|| || || || || ||

---

Impala::Core::Training::Svm Class Reference

SVM is a well known Classifier, see internet for details. More...

#include <Svm.h>

Inheritance diagram for Impala::Core::Training::Svm:

[legend]Collaboration diagram for Impala::Core::Training::Svm:

[legend]List of all members.

Public Member Functions
	Svm ()
	Svm (String s)
virtual	~Svm ()
virtual void	Train (Util::PropertySet *properties, TrainDataSrc *data)
	Trains a model.
virtual Table::ScoreTable *	Predict (TrainDataSrc *data)
virtual void	PredictForActiveLearn (Matrix::DistributedAccess *da, Table::QuidTable *columnQuids, Core::Table::SimilarityTableSet::SimTableType *result)
void	SetModel (svm_model *model)
virtual void	OverrideModelOptions (Util::PropertySet *properties)
const svm_model *	GetModel ()
bool	Equals (const Svm *other) const
int	ReferenceDiff (const Svm *other) const
int	Diff (const Svm *other) const
Private Member Functions
void	CheckTestMode (svm_problem *problem)
double	PredictSingle (const svm_node *problem)
Private Attributes
svm_model *	mModel
bool	mIsProbabilityModel
int	mProbilityIndex
	ILOG_VAR_DEC

---

Detailed Description

SVM is a well known Classifier, see internet for details.

Todo:

Fix the machanism to set gamma to 1/vectorlength. See OverrideModelOptions(), Train() and MakeSvmParams().

parameters for the model: {w1, w2, C, gamma, cache, degree, coef0, nu, eps, p, shrinking, probability}

Note:

Warning about libsvm! (Just in case other people ever want to change this part of Impala) Things I (Michiel) learned the hard way:

• LibSvm is not meant to be used the way it is used in impala. The memory model (if you can call it that) assumes usage by either svm-train.c or svm-predict.c from the libsvm package.
• encountered problems:
  • pointers to svm_nodes are shared between the problem and the model. So don't delete the data until you're done using the model. This has cost me a few days :) I have created a fix for this: FixModelDependency in LinkSvm.h
  • in the inner wrokings of svm_train() callers of SVC_Q::get_Q expect a pointer to data that remains valid even after another call to get_Q returned another pointer. (This is possible because of the caching mechanisms). Beware when you start tinkering here. This one has cost me a few weeks :)

Definition at line 50 of file Svm.h.

---

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

• Core/Training/Svm.h

---