lbann::ltfb::RegularizedEvolution Class Reference

#include <regularized_evolution.hpp>

Inheritance diagram for lbann::ltfb::RegularizedEvolution:

Collaboration diagram for lbann::ltfb::RegularizedEvolution:

Public Types
enum	metric_strategy { metric_strategy::LOWER_IS_BETTER, metric_strategy::HIGHER_IS_BETTER }

Public Member Functions
	RegularizedEvolution (std::string metric_name, metric_strategy winner_strategy, std::unique_ptr< MutationStrategy > mutate_algo, int sample_size)
	~RegularizedEvolution ()=default
	RegularizedEvolution (RegularizedEvolution const &other)
void	select_next (model &m, ltfb::LTFBExecutionContext &ctxt, data_coordinator &dc) const final
Public Member Functions inherited from lbann::Cloneable< RegularizedEvolution, MetaLearningStrategy >
std::unique_ptr< RegularizedEvolution >	clone () const
	Return an exception-safe, memory-safe copy of this object. More...

Private Member Functions
EvalType	evaluate_model (model &m, LTFBExecutionContext &ctxt, data_coordinator &dc) const
	Get the value of the given metric from the model. More...

Private Attributes
std::unique_ptr< MutationStrategy >	m_mutate_algo
	The strategy for mutation of a model. More...
std::string	m_metric_name
	Name of the metric for evaluation. More...
metric_strategy	m_metric_strategy
	Strategy to consider for evaluating the metric e.g., HIGHER_IS_BETTER or LOWER_IS_BETTER. More...
int	m_sample_size
	The size of the sample to choose from the population in every step. More...

Detailed Description

This is a meta-learning strategy in population-based training. A sample of trainers is chosen from a population in every tournament. The best trainer is chosen from that sample according to an evaluation metric. Then the model from that best trainer is mutated and replaces the oldest model.

Definition at line 51 of file regularized_evolution.hpp.

Member Enumeration Documentation

metric_strategy

enum lbann::ltfb::RegularizedEvolution::metric_strategy

strong

Enumerator
LOWER_IS_BETTER
HIGHER_IS_BETTER

Definition at line 55 of file regularized_evolution.hpp.

Constructor & Destructor Documentation

RegularizedEvolution() [1/2]

lbann::ltfb::RegularizedEvolution::RegularizedEvolution	(	std::string	metric_name,
		metric_strategy	winner_strategy,
		std::unique_ptr< MutationStrategy >	mutate_algo,
		int	sample_size
	)

~RegularizedEvolution()

lbann::ltfb::RegularizedEvolution::~RegularizedEvolution ( )

default

RegularizedEvolution() [2/2]

lbann::ltfb::RegularizedEvolution::RegularizedEvolution

(

RegularizedEvolution const &

other

)

Member Function Documentation

evaluate_model()

EvalType lbann::ltfb::RegularizedEvolution::evaluate_model ( model &  m, LTFBExecutionContext &  ctxt, data_coordinator &  dc  ) const

private

Get the value of the given metric from the model.

select_next()

void lbann::ltfb::RegularizedEvolution::select_next ( model &  m, ltfb::LTFBExecutionContext &  ctxt, data_coordinator &  dc  ) const

final

Member Data Documentation

m_metric_name

std::string lbann::ltfb::RegularizedEvolution::m_metric_name

private

Name of the metric for evaluation.

Definition at line 91 of file regularized_evolution.hpp.

m_metric_strategy

metric_strategy lbann::ltfb::RegularizedEvolution::m_metric_strategy

private

Strategy to consider for evaluating the metric e.g., HIGHER_IS_BETTER or LOWER_IS_BETTER.

Definition at line 96 of file regularized_evolution.hpp.

m_mutate_algo

std::unique_ptr<MutationStrategy> lbann::ltfb::RegularizedEvolution::m_mutate_algo

private

The strategy for mutation of a model.

When a trainer loses in a LTFB tournament, the winning model is copied over to it and this mutation strategy is applied to the copied model to explore a new model. This is relevant to neural architecture search (NAS).

Definition at line 87 of file regularized_evolution.hpp.

m_sample_size

int lbann::ltfb::RegularizedEvolution::m_sample_size

private

The size of the sample to choose from the population in every step.

Definition at line 100 of file regularized_evolution.hpp.

---

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

• regularized_evolution.hpp