Question

我是机器学习的新手。我在大学修过ML课程，作为一项家庭作业，我有使用朴素贝叶斯分类器对cifar10进行分类的任务。我已经尝试了两天才能完成这项作业，但是总有些错误。老师告诉我们，分类准确度应该在60-70％之间。我的分类精度为20％。作为特征，我们具有每个通道的平均值（红色平均值，绿色平均值，蓝色平均值）。因此，我们拥有大小为50000x3的训练数据矩阵。我按类别对矩阵的条目进行分组，并为每个矩阵找到了mu和sigma。这是代码：

import numpy as np
import math
from ex2_class import DataMatrix
import matplotlib.pyplot as plt
from keras.datasets import cifar10

def cifar_10_evaluate(pred, gt):
    correct = 0
    for i in range(len(pred)):
        if pred[i] == gt[i]:
            correct +=1
    return (correct/len(pred))*100

def normpdf(x,mu,var):
    nom = math.exp(-((x-mu)**2)/(2*var))
    denom = math.sqrt(2*math.pi*var)
    return nom/denom

def cifar_10_features(x):
    r = x[0:1024]
    g = x[1024:2048]
    b = x[2048:3072]
    return [sum(r)/1024, sum(g)/1024, sum(b)/1024]

def cifar_10_bayes_learn(F,label):
    ms = []
    sg = []
    for i in range(0,10):
        idx = np.where(label==i) #get from label-vector ith entries indecies
        class_i = F[idx] #get data according to gotten indecies
        #calculate mean of each component
        mur = np.mean(class_i[:,0])
        mug = np.mean(class_i[:,1])
        mub = np.mean(class_i[:,2])
        ms.append([mur,mug,mub])

        #calculate sigma squared of each component
        sg_r = np.var(class_i[:,0])
        sg_g = np.var(class_i[:,1])
        sg_b = np.var(class_i[:,2])
        sg.append([sg_r,sg_g,sg_b])
    p = len(class_i)/len(F)
    return [np.array(ms),np.array(sg),1/10]

def cifar_10_classify(f,ms,gs,p):
    test = cifar_10_features(f)
    prob = {}
    for i in range(0,10):
        pr = normpdf(test[0],ms[i,0],gs[i,0])
        pg = normpdf(test[1],ms[i,1],gs[i,1])
        pb = normpdf(test[2],ms[i,2],gs[i,2])
        prob[i] = pr*pg*pb*p
    mx = max(prob.values())
    cl = [c for c,nd in prob.items() if mx == nd]
    return cl[0]

def save(mus, sigmas):
    np.save('class_mus',mus)
    np.save('class_variances',sigmas)

def main():
    mydata = DataMatrix(path_to_data)
    T = mydata.get_training_data() #training data
    L = mydata.get_training_labels()
    C = mydata.get_test_data()
    CL = mydata.get_test_labels()
    C = C[0:10000,:]
    CL = CL[0:10000]

    tr_set = np.load('training_mus.npy')
    [ms, gs, p] = cifar_10_bayes_learn(tr_set, L)

    prob = []
    for i in range(len(C)):
        prob.append(cifar_10_classify(C[i],ms,gs,p))
    print(cifar_10_evaluate(prob,CL))

请帮助我解决问题，因为两天后我仍然没有发现我的代码有什么问题。谢谢

使用朴素贝叶斯分类器进行cifar10分类

0 个答案: