opennn::TrainingStrategy Class Reference

Coordinates the training of a NeuralNetwork on a Dataset. More...

#include <training_strategy.h>

Public Member Functions
	TrainingStrategy (NeuralNetwork *new_neural_network=nullptr, Dataset *new_dataset=nullptr)
	Constructs a training strategy for a given network and dataset.
const Dataset *	get_dataset () const
	Returns the dataset associated with the training strategy.
Dataset *	get_dataset ()
	Returns the dataset associated with the training strategy.
const NeuralNetwork *	get_neural_network () const
	Returns the neural network being trained.
NeuralNetwork *	get_neural_network ()
	Returns the neural network being trained.
const Loss *	get_loss () const
	Returns the loss term used during training.
Loss *	get_loss ()
	Returns the loss term used during training.
const Optimizer *	get_optimization_algorithm () const
	Returns the optimizer that performs parameter updates.
Optimizer *	get_optimization_algorithm ()
	Returns the optimizer that performs parameter updates.
void	set (NeuralNetwork *new_neural_network=nullptr, Dataset *new_dataset=nullptr)
	Resets the strategy to point at a new network and dataset.
void	set_default ()
	Picks a default loss and optimizer based on the network architecture.
void	set_dataset (Dataset *new_dataset)
	Replaces the dataset pointer without rebuilding loss/optimizer.
void	set_neural_network (NeuralNetwork *new_neural_network)
	Replaces the neural network pointer without rebuilding loss/optimizer.
void	set_loss (const string &new_loss)
	Selects the loss term by name.
void	set_optimization_algorithm (const string &new_optimization_algorithm)
	Selects the optimization algorithm by name.
TrainingResults	train ()
	Runs the training loop.
void	from_JSON (const JsonDocument &document)
	Restores the strategy state from a JSON document.
void	to_JSON (JsonWriter &writer) const
	Serializes the strategy state to JSON.
void	save (const filesystem::path &file_name) const
	Saves the strategy state to a JSON file on disk.
void	load (const filesystem::path &file_name)
	Loads the strategy state from a JSON file on disk.

Detailed Description

Coordinates the training of a NeuralNetwork on a Dataset.

Aggregates a Loss (error term) and an Optimizer (parameter-update rule), keeps non-owning pointers to the network and dataset they operate on, and exposes a single train() entry point that delegates to the optimizer.

Sensible defaults are picked automatically based on the network's layer composition: AdaptiveMomentEstimation + appropriate loss for recurrent, convolutional, transformer-like and text-classification networks; QuasiNewtonMethod

• CrossEntropy / WeightedSquaredError / MeanSquaredError for the classification and approximation cases.

Constructor & Destructor Documentation

TrainingStrategy()

opennn::TrainingStrategy::TrainingStrategy	(	NeuralNetwork *	new_neural_network = nullptr,
		Dataset *	new_dataset = nullptr )

Constructs a training strategy for a given network and dataset.

Both pointers may be null and supplied later via set(). When both are provided the constructor calls set_default() to pick a loss and optimizer appropriate for the network's architecture.

Parameters

new_neural_network	Non-owning pointer to the network to be trained.
new_dataset	Non-owning pointer to the dataset providing samples.

Member Function Documentation

from_JSON()

void opennn::TrainingStrategy::from_JSON

(

const JsonDocument &

document

)

Restores the strategy state from a JSON document.

Parameters

document

Parsed JSON document produced by to_JSON().

get_dataset() [1/2]

Dataset * opennn::TrainingStrategy::get_dataset ( )

inline

Returns the dataset associated with the training strategy.

Returns

Mutable pointer to the dataset (may be nullptr).

get_dataset() [2/2]

const Dataset * opennn::TrainingStrategy::get_dataset ( ) const

inline

Returns the dataset associated with the training strategy.

Returns

Const pointer to the dataset (may be nullptr).

get_loss() [1/2]

Loss * opennn::TrainingStrategy::get_loss ( )

inline

Returns the loss term used during training.

Returns

Mutable pointer to the Loss instance owned by this strategy.

get_loss() [2/2]

const Loss * opennn::TrainingStrategy::get_loss ( ) const

inline

Returns the loss term used during training.

Returns

Const pointer to the Loss instance owned by this strategy.

get_neural_network() [1/2]

NeuralNetwork * opennn::TrainingStrategy::get_neural_network ( )

inline

Returns the neural network being trained.

Returns

Mutable pointer to the network (may be nullptr).

get_neural_network() [2/2]

const NeuralNetwork * opennn::TrainingStrategy::get_neural_network ( ) const

inline

Returns the neural network being trained.

Returns

Const pointer to the network (may be nullptr).

get_optimization_algorithm() [1/2]

Optimizer * opennn::TrainingStrategy::get_optimization_algorithm ( )

inline

Returns the optimizer that performs parameter updates.

Returns

Mutable pointer to the Optimizer instance owned by this strategy.

get_optimization_algorithm() [2/2]

const Optimizer * opennn::TrainingStrategy::get_optimization_algorithm ( ) const

inline

Returns the optimizer that performs parameter updates.

Returns

Const pointer to the Optimizer instance owned by this strategy.

load()

void opennn::TrainingStrategy::load

(

const filesystem::path &

file_name

)

Loads the strategy state from a JSON file on disk.

Calls set_default() before reading, so any fields missing from the file keep architecture-appropriate defaults.

Parameters

file_name

Source path.

save()

void opennn::TrainingStrategy::save

(

const filesystem::path &

file_name

)

const

Saves the strategy 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::TrainingStrategy::set	(	NeuralNetwork *	new_neural_network = nullptr,
		Dataset *	new_dataset = nullptr )

Resets the strategy to point at a new network and dataset.

Stores both pointers and calls set_default() to rebuild the loss and optimizer to match the network's architecture.

Parameters

new_neural_network	Non-owning pointer to the network.
new_dataset	Non-owning pointer to the dataset.

set_dataset()

void opennn::TrainingStrategy::set_dataset ( Dataset * new_dataset )

inline

Replaces the dataset pointer without rebuilding loss/optimizer.

Parameters

new_dataset

Non-owning pointer to the new dataset.

set_default()

void opennn::TrainingStrategy::set_default

(

)

Picks a default loss and optimizer based on the network architecture.

Inspects the layers of the configured neural network and selects:

• Recurrent networks: MeanSquaredError + AdaptiveMomentEstimation.
• Convolutional networks: CrossEntropy + AdaptiveMomentEstimation.
• Transformer-style networks (Dense over rank-2 inputs): CrossEntropy + Adam.
• Text classification (Embedding or MultiHeadAttention): CrossEntropy + Adam.
• Multi-class classification (Softmax output): CrossEntropy + QuasiNewtonMethod.
• Binary classification (Sigmoid output): WeightedSquaredError + QuasiNewtonMethod.
• Regression (default): MeanSquaredError + QuasiNewtonMethod.

set_loss()

void opennn::TrainingStrategy::set_loss

(

const string &

new_loss

)

Selects the loss term by name.

Constructs a fresh Loss instance bound to the current network and dataset, sets the requested error type and re-binds the optimizer (if any).

Parameters

new_loss

One of: "MeanSquaredError", "NormalizedSquaredError", "WeightedSquaredError", "CrossEntropy", "CrossEntropyError3d", "MinkowskiError".

set_neural_network()

void opennn::TrainingStrategy::set_neural_network ( NeuralNetwork * new_neural_network )

inline

Replaces the neural network pointer without rebuilding loss/optimizer.

Parameters

new_neural_network

Non-owning pointer to the new network.

set_optimization_algorithm()

void opennn::TrainingStrategy::set_optimization_algorithm

(

const string &

new_optimization_algorithm

)

Selects the optimization algorithm by name.

Looks up the optimizer in the registry, instantiates it and binds it to the current loss.

Parameters

new_optimization_algorithm

One of: "AdaptiveMomentEstimation", "QuasiNewtonMethod", "StochasticGradientDescent", "LevenbergMarquardtAlgorithm".

to_JSON()

void opennn::TrainingStrategy::to_JSON

(

JsonWriter &

writer

)

const

Serializes the strategy state to JSON.

Parameters

writer

JSON writer that receives the strategy's element tree.

train()

TrainingResults opennn::TrainingStrategy::train

(

)

Runs the training loop.

Validates that network, dataset, loss and optimizer are all set, applies the forecasting batch-size adjustment when relevant, then delegates to the optimizer's train() method.

Returns

TrainingResults with the per-epoch loss/selection-error history and the final stopping condition.

Exceptions

runtime_error

if any required component is not configured.