numpy实现线性可分支持向量机
import torch
import torch.nn as nn
import numpy as np
import pandas as pd
a=np.zeros(128)
b=0
def take(i1,i2):
global b
if i1==i2:
return 0
w=0
for i in range(128):
w=w+a[i]*train[i]*train_label[i]
e1=np.dot(w,train[i1])+b-train_label[i1]
e2=np.dot(w,train[i2])+b-train_label[i2]
eta=np.dot(train[i1],train[i1])+np.dot(train[i2],train[i2])-2*np.dot(train[i1],train[i2])
alpha2=a[i2]+train_label[i2]*(e1-e2)/eta
L=0
H=0
old2=a[i2]
if train_label[i1]!=train_label[i2]:
L=np.maximum(0,a[i2]-a[i1])
if alpha2<L:
a[i2]=L
else:
a[i2]=alpha2
else:
L=0
H=np.maximum(0,a[i2]+a[i1])
if alpha2<L:
a[i2]=L
elif alpha2>H:
a[i2]=H
else:
a[i2]=alpha2
a[i1]=a[i1]+train_label[i1]*train_label[i2]*(old2-a[i2])
b1=train_label[i1]-np.dot(w,train[i1])
b2=train_label[i2]-np.dot(w,train[i2])
if a[i1]>0:
b=train_label[i1]-np.dot(w,train[i1])
elif a[i2]>0:
b=train_label[i2]-np.dot(w,train[i2])
else:
b=(b1+b2)/2
if abs(old2-a[i2])<=0.0001:
return 0
return 1
def examineExample(i2):
for suo in range(128):
if(a[suo]>0):
if (take(suo,i2)):
return 1
for suo in range(128):
if(a[suo]==0):
if (take(suo,i2)):
return 1
return 0
train=np.random.randint(-300,300,(128,4))
test=np.random.randint(-300,300,(30,4))
train=train.astype(float)
test=test.astype(float)
for i in range(4):
train[:,i]=(train[:,i]-train[:,i].mean())/train[:,i].std()
for i in range(4):
test[:,i]=(test[:,i]-test[:,i].mean())/test[:,i].std()
print(train)
w=[1,2,3,4]
train_label=np.dot(train,w)+5
test_label=np.dot(test,w)+5
train_label=train_label>0
train_label=train_label.astype(int)
train_label=train_label*2-1
test_label=test_label>0
test_label=test_label.astype(int)
test_label=test_label*2-1
numChanged=0
examineAll=1
while(numChanged>0 or examineAll):
numChanged=0
if examineAll:
for i in range(128):
numChanged=numChanged+examineExample(i)
else:
for i in range(128):
if(a[i]>0):
numChanged=numChanged+examineExample(i)
if examineAll==1:
examineAll=0
elif numChanged==0:
examineAll=1
w=0
for i in range(128):
w=w+a[i]*train[i]*train_label[i]
m=0
max=0
li=0
for i in range(128):
if i==0:
max=a[i]
li=0
if(a[i]!=0):
b=train_label[i]-np.dot(w,train[i])
print(b)
if i==0:
li=0
max=a[i]
elif max<np.float(a[i]):
max=a[i]
li=i
b=train_label[li]-np.dot(w,train[li])
print(b)
sum1=0
loss=0
for i in range(128):
targer=0
if np.dot(w,train[i])+b>0:
targer=1
elif np.dot(w,train[i])+b<0:
targer=-1
if targer!=train_label[i]:
sum1=sum1+1
print("训练出错数",sum1,"训练误差率：",sum1/128)
for i in range(30):
targer=0
if np.dot(w,train[i])+b>0:
targer=1
elif np.dot(w,train[i])+b<0:
targer=-1
if targer!=train_label[i]:
loss=loss+1
print("测试出错数",loss,"测试误差率：",loss/30)
print(a)
print(w,b)
