shark::McPegasos< VectorType > Class Template Reference

Pegasos solver for linear multi-class support vector machines. More...

#include <shark/Algorithms/Pegasos.h>

Public Types
enum	eMarginType { emRelative , emAbsolute }
	Multi-class margin type. More...
enum	eLossType {   elNaiveHinge , elDiscriminativeMax , elDiscriminativeSum , elTotalMax ,   elTotalSum }
	Multi-class loss function type. More...

Static Public Member Functions
template<class WeightType >
static std::size_t	solve (LabeledData< VectorType, unsigned int > const &data, eMarginType margintype, eLossType losstype, bool sumToZero, double C, std::vector< WeightType > &w, std::size_t batchsize=1, double varepsilon=0.001)
	Solve the primal multi-class SVM problem.

Protected Types
typedef bool(*	LossGradientFunction) (VectorType const &, unsigned int, RealVector const &, std::vector< VectorType > &, bool)

Static Protected Member Functions
static bool	lossGradientANH (VectorType const &x, unsigned int y, RealVector const &f, std::vector< VectorType > &gradient, bool sumToZero)
static bool	lossGradientRDM (VectorType const &x, unsigned int y, RealVector const &f, std::vector< VectorType > &gradient, bool sumToZero)
static bool	lossGradientRDS (VectorType const &x, unsigned int y, RealVector const &f, std::vector< VectorType > &gradient, bool sumToZero)
static bool	lossGradientADS (VectorType const &x, unsigned int y, RealVector const &f, std::vector< VectorType > &gradient, bool sumToZero)
static bool	lossGradientADM (VectorType const &x, unsigned int y, RealVector const &f, std::vector< VectorType > &gradient, bool sumToZero)
static bool	lossGradientATS (VectorType const &x, unsigned int y, RealVector const &f, std::vector< VectorType > &gradient, bool sumToZero)
static bool	lossGradientATM (VectorType const &x, unsigned int y, RealVector const &f, std::vector< VectorType > &gradient, bool sumToZero)

Detailed Description

template<class VectorType>
class shark::McPegasos< VectorType >

Pegasos solver for linear multi-class support vector machines.

Definition at line 192 of file Pegasos.h.

Member Typedef Documentation

LossGradientFunction

template<class VectorType >

typedef bool(* shark::McPegasos< VectorType >::LossGradientFunction) (VectorType const &, unsigned int, RealVector const &, std::vector< VectorType > &, bool)

protected

Definition at line 384 of file Pegasos.h.

Member Enumeration Documentation

eLossType

template<class VectorType >

enum shark::McPegasos::eLossType

Multi-class loss function type.

Enumerator
elNaiveHinge
elDiscriminativeMax
elDiscriminativeSum
elTotalMax
elTotalSum

Definition at line 203 of file Pegasos.h.

eMarginType

template<class VectorType >

enum shark::McPegasos::eMarginType

Multi-class margin type.

Enumerator
emRelative
emAbsolute

Definition at line 196 of file Pegasos.h.

Member Function Documentation

lossGradientADM()

template<class VectorType >

static bool shark::McPegasos< VectorType >::lossGradientADM ( VectorType const &  x, unsigned int  y, RealVector const &  f, std::vector< VectorType > &  gradient, bool  sumToZero  )

inlinestaticprotected

Definition at line 483 of file Pegasos.h.

Referenced by shark::McPegasos< VectorType >::solve().

lossGradientADS()

template<class VectorType >

static bool shark::McPegasos< VectorType >::lossGradientADS ( VectorType const &  x, unsigned int  y, RealVector const &  f, std::vector< VectorType > &  gradient, bool  sumToZero  )

inlinestaticprotected

Definition at line 456 of file Pegasos.h.

References shark::mean().

Referenced by shark::McPegasos< VectorType >::solve().

lossGradientANH()

template<class VectorType >

static bool shark::McPegasos< VectorType >::lossGradientANH ( VectorType const &  x, unsigned int  y, RealVector const &  f, std::vector< VectorType > &  gradient, bool  sumToZero  )

inlinestaticprotected

Definition at line 387 of file Pegasos.h.

Referenced by shark::McPegasos< VectorType >::solve().

lossGradientATM()

template<class VectorType >

static bool shark::McPegasos< VectorType >::lossGradientATM ( VectorType const &  x, unsigned int  y, RealVector const &  f, std::vector< VectorType > &  gradient, bool  sumToZero  )

inlinestaticprotected

Definition at line 553 of file Pegasos.h.

Referenced by shark::McPegasos< VectorType >::solve().

lossGradientATS()

template<class VectorType >

static bool shark::McPegasos< VectorType >::lossGradientATS ( VectorType const &  x, unsigned int  y, RealVector const &  f, std::vector< VectorType > &  gradient, bool  sumToZero  )

inlinestaticprotected

Definition at line 515 of file Pegasos.h.

References shark::mean().

Referenced by shark::McPegasos< VectorType >::solve().

lossGradientRDM()

template<class VectorType >

static bool shark::McPegasos< VectorType >::lossGradientRDM ( VectorType const &  x, unsigned int  y, RealVector const &  f, std::vector< VectorType > &  gradient, bool  sumToZero  )

inlinestaticprotected

Definition at line 408 of file Pegasos.h.

Referenced by shark::McPegasos< VectorType >::solve().

lossGradientRDS()

template<class VectorType >

static bool shark::McPegasos< VectorType >::lossGradientRDS ( VectorType const &  x, unsigned int  y, RealVector const &  f, std::vector< VectorType > &  gradient, bool  sumToZero  )

inlinestaticprotected

Definition at line 435 of file Pegasos.h.

Referenced by shark::McPegasos< VectorType >::solve().

solve()

template<class VectorType >

template<class WeightType >

static std::size_t shark::McPegasos< VectorType >::solve ( LabeledData< VectorType, unsigned int > const &  data, eMarginType  margintype, eLossType  losstype, bool  sumToZero, double  C, std::vector< WeightType > &  w, std::size_t  batchsize = 1, double  varepsilon = 0.001  )

inlinestatic

Solve the primal multi-class SVM problem.

In addition to "standard" Pegasos this solver checks a meaningful stopping criterion.

The function returns the number of model predictions during training (this is comparable to SMO iterations).

Parameters

data	training data
margintype	margin function type
losstype	loss function type
sumToZero	enforce the sum-to-zero constraint?
C	SVM regularization parameter
w	class-wise weight vectors
batchsize	number of samples in each gradient estimate
varepsilon	solution accuracy (factor by which the primal gradient should be reduced)

Definition at line 220 of file Pegasos.h.

References shark::random::discrete(), shark::McPegasos< VectorType >::elDiscriminativeMax, shark::McPegasos< VectorType >::elDiscriminativeSum, shark::McPegasos< VectorType >::elNaiveHinge, shark::McPegasos< VectorType >::elTotalMax, shark::McPegasos< VectorType >::elTotalSum, shark::McPegasos< VectorType >::emAbsolute, shark::McPegasos< VectorType >::emRelative, shark::McPegasos< VectorType >::lossGradientADM(), shark::McPegasos< VectorType >::lossGradientADS(), shark::McPegasos< VectorType >::lossGradientANH(), shark::McPegasos< VectorType >::lossGradientATM(), shark::McPegasos< VectorType >::lossGradientATS(), shark::McPegasos< VectorType >::lossGradientRDM(), shark::McPegasos< VectorType >::lossGradientRDS(), shark::LabeledData< InputT, LabelT >::numberOfElements(), SHARK_ASSERT, and SHARK_RUNTIME_CHECK.

---

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

• include/shark/Algorithms/Pegasos.h