深度学习：文本预测

代码：

#coding=utf-8

import os
import sys
import re
import random
import numpy as np
from keras.models import Sequential,load_model
from keras.layers import Dense, LSTM, Bidirectional
from keras.optimizers import Adam

def sample(preds, gamma=1.0):
    preds **= 1 / gamma
    preds /= np.sum(preds)
    # 从多项分布中抽取样本
    probas = np.random.multinomial(1, preds, 1)
    return np.argmax(probas)


def getyucetext(generated_text, sent_len, N, char_indices, char_set, model):
    #[0.2, 0.5, 1.0, 1.2, 1.5]
    gamma = 1.0
    next_chars = []
    for j in range(100):
        sampled = np.zeros((1, sent_len, N))  # 第一个维度为1，取单个样本
        for t, char in enumerate(generated_text):
            sampled[0, t, char_indices[char]] = 1
        preds = model.predict(sampled, verbose=0)[0]  # 返回3843个概率值的向量
        preds = preds.astype(float)
        next_index = sample(preds, gamma)
        next_char = char_set[next_index]
        generated_text = (generated_text + next_char)[1:]
        next_chars.append(next_char)
    
    return ''.join(next_chars)


fp = open(r'data/sanguoyanyi.txt', 'r', encoding='utf-8')
text = fp.read()
fp.close()

print('语料总长度：', len(text))
sent_len = 4  # 样本长度，用前4个字来预测第5个字
step = 1  # 在总文本上面平移取样本，步长为1
sentences = []  # 样本集X，用来预测
next_chars = []  # 样本集Y，真实值
for i in range(0, len(text) - sent_len, step):
    sentences.append(text[i:i + sent_len])
    next_chars.append(text[i + sent_len])

sentenceslen = len(sentences)
print('训练样本数目：', sentenceslen)

# 创建字典，文本的字符集,sorted()对文本进行排序
char_set = sorted(list(set(text)))
N = len(char_set)
print('字典中字的个数：', N)

# 根据值来获取索引，存入字典
char_indices = dict((char, index) for index, char in enumerate(char_set))
#print(char_indices)

x = np.zeros((sentenceslen, sent_len, N), dtype=bool)
y = np.zeros((sentenceslen, N), dtype=bool)

# 对每一个字做one-hot编码，构成词向量
for i, sentence in enumerate(sentences):
    for t, char in enumerate(sentence):
        x[i, t, char_indices[char]] = 1  # 将x索引值的文本置1，形成one——hot编码
    y[i, char_indices[next_chars[i]]] = 1  # 将Y索引值的文本置1，形成one——hot编码
    
xsplit = int(len(x) * 0.95)
ysplit = int(len(y) * 0.95)
x_data = x[:xsplit]
y_data = y[:ysplit]
x_test_data = x[xsplit:]
y_test_data = y[ysplit:]

model_path = 'models/wenben_yuce_lstm.h5'
if os.path.exists(model_path):
    model = load_model(model_path)
else:
    # 构建模型
    model = Sequential(name='text_generation')
    # 将3843维降维到300，输入维度，样本：3个字，每个字3843维
    model.add(LSTM(300, return_sequences=True, input_shape=(sent_len, N)))
    model.add(Bidirectional(LSTM(64, return_sequences=True)))  # LSTM双向传递
    model.add(Bidirectional(LSTM(32, return_sequences=False)))
    model.add(Dense(N, activation="softmax"))  # 全连接层，有N个分类
    model.compile(loss='categorical_crossentropy', optimizer=Adam(learning_rate=0.01))
    model.fit(x_data, y_data, batch_size=128, epochs=10, validation_data=(x_test_data, y_test_data))
    model.save(model_path)
    
model.summary()

generated_texts = ['诸葛村夫', '安敢望此', '一将飞出', '背义之徒', '你大爷的']
for generated_text in generated_texts:
    print('输入值:%s' % generated_text)
    yucetext = getyucetext(generated_text, sent_len, N, char_indices, char_set, model)
    print('预测值:%s' % yucetext)
    print()

data/sanguoyanyi.txt  是三国演义文字版

执行输出：

语料总长度： 603705
训练样本数目： 603701
字典中字的个数： 3843

Model: "text_generation"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 lstm (LSTM)                 (None, 4, 300)            4972800   
                                                                 
 bidirectional (Bidirectiona  (None, 4, 128)           186880    
 l)                                                              
                                                                 
 bidirectional_1 (Bidirectio  (None, 64)               41216     
 nal)                                                            
                                                                 
 dense (Dense)               (None, 3843)              249795    
                                                                 
=================================================================
Total params: 5,450,691
Trainable params: 5,450,691
Non-trainable params: 0
_________________________________________________________________
输入值:诸葛村夫
预测值:里，用将鹤威瘤矣，故其成州，何显伐何？”昭曰：“令然引相扶看，闻被袁瓒按两州。孙宫暗见迷间。曹德好到也，足期于中，曜休前意！”默笑，图教而去曰：“主面有此幕，不欲而走。布同、赵咨回柩，则恨骂姬女两处，

输入值:安敢望此
预测值:，必敢容怒，果得轮运爪若，非第民成义来章礼矣。前弟词保天舅兮之者宗，安可图此之？”孔明曰：“败不若当官居城！但相起士。此急怕师，正得所商明耳。”绍见入之。”玄德曰：“国化众近刘褚。襄议霸而去。尽正附到

输入值:一将飞出
预测值:而走。突空败，鼓三里围刺，翻见李曰：“涓相夜成将口。手大惊灵于，姓刘大江谷关，胜足欲用邳之，用后悲失，彤不携吞下，黄言有诸将于，傲连英苦，荆更追卒，引百万擒疑，所投东面所神，忽饮国地主，与陈音回书统卜

输入值:背义之徒
预测值:。此处后国炎众也。字昭鼓，旌面攻回，岂要举功，止后惊盔而无言。今日连一大将势佯瓒，鼻受三尺，随李砍往傍海，背崇陆主，足兵为此，都败击钺。只说张权去马，使进世军旬唇者也。”其人叹曰：“伯兄轻辅为疾之兵，

输入值:你大爷的
预测值:杀，景赛敢到术，欲勿辞也。”二笑曰：“此从之？”赵芳皆退，乃须唤敌子出中。魏延等可教曰：“前将有父。”又藏三个住平之，房南相居兵无说，何三处急肉段，有一常索五阳，生曰：“马不可如魏耳。以曰：“公孙春观

还是挺好玩的

参考：https://blog.csdn.net/zzjcymbq/article/details/125666251