opennn::BackPropagationLM Struct Reference

Scratch state used by LevenbergMarquardtAlgorithm. More...

#include <levenberg_marquardt_algorithm.h>

Public Member Functions
	BackPropagationLM (const Index samples_number=0, Loss *loss=nullptr)
	Constructs the scratch state.
virtual	~BackPropagationLM ()=default
	Virtual destructor.
void	set (const Index samples_number=0, Loss *loss=nullptr)
	(Re)allocates buffers for the given dataset.

Public Attributes
Index	samples_number = 0
	Number of samples this scratch is sized for.
VectorR	output_deltas
	Per-sample output deltas.
Shape	output_delta_dimensions
	Shape of output_deltas as a multi-dimensional view.
Loss *	loss = nullptr
	Loss to which this scratch is bound; not owned.
float	error
	Mean error over the batch.
float	regularization = 0.0f
	Regularization contribution to the loss.
float	loss_value = 0.0f
	Total loss = error + regularization.
VectorR	errors
	Per-sample residual errors.
VectorR	squared_errors
	Per-sample squared errors.
MatrixR	squared_errors_jacobian
	Jacobian of squared_errors with respect to the parameters.
VectorR	gradient
	Approximate gradient (J^T * e).
MatrixR	hessian
	Approximate Hessian (J^T * J + lambda * I).

Detailed Description

Scratch state used by LevenbergMarquardtAlgorithm.

Holds per-sample errors and squared errors, the Jacobian of the squared errors with respect to parameters, and the resulting approximate gradient and Hessian used by the LM update.

Constructor & Destructor Documentation

BackPropagationLM()

opennn::BackPropagationLM::BackPropagationLM	(	const Index	samples_number = 0,
		Loss *	loss = nullptr )

Constructs the scratch state.

Parameters

samples_number	Number of samples that will be processed.
loss	Loss to which the state is bound.

~BackPropagationLM()

virtual opennn::BackPropagationLM::~BackPropagationLM ( )

virtualdefault

Virtual destructor.

Member Function Documentation

set()

void opennn::BackPropagationLM::set	(	const Index	samples_number = 0,
		Loss *	loss = nullptr )

(Re)allocates buffers for the given dataset.

Parameters

samples_number	Number of samples to allocate for.
loss	Loss to bind.

Member Data Documentation

error

float opennn::BackPropagationLM::error

Mean error over the batch.

errors

VectorR opennn::BackPropagationLM::errors

Per-sample residual errors.

gradient

VectorR opennn::BackPropagationLM::gradient

Approximate gradient (J^T * e).

hessian

MatrixR opennn::BackPropagationLM::hessian

Approximate Hessian (J^T * J + lambda * I).

loss

Loss* opennn::BackPropagationLM::loss = nullptr

Loss to which this scratch is bound; not owned.

loss_value

float opennn::BackPropagationLM::loss_value = 0.0f

Total loss = error + regularization.

output_delta_dimensions

Shape opennn::BackPropagationLM::output_delta_dimensions

Shape of output_deltas as a multi-dimensional view.

output_deltas

VectorR opennn::BackPropagationLM::output_deltas

Per-sample output deltas.

regularization

float opennn::BackPropagationLM::regularization = 0.0f

Regularization contribution to the loss.

samples_number

Index opennn::BackPropagationLM::samples_number = 0

Number of samples this scratch is sized for.

squared_errors

VectorR opennn::BackPropagationLM::squared_errors

Per-sample squared errors.

squared_errors_jacobian

MatrixR opennn::BackPropagationLM::squared_errors_jacobian

Jacobian of squared_errors with respect to the parameters.