影评文本分类
数据集:IMDB 数据集
模型建立:tf.keras
1. 下载数据集
import numpy as np
import tensorflow as tf
from tensorflow import keras
#也可以直接 import keras
#下载数据
imdb = keras.datasets.imdb
(train_data, train_labels), (test_data, test_labels) = imdb.load_data(num_words=10000)
- 参数
num_words=10000
会保留训练数据中出现频次在前 10000 位的字词。为确保数据规模处于可管理的水平,罕见字词将被舍弃
运行结果:
2. 探索数据
- 数据集已经过预处理:每个样本都是一个整数数组,表示影评中的字词。每个标签都是整数值 0 或 1,其中 0 表示负面影评,1 表示正面影评
- 不同的影评通常具有不同的长度,所以我们需要对影评进行填充,使得它们具有同样的长度
#数据探索
print('Training entries: {}, labels: {}'.format(len(train_data), len(train_labels)))
#运行结果:Training entries: 25000, labels: 25000
print(train_data[0])
运行结果:
3. 将整数转换回字词
- 将整数转换回字词可以让我们观察到每条影评的具体含义,但是用于影评中可能包含一些错误或者无效字符,所以需要对其进行标示。下面是一种实现这种标示功能:
word_index = imdb.get_word_index()
word_index = {k:(v+3) for k, v in word_index.items()}
word_index['<PAD>'] = 0
word_index['<START>'] = 1
word_index['<UNK>'] = 2
word_index['<UNUSED>'] = 3
reverse_word_index = dict([(value, key) for (key, value) in word_index.items()])
运行结果:
def decode_review(text):
return ' '.join([reverse_word_index.get(i, '?') for i in text])
decode_review(train_data[0])
运行结果:
查看数据集维度:
len(train_data[0]), len(train_data[1])
#运行结果:(218, 189)
4. 准备数据
影评(整数数组)必须转换为张量,然后才能馈送到神经网络中。这里有两种方法实现这种转换:
- 对数组进行独热编码,将它们转换为由 0 和 1 构成的向量。例如,序列 [3, 5] 将变成一个 10000 维的向量,除索引 3 和 5 转换为 1 之外,其余全转换为 0。然后,将它作为网络的第一层,一个可以处理浮点向量数据的密集层。不过,这种方法会占用大量内存,需要一个大小为 num_words * num_reviews 的矩阵。
- 或者,我们可以填充数组,使它们都具有相同的长度,然后创建一个形状为 max_length * num_reviews 的整数张量。我们可以使用一个能够处理这种形状的嵌入层作为网络中的第一层
这里使用第二种方法进行转换。由于影评的长度必须相同,可以使用 keras 的 pad_sequences
函数将长度标准化:
train_data = keras.preprocessing.sequence.pad_sequences(train_data,
value=word_index['<PAD>'],
padding='post',
maxlen=256)
test_data = keras.preprocessing.sequence.pad_sequences(test_data,
value=word_index['<PAD>'],
padding='post',
maxlen=256)
再一次查看数据集维度:
len(train_data[0]), len(train_data[1])
#运行结果:(256, 256)
查看第一条影评,发现后面已经填充上了 0:
train_data[0]
5. 构建模型
神经网络通过堆叠层创建而成,这需要做出两个架构方面的主要决策:
- 要在模型中使用多少个层?
- 要针对每个层使用多少个隐藏单元?
这里按照顺序堆叠各个层以构建分类器:
- 第一层是 Embedding 层。该层会在整数编码的词汇表中查找每个字词-索引的嵌入向量。模型在接受训练时会学习这些向量。这些向量会向输出数组添加一个维度。生成的维度为:(batch, sequence, embedding)
- 一个 GlobalAveragePooling1D 层通过对序列维度求平均值,针对每个样本返回一个长度固定的输出向量。这样,模型便能够以尽可能简单的方式处理各种长度的输入
- 该长度固定的输出向量会传入一个全连接 (Dense) 层(包含 16 个隐藏单元)
- 最后一层与单个输出节点密集连接。应用 sigmoid 激活函数后,结果是介于 0 到 1 之间的浮点值,表示概率或置信水平
vocab_size = 10000
model = keras.Sequential()
model.add(keras.layers.Embedding(vocab_size, 16))
model.add(keras.layers.GlobalAveragePooling1D())
model.add(keras.layers.Dense(16, activation=tf.nn.relu))
model.add(keras.layers.Dense(1, activation=tf.nn.sigmoid))
model.summary()
模型结构如下:
上述模型在输入和输出之间有两个中间层(也称为“隐藏”层)。输出(单元、节点或神经元)的数量是相应层的表示法空间的维度。换句话说,该数值表示学习内部表示法时网络所允许的自由度。
如果模型具有更多隐藏单元(更高维度的表示空间)和/或更多层,则说明网络可以学习更复杂的表示法。不过,这会使网络耗费更多计算资源,并且可能导致学习不必要的模式(可以优化在训练数据上的表现,但不会优化在测试数据上的表现),这说明模型已经过拟合了。
6. 损失函数和优化器
模型在训练时需要一个损失函数和一个优化器。由于这是一个二元分类问题且模型会输出一个概率(应用 S 型激活函数的单个单元层),因此我们将使用 binary_crossentropy
损失函数
model.compile(optimizer=tf.train.AdamOptimizer(),
loss='binary_crossentropy',
metrics=['accuracy'])
7. 创建验证集
这里我们使用10000个样本作为验证集
x_val = train_data[:10000]
partial_x_train = train_data[10000:]
y_val = train_labels[:10000]
partial_y_train = train_labels[10000:]
8. 训练模型
用有 512 个样本的小批次训练模型 40 个周期。这将对 x_train 和 y_train 张量中的所有样本进行 40 次迭代。在训练期间,监控模型在验证集的 10000 个样本上的损失和准确率:
history = model.fit(partial_x_train,
partial_y_train,
epochs=40,
batch_size=512,
validation_data=(x_val, y_val),
verbose=1)
部分训练截图:
9. 评估模型
使用模型的 evaluate
方法来得到模型的损失和准确率:
results = model.evaluate(test_data, test_labels)
results
参考文献:影评文本分类
THUCNews 数据集
使用THUCNews的一个子集进行训练与测试,数据集请自行到THUCTC:一个高效的中文文本分类工具包下载,请遵循数据提供方的开源协议
本次训练使用了其中的10个分类,每个分类6500条数据,此子集的下载链接: https://pan.baidu.com/s/1hugrfRu 密码: qfud
数据集划分如下:
- 训练集: 5000*10
- 验证集: 500*10
- 测试集: 1000*10
1. 数据读取
"""
将文本整合到 train、test、val 三个文件中
"""
import os
def _read_file(filename):
"""读取一个文件并转换为一行"""
with open(filename, 'r', encoding='utf-8') as f:
return f.read().replace('\n', '').replace('\t', '').replace('\u3000', '')
def save_file(dirname):
"""
将多个文件整合并存到3个文件中
dirname: 原数据目录
文件内容格式: 类别\t内容
"""
f_train = open('data/cnews/cnews.train.txt', 'w', encoding='utf-8')
f_test = open('data/cnews/cnews.test.txt', 'w', encoding='utf-8')
f_val = open('data/cnews/cnews.val.txt', 'w', encoding='utf-8')
for category in os.listdir(dirname): # 分类目录
cat_dir = os.path.join(dirname, category)
if not os.path.isdir(cat_dir):
continue
files = os.listdir(cat_dir)
count = 0
for cur_file in files:
filename = os.path.join(cat_dir, cur_file)
content = _read_file(filename)
if count < 5000:
f_train.write(category + '\t' + content + '\n')
elif count < 6000:
f_test.write(category + '\t' + content + '\n')
else:
f_val.write(category + '\t' + content + '\n')
count += 1
print('Finished:', category)
f_train.close()
f_test.close()
f_val.close()
if __name__ == '__main__':
save_file('data/thucnews')
print(len(open('data/cnews/cnews.train.txt', 'r', encoding='utf-8').readlines()))
print(len(open('data/cnews/cnews.test.txt', 'r', encoding='utf-8').readlines()))
print(len(open('data/cnews/cnews.val.txt', 'r', encoding='utf-8').readlines()))
2. 数据预处理
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
def native_word(word, encoding='utf-8'):
"""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"""
if not is_py3:
return word.encode(encoding)
else:
return word
def native_content(content):
if not is_py3:
return content.decode('utf-8')
else:
return content
def open_file(filename, mode='r'):
"""
常用文件操作,可在python2和python3间切换.
mode: 'r' or 'w' for read or write
"""
if is_py3:
return open(filename, mode, encoding='utf-8', errors='ignore')
else:
return open(filename, mode)
def read_file(filename):
"""读取文件数据"""
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
def build_vocab(train_dir, vocab_dir, vocab_size=5000):
"""根据训练集构建词汇表,存储"""
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))
# 添加一个 <PAD> 来将所有文本pad为同一长度
words = ['<PAD>'] + list(words)
open_file(vocab_dir, mode='w').write('\n'.join(words) + '\n')
def read_vocab(vocab_dir):
"""读取词汇表"""
# words = open_file(vocab_dir).read().strip().split('\n')
with open_file(vocab_dir) as fp:
# 如果是py2 则每个值都转化为unicode
words = [native_content(_.strip()) for _ in fp.readlines()]
word_to_id = dict(zip(words, range(len(words))))
return words, word_to_id
def read_category():
"""读取分类目录,固定"""
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
def to_words(content, words):
"""将id表示的内容转换为文字"""
return ''.join(words[x] for x in content)
def process_file(filename, word_to_id, cat_to_id, max_length=600):
"""将文件转换为id表示"""
contents, labels = read_file(filename)
data_id, label_id = [], []
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]])
# 使用keras提供的pad_sequences来将文本pad为固定长度
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)) # 将标签转换为one-hot表示
return x_pad, y_pad
def batch_iter(x, y, batch_size=64):
"""生成批次数据"""
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]
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]
3. CNN模型
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 cnn_model import TCNNConfig, TextCNN
from data.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/textcnn'
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)
loss, acc = sess.run([model.loss, model.acc], feed_dict=feed_dict)
total_loss += loss * batch_len
total_acc += acc * batch_len
return total_loss / data_len, total_acc / data_len
def train():
print("Configuring TensorBoard and Saver...")
# 配置 Tensorboard,重新训练时,请将tensorboard文件夹删除,不然图会覆盖
tensorboard_dir = 'tensorboard/textcnn'
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 # 总批次
best_acc_val = 0.0 # 最佳验证集准确率
last_improved = 0 # 记录上一次提升批次
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:
# 每多少轮次将训练结果写入tensorboard scalar
s = session.run(merged_summary, feed_dict=feed_dict)
writer.add_summary(s, total_batch)
if total_batch % config.print_per_batch == 0:
# 每多少轮次输出在训练集和验证集上的性能
feed_dict[model.keep_prob] = 1.0
loss_train, acc_train = session.run([model.loss, model.acc], feed_dict=feed_dict)
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 # 跳出循环
if flag: # 同上
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...')
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) # 保存预测结果
for i in range(num_batch): # 逐批次处理
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__':
if len(sys.argv) != 2 or sys.argv[1] not in ['train', 'test']:
raise ValueError("""usage: python run_cnn.py [train / test]""")
print('Configuring CNN model...')
config = TCNNConfig()
if not os.path.exists(vocab_dir): # 如果不存在词汇表,重建
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 = TextCNN(config)
if sys.argv[1] == 'train':
train()
else:
test()
4. RNN模型
import tensorflow as tf
class TRNNConfig(object):
"""RNN配置参数"""
# 模型参数
embedding_dim = 64 # 词向量维度
seq_length = 600 # 序列长度
num_classes = 10 # 类别数
vocab_size = 5000 # 词汇表达小
num_layers= 2 # 隐藏层层数
hidden_dim = 128 # 隐藏层神经元
rnn = 'gru' # lstm 或 gru
dropout_keep_prob = 0.8 # dropout保留比例
learning_rate = 1e-3 # 学习率
batch_size = 128 # 每批训练大小
num_epochs = 10 # 总迭代轮次
print_per_batch = 100 # 每多少轮输出一次结果
save_per_batch = 10 # 每多少轮存入tensorboard
class TextRNN(object):
"""文本分类,RNN模型"""
def __init__(self, config):
self.config = config
# 三个待输入的数据
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(): # lstm核
return tf.contrib.rnn.BasicLSTMCell(self.config.hidden_dim, state_is_tuple=True)
def gru_cell(): # gru核
return tf.contrib.rnn.GRUCell(self.config.hidden_dim)
def dropout(): # 为每一个rnn核后面加一个dropout层
if (self.config.rnn == 'lstm'):
cell = lstm_cell()
else:
cell = gru_cell()
return tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=self.keep_prob)
# 词向量映射
with tf.device('/cpu: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"):
# 多层rnn网络
cells = [dropout() for _ in range(self.config.num_layers)]
rnn_cell = tf.contrib.rnn.MultiRNNCell(cells, state_is_tuple=True)
_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
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)
# 分类器
self.logits = tf.layers.dense(fc, self.config.num_classes, name='fc2')
self.y_pred_cls = tf.argmax(tf.nn.softmax(self.logits), 1) # 预测类别
with tf.name_scope("optimize"):
# 损失函数,交叉熵
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=self.logits, labels=self.input_y)
self.loss = tf.reduce_mean(cross_entropy)
# 优化器
self.optim = tf.train.AdamOptimizer(learning_rate=self.config.learning_rate).minimize(self.loss)
with tf.name_scope("accuracy"):
# 准确率
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))
- 个人心得:其中CNN、RNN的架构值得我们好好学习,还有就是TensorFlow的语法需要多练习,才能熟练掌握。