SMO的Python代码实现（仅供参考）

一起学ML之支持向量机 - SMO算法的Python实现 - 代码

 

 

import numpy as np
import matplotlib.pyplot as plt

from sklearn.datasets import make_blobs,make_circles,make_moons
from sklearn.preprocessing import StandardScaler

class SMOStruct:
    """ 按照John Platt的论文构造SMO的数据结构"""
    def __init__(self, X, y, C, kernel, alphas, b, errors, user_linear_optim):
        self.X = X              # 训练样本
        self.y = y              # 类别 label
        self.C = C              # regularization parameter  正则化常量，用于调整（过）拟合的程度
        self.kernel = kernel    # kernel function   核函数，实现了两个核函数，线性和高斯（RBF）
        self.alphas = alphas    # lagrange multiplier 拉格朗日乘子，与样本一一相对
        self.b = b              # scalar bias term 标量，偏移量
        self.errors = errors    # error cache  用于存储alpha值实际与预测值得差值，与样本数量一一相对
        
        self.m, self.n = np.shape(self.X)    # store size(m) of training set and the number of features(n) for each example  
                                             #训练样本的个数和每个样本的features数量

        self.user_linear_optim = user_linear_optim    # 判断模型是否使用线性核函数
        self.w = np.zeros(self.n)     # 初始化权重w的值，主要用于线性核函数
        #self.b = 0               


   

def linear_kernel(x,y,b=1):
    #线性核函数
    """ returns the linear combination of arrays 'x' and 'y' with
    the optional bias term 'b' (set to 1 by default). """
    result = x @ y.T + b
    return result # Note the @ operator for matrix multiplications


def gaussian_kernel(x,y, sigma=1):
    #高斯核函数
    """Returns the gaussian similarity of arrays 'x' and 'y' with
    kernel width paramenter 'sigma' (set to 1 by default)"""

    if np.ndim(x) == 1 and np.ndim(y) == 1:
        result = np.exp(-(np.linalg.norm(x-y,2))**2/(2*sigma**2))
    elif(np.ndim(x)>1 and np.ndim(y) == 1) or (np.ndim(x) == 1 and np.ndim(y)>1):
        result = np.exp(-(np.linalg.norm(x-y, 2, axis=1)**2)/(2*sigma**2))
    elif np.ndim(x) > 1 and np.ndim(y) > 1 :
        result = np.exp(-(np.linalg.norm(x[:, np.newaxis]- y[np.newaxis, :], 2, axis = 2) ** 2)/(2*sigma**2))
    return result


# 判别函数一，用于单一样本
def decision_function_output(model,i):
    if model.user_linear_optim:
        #Equation (J1)
        #return float(np.dot(model.w.T, model.X[i])) - model.b
        return float(model.w.T @ model.X[i]) - model.b
    else:
        #Equation (J10)
        return np.sum([model.alphas[j] * model.y[j] * model.kernel(model.X[j], model.X[i]) for j in range(model.m)]) - model.b


# 判别函数二，用于多个样本
def decision_function(