The interface of objective function. More...

#include <objective.h>

Inheritance diagram for xgboost::ObjFunction:

Collaboration diagram for xgboost::ObjFunction:

Public Member Functions
	~ObjFunction () override=default

virtual void	Configure (Args const &args)=0
	Configure the objective with the specified parameters. More...

virtual void	GetGradient (HostDeviceVector< float > const &preds, MetaInfo const &info, std::int32_t iter, linalg::Matrix< GradientPair > *out_gpair)=0
	Get gradient over each of predictions, given existing information. More...

virtual const char *	DefaultEvalMetric () const =0

virtual Json	DefaultMetricConfig () const
	Return the configuration for the default metric. More...

virtual void	PredTransform (HostDeviceVector< float > *) const
	Apply inverse link (activation) function to prediction values. More...

virtual void	EvalTransform (HostDeviceVector< float > *io_preds)
	Apply inverse link (activation) function to prediction values. More...

virtual void	ProbToMargin (linalg::Vector< float > *) const
	Apply the link function to the intercept. More...

virtual void	InitEstimation (MetaInfo const &info, linalg::Vector< float > *base_score) const
	Obtain the initial estimation of prediction (intercept). More...

virtual struct ObjInfo	Task () const =0
	Return task of this objective. More...

virtual bst_target_t	Targets (MetaInfo const &info) const
	Return number of targets for input matrix. Right now XGBoost supports only multi-target regression. More...

Context const *	Ctx () const
	Getter of the context. More...

virtual void	UpdateTreeLeaf (HostDeviceVector< bst_node_t > const &, MetaInfo const &, float, HostDeviceVector< float > const &, bst_target_t, RegTree *) const
	Update the leaf values after a tree is built. Needed for objectives with 0 hessian. More...

Public Member Functions inherited from xgboost::Configurable
virtual	~Configurable ()=default

virtual void	LoadConfig (Json const &in)=0
	Load configuration from JSON object. More...

virtual void	SaveConfig (Json *out) const =0
	Save configuration to JSON object. More...

Static Public Member Functions
static constexpr float	DefaultBaseScore ()

static ObjFunction *	Create (const std::string &name, Context const *ctx)
	Create an objective function according to the name. More...

Protected Attributes
Context const *	ctx_ {nullptr}

Detailed Description

The interface of objective function.

Constructor & Destructor Documentation

◆ ~ObjFunction()

xgboost::ObjFunction::~ObjFunction ( )

overridedefault

Member Function Documentation

◆ Configure()

virtual void xgboost::ObjFunction::Configure ( Args const & args )

pure virtual

Configure the objective with the specified parameters.

Parameters

args	arguments to the objective function.

◆ Create()

static ObjFunction* xgboost::ObjFunction::Create	(	const std::string &	name,
		Context const *	ctx
	)

static

Create an objective function according to the name.

Parameters

name	Name of the objective.
ctx	Pointer to the context.

◆ Ctx()

Context const* xgboost::ObjFunction::Ctx ( ) const

inline

Getter of the context.

◆ DefaultBaseScore()

static constexpr float xgboost::ObjFunction::DefaultBaseScore ( )

inlinestaticconstexpr

◆ DefaultEvalMetric()

virtual const char* xgboost::ObjFunction::DefaultEvalMetric ( ) const

pure virtual

Returns: the default evaluation metric for the objective

◆ DefaultMetricConfig()

virtual Json xgboost::ObjFunction::DefaultMetricConfig ( ) const

inlinevirtual

Return the configuration for the default metric.

◆ EvalTransform()

virtual void xgboost::ObjFunction::EvalTransform ( HostDeviceVector< float > * io_preds )

inlinevirtual

Apply inverse link (activation) function to prediction values.

This is only called when Eval is called, usually it redirect to PredTransform

Parameters

[in,out] io_preds prediction values, saves to this vector as well.

◆ GetGradient()

virtual void xgboost::ObjFunction::GetGradient	(	HostDeviceVector< float > const &	preds,
		MetaInfo const &	info,
		std::int32_t	iter,
		linalg::Matrix< GradientPair > *	out_gpair
	)

pure virtual

Get gradient over each of predictions, given existing information.

Parameters

preds	Raw prediction (before applying the inverse link) of the current round.
info	information about labels, weights, groups in rank.
iteration	current iteration number.
out_gpair	output of get gradient, saves gradient and second order gradient in

◆ InitEstimation()

virtual void xgboost::ObjFunction::InitEstimation	(	MetaInfo const &	info,
		linalg::Vector< float > *	base_score
	)		const

virtual

Obtain the initial estimation of prediction (intercept).

The output in base_score represents prediction after apply the inverse link function (valid prediction instead of raw).

Parameters

info	MetaInfo that contains label.
base_score	Output estimation.

◆ PredTransform()

virtual void xgboost::ObjFunction::PredTransform ( HostDeviceVector< float > * ) const

inlinevirtual

Apply inverse link (activation) function to prediction values.

This is only called when Prediction is called

Parameters

[in,out] io_preds prediction values, saves to this vector as well.

◆ ProbToMargin()

virtual void xgboost::ObjFunction::ProbToMargin ( linalg::Vector< float > * ) const

inlinevirtual

Apply the link function to the intercept.

This is an inverse of PredTransform for most of the objectives (if there's a valid inverse). It's used to transform user-set base_score back to margin used by gradient boosting. The method converts objective-based valid outputs like probability back to raw model outputs.

Parameters

[in,out] base_score The intercept to transform.

◆ Targets()

virtual bst_target_t xgboost::ObjFunction::Targets ( MetaInfo const & info ) const

inlinevirtual

Return number of targets for input matrix. Right now XGBoost supports only multi-target regression.

◆ Task()

virtual struct ObjInfo xgboost::ObjFunction::Task ( ) const

pure virtual

Return task of this objective.

◆ UpdateTreeLeaf()

virtual void xgboost::ObjFunction::UpdateTreeLeaf	(	HostDeviceVector< bst_node_t > const &	,
		MetaInfo const &	,
		float	,
		HostDeviceVector< float > const &	,
		bst_target_t	,
		RegTree *
	)		const

inlinevirtual

Update the leaf values after a tree is built. Needed for objectives with 0 hessian.

Note that the leaf update is not well defined for distributed training as XGBoost computes only an average of quantile between workers. This breaks when some leaf have no sample assigned in a local worker.

Parameters

position	The leaf index for each rows.
info	MetaInfo providing labels and weights.
learning_rate	The learning rate for current iteration.
prediction	Model prediction after transformation.
group_idx	The group index for this tree, 0 when it's not multi-target or multi-class.
p_tree	Tree that needs to be updated.

Member Data Documentation

◆ ctx_

Context const* xgboost::ObjFunction::ctx_ {nullptr}

protected

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

/home/docs/checkouts/readthedocs.org/user_builds/xgboost/checkouts/stable/include/xgboost/objective.h

Public Member Functions

Static Public Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ ~ObjFunction()

Member Function Documentation

◆ Configure()

◆ Create()

◆ Ctx()

◆ DefaultBaseScore()

◆ DefaultEvalMetric()

◆ DefaultMetricConfig()

◆ EvalTransform()

◆ GetGradient()

◆ InitEstimation()

◆ PredTransform()

◆ ProbToMargin()

◆ Targets()

◆ Task()

◆ UpdateTreeLeaf()

Member Data Documentation

◆ ctx_