Base class for arbitrary machine learning models based on adaptive sparse grids. More...

#include <ModelFittingBase.hpp>

Inheritance diagram for sgpp::datadriven::ModelFittingBase:

Public Member Functions
virtual bool	adapt ()=0
	Improve accuracy of the model on the given training data by adaptive refinement of the grid.

virtual double	computeResidual (DataMatrix &validationData) const =0
	Should compute some kind of Residual to evaluate the fit of the model.

virtual double	evaluate (const DataVector &sample)=0
	Evaluate the fitted model at a single data point.

virtual void	evaluate (DataMatrix &samples, DataVector &results)=0
	Evaluate the fitted model on a set of data points.

virtual void	fit (Dataset &dataset)=0
	Fit the grid to the dataset by determining the weights of an initial grid.

virtual void	fit (Dataset &datasetP, Dataset &datasetQ)=0

Dataset *	getDataset ()

FitterConfiguration &	getFitterConfiguration ()
	Get or set the configuration of the fitter object.

const FitterConfiguration &	getFitterConfiguration () const
	Get the configuration of the fitter object.

virtual Grid &	getGrid ()
	Return the learned grid.

virtual std::shared_ptr< BlacsProcessGrid >	getProcessGrid () const

virtual DataVector &	getSurpluses ()
	Return the learned hierarchical surpluses.

	ModelFittingBase ()
	Default constructor.

	ModelFittingBase (const ModelFittingBase &rhs)=delete
	Copy constructor - we cannot deep copy all member variables yet.

	ModelFittingBase (ModelFittingBase &&rhs)=default
	Move constructor.

ModelFittingBase &	operator= (const ModelFittingBase &rhs)=delete
	Copy assign operator - we cannot deep copy all member variables yet.

ModelFittingBase &	operator= (ModelFittingBase &&rhs)=default
	Move assign operator.

virtual void	reset ()=0
	Resets the state of the entire model.

virtual void	resetTraining ()=0
	Resets any trained representations of the model, but does not reset the entire state.

virtual void	update (Dataset &dataset)=0
	Train the grid of an existing model with new samples.

virtual void	update (Dataset &datasetP, Dataset &datasetQ)=0

virtual void	updateRegularization (double lambda)=0
	Updates the regularization parameter lambda of the underlying model.

virtual	~ModelFittingBase ()=default
	virtual destructor.

Public Attributes
bool	verboseSolver
	Whether the Solver produces output or not.

Protected Member Functions
Grid *	buildGrid (const sgpp::base::GeneralGridConfiguration &gridConfig) const
	Factory member function that generates a grid from configuration.

Grid *	buildGrid (const sgpp::base::GeneralGridConfiguration &gridConfig, const GeometryConfiguration &geometryConfig) const
	Factory member function that generates a grid from configuration.

SLESolver *	buildSolver (const SLESolverConfiguration &config) const
	Factory member function to build the solver for the least squares regression problem according to the config.

std::set< std::set< size_t > >	getInteractions (const GeometryConfiguration &geometryConfig)

void	reconfigureSolver (SLESolver &solver, const SLESolverConfiguration &config) const
	Configure solver based on the desired configuration.

Protected Attributes
std::unique_ptr< FitterConfiguration >	config
	Configuration object for the fitter.

Dataset *	dataset
	Pointer to sgpp::datadriven::Dataset.

Dataset *	extraDataset

std::unique_ptr< std::set< std::set< size_t > > >	interactions

std::unique_ptr< SLESolver >	solver
	Solver for the learning problem.

Detailed Description

Base class for arbitrary machine learning models based on adaptive sparse grids.

A model tries to generalize high dimensional training data by using sparse grids. An underlying model can be trained using training data, its accuracy can be improved by using the adaptivity of sparse grids and the underlying grid(s) of a model can be retrained on other data. Once a model is trained it can be evaluated on unseen data.

Constructor & Destructor Documentation

◆ ModelFittingBase() [1/3]

sgpp::datadriven::ModelFittingBase::ModelFittingBase ( )

Default constructor.

◆ ModelFittingBase() [2/3]

sgpp::datadriven::ModelFittingBase::ModelFittingBase ( const ModelFittingBase & rhs )

delete

Copy constructor - we cannot deep copy all member variables yet.

Parameters

rhs	const reference to the scorer object to copy from.

◆ ModelFittingBase() [3/3]

sgpp::datadriven::ModelFittingBase::ModelFittingBase ( ModelFittingBase && rhs )

default

Move constructor.

Parameters

rhs	R-value reference to a scorer object to moved from.

◆ ~ModelFittingBase()

virtual sgpp::datadriven::ModelFittingBase::~ModelFittingBase ( )

virtualdefault

virtual destructor.

Member Function Documentation

◆ adapt()

virtual bool sgpp::datadriven::ModelFittingBase::adapt ( )

pure virtual

Improve accuracy of the model on the given training data by adaptive refinement of the grid.

Returns: true if refinement was performed, else false.

Implemented in sgpp::datadriven::ModelFittingClassification, sgpp::datadriven::ModelFittingDensityEstimation, sgpp::datadriven::ModelFittingDensityEstimationCombi, sgpp::datadriven::ModelFittingDensityRatioEstimation, and sgpp::datadriven::ModelFittingLeastSquares.

◆ buildGrid() [1/2]

Grid * sgpp::datadriven::ModelFittingBase::buildGrid ( const sgpp::base::GeneralGridConfiguration & gridConfig ) const

protected

Factory member function that generates a grid from configuration.

Parameters

gridConfig configuration for the grid object

Returns: new grid object that is owned by the caller.

References sgpp::datadriven::GridFactory::createGrid(), and interactions.

Referenced by sgpp::datadriven::ModelFittingDensityEstimationCG::fit(), sgpp::datadriven::ModelFittingDensityEstimationOnOff::fit(), sgpp::datadriven::ModelFittingDensityEstimationOnOffParallel::fit(), sgpp::datadriven::ModelFittingDensityDifferenceEstimationCG::fit(), sgpp::datadriven::ModelFittingDensityDifferenceEstimationOnOff::fit(), sgpp::datadriven::ModelFittingLeastSquares::fit(), and sgpp::datadriven::ModelFittingDensityRatioEstimation::fit().

◆ buildGrid() [2/2]

Grid * sgpp::datadriven::ModelFittingBase::buildGrid	(	const sgpp::base::GeneralGridConfiguration &	gridConfig,
		const GeometryConfiguration &	geometryConfig
	)		const

protected

Factory member function that generates a grid from configuration.

Parameters

gridConfig	configuration for the grid object
geometryConfig	configuration for the geometry parameters

Returns: new grid object that is owned by the caller.

References sgpp::datadriven::GridFactory::createGrid(), sgpp::datadriven::GridFactory::getInteractions(), interactions, and sgpp::datadriven::GeometryConfiguration::stencils_.

◆ buildSolver()

SLESolver * sgpp::datadriven::ModelFittingBase::buildSolver ( const SLESolverConfiguration & config ) const

protected

Factory member function to build the solver for the least squares regression problem according to the config.

Parameters

config configuratin for the solver object

References sgpp::solver::SLESolverConfiguration::eps_, sgpp::solver::SLESolverConfiguration::maxIterations_, and sgpp::solver::SLESolverConfiguration::type_.

Referenced by sgpp::datadriven::ModelFittingDensityRatioEstimation::ModelFittingDensityRatioEstimation(), and sgpp::datadriven::ModelFittingLeastSquares::ModelFittingLeastSquares().

◆ computeResidual()

virtual double sgpp::datadriven::ModelFittingBase::computeResidual ( DataMatrix & validationData ) const

pure virtual

Should compute some kind of Residual to evaluate the fit of the model.

In the case of density estimation, this is || R * alpha_lambda - b_val ||_2

This is useful for unsupervised learning models, where normal evaluation cannot be used as there are no targets.

Parameters

validationData Matrix for validation data

Returns: the residual score

Referenced by sgpp::datadriven::ResidualScore::measure().

◆ evaluate() [1/2]

virtual double sgpp::datadriven::ModelFittingBase::evaluate ( const DataVector & sample )

pure virtual

Evaluate the fitted model at a single data point.

Parameters

sample vector with the coordinates in all dimensions of that sample.

Returns: evaluation of the model.

◆ evaluate() [2/2]

virtual void sgpp::datadriven::ModelFittingBase::evaluate	(	DataMatrix &	samples,
		DataVector &	results
	)

pure virtual

Evaluate the fitted model on a set of data points.

Parameters

samples	matrix where each row represents a sample and the columns contain the coordinates in all dimensions of that sample.
results	vector where each row will contain the evaluation of the respective sample on the current model.

◆ fit() [1/2]

virtual void sgpp::datadriven::ModelFittingBase::fit ( Dataset & dataset )

pure virtual

Fit the grid to the dataset by determining the weights of an initial grid.

Parameters

dataset the training dataset that is used to fit the model.

◆ fit() [2/2]

virtual void sgpp::datadriven::ModelFittingBase::fit	(	Dataset &	datasetP,
		Dataset &	datasetQ
	)

pure virtual

◆ getDataset()

Dataset * sgpp::datadriven::ModelFittingBase::getDataset ( )

References dataset.

Referenced by sgpp::datadriven::VisualizerClassification::getHeatmapsClassification(), sgpp::datadriven::VisualizerClassification::initializeMatrices(), and sgpp::datadriven::VisualizerDensityEstimation::initializeMatrices().

◆ getFitterConfiguration() [1/2]

FitterConfiguration & sgpp::datadriven::ModelFittingBase::getFitterConfiguration ( )

Get or set the configuration of the fitter object.

Returns: configuration of the fitter object

References config.

◆ getFitterConfiguration() [2/2]

const FitterConfiguration & sgpp::datadriven::ModelFittingBase::getFitterConfiguration ( ) const

Get the configuration of the fitter object.

Returns: configuration of the fitter object

References config.

Referenced by sgpp::datadriven::ModelFittingDensityEstimationCombi::fit(), sgpp::datadriven::VisualizerDensityEstimation::getHeatmap(), sgpp::datadriven::VisualizerDensityEstimation::getLinearCuts(), sgpp::datadriven::VisualizerClassification::runVisualization(), sgpp::datadriven::VisualizerDensityEstimation::runVisualization(), and sgpp::datadriven::Scorer::test().

◆ getGrid()

virtual Grid & sgpp::datadriven::ModelFittingBase::getGrid ( )

inlinevirtual

Return the learned grid.

Has to be rewritten to modify default behavior. Only available for single grid models. Otherwise, an error message is shown.

Reimplemented in sgpp::datadriven::ModelFittingBaseSingleGrid.

Referenced by sgpp::datadriven::ModelFittingClassification::adapt(), python.uq.analysis.asgc.ASGCKnowledge.ASGCKnowledge::getSparseGridFunction(), python.uq.learner.SimulationLearner.SimulationLearner::learnData(), python.uq.learner.SimulationLearner.SimulationLearner::learnDataWithFolding(), python.uq.learner.SimulationLearner.SimulationLearner::learnDataWithTest(), and python.uq.learner.SimulationLearner.SimulationLearner::refineGrid().

◆ getInteractions()

std::set< std::set< size_t > > sgpp::datadriven::ModelFittingBase::getInteractions ( const GeometryConfiguration & geometryConfig )

protected

References sgpp::datadriven::GridFactory::getInteractions(), interactions, and sgpp::datadriven::GeometryConfiguration::stencils_.

Referenced by sgpp::datadriven::ModelFittingDensityEstimationOnOff::fit().

◆ getProcessGrid()

virtual std::shared_ptr< BlacsProcessGrid > sgpp::datadriven::ModelFittingBase::getProcessGrid ( ) const

inlinevirtual

Returns: the BLACS process grid, useful if the fitter uses ScaLAPACK

Reimplemented in sgpp::datadriven::ModelFittingDensityEstimationOnOffParallel.

◆ getSurpluses()

virtual DataVector & sgpp::datadriven::ModelFittingBase::getSurpluses ( )

inlinevirtual

Return the learned hierarchical surpluses.

Reimplemented in sgpp::datadriven::ModelFittingBaseSingleGrid.

Referenced by sgpp::datadriven::ModelFittingClassification::adapt().

◆ operator=() [1/2]

ModelFittingBase & sgpp::datadriven::ModelFittingBase::operator= ( const ModelFittingBase & rhs )

delete

Copy assign operator - we cannot deep copy all member variables yet.

Parameters

rhs	const reference to the scorer object to copy from.

Returns: rerefernce to this with updated values.

◆ operator=() [2/2]

ModelFittingBase & sgpp::datadriven::ModelFittingBase::operator= ( ModelFittingBase && rhs )

default

Move assign operator.

Parameters

rhs	R-value reference to an a scorer object to move from.

Returns: rerefernce to this with updated values.

◆ reconfigureSolver()

void sgpp::datadriven::ModelFittingBase::reconfigureSolver	(	SLESolver &	solver,
		const SLESolverConfiguration &	config
	)		const

protected

Configure solver based on the desired configuration.

Parameters

solver	the solver object to be modified.
config	configuration updating the for the solver.

References sgpp::solver::SLESolverConfiguration::eps_, sgpp::solver::SLESolverConfiguration::maxIterations_, and solver.

◆ reset()

virtual void sgpp::datadriven::ModelFittingBase::reset ( )

pure virtual

Resets the state of the entire model.

Referenced by sgpp::datadriven::SparseGridMiner::setModel().

◆ resetTraining()

virtual void sgpp::datadriven::ModelFittingBase::resetTraining ( )

pure virtual

Resets any trained representations of the model, but does not reset the entire state.

◆ update() [1/2]

virtual void sgpp::datadriven::ModelFittingBase::update ( Dataset & dataset )

pure virtual

Train the grid of an existing model with new samples.

Parameters

dataset the training dataset that is used to fit the model.

Referenced by python.uq.refinement.RefinementStrategy.Ranking::rank(), and python.learner.LearnedKnowledge.LearnedKnowledge::setMemento().

◆ update() [2/2]

virtual void sgpp::datadriven::ModelFittingBase::update	(	Dataset &	datasetP,
		Dataset &	datasetQ
	)

pure virtual

Referenced by python.uq.refinement.RefinementStrategy.Ranking::rank(), and python.learner.LearnedKnowledge.LearnedKnowledge::setMemento().

◆ updateRegularization()

virtual void sgpp::datadriven::ModelFittingBase::updateRegularization ( double lambda )

pure virtual

Updates the regularization parameter lambda of the underlying model.

Parameters

lambda the new lambda parameter

Member Data Documentation

◆ config

std::unique_ptr<FitterConfiguration> sgpp::datadriven::ModelFittingBase::config

protected

Configuration object for the fitter.

◆ dataset

Dataset* sgpp::datadriven::ModelFittingBase::dataset

protected

Pointer to sgpp::datadriven::Dataset.

The initial grid is fitted on the given data. Adaptive refinement is then performed on the very same data. The used dataset used for refinement overwritten once either fit() or update() introduce a new dataset.

◆ extraDataset

Dataset* sgpp::datadriven::ModelFittingBase::extraDataset

protected

Referenced by sgpp::datadriven::ModelFittingDensityDifferenceEstimationCG::fit(), sgpp::datadriven::ModelFittingDensityDifferenceEstimationOnOff::fit(), sgpp::datadriven::ModelFittingDensityRatioEstimation::fit(), sgpp::datadriven::ModelFittingDensityDifferenceEstimationCG::update(), sgpp::datadriven::ModelFittingDensityDifferenceEstimationOnOff::update(), and sgpp::datadriven::ModelFittingDensityRatioEstimation::update().

◆ interactions

std::unique_ptr<std::set<std::set<size_t> > > sgpp::datadriven::ModelFittingBase::interactions

protected

Referenced by buildGrid(), buildGrid(), and getInteractions().

◆ solver

std::unique_ptr<SLESolver> sgpp::datadriven::ModelFittingBase::solver

protected

Solver for the learning problem.

Referenced by python.uq.learner.Learner.Learner::copy(), python.learner.Learner.Learner::doLearningIteration(), python.learner.Learner.Learner::fromJson(), python.uq.learner.Learner.Learner::fromJson(), sgpp::datadriven::ModelFittingDensityRatioEstimation::ModelFittingDensityRatioEstimation(), sgpp::datadriven::ModelFittingLeastSquares::ModelFittingLeastSquares(), reconfigureSolver(), python.learner.Learner.Learner::setSolver(), and python.uq.learner.Learner.Learner::setSolver().

◆ verboseSolver

bool sgpp::datadriven::ModelFittingBase::verboseSolver

Whether the Solver produces output or not.

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

datadriven/src/sgpp/datadriven/datamining/modules/fitting/ModelFittingBase.hpp
datadriven/src/sgpp/datadriven/datamining/modules/fitting/ModelFittingBase.cpp

Public Member Functions

Public Attributes

Protected Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ ModelFittingBase() [1/3]

◆ ModelFittingBase() [2/3]

◆ ModelFittingBase() [3/3]

◆ ~ModelFittingBase()

Member Function Documentation

◆ adapt()

◆ buildGrid() [1/2]

◆ buildGrid() [2/2]

◆ buildSolver()

◆ computeResidual()

◆ evaluate() [1/2]

◆ evaluate() [2/2]

◆ fit() [1/2]

◆ fit() [2/2]

◆ getDataset()

◆ getFitterConfiguration() [1/2]

◆ getFitterConfiguration() [2/2]

◆ getGrid()

◆ getInteractions()

◆ getProcessGrid()

◆ getSurpluses()

◆ operator=() [1/2]

◆ operator=() [2/2]

◆ reconfigureSolver()

◆ reset()

◆ resetTraining()

◆ update() [1/2]

◆ update() [2/2]

◆ updateRegularization()

Member Data Documentation

◆ config

◆ dataset

◆ extraDataset

◆ interactions

◆ solver

◆ verboseSolver