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

Adagrad Optimizer class.

This class represents the Adagrad Optimizer.

Definition at line 45 of file Adagrad.h.

Public Types
using	Matrix_t = typename Architecture_t::Matrix_t
using	Scalar_t = typename Architecture_t::Scalar_t

Public Member Functions
	TAdagrad (DeepNet_t &deepNet, Scalar_t learningRate=0.01, Scalar_t epsilon=1e-8)
	Constructor.
	~TAdagrad ()=default
	Destructor.
Scalar_t	GetEpsilon () const
	Getters.
std::vector< std::vector< Matrix_t > > &	GetPastSquaredBiasGradients ()
std::vector< Matrix_t > &	GetPastSquaredBiasGradientsAt (size_t i)
std::vector< std::vector< Matrix_t > > &	GetPastSquaredWeightGradients ()
std::vector< Matrix_t > &	GetPastSquaredWeightGradientsAt (size_t i)

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	fEpsilon
	The Smoothing term used to avoid division by zero.
std::vector< std::vector< Matrix_t > >	fPastSquaredBiasGradients
	The sum of the square of the past bias gradients associated with the deep net.
std::vector< std::vector< Matrix_t > >	fPastSquaredWeightGradients
	The sum of the square of the past weight gradients associated with the deep net.
std::vector< std::vector< Matrix_t > >	fWorkBiasTensor
	working tensor used to keep a temporary copy of bias or bias gradients
std::vector< std::vector< Matrix_t > >	fWorkWeightTensor
	working tensor used to keep a temporary copy of weights or weight gradients

#include <TMVA/DNN/Adagrad.h>

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

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::TAdagrad< Architecture_t, Layer_t, DeepNet_t >::Matrix_t = typename Architecture_t::Matrix_t

Definition at line 47 of file Adagrad.h.

◆ Scalar_t

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

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

Definition at line 48 of file Adagrad.h.

Constructor & Destructor Documentation

◆ TAdagrad()

template<typename Architecture_t, typename Layer_t, typename DeepNet_t>

TMVA::DNN::TAdagrad< Architecture_t, Layer_t, DeepNet_t >::TAdagrad	(	DeepNet_t &	deepNet,
		Scalar_t	learningRate = 0.01,
		Scalar_t	epsilon = 1e-8 )

Constructor.

Definition at line 90 of file Adagrad.h.

◆ ~TAdagrad()

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

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

default

Destructor.

Member Function Documentation

◆ GetEpsilon()

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

Scalar_t TMVA::DNN::TAdagrad< Architecture_t, Layer_t, DeepNet_t >::GetEpsilon ( ) const

inline

Getters.

Definition at line 76 of file Adagrad.h.

◆ GetPastSquaredBiasGradients()

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::TAdagrad< Architecture_t, Layer_t, DeepNet_t >::GetPastSquaredBiasGradients ( )

inline

Definition at line 81 of file Adagrad.h.

◆ GetPastSquaredBiasGradientsAt()

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

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

inline

Definition at line 82 of file Adagrad.h.

◆ GetPastSquaredWeightGradients()

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::TAdagrad< Architecture_t, Layer_t, DeepNet_t >::GetPastSquaredWeightGradients ( )

inline

Definition at line 78 of file Adagrad.h.

◆ GetPastSquaredWeightGradientsAt()

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

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

inline

Definition at line 79 of file Adagrad.h.

◆ UpdateBiases()

template<typename Architecture_t, typename Layer_t, typename DeepNet_t>

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

overrideprotected

Update the biases, given the current bias gradients.

Definition at line 158 of file Adagrad.h.

◆ UpdateWeights()

template<typename Architecture_t, typename Layer_t, typename DeepNet_t>

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

overrideprotected

Update the weights, given the current weight gradients.

Definition at line 126 of file Adagrad.h.

Member Data Documentation

◆ fEpsilon

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

Scalar_t TMVA::DNN::TAdagrad< Architecture_t, Layer_t, DeepNet_t >::fEpsilon

protected

The Smoothing term used to avoid division by zero.

Definition at line 51 of file Adagrad.h.

◆ fPastSquaredBiasGradients

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::TAdagrad< Architecture_t, Layer_t, DeepNet_t >::fPastSquaredBiasGradients

protected

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

Definition at line 56 of file Adagrad.h.

◆ fPastSquaredWeightGradients

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::TAdagrad< Architecture_t, Layer_t, DeepNet_t >::fPastSquaredWeightGradients

protected

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

Definition at line 54 of file Adagrad.h.

◆ fWorkBiasTensor

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::TAdagrad< Architecture_t, Layer_t, DeepNet_t >::fWorkBiasTensor

protected

working tensor used to keep a temporary copy of bias or bias gradients

Definition at line 60 of file Adagrad.h.

◆ fWorkWeightTensor

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::TAdagrad< Architecture_t, Layer_t, DeepNet_t >::fWorkWeightTensor

protected

working tensor used to keep a temporary copy of weights or weight gradients

Definition at line 58 of file Adagrad.h.

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

tmva/tmva/inc/TMVA/DNN/Adagrad.h

Public Types

Public Member Functions

Protected Member Functions

Protected Attributes

Member Typedef Documentation

◆ Matrix_t

◆ Scalar_t

Constructor & Destructor Documentation

◆ TAdagrad()

◆ ~TAdagrad()

Member Function Documentation

◆ GetEpsilon()

◆ GetPastSquaredBiasGradients()

◆ GetPastSquaredBiasGradientsAt()

◆ GetPastSquaredWeightGradients()

◆ GetPastSquaredWeightGradientsAt()

◆ UpdateBiases()

◆ UpdateWeights()

Member Data Documentation

◆ fEpsilon

◆ fPastSquaredBiasGradients

◆ fPastSquaredWeightGradients

◆ fWorkBiasTensor

◆ fWorkWeightTensor