【机器学习】——朴素贝叶斯 python实现

话不多说，上代码

【注意】p（x | y）在计算的时候会有零值，需要用贝叶斯估计来解决，这里采用拉普拉斯平滑，Sj为特征取值个数，我尝试了二值和八值，效果差不多，还有一点很重要的是，因为数值太小，相差的化最后都会变成零，所以用log函数来处理，最后的连乘就变成了累加，因为用的argmax只需要比较相对大小，所以最后的P（X）不用除，否则要计算的话得先将各个p[k]取指数再累加，此外，log函数也不用再变回去，因为它是单调递增的。

#图像设为二值，accuracy=84.27%，time=50.65s
#图像设为八值，accuracy=84.77%，time=52.44s


import numpy as np
import time

def loadData(filename):
    image_array=[]
    label_array=[]
    
    file=open(filename,'r')
    for line in file.readlines():
        curline=line.strip().split(',')
        image=[int(item)//32 for item in curline[1:]]
        image_array.append(image)
        label_array.append(int(curline[0]))
    return np.array(image_array),np.array(label_array)
    
def calculate_probability(train_image_array,train_label_array):
    feature_num=784
    class_num=10
    instance_num=train_image_array.shape[0]
    
    pixy=np.zeros((class_num,feature_num,8))
    py=np.zeros((class_num))
    
    for i in range(class_num):
        py[i]=(np.sum(train_label_array==i))
        
    for i in range(instance_num):
        image=train_image_array[i]
        #print('sum:',np.sum(image))
        label=train_label_array[i]
        for j in range(feature_num):
            pixy[label,j,image[j]]+=1
    
    #for i in range(class_num):
        #print(np.mat(pixy[i,200:250,:]).T)
    for i in range(class_num):
        pixy[i,:,:]=np.log((pixy[i,:,:]+1)/(py[i]+2))
        py[i]=np.log((py[i]+1)/(instance_num+10))
        #pixy[i,:,:]=pixy[i,:,:]-py[i]
    #class*feature*2
    #print('x',pixy[0,:,5])
    #py=np.sum(pixy,axis=(1,2))
    #pixy=pixy/py[:,np.newaxis,np.newaxis]
    #print('px',pixy)
    #py=py/instance_num
    #print('py',py)
    #print(pixy,py)
    return pixy,py

def naive_bayes(pixy,py,x):
    #P(Y=Yk/Xi1=j，Xi2=j)=P(Xij,Yk)/P(Xij)=P(Yk)*P(Xij/Yk)/sigama(k)P(Yk)*P(Xij/Yk)
    #上式要对所有的Xi值相乘
    class_num=10
    feature_num=784
    p=[0]*class_num
    
    for k in range(class_num):
        p[k]=py[k]
        for i in range(feature_num):
            p[k]=p[k]+pixy[k,i,x[i]]
    #print(p)
    return np.argmax(p)

def test(pixy,py,test_image_array,test_label_array):
    num=0
    for i in range(len(test_image_array)):
        image=test_image_array[i]
        pred=naive_bayes(pixy,py,image)
        label=test_label_array[i]
        if pred==label:
            num+=1
        #print('pred:',pred)
    accuracy=num/len(test_image_array)
    return accuracy


if __name__=="__main__":
    start=time.time()
    
    print('start to read data')
    train_image_array,train_label_array=loadData('../mnist/mnist_train.csv')
    test_image_array,test_label_array=loadData('../mnist/mnist_test.csv')
    
    print('start to test')
    pixy,py=calculate_probability(train_image_array,train_label_array)
    accuracy=test(pixy,py,test_image_array,test_label_array)
    
    end=time.time()
    print('accuracy:%f:'%accuracy)
    print('time spend: ',end-start)