Analytic Derivatives

Example of Analytic Derivatives

考虑通过样本数据对如下问题 (Rat43) 进行曲线拟合

y=b1(1+eb2b3x)1/b4y=\frac{b_{1}}{\left(1+e^{b_{2}-b_{3} x}\right)^{1 / b_{4}}}

给定一些样本数据 {xi,yi},i=1,,n\left\{x_{i}, y_{i}\right\}, \forall i=1, \ldots, n ,估计参数 b1,b2,b3,b4b_{1}, b_{2}, b_{3},b_{4} 。可以建模成寻找使得如下目标函数取得最小值时, b1,b2,b3,b4b_{1}, b_{2}, b_{3},b_{4} 的值:

E(b1,b2,b3,b4)=if2(b1,b2,b3,b4;xi,yi)=i(b1(1+eb2b3xi)1/b4yi)2\begin{aligned} E\left(b_{1}, b_{2}, b_{3}, b_{4}\right) & =\sum_{i} f^{2}\left(b_{1}, b_{2}, b_{3}, b_{4} ; x_{i}, y_{i}\right) \\ & =\sum_{i}\left(\frac{b_{1}}{\left(1+e^{b_{2}-b_{3} x_{i}}\right)^{1 / b_{4}}}-y_{i}\right)^{2} \end{aligned}

使用 Ceres 解决该问题,我们需要定义一个 costFunction,用于计算给定 xxyy 的残差 ff 及其相对于 b1b2b3b_{1}、b_{2}、b_{3}b4b_{4} 的导数。

D1f(b1,b2,b3,b4;x,y)=1(1+eb2b3x)1/b4D2f(b1,b2,b3,b4;x,y)=b1eb2b3xb4(1+eb2b3x)1/b4+1D3f(b1,b2,b3,b4;x,y)=b1xeb2b3xb4(1+eb2b3x)1/b4+1D4f(b1,b2,b3,b4;x,y)=b1log(1+eb2b3x)b42(1+eb2b3x)1/b4\begin{array}{l} D_{1} f\left(b_{1}, b_{2}, b_{3}, b_{4} ; x, y\right)=\frac{1}{\left(1+e^{b_{2}-b_{3} x}\right)^{1 / b_{4}}} \\ D_{2} f\left(b_{1}, b_{2}, b_{3}, b_{4} ; x, y\right)=\frac{-b_{1} e^{b_{2}-b_{3} x}}{b_{4}\left(1+e^{b_{2}-b_{3} x}\right)^{1 / b_{4}+1}} \\ D_{3} f\left(b_{1}, b_{2}, b_{3}, b_{4} ; x, y\right)=\frac{b_{1} x e^{b_{2}-b_{3} x}}{b_{4}\left(1+e^{b_{2}-b_{3} x}\right)^{1 / b_{4}+1}} \\ D_{4} f\left(b_{1}, b_{2}, b_{3}, b_{4} ; x, y\right)=\frac{b_{1} \log \left(1+e^{b_{2}-b_{3} x}\right)}{b_{4}^{2}\left(1+e^{b_{2}-b_{3} x}\right)^{1 / b_{4}}} \end{array}

有了这些导数,我们现在可以将 CostFunction 实现为:

class Rat43Analytic : public SizedCostFunction<1,4> {
   public:
     Rat43Analytic(const double x, const double y) : x_(x), y_(y) {}
     virtual ~Rat43Analytic() {}
     virtual bool Evaluate(double const* const* parameters,
                           double* residuals,
                           double** jacobians) const {
       const double b1 = parameters[0][0];
       const double b2 = parameters[0][1];
       const double b3 = parameters[0][2];
       const double b4 = parameters[0][3];

       residuals[0] = b1 *  pow(1 + exp(b2 -  b3 * x_), -1.0 / b4) - y_;

       if (!jacobians) return true;
       double* jacobian = jacobians[0];
       if (!jacobian) return true;

       jacobian[0] = pow(1 + exp(b2 - b3 * x_), -1.0 / b4);
       jacobian[1] = -b1 * exp(b2 - b3 * x_) *
                     pow(1 + exp(b2 - b3 * x_), -1.0 / b4 - 1) / b4;
       jacobian[2] = x_ * b1 * exp(b2 - b3 * x_) *
                     pow(1 + exp(b2 - b3 * x_), -1.0 / b4 - 1) / b4;
       jacobian[3] = b1 * log(1 + exp(b2 - b3 * x_)) *
                     pow(1 + exp(b2 - b3 * x_), -1.0 / b4) / (b4 * b4);
       return true;
     }

    private:
     const double x_;
     const double y_;
 };

这些难以阅读并且有很多冗余。因此在实践中我们会用一些子表达式来提高其效率,这会给我们带来类似的结果:

class Rat43AnalyticOptimized : public SizedCostFunction<1,4> {
   public:
     Rat43AnalyticOptimized(const double x, const double y) : x_(x), y_(y) {}
     virtual ~Rat43AnalyticOptimized() {}
     virtual bool Evaluate(double const* const* parameters,
                           double* residuals,
                           double** jacobians) const {
       const double b1 = parameters[0][0];
       const double b2 = parameters[0][1];
       const double b3 = parameters[0][2];
       const double b4 = parameters[0][3];

       const double t1 = exp(b2 -  b3 * x_);
       const double t2 = 1 + t1;
       const double t3 = pow(t2, -1.0 / b4);
       residuals[0] = b1 * t3 - y_;

       if (!jacobians) return true;
       double* jacobian = jacobians[0];
       if (!jacobian) return true;

       const double t4 = pow(t2, -1.0 / b4 - 1);
       jacobian[0] = t3;
       jacobian[1] = -b1 * t1 * t4 / b4;
       jacobian[2] = -x_ * jacobian[1];
       jacobian[3] = b1 * log(t2) * t3 / (b4 * b4);
       return true;
     }

   private:
     const double x_;
     const double y_;
 };

两种实现的区别如下:

CostFunctionTimes(ns)

Rat43Analytic

255

Rat43Analytic

92

When should you use analytical derivatives?

  • 表达式很简单,例如大多是线性的。

  • 可以使用 MapleMathematicaSymPy 等计算机代数系统对目标函数进行微分,并生成 C++ 对其进行评估。

  • 性能是最先考虑的,可以利用计算中的代数结构来获得比自动微分更好的性能。尽管如此,要想从 analytical derivatives 中获得最佳性能,还需要做大量的工作。在采用这种方法之前,最好先估算一下求解 jacobian 所花费的时间占总求解时间的比例,使用 Amdahl’s Law 即可。

  • 您喜欢链式法则,并且喜欢手工计算所有导数。

  • There is no other way to compute the derivatives, e.g. you wish to compute the derivative of the root of a polynomial, a3(x,y)z3+a2(x,y)z2+a1(x,y)z+a0(x,y)=0a_{3}(x, y) z^{3}+a_{2}(x, y) z^{2}+a_{1}(x, y) z+a_{0}(x, y)=0, with respect to x,yx,y, This requires the use of the Inverse Function Theorem.

Last updated