str1 += '\t'+pre_Yname[j]+'('+str(y_value[i][j]) +')' str1+='\n' fo.write(str1) if __name__ == "__main__": if len(sys.argv) != 2: print ('python cnn_predict.py batch(debug)') exit(1) if sys.argv[1] == 'batch': _time = time.strftime("%m%d%H%M", time.localtime()) fo = codecs.open('test_result'+_time+'.txt', 'w', 'utf8') batch_line = [] #批处理检测 满足1000条进行一次批处理 b_size = 0 with codecs.open('test.txt', 'rb', 'utf8', 'ignore') as file: for line in file: datalist1 = line.strip().split('\t') if len(datalist1) != 3: # or datalist[1] not in train_label_list: continue batch_line.append(line.strip()) b_size+=1 if b_size == 1000: test_id_list = [k.split('\t')[0] for k in batch_line] test_label_list = [k.split('\t')[2] for k in batch_line] test_content_list = [k.split('\t')[1] for k in batch_line] test_idlist_list = [content2idList(content) for content in test_content_list] test_X = kr.preprocessing.sequence.pad_sequences(test_idlist_list, seq_length) #预测 Y = predictAll(test_X) format_predict5(Y) batch_line.clear() b_size = 0 fo.close() elif sys.argv[1] == 'debug': while(1): title = input("title:") if not title.strip(): continue title_idlist_list = [content2idList(title.strip())] test_X = kr.preprocessing.sequence.pad_sequences(title_idlist_list, seq_length) selected_X = test_X[0:10] predict_value = session.run(predict_Y, {X_holder:selected_X}) index_l = heapq.nlargest(10, range(len(predict_value[0])), predict_value[0].take) #获取前十个数下标 value_l = heapq.nlargest(10, predict_value[0]) for i in range(10): line = ' '+str(i+1)+'. '+labelEncoder.classes_[index_l[i]]+'('+str(value_l[i]) +')' print (line)
最后贴几张结果:
训练(上图)
准确率 召回率 f1(上图)
debug单条给出十个预测结果(上图)
分类代码以字为单位,没有进行分词,两部分代码可单独运行,100个类的测试集平均准确率可达到95%,有问题欢迎留言。