Training configuration for projection pursuit trees and forests. More...

#include <TrainingSpec.hpp>

Classes
class	Builder
	Fluent builder for TrainingSpec. More...

Public Types
using	Ptr = std::shared_ptr<TrainingSpec>

Public Member Functions
	TrainingSpec (pp::ProjectionPursuit::Ptr pp, vars::VariableSelection::Ptr vars, cutpoint::Cutpoint::Ptr cutpoint, stop::StopRule::Ptr stop, binarize::Binarization::Ptr binarization, grouping::Grouping::Ptr grouping, leaf::LeafStrategy::Ptr leaf, types::Mode mode, int size, int seed, int threads, int max_retries)
	Construct a training specification.

TreeNode::Ptr	create_leaf (NodeContext const &ctx, stats::RNG &rng) const
	Create a leaf node from the current node context.

void	find_cutpoint (NodeContext &ctx, stats::RNG &rng) const
	Compute the split cutpoint. Asserts postcondition: `ctx.cutpoint` is set.

void	find_projection (NodeContext &ctx, stats::RNG &rng) const
	Run projection pursuit optimization. Asserts postcondition: `ctx.projector` and `ctx.pp_index_value` are set.

void	group (NodeContext &ctx, stats::RNG &rng) const
	Split observations into two child partitions.

stats::GroupPartition	init_groups (types::OutcomeVector const &y) const
	Create the initial group partition from the training response.

bool	is_forest () const
	Whether this specification describes a forest (size > 0).

void	regroup (NodeContext &ctx, stats::RNG &rng) const
	Reduce multiclass partition to binary. Asserts postcondition: `ctx.y_bin` is set.

int	resolve_threads () const
	Get the number of threads to use for training.

void	select_vars (NodeContext &ctx, stats::RNG &rng) const
	Run variable selection. Asserts postcondition: `ctx.var_selection` is set.

bool	should_stop (NodeContext const &ctx, stats::RNG &rng) const
	Check whether the node should stop growing.

nlohmann::json	to_json () const
	Serialize the training spec to JSON.

Static Public Member Functions
static Builder	builder (types::Mode mode)
	Create a builder for the given mode.

static Ptr	from_json (nlohmann::json const &j)
	Deserialize a training spec from JSON.

template<typename... Args>
static Ptr	make (Args &&... args)
	Create a shared pointer to a TrainingSpec.

Public Attributes
binarize::Binarization::Ptr const	binarization
	Binarization strategy.

cutpoint::Cutpoint::Ptr const	cutpoint
	Split cutpoint strategy.

grouping::Grouping::Ptr const	grouping
	Grouping strategy.

leaf::LeafStrategy::Ptr const	leaf
	Leaf creation strategy.

int const	max_retries
	Maximum retry attempts for degenerate trees.

types::Mode const	mode
	Training mode (classification or regression).

pp::ProjectionPursuit::Ptr const	pp
	Projection pursuit optimization strategy.

int const	seed
	RNG seed.

int const	size
	Number of trees (0 = single tree).

stop::StopRule::Ptr const	stop
	Stop rule strategy.

int const	threads
	Number of threads for parallel forest training.

vars::VariableSelection::Ptr const	vars
	Variable selection strategy.

Detailed Description

Training configuration for projection pursuit trees and forests.

Composes seven strategies (projection pursuit, variable selection, split cutpoint, stop rule, binarization, grouping, leaf) together with forest-level parameters (size, seed, threads, max retries).

TrainingSpec is a concrete class — new strategies are plugged in via the builder, not by subclassing:

// Single tree with PDA (lambda = 0.5):
auto spec = TrainingSpec::builder(types::Mode::Classification)
    .pp(pp::pda(0.5))
    .build();
 
// Random forest with uniform variable selection:
auto spec = TrainingSpec::builder(types::Mode::Classification)
    .size(100)
    .pp(pp::pda(0.0))
    .vars(vars::uniform(3))
    .build();

Strategies are held via shared_ptr and are immutable after construction, so TrainingSpec can be freely copied and shared across trees without deep cloning.

Member Typedef Documentation

◆ Ptr

using ppforest2::TrainingSpec::Ptr = std::shared_ptr<TrainingSpec>

Constructor & Destructor Documentation

◆ TrainingSpec()

ppforest2::TrainingSpec::TrainingSpec	(	pp::ProjectionPursuit::Ptr	pp,
		vars::VariableSelection::Ptr	vars,
		cutpoint::Cutpoint::Ptr	cutpoint,
		stop::StopRule::Ptr	stop,
		binarize::Binarization::Ptr	binarization,
		grouping::Grouping::Ptr	grouping,
		leaf::LeafStrategy::Ptr	leaf,
		types::Mode	mode,
		int	size,
		int	seed,
		int	threads,
		int	max_retries )

Construct a training specification.

Parameters

pp	Projection pursuit strategy.
vars	Variable selection strategy.
cutpoint	Split cutpoint strategy.
stop	Stop rule strategy.
binarization	Binarization strategy.
grouping	Grouping strategy.
leaf	Leaf creation strategy.
size	Number of trees (0 = single tree).
seed	RNG seed.
threads	Number of threads (0 = hardware concurrency).
max_retries	Maximum retry attempts for degenerate trees.

Member Function Documentation

◆ builder()

static Builder ppforest2::TrainingSpec::builder ( types::Mode mode )

inlinestatic

Create a builder for the given mode.

Mode is required because it's structurally primary: it determines the default binarize strategy and the mode-compatibility checks for every other strategy. There is no "no-mode" TrainingSpec.

◆ create_leaf()

TreeNode::Ptr ppforest2::TrainingSpec::create_leaf	(	NodeContext const &	ctx,
		stats::RNG &	rng ) const

inline

Create a leaf node from the current node context.

◆ find_cutpoint()

void ppforest2::TrainingSpec::find_cutpoint	(	NodeContext &	ctx,
		stats::RNG &	rng ) const

Compute the split cutpoint. Asserts postcondition: ctx.cutpoint is set.

◆ find_projection()

void ppforest2::TrainingSpec::find_projection	(	NodeContext &	ctx,
		stats::RNG &	rng ) const

Run projection pursuit optimization. Asserts postcondition: ctx.projector and ctx.pp_index_value are set.

◆ from_json()

static Ptr ppforest2::TrainingSpec::from_json ( nlohmann::json const & j )

static

Deserialize a training spec from JSON.

◆ group()

void ppforest2::TrainingSpec::group	(	NodeContext &	ctx,
		stats::RNG &	rng ) const

Split observations into two child partitions.

Writes ctx.lower_y_part / ctx.upper_y_part. Sets ctx.aborted if the grouping produced no progress (i.e. one child covers the whole parent).

◆ init_groups()

stats::GroupPartition ppforest2::TrainingSpec::init_groups ( types::OutcomeVector const & y ) const

inline

Create the initial group partition from the training response.

◆ is_forest()

bool ppforest2::TrainingSpec::is_forest ( ) const

inline

Whether this specification describes a forest (size > 0).

◆ make()

template<typename... Args>

static Ptr ppforest2::TrainingSpec::make ( Args &&... args )

inlinestatic

Create a shared pointer to a TrainingSpec.

◆ regroup()

void ppforest2::TrainingSpec::regroup	(	NodeContext &	ctx,
		stats::RNG &	rng ) const

Reduce multiclass partition to binary. Asserts postcondition: ctx.y_bin is set.

◆ resolve_threads()

int ppforest2::TrainingSpec::resolve_threads ( ) const

inline

Get the number of threads to use for training.

If the number of threads is not specified, the number of hardware concurrency is returned.

Returns: The number of threads to use for training.

◆ select_vars()

void ppforest2::TrainingSpec::select_vars	(	NodeContext &	ctx,
		stats::RNG &	rng ) const

Run variable selection. Asserts postcondition: ctx.var_selection is set.

◆ should_stop()

bool ppforest2::TrainingSpec::should_stop	(	NodeContext const &	ctx,
		stats::RNG &	rng ) const

Check whether the node should stop growing.

Returns true if the configured stop rule fires, OR if the node has fewer than 2 groups (a tree-level invariant — no configured rule can split a single group, so stopping is the only sane option).

◆ to_json()

nlohmann::json ppforest2::TrainingSpec::to_json ( ) const

Serialize the training spec to JSON.

Member Data Documentation

◆ binarization

binarize::Binarization::Ptr const ppforest2::TrainingSpec::binarization

Binarization strategy.

◆ cutpoint

cutpoint::Cutpoint::Ptr const ppforest2::TrainingSpec::cutpoint

Split cutpoint strategy.

◆ grouping

grouping::Grouping::Ptr const ppforest2::TrainingSpec::grouping

Grouping strategy.

◆ leaf

leaf::LeafStrategy::Ptr const ppforest2::TrainingSpec::leaf

Leaf creation strategy.

◆ max_retries

int const ppforest2::TrainingSpec::max_retries

Maximum retry attempts for degenerate trees.

◆ mode

types::Mode const ppforest2::TrainingSpec::mode

Training mode (classification or regression).

◆ pp

pp::ProjectionPursuit::Ptr const ppforest2::TrainingSpec::pp

Projection pursuit optimization strategy.

◆ seed

int const ppforest2::TrainingSpec::seed

RNG seed.

◆ size

int const ppforest2::TrainingSpec::size

Number of trees (0 = single tree).

◆ stop

stop::StopRule::Ptr const ppforest2::TrainingSpec::stop

Stop rule strategy.

◆ threads

int const ppforest2::TrainingSpec::threads

Number of threads for parallel forest training.

◆ vars

vars::VariableSelection::Ptr const ppforest2::TrainingSpec::vars

Variable selection strategy.

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

core/src/models/TrainingSpec.hpp

Classes

Public Types

Public Member Functions

Static Public Member Functions

Public Attributes

Detailed Description

Member Typedef Documentation

◆ Ptr

Constructor & Destructor Documentation

◆ TrainingSpec()

Member Function Documentation

◆ builder()

◆ create_leaf()

◆ find_cutpoint()

◆ find_projection()

◆ from_json()

◆ group()

◆ init_groups()

◆ is_forest()

◆ make()

◆ regroup()

◆ resolve_threads()

◆ select_vars()

◆ should_stop()

◆ to_json()

Member Data Documentation

◆ binarization

◆ cutpoint

◆ grouping

◆ leaf

◆ max_retries

◆ mode

◆ pp

◆ seed

◆ size

◆ stop

◆ threads

◆ vars