DGtal::AngleLinearMinimizerByGradientDescent Class Reference

#include <AngleLinearMinimizer.h>

Inheritance diagram for DGtal::AngleLinearMinimizerByGradientDescent:

Collaboration diagram for DGtal::AngleLinearMinimizerByGradientDescent:

Public Member Functions
	AngleLinearMinimizerByGradientDescent (double step=0.1)
virtual	~AngleLinearMinimizerByGradientDescent ()
virtual double	lastDelta () const
virtual void	selfDisplay (std::ostream &aStream) const
Public Member Functions inherited from DGtal::AngleLinearMinimizer
virtual	~AngleLinearMinimizer ()
	AngleLinearMinimizer ()
void	reset ()
void	init (unsigned int nbMax)
ValueInfo &	rw (unsigned int i)
const ValueInfo &	ro (unsigned int i) const
unsigned int	maxSize () const
unsigned int	size () const
void	setSize (unsigned int nb)
void	setIsCurveOpen (bool is_curve_open=false)
double	getEnergy (unsigned int i1, unsigned int i2) const
double	getFormerEnergy (unsigned int i1, unsigned int i2) const
std::vector< double >	getGradient () const
std::vector< double >	getFormerGradient () const
double	optimize ()
double	optimize (unsigned int i1, unsigned int i2)
double	sum () const
double	max () const
std::string	className () const
bool	isValid () const

Protected Member Functions
virtual void	oneStep (unsigned int i1, unsigned int i2)

Private Attributes
double	myStep

Additional Inherited Members
Protected Attributes inherited from DGtal::AngleLinearMinimizer
bool	myIsCurveOpen
ValueInfo *	myValues
unsigned int	mySize

---

Detailed Description

Specializes AngleLinearMinimizer to optimize with a gradient descent method.

Definition at line 423 of file AngleLinearMinimizer.h.

---

Constructor & Destructor Documentation

DGtal::AngleLinearMinimizerByGradientDescent::AngleLinearMinimizerByGradientDescent ( double  aStep = 0.1 )

inline

Default constructor. Does nothing.

Definition at line 89 of file AngleLinearMinimizer.ih.

  : myStep( aStep )
{}

DGtal::AngleLinearMinimizerByGradientDescent::~AngleLinearMinimizerByGradientDescent ( )

inlinevirtual

Destructor. Does nothing.

Definition at line 98 of file AngleLinearMinimizer.ih.

{}

---

Member Function Documentation

double DGtal::AngleLinearMinimizerByGradientDescent::lastDelta ( ) const

virtual

Should be used to stop the minimization process. The smaller is this value, the more the optimization is at an end. May have several meanings, like the infinite norm of the last displacement or the infinite norm of the projected gradient.

Returns:

an upper bound on the norm of the last displacement.

Reimplemented from DGtal::AngleLinearMinimizer.

Definition at line 389 of file AngleLinearMinimizer.cpp.

References DGtal::AngleLinearMinimizer::ValueInfo::oldValue, and DGtal::AngleLinearMinimizer::ValueInfo::value.

{
  vector<double> grad ( getFormerGradient() );
  double ninf = 0.0;
  for ( unsigned int i = 0; i < size(); i++ ) 
    {
      const ValueInfo & vi = this->ro( i );
      if ( vi.value != vi.oldValue )
        {
          double n = fabs( grad[ i ] );
          if ( n > ninf ) ninf = n;
        }
    }
  return ninf;
}

void DGtal::AngleLinearMinimizerByGradientDescent::oneStep ( unsigned int  i1, unsigned int  i2  )

protectedvirtual

The method which performs the optimization effectively. Use a relaxation technique. The optimization is performed on values [i1] included to [i2] excluded.

Parameters:

i1	the first value to be optimized (between 0 and 'size()-1').
i2	the value after the last to be optimized (between 0 and 'size()-1').

Reimplemented from DGtal::AngleLinearMinimizer.

Definition at line 361 of file AngleLinearMinimizer.cpp.

References DGtal::AngleComputer::cast(), DGtal::AngleComputer::less(), DGtal::AngleLinearMinimizer::ValueInfo::max, DGtal::AngleLinearMinimizer::ValueInfo::min, DGtal::ModuloComputer< TInteger >::next(), DGtal::AngleLinearMinimizer::ValueInfo::oldValue, and DGtal::AngleLinearMinimizer::ValueInfo::value.

{
  ModuloComputer<unsigned int> mc( size() );

  vector<double> grad ( getFormerGradient() );
  double mid;
  unsigned int i = i1; 
  do 
    {
      unsigned int inext = mc.next( i );
      ValueInfo & vi = this->rw( i );
      double val = vi.oldValue;
      mid = AngleComputer::cast( val - myStep*grad[ i ] );
      if ( AngleComputer::less( mid, vi.min ) ) mid = vi.min;
      if ( AngleComputer::less( vi.max, mid ) ) mid = vi.max;
      vi.value = mid;
      // go to next.
      i = inext;
    }
  while ( i != i2 );
  double E1 = getFormerEnergy( i1, i2 );
  double E2 = getEnergy( i1, i2 );  
  cerr << "E1=" << E1 << " E2=" << E2 << " s=" << myStep << endl;
}

void DGtal::AngleLinearMinimizerByGradientDescent::selfDisplay ( std::ostream &  aStream ) const

virtual

Writes/Displays the object on an output stream.

Parameters:

aStream

the output stream where the object is written.

Reimplemented from DGtal::AngleLinearMinimizer.

Definition at line 495 of file AngleLinearMinimizer.cpp.

{
  aStream << "[LinearMinimizer::gradient descent " << myStep << "]";
}

---

Field Documentation

double DGtal::AngleLinearMinimizerByGradientDescent::myStep

private

The step for the gradient descent.

Definition at line 430 of file AngleLinearMinimizer.h.

---

The documentation for this class was generated from the following files:

• AngleLinearMinimizer.h
• AngleLinearMinimizer.cpp
• AngleLinearMinimizer.ih