【丢弃】【NLP1】词袋

['others','TensorFlow','makes','machine','learning','easy']

sentence1='TensorFlow makes machine learning easy'
sentence1_list=[0,1,1,1,1,1]

sentence2='Machine learning is easy'
sentence2=[1,0,0,1,1,1]
# 这里，'is'没有出现在单词向量中，为0号

import pandas as pd
df=pd.read_csv('http://www.guofei.site/datasets_for_ml/SMSSpamCollection/SMSSpamCollection.csv',sep='\t',header=None,names=['label','sentences'])
Y=(df.label=='spam')*1

import re
regex=re.compile('[a-zA-Z]+')
X1=[regex.findall(sentence) for sentence in df.sentences] # 只挑出单词，去除标点符号和数字
X2=[[i.lower() for i in j if len(i)>1] for j in X1] # 改成小写，并剔除单个字母
X=[' '.join(i) for i in X2] # 重新整合成句子


from sklearn.model_selection import train_test_split
X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.2)

import tensorflow as tf
from tensorflow.contrib import learn
vocab_processor=learn.preprocessing.VocabularyProcessor(max_document_length=25,min_frequency=3) # max_document_length 是每个句子最大的长度，短的补空值，长的截取
vocab_processor.fit(X)


embedding_size=len(vocab_processor.vocabulary_) # 词典向量的长度

# 使用案例：
vocab_processor.transform(['hello world','good morning']) # 返回一个generator，每个next对应入参中的一项（入参也必须是<iterable>）
# 每项是一个与入参对应项大小相同的向量，向量的值是每个word在vocab_processor.vocabulary_中的序号
# 例如，案例返回[237 327   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0]
# 前两个数字分别是'hello', 'world' 在 vocab_processor.vocabulary_中的位置，后面的0是用'其它'补到25

# vocab_processor.fit_transfrom(X)

import tensorflow as tf
sess = tf.Session()
identity_mat = tf.diag(tf.ones(shape=[embedding_size]))

A = tf.Variable(tf.random_normal(shape=(embedding_size, 1)))
b = tf.Variable(tf.random_normal(shape=[1, 1]))
x = tf.placeholder(shape=[25], dtype=tf.int32)  # 序号组成的向量，将来用vocab_processor.transform()来feed
y = tf.placeholder(shape=[1, 1], dtype=tf.float32)

x_embed = tf.nn.embedding_lookup(identity_mat, x)
# 用法解释：
# sess.run(x_embed,feed_dict={x:next(vocab_processor.transform(['hello world']))})
# 一个shape=(max_document_length,vocab_processor.vocabulary_)的向量，第i行是第i个单词的OneHot编码

x_col_sum=tf.reduce_sum(x_embed,axis=0) # 上一步是每个单词的OneHot编码，现在加总成句子的OneHot编码
x_col_sum_2D=tf.expand_dims(x_col_sum,axis=0)

# 下面做softmax
a=tf.matmul(x_col_sums_2D,A)+b
y_hat=tf.nn.sigmoid(a)
cross_entropy=tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=a, labels=y))
# 等价写法：
# y_hat=tf.nn.sigmoid(tf.matmul(x_col_2D,A)+b)
# cross_entropy = tf.reduce_mean(-tf.reduce_sum(y * tf.log(y_hat), axis=1))

train_step = tf.train.GradientDescentOptimizer(learning_rate=0.001).minimize(cross_entropy)

sess.run(tf.global_variables_initializer())
for i,j in enumerate(vocab_processor.transform(X)):
    sess.run(train_step,feed_dict={x:j,y:[[Y[i]]]})


# 预测阶段
Y_predict=[]
for i,j in enumerate(vocab_processor.transform(X)):
    if i % 100 == 0: print(i)
    Y_predict.append(sess.run(y_hat,feed_dict={x:j})[0][0])


# 评价阶段
from sklearn.metrics import roc_auc_score
roc_auc_score(Y,Y_predict)

import pandas as pd
df=pd.read_csv('SMSSpamCollection.csv',sep='\t',header=None,names=['label','sentences'])
Y=(df.label=='spam')*1

import re
regex=re.compile('[a-zA-Z]+')
X1=[regex.findall(sentence) for sentence in df.sentences] # 只挑出单词，去除标点符号和数字
X2=[[i.lower() for i in j if len(i)>1] for j in X1] # 改成小写，并剔除单个字母
X=[' '.join(i) for i in X2] # 重新整合成句子


from tensorflow.contrib import learn
vocab_processor=learn.preprocessing.VocabularyProcessor(max_document_length=25,min_frequency=3) # max_document_length 是每个句子最大的长度，短的补空值，长的截取
vocab_processor.fit(X)


embedding_size=len(vocab_processor.vocabulary_) # 词典向量的长度



import tensorflow as tf
sess = tf.Session()
identity_mat = tf.diag(tf.ones(shape=[embedding_size]))

A = tf.Variable(tf.random_normal(shape=(embedding_size, 1)))
b = tf.Variable(tf.random_normal(shape=[1, 1]))
x = tf.placeholder(shape=[25], dtype=tf.int32)  # 序号组成的向量，将来用vocab_processor.transform()来feed
y = tf.placeholder(shape=[1, 1], dtype=tf.float32)

x_embed = tf.nn.embedding_lookup(identity_mat, x)
x_col_sum=tf.reduce_sum(x_embed,axis=0) # 上一步是每个单词的OneHot编码，现在加总成句子的OneHot编码
x_col_sum_2D=tf.expand_dims(x_col_sum,axis=0)


f_X=open('X_embed.csv','w')
for i in vocab_processor.transform(X):
    x_col_sums_2D_values=sess.run(x_col_sum_2D,feed_dict={x:i})[0]
    f_X.writelines(','.join([str(j) for j in x_col_sums_2D_values])+'\n')
f_X.close()

f_Y=open('Y.csv','w')
f_Y.writelines(','.join([str(i) for i in Y]))
f_Y.close()

import pandas as pd
X_embed=pd.read_csv('X_embed.csv',header=None).values
Y=pd.read_csv('Y.csv',header=None).values[0].reshape(-1,1)

num_sample,num_features=X_embed.shape


# 定义变量
import tensorflow as tf

x_embed=tf.placeholder(shape=(None,num_features),dtype=tf.float32)
y=tf.placeholder(shape=(None,1),dtype=tf.float32)
A=tf.Variable(tf.truncated_normal(shape=(num_features,1),dtype=tf.float32))
b=tf.Variable(tf.truncated_normal(shape=(1,1),dtype=tf.float32))
a=tf.matmul(x_embed,A)+b
y_hat=tf.nn.sigmoid(a)
cross_entropy=tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=a, labels=y))
train_step=tf.train.GradientDescentOptimizer(learning_rate=0.001).minimize(cross_entropy)

sess=tf.Session()
sess.run(tf.global_variables_initializer())


a_list=[]
import numpy as np
batch_size=30
for i in range(200):
    choice_index=np.random.choice(num_sample,size=batch_size)
    sess.run(train_step,feed_dict={x_embed:X_embed[choice_index],y:Y[choice_index]})