原文地址：Keras 实现 LSTM

本文在原文的基础上添加了一些注释、运行结果和修改了少量的代码。

1. 介绍

LSTM(Long Short Term Memory)是一种特殊的循环神经网络，在许多任务中，LSTM表现得比标准的RNN要出色得多。

关于LSTM的介绍可以看参考文献1和2。本文重点在使用LSTM实现一个分类器。

2. 如何在 keras 中使用LSTM

本文主要测试 keras
使用Word Embeddings
并进行分类的测试。代码是在keras
官方文档的示例中修改而来。IPython代码链接

2.1 Word Embeddings 数据集

使用了stanford的GloVe作为词向量集，这个直接下载训练好的词向量文件。直接字典搜索，得到文本词向量。Glove数据集下载文本测试数据是20_newsgroup

This data set is a collection of 20,000 messages, collected from 20 different netnews newsgroups. One thousand messages from each of the twenty newsgroups were chosen at random and partitioned by newsgroup name. The list of newsgroups from which the messages were chose is as follows:
alt.atheism
talk.politics.guns
talk.politics.mideast
talk.politics.misc
talk.religion.misc
soc.religion.christian
comp.sys.ibm.pc.hardware
comp.graphics
comp.os.ms-windows.misc
comp.sys.mac.hardware
comp.windows.x
rec.autos
rec.motorcycles
rec.sport.baseball
rec.sport.hockey
sci.crypt
sci.electronics
sci.space
sci.med
misc.forsale

我们通过label标注把message分成不同的20个类别。每个newsgroup被map到一个数值label上。

需要用到的模块

import numpy as np
import os
import sys
import random
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
from keras.utils.np_utils import to_categorical
from keras.models import Sequential
from keras.layers import Embedding, LSTM, Dense, Activation

2.2 数据预处理

这部分是设定训练相关参数，并且读入训练好的GloVe词向量文件。把文本读入进list里，一个文本存成一个str，变成一个[str]

BASE_DIR = '/home/lich/Workspace/Learning'
GLOVE_DIR = BASE_DIR + '/glove.6B/'
TEXT_DATA_DIR = BASE_DIR + '/20_newsgroup/'
MAX_SEQUENCE_LENGTH = 1000
MAX_NB_WORDS = 20000
EMBEDDING_DIM = 100
VALIDATION_SPLIT = 0.2
batch_size = 32

# first, build index mapping words in the embeddings set
# to their embedding vector

embeddings_index = {}
f = open(os.path.join(GLOVE_DIR, 'glove.6B.100d.txt'))
for line in f:
    values = line.split()
    word = values[0]
    coefs = np.asarray(values[1:], dtype='float32')
    embeddings_index[word] = coefs
f.close()

print('Found %s word vectors.' % len(embeddings_index))
#Found 400000 word vectors.

# second, prepare text samples and their labels
print('Processing text dataset')

texts = []  # list of text samples
labels_index = {}  # dictionary mapping label name to numeric id
labels = []  # list of label ids
for name in sorted(os.listdir(TEXT_DATA_DIR)):
    path = os.path.join(TEXT_DATA_DIR, name)
    if os.path.isdir(path):
        label_id = len(labels_index)
        labels_index[name] = label_id
        for fname in sorted(os.listdir(path)):
            if fname.isdigit():
                fpath = os.path.join(path, fname)
                if sys.version_info < (3,):
                    f = open(fpath)
                else:
                    f = open(fpath, encoding='latin-1')
                texts.append(f.read())
                f.close()
                labels.append(label_id)

print('Found %s texts.' % len(texts))
#Found 19997 texts.

embeddings_index 里面是这样：

embeddings_index['hi']
"""
array([ 0.1444    ,  0.23978999,  0.96692997,  0.31628999, -0.36063999,
       -0.87673998,  0.098512  ,  0.31077999,  0.47929001,  0.27175   ,
        0.30004999, -0.23732001, -0.31516999,  0.17925   ,  0.61773002,
        0.59820998,  0.49489   ,  0.3423    , -0.078034  ,  0.60211998,
        0.18683   ,  0.52069998, -0.12331   ,  0.48313001, -0.24117   ,
        0.59696001,  0.61078   , -0.84413999,  0.27660999,  0.068767  ,
       -1.13880002,  0.089544  ,  0.89841998,  0.53788   ,  0.10841   ,
       -0.10038   ,  0.12921   ,  0.11476   , -0.47400001, -0.80489999,
        0.95999998, -0.36601999, -0.43019   , -0.39807999, -0.096782  ,
       -0.71183997, -0.31494001,  0.82345998,  0.42179   , -0.69204998,
       -1.48640001,  0.29497999, -0.30875   , -0.49994999, -0.46489999,
       -0.44523999,  0.81059998,  1.47570002,  0.53781998, -0.28270999,
       -0.045796  ,  0.14454   , -0.74484998,  0.35495001, -0.40961   ,
        0.35778999,  0.40061   ,  0.37338999,  0.72162998,  0.40812999,
        0.26155001, -0.14239   , -0.020514  , -1.11059999, -0.47670001,
        0.37832001,  0.89612001, -0.17323001, -0.50137001,  0.22991   ,
        1.53240001, -0.82032001, -0.10096   ,  0.45201999, -0.88638997,
        0.089056  , -0.19347   , -0.42253   ,  0.022429  ,  0.29444   ,
        0.020747  ,  0.48934999,  0.35991001,  0.092758  , -0.22428   ,
        0.60038   , -0.31850001, -0.72424001, -0.22632   , -0.030972  ], dtype=float32)
"""

embeddings_index['hi'].shape
# (100,)

labels_index 与 20_newsgroup 的20个分类一一对应

labels_index['alt.atheism']
#0

labels_index['comp.sys.ibm.pc.hardware']
#3

labels[:10]
#[0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

labels[1000:1010]
#[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]

labels[2000:2010]
#[2, 2, 2, 2, 2, 2, 2, 2, 2, 2]

打开其中一个 texts 看看

len(texts[2])
#4550

texts[2]
"""
Organization: Technical University Braunschweig, Germany
References: <16BA1E197.I3150101@dbstu1.rz.tu-bs.de> <65974@mimsy.umd.edu>
Date: Mon, 5 Apr 1993 19:08:25 GMT
Lines: 93
In article <65974@mimsy.umd.edu>
mangoe@cs.umd.edu (Charley Wingate) writes:
 
Well, John has a quite different, not necessarily more elaborated theology.
There is some evidence that he must have known Luke, and that the content
of Q was known to him, but not in a 'canonized' form.

This is a new argument to me.  Could you elaborate a little?
 
The argument goes as follows: Q-oid quotes appear in John, but not in
the almost codified way they were in Matthew or Luke. However, they are
considered to be similar enough to point to knowledge of Q as such, and
not an entirely different source.
 
Assuming that he knew Luke would obviously put him after Luke, and would
give evidence for the latter assumption.

I don't think this follows.  If you take the most traditional attributions,
then Luke might have known John, but John is an elder figure in either case.
We're talking spans of time here which are well within the range of
lifetimes.
 
We are talking date of texts here, not the age of the authors. The usual
explanation for the time order of Mark, Matthew and Luke does not consider
their respective ages. It says Matthew has read the text of Mark, and Luke
that of Matthew (and probably that of Mark).
 
As it is assumed that John knew the content of Luke's text. The evidence
for that is not overwhelming, admittedly.
 
(1)  Earlier manuscripts of John have been discovered.

Interesting, where and which? How are they dated? How old are they?

Unfortunately, I haven't got the info at hand.  It was (I think) in the late
'70s or early '80s, and it was possibly as old as CE 200.

When they are from about 200, why do they shed doubt on the order on
putting John after the rest of the three?
 
 
I don't see your point, it is exactly what James Felder said.  They had no
first hand knowledge of the events, and it obvious that at least two of them
used older texts as the base of their account.  And even the association of
Luke to Paul or Mark to Peter are not generally accepted.

Well, a genuine letter of Peter would be close enough, wouldn't it?
 
Sure, an original together with Id card of sender and receiver would be
fine. So what's that supposed to say? Am I missing something?
 
 
And I don't think a "one step removed" source is that bad.  If Luke and Mark
and Matthew learned their stories directly from diciples, then I really
cannot believe in the sort of "big transformation from Jesus to gospel" that
some people posit.  In news reports, one generally gets no better
information than this.

And if John IS a diciple, then there's nothing more to be said.

That John was a disciple is not generally accepted. The style and language
together with the theology are usually used as counterargument.

The argument that John was a disciple relies on the claim in the gospel
of John itself. Is there any other evidence for it?

One step and one generation removed is bad even in our times. Compare that
to reports of similar events in our century in almost illiterate societies.
Not even to speak off that believers are not necessarily the best sources.

It is also obvious that Mark has been edited. How old are the oldest
manuscripts? To my knowledge (which can be antiquated) the oldest is
quite after any of these estimates, and it is not even complete.

The only clear "editing" is problem of the ending, and it's basically a
hopeless mess.  The oldest versions give a strong sense of incompleteness,
to the point where the shortest versions seem to break off in midsentence.
The most obvious solution is that at some point part of the text was lost.
The material from verse 9 on is pretty clearly later and seems to represent
a synopsys of the end of Luke.

In other words, one does not know what the original of Mark did look like
and arguments based on Mark are pretty weak.
 
But how is that connected to a redating of John?
   Benedikt
"""

2.3Tokenize

Tokenizer 所有文本，并且把texts里面的str值先tokenizer然后映射到相应index。下面是举出的一个例子(只是形式一样)：

“he is a professor”
变成：
[143, 12, 1, 23]

# finally, vectorize the text samples into a 2D integer tensor
tokenizer = Tokenizer(nb_words=MAX_NB_WORDS)
tokenizer.fit_on_texts(texts)
sequences = tokenizer.texts_to_sequences(texts)

word_index = tokenizer.word_index
print('Found %s unique tokens.' % len(word_index))
#Found 214909 unique tokens.

上面的代码吧所有的单词都转换成了数字

word_index['newsgroups']
# 43

sequences[2][:20]
"""
[43,
 127,
 357,
 44,
 29,
 24,
 16,
 12,
 2,
 160,
 24,
 16,
 12,
 2,
 195,
 185,
 12,
 2,
 182,
 144]
"""

2.4 生成Train和Validate数据集

使用random.shuffle进行随机分割数据集，并声称相关训练验证集。

data = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH)

labels = to_categorical(np.asarray(labels))

print('Shape of data tensor:', data.shape)
print('Shape of label tensor:', labels.shape)
# ('Shape of data tensor:', (19997, 1000))
# ('Shape of label tensor:', (19997, 20))

# split the data into a training set and a validation set
indices = np.arange(data.shape[0])
np.random.shuffle(indices)
data = data[indices]
labels = labels[indices]
nb_validation_samples = int(VALIDATION_SPLIT * data.shape[0])

x_train = data[:-nb_validation_samples]
y_train = labels[:-nb_validation_samples]

x_train.shape
#(15998, 1000)

y_train.shape
#(15998, 20)

x_val = data[-nb_validation_samples:]
y_val = labels[-nb_validation_samples:]

print('Preparing embedding matrix.')

data 是一个长度为 1000 的 array，sequences 中不够长的部分被补0了。
labels 被转换成了 one-hot 编码的形式。

len(data)
#1000

data[2]
"""
array([    0,     0,     0,     0,     0,     0,     0,     0,     0,
           0,     0,     0,     0,     0,     0,     0,     0,     0,
                                      ...
                                      ...
          93,     6,  1818,   480,    19,   471,    25,   668,  2797,
          35,   111,     9,    10,  2425,     3,     5,     4,   370,  5271], dtype=int32)
"""


labels[0]
"""
array([ 1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,
        0.,  0.,  0.,  0.,  0.,  0.,  0.])
"""

labels[1000]
"""
array([ 0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,
        0.,  0.,  0.,  0.,  0.,  0.,  0.])
"""
labels[2000]
"""
array([ 0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,
        0.,  0.,  0.,  0.,  0.,  0.,  0.])
"""

2.5 生成Embedding Matrix

把有效出现次数在前面的通过GloVe生成的字典，以及本身所有的Token串进行比对，得到出现在训练集中每个词的词向量。

nb_words = min(MAX_NB_WORDS, len(word_index))
#20000
embedding_matrix = np.zeros((nb_words + 1, EMBEDDING_DIM))

for word, i in word_index.items():
    if i > MAX_NB_WORDS:
        continue
    embedding_vector = embeddings_index.get(word)
    if embedding_vector is not None:
        # words not found in embedding index will be all-zeros.
        embedding_matrix[i] = embedding_vector

print(embedding_matrix.shape)
#(20001, 100)

embedding_matrix 和 embeddings_index 是这样的：

embedding_matrix[76]
'''
array([ 0.1225    , -0.058833  ,  0.23658   , -0.28876999, -0.028181  ,
        0.31524   ,  0.070229  ,  0.16447   , -0.027623  ,  0.25213999,
        0.21174   , -0.059674  ,  0.36133   ,  0.13607   ,  0.18754999,
       -0.1487    ,  0.31314999,  0.13368   , -0.59702998, -0.030161  ,
        0.080656  ,  0.26161999, -0.055924  , -0.35350999,  0.34722   ,
       -0.0055801 , -0.57934999, -0.88006997,  0.42930999, -0.15695   ,
       -0.51256001,  1.26839995, -0.25228   ,  0.35264999, -0.46419001,
        0.55647999, -0.57555997,  0.32574001, -0.21893001, -0.13178   ,
       -1.1027    , -0.039591  ,  0.89643002, -0.98449999, -0.47393   ,
       -0.12854999,  0.63506001, -0.94888002,  0.40088001, -0.77542001,
       -0.35152999, -0.27788001,  0.68747002,  1.45799994, -0.38474   ,
       -2.89369988, -0.29523   , -0.38835999,  0.94880998,  1.38909996,
        0.054591  ,  0.70485997, -0.65698999,  0.075648  ,  0.76550001,
       -0.63365   ,  0.86556   ,  0.42440999,  0.14796001,  0.4156    ,
        0.29354   , -0.51295   ,  0.19634999, -0.45568001,  0.0080246 ,
        0.14528   , -0.15395001,  0.11406   , -1.21669996, -0.1111    ,
        0.82639998,  0.21738   , -0.63775998, -0.074874  , -1.71300006,
       -0.88270003, -0.0073058 , -0.37623   , -0.50208998, -0.58844   ,
       -0.24943   , -1.04250002,  0.27678001,  0.64142001, -0.64604998,
        0.43559   , -0.37276   , -0.0032068 ,  0.18743999,  0.30702001])
'''

embeddings_index.get('he')
'''
array([ 0.1225    , -0.058833  ,  0.23658   , -0.28876999, -0.028181  ,
        0.31524   ,  0.070229  ,  0.16447   , -0.027623  ,  0.25213999,
        0.21174   , -0.059674  ,  0.36133   ,  0.13607   ,  0.18754999,
       -0.1487    ,  0.31314999,  0.13368   , -0.59702998, -0.030161  ,
        0.080656  ,  0.26161999, -0.055924  , -0.35350999,  0.34722   ,
       -0.0055801 , -0.57934999, -0.88006997,  0.42930999, -0.15695   ,
       -0.51256001,  1.26839995, -0.25228   ,  0.35264999, -0.46419001,
        0.55647999, -0.57555997,  0.32574001, -0.21893001, -0.13178   ,
       -1.1027    , -0.039591  ,  0.89643002, -0.98449999, -0.47393   ,
       -0.12854999,  0.63506001, -0.94888002,  0.40088001, -0.77542001,
       -0.35152999, -0.27788001,  0.68747002,  1.45799994, -0.38474   ,
       -2.89369988, -0.29523   , -0.38835999,  0.94880998,  1.38909996,
        0.054591  ,  0.70485997, -0.65698999,  0.075648  ,  0.76550001,
       -0.63365   ,  0.86556   ,  0.42440999,  0.14796001,  0.4156    ,
        0.29354   , -0.51295   ,  0.19634999, -0.45568001,  0.0080246 ,
        0.14528   , -0.15395001,  0.11406   , -1.21669996, -0.1111    ,
        0.82639998,  0.21738   , -0.63775998, -0.074874  , -1.71300006,
       -0.88270003, -0.0073058 , -0.37623   , -0.50208998, -0.58844   ,
       -0.24943   , -1.04250002,  0.27678001,  0.64142001, -0.64604998,
        0.43559   , -0.37276   , -0.0032068 ,  0.18743999,  0.30702001], dtype=float32)
'''

embeddings_index.get('he') == embedding_matrix[76]
'''
array([ True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,  True], dtype=bool)
'''

2.6 LSTM训练

注意训练集data的shape是(N_SAMPLES, MAX_SEQUENCE_LENGT)，100是词向量长度，然后根据Embedding层会变成3D的Matrix

如果不清楚 Word Embedding 可以参考在Keras模型中使用预训练的词向量

因为 keras 版本的问题，运行原文的代码会出了一个错误，本文根据这里进行了更改。将：

embedding_layer = Embedding(nb_words + 1,
                            EMBEDDING_DIM,
                            weights=[embedding_matrix],
                            input_length=MAX_SEQUENCE_LENGTH,
                            trainable=False,
                            dropout=0.2)

中的 trainable=False 去掉，在后面加上 model.layers[1].trainable=False

embedding_layer = Embedding(nb_words + 1,
                            EMBEDDING_DIM,
                            weights=[embedding_matrix],
                            input_length=MAX_SEQUENCE_LENGTH,
                            dropout=0.2)

print('Build model...')
# sequence_input = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='int32')
# embedded_sequences = embedding_layer()
model = Sequential()
model.add(embedding_layer)
model.add(LSTM(100, dropout_W=0.2, dropout_U=0.2))  # try using a GRU instead, for fun
model.add(Dense(1))
model.add(Activation('sigmoid'))
model.add(Dense(len(labels_index), activation='softmax'))
model.layers[1].trainable=False

网络的模型是个样子的：

model.summary()

"""
____________________________________________________________________________________________________
Layer (type)                     Output Shape          Param #     Connected to                     
====================================================================================================
embedding_1 (Embedding)          (None, 1000, 100)     2000100     embedding_input_1[0][0]          
____________________________________________________________________________________________________
lstm_1 (LSTM)                    (None, 100)           80400       embedding_1[0][0]                
____________________________________________________________________________________________________
dense_1 (Dense)                  (None, 1)             101         lstm_1[0][0]                     
____________________________________________________________________________________________________
activation_1 (Activation)        (None, 1)             0           dense_1[0][0]                    
____________________________________________________________________________________________________
dense_2 (Dense)                  (None, 20)            40          activation_1[0][0]               
====================================================================================================
Total params: 2,080,641
Trainable params: 2,000,241
Non-trainable params: 80,400
____________________________________________________________________________________________________
"""

2.6 LSTM训练

注意训练集data的shape是(N_SAMPLES, MAX_SEQUENCE_LENGT)，100是词向量长度，然后根据Embedding层会变成3D的Matrix。

# try using different optimizers and different optimizer configs
model.compile(loss='categorical_crossentropy',
              optimizer='adam',
              metrics=['accuracy'])

print('Train...')
model.fit(x_train, y_train, batch_size=batch_size, nb_epoch=5,
          validation_data=(x_val, y_val))
score, acc = model.evaluate(x_val, y_val,
                            batch_size=batch_size)
print('Test score:', score)
print('Test accuracy:', acc)

"""
Train on 15998 samples, validate on 3999 samples
Epoch 1/5
  608/15998 [>.............................] - ETA: 833s - loss: 0.1992 - acc: 0.9500 
"""

后面我就懒得训练了,你们也看到了，渣渣电脑太慢了。

参考
[1] Understanding LSTM：http://colah.github.io/posts/2015-08-Understanding-LSTMs/
[2] 理解 LSTM 网络：https://www.yunaitong.cn/understanding-lstm-networks.html
[2] GloVe: Global Vectors for Word Representation：http://nlp.stanford.edu/projects/glove