误分类点:
设一个点
x
⃗
\vec x
x 被误分类了
n
n
n 次,则由于
∇
L
=
<
x
⃗
y
,
y
>
\nabla L=<\vec xy,y>
∇L=<xy,y>,更新了
n
n
n 次后的
w
⃗
\vec w
w 为:
w
⃗
=
α
n
x
⃗
y
\begin{aligned}\vec w&=\alpha n\vec xy\end{aligned}
w=αnxy
对所有的点来说:
w
⃗
=
∑
α
n
i
x
⃗
i
y
i
\begin{aligned}\vec w&=\sum\alpha n_i\vec x_iy_i\end{aligned}
w=∑αnixiyi
所以当 ( ∑ α n i x ⃗ i y i x ⃗ j + b ) y j < 0 (\sum\alpha n_i\vec x_iy_i\vec x_j+b)y_j<0 (∑αnixiyixj+b)yj<0 时, x ⃗ j \vec x_j xj 就是误分类点。
为了简化计算,我们可以将 x ⃗ i x ⃗ j \vec x_i\vec x_j xixj 预先取出,形成一个矩阵,这样可以免去重复的计算。
更新权重项与偏置项:
由于
n
n
n 代表一个点被误分类的次数,所以每次找到误分类点后,只需要
n
+
1
n+1
n+1 即相当于更新了权重项。或者简化一下,设
a
=
α
n
a=\alpha n
a=αn 则每次更新时只需要:
a
+
=
α
b
+
=
α
y
i
\begin{aligned}a&+=\alpha\\b&+=\alpha y_i\end{aligned}
ab+=α+=αyi
模板:
import numpy as np
X_train = np.array([[1, 1], [1, 0], [0, 1], [0, 0]])
Y_train = np.array([1, 1, 1, -1])
class perceptron:
def __init__(self, X, Y, Alpha = 0.01):
self.alpha = Alpha
self.X_trn = np.array(X)
self.Y_trn = np.array(Y)
self.B = np.random.rand()
self.lenth = len(self.X_trn)
self.a = np.zeros(self.lenth)
self.mat = np.zeros((self.lenth, self.lenth))
for i in range(self.lenth):
for j in range(self.lenth):
self.mat[i, j] = np.dot(self.X_trn[i], self.X_trn[j])
def check(self, j):
Sum = np.dot(self.a * self.Y_trn, self.mat[j]) + self.B
if (Sum * self.Y_trn[j] >= 0):
return True
else:
return False
def Loss_func(self, j):
self.a[j] += self.alpha
self.B += self.alpha * self.Y_trn[j]
def train(self):
Error = True
k = 0
while (Error):
Error = False
for i in range(self.lenth):
if (self.check(i) == False):
self.Loss_func(i)
Error = True
print(self.a)
print("W: ", np.dot(self.a * self.Y_trn, self.X_trn), " B: ", self.B)
per = perceptron(X_train, Y_train, 0.01)
per.train()
同步更新于:SP-FA 的博客