Per-optimizer scratch state shared across iterations. More...

#include <optimizer.h>

Public Member Functions
	OptimizerData ()=default
	Default constructor; data buffer left empty.

virtual	~OptimizerData ()=default
	Virtual destructor.

virtual void	print () const
	Prints a human-readable summary of the scratch state.

void	set (const vector< Shape > &slot_shapes, Device device=Device::CPU)
	Allocates the scratch buffer and slices it into views.

Public Attributes
Buffer	data
	Owning storage for the per-slot scratch tensors.

vector< TensorView >	views
	Per-slot non-owning views into data.

VectorR	potential_parameters
	Candidate parameter vector used by line searches.

VectorR	training_direction
	Current search direction (e.g. quasi-Newton step).

float	initial_learning_rate = 0.0f
	Initial learning rate at the start of a line search.

Index	iteration = 0
	Iteration counter used by adaptive optimizers.

Detailed Description

Per-optimizer scratch state shared across iterations.

Holds an owning data buffer plus a vector of TensorViews into it (slot shapes are decided by the subclass), and three small fields used by line-search-based methods.

Constructor & Destructor Documentation

◆ OptimizerData()

opennn::OptimizerData::OptimizerData ( )

default

Default constructor; data buffer left empty.

◆ ~OptimizerData()

virtual opennn::OptimizerData::~OptimizerData ( )

virtualdefault

Virtual destructor.

Member Function Documentation

◆ print()

virtual void opennn::OptimizerData::print ( ) const

virtual

Prints a human-readable summary of the scratch state.

◆ set()

void opennn::OptimizerData::set	(	const vector< Shape > &	slot_shapes,
		Device	device = Device::CPU )

Allocates the scratch buffer and slices it into views.

Parameters

slot_shapes	Per-slot tensor shapes.
device	CPU or GPU memory placement.

Member Data Documentation

◆ data

Buffer opennn::OptimizerData::data

Owning storage for the per-slot scratch tensors.

◆ initial_learning_rate

float opennn::OptimizerData::initial_learning_rate = 0.0f

Initial learning rate at the start of a line search.

◆ iteration

Index opennn::OptimizerData::iteration = 0

Iteration counter used by adaptive optimizers.

◆ potential_parameters

VectorR opennn::OptimizerData::potential_parameters

Candidate parameter vector used by line searches.

◆ training_direction

VectorR opennn::OptimizerData::training_direction

Current search direction (e.g. quasi-Newton step).

◆ views

vector<TensorView> opennn::OptimizerData::views

Per-slot non-owning views into data.

Public Member Functions

Public Attributes