logRegIris.py

# -*- coding: utf-8 -*-
"""
Machine Learning lab1
comparison with the Iris dataset
chose two classes from all three classes for computaion
used 100 samples in total
each two validation method has 100 validation data sample
@author: 10152510119 徐紫琦
"""

from numpy import *
import scipy
import matplotlib.pyplot as plt

filename='pima-indians-diabetes.data' #file directory

def loadFile():   #read the .data file of the dataset
    fr = open(filename)
    linestr = fr.readlines()
    return linestr

def appendData(linestr,dataMat,labelMat,i):#append attributes into the data matrix and append class label into the label matrix
    lineArr = linestr[i].strip().split(",")
    w2=float(lineArr[0])
    w3=float(lineArr[1])
    w4=float(lineArr[2])
    w5=float(lineArr[3])
    dataMat.append([1.0, w2, w2,w4,w5])   #4 attributes given, w1=1
    if lineArr[4]=="Iris-setosa":
        labelMat.append([0])
    else:
        labelMat.append([1])
    
def loadDataSet10(linestr,j):   #turn raw data into matrixes for the j-th computation in 10-fold cross-validation
    dataMat = []
    labelMat = []
    length=len(linestr)
    for i in range(0,10*j): #100 data samples->leave 10 samples out and append the rest samples
       appendData(linestr,dataMat,labelMat,i)
    if(j<9):    
        for i in range(10*(j+1),100):
            appendData(linestr,dataMat,labelMat,i)
    return dataMat,labelMat


def loadDataSet1(linestr,j):   #turn raw data into matrixes for the j-th computation in leave-one-out cross-validation
    dataMat = []
    labelMat = []
    length=len(linestr)
    for i in range(0,j-1):   #100 data samples->leave 1 sample out and append the rest samples
        appendData(linestr,dataMat,labelMat,i)
    if(j<99):    
        for i in range(j+1,100):
            appendData(linestr,dataMat,labelMat,i)
    return dataMat,labelMat

def sigmoid(inX):  #sigmoid function: special exp function to prevent overflow
    return 1.0/(1+scipy.special.expit(-inX))

def gradAscent(dataMat, labelMat):
    dataMatrix=mat(dataMat)
    alpha = 0.001  #learning rate
    maxCycles = 500 #the number of circles of iteration
    weights = ones((5,1)) #initialize the parameters
    for k in range(maxCycles):
        h = sigmoid(dataMatrix*weights)
        error = (labelMat - h)     
        weights = weights + alpha * dataMatrix.transpose()* error #refresh the parameters
    return weights

def countErr10(weights,linestr,j):  #count the numver of errors in the j-th computation in 10-fold cross-validation
    length=len(linestr)
    dataMat = []
    labelMat = []
    for i in range(10*(j),10*(j+1)):#construct the validation data matrix
        appendData(linestr,dataMat,labelMat,i)
    y=sigmoid(mat(dataMat)*weights)
    err=0
    for i in range(0,10):  #compare the class labels with the pridiction values to count errors
        if (round(y[(i,0)])!=labelMat[i][0]) or (y[(i,0)]==0.5 and labelMat[i][0]==0):  #to solve the problem that round(0.5,0)==0.0
            err+=1
    return err


def countErr1(weights,linestr,j):  #judge if the validation data is an error
    length=len(linestr)
    dataMat = []
    labelMat = []
    appendData(linestr,dataMat,labelMat,j)
    y=sigmoid(mat(dataMat)*weights)
    err=0
    if (round(y[(0,0)])!=labelMat[0][0]) or (y[(0,0)]==0.5 and labelMat[0][0]==0):  
        err+=1
    return err

def printFigure(err10,err1):
    n_groups = 2
    list=[err10,err1]
    fig, ax = plt.subplots()  
    index = np.arange(n_groups)  
    bar_width = 0.35     
    opacity = 0.4  
    rects1 = plt.bar(index, list, bar_width,alpha=opacity, color='b')  
    plt.ylabel('error rate')  
    plt.title('Error rate comparison between 10-fold cross-validation\n and leave-one-out cross-validation')  
    plt.xticks(index, ('10-fold', 'leave-one-out'))  
    plt.ylim(0,1)  
    plt.legend()  
    plt.tight_layout()  
    plt.show()

def main():
    totalErr10=0
    totalErr1=0
    linestr=loadFile()
    for i in range(0,10):  #run 10 times 10-fold cross-validation
        dataMat, labelMat = loadDataSet10(linestr,i)
        weights=gradAscent(dataMat, labelMat).getA()
        totalErr10+=countErr10(weights,loadFile(),i)
        
    for i in range(0,100):  #run the leave-one-out vross-validation 100 times
        dataMat, labelMat = loadDataSet1(linestr,i)
        weights=gradAscent(dataMat, labelMat).getA()
        totalErr1+=countErr1(weights,loadFile(),i)
    printFigure(totalErr10/100,totalErr1/100)

if __name__=='__main__':
    main()