六.深度学习在NLP上的应用

四.机器翻译（未听）
五.马尔科夫模型（未听）

Intro

文本分类：Text-Label
行业Baseline：
用Bow表示sentences，然后用LR或者SVM做回归。

库：keras.io、gensim

Auto-Encoder

历史使命：数据降噪、数据降维
encoder+decoder
不需要label标注的算法

CNN4Text

用卷积可以实现图片直接的转换
如何应用于文字？

1.将文字转化成图片

抵用Word2Vec将文字转换成向量，堆积成图片
结构设计：
输入文字向量图片，使用pooling层进行降维（maxpool）变成一维排布，放入DL进行学习

2.将CNN做成1D

将一维的向量输入1D的CNN

CNN hypothesis

人眼在看文字时，第一眼也是看成图片，小的语序变化并不会影响阅读

案例：每日新闻

NLP
预处理：去除非字母、全小写、去除STOPWORDS、lemma(归一化)
可能的NLP模型：
1.TF-IDF
2.AutoEncoder(simple\deep\conv)
3.W2V
4.W2V+CNN(1D)
5.W2V+CNN
…

RNN

RNN的目的是让有时间上的前后信息得到考虑。
每一个神经元除了上一层的输入，还有上一个时间的神经元的记忆输入。
一个问题是记忆不能保持的很久。

LSTM长效记忆网络

每个神经元变得更加复杂
有一个记忆纽带Cell State，用异或和相加来进行处理。
控制的阀门叫做Gate，输出1代表需要记忆此次信息，0代表忘记。

第一步：忘记门

第二部：记忆门

用sigmoid决定什么信息需要更新；
用tanh造一个新的Cell State进行记忆

第三步：更新门

用异或和加和来进行更新

第四步：输出门

输出会被用来下一个的输入

LSTM 做文本生成

char级别

#举个小小的例子，来看看LSTM是怎么玩的

#我们这里用温斯顿丘吉尔的人物传记作为我们的学习语料。

#第一步，一样，先导入各种库

import numpy
from keras.models import Sequential # 层级
from keras.layers import Dense
from keras.layers import Dropout #去除放过拟合
from keras.layers import LSTM # LSTM
from keras.callbacks import ModelCheckpoint 
from keras.utils import np_utils

#接下来，我们把文本读入

raw_text = open('../input/Winston_Churchil.txt').read()
raw_text = raw_text.lower() #小写化
#既然我们是以每个字母为层级，字母总共才26个，所以我们可以很方便的用One-Hot来编码出所有的字母（当然，可能还有些标点符号和其他noise）

chars = sorted(list(set(raw_text))) # 去重、列表、排序
char_to_int = dict((c, i) for i, c in enumerate(chars)) # 形成char和索引的一一对照表
int_to_char = dict((i, c) for i, c in enumerate(chars))
#我们看到，全部的chars：

chars
['\n',
 ' ',
 '!',
 '#',
 '$',
 '%',
 '(',
 ')',
 '*',
 ',',
 '-',
 '.',
 '/',
 '0',
 '1',
 '2',
 '3',
 '4',
 '5',
 '6',
 '7',
 '8',
 '9',
 ':',
 ';',
 '?',
 '@',
 '[',
 ']',
 '_',
 'a',
 'b',
 'c',
 'd',
 'e',
 'f',
 'g',
 'h',
 'i',
 'j',
 'k',
 'l',
 'm',
 'n',
 'o',
 'p',
 'q',
 'r',
 's',
 't',
 'u',
 'v',
 'w',
 'x',
 'y',
 'z',
 '‘',
 '’',
 '“',
 '”',
 '\ufeff']
#一共有：

len(chars)
61
#同时，我们的原文本一共有:

len(raw_text)
276830
#我们这里简单的文本预测就是，给了前置的字母以后，下一个字母是谁？

#比如，Winsto, 给出 n Britai 给出 n

#构造训练测试集
#我们需要把我们的raw text变成可以用来训练的x,y:

#x 是前置字母们 y 是后一个字母

seq_length = 100 # 前后语境的长度
x = []
y = []
for i in range(0, len(raw_text) - seq_length): 
    given = raw_text[i:i + seq_length] # 前1000个作为已经给出的
    predict = raw_text[i + seq_length] # 第101个作为被预测的结果
    x.append([char_to_int[char] for char in given]) # 将上面的字符转成数字放进x中
    y.append(char_to_int[predict])
#我们可以看看我们做好的数据集的长相：

print(x[:3]) #x的一个数据点有100个
print(y[:3]) #y的一个数据点有1个
[[60, 45, 47, 44, 39, 34, 32, 49, 1, 36, 50, 49, 34, 43, 31, 34, 47, 36, 57, 48, 1, 47, 34, 30, 41, 1, 48, 44, 41, 33, 38, 34, 47, 48, 1, 44, 35, 1, 35, 44, 47, 49, 50, 43, 34, 9, 1, 31, 54, 1, 47, 38, 32, 37, 30, 47, 33, 1, 37, 30, 47, 33, 38, 43, 36, 1, 33, 30, 51, 38, 48, 0, 0, 49, 37, 38, 48, 1, 34, 31, 44, 44, 40, 1, 38, 48, 1, 35, 44, 47, 1, 49, 37, 34, 1, 50, 48, 34, 1, 44], [45, 47, 44, 39, 34, 32, 49, 1, 36, 50, 49, 34, 43, 31, 34, 47, 36, 57, 48, 1, 47, 34, 30, 41, 1, 48, 44, 41, 33, 38, 34, 47, 48, 1, 44, 35, 1, 35, 44, 47, 49, 50, 43, 34, 9, 1, 31, 54, 1, 47, 38, 32, 37, 30, 47, 33, 1, 37, 30, 47, 33, 38, 43, 36, 1, 33, 30, 51, 38, 48, 0, 0, 49, 37, 38, 48, 1, 34, 31, 44, 44, 40, 1, 38, 48, 1, 35, 44, 47, 1, 49, 37, 34, 1, 50, 48, 34, 1, 44, 35], [47, 44, 39, 34, 32, 49, 1, 36, 50, 49, 34, 43, 31, 34, 47, 36, 57, 48, 1, 47, 34, 30, 41, 1, 48, 44, 41, 33, 38, 34, 47, 48, 1, 44, 35, 1, 35, 44, 47, 49, 50, 43, 34, 9, 1, 31, 54, 1, 47, 38, 32, 37, 30, 47, 33, 1, 37, 30, 47, 33, 38, 43, 36, 1, 33, 30, 51, 38, 48, 0, 0, 49, 37, 38, 48, 1, 34, 31, 44, 44, 40, 1, 38, 48, 1, 35, 44, 47, 1, 49, 37, 34, 1, 50, 48, 34, 1, 44, 35, 1]]
[35, 1, 30]
#此刻，楼上这些表达方式，类似就是一个词袋，或者说 index。

#接下来我们做两件事：

#我们已经有了一个input的数字表达（index），我们要把它变成LSTM需要的数组格式： [样本数，时间步伐，特征]，时间步伐是指想要记忆链条的时间长度，在这里就是100个。

#第二，对于output，我们在Word2Vec里学过，用one-hot做output的预测可以给我们更好的效果，相对于直接预测一个准确的y数值的话。

n_patterns = len(x)
n_vocab = len(chars)

# 处理x和y
# 把x变成LSTM需要的样子，二维数据变成三位数组
x = numpy.reshape(x, (n_patterns, seq_length, 1))
# 简单normal到0-1之间，归一化
x = x / float(n_vocab)
# output变成one-hot
y = np_utils.to_categorical(y)

print(x[11])
print(y[11])
'''
[[ 0.80327869]
 [ 0.55737705]
 [ 0.70491803]
 [ 0.50819672]
 [ 0.55737705]
 [ 0.7704918 ]
 [ 0.59016393]
 [ 0.93442623]
 [ 0.78688525]
 [ 0.01639344]
 [ 0.7704918 ]
 [ 0.55737705]
 [ 0.49180328]
 [ 0.67213115]
 [ 0.01639344]
 [ 0.78688525]
 [ 0.72131148]
 [ 0.67213115]
 [ 0.54098361]
 [ 0.62295082]
 [ 0.55737705]
 [ 0.7704918 ]
 [ 0.78688525]
 [ 0.01639344]
 [ 0.72131148]
 [ 0.57377049]
 [ 0.01639344]
 [ 0.57377049]
 [ 0.72131148]
 [ 0.7704918 ]
 [ 0.80327869]
 [ 0.81967213]
 [ 0.70491803]
 [ 0.55737705]
 [ 0.14754098]
 [ 0.01639344]
 [ 0.50819672]
 [ 0.8852459 ]
 [ 0.01639344]
 [ 0.7704918 ]
 [ 0.62295082]
 [ 0.52459016]
 [ 0.60655738]
 [ 0.49180328]
 [ 0.7704918 ]
 [ 0.54098361]
 [ 0.01639344]
 [ 0.60655738]
 [ 0.49180328]
 [ 0.7704918 ]
 [ 0.54098361]
 [ 0.62295082]
 [ 0.70491803]
 [ 0.59016393]
 [ 0.01639344]
 [ 0.54098361]
 [ 0.49180328]
 [ 0.83606557]
 [ 0.62295082]
 [ 0.78688525]
 [ 0.        ]
 [ 0.        ]
 [ 0.80327869]
 [ 0.60655738]
 [ 0.62295082]
 [ 0.78688525]
 [ 0.01639344]
 [ 0.55737705]
 [ 0.50819672]
 [ 0.72131148]
 [ 0.72131148]
 [ 0.6557377 ]
 [ 0.01639344]
 [ 0.62295082]
 [ 0.78688525]
 [ 0.01639344]
 [ 0.57377049]
 [ 0.72131148]
 [ 0.7704918 ]
 [ 0.01639344]
 [ 0.80327869]
 [ 0.60655738]
 [ 0.55737705]
 [ 0.01639344]
 [ 0.81967213]
 [ 0.78688525]
 [ 0.55737705]
 [ 0.01639344]
 [ 0.72131148]
 [ 0.57377049]
 [ 0.01639344]
 [ 0.49180328]
 [ 0.70491803]
 [ 0.8852459 ]
 [ 0.72131148]
 [ 0.70491803]
 [ 0.55737705]
 [ 0.01639344]
 [ 0.49180328]
 [ 0.70491803]]
 # y全变成了由0和1构成的数组，loss也可以由向量进行差值运算
[ 0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.
  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.
  1.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.
  2.  0.  0.  0.  0.  0.]
'''

#模型建造
#LSTM模型构建

model = Sequential() # 表明每层都是前后关系（还有图像型关系）
model.add(LSTM(256, input_shape=(x.shape[1], x.shape[2]))) # 只考虑两个维度，把x的后面的维度提取就可以（时间步伐）
model.add(Dropout(0.2)) # 舍弃0.8的
model.add(Dense(y.shape[1], activation='softmax')) # 普通神经网络
model.compile(loss='categorical_crossentropy', optimizer='adam') # 多元熵值作为loss
#跑模型设置轮数和每次跑的数量
model.fit(x, y, nb_epoch=50, batch_size=4096) 
'''
Epoch 1/50
276730/276730 [==============================] - 197s - loss: 3.1120   
Epoch 2/50
276730/276730 [==============================] - 197s - loss: 3.0227   
Epoch 3/50
276730/276730 [==============================] - 197s - loss: 2.9910   
Epoch 4/50
276730/276730 [==============================] - 197s - loss: 2.9337   
Epoch 5/50
276730/276730 [==============================] - 197s - loss: 2.8971   
Epoch 6/50
276730/276730 [==============================] - 197s - loss: 2.8784   
Epoch 7/50
276730/276730 [==============================] - 197s - loss: 2.8640   
Epoch 8/50
276730/276730 [==============================] - 197s - loss: 2.8516   
Epoch 9/50
276730/276730 [==============================] - 197s - loss: 2.8384   
Epoch 10/50
276730/276730 [==============================] - 197s - loss: 2.8254   
Epoch 11/50
276730/276730 [==============================] - 197s - loss: 2.8133   
Epoch 12/50
276730/276730 [==============================] - 197s - loss: 2.8032   
Epoch 13/50
276730/276730 [==============================] - 197s - loss: 2.7913   
Epoch 14/50
276730/276730 [==============================] - 197s - loss: 2.7831   
Epoch 15/50
276730/276730 [==============================] - 197s - loss: 2.7744   
Epoch 16/50
276730/276730 [==============================] - 197s - loss: 2.7672   
Epoch 17/50
276730/276730 [==============================] - 197s - loss: 2.7601   
Epoch 18/50
276730/276730 [==============================] - 197s - loss: 2.7540   
Epoch 19/50
276730/276730 [==============================] - 197s - loss: 2.7477   
Epoch 20/50
276730/276730 [==============================] - 197s - loss: 2.7418   
Epoch 21/50
276730/276730 [==============================] - 197s - loss: 2.7360   
Epoch 22/50
276730/276730 [==============================] - 197s - loss: 2.7296   
Epoch 23/50
276730/276730 [==============================] - 197s - loss: 2.7238   
Epoch 24/50
276730/276730 [==============================] - 197s - loss: 2.7180   
Epoch 25/50
276730/276730 [==============================] - 197s - loss: 2.7113   
Epoch 26/50
276730/276730 [==============================] - 197s - loss: 2.7055   
Epoch 27/50
276730/276730 [==============================] - 197s - loss: 2.7000   
Epoch 28/50
276730/276730 [==============================] - 197s - loss: 2.6934   
Epoch 29/50
276730/276730 [==============================] - 197s - loss: 2.6859   
Epoch 30/50
276730/276730 [==============================] - 197s - loss: 2.6800   
Epoch 31/50
276730/276730 [==============================] - 197s - loss: 2.6741   
Epoch 32/50
276730/276730 [==============================] - 197s - loss: 2.6669   
Epoch 33/50
276730/276730 [==============================] - 197s - loss: 2.6593   
Epoch 34/50
276730/276730 [==============================] - 197s - loss: 2.6529   
Epoch 35/50
276730/276730 [==============================] - 197s - loss: 2.6461   
Epoch 36/50
276730/276730 [==============================] - 197s - loss: 2.6385   
Epoch 37/50
276730/276730 [==============================] - 197s - loss: 2.6320   
Epoch 38/50
276730/276730 [==============================] - 197s - loss: 2.6249   
Epoch 39/50
276730/276730 [==============================] - 197s - loss: 2.6187   
Epoch 40/50
276730/276730 [==============================] - 197s - loss: 2.6110   
Epoch 41/50
276730/276730 [==============================] - 192s - loss: 2.6039   
Epoch 42/50
276730/276730 [==============================] - 141s - loss: 2.5969   
Epoch 43/50
276730/276730 [==============================] - 140s - loss: 2.5909   
Epoch 44/50
276730/276730 [==============================] - 140s - loss: 2.5843   
Epoch 45/50
276730/276730 [==============================] - 140s - loss: 2.5763   
Epoch 46/50
276730/276730 [==============================] - 140s - loss: 2.5697   
Epoch 47/50
276730/276730 [==============================] - 141s - loss: 2.5635   
Epoch 48/50
276730/276730 [==============================] - 140s - loss: 2.5575   
Epoch 49/50
276730/276730 [==============================] - 140s - loss: 2.5496   
Epoch 50/50
276730/276730 [==============================] - 140s - loss: 2.5451   
<keras.callbacks.History at 0x7fb6121b6e48>
'''
#我们来写个程序，看看我们训练出来的LSTM的效果：

def predict_next(input_array): # 给输入的数组放入模型中求输出
    x = numpy.reshape(input_array, (1, seq_length, 1))
    x = x / float(n_vocab)
    y = model.predict(x)
    return y

def string_to_index(raw_input): # 将输入的字符变成数字的集合
    res = []
    for c in raw_input[(len(raw_input)-seq_length):]:
        res.append(char_to_int[c])
    return res

def y_to_char(y): # 将one-hot编码转换回char
    largest_index = y.argmax() # y中数值最大对应的序列号
    c = int_to_char[largest_index]
    return c
好，写成一个大程序：

def generate_article(init, rounds=200):
    in_string = init.lower()
    for i in range(rounds):
        n = y_to_char(predict_next(string_to_index(in_string)))
        in_string += n
    return in_string
# 进行测试
init = 'His object in coming to New York was to engage officers for that service. He came at an opportune moment'
article = generate_article(init)
print(article)
'''
his object in coming to new york was to engage officers for that service. he came at an opportune moment th the toote of the carie and the soote of the carie and the soote of the carie and the soote of the carie and the soote of the carie and the soote of the carie and the soote of the carie and the soo
'''

word级别

设计到一个word2vec

#我们这里不再用char级别，我们用word级别来做。

#第一步，一样，先导入各种库

import os
import numpy as np
import nltk
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import Dropout
from keras.layers import LSTM
from keras.callbacks import ModelCheckpoint
from keras.utils import np_utils
from gensim.models.word2vec import Word2Vec

#接下来，我们把文本读入

raw_text = ''
for file in os.listdir("../input/"):
    if file.endswith(".txt"):
        raw_text += open("../input/"+file, errors='ignore').read() + '\n\n'
# raw_text = open('../input/Winston_Churchil.txt').read()
raw_text = raw_text.lower()
sentensor = nltk.data.load('tokenizers/punkt/english.pickle')        
sents = sentensor.tokenize(raw_text) # 将句子相互分开
corpus = [] #二维数组，句子和单词
for sen in sents:
    corpus.append(nltk.word_tokenize(sen)) # 将每个句子再分词

print(len(corpus)) 
print(corpus[:3])
'''
91007
[['\ufeffthe', 'project', 'gutenberg', 'ebook', 'of', 'great', 'expectations', ',', 'by', 'charles', 'dickens', 'this', 'ebook', 'is', 'for', 'the', 'use', 'of', 'anyone', 'anywhere', 'at', 'no', 'cost', 'and', 'with', 'almost', 'no', 'restrictions', 'whatsoever', '.'], ['you', 'may', 'copy', 'it', ',', 'give', 'it', 'away', 'or', 're-use', 'it', 'under', 'the', 'terms', 'of', 'the', 'project', 'gutenberg', 'license', 'included', 'with', 'this', 'ebook', 'or', 'online', 'at', 'www.gutenberg.org', 'title', ':', 'great', 'expectations', 'author', ':', 'charles', 'dickens', 'posting', 'date', ':', 'august', '20', ',', '2008', '[', 'ebook', '#', '1400', ']', 'release', 'date', ':', 'july', ',', '1998', 'last', 'updated', ':', 'september', '25', ',', '2016', 'language', ':', 'english', 'character', 'set', 'encoding', ':', 'utf-8', '***', 'start', 'of', 'this', 'project', 'gutenberg', 'ebook', 'great', 'expectations', '***', 'produced', 'by', 'an', 'anonymous', 'volunteer', 'great', 'expectations', '[', '1867', 'edition', ']', 'by', 'charles', 'dickens', '[', 'project', 'gutenberg', 'editor’s', 'note', ':', 'there', 'is', 'also', 'another', 'version', 'of', 'this', 'work', 'etext98/grexp10.txt', 'scanned', 'from', 'a', 'different', 'edition', ']', 'chapter', 'i', 'my', 'father’s', 'family', 'name', 'being', 'pirrip', ',', 'and', 'my', 'christian', 'name', 'philip', ',', 'my', 'infant', 'tongue', 'could', 'make', 'of', 'both', 'names', 'nothing', 'longer', 'or', 'more', 'explicit', 'than', 'pip', '.'], ['so', ',', 'i', 'called', 'myself', 'pip', ',', 'and', 'came', 'to', 'be', 'called', 'pip', '.']]
'''
#好，w2v乱炖：

w2v_model = Word2Vec(corpus, size=128, window=5, min_count=5, workers=4) # 窗口长度5，线程4
#可以了

w2v_model['office']
'''
array([-0.01398709,  0.15975526,  0.03589381, -0.4449192 ,  0.365403  ,
        0.13376504,  0.78731823,  0.01640314, -0.29723561, -0.21117583,
        0.13451998, -0.65348488,  0.06038611, -0.02000343,  0.05698346,
        0.68013376,  0.19010596,  0.56921762,  0.66904438, -0.08069923,
       -0.30662233,  0.26082459, -0.74816126, -0.41383636, -0.56303871,
       -0.10834043, -0.10635001, -0.7193433 ,  0.29722607, -0.83104628,
        1.11914253, -0.34119046, -0.39490014, -0.34709939, -0.00583572,
        0.17824887,  0.43295503,  0.11827419, -0.28707108, -0.02838829,
        0.02565269,  0.10328653, -0.19100265, -0.24102989,  0.23023468,
        0.51493132,  0.34759828,  0.05510307,  0.20583512, -0.17160387,
       -0.10351282,  0.19884749, -0.03935663, -0.04055062,  0.38888735,
       -0.02003323, -0.16577065, -0.15858875,  0.45083243, -0.09268586,
       -0.91098118,  0.16775337,  0.3432925 ,  0.2103184 , -0.42439541,
        0.26097715, -0.10714807,  0.2415273 ,  0.2352251 , -0.21662289,
       -0.13343927,  0.11787982, -0.31010333,  0.21146733, -0.11726214,
       -0.65574747,  0.04007725, -0.12032496, -0.03468512,  0.11063002,
        0.33530036, -0.64098376,  0.34013858, -0.08341357, -0.54826909,
        0.0723564 , -0.05169795, -0.19633259,  0.08620321,  0.05993884,
       -0.14693044, -0.40531522, -0.07695422,  0.2279872 , -0.12342903,
       -0.1919964 , -0.09589464,  0.4433476 ,  0.38304719,  1.0319351 ,
        0.82628119,  0.3677327 ,  0.07600326,  0.08538571, -0.44261214,
       -0.10997667, -0.03823839,  0.40593523,  0.32665277, -0.67680383,
        0.32504487,  0.4009226 ,  0.23463745, -0.21442334,  0.42727917,
        0.19593567, -0.10731711, -0.01080817, -0.14738144,  0.15710345,
       -0.01099576,  0.35833639,  0.16394758, -0.10431164, -0.28202233,
        0.24488974,  0.69327635, -0.29230621], dtype=float32)
'''

#接下来，其实我们还是以之前的方式来处理我们的training data，把源数据变成一个长长的x，好让LSTM学会predict下一个单词：

raw_input = [item for sublist in corpus for item in sublist] # 将二维数组变成一维数组
len(raw_input)
'''
2115170
'''
raw_input[12]
'ebook'
text_stream = []
vocab = w2v_model.vocab
for word in raw_input:
    if word in vocab:
        text_stream.append(word)
len(text_stream)
'''
2058753
'''
#我们这里的文本预测就是，给了前面的单词以后，下一个单词是谁？

#比如，hello from the other, 给出 side

#构造训练测试集
#我们需要把我们的raw text变成可以用来训练的x,y:

#x 是前置字母们 y 是后一个字母

seq_length = 10 # 语境长度10
x = []
y = []
for i in range(0, len(text_stream) - seq_length):

    given = text_stream[i:i + seq_length]
    predict = text_stream[i + seq_length]
    x.append(np.array([w2v_model[word] for word in given]))
    y.append(w2v_model[predict])
我们可以看看我们做好的数据集的长相：

print(x[10])
print(y[10])
[[-0.02218935  0.04861801 -0.03001036 ...,  0.07096259  0.16345282
  -0.18007144]
 [ 0.1663752   0.67981642  0.36581406 ...,  1.03355932  0.94110376
  -1.02763569]
 [-0.12611888  0.75773817  0.00454156 ...,  0.80544478  2.77890372
  -1.00110698]
 ..., 
 [ 0.34167829 -0.28152692 -0.12020591 ...,  0.19967555  1.65415502
  -1.97690392]
 [-0.66742641  0.82389861 -1.22558379 ...,  0.12269551  0.30856156
   0.29964617]
 [-0.17075984  0.0066567  -0.3894183  ...,  0.23729582  0.41993639
  -0.12582727]]
[ 0.18125793 -1.72401989 -0.13503326 -0.42429626  1.40763748 -2.16775346
  2.26685596 -2.03301549  0.42729807 -0.84830129  0.56945151  0.87243706
  3.01571465 -0.38155749 -0.99618471  1.1960727   1.93537641  0.81187075
 -0.83017075 -3.18952608  0.48388934 -0.03766865 -1.68608069 -1.84907544
 -0.95259917  0.49039507 -0.40943271  0.12804921  1.35876858  0.72395176
  1.43591952 -0.41952157  0.38778016 -0.75301784 -2.5016799  -0.85931653
 -1.39363682  0.42932403  1.77297652  0.41443667 -1.30974782 -0.08950856
 -0.15183811 -1.59824061 -1.58920395  1.03765178  2.07559252  2.79692245
  1.11855054 -0.25542653 -1.04980111 -0.86929852 -1.26279402 -1.14124119
 -1.04608357  1.97869778 -2.23650813 -2.18115139 -0.26534671  0.39432198
 -0.06398458 -1.02308178  1.43372631 -0.02581184 -0.96472031 -3.08931994
 -0.67289352  1.06766248 -1.95796657  1.40857184  0.61604798 -0.50270212
 -2.33530831  0.45953822  0.37867084 -0.56957626 -1.90680516 -0.57678169
  0.50550407 -0.30320352  0.19682285  1.88185465 -1.40448165 -0.43952951
  1.95433044  2.07346153  0.22390689 -0.95107335 -0.24579825 -0.21493609
  0.66570002 -0.59126669 -1.4761591   0.86431485  0.36701021  0.12569368
  1.65063572  2.048352    1.81440067 -1.36734581  2.41072559  1.30975604
 -0.36556485 -0.89859813  1.28804696 -2.75488496  1.5667206  -1.75327337
  0.60426879  1.77851915 -0.32698369  0.55594021  2.01069188 -0.52870172
 -0.39022744 -1.1704396   1.28902853 -0.89315164  1.41299319  0.43392688
 -2.52578211 -1.13480854 -1.05396986 -0.85470092  0.6618616   1.23047733
 -0.28597715 -2.35096407]
print(len(x))
print(len(y))
print(len(x[12]))
print(len(x[12][0]))
print(len(y[12]))
2058743
2058743
10
128
128
x = np.reshape(x, (-1, seq_length, 128))
y = np.reshape(y, (-1,128))
接下来我们做两件事：

我们已经有了一个input的数字表达（w2v），我们要把它变成LSTM需要的数组格式： [样本数，时间步伐，特征]

第二，对于output，我们直接用128维的输出

模型建造
LSTM模型构建

model = Sequential()
model.add(LSTM(256, dropout_W=0.2, dropout_U=0.2, input_shape=(seq_length, 128)))
model.add(Dropout(0.2))
model.add(Dense(128, activation='sigmoid'))
model.compile(loss='mse', optimizer='adam')
跑模型
model.fit(x, y, nb_epoch=50, batch_size=4096)
Epoch 1/50
2058743/2058743 [==============================] - 150s - loss: 0.6839   
Epoch 2/50
2058743/2058743 [==============================] - 150s - loss: 0.6670   
Epoch 3/50
2058743/2058743 [==============================] - 150s - loss: 0.6625   
Epoch 4/50
2058743/2058743 [==============================] - 150s - loss: 0.6598   
Epoch 5/50
2058743/2058743 [==============================] - 150s - loss: 0.6577   
Epoch 6/50
2058743/2058743 [==============================] - 150s - loss: 0.6562   
Epoch 7/50
2058743/2058743 [==============================] - 150s - loss: 0.6549   
Epoch 8/50
2058743/2058743 [==============================] - 150s - loss: 0.6537   
Epoch 9/50
2058743/2058743 [==============================] - 150s - loss: 0.6527   
Epoch 10/50
2058743/2058743 [==============================] - 150s - loss: 0.6519   
Epoch 11/50
2058743/2058743 [==============================] - 150s - loss: 0.6512   
Epoch 12/50
2058743/2058743 [==============================] - 150s - loss: 0.6506   
Epoch 13/50
2058743/2058743 [==============================] - 150s - loss: 0.6500   
Epoch 14/50
2058743/2058743 [==============================] - 150s - loss: 0.6496   
Epoch 15/50
2058743/2058743 [==============================] - 150s - loss: 0.6492   
Epoch 16/50
2058743/2058743 [==============================] - 150s - loss: 0.6488   
Epoch 17/50
2058743/2058743 [==============================] - 151s - loss: 0.6485   
Epoch 18/50
2058743/2058743 [==============================] - 150s - loss: 0.6482   
Epoch 19/50
2058743/2058743 [==============================] - 150s - loss: 0.6480   
Epoch 20/50
2058743/2058743 [==============================] - 150s - loss: 0.6477   
Epoch 21/50
2058743/2058743 [==============================] - 150s - loss: 0.6475   
Epoch 22/50
2058743/2058743 [==============================] - 150s - loss: 0.6473   
Epoch 23/50
2058743/2058743 [==============================] - 150s - loss: 0.6471   
Epoch 24/50
2058743/2058743 [==============================] - 150s - loss: 0.6470   
Epoch 25/50
2058743/2058743 [==============================] - 150s - loss: 0.6468   
Epoch 26/50
2058743/2058743 [==============================] - 150s - loss: 0.6466   
Epoch 27/50
2058743/2058743 [==============================] - 150s - loss: 0.6464   
Epoch 28/50
2058743/2058743 [==============================] - 150s - loss: 0.6463   
Epoch 29/50
2058743/2058743 [==============================] - 150s - loss: 0.6462   
Epoch 30/50
2058743/2058743 [==============================] - 150s - loss: 0.6461   
Epoch 31/50
2058743/2058743 [==============================] - 150s - loss: 0.6460   
Epoch 32/50
2058743/2058743 [==============================] - 150s - loss: 0.6458   
Epoch 33/50
2058743/2058743 [==============================] - 150s - loss: 0.6458   
Epoch 34/50
2058743/2058743 [==============================] - 150s - loss: 0.6456   
Epoch 35/50
2058743/2058743 [==============================] - 150s - loss: 0.6456   
Epoch 36/50
2058743/2058743 [==============================] - 150s - loss: 0.6455   
Epoch 37/50
2058743/2058743 [==============================] - 150s - loss: 0.6454   
Epoch 38/50
2058743/2058743 [==============================] - 150s - loss: 0.6453   
Epoch 39/50
2058743/2058743 [==============================] - 150s - loss: 0.6452   
Epoch 40/50
2058743/2058743 [==============================] - 150s - loss: 0.6452   
Epoch 41/50
2058743/2058743 [==============================] - 150s - loss: 0.6451   
Epoch 42/50
2058743/2058743 [==============================] - 150s - loss: 0.6450   
Epoch 43/50
2058743/2058743 [==============================] - 150s - loss: 0.6450   
Epoch 44/50
2058743/2058743 [==============================] - 150s - loss: 0.6449   
Epoch 45/50
2058743/2058743 [==============================] - 150s - loss: 0.6448   
Epoch 46/50
2058743/2058743 [==============================] - 150s - loss: 0.6447   
Epoch 47/50
2058743/2058743 [==============================] - 150s - loss: 0.6447   
Epoch 48/50
2058743/2058743 [==============================] - 150s - loss: 0.6446   
Epoch 49/50
2058743/2058743 [==============================] - 150s - loss: 0.6446   
Epoch 50/50
2058743/2058743 [==============================] - 150s - loss: 0.6445   
<keras.callbacks.History at 0x7f6ed8816a58>
我们来写个程序，看看我们训练出来的LSTM的效果：

def predict_next(input_array):
    x = np.reshape(input_array, (-1,seq_length,128))
    y = model.predict(x)
    return y

def string_to_index(raw_input):
    raw_input = raw_input.lower()
    input_stream = nltk.word_tokenize(raw_input)
    res = []
    for word in input_stream[(len(input_stream)-seq_length):]:
        res.append(w2v_model[word])
    return res

def y_to_word(y):
    word = w2v_model.most_similar(positive=y, topn=1)
    return word
好，写成一个大程序：

def generate_article(init, rounds=30):
    in_string = init.lower()
    for i in range(rounds):
        n = y_to_word(predict_next(string_to_index(in_string)))
        in_string += ' ' + n[0][0]
    return in_string
init = 'Language Models allow us to measure how likely a sentence is, which is an important for Machine'
article = generate_article(init)
print(article)
language models allow us to measure how likely a sentence is, which is an important for machine engagement . to-day good-for-nothing fit job job job job job . i feel thing job job job ; thing really done certainly job job ; but i need not say