Keras文本分类

内置样本

影评正负2分类（已编码）

from keras.datasets import imdb  # Internet Movie Database
# num_words设定较小时，会发现高频词多是停词
(x, y), _ = imdb.load_data(num_words=1)
print(x.shape, y.shape)  # (25000,) (25000,)
# 词与ID间的映射
word2id = imdb.get_word_index()
id2word = {i: w for w, i in word2id.items()}
print(x[0])
print(' '.join([id2word[i] for i in x[0]]))

序列截断或补齐为等长

from keras.preprocessing.sequence import pad_sequences
maxlen = 2
print(pad_sequences([[1, 2, 3], [1]], maxlen))
"""[[2 3] [0 1]]"""
print(pad_sequences([[1, 2, 3], [1]], maxlen, value=9))
"""[[2 3] [9 1]]"""
print(pad_sequences([[1, 2, 3], [1]], maxlen, padding='post'))
"""[[2 3] [1 0]]"""
print(pad_sequences([[1, 2, 3], [1]], maxlen, truncating='post'))
"""[[1 2] [0 1]]"""

词嵌入

from keras.datasets.imdb import load_data  # 影评情感2分类
from keras.preprocessing.sequence import pad_sequences
from keras.models import Sequential
from keras.layers import Embedding, Flatten, Dense

"""配置"""
num_words = 10000  # 按词频大小取样本前10000个词
input_dim = num_words  # 词库大小（必须>=num_words）
maxlen = 25  # 序列长度
output_dim = 40  # 词向量维度
batch_size = 128
epochs = 2

"""数据读取与处理"""
(x, y), _ = load_data(num_words=num_words)
x = pad_sequences(x, maxlen)

"""建模"""
model = Sequential()
# 词嵌入：词库大小、词向量维度、固定序列长度
model.add(Embedding(input_dim, output_dim, input_length=maxlen))
# 平坦化：maxlen * output_dim
model.add(Flatten())
# 输出层：2分类
model.add(Dense(units=1, activation='sigmoid'))
# RMSprop优化器、二元交叉熵损失
model.compile('rmsprop', 'binary_crossentropy', ['acc'])
# 训练
model.fit(x, y, batch_size, epochs)

"""模型可视化"""
model.summary()

词嵌入+RNN

from keras.datasets.imdb import load_data  # 影评情感2分类
from keras.preprocessing.sequence import pad_sequences
from keras.models import Sequential
from keras.layers import Embedding, Dense, SimpleRNN
import matplotlib.pyplot as mp

"""配置"""
num_words = 10000  # 按词频大小取样本前10000个词
input_dim = num_words  # 词库大小（必须>=num_words）
maxlen = 25  # 序列长度
output_dim = 40  # 词向量维度
units = 32  # RNN神经元数量
batch_size = 128
epochs = 3

"""数据读取与处理"""
(x, y), _ = load_data(num_words=num_words)
x = pad_sequences(x, maxlen)

"""建模"""
model = Sequential()
model.add(Embedding(input_dim, output_dim, input_length=maxlen))
model.add(SimpleRNN(units, return_sequences=True))  # 返回序列全部结果
model.add(SimpleRNN(units, return_sequences=False))  # 返回序列最尾结果
model.add(Dense(units=1, activation='sigmoid'))
model.summary()

"""编译、训练"""
model.compile('rmsprop', 'binary_crossentropy', ['acc'])
history = model.fit(x, y, batch_size, epochs, verbose=2,
                    validation_split=.1)  # 取10%样本作验证

"""精度曲线"""
acc = history.history['acc']
val_acc = history.history['val_acc']
mp.plot(range(epochs), acc)
mp.plot(range(epochs), val_acc)
mp.show()

词嵌入+CNN

from keras.datasets.imdb import load_data  # 影评情感2分类
from keras.preprocessing.sequence import pad_sequences
from keras.models import Sequential
from keras.layers import Embedding, Dense,\
    Conv1D, MaxPool1D, GlobalMaxPool1D

"""配置"""
num_words = 10000  # 按词频大小取样本前10000个词
input_dim = num_words  # 词库大小（必须>=num_words）
maxlen = 25  # 序列长度
output_dim = 40  # 词向量维度
filters = 32  # 卷积层滤波器数量
kernel_size = 7  # 卷积层滤波器大小
batch_size = 128
epochs = 3

"""数据读取与处理"""
(x, y), _ = load_data(num_words=num_words)
x = pad_sequences(x, maxlen)

"""建模"""
model = Sequential()
model.add(Embedding(input_dim, output_dim, input_length=maxlen))
model.add(Conv1D(filters, kernel_size, activation='relu'))
model.add(MaxPool1D(pool_size=2))  # strides默认等于pool_size
model.add(Conv1D(filters, kernel_size, activation='relu'))
model.add(GlobalMaxPool1D())  # 对于时序数据的全局最大池化
model.add(Dense(units=1, activation='sigmoid'))
model.summary()

"""编译、训练"""
model.compile('rmsprop', 'binary_crossentropy', ['acc'])
model.fit(x, y, batch_size, epochs, 2)

CNN+RNN

from keras.datasets.imdb import load_data  # 影评情感2分类
from keras.preprocessing.sequence import pad_sequences
from keras.models import Sequential
from keras.layers import Embedding, Dense,\
    Conv1D, MaxPool1D, GRU  # Gated Recurrent Unit

"""配置"""
num_words = 10000  # 按词频大小取样本前10000个词
input_dim = num_words  # 词库大小（必须>=num_words）
maxlen = 25  # 序列长度
output_dim = 40  # 词向量维度
filters = 32  # 卷积层滤波器数量
kernel_size = 7  # 卷积层滤波器大小
units = 32  # RNN神经元数量
batch_size = 128
epochs = 3

"""数据读取与处理"""
(x, y), _ = load_data(num_words=num_words)
x = pad_sequences(x, maxlen)

"""建模"""
model = Sequential()
model.add(Embedding(input_dim, output_dim, input_length=maxlen))
model.add(Conv1D(filters, kernel_size, activation='relu'))
model.add(MaxPool1D())
model.add(Conv1D(filters, kernel_size, activation='relu'))
model.add(GRU(units))  # 门限循环单元网络
model.add(Dense(1, activation='sigmoid'))
model.summary()

"""编译、训练"""
model.compile('rmsprop', 'binary_crossentropy', ['acc'])
model.fit(x, y, batch_size, epochs, 2)

双向RNN

Bidirectional

from keras.datasets.imdb import load_data  # 影评情感2分类
from keras.preprocessing.sequence import pad_sequences
from keras.models import Sequential
from keras.layers import Embedding, Dense, GRU, Bidirectional
import matplotlib.pyplot as mp

"""配置"""
num_words = 10000  # 按词频大小取样本前10000个词
input_dim = num_words  # 词库大小（必须>=num_words）
maxlen = 25  # 序列长度
output_dim = 40  # 词向量维度
units = 32  # RNN神经元数量
batch_size = 128
epochs = 3
verbose = 2

"""数据读取与处理"""
(x, y), _ = load_data(num_words=num_words)
x = pad_sequences(x, maxlen)

"""建模"""
model = Sequential()
model.add(Embedding(input_dim, output_dim, input_length=maxlen))
model.add(Bidirectional(GRU(units)))  # 双向RNN
model.add(Dense(units=1, activation='sigmoid'))
model.summary()

"""编译、训练"""
model.compile('rmsprop', 'binary_crossentropy', ['acc'])
history = model.fit(x, y, batch_size, epochs, verbose, validation_split=.1)

"""精度曲线"""
acc = history.history['acc']
val_acc = history.history['val_acc']
mp.plot(range(epochs), acc)
mp.plot(range(epochs), val_acc)
mp.show()

双向RNN+Word2Vector

from jieba import lcut
from sklearn.model_selection import train_test_split
from gensim.models import Word2Vec
from tensorflow.python.keras.preprocessing.sequence import pad_sequences
from tensorflow.python.keras.models import Sequential
from tensorflow.python.keras.layers import Dense, GRU, Bidirectional
from tensorflow.python.keras.callbacks import EarlyStopping
from numpy import zeros

"""配置"""
size = 20  # 词向量维度
units = 30  # RNN神经元数量
maxlen = 40  # 序列长度
batch_size = 1  # 每批数据量大小
epochs = 9  # 训练最大轮数
verbose = 2  # 训练过程展示
patience = 1  # 没有进步的训练轮数
callbacks = [EarlyStopping('val_acc', patience=patience)]

"""训练数据、分词"""
data = [
    [0, '小米粥是以小米作为主要食材熬制而成的粥，口味清淡，清香味，具有简单易制，健胃消食的特点'],
    [0, '煮粥时一定要先烧开水然后放入洗净后的小米'], [0, '蛋白质及氨基酸、脂肪、维生素、矿物质'],
    [0, '小米是传统健康食品，可单独焖饭和熬粥'], [0, '苹果，是水果中的一种'],
    [0, '粥的营养价值很高，富含矿物质和维生素，含钙量丰富，有助于代谢掉体内多余盐分'],
    [0, '鸡蛋有很高的营养价值，是优质蛋白质、B族维生素的良好来源，还能提供脂肪、维生素和矿物质'],
    [0, '这家超市的苹果都非常新鲜'], [0, '在北方小米是主要食物之一，很多地区有晚餐吃小米粥的习俗'],
    [0, '小米营养价值高，营养全面均衡 ，主要含有碳水化合物'], [0, '蛋白质及氨基酸、脂肪、维生素、盐分'],
    [1, '小米、三星、华为，作为安卓三大手机旗舰'], [1, '别再管小米华为了！魅族手机再曝光：这次真的完美了'],
    [1, '苹果手机或将重陷2016年困境，但这次它无法再大幅提价了'], [1, '三星想要继续压制华为，仅凭A70还不够'],
    [1, '三星手机屏占比将再创新高，超华为及苹果旗舰'], [1, '华为P30、三星A70爆卖，斩获苏宁最佳手机营销奖'],
    [1, '雷军，用一张图告诉你：小米和三星的差距在哪里'], [1, '小米米聊APP官方Linux版上线，适配深度系统'],
    [1, '三星刚刚更新了自家的可穿戴设备APP'], [1, '华为、小米跨界并不可怕，可怕的打不破内心的“天花板”'],
]
X, Y = [lcut(i[1]) for i in data], [i[0] for i in data]
X_train, X_test, y_train, y_test = train_test_split(X, Y)

"""词向量"""
word2vec = Word2Vec(X_train, size, min_count=1)
w2i = {w: i for i, w in enumerate(word2vec.wv.index2word)}
vectors = word2vec.wv.vectors
def w2v(w):
    i = w2i.get(w)
    return vectors[i] if i else zeros(size)
def pad(ls_of_words):
    a = [[w2v(i) for i in x] for x in ls_of_words]
    a = pad_sequences(a, maxlen, dtype='float')
    return a
X_train, X_test = pad(X_train), pad(X_test)

"""建模"""
model = Sequential()
model.add(Bidirectional(GRU(units), input_shape=(maxlen, size)))  # 双向RNN
model.add(Dense(units=1, activation='sigmoid'))
model.summary()

"""训练"""
model.compile('rmsprop', 'binary_crossentropy', ['acc'])
model.fit(X_train, y_train, batch_size, epochs, verbose, callbacks, validation_data=(X_test, y_test))