opennn::ModelSelection Class Reference

Searches for the best generalizing architecture for a model. More...

#include <model_selection.h>

Public Member Functions
	ModelSelection (TrainingStrategy *new_training_strategy=nullptr)
	Constructs a model selection bound to a training strategy.
const TrainingStrategy *	get_training_strategy () const
	Returns the underlying training strategy.
bool	has_training_strategy () const
	Reports whether a training strategy is configured.
void	set (TrainingStrategy *new_training_strategy)
	Replaces the underlying training strategy.
void	set_default ()
	Picks default selection algorithms.
void	check () const
	Validates that all components needed for selection are set.
NeuronsSelectionResults	perform_neurons_selection ()
	Runs the configured neuron-selection algorithm.
InputsSelectionResults	perform_input_selection ()
	Runs the configured input-selection algorithm.
void	from_JSON (const JsonDocument &document)
	Restores the model-selection state from a JSON document.
void	to_JSON (JsonWriter &writer) const
	Serializes the model-selection state to JSON.
void	save (const filesystem::path &file_name) const
	Saves the model-selection state to a JSON file on disk.
void	load (const filesystem::path &file_name)
	Loads the model-selection state from a JSON file on disk.

Detailed Description

Searches for the best generalizing architecture for a model.

Wraps two pluggable algorithms:

• an InputsSelection strategy that decides which input features to keep,
• a NeuronSelection strategy that decides how many hidden neurons to use.

Both algorithms run on top of a user-supplied TrainingStrategy, which provides the underlying NeuralNetwork, Dataset and Loss they consume to score candidate architectures.

Constructor & Destructor Documentation

ModelSelection()

opennn::ModelSelection::ModelSelection

(

TrainingStrategy *

new_training_strategy = nullptr

)

Constructs a model selection bound to a training strategy.

Picks default selection algorithms (GrowingInputs and GrowingNeurons).

Parameters

new_training_strategy

Non-owning pointer to the strategy used to evaluate candidate architectures. May be null and supplied later via set().

Member Function Documentation

check()

void opennn::ModelSelection::check

(

)

const

Validates that all components needed for selection are set.

Verifies the training strategy, its loss, the network behind the loss, the dataset, and that the dataset has at least one validation sample.

Exceptions

runtime_error

if any required component is missing.

from_JSON()

void opennn::ModelSelection::from_JSON

(

const JsonDocument &

document

)

Restores the model-selection state from a JSON document.

Parameters

document

Parsed JSON produced by to_JSON().

get_training_strategy()

const TrainingStrategy * opennn::ModelSelection::get_training_strategy ( ) const

inline

Returns the underlying training strategy.

Returns

Const pointer to the strategy (may be nullptr).

has_training_strategy()

bool opennn::ModelSelection::has_training_strategy ( ) const

inline

Reports whether a training strategy is configured.

Returns

true if a non-null TrainingStrategy is set.

load()

void opennn::ModelSelection::load

(

const filesystem::path &

file_name

)

Loads the model-selection state from a JSON file on disk.

Parameters

file_name

Source path.

perform_input_selection()

InputsSelectionResults opennn::ModelSelection::perform_input_selection

(

)

Runs the configured input-selection algorithm.

Returns

Results of the search (selected feature subset, error history, etc.).

perform_neurons_selection()

NeuronsSelectionResults opennn::ModelSelection::perform_neurons_selection

(

)

Runs the configured neuron-selection algorithm.

Returns

Results of the search (best size found, error history, etc.).

save()

void opennn::ModelSelection::save

(

const filesystem::path &

file_name

)

const

Saves the model-selection state to a JSON file on disk.

Parameters

file_name

Destination path.

Exceptions

runtime_error

if the file cannot be opened for writing.

set()

void opennn::ModelSelection::set ( TrainingStrategy * new_training_strategy )

inline

Replaces the underlying training strategy.

Parameters

new_training_strategy

Non-owning pointer to the new strategy.

set_default()

void opennn::ModelSelection::set_default

(

)

Picks default selection algorithms.

Sets the neurons-selection algorithm to "GrowingNeurons" and the inputs-selection algorithm to "GrowingInputs".

to_JSON()

void opennn::ModelSelection::to_JSON

(

JsonWriter &

writer

)

const

Serializes the model-selection state to JSON.

Parameters

writer

JSON writer that receives the configuration tree.