sgpp::datadriven::LearnerSGD Class Reference

LearnerSGD learns the data using stochastic gradient descent. More...

#include <LearnerSGD.hpp>

Public Member Functions
double	getAccuracy (sgpp::base::DataMatrix &testData, sgpp::base::DataVector &testLabels, double threshold)
	Computes the classification accuracy on the given dataset. More...
void	initialize ()
	Initializes the SGD learner (creates grid etc.). More...
	LearnerSGD (base::RegularGridConfiguration &gridConfig, base::AdaptivityConfiguration &adaptivityConfig, base::DataMatrix &pTrainData, base::DataVector &pTrainLabels, base::DataMatrix &pTestData, base::DataVector &pTestLabels, base::DataMatrix *pValData, base::DataVector *pValLabels, double lambda, double gamma, size_t batchSize, bool useValidData)
	Constructor. More...
void	storeResults (base::DataMatrix &testDataset)
	Stores classified data, grids and function evaluations to csv files. More...
void	train (size_t maxDataPasses, std::string refType, std::string refMonitor, size_t refPeriod, double errorDeclineThreshold, size_t errorDeclineBufferSize, size_t minRefInterval)
	Implements online learning using stochastic gradient descent. More...
	~LearnerSGD ()
	Destructor. More...

Public Attributes
sgpp::base::DataVector	avgErrors
double	error

Protected Member Functions
std::unique_ptr< base::Grid >	createRegularGrid ()
	Generates a regular grid. More...
double	getAccuracy (sgpp::base::DataVector &testLabels, double threshold, sgpp::base::DataVector &predictedLabels)
	Computes the classification accuracy. More...
void	getBatchError (sgpp::base::DataMatrix &data, const sgpp::base::DataVector &labels)
	Computes error contribution for each data point of the given data set (required for predictive refinement indicator). More...
double	getError (sgpp::base::DataMatrix &data, sgpp::base::DataVector &labels, std::string errorType)
	Computes specified error type (e.g. More...
void	predict (base::DataMatrix &testData, base::DataVector &predictedLabels)
	Predicts class labels based on the trained model. More...
void	pushToBatch (sgpp::base::DataVector &x, double y)
	Stores the last 'batchSize' processed data points if no validation data is provided. More...

Protected Attributes
base::AdaptivityConfiguration	adaptivityConfig
base::DataVector	alpha
base::DataVector	alphaAvg
base::DataMatrix *	batchData
base::DataVector	batchError
base::DataVector *	batchLabels
size_t	batchSize
double	currentGamma
double	gamma
std::unique_ptr< base::Grid >	grid
base::RegularGridConfiguration	gridConfig
double	lambda
base::DataMatrix &	testData
base::DataVector &	testLabels
base::DataMatrix &	trainData
base::DataVector &	trainLabels
bool	useValidData

Detailed Description

LearnerSGD learns the data using stochastic gradient descent.

Constructor & Destructor Documentation

LearnerSGD()

sgpp::datadriven::LearnerSGD::LearnerSGD	(	base::RegularGridConfiguration &	gridConfig,
		base::AdaptivityConfiguration &	adaptivityConfig,
		base::DataMatrix &	pTrainData,
		base::DataVector &	pTrainLabels,
		base::DataMatrix &	pTestData,
		base::DataVector &	pTestLabels,
		base::DataMatrix *	pValData,
		base::DataVector *	pValLabels,
		double	lambda,
		double	gamma,
		size_t	batchSize,
		bool	useValidData
	)

Constructor.

Parameters

gridConfig	The grid configuration
adaptivityConfig	The refinement configuration
pTrainData	The training dataset
pTrainLabels	The corresponding training labels
pTestData	The test dataset
pTestLabels	The corresponding test labels
pValData	The validation dataset
pValLabels	The corresponding validation labels
lambda	The regularization parameter
gamma	The learning parameter (i.e. step width)
batchSize	The number of data points which are considered to compute the error contributions for predictive refinement
useValidData	Specifies if validation data should be used for all error computations

References batchData, batchLabels, sgpp::base::DataMatrix::getNcols(), sgpp::base::DataMatrix::reserveAdditionalRows(), sgpp::base::DataVector::setAll(), and trainData.

~LearnerSGD()

sgpp::datadriven::LearnerSGD::~LearnerSGD

(

)

Destructor.

Member Function Documentation

createRegularGrid()

std::unique_ptr< base::Grid > sgpp::datadriven::LearnerSGD::createRegularGrid ( )

protected

Generates a regular grid.

Returns

The created grid

References sgpp::base::Grid::createLinearGrid(), sgpp::base::Grid::createModLinearGrid(), sgpp::base::GeneralGridConfiguration::dim_, gridConfig, sgpp::base::GeneralGridConfiguration::level_, sgpp::base::Linear, sgpp::base::ModLinear, and sgpp::base::GeneralGridConfiguration::type_.

Referenced by initialize().

getAccuracy() [1/2]

double sgpp::datadriven::LearnerSGD::getAccuracy	(	sgpp::base::DataMatrix &	testData,
		sgpp::base::DataVector &	testLabels,
		double	threshold
	)

Computes the classification accuracy on the given dataset.

Parameters

testData	The data for which class labels should be predicted
testLabels	The corresponding actual class labels
threshold	The decision threshold (e.g. for class labels -1, 1 -> threshold = 0)

Returns

The resulting accuracy

References sgpp::base::DataMatrix::getNrows(), and predict().

Referenced by train().

getAccuracy() [2/2]

double sgpp::datadriven::LearnerSGD::getAccuracy ( sgpp::base::DataVector &  testLabels, double  threshold, sgpp::base::DataVector &  predictedLabels  )

protected

Computes the classification accuracy.

Parameters

testLabels	The actual class labels
threshold	The decision threshold (e.g. for class labels -1, 1 -> threshold = 0)
predictedLabels	The predicted class labels

Returns

The resulting accuracy

References sgpp::base::DataVector::get(), sgpp::base::DataVector::getSize(), and python.statsfileInfo::i.

getBatchError()

void sgpp::datadriven::LearnerSGD::getBatchError ( sgpp::base::DataMatrix &  data, const sgpp::base::DataVector &  labels  )

protected

Computes error contribution for each data point of the given data set (required for predictive refinement indicator).

Parameters

data	The data points
labels	The corresponding class labels

References alphaAvg, batchError, sgpp::op_factory::createOperationMultipleEval(), sgpp::base::DataVector::get(), sgpp::base::DataMatrix::getNrows(), grid, python.statsfileInfo::i, sgpp::combigrid::pow(), and sgpp::base::DataVector::set().

Referenced by train().

getError()

double sgpp::datadriven::LearnerSGD::getError ( sgpp::base::DataMatrix &  data, sgpp::base::DataVector &  labels, std::string  errorType  )

protected

Computes specified error type (e.g.

MSE).

Parameters

data	The data points
labels	The corresponding class labels
errorType	The type of the error measurement (MSE or Hinge loss)

Returns

The error estimation

References alphaAvg, sgpp::op_factory::createOperationMultipleEval(), error, sgpp::base::DataVector::get(), sgpp::base::DataMatrix::getNrows(), grid, python.statsfileInfo::i, sgpp::base::DataVector::set(), and sgpp::base::DataVector::setAll().

Referenced by train().

initialize()

void sgpp::datadriven::LearnerSGD::initialize

(

)

Initializes the SGD learner (creates grid etc.).

References alpha, alphaAvg, batchError, batchSize, createRegularGrid(), sgpp::base::GeneralGridConfiguration::dim_, sgpp::base::DataMatrix::getNcols(), grid, gridConfig, and trainData.

predict()

void sgpp::datadriven::LearnerSGD::predict ( base::DataMatrix &  testData, base::DataVector &  predictedLabels  )

protected

Predicts class labels based on the trained model.

Parameters

testData	The data for which class labels should be predicted
predictedLabels	The predicted class labels

References alphaAvg, sgpp::op_factory::createOperationMultipleEval(), sgpp::base::DataMatrix::getNrows(), grid, python.statsfileInfo::i, and sgpp::base::DataVector::set().

Referenced by getAccuracy(), storeResults(), and train().

pushToBatch()

void sgpp::datadriven::LearnerSGD::pushToBatch ( sgpp::base::DataVector &  x, double  y  )

protected

Stores the last 'batchSize' processed data points if no validation data is provided.

Parameters

x	The current data point
y	The corresponding class label

References sgpp::base::DataMatrix::appendRow(), batchData, batchSize, sgpp::base::DataMatrix::getNrows(), and sgpp::base::DataMatrix::setRow().

Referenced by train().

storeResults()

void sgpp::datadriven::LearnerSGD::storeResults

(

base::DataMatrix &

testDataset

)

Stores classified data, grids and function evaluations to csv files.

Parameters

testDataset

Data points for which the model is evaluated

References alphaAvg, sgpp::base::DataMatrix::appendRow(), sgpp::base::HashGridStorage::begin(), sgpp::op_factory::createOperationEval(), sgpp::base::HashGridStorage::end(), sgpp::base::DataVector::get(), sgpp::base::HashGridStorage::getCoordinates(), sgpp::base::DataMatrix::getNcols(), sgpp::base::DataMatrix::getNrows(), sgpp::base::DataMatrix::getRow(), sgpp::base::DataVector::getSize(), grid, python.statsfileInfo::i, python.utils.statsfile2gnuplot::j, predict(), and sgpp::base::DataVector::set().

train()

void sgpp::datadriven::LearnerSGD::train	(	size_t	maxDataPasses,
		std::string	refType,
		std::string	refMonitor,
		size_t	refPeriod,
		double	errorDeclineThreshold,
		size_t	errorDeclineBufferSize,
		size_t	minRefInterval
	)

Implements online learning using stochastic gradient descent.

Parameters

maxDataPasses	The number of passes over the whole training data
refType	The refinement indicator (surplus, zero-crossings or data-based)
refMonitor	The refinement strategy (periodic or convergence-based)
refPeriod	The refinement interval (if periodic refinement is chosen)
errorDeclineThreshold	The convergence threshold (if convergence-based refinement is chosen)
errorDeclineBufferSize	The number of error measurements which are used to check convergence (if convergence-based refinement is chosen)
minRefInterval	The minimum number of data points which have to be processed before next refinement can be scheduled (if convergence-based refinement is chosen)

References adaptivityConfig, alpha, alphaAvg, sgpp::base::DataVector::append(), avgErrors, sgpp::base::DataVector::axpy(), batchData, batchError, batchLabels, sgpp::op_factory::createOperationMultipleEval(), currentGamma, chess::dim, error, sgpp::base::ImpurityRefinement::free_refine(), sgpp::base::PredictiveRefinement::free_refine(), gamma, sgpp::base::DataVector::get(), getAccuracy(), getBatchError(), getError(), sgpp::base::DataMatrix::getNcols(), sgpp::base::DataMatrix::getNrows(), sgpp::base::DataVector::getPointer(), sgpp::base::DataMatrix::getRow(), sgpp::base::DataVector::getSize(), grid, python.utils.sg_projections::gridStorage, lambda, mu, sgpp::base::DataVector::mult(), sgpp::base::AdaptivityConfiguration::noPoints_, sgpp::base::AdaptivityConfiguration::numRefinements_, sgpp::combigrid::pow(), predict(), pushToBatch(), sgpp::datadriven::RefinementMonitor::pushToBuffer(), sgpp::datadriven::RefinementMonitor::refinementsNecessary(), sgpp::base::DataVector::resizeZero(), testData, testLabels, sgpp::base::AdaptivityConfiguration::threshold_, trainData, trainLabels, and useValidData.

Member Data Documentation

adaptivityConfig

base::AdaptivityConfiguration sgpp::datadriven::LearnerSGD::adaptivityConfig

protected

Referenced by train().

alpha

base::DataVector sgpp::datadriven::LearnerSGD::alpha

protected

Referenced by initialize(), python.learner.Classifier.Classifier::refineGrid(), and train().

alphaAvg

base::DataVector sgpp::datadriven::LearnerSGD::alphaAvg

protected

Referenced by getBatchError(), getError(), initialize(), predict(), storeResults(), and train().

avgErrors

sgpp::base::DataVector sgpp::datadriven::LearnerSGD::avgErrors

Referenced by train().

batchData

base::DataMatrix* sgpp::datadriven::LearnerSGD::batchData

protected

Referenced by LearnerSGD(), pushToBatch(), and train().

batchError

base::DataVector sgpp::datadriven::LearnerSGD::batchError

protected

Referenced by getBatchError(), initialize(), and train().

batchLabels

base::DataVector* sgpp::datadriven::LearnerSGD::batchLabels

protected

Referenced by LearnerSGD(), and train().

batchSize

size_t sgpp::datadriven::LearnerSGD::batchSize

protected

Referenced by initialize(), and pushToBatch().

currentGamma

double sgpp::datadriven::LearnerSGD::currentGamma

protected

Referenced by train().

error

double sgpp::datadriven::LearnerSGD::error

Referenced by getError(), and train().

gamma

double sgpp::datadriven::LearnerSGD::gamma

protected

Referenced by train().

grid

std::unique_ptr<base::Grid> sgpp::datadriven::LearnerSGD::grid

protected

Referenced by python.uq.learner.Interpolant.Interpolant::doLearningIteration(), python.learner.Classifier.Classifier::evalError(), python.uq.learner.Interpolant.Interpolant::evalError(), getBatchError(), python.uq.learner.SimulationLearner.SimulationLearner::getCollocationNodes(), getError(), python.uq.learner.SimulationLearner.SimulationLearner::getGrid(), python.uq.learner.SimulationLearner.SimulationLearner::getLearner(), initialize(), python.uq.learner.Regressor.Regressor::learnData(), python.uq.learner.Regressor.Regressor::learnDataWithFolding(), python.uq.learner.Regressor.Regressor::learnDataWithTest(), predict(), python.learner.Classifier.Classifier::refineGrid(), python.learner.Regressor.Regressor::refineGrid(), python.uq.learner.Regressor.Regressor::refineGrid(), python.uq.learner.SimulationLearner.SimulationLearner::refineGrid(), storeResults(), train(), python.learner.Classifier.Classifier::updateResults(), python.learner.Regressor.Regressor::updateResults(), and python.uq.learner.Regressor.Regressor::updateResults().

gridConfig

base::RegularGridConfiguration sgpp::datadriven::LearnerSGD::gridConfig

protected

Referenced by createRegularGrid(), and initialize().

lambda

double sgpp::datadriven::LearnerSGD::lambda

protected

Referenced by train().

testData

base::DataMatrix& sgpp::datadriven::LearnerSGD::testData

protected

Referenced by train().

testLabels

base::DataVector& sgpp::datadriven::LearnerSGD::testLabels

protected

Referenced by train().

trainData

base::DataMatrix& sgpp::datadriven::LearnerSGD::trainData

protected

Referenced by python.uq.dists.KDEDist.KDEDist::__init__(), initialize(), LearnerSGD(), python.uq.dists.KDEDist.KDEDist::marginalize(), python.uq.dists.KDEDist.KDEDist::marginalizeToDimX(), python.uq.dists.KDEDist.KDEDist::toJson(), and train().

trainLabels

base::DataVector& sgpp::datadriven::LearnerSGD::trainLabels

protected

Referenced by train().

useValidData

bool sgpp::datadriven::LearnerSGD::useValidData

protected

Referenced by train().

---

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

• datadriven/src/sgpp/datadriven/application/LearnerSGD.hpp
• datadriven/src/sgpp/datadriven/application/LearnerSGD.cpp