template<typename Architecture_t, typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>
class TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >

Stochastic Batch Gradient Descent Optimizer class.

This class represents the Stochastic Batch Gradient Descent Optimizer with options for applying momentum and nesterov momentum.

Definition at line 46 of file SGD.h.

Public Types
using	Matrix_t = typename Architecture_t::Matrix_t

using	Scalar_t = typename Architecture_t::Scalar_t

Public Types inherited from TMVA::DNN::VOptimizer< Architecture_t, Layer_t, DeepNet_t >
using	Matrix_t = typename Architecture_t::Matrix_t

using	Scalar_t = typename Architecture_t::Scalar_t

Public Member Functions
	TSGD (Scalar_t learningRate, DeepNet_t &deepNet, Scalar_t momentum)
	Constructor.

	~TSGD ()=default
	Destructor.

Scalar_t	GetMomentum () const
	Getters.

std::vector< std::vector< Matrix_t > > &	GetPastBiasGradients ()

std::vector< Matrix_t > &	GetPastBiasGradientsAt (size_t i)

std::vector< std::vector< Matrix_t > > &	GetPastWeightGradients ()

std::vector< Matrix_t > &	GetPastWeightGradientsAt (size_t i)

Public Member Functions inherited from TMVA::DNN::VOptimizer< Architecture_t, Layer_t, DeepNet_t >
	VOptimizer (Scalar_t learningRate, DeepNet_t &deepNet)
	Constructor.

virtual	~VOptimizer ()=default
	Virtual Destructor.

size_t	GetGlobalStep () const

Layer_t *	GetLayerAt (size_t i)

std::vector< Layer_t * > &	GetLayers ()

Scalar_t	GetLearningRate () const
	Getters.

void	IncrementGlobalStep ()
	Increments the global step.

void	SetLearningRate (size_t learningRate)
	Setters.

void	Step ()
	Performs one step of optimization.

Protected Member Functions
void	UpdateBiases (size_t layerIndex, std::vector< Matrix_t > &biases, const std::vector< Matrix_t > &biasGradients) override
	Update the biases, given the current bias gradients.

void	UpdateWeights (size_t layerIndex, std::vector< Matrix_t > &weights, const std::vector< Matrix_t > &weightGradients) override
	Update the weights, given the current weight gradients.

Protected Attributes
Scalar_t	fMomentum
	The momentum used for training.

std::vector< std::vector< Matrix_t > >	fPastBiasGradients
	The sum of the past bias gradients associated with the deep net.

std::vector< std::vector< Matrix_t > >	fPastWeightGradients
	The sum of the past weight gradients associated with the deep net.

Protected Attributes inherited from TMVA::DNN::VOptimizer< Architecture_t, Layer_t, DeepNet_t >
DeepNet_t &	fDeepNet
	The reference to the deep net.

size_t	fGlobalStep
	The current global step count during training.

Scalar_t	fLearningRate
	The learning rate used for training.

#include <TMVA/DNN/SGD.h>

Inheritance diagram for TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >:

[legend]

Member Typedef Documentation

◆ Matrix_t

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

using TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::Matrix_t = typename Architecture_t::Matrix_t

Definition at line 48 of file SGD.h.

◆ Scalar_t

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

using TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::Scalar_t = typename Architecture_t::Scalar_t

Definition at line 49 of file SGD.h.

Constructor & Destructor Documentation

◆ TSGD()

template<typename Architecture_t , typename Layer_t , typename DeepNet_t >

TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::TSGD	(	Scalar_t	learningRate,
		DeepNet_t &	deepNet,
		Scalar_t	momentum )

Constructor.

Definition at line 86 of file SGD.h.

◆ ~TSGD()

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::~TSGD ( )

default

Destructor.

Member Function Documentation

◆ GetMomentum()

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

Scalar_t TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::GetMomentum ( ) const

inline

Getters.

Definition at line 72 of file SGD.h.

◆ GetPastBiasGradients()

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

std::vector< std::vector< Matrix_t > > & TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::GetPastBiasGradients ( )

inline

Definition at line 77 of file SGD.h.

◆ GetPastBiasGradientsAt()

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

std::vector< Matrix_t > & TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::GetPastBiasGradientsAt ( size_t i )

inline

Definition at line 78 of file SGD.h.

◆ GetPastWeightGradients()

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

std::vector< std::vector< Matrix_t > > & TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::GetPastWeightGradients ( )

inline

Definition at line 74 of file SGD.h.

◆ GetPastWeightGradientsAt()

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

std::vector< Matrix_t > & TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::GetPastWeightGradientsAt ( size_t i )

inline

Definition at line 75 of file SGD.h.

◆ UpdateBiases()

template<typename Architecture_t , typename Layer_t , typename DeepNet_t >

auto TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::UpdateBiases	(	size_t	layerIndex,
		std::vector< Matrix_t > &	biases,
		const std::vector< Matrix_t > &	biasGradients )

overrideprotectedvirtual

Update the biases, given the current bias gradients.

Implements TMVA::DNN::VOptimizer< Architecture_t, Layer_t, DeepNet_t >.

Definition at line 136 of file SGD.h.

◆ UpdateWeights()

template<typename Architecture_t , typename Layer_t , typename DeepNet_t >

auto TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::UpdateWeights	(	size_t	layerIndex,
		std::vector< Matrix_t > &	weights,
		const std::vector< Matrix_t > &	weightGradients )

overrideprotectedvirtual

Update the weights, given the current weight gradients.

Implements TMVA::DNN::VOptimizer< Architecture_t, Layer_t, DeepNet_t >.

Definition at line 114 of file SGD.h.

Member Data Documentation

◆ fMomentum

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

Scalar_t TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::fMomentum

protected

The momentum used for training.

Definition at line 52 of file SGD.h.

◆ fPastBiasGradients

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

std::vector<std::vector<Matrix_t> > TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::fPastBiasGradients

protected

The sum of the past bias gradients associated with the deep net.

Definition at line 56 of file SGD.h.

◆ fPastWeightGradients

template<typename Architecture_t , typename Layer_t = VGeneralLayer<Architecture_t>, typename DeepNet_t = TDeepNet<Architecture_t, Layer_t>>

std::vector<std::vector<Matrix_t> > TMVA::DNN::TSGD< Architecture_t, Layer_t, DeepNet_t >::fPastWeightGradients

protected

The sum of the past weight gradients associated with the deep net.

Definition at line 54 of file SGD.h.

tmva/tmva/inc/TMVA/DNN/SGD.h

Public Types

Public Member Functions

Protected Member Functions

Protected Attributes

Member Typedef Documentation

◆ Matrix_t

◆ Scalar_t

Constructor & Destructor Documentation

◆ TSGD()

◆ ~TSGD()

Member Function Documentation

◆ GetMomentum()

◆ GetPastBiasGradients()

◆ GetPastBiasGradientsAt()

◆ GetPastWeightGradients()

◆ GetPastWeightGradientsAt()

◆ UpdateBiases()

◆ UpdateWeights()

Member Data Documentation

◆ fMomentum

◆ fPastBiasGradients

◆ fPastWeightGradients