转换仿函数结构以采用不同的参数

Transform functor struct to take a different argument

我从 Eigen 库中得到了以下(无约束二次方 objective)定义的借用矩阵和向量:

#ifndef QP_UNCON_HPP
#define QP_UNCON_HPP
#include "EigenDataTypes.hpp"

template <int Nx>
struct objective
{
private:
    const spMat Q;
    const Vec<Nx> c;

public:
    objective(spMat Q_, Vec<Nx> c_) : Q(Q_), c(c_) {}

    inline scalar operator()(const Vec<Nx> &x)
    {
        return (.5 * x.transpose() * Q * x + c.transpose() * x);
    }

    inline Vec<Nx> Eval_grad(const Vec<Nx> &x)
    {
        return Q.transpose() * x;
    }

    inline Mat<Nx, Nx> Eval_hessian(const Vec<Nx> &x)
    {
        return Q;
    }
};

这使得在不同状态 x:

下评估 objective 成为可能
objective<2> f(Q, c);
Vec<2> x {0,1};

#Objective-value
f(x);
#gradient:
f.Eval_grad(x);
#hessian:
f.Eval_hessian(x);

我想创建一个新的 struct Phi(用于行搜索),其中 scalar 用作输入参数,如下所示:

p_objective<2> Phi(f, x0, p);
double alpha = 0.9;

#Objective-value
Phi(alpha);
#gradient:
Phi.Eval_grad(alpha);
#hessian:
Phi.Eval_hessian(alpha);

对应的是:

#Objective-value
f(x + alpha*p);
#gradient:
f.Eval_grad(x + alpha*p);
#hessian:
f.Eval_hessian(x + alpha*p);

这对于使用 lambda 函数的单个函数来说很简单,但是 'lambdifying' 仿函数结构有平滑的方法吗?

EigenDataTypes.hpp

#ifndef EIGENDATATYPES_H
#define EIGENDATATYPES_H

#include <Eigen/Dense>
#include <Eigen/SparseCore>

#ifdef CSOLVER_USE_SINGLE
typedef float real_t;
#else
typedef double real_t;
#endif

using scalar = Eigen::Matrix<double, 1, 1>;
template <int Rows>
using Vec = Eigen::Matrix<double, Rows, 1>;
template <int Rows, int Cols>
using Mat = Eigen::Matrix<double, Rows, Cols>;
using spVec = Eigen::SparseVector<double>;
using spMat = Eigen::SparseMatrix<double>;
using Triplet = Eigen::Triplet<double>;
#endif

没有特征库的一维示例:

struct objective_1D
{
private:
    const double Q;
    const double c;

public:
    objective_1D(double Q_, double c_) : Q(Q_), c(c_) {}

    double operator()(const double &x)
    {
        return (.5 * x * Q * x + c* x);
    }
    double Eval_grad(const double &x)
    {
        return Q * x;
    }
    double Eval_hessian(const double &x)
    {
        return Q;
    }
};

TLDR; 我想创建一个仿函数结构 p_objective 用作 struct objective:

的 lambda
p_objective = [&x, &p] (double alpha) (return objective-methods at (x + alpha*p))

I want to create a functor struct p_objective that works as a lambda for struct objective.

如果你已经写了objective,你也可以同样写一个p_objective。下面是一个例子。

这里是 (demo)

template <int Nx> class p_objective
{
    objective<Nx> f;
    const spMat x;
    const Vec<Nx> p;

public:
    explicit p_objective(const objective<Nx>& ob, spMat const& Q_, const Vec<Nx>& c_)
        : f{ ob }
        , x{ Q_ }
        , p{ c_ }
    {}

    scalar operator()(const double alpha)
    {
        return f(x + (alpha*p));
    }

    Vec<Nx> Eval_grad(const double alpha)
    {
        return f.Eval_grad(x + alpha * p);
    }

    Mat<Nx, Nx> Eval_hessian(const double alpha)
    {
        return f.Eval_hessian(x + alpha * p);
    }
};

由于objectivep_objective都有成员函数,(即Eval_gradEval_hessian),当它们正常时,它会更具可读性类,而不是 lambda 函数。