中文文本分类之TextRNN
RNN模型由于具有短期记忆功能，因此天然就比较适合处理自然语言等序列问题，尤其是引入门控机制后，能够解决长期依赖问题，捕获输入样本之间的长距离联系。本文的模型是堆叠两层的LSTM和GRU模型，模型的结构为：LSTM（GRU）—dropout—LSTM（GRU）—dropout—全连接层—输出层，比较简单。关于TensorFlow搭建RNN模型有关的内容，在这篇《TensorFlow之RNN：堆叠RNN、LSTM、GRU及双向LSTM》博客里阐述得比较清楚了，这里不赘述。
尽管RNN模型天然比较适合处理自然语言的问题，可是最近CNN模型有迎头赶上之势。为什么呢？从这次文本分类的任务中可以体会到，RNN模型在运行速度上丝毫不占优势，比CNN模型要慢几倍到十几倍。后一个时间步的输出依赖于前一个时间步的输出，无法进行并行处理，导致模型训练的速度慢，这是一个致命的弱点。而RNN模型引以为傲的能够捕获序列中的长距离依赖关系，已经不再是独门秘诀，因为CNN模型的卷积操作就类似于N-gram，可以捕获上下文关系，而且通过把构建更深层的卷积层，可以捕获更长距离的依赖关系。此外，Transformer横空出世，不仅能够进行并行处理，而且通过自注意力机制能够在任意距离的两个词之间建立依赖关系，大有后浪把前浪拍死在沙滩上的趋势。
另外，从模型预测的准确性来讲，CNN模型的准确性不比RNN模型低，甚至超过了RNN模型。
TextRNN模型依然分为四个模块：1、数据处理模块；2、模型构建模块；3、模型训练模快；4、模型预测模块。
GitHub地址：https://github.com/DengYangyong/Chinese_Text_Classification/tree/master/Text-Classification-On_RNN
好，下面看代码。
一、数据处理
数据处理部分和上一篇CharCNN是一样的，尽管我们说RNN模型可以处理任意长度的序列，但是在这个TextRNN模型中，我们还是把输入处理成了固定长度的序列。
#coding: utf-8 import sys from collections import Counter import numpy as np import tensorflow.contrib.keras as kr if sys.version_info[0] > 2: is_py3 = True else: reload(sys) sys.setdefaultencoding("utf-8") is_py3 = False # 判断软件的版本，如果版本为3.6.5，那么sys.version_info的输出为：sys.version_info(major=3, minor=6, micro=5)。 """如果在python2下面使用python3训练的模型，可考虑调用此函数转化一下字符编码""" def native_word(word, encoding='utf-8'): if not is_py3: return word.encode(encoding) else: return word """is_py3函数当版本为3时返回True，否则返回False。if not 后面的值为False则将“utf-8”编码转换为'unicode'.""" def native_content(content): if not is_py3: return content.decode('utf-8') else: return content """ 常用文件操作，可在python2和python3间切换.""" def open_file(filename, mode='r'): if is_py3: return open(filename, mode, encoding='utf-8', errors='ignore') else: return open(filename, mode) """ 读取文件数据""" def read_file(filename): contents, labels = [], [] with open_file(filename) as f: for line in f: try: label, content = line.strip().split('\t') if content: contents.append(list(native_content(content))) labels.append(native_content(label)) except: pass return contents, labels # line.strip().split('\t')的输出为两个元素的列表：['体育', '黄蜂vs湖人首发：科比带伤战保罗 加索尔救赎之战 新浪体育讯...']。 # 注意这个list()函数，把一段文字转化为了列表，元素为每个字和符号：['黄', '蜂', 'v', 's', '湖', '人', '首', '发', '：', '科', '比',...] # contents的元素为每段新闻转化成的列表：[['黄', '蜂', 'v', 's', '湖', '人', '首', '发', '：', '科', '比',...],[],...] # labels为['体育', '体育',...] """根据训练集构建词汇表，存储""" def build_vocab(train_dir, vocab_dir, vocab_size=5000): data_train, _ = read_file(train_dir) all_data = [] for content in data_train: all_data.extend(content) counter = Counter(all_data) count_pairs = counter.most_common(vocab_size - 1) words, _ = list(zip(*count_pairs)) words = ['<PAD>'] + list(words) open_file(vocab_dir, mode='w').write('\n'.join(words) + '\n') '''读取词汇表''' def read_vocab(vocab_dir): with open_file(vocab_dir) as fp: words = [native_content(_.strip()) for _ in fp.readlines()] word_to_id = dict(zip(words, range(len(words)))) return words, word_to_id # readlines()读取所有行然后把它们作为一个字符串列表返回:['头\n', '天\n', ...]。strip()函数去掉"\n"。 # words: ['<PAD>', '，', '的', '。', '一', '是', '在', '0', '有',...] # word_to_id：{'<PAD>': 0, '，': 1, '的': 2, '。': 3, '一': 4, '是': 5,..}，每个类别对应的value值为其索引ID """读取分类目录""" def read_category(): categories = ['体育', '财经', '房产', '家居', '教育', '科技', '时尚', '时政', '游戏', '娱乐'] categories = [native_content(x) for x in categories] cat_to_id = dict(zip(categories, range(len(categories)))) return categories, cat_to_id # cat_to_id的输出为：{'体育': 0, '财经': 1, '房产': 2, '家居': 3,...}，每个类别对应的value值为其索引ID. """ 将id表示的内容转换为文字 """ def to_words(content, words): return ''.join(words[x] for x in content) """ 将文件转换为id表示,进行pad """ def process_file(filename, word_to_id, cat_to_id, max_length=600): contents, labels = read_file(filename) data_id, label_id = [], [] #contents的形式为：[['黄', '蜂', 'v', 's', '湖', '人',...],[],[],...]，每一个元素是一个列表，该列表的元素是每段新闻的字和符号。 #labels的形式为：['体育', '体育', '体育', '体育', '体育', ...] for i in range(len(contents)): data_id.append([word_to_id[x] for x in contents[i] if x in word_to_id]) label_id.append(cat_to_id[labels[i]]) x_pad = kr.preprocessing.sequence.pad_sequences(data_id, max_length) y_pad = kr.utils.to_categorical(label_id, num_classes=len(cat_to_id)) return x_pad, y_pad # word_to_id是一个字典：{'<PAD>': 0, '，': 1, '的': 2, '。': 3, '一': 4, '是': 5,...} # 对于每一段新闻转化的字列表，把每个字在字典中对应的索引找到： # data_id: 将[['黄', '蜂', 'v', 's', '湖', '人',...],[],[],...] 转化为 [[387, 1197, 2173, 215, 110, 264,...],[],[],...]的形式 # label_id : ['体育', '体育', '体育', '体育', '体育', ...] 转化为[0, 0, 0, 0, 0, ...] # data_id的行数为50000，即为新闻的条数，每个元素为由每段新闻的字的数字索引构成的列表； # data_id长这样：[[387, 1197, 2173, 215, 110, 264,...],[],[],...] # 由于每段新闻的字数不一样，因此每个元素（列表）的长度不一样，可能大于600，也可能小于600，需要统一长度为600。 # 使用keras提供的pad_sequences来将文本pad为固定长度，x_pad的形状为（50000，600）. # label_id是形如[0, 0, 0, 0, 0, ...]的整形数组，cat_to_id是形如{'体育': 0, '财经': 1, '房产': 2, '家居': 3,...}的字典 # to_categorical是对标签进行one-hot编码，num-classes是类别数10，y_pad的维度是（50000，10） """生成批次数据""" def batch_iter(x, y, batch_size=64): data_len = len(x) num_batch = int((data_len - 1) / batch_size) + 1 indices = np.random.permutation(np.arange(data_len)) x_shuffle = x[indices] y_shuffle = y[indices] # 样本长度为50000 # int()可以将其他类型转化为整型，也可以用于向下取整，这里为782. # indices元素的范围是0-49999，形如[256,189,2,...]的拥有50000个元素的列表 # 用indices对样本和标签按照行进行重新洗牌，接着上面的例子，把第256行(从0开始计)放在第0行，第189行放在第1行. for i in range(num_batch): start_id = i * batch_size end_id = min((i + 1) * batch_size, data_len) yield x_shuffle[start_id:end_id], y_shuffle[start_id:end_id] # i=780时，end_id=781*64=49984; # 当i=781时，end_id=50000，因为782*64=50048>50000,所以最后一批取[49984:50000] # yield是生成一个迭代器，用for循环来不断生成下一个批量。 # 为了防止内存溢出，每次只取64个，内存占用少。
二、模型搭建
模型比较简单，下面的代码也就是按照LSTM（GRU）—dropout—LSTM（GRU）—dropout—全连接层—输出层这样的结构来进行组织的。要注意的是对每层的LSTM或GRU核中的神经元进行dropout，还有取最后时刻和最后一层的LSTM或GRU的隐状态作为全连接层的输入。
#!/usr/bin/python # -*- coding: utf-8 -*- import tensorflow as tf class TRNNConfig(object): """RNN配置参数""" embedding_dim = 64 seq_length = 600 num_classes = 10 vocab_size = 5000 num_layers= 2 hidden_dim = 128 rnn = 'gru' # 隐藏层层数为2 # 选择lstm 或 gru dropout_keep_prob = 0.8 learning_rate = 1e-3 batch_size = 128 num_epochs = 10 print_per_batch = 100 save_per_batch = 10 class TextRNN(object): """文本分类，RNN模型""" def __init__(self, config): self.config = config self.input_x = tf.placeholder(tf.int32, [None, self.config.seq_length], name='input_x') self.input_y = tf.placeholder(tf.float32, [None, self.config.num_classes], name='input_y') self.keep_prob = tf.placeholder(tf.float32, name='keep_prob') self.rnn() def rnn(self): """rnn模型""" def lstm_cell(): return tf.nn.rnn_cell.LSTMCell(self.config.hidden_dim, state_is_tuple=True) def gru_cell(): return tf.nn.rnn_cell.GRUCell(self.config.hidden_dim) def dropout(): if (self.config.rnn == 'lstm'): cell = lstm_cell() else: cell = gru_cell() return tf.nn.rnn_cell.DropoutWrapper(cell, output_keep_prob=self.keep_prob) # 为每一个rnn核后面加一个dropout层 with tf.device('/gpu:0'): embedding = tf.get_variable('embedding', [self.config.vocab_size, self.config.embedding_dim]) embedding_inputs = tf.nn.embedding_lookup(embedding, self.input_x) with tf.name_scope("rnn"): cells = [dropout() for _ in range(self.config.num_layers)] rnn_cell = tf.nn.rnn_cell.MultiRNNCell(cells, state_is_tuple=True) # 堆叠了2层的RNN模型。 _outputs, _ = tf.nn.dynamic_rnn(cell=rnn_cell, inputs=embedding_inputs, dtype=tf.float32) last = _outputs[:, -1, :] # 取最后一个时序输出作为结果，也就是最后时刻和第2层的LSTM或GRU的隐状态。 with tf.name_scope("score"): # 全连接层，后面接dropout以及relu激活 fc = tf.layers.dense(last, self.config.hidden_dim, name='fc1') fc = tf.contrib.layers.dropout(fc, self.keep_prob) fc = tf.nn.relu(fc) self.logits = tf.layers.dense(fc, self.config.num_classes, name='fc2') self.y_pred_cls = tf.argmax(tf.nn.softmax(self.logits), 1) with tf.name_scope("optimize"): cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=self.logits, labels=self.input_y) self.loss = tf.reduce_mean(cross_entropy) self.optim = tf.train.AdamOptimizer(learning_rate=self.config.learning_rate).minimize(self.loss) with tf.name_scope("accuracy"): correct_pred = tf.equal(tf.argmax(self.input_y, 1), self.y_pred_cls) self.acc = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
三、模型训练、验证和测试
这一部分的代码和CharCNN可以说没啥区别。注意在验证和测试时不用做dropout，还有用早停防止过拟合。
# coding: utf-8 from __future__ import print_function import os import sys import time from datetime import timedelta import numpy as np import tensorflow as tf from sklearn import metrics from rnn_model import TRNNConfig, TextRNN from cnews_loader import read_vocab, read_category, batch_iter, process_file, build_vocab base_dir = 'data/cnews' train_dir = os.path.join(base_dir, 'cnews.train.txt') test_dir = os.path.join(base_dir, 'cnews.test.txt') val_dir = os.path.join(base_dir, 'cnews.val.txt') vocab_dir = os.path.join(base_dir, 'cnews.vocab.txt') save_dir = 'checkpoints/textrnn' save_path = os.path.join(save_dir, 'best_validation') # 最佳验证结果保存路径 def get_time_dif(start_time): """获取已使用时间""" end_time = time.time() time_dif = end_time - start_time return timedelta(seconds=int(round(time_dif))) def feed_data(x_batch, y_batch, keep_prob): feed_dict = { model.input_x: x_batch, model.input_y: y_batch, model.keep_prob: keep_prob } return feed_dict def evaluate(sess, x_, y_): """评估在某一数据上的准确率和损失""" data_len = len(x_) batch_eval = batch_iter(x_, y_, 128) total_loss = 0.0 total_acc = 0.0 for x_batch, y_batch in batch_eval: batch_len = len(x_batch) feed_dict = feed_data(x_batch, y_batch, 1.0) # 在测试时不用进行dropout y_pred_class,loss, acc = sess.run([model.y_pred_cls,model.loss, model.acc], feed_dict=feed_dict) total_loss += loss * batch_len total_acc += acc * batch_len return y_pred_class,total_loss / data_len, total_acc / data_len def train(): print("Configuring TensorBoard and Saver...") tensorboard_dir = 'tensorboard/textrnn' if not os.path.exists(tensorboard_dir): os.makedirs(tensorboard_dir) tf.summary.scalar("loss", model.loss) tf.summary.scalar("accuracy", model.acc) merged_summary = tf.summary.merge_all() writer = tf.summary.FileWriter(tensorboard_dir) # 配置 Saver saver = tf.train.Saver() if not os.path.exists(save_dir): os.makedirs(save_dir) print("Loading training and validation data...") # 载入训练集与验证集 start_time = time.time() x_train, y_train = process_file(train_dir, word_to_id, cat_to_id, config.seq_length) x_val, y_val = process_file(val_dir, word_to_id, cat_to_id, config.seq_length) time_dif = get_time_dif(start_time) print("Time usage:", time_dif) # 创建session session = tf.Session() session.run(tf.global_variables_initializer()) writer.add_graph(session.graph) print('Training and evaluating...') start_time = time.time() total_batch = 0 best_acc_val = 0.0 last_improved = 0 require_improvement = 1000 # 如果超过1000轮未提升，提前结束训练 flag = False for epoch in range(config.num_epochs): print('Epoch:', epoch + 1) batch_train = batch_iter(x_train, y_train, config.batch_size) for x_batch, y_batch in batch_train: feed_dict = feed_data(x_batch, y_batch, config.dropout_keep_prob) if total_batch % config.save_per_batch == 0: s = session.run(merged_summary, feed_dict=feed_dict) writer.add_summary(s, total_batch) if total_batch % config.print_per_batch == 0: feed_dict[model.keep_prob] = 1.0 loss_train, acc_train = session.run([model.loss, model.acc], feed_dict=feed_dict) y_pred_cls_1,loss_val, acc_val = evaluate(session, x_val, y_val) # todo if acc_val > best_acc_val: # 保存最好结果 best_acc_val = acc_val last_improved = total_batch saver.save(sess=session, save_path=save_path) improved_str = '*' else: improved_str = '' time_dif = get_time_dif(start_time) msg = 'Iter: {0:>6}, Train Loss: {1:>6.2}, Train Acc: {2:>7.2%},' \ + ' Val Loss: {3:>6.2}, Val Acc: {4:>7.2%}, Time: {5} {6}' print(msg.format(total_batch, loss_train, acc_train, loss_val, acc_val, time_dif, improved_str)) session.run(model.optim, feed_dict=feed_dict) # 运行优化 total_batch += 1 if total_batch - last_improved > require_improvement: # 验证集正确率长期不提升，提前结束训练 print("No optimization for a long time, auto-stopping...") flag = True break if flag: break def test(): print("Loading test data...") start_time = time.time() x_test, y_test = process_file(test_dir, word_to_id, cat_to_id, config.seq_length) session = tf.Session() session.run(tf.global_variables_initializer()) saver = tf.train.Saver() saver.restore(sess=session, save_path=save_path) # 读取保存的模型 print('Testing...') y_pred,loss_test, acc_test = evaluate(session, x_test, y_test) msg = 'Test Loss: {0:>6.2}, Test Acc: {1:>7.2%}' print(msg.format(loss_test, acc_test)) batch_size = 128 data_len = len(x_test) num_batch = int((data_len - 1) / batch_size) + 1 y_test_cls = np.argmax(y_test, 1) y_pred_cls = np.zeros(shape=len(x_test), dtype=np.int32) for i in range(num_batch): start_id = i * batch_size end_id = min((i + 1) * batch_size, data_len) feed_dict = { model.input_x: x_test[start_id:end_id], model.keep_prob: 1.0 } y_pred_cls[start_id:end_id] = session.run(model.y_pred_cls, feed_dict=feed_dict) # 评估 print("Precision, Recall and F1-Score...") print(metrics.classification_report(y_test_cls, y_pred_cls, target_names=categories)) # 混淆矩阵 print("Confusion Matrix...") cm = metrics.confusion_matrix(y_test_cls, y_pred_cls) print(cm) time_dif = get_time_dif(start_time) print("Time usage:", time_dif) if __name__ == '__main__': print('Configuring RNN model...') config = TRNNConfig() if not os.path.exists(vocab_dir): build_vocab(train_dir, vocab_dir, config.vocab_size) categories, cat_to_id = read_category() words, word_to_id = read_vocab(vocab_dir) config.vocab_size = len(words) model = TextRNN(config) option='train' if option == 'train': train() else: test()
采用GRU核，训练了46分钟47秒，最好的验证精度为91.54%。测试精度为94.67%。上一篇的CNN模型训练只花了3分21秒，可见RNN模型在速度上慢了十几倍。
Iter: 3500, Train Loss: 0.034, Train Acc: 98.44%, Val Loss: 0.35, Val Acc: 91.54%, Time: 0:46:47 *
Testing...
Test Loss: 0.2, Test Acc: 94.67%
Precision, Recall and F1-Score...
precision recall f1-score support
体育 0.99 0.99 0.99 1000
财经 0.93 0.99 0.96 1000
房产 1.00 1.00 1.00 1000
家居 0.95 0.83 0.89 1000
教育 0.88 0.93 0.90 1000
科技 0.95 0.96 0.95 1000
时尚 0.95 0.95 0.95 1000
时政 0.95 0.91 0.93 1000
游戏 0.94 0.96 0.95 1000
娱乐 0.94 0.96 0.95 1000
micro avg 0.95 0.95 0.95 10000
macro avg 0.95 0.95 0.95 10000
weighted avg 0.95 0.95 0.95 10000
Confusion Matrix...
[[990 0 0 0 5 1 0 0 4 0]
[ 0 987 1 0 2 3 0 6 1 0]
[ 0 0 996 2 2 0 0 0 0 0]
[ 0 22 2 834 60 20 25 20 10 7]
[ 1 6 0 6 925 7 5 12 4 34]
[ 0 5 0 8 8 959 2 2 16 0]
[ 0 0 0 13 9 2 948 4 12 12]
[ 0 33 1 15 21 11 1 910 4 4]
[ 1 1 0 2 10 5 11 0 962 8]
[ 4 2 0 1 15 3 5 2 12 956]]
Time usage: 0:00:40
四、模型预测
从两个新闻中各摘取了一段内容，进行预测。结果预测为：科技、体育。
# coding: utf-8 from __future__ import print_function import os import tensorflow as tf import tensorflow.contrib.keras as kr from rnn_model import TRNNConfig, TextRNN from cnews_loader import read_category, read_vocab try: bool(type(unicode)) except NameError: unicode = str base_dir = 'data/cnews' vocab_dir = os.path.join(base_dir, 'cnews.vocab.txt') save_dir = 'checkpoints/textrnn' save_path = os.path.join(save_dir, 'best_validation') # 最佳验证结果保存路径 class RnnModel: def __init__(self): self.config = TRNNConfig() self.categories, self.cat_to_id = read_category() self.words, self.word_to_id = read_vocab(vocab_dir) self.config.vocab_size = len(self.words) self.model = TextRNN(self.config) self.session = tf.Session() self.session.run(tf.global_variables_initializer()) saver = tf.train.Saver() saver.restore(sess=self.session, save_path=save_path) # 读取保存的模型 def predict(self, message): content = unicode(message) data = [self.word_to_id[x] for x in content if x in self.word_to_id] feed_dict = { self.model.input_x: kr.preprocessing.sequence.pad_sequences([data], self.config.seq_length), self.model.keep_prob: 1.0 } y_pred_cls = self.session.run(self.model.y_pred_cls, feed_dict=feed_dict) return self.categories[y_pred_cls[0]] if __name__ == '__main__': rnn_model = RnnModel() test_demo = ['三星ST550以全新的拍摄方式超越了以往任何一款数码相机', '热火vs骑士前瞻：皇帝回乡二番战 东部次席唾手可得新浪体育讯北京时间3月30日7:00'] for i in test_demo: print(rnn_model.predict(i))
