【凸优化】交替方向乘子法的PyTorch实现

Harvey Chui

已于 2023-11-28 15:45:23 修改

阅读量587

点赞数 11

分类专栏：凸优化文章标签： pytorch 人工智能 python

于 2023-11-28 11:23:34 首次发布

本文链接：https://blog.csdn.net/CalvinTri/article/details/134663005

版权

凸优化专栏收录该内容

4 篇文章 0 订阅

订阅专栏

【凸优化】交替方向乘子法的PyTorch实现

本实验使用Python作为编程语言，PyTorch作为机器学习库，实现了交替方向乘子法

基本步骤

算法基本步骤

优化目标

本次实验的优化目标如下：
$f(x)=\frac{3}{2} x_1^2+\frac{1}{2} x_2^2-x_1 x_2-2 x_1 \\ g(z)=\frac{1}{2} z_1^2+2 z_2^2+z_1 z_2+z_1+z_2 \\ A=(1,1) \ B=(1,1) \ C=1$

代码

import torch
import numpy as np
import matplotlib.pyplot as plt
from math import sqrt, log

x = torch.tensor([1., 1.], requires_grad=True).reshape(2, 1)
z = torch.tensor([1., 1.], requires_grad=True).reshape(2, 1)

rho = 2
epsilon = 0.01
u = torch.tensor([0.])
c = 1
A_ = torch.tensor([1., 1.])
B_ = torch.tensor([1., 1.])

A = torch.tensor([[3., -1], [-1, 1]])
b1 = torch.tensor([2., 0])


# 定义初始函数
def f(x, z, u):
    return 1 / 2 * x.t() @ A @ x - b1.t() @ x + 1 / 2 * rho * torch.norm(
        x @ torch.tensor([1., 1.]).reshape(1, 2) + z @ torch.tensor([1., 1.]).reshape(1, 2) - 1 + u)


B = torch.tensor([[1., 4.], [1., 1.]])
b2 = torch.tensor([1., 1.])


def g(x, z, u):
    return 1 / 2 * z.t() @ B @ z + b2.t() @ z + 1 / 2 * rho * torch.norm(
        x @ torch.tensor([1., 1.]).reshape(1, 2) + z @ torch.tensor([1., 1.]).reshape(1, 2) - 1 + u)


def argmin_f(x, z, u):
    def f_grad(x, z, u):  # 计算梯度
        y = f(x, z, u)
        grad = torch.autograd.grad(y, x, retain_graph=True, create_graph=True)[0]
        return grad

    alpha = torch.tensor([0.25])
    beta = torch.tensor([0.5])
    eta = 0.1
    while torch.norm(f_grad(x, z, u)) > eta:
        delta_x = - f_grad(x, z, u)
        t = torch.tensor([1.])
        while f(x + t.mul(delta_x), z, u) > (f(x, z, u) + alpha.mul(t).mul((f_grad(x, z, u).t()) @ delta_x)):
            t = beta.mul(t)
        x = x + t.mul(delta_x)
    return x


def argmin_g(x, z, u):
    def g_grad(x, z, u):  # 计算梯度
        y = g(x, z, u)
        grad = torch.autograd.grad(y, z, retain_graph=True, create_graph=True)[0]
        return grad

    alpha = torch.tensor([0.25])
    beta = torch.tensor([0.5])
    eta = 0.1
    while torch.norm(g_grad(x, z, u)) > eta:
        delta_z = - g_grad(x, z, u)
        t = torch.tensor([1.])
        while g(x, z + t.mul(delta_z), u) > (g(x, z, u) + alpha.mul(t).mul((g_grad(x, z, u).t()) @ delta_z)):
            t = beta.mul(t)
        z = z + t.mul(delta_z)
    return z


def admm(x, z, u):
    while torch.norm(A_@x+B@z-c) > epsilon or torch.norm(rho*A_.t()@B@(z_last-z)) > epsilon:
        x = argmin_f(x, z, u)
        z_last = z
        z = argmin_g(x, z, u)
        u = u + (torch.tensor([1., 1.]).reshape(1, 2)@x - torch.tensor([1., 1.]).reshape(1, 2)@z)
        opt = 1/2*x.t()@A@x - b1.t()@x + 1/2*z.t()@B@z + b2.t()@z_last
        print('Function value is: ' + str(opt.tolist()))
    opt = 1/2*x.t()@A@x - b1.t()@x + 1/2*z.t()@B@z + b2.t()@z
    print('Optimal solution is: ' + str(opt.tolist()))
   
if __name__ == '__main__':
	admm(x, z, u)