栏目分类:
子分类:
返回
名师互学网用户登录
快速导航关闭
当前搜索
当前分类
子分类
实用工具
热门搜索
名师互学网 > IT > 软件开发 > 后端开发 > Python

梯度下降法

Python 更新时间: 发布时间: IT归档 最新发布 模块sitemap 名妆网 法律咨询 聚返吧 英语巴士网 伯小乐 网商动力

梯度下降法

原理参考:梯度-牛顿-拟牛顿优化算法和实现

python实现:

import sympy
import numpy as np

#f为要求极值的函数,x0为初始位置,max_iter为最大迭代次数,epsilon为相邻两次迭代的x改变量
def gradient_x(f, x0, max_iter, epsilon):
    i = 0 #记录迭代次数的变量
    x0 = float(x0) #浮点数计算更快
    df = sympy.diff(f, x) #定义一阶导数
    #alpha = 0.001
    beta = 0.5 #beta 0~1
    delta = 0.25 #delta 0~0.5
    while i < max_iter:
        gk = df.subs(x, x0)
        dk = -gk
        #xnew = x0 + alpha*dk
        mk = 0
        while mk < 10:
            if f.subs(x, x0+beta**mk*dk) < f.subs(x,x0) + delta*beta**mk*gk*dk:
                break
            mk += 1
        xnew = x0 + beta**mk*dk
        i += 1
        print('迭代第%d次:%.5f' %(i, xnew))
        if abs(df.subs(x, xnew)-df.subs(x, x0)) < epsilon:
            break
        x0 = xnew
    return xnew

#f为要求极值的函数,X0为初始位置,max_iter为最大迭代次数,epsilon为相邻两次迭代的x改变量
def gradient_x0x1(f, X0, max_iter, epsilon):
    i = 0 #记录迭代次数的变量
    X0[0], X0[1] = float(X0[0]), float(X0[1]) #浮点数计算更快
    df0 = sympy.diff(f, x0) #定义一阶导数
    df1 = sympy.diff(f, x1)
    #alpha = 0.001
    beta = 0.5 #beta 0~1
    delta = 0.25 #delta 0~0.5
    while i < max_iter:
        gk = np.mat([(float)(df0.subs([(x0, X0[0]), (x1, X0[1])])), float(df1.subs([(x0, X0[0]), (x1, X0[1])]))]) #梯度矩阵
        dk = -gk      
        #Xnew = [X0[0] + alpha*dk[0,0], X0[1] + alpha*dk[0,1]]
        mk = 0
        while mk < 10:
            if f.subs([(x0, X0[0]+beta**mk*dk[0,0]), (x1, X0[1]+beta**mk*dk[0,1])]) < f.subs([(x0, X0[0]), (x1, X0[1])]) + (delta*beta**mk*gk*dk.T)[0,0]:
                break
            mk += 1
        Xnew = [X0[0] + beta**mk*dk[0,0], X0[1] + beta**mk*dk[0,1]]
        i += 1
        print('迭代第%d次:[%.5f, %.5f]' %(i, Xnew[0], Xnew[1]))      
        if abs(f.subs([(x0, Xnew[0]), (x1, Xnew[1])])-f.subs([(x0, X0[0]), (x1, X0[1])])) < epsilon:
            break
        X0 = Xnew
    return Xnew


if __name__ == '__main__':      
    x = sympy.symbols("x") 
    x0 = sympy.symbols("x0")
    x1 = sympy.symbols("x1")
    result = gradient_x(x**4-4*x, 10, 50, 1e-5)
    print('最佳迭代的位置:%.5f' %result)
    result = gradient_x0x1((x0-1)**2+(x1-1)**4, [10,10], 50, 1e-5)
    print('最佳迭代位置:[%.5f, %.5f]' %(result[0], result[1]))

C++实现:

#include 
#include 
#include 

const float dx = 1e-3;

float f(float x)
{
	return pow(x, 4) - 4 * x;
}

float df(float x)
{
	//return 4 * pow(x, 3) - 4;
	return (f(x + dx) - f(x)) / dx;
}

float f(std::vector X)
{
	return pow(X[0] - 1, 2) + pow(X[1] - 1, 4);
}

float df0(std::vector X)
{
	//return 2 * (X[0] - 1);
	return (f({ X[0] + dx,X[1] }) - f(X)) / dx;
}

float df1(std::vector X)
{
	//return 4 * pow(X[1] - 1, 3);
	return (f({ X[0], X[1] + dx }) - f(X)) / dx;
}

float gradient_x(float x0, int max_iter, float epsilon)
{
	int i = 0;
	float alpha = 0.001;
	float beta = 0.5;
	float delta = 0.25;
	float xnew;
	while (i < max_iter)
	{
		float gk = df(x0);
		float dk = -gk;
		//xnew = x0 + alpha*dk;
		int mk = 0;
		while (mk < 10)
		{
			if (f(x0 + pow(beta, mk)*dk) < f(x0) + delta*pow(beta, mk)*gk*dk)
				break;
			++mk;
		}
		xnew = x0 + pow(beta, mk)*dk;

		++i;
		std::cout << "迭代次数:" << i << " " << x0 << std::endl;
		if (abs(f(xnew) - f(x0)) < epsilon)
			break;
		x0 = xnew;
	}
	return xnew;
}

std::vector gradient_x0x1(std::vector X0, int max_iter, float epsilon)
{
	int i = 0;
	float alpha = 0.001;
	float beta = 0.5;
	float delta = 0.25;
	std::vector Xnew;
	while (i < max_iter)
	{
		Eigen::Vector2f gk;
		gk << df0(X0), df1(X0);
		
		Eigen::Vector2f dk = -gk;
		//Xnew = { X0[0] + alpha*dk(0), X0[1] + alpha*dk(1) };
		int mk = 0;
		while (mk < 10)
		{
			Xnew = { X0[0] + pow(beta, mk)*dk(0), X0[1] + pow(beta, mk)*dk(1) };
			if (f(Xnew) < f(X0) + delta*pow(beta, mk)*gk.transpose()*dk)
				break;
			++mk;
		}
		Xnew = { X0[0] + pow(beta, mk)*dk(0), X0[1] + pow(beta, mk)*dk(1) };

		++i;
		std::cout << "迭代次数:" << i << " " << X0[0] << " " << X0[1] << std::endl;
		if (abs(f(Xnew) - f(X0)) < epsilon)
			break;
		X0 = Xnew;
	}
	return X0;
}


int main(int argc, char* argv[])
{
	float result = gradient_x(10, 50000, 1e-5);
	std::cout << "最佳迭代位置:" << result << std::endl;

	std::vector results = gradient_x0x1({ 10,10 }, 50000, 1e-5);
	std::cout << "最佳迭代位置:" << results[0] << " " << results[1] << std::endl;

	system("pause");
	return EXIT_SUCCESS;
}
转载请注明:文章转载自 www.mshxw.com
本文地址:https://www.mshxw.com/it/893449.html
我们一直用心在做
关于我们 文章归档 网站地图 联系我们

版权所有 (c)2021-2022 MSHXW.COM

ICP备案号:晋ICP备2021003244-6号