一、概述
在英文分類的基礎上，再看看中文分類的，是一種10分類問題（體育，科技，游戲，財經，房產，家居等）的處理。
二、數據集合
數據集為新聞，總共有四個數據文件，在/data/cnews目錄下，包括內容如下圖所示測試集，訓練集和驗證集，和單詞表(最后的單詞表cnews.vocab.txt可以不要，因為訓練可以自動產生)。數據格式：前面為類別，后面為描述內容。
訓練數據地址：鏈接： https://pan.baidu.com/s/1ZHh98RrjQpG5Tm-yq73vBQ 提取碼：2r04
其中訓練集的格式：
vocab.txt的格式：每個字一行，其中前面加上PAD。
三、代碼
3.1 數據采集cnews_loader.py
1 # coding: utf-8 2 import sys 3 from collections import Counter 4 import numpy as np 5 import tensorflow.contrib.keras as kr 6 7 if sys.version_info[0] > 2: 8 is_py3 = True 9 else: 10 reload(sys) 11 sys.setdefaultencoding("utf-8") 12 is_py3 = False 13 14 def native_word(word, encoding='utf-8'): 15 """如果在python2下面使用python3訓練的模型，可考慮調用此函數轉化一下字符編碼""" 16 if not is_py3: 17 return word.encode(encoding) 18 else: 19 return word 20 21 def native_content(content): 22 if not is_py3: 23 return content.decode('utf-8') 24 else: 25 return content 26 27 def open_file(filename, mode='r'): 28 """ 29 常用文件操作，可在python2和python3間切換. 30 mode: 'r' or 'w' for read or write 31 """ 32 if is_py3: 33 return open(filename, mode, encoding='utf-8', errors='ignore') 34 else: 35 return open(filename, mode) 36 37 def read_file(filename): 38 """讀取文件數據""" 39 contents, labels = [], [] 40 with open_file(filename) as f: 41 for line in f: 42 try: 43 label, content = line.strip().split('\t') 44 if content: 45 contents.append(list(native_content(content))) 46 labels.append(native_content(label)) 47 except: 48 pass 49 return contents, labels 50 51 def build_vocab(train_dir, vocab_dir, vocab_size=5000): 52 """根據訓練集構建詞匯表，存儲""" 53 data_train, _ = read_file(train_dir) 54 55 all_data = [] 56 for content in data_train: 57 all_data.extend(content) 58 counter = Counter(all_data) 59 count_pairs = counter.most_common(vocab_size - 1) 60 words, _ = list(zip(*count_pairs)) 61 # 添加一個 <PAD> 來將所有文本pad為同一長度 62 words = ['<PAD>'] + list(words) 63 open_file(vocab_dir, mode='w').write('\n'.join(words) + '\n') 64 65 def read_vocab(vocab_dir): 66 """讀取詞匯表""" 67 # words = open_file(vocab_dir).read().strip().split('\n') 68 with open_file(vocab_dir) as fp: 69 # 如果是py2 則每個值都轉化為unicode 70 words = [native_content(_.strip()) for _ in fp.readlines()] 71 word_to_id = dict(zip(words, range(len(words)))) 72 return words, word_to_id 73 74 def read_category(): 75 """讀取分類目錄，固定""" 76 categories = ['體育', '財經', '房產', '家居', '教育', '科技', '時尚', '時政', '游戲', '娛樂'] 77 categories = [native_content(x) for x in categories] 78 cat_to_id = dict(zip(categories, range(len(categories)))) 79 return categories, cat_to_id 80 81 def to_words(content, words): 82 """將id表示的內容轉換為文字""" 83 return ''.join(words[x] for x in content) 84 85 def process_file(filename, word_to_id, cat_to_id, max_length=600): 86 """將文件轉換為id表示""" 87 contents, labels = read_file(filename) 88 89 data_id, label_id = [], [] 90 for i in range(len(contents)): 91 data_id.append([word_to_id[x] for x in contents[i] if x in word_to_id]) 92 label_id.append(cat_to_id[labels[i]]) 93 # 使用keras提供的pad_sequences來將文本pad為固定長度 94 x_pad = kr.preprocessing.sequence.pad_sequences(data_id, max_length) 95 y_pad = kr.utils.to_categorical(label_id, num_classes=len(cat_to_id)) # 將標簽轉換為one-hot表示 96 97 return x_pad, y_pad 98 99 def batch_iter(x, y, batch_size=64): 100 """生成批次數據""" 101 data_len = len(x) 102 num_batch = int((data_len - 1) / batch_size) + 1 103 104 indices = np.random.permutation(np.arange(data_len)) 105 x_shuffle = x[indices] 106 y_shuffle = y[indices] 107 108 for i in range(num_batch): 109 start_id = i * batch_size 110 end_id = min((i + 1) * batch_size, data_len) 111 yield x_shuffle[start_id:end_id], y_shuffle[start_id:end_id
3.2 模型搭建cnn_model.py
1 #!/usr/bin/python 2 # -*- coding: utf-8 -*- 3 import tensorflow as tf 4 5 class TRNNConfig(object): 6 """RNN配置參數""" 7 # 模型參數 8 embedding_dim = 64 # 詞向量維度 9 seq_length = 600 # 序列長度 10 num_classes = 10 # 類別數 11 vocab_size = 5000 # 詞匯表達小 12 num_layers= 2 # 隱藏層層數 13 hidden_dim = 128 # 隱藏層神經元 14 rnn = 'gru' # lstm 或 gru 15 dropout_keep_prob = 0.8 # dropout保留比例 16 learning_rate = 1e-3 # 學習率 17 batch_size = 128 # 每批訓練大小 18 num_epochs = 10 # 總迭代輪次 19 print_per_batch = 100 # 每多少輪輸出一次結果 20 save_per_batch = 10 # 每多少輪存入tensorboard 21 22 class TextRNN(object): 23 """文本分類，RNN模型""" 24 def __init__(self, config): 25 self.config = config 26 # 三個待輸入的數據 27 self.input_x = tf.placeholder(tf.int32, [None, self.config.seq_length], name='input_x') 28 self.input_y = tf.placeholder(tf.float32, [None, self.config.num_classes], name='input_y') 29 self.keep_prob = tf.placeholder(tf.float32, name='keep_prob') 30 self.rnn() 31 32 def rnn(self): 33 """rnn模型""" 34 def lstm_cell(): # lstm核 35 return tf.contrib.rnn.BasicLSTMCell(self.config.hidden_dim, state_is_tuple=True) 36 def gru_cell(): # gru核 37 return tf.contrib.rnn.GRUCell(self.config.hidden_dim) 38 def dropout(): # 為每一個rnn核后面加一個dropout層 39 if (self.config.rnn == 'lstm'): 40 cell = lstm_cell() 41 else: 42 cell = gru_cell() 43 return tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=self.keep_prob) 44 45 # 詞向量映射 46 with tf.device('/cpu:0'): 47 embedding = tf.get_variable('embedding', [self.config.vocab_size, self.config.embedding_dim]) 48 embedding_inputs = tf.nn.embedding_lookup(embedding, self.input_x) 49 50 with tf.name_scope("rnn"): 51 # 多層rnn網絡 52 cells = [dropout() for _ in range(self.config.num_layers)] 53 rnn_cell = tf.contrib.rnn.MultiRNNCell(cells, state_is_tuple=True) 54 55 _outputs, _ = tf.nn.dynamic_rnn(cell=rnn_cell, inputs=embedding_inputs, dtype=tf.float32) 56 last = _outputs[:, -1, :] # 取最后一個時序輸出作為結果 57 58 with tf.name_scope("score"): 59 # 全連接層，后面接dropout以及relu激活 60 fc = tf.layers.dense(last, self.config.hidden_dim, name='fc1') 61 fc = tf.contrib.layers.dropout(fc, self.keep_prob) 62 fc = tf.nn.relu(fc) 63 # 分類器 64 self.logits = tf.layers.dense(fc, self.config.num_classes, name='fc2') 65 self.y_pred_cls = tf.argmax(tf.nn.softmax(self.logits), 1) # 預測類別 66 67 with tf.name_scope("optimize"): 68 # 損失函數，交叉熵 69 cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=self.logits, labels=self.input_y) 70 self.loss = tf.reduce_mean(cross_entropy) 71 # 優化器 72 self.optim = tf.train.AdamOptimizer(learning_rate=self.config.learning_rate).minimize(self.loss) 73 74 with tf.name_scope("accuracy"): 75 # 准確率 76 correct_pred = tf.equal(tf.argmax(self.input_y, 1), self.y_pred_cls) 77 self.acc = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
3.3 運行代碼run_cnn.py
1 # coding: utf-8 2 from __future__ import print_function 3 import os 4 import sys 5 import time 6 from datetime import timedelta 7 import numpy as np 8 import tensorflow as tf 9 from sklearn import metrics 10 from rnn_model import TRNNConfig, TextRNN 11 from cnews_loader import read_vocab, read_category, batch_iter, process_file, build_vocab 12 13 base_dir = '../data/cnews' 14 train_dir = os.path.join(base_dir, 'cnews.train.txt') 15 test_dir = os.path.join(base_dir, 'cnews.test.txt') 16 val_dir = os.path.join(base_dir, 'cnews.val.txt') 17 vocab_dir = os.path.join(base_dir, 'cnews.vocab.txt') 18 save_dir = '../checkpoints/textrnn' 19 save_path = os.path.join(save_dir, 'best_validation') # 最佳驗證結果保存路徑 20 21 def get_time_dif(start_time): 22 """獲取已使用時間""" 23 end_time = time.time() 24 time_dif = end_time - start_time 25 return timedelta(seconds=int(round(time_dif))) 26 27 def feed_data(x_batch, y_batch, keep_prob): 28 feed_dict = { 29 model.input_x: x_batch, 30 model.input_y: y_batch, 31 model.keep_prob: keep_prob 32 } 33 return feed_dict 34 35 def evaluate(sess, x_, y_): 36 """評估在某一數據上的准確率和損失""" 37 data_len = len(x_) 38 batch_eval = batch_iter(x_, y_, 128) 39 total_loss = 0.0 40 total_acc = 0.0 41 for x_batch, y_batch in batch_eval: 42 batch_len = len(x_batch) 43 feed_dict = feed_data(x_batch, y_batch, 1.0) 44 y_pred_class,loss, acc = sess.run([model.y_pred_cls,model.loss, model.acc], feed_dict=feed_dict) 45 total_loss += loss * batch_len 46 total_acc += acc * batch_len 47 return y_pred_class,total_loss / data_len, total_acc / data_len 48 49 def train(): 50 print("Configuring TensorBoard and Saver...") 51 # 配置 Tensorboard，重新訓練時，請將tensorboard文件夾刪除，不然圖會覆蓋 52 tensorboard_dir = '../tensorboard/textrnn' 53 if not os.path.exists(tensorboard_dir): 54 os.makedirs(tensorboard_dir) 55 tf.summary.scalar("loss", model.loss) 56 tf.summary.scalar("accuracy", model.acc) 57 merged_summary = tf.summary.merge_all() 58 writer = tf.summary.FileWriter(tensorboard_dir) 59 # 配置 Saver 60 saver = tf.train.Saver() 61 if not os.path.exists(save_dir): 62 os.makedirs(save_dir) 63 64 print("Loading training and validation data...") 65 # 載入訓練集與驗證集 66 start_time = time.time() 67 x_train, y_train = process_file(train_dir, word_to_id, cat_to_id, config.seq_length) 68 x_val, y_val = process_file(val_dir, word_to_id, cat_to_id, config.seq_length) 69 time_dif = get_time_dif(start_time) 70 print("Time usage:", time_dif) 71 72 # 創建session 73 session = tf.Session() 74 session.run(tf.global_variables_initializer()) 75 writer.add_graph(session.graph) 76 print('Training and evaluating...') 77 start_time = time.time() 78 total_batch = 0 # 總批次 79 best_acc_val = 0.0 # 最佳驗證集准確率 80 last_improved = 0 # 記錄上一次提升批次 81 require_improvement = 1000 # 如果超過1000輪未提升，提前結束訓練 82 83 flag = False 84 for epoch in range(config.num_epochs): 85 print('Epoch:', epoch + 1) 86 batch_train = batch_iter(x_train, y_train, config.batch_size) 87 for x_batch, y_batch in batch_train: 88 feed_dict = feed_data(x_batch, y_batch, config.dropout_keep_prob) 89 90 if total_batch % config.save_per_batch == 0: 91 # 每多少輪次將訓練結果寫入tensorboard scalar 92 s = session.run(merged_summary, feed_dict=feed_dict) 93 writer.add_summary(s, total_batch) 94 95 if total_batch % config.print_per_batch == 0: 96 # 每多少輪次輸出在訓練集和驗證集上的性能 97 feed_dict[model.keep_prob] = 1.0 98 loss_train, acc_train = session.run([model.loss, model.acc], feed_dict=feed_dict) 99 y_pred_class,loss_val, acc_val = evaluate(session, x_val, y_val) # todo 100 101 if acc_val > best_acc_val: 102 # 保存最好結果 103 best_acc_val = acc_val 104 last_improved = total_batch 105 saver.save(sess=session, save_path=save_path) 106 improved_str = '*' 107 else: 108 improved_str = '' 109 110 time_dif = get_time_dif(start_time) 111 msg = 'Iter: {0:>6}, Train Loss: {1:>6.2}, Train Acc: {2:>7.2%},' \ 112 + ' Val Loss: {3:>6.2}, Val Acc: {4:>7.2%}, Time: {5} {6}' 113 print(msg.format(total_batch, loss_train, acc_train, loss_val, acc_val, time_dif, improved_str)) 114 115 session.run(model.optim, feed_dict=feed_dict) # 運行優化 116 total_batch += 1 117 118 if total_batch - last_improved > require_improvement: 119 # 驗證集正確率長期不提升，提前結束訓練 120 print("No optimization for a long time, auto-stopping...") 121 flag = True 122 break # 跳出循環 123 if flag: # 同上 124 break 125 126 def test(): 127 print("Loading test data...") 128 start_time = time.time() 129 x_test, y_test = process_file(test_dir, word_to_id, cat_to_id, config.seq_length) 130 session = tf.Session() 131 session.run(tf.global_variables_initializer()) 132 saver = tf.train.Saver() 133 saver.restore(sess=session, save_path=save_path) # 讀取保存的模型 134 135 print('Testing...') 136 y_pred,loss_test, acc_test = evaluate(session, x_test, y_test) 137 msg = 'Test Loss: {0:>6.2}, Test Acc: {1:>7.2%}' 138 print(msg.format(loss_test, acc_test)) 139 140 batch_size = 128 141 data_len = len(x_test) 142 num_batch = int((data_len - 1) / batch_size) + 1 143 144 y_test_cls = np.argmax(y_test, 1) 145 y_pred_cls = np.zeros(shape=len(x_test), dtype=np.int32) # 保存預測結果 146 for i in range(num_batch): # 逐批次處理 147 start_id = i * batch_size 148 end_id = min((i + 1) * batch_size, data_len) 149 feed_dict = { 150 model.input_x: x_test[start_id:end_id], 151 model.keep_prob: 1.0 152 } 153 y_pred_cls[start_id:end_id] = session.run(model.y_pred_cls, feed_dict=feed_dict) 154 155 # 評估 156 print("Precision, Recall and F1-Score...") 157 print(metrics.classification_report(y_test_cls, y_pred_cls, target_names=categories)) 158 # 混淆矩陣 159 print("Confusion Matrix...") 160 cm = metrics.confusion_matrix(y_test_cls, y_pred_cls) 161 print(cm) 162 time_dif = get_time_dif(start_time) 163 print("Time usage:", time_dif) 164 165 if __name__ == '__main__': 166 print('Configuring RNN model...') 167 config = TRNNConfig() 168 if not os.path.exists(vocab_dir): # 如果不存在詞匯表，重建 169 build_vocab(train_dir, vocab_dir, config.vocab_size) 170 categories, cat_to_id = read_category() 171 words, word_to_id = read_vocab(vocab_dir) 172 config.vocab_size = len(words) 173 model = TextRNN(config) 174 option='train' 175 if option == 'train': 176 train() 177 else: 178 test()
3.4 預測predict.py
1 # coding: utf-8 2 from __future__ import print_function 3 import os 4 import tensorflow as tf 5 import tensorflow.contrib.keras as kr 6 from rnn_model import TRNNConfig, TextRNN 7 from cnews_loader import read_category, read_vocab 8 try: 9 bool(type(unicode)) 10 except NameError: 11 unicode = str 12 13 base_dir = '../data/cnews' 14 vocab_dir = os.path.join(base_dir, 'cnews.vocab.txt') 15 save_dir = '../checkpoints/textrnn' 16 save_path = os.path.join(save_dir, 'best_validation') # 最佳驗證結果保存路徑 17 18 class RnnModel: 19 def __init__(self): 20 self.config = TRNNConfig() 21 self.categories, self.cat_to_id = read_category() 22 self.words, self.word_to_id = read_vocab(vocab_dir) 23 self.config.vocab_size = len(self.words) 24 self.model = TextRNN(self.config) 25 self.session = tf.Session() 26 self.session.run(tf.global_variables_initializer()) 27 saver = tf.train.Saver() 28 saver.restore(sess=self.session, save_path=save_path) # 讀取保存的模型 29 30 def predict(self, message): 31 # 支持不論在python2還是python3下訓練的模型都可以在2或者3的環境下運行 32 content = unicode(message) 33 data = [self.word_to_id[x] for x in content if x in self.word_to_id] 34 35 feed_dict = { 36 self.model.input_x: kr.preprocessing.sequence.pad_sequences([data], self.config.seq_length), 37 self.model.keep_prob: 1.0 38 } 39 y_pred_cls = self.session.run(self.model.y_pred_cls, feed_dict=feed_dict) 40 return self.categories[y_pred_cls[0]] 41 42 if __name__ == '__main__': 43 rnn_model = RnnModel() 44 test_demo = ['三星ST550以全新的拍攝方式超越了以往任何一款數碼相機', 45 '熱火vs騎士前瞻：皇帝回鄉二番戰 東部次席唾手可得新浪體育訊北京時間3月30日7:00'] 46 for i in test_demo: 47 print(rnn_model.predict(i))
四、結果展示
訓練時長，接近2小時
相關代碼可見：https://github.com/yifanhunter/textClassifier_chinese