超级会员免费看
请勿直接复制代码
代码地址 https://github.com/W1412X/SDU_mechine_learning/tree/main/lab3
点击此处直接进入
两个类别
导入数据
import numpy as np
red_class=np.loadtxt('/home/wangxv/Files/hw/ml/lab3/data/ex3Data/ex3red.dat')
blue_class=np.loadtxt('/home/wangxv/Files/hw/ml/lab3/data/ex3Data/ex3blue.dat')
red_class
array([[2.95, 6.63],
[2.53, 7.79],
[3.57, 5.65],
[3.16, 5.47],
[2.78, 6.42],
[2.78, 7.42],
[3.18, 8.27],
[3.28, 6.42],
[3.68, 7.42],
[3.58, 8.16],
[2.78, 8.42],
[2.78, 6.42],
[2.78, 7.42],
[2.78, 8.42]])
计算平均向量
red_mean=np.array([red_class[:,0].mean(),red_class[:,1].mean()])
blue_mean=np.array([blue_class[:,0].mean(),blue_class[:,1].mean()])
计算red/blue点类协方差矩阵
s_red=np.dot((red_class-red_mean).T,(red_class-red_mean))
s_blue=np.dot((blue_class-blue_mean).T,(blue_class-blue_mean))
计算类内散度矩阵以及类间散度矩阵
s_w=np.array(s_blue+s_red)
s_b=np.dot(np.array([(red_mean-blue_mean)]).T,np.array([(red_mean-blue_mean)]))
print(s_w)
print(s_b)
[[ 4.80081429 -1.41724286]
[-1.41724286 29.07295714]]
[[1.54824694 3.28313673]
[3.28313673 6.96205918]]
得到方程
S w − 1 ⋅ S B ⋅ w = λ ⋅ w {S_w}^{-1} \cdot S_B \cdot w = \lambda \cdot w Sw−1⋅SB⋅w=λ⋅w
进一步
S w − 1 ⋅ S B = λ ⋅ I {S_w}^{-1} \cdot S_B = \lambda \cdot I Sw−1⋅SB=λ⋅I
m=np.dot(np.linalg.inv(s_w),s_b)
lamda=[0,m[0,0]+m[1,1]]
lamda
[0, 0.6378184048755453]
直接计算 w
w=np.dot(np.linalg.inv(s_w),np.array([red_mean-blue_mean]).T)
length = np.sqrt(w[0]**2 + w[1]**2)
w=w/length
w
array([[0.94042471],
[0.34000199]])
画图
import matplotlib.pyplot as plt
fig1=plt.figure(0)
plt.scatter(red_class[:,0],red_class[:,1],marker='o',color='red')
plt.scatter(blue_class[:,0],blue_class[:,1],marker='o',color='blue')
plt.axline((0,0),(w[0][0],w[1][0]), linestyle='-', color='black')
#映射点
red_projection_len=np.array(np.dot(red_class[:,0],w[0][0])+np.dot(red_class[:,1],w[1][0]))
blue_projection_len=np.array(np.dot(blue_class[:,0],w[0][0])+np.dot(blue_class[:,1],w[1][0]))
red_projection=np.array([red_projection_len*w[0][0],red_projection_len*w[1][0]])
blue_projection=np.array([np.multiply(blue_projection_len,w[0][0]),np.multiply(blue_projection_len,w[1][0])])
plt.scatter(red_projection[0],red_projection[1],marker='o',color='red')
plt.scatter(blue_projection[0],blue_projection[1],marker='o',color='blue')
<matplotlib.collections.PathCollection at 0x7f5696a26f70>
多个类别
导入数据
import numpy as np
red_class=np.loadtxt('/home/wangxv/Files/hw/ml/lab3/data/ex3Data/ex3red.dat')
green_class=np.loadtxt('/home/wangxv/Files/hw/ml/lab3/data/ex3Data/ex3green.dat')
blue_class=np.loadtxt('/home/wangxv/Files/hw/ml/lab3/data/ex3Data/ex3blue.dat')
red_class
array([[2.95, 6.63],
[2.53, 7.79],
[3.57, 5.65],
[3.16, 5.47],
[2.78, 6.42],
[2.78, 7.42],
[3.18, 8.27],
[3.28, 6.42],
[3.68, 7.42],
[3.58, 8.16],
[2.78, 8.42],
[2.78, 6.42],
[2.78, 7.42],
[2.78, 8.42]])
计算平均向量
red_mean=np.array([red_class[:,0].mean(),red_class[:,1].mean()])
blue_mean=np.array([blue_class[:,0].mean(),blue_class[:,1].mean()])
green_mean=np.array([green_class[:,0].mean(),green_class[:,1].mean()])
all_mean=(red_mean+blue_mean+green_mean)/3.0
计算协方差
s_red=np.cov(red_class.T)
s_blue=np.cov(blue_class.T)
s_green=np.cov(green_class.T)
计算类内散度矩阵(这里类内散度矩阵被重新定义为每个类别的散度矩阵之和)
s_w=s_red+s_green+s_blue
s_w
array([[ 0.83919725, -0.18908407],
[-0.18908407, 2.66521923]])
计算类间散度矩阵
m=red_class.shape[0]
sb_red=m*np.dot(np.array([red_mean-all_mean]).T,np.array([red_mean-all_mean]))
sb_blue=m*np.dot(np.array([blue_mean-all_mean]).T,np.array([blue_mean-all_mean]))
sb_green=m*np.dot(np.array([green_mean-all_mean]).T,np.array([green_mean-all_mean]))
s_b=sb_red+sb_blue+sb_green
s_b
array([[ 78.55411429, 84.16194286],
[ 84.16194286, 104.00893333]])
S w − 1 ⋅ S B {S_w}^{-1}\cdot S_B Sw−1⋅SB
inv_sw_sb=np.dot(np.linalg.inv(s_w),s_b)
inv_sw_sb
array([[102.35742211, 110.85345158],
[ 38.83961935, 46.88903383]])
计算特征向量
values, vectors = np.linalg.eig(inv_sw_sb)
vectors
w1=vectors[:,0]
w2=vectors[:,1]
print(w1)
print(w2)
[0.93088546 0.36531119]
[-0.74595516 0.66599617]
画图
import matplotlib.pyplot as plt
fig, ax = plt.subplots()
ax.scatter(red_class[:,0],red_class[:,1],marker='o',color='red')
ax.scatter(blue_class[:,0],blue_class[:,1],marker='o',color='blue')
ax.scatter(green_class[:,0],green_class[:,1],marker='o',color='green')
k1=w1[1]/w1[0]
k2=w2[1]/w2[0]
ax.axline((0,0),(1,k1),linestyle='-', color='black')
ax.axline((0,0),(1,k2),linestyle='-', color='black')
ax.set_xlim(-9,9)
ax.set_ylim(-10,20)
###
#映射点
red_projection_len=np.array(np.dot(red_class[:,0],w1[0])+np.dot(red_class[:,1],w1[1]))
blue_projection_len=np.array(np.dot(blue_class[:,0],w1[0])+np.dot(blue_class[:,1],w1[1]))
green_projection_len=np.array(np.dot(green_class[:,0],w1[0])+np.dot(green_class[:,1],w1[1]))
red_projection=np.array([red_projection_len*w1[0],red_projection_len*w1[1]])
blue_projection=np.array([np.multiply(blue_projection_len,w1[0]),np.multiply(blue_projection_len,w1[1])])
green_projection=np.array([green_projection_len*w1[0],green_projection_len*w1[1]])
plt.scatter(red_projection[0],red_projection[1],marker='o',color='red')
plt.scatter(blue_projection[0],blue_projection[1],marker='o',color='blue')
plt.scatter(green_projection[0],green_projection[1],marker='o',color='green')
red_projection_len1=np.abs(np.array(np.dot(red_class[:,0],w2[0])+np.dot(red_class[:,1],w2[1])))
blue_projection_len1=np.abs(np.array(np.dot(blue_class[:,0],w2[0])+np.dot(blue_class[:,1],w2[1])))
green_projection_len1=np.abs(np.array(np.dot(green_class[:,0],w2[0])+np.dot(green_class[:,1],w2[1])))
red_projection1=np.array([red_projection_len1*w2[0],red_projection_len1*w2[1]])
blue_projection1=np.array([np.multiply(blue_projection_len1,w2[0]),np.multiply(blue_projection_len1,w2[1])])
green_projection1=np.array([green_projection_len1*w2[0],green_projection_len1*w2[1]])
plt.scatter(red_projection1[0],red_projection1[1],marker='o',color='red')
plt.scatter(blue_projection1[0],blue_projection1[1],marker='o',color='blue')
plt.scatter(green_projection1[0],green_projection1[1],marker='o',color='green')
fig.show()
/tmp/ipykernel_4571/3613756043.py:35: UserWarning: FigureCanvasAgg is non-interactive, and thus cannot be shown
fig.show()
可以看到使用第一个特征向量的分类效果明显好于第二个
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
