基于DNN的semi-supervised learning

半监督学习：就是通过部分有label的数据，学习得到其余没有label数据的label。

本文用的方法是 low-density separation (低密度的边界分类)中 self-training 方法；low-density separation 是按照“非黑即白”的观念进行分类。在对已有label数据建立模型时，本文采用的是DNN。整体基本步骤如下：

repeat:

1）将已有label数据带入DNN构建模型f*

2）将没有label数据带入f*预测其结果u

3）将u中满足hard-label条件的data拿出来，再带入更新f*

直到所有的unlabelled数据全都打上label

注意：对预测结果要采用hard-label重新打label，例如预测得到结果为[0.7,0.3]要变成[1,0]，再重新带入模型。实验数据还是用的Iris，贴代码：

# -*- coding: utf-8 -*-
"""
Created on Mon Sep 25 10:22:49 2017

@author: wjw
"""
#基于DNN的 semi-supervised learning

import keras
import numpy as np
from keras.layers.core import Dense, Activation #直接import函数

def readText(filePath):
    lines = open(filePath,'r').readlines()
    data = []
    dataClass = []
    for line in lines:
        dataList = line.split(',')
        data.append([float(dataList[0]),float(dataList[1]),float(dataList[2]),float(dataList[3])])
        dataClass.append(dataList[4].split("\n")[0])
   
    new_class = []
    
    for name in dataClass:
        if name=="Iris-setosa":
            new_class.append(0)
        elif name=="Iris-versicolor":
            new_class.append(1)
        else:
            new_class.append(2)
            
    return np.array(data),np.array(new_class)

class network:
    def createDnn(self):
        model = keras.models.Sequential()#初始化一个神经网络
        model.add(Dense(input_dim=4,output_dim=10))#Dense表示fully connected 的神经网络
        model.add(Activation("sigmoid"))
        model.add(Dense(output_dim=10))
        model.add(Activation("sigmoid"))
        model.add(Dense(output_dim=3))
        model.add(Activation("softmax"))
        model.compile(loss="categorical_crossentropy",optimizer="adam",metrics=["accuracy"])
        #"categorical crossentropy"按照crossenrropy的方法定义损失函数
        #优化方法是"adam"
        return model
    
def low_density_seperation(classes):
    for class_line in classes:
        if max(class_line)>=0.7:
            max_index = np.argmax(class_line)
            class_line[max_index]=1
            class_line[list(filter(lambda x : x != max_index, range(class_line.size)))]=0
    return classes
    
def getTraindata():
    filePath = r"E:\学习\研究生\李宏毅\data\iris.txt"
    data, Class = readText(filePath)
    train_index = np.random.randint(0,data.shape[0],size =int(data.shape[0]*0.2))
    traindata = data[train_index]
    dataClass = Class[train_index]
    rest_index = list(filter(lambda x:x not in train_index,range(data.shape[0])))#filter过滤器，和map类似,将有标签的数据过滤掉
    restdata = data[rest_index]
    restClass = Class[rest_index]
    return traindata,dataClass,restdata,restClass

def train(model,traindata,dataClass,restdata):
    
    if len(restdata)==0:
        return 
    
    model.fit(traindata,dataClass,batch_size=15,epochs=1000)#traindata是输入数据，dataclass是输出结果，对应的都是神经元
    testClasses = model.predict(restdata)
    testClasses = low_density_seperation(testClasses)
    low_density_index = list(filter(lambda x : max(testClasses[x])==1,range(len(testClasses))))
    traindata = np.vstack((traindata , restdata[low_density_index]))  # np.vstack将两个array进行纵向合并
    dataClass = np.vstack((dataClass , testClasses[low_density_index])) #np.hstack 将两个array进行横向合并
    
    rest_index = list(filter(lambda x:x not in low_density_index,range(len(restdata))))
    if rest_index == '':
        return 
    restdata = restdata[rest_index]
    train(model,traindata,dataClass,restdata)
    
if __name__ == '__main__':
    traindata,dataClass,restdata,restClass = getTraindata()
    dataClass = keras.utils.to_categorical(dataClass)#装换成one-hot类型的数据比如说3就变成了[0,0,0,1]

    restClass = keras.utils.to_categorical(restClass)
    model = network().createDnn()
    train(model,traindata,dataClass,restdata)
    score = model.evaluate(np.vstack((traindata,restdata)),np.vstack((dataClass,restClass)),batch_size=20)

    print(score)

差不多要进行3次递归，得到的score为：

20/155 [==>...........................] - ETA: 2s[0.22960705333980513, 0.94838709792783182]