A manifold learning algorithm that uses back-propagation to train a neural net model to map from low-dimensional space to high-dimensional space. More...

#include <GManifold.h>

Inheritance diagram for GClasses::GUnsupervisedBackProp:

Public Member Functions
	GUnsupervisedBackProp (size_t intrinsicDims, GRand *pRand)

	GUnsupervisedBackProp (GDomNode *pNode, GLearnerLoader &ll)

virtual	~GUnsupervisedBackProp ()

size_t	featureDims ()

void	hiToLow (const double pIn, double pOut)
	Given a high-dimensional vector, computes and returns a corresponding low-dimensional vector. More...

GImageJitterer *	jitterer ()
	Returns the current image jitterer. More...

size_t	labelDims ()

void	lowToHi (const double pIn, double pOut)
	Given a low-dimensional vector, computes and returns the corresponding high-dimensional vector. More...

double *	mins ()

GNeuralNet *	neuralNet ()
	Returns a pointer to the neural network used to model the manifold. Typically, this is used to add layers to the neural network, or set the learning rate (etc.) before calling reduce. (You must add at least one layer.) More...

void	onePass ()
	Use only single-pass training. More...

GMatrix &	progress ()

double *	ranges ()

virtual GMatrix *	reduce (const GMatrix &in)
	Perform NLDR. (This also trains the internal neural network to map from low-dimensional space to high-dimensional space.) Returns a pointer to the intrinsic values (which you are responsible to delete). More...

GDomNode *	serialize (GDom *pDoc) const
	Marshall this object to a DOM that can be serialized. More...

void	setIntrinsic (GMatrix *pIntrinsic)
	Specify initial values for the intrinsic variables. This method takes ownership of pIntrinsic. If this method is not called prior to "reduce", then the intrinsic variables will be initialized with small random values. More...

void	setJitterer (GImageJitterer *pJitterer)
	Takes ownership of pJitterer. Specify an image jitterer to use during training to make it robust to rotation, translation, and scale. If an image jitterer is used, then there must be exactly two dimensional parameters, and the parameters used to construct the image jitterer must be consistent with those used to construct this object. More...

void	setNeuralNet (GNeuralNet *pNN)
	Takes ownership of pNN. Replaces the internal neural net with the one specified. This method assumes that pNN has already been trained. If m_updateWeights is true, then it will further-refinde this model when reduce is called. (You can pass NULL to this method to discard the current model, so that a new model will be trained next time reduce is called.) More...

void	setParams (std::vector< size_t > &paramRanges)
	Parameterize the output values. This feature is typically used when the output is an image, and the width and height are specified for the paramRanges. More...

void	setUseInputBias (bool b)
	Specify whether to use one of the input values as a bias. More...

void	trackProgress ()

void	useJitterer ()

Public Member Functions inherited from GClasses::GTransform
	GTransform ()

	GTransform (GDomNode *pNode, GLearnerLoader &ll)

virtual	~GTransform ()

Protected Attributes
GCoordVectorIterator	m_cvi

size_t	m_intrinsicDims

size_t	m_jitterDims

bool	m_onePass

size_t	m_paramDims

GMatrix *	m_pIntrinsic

GImageJitterer *	m_pJitterer

double *	m_pMins

GNeuralNet *	m_pNN

size_t *	m_pParamRanges

GMatrix *	m_pProgress

double *	m_pRanges

bool	m_useInputBias

Additional Inherited Members
Protected Member Functions inherited from GClasses::GTransform
virtual GDomNode *	baseDomNode (GDom pDoc, const char szClassName) const
	Child classes should use this in their implementation of serialize. More...

Detailed Description

A manifold learning algorithm that uses back-propagation to train a neural net model to map from low-dimensional space to high-dimensional space.

Constructor & Destructor Documentation

GClasses::GUnsupervisedBackProp::GUnsupervisedBackProp	(	size_t	intrinsicDims,
		GRand *	pRand
	)

GClasses::GUnsupervisedBackProp::GUnsupervisedBackProp	(	GDomNode *	pNode,
		GLearnerLoader &	ll
	)

virtual GClasses::GUnsupervisedBackProp::~GUnsupervisedBackProp ( )

virtual

Member Function Documentation

size_t GClasses::GUnsupervisedBackProp::featureDims ( )

inline

void GClasses::GUnsupervisedBackProp::hiToLow	(	const double *	pIn,
		double *	pOut
	)

Given a high-dimensional vector, computes and returns a corresponding low-dimensional vector.

GImageJitterer* GClasses::GUnsupervisedBackProp::jitterer ( )

inline

Returns the current image jitterer.

size_t GClasses::GUnsupervisedBackProp::labelDims ( )

void GClasses::GUnsupervisedBackProp::lowToHi	(	const double *	pIn,
		double *	pOut
	)

Given a low-dimensional vector, computes and returns the corresponding high-dimensional vector.

double* GClasses::GUnsupervisedBackProp::mins ( )

GNeuralNet* GClasses::GUnsupervisedBackProp::neuralNet ( )

inline

Returns a pointer to the neural network used to model the manifold. Typically, this is used to add layers to the neural network, or set the learning rate (etc.) before calling reduce. (You must add at least one layer.)

void GClasses::GUnsupervisedBackProp::onePass ( )

inline

Use only single-pass training.

GMatrix& GClasses::GUnsupervisedBackProp::progress ( )

inline

double* GClasses::GUnsupervisedBackProp::ranges ( )

virtual GMatrix* GClasses::GUnsupervisedBackProp::reduce ( const GMatrix & in )

virtual

Perform NLDR. (This also trains the internal neural network to map from low-dimensional space to high-dimensional space.) Returns a pointer to the intrinsic values (which you are responsible to delete).

Implements GClasses::GTransform.

GDomNode* GClasses::GUnsupervisedBackProp::serialize ( GDom * pDoc ) const

Marshall this object to a DOM that can be serialized.

void GClasses::GUnsupervisedBackProp::setIntrinsic ( GMatrix * pIntrinsic )

Specify initial values for the intrinsic variables. This method takes ownership of pIntrinsic. If this method is not called prior to "reduce", then the intrinsic variables will be initialized with small random values.

void GClasses::GUnsupervisedBackProp::setJitterer ( GImageJitterer * pJitterer )

Takes ownership of pJitterer. Specify an image jitterer to use during training to make it robust to rotation, translation, and scale. If an image jitterer is used, then there must be exactly two dimensional parameters, and the parameters used to construct the image jitterer must be consistent with those used to construct this object.

void GClasses::GUnsupervisedBackProp::setNeuralNet ( GNeuralNet * pNN )

Takes ownership of pNN. Replaces the internal neural net with the one specified. This method assumes that pNN has already been trained. If m_updateWeights is true, then it will further-refinde this model when reduce is called. (You can pass NULL to this method to discard the current model, so that a new model will be trained next time reduce is called.)

void GClasses::GUnsupervisedBackProp::setParams ( std::vector< size_t > & paramRanges )

Parameterize the output values. This feature is typically used when the output is an image, and the width and height are specified for the paramRanges.

void GClasses::GUnsupervisedBackProp::setUseInputBias ( bool b )

inline

Specify whether to use one of the input values as a bias.

void GClasses::GUnsupervisedBackProp::trackProgress ( )

void GClasses::GUnsupervisedBackProp::useJitterer ( )

inline

Member Data Documentation

GCoordVectorIterator GClasses::GUnsupervisedBackProp::m_cvi

protected

size_t GClasses::GUnsupervisedBackProp::m_intrinsicDims

protected

size_t GClasses::GUnsupervisedBackProp::m_jitterDims

protected

bool GClasses::GUnsupervisedBackProp::m_onePass

protected

size_t GClasses::GUnsupervisedBackProp::m_paramDims

protected

GMatrix* GClasses::GUnsupervisedBackProp::m_pIntrinsic

protected

GImageJitterer* GClasses::GUnsupervisedBackProp::m_pJitterer

protected

double* GClasses::GUnsupervisedBackProp::m_pMins

protected

GNeuralNet* GClasses::GUnsupervisedBackProp::m_pNN

protected

size_t* GClasses::GUnsupervisedBackProp::m_pParamRanges

protected

GMatrix* GClasses::GUnsupervisedBackProp::m_pProgress

protected

double* GClasses::GUnsupervisedBackProp::m_pRanges

protected

bool GClasses::GUnsupervisedBackProp::m_useInputBias

protected

Public Member Functions

Protected Attributes

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

Member Function Documentation

Member Data Documentation