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

RMSProp Optimizer class.

This class represents the RMSProp Optimizer with options for applying momentum.

Definition at line 44 of file RMSProp.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, VGeneralLayer< Architecture_t >, TDeepNet< Architecture_t, VGeneralLayer< Architecture_t > > >
using	Matrix_t = typename Architecture_t::Matrix_t

using	Scalar_t = typename Architecture_t::Scalar_t

Public Member Functions
	TRMSProp (DeepNet_t &deepNet, Scalar_t learningRate=0.001, Scalar_t momentum=0.0, Scalar_t rho=0.9, Scalar_t epsilon=1e-7)
	Constructor. More...

	~TRMSProp ()=default
	Destructor. More...

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

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

Scalar_t	GetEpsilon () const

Scalar_t	GetMomentum () const
	Getters. More...

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)

Scalar_t	GetRho () const

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

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

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

virtual	~VOptimizer ()=default
	Virtual Destructor. More...

size_t	GetGlobalStep () const

VGeneralLayer< Architecture_t > *	GetLayerAt (size_t i)

std::vector< VGeneralLayer< Architecture_t > * > &	GetLayers ()

Scalar_t	GetLearningRate () const
	Getters. More...

void	IncrementGlobalStep ()
	Increments the global step. More...

void	SetLearningRate (size_t learningRate)
	Setters. More...

void	Step ()
	Performs one step of optimization. More...

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

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

virtual void	UpdateBiases (size_t layerIndex, std::vector< Matrix_t > &biases, const std::vector< Matrix_t > &biasGradients)=0
	Update the biases, given the current bias gradients. More...

virtual void	UpdateWeights (size_t layerIndex, std::vector< Matrix_t > &weights, const std::vector< Matrix_t > &weightGradients)=0
	Update the weights, given the current weight gradients. More...

Protected Attributes
std::vector< std::vector< Matrix_t > >	fBiasUpdates
	The accumulation of the past Biases for performing updates. More...

Scalar_t	fEpsilon
	The Smoothing term used to avoid division by zero. More...

Scalar_t	fMomentum
	The momentum used for training. More...

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

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

Scalar_t	fRho
	The Rho constant used by the optimizer. More...

std::vector< std::vector< Matrix_t > >	fWeightUpdates
	The accumulation of the past Weights for performing updates. More...

std::vector< std::vector< Matrix_t > >	fWorkBiasTensor1
	working tensor used to keep a temporary copy of bias or bias gradients More...

std::vector< std::vector< Matrix_t > >	fWorkBiasTensor2
	working tensor used to keep a temporary copy of bias or bias gradients More...

std::vector< std::vector< Matrix_t > >	fWorkWeightTensor1
	working tensor used to keep a temporary copy of weights or weight gradients More...

std::vector< std::vector< Matrix_t > >	fWorkWeightTensor2
	working tensor used to keep a temporary copy of weights or weight gradients More...

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

size_t	fGlobalStep
	The current global step count during training. More...

Scalar_t	fLearningRate
	The learning rate used for training. More...

#include <TMVA/DNN/RMSProp.h>

Inheritance diagram for TMVA::DNN::TRMSProp< 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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::Matrix_t = typename Architecture_t::Matrix_t

Definition at line 46 of file RMSProp.h.

◆ Scalar_t

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

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

Definition at line 47 of file RMSProp.h.

Constructor & Destructor Documentation

◆ TRMSProp()

template<typename Architecture_t , typename Layer_t , typename DeepNet_t >

TMVA::DNN::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::TRMSProp	(	DeepNet_t &	deepNet,
		Scalar_t	learningRate = `0.001`,
		Scalar_t	momentum = `0.0`,
		Scalar_t	rho = `0.9`,
		Scalar_t	epsilon = `1e-7`
	)

Constructor.

Definition at line 106 of file RMSProp.h.

◆ ~TRMSProp()

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

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

default

Destructor.

Member Function Documentation

◆ GetBiasUpdates()

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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::GetBiasUpdates ( )

inline

Definition at line 97 of file RMSProp.h.

◆ GetBiasUpdatesAt()

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

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

inline

Definition at line 98 of file RMSProp.h.

◆ GetEpsilon()

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

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

inline

Definition at line 86 of file RMSProp.h.

◆ GetMomentum()

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

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

inline

Getters.

Definition at line 84 of file RMSProp.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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::GetPastSquaredBiasGradients ( )

inline

Definition at line 91 of file RMSProp.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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::GetPastSquaredBiasGradientsAt ( size_t i )

inline

Definition at line 92 of file RMSProp.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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::GetPastSquaredWeightGradients ( )

inline

Definition at line 88 of file RMSProp.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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::GetPastSquaredWeightGradientsAt ( size_t i )

inline

Definition at line 89 of file RMSProp.h.

◆ GetRho()

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

Scalar_t TMVA::DNN::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::GetRho ( ) const

inline

Definition at line 85 of file RMSProp.h.

◆ GetWeightUpdates()

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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::GetWeightUpdates ( )

inline

Definition at line 94 of file RMSProp.h.

◆ GetWeightUpdatesAt()

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

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

inline

Definition at line 95 of file RMSProp.h.

◆ UpdateBiases()

template<typename Architecture_t , typename Layer_t , typename DeepNet_t >

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

protectedvirtual

Update the biases, given the current bias gradients.

Implements TMVA::DNN::VOptimizer< Architecture_t, VGeneralLayer< Architecture_t >, TDeepNet< Architecture_t, VGeneralLayer< Architecture_t > > >.

Definition at line 195 of file RMSProp.h.

◆ UpdateWeights()

template<typename Architecture_t , typename Layer_t , typename DeepNet_t >

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

protectedvirtual

Update the weights, given the current weight gradients.

Implements TMVA::DNN::VOptimizer< Architecture_t, VGeneralLayer< Architecture_t >, TDeepNet< Architecture_t, VGeneralLayer< Architecture_t > > >.

Definition at line 151 of file RMSProp.h.

Member Data Documentation

◆ fBiasUpdates

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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::fBiasUpdates

protected

The accumulation of the past Biases for performing updates.

Definition at line 59 of file RMSProp.h.

◆ fEpsilon

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

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

protected

The Smoothing term used to avoid division by zero.

Definition at line 52 of file RMSProp.h.

◆ fMomentum

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

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

protected

The momentum used for training.

Definition at line 50 of file RMSProp.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::TRMSProp< 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 RMSProp.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::TRMSProp< 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 RMSProp.h.

◆ fRho

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

Scalar_t TMVA::DNN::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::fRho

protected

The Rho constant used by the optimizer.

Definition at line 51 of file RMSProp.h.

◆ fWeightUpdates

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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::fWeightUpdates

protected

The accumulation of the past Weights for performing updates.

Definition at line 58 of file RMSProp.h.

◆ fWorkBiasTensor1

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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::fWorkBiasTensor1

protected

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

Definition at line 63 of file RMSProp.h.

◆ fWorkBiasTensor2

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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::fWorkBiasTensor2

protected

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

Definition at line 67 of file RMSProp.h.

◆ fWorkWeightTensor1

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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::fWorkWeightTensor1

protected

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

Definition at line 61 of file RMSProp.h.

◆ fWorkWeightTensor2

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::TRMSProp< Architecture_t, Layer_t, DeepNet_t >::fWorkWeightTensor2

protected

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

Definition at line 65 of file RMSProp.h.

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

tmva/tmva/inc/TMVA/DNN/RMSProp.h

Public Types

Public Member Functions

Protected Member Functions

Protected Attributes

Member Typedef Documentation

◆ Matrix_t

◆ Scalar_t

Constructor & Destructor Documentation

◆ TRMSProp()

◆ ~TRMSProp()

Member Function Documentation

◆ GetBiasUpdates()

◆ GetBiasUpdatesAt()

◆ GetEpsilon()

◆ GetMomentum()

◆ GetPastSquaredBiasGradients()

◆ GetPastSquaredBiasGradientsAt()

◆ GetPastSquaredWeightGradients()

◆ GetPastSquaredWeightGradientsAt()

◆ GetRho()

◆ GetWeightUpdates()

◆ GetWeightUpdatesAt()

◆ UpdateBiases()

◆ UpdateWeights()

Member Data Documentation

◆ fBiasUpdates

◆ fEpsilon

◆ fMomentum

◆ fPastSquaredBiasGradients

◆ fPastSquaredWeightGradients

◆ fRho

◆ fWeightUpdates

◆ fWorkBiasTensor1

◆ fWorkBiasTensor2

◆ fWorkWeightTensor1

◆ fWorkWeightTensor2