An implementation of fast exact max-kernel search. More...

Inheritance diagram for FastMKS< KernelType, MatType, TreeType >:

[legend]

Public Types
typedef TreeType< metric::IPMetric< KernelType >, FastMKSStat, MatType >	Tree
	Convenience typedef. More...

Public Member Functions
	FastMKS (const bool singleMode=false, const bool naive=false)
	Create the FastMKS object with an empty reference set and default kernel. More...

	FastMKS (const MatType &referenceSet, const bool singleMode=false, const bool naive=false)
	Create the FastMKS object with the given reference set (this is the set that is searched). More...

	FastMKS (const MatType &referenceSet, KernelType &kernel, const bool singleMode=false, const bool naive=false)
	Create the FastMKS object using the reference set (this is the set that is searched) with an initialized kernel. More...

	FastMKS (MatType &&referenceSet, const bool singleMode=false, const bool naive=false)
	Create the FastMKS object with the given reference set (this is the set that is searched), taking ownership of the reference set. More...

	FastMKS (MatType &&referenceSet, KernelType &kernel, const bool singleMode=false, const bool naive=false)
	Create the FastMKS object using the reference set (this is the set that is searched) with an initialized kernel, taking ownership of the reference set. More...

	FastMKS (Tree *referenceTree, const bool singleMode=false)
	Create the FastMKS object with an already-initialized tree built on the reference points. More...

	FastMKS (const FastMKS &other)
	Copy the parameters of the given model. More...

	FastMKS (FastMKS &&other)
	Take ownership of the given model. More...

	~FastMKS ()
	Destructor for the FastMKS object. More...

const metric::IPMetric< KernelType > &	Metric () const
	Get the inner-product metric induced by the given kernel. More...

metric::IPMetric< KernelType > &	Metric ()
	Modify the inner-product metric induced by the given kernel. More...

bool	Naive () const
	Get whether or not brute-force (naive) search is used. More...

bool &	Naive ()
	Modify whether or not brute-force (naive) search is used. More...

FastMKS &	operator= (const FastMKS &other)
	Assign this model to be a copy of the given model. More...

FastMKS &	operator= (FastMKS &&other)
	Move assignment operator. More...

void	Search (const MatType &querySet, const size_t k, arma::Mat< size_t > &indices, arma::mat &kernels)
	Search for the points in the reference set with maximum kernel evaluation to each point in the given query set. More...

void	Search (Tree *querySet, const size_t k, arma::Mat< size_t > &indices, arma::mat &kernels)
	Search for the points in the reference set with maximum kernel evaluation to each point in the query set corresponding to the given pre-built query tree. More...

void	Search (const size_t k, arma::Mat< size_t > &indices, arma::mat &products)
	Search for the maximum inner products of the query set (or if no query set was passed, the reference set is used). More...

template < typename Archive >
void	serialize (Archive &ar, const uint32_t)
	Serialize the model. More...

bool	SingleMode () const
	Get whether or not single-tree search is used. More...

bool &	SingleMode ()
	Modify whether or not single-tree search is used. More...

void	Train (const MatType &referenceSet)
	"Train" the FastMKS model on the given reference set (this will just build a tree, if the current search mode is not naive mode). More...

void	Train (const MatType &referenceSet, KernelType &kernel)
	"Train" the FastMKS model on the given reference set and use the given kernel. More...

void	Train (MatType &&referenceSet)
	"Train" the FastMKS model on the given reference set (this will just build a tree, if the current search mode is not naive mode). More...

void	Train (MatType &&referenceSet, KernelType &kernel)
	"Train" the FastMKS model on the given reference set and use the given kernel. More...

void	Train (Tree *referenceTree)
	Train the FastMKS model on the given reference tree. More...

Detailed Description

template
<
typename
KernelType
,
typename
MatType
=
arma::mat
,
template
<
typename
TreeMetricType
,
typename
TreeStatType
,
typename
TreeMatType
>
class
TreeType
=
tree::StandardCoverTree
>

class mlpack::fastmks::FastMKS< KernelType, MatType, TreeType >

An implementation of fast exact max-kernel search.

Given a query dataset and a reference dataset (or optionally just a reference dataset which is also used as the query dataset), fast exact max-kernel search finds, for each point in the query dataset, the k points in the reference set with maximum kernel value K(p_q, p_r), where k is a specified parameter and K() is a Mercer kernel.

For more information, see the following paper.

@inproceedings{curtin2013fast,
  title={Fast Exact Max-Kernel Search},
  author={Curtin, Ryan R. and Ram, Parikshit and Gray, Alexander G.},
  booktitle={Proceedings of the 2013 SIAM International Conference on Data
      Mining (SDM 13)},
  year={2013}
}

This class allows specification of the type of kernel and also of the type of tree. FastMKS can be run on kernels that work on arbitrary objects – however, this only works with cover trees and other trees that are built only on points in the dataset (and not centroids of regions or anything like that).

Template Parameters

KernelType	Type of kernel to run FastMKS with.
MatType	Type of data matrix (usually arma::mat).
TreeType	Type of tree to run FastMKS with; it must satisfy the TreeType policy API.

Definition at line 63 of file fastmks.hpp.

Member Typedef Documentation

◆ Tree

typedef TreeType<metric::IPMetric<KernelType>, FastMKSStat, MatType> Tree

Convenience typedef.

Definition at line 67 of file fastmks.hpp.

Constructor & Destructor Documentation

◆ FastMKS() [1/8]

FastMKS	(	const bool	singleMode = `false`,
		const bool	naive = `false`
	)

Create the FastMKS object with an empty reference set and default kernel.

Make sure to call Train() before Search() is called!

Parameters

singleMode	Whether or not to run single-tree search.
naive	Whether or not to run brute-force (naive) search.

◆ FastMKS() [2/8]

FastMKS	(	const MatType &	referenceSet,
		const bool	singleMode = `false`,
		const bool	naive = `false`
	)

Create the FastMKS object with the given reference set (this is the set that is searched).

Optionally, specify whether or not single-tree search or naive (brute-force) search should be used.

Parameters

referenceSet	Set of reference data.
singleMode	Whether or not to run single-tree search.
naive	Whether or not to run brute-force (naive) search.

◆ FastMKS() [3/8]

FastMKS	(	const MatType &	referenceSet,
		KernelType &	kernel,
		const bool	singleMode = `false`,
		const bool	naive = `false`
	)

Create the FastMKS object using the reference set (this is the set that is searched) with an initialized kernel.

This is useful for when the kernel stores state. Optionally, specify whether or not single-tree search or naive (brute-force) search should be used.

Parameters

referenceSet	Reference set of data for FastMKS.
kernel	Initialized kernel.
singleMode	Whether or not to run single-tree search.
naive	Whether or not to run brute-force (naive) search.

◆ FastMKS() [4/8]

FastMKS	(	MatType &&	referenceSet,
		const bool	singleMode = `false`,
		const bool	naive = `false`
	)

Create the FastMKS object with the given reference set (this is the set that is searched), taking ownership of the reference set.

Optionally, specify whether or not single-tree search or naive (brute-force) search should be used.

Parameters

referenceSet	Set of reference data.
singleMode	Whether or not to run single-tree search.
naive	Whether or not to run brute-force (naive) search.

◆ FastMKS() [5/8]

FastMKS	(	MatType &&	referenceSet,
		KernelType &	kernel,
		const bool	singleMode = `false`,
		const bool	naive = `false`
	)

Create the FastMKS object using the reference set (this is the set that is searched) with an initialized kernel, taking ownership of the reference set.

This is useful for when the kernel stores state. Optionally, specify whether or not single-tree search or naive (brute-force) search should be used.

Parameters

referenceSet	Reference set of data for FastMKS.
kernel	Initialized kernel.
singleMode	Whether or not to run single-tree search.
naive	Whether or not to run brute-force (naive) search.

◆ FastMKS() [6/8]

FastMKS	(	Tree *	referenceTree,
		const bool	singleMode = `false`
	)

Create the FastMKS object with an already-initialized tree built on the reference points.

Be sure that the tree is built with the metric type IPMetric<KernelType>. Optionally, whether or not to run single-tree search can be specified. Brute-force search is not available with this constructor since a tree is given (use one of the other constructors).

Parameters

referenceTree	Tree built on reference data.
singleMode	Whether or not to run single-tree search.

◆ FastMKS() [7/8]

FastMKS ( const FastMKS< KernelType, MatType, TreeType > & other )

Copy the parameters of the given model.

◆ FastMKS() [8/8]

FastMKS ( FastMKS< KernelType, MatType, TreeType > && other )

Take ownership of the given model.

◆ ~FastMKS()

~FastMKS ( )

Destructor for the FastMKS object.

Member Function Documentation

◆ Metric() [1/2]

const metric::IPMetric<KernelType>& Metric ( ) const

inline

Get the inner-product metric induced by the given kernel.

Definition at line 291 of file fastmks.hpp.

◆ Metric() [2/2]

metric::IPMetric<KernelType>& Metric ( )

inline

Modify the inner-product metric induced by the given kernel.

Definition at line 293 of file fastmks.hpp.

◆ Naive() [1/2]

bool Naive ( ) const

inline

Get whether or not brute-force (naive) search is used.

Definition at line 301 of file fastmks.hpp.

◆ Naive() [2/2]

bool& Naive ( )

inline

Modify whether or not brute-force (naive) search is used.

Definition at line 303 of file fastmks.hpp.

◆ operator=() [1/2]

FastMKS& operator= ( const FastMKS< KernelType, MatType, TreeType > & other )

Assign this model to be a copy of the given model.

◆ operator=() [2/2]

FastMKS& operator= ( FastMKS< KernelType, MatType, TreeType > && other )

Move assignment operator.

◆ Search() [1/3]

void Search	(	const MatType &	querySet,
		const size_t	k,
		arma::Mat< size_t > &	indices,
		arma::mat &	kernels
	)

Search for the points in the reference set with maximum kernel evaluation to each point in the given query set.

The resulting kernel evaluations are stored in the kernels matrix, and the corresponding point indices are stored in the indices matrix. The results for each point in the query set are stored in the corresponding column of the kernels and products matrices; for instance, the index of the point with maximum kernel evaluation to point 4 in the query set will be stored in row 0 and column 4 of the indices matrix.

If querySet only contains a few points, the extra overhead of building a tree to perform dual-tree search may not be warranted, and it may be faster to use single-tree search, either by setting singleMode to false in the constructor or with SingleMode().

Parameters

querySet	Set of query points (can be a single point).
k	The number of maximum kernels to find.
indices	Matrix to store resulting indices of max-kernel search in.
kernels	Matrix to store resulting max-kernel values in.

◆ Search() [2/3]

void Search	(	Tree *	querySet,
		const size_t	k,
		arma::Mat< size_t > &	indices,
		arma::mat &	kernels
	)

Search for the points in the reference set with maximum kernel evaluation to each point in the query set corresponding to the given pre-built query tree.

The resulting kernel evaluations are stored in the kernels matrix, and the corresponding point indices are stored in the indices matrix. The results for each point in the query set are stored in the corresponding column of the kernels and products matrices; for instance, the index of the point with maximum kernel evaluation to point 4 in the query set will be stored in row 0 and column 4 of the indices matrix.

This will throw an exception if called while the FastMKS object has 'single' set to true.

Be aware that if your tree modifies the original input matrix, the results here are with respect to the modified input matrix (that is, queryTree->Dataset()).

Parameters

querySet	Tree built on query points.
k	The number of maximum kernels to find.
indices	Matrix to store resulting indices of max-kernel search in.
kernels	Matrix to store resulting max-kernel values in.

◆ Search() [3/3]

void Search	(	const size_t	k,
		arma::Mat< size_t > &	indices,
		arma::mat &	products
	)

Search for the maximum inner products of the query set (or if no query set was passed, the reference set is used).

The resulting maximum inner products are stored in the products matrix and the corresponding point indices are stores in the indices matrix. The results for each point in the query set are stored in the corresponding column of the indices and products matrices; for instance, the index of the point with maximum inner product to point 4 in the query set will be stored in row 0 and column 4 of the indices matrix.

Parameters

k	The number of maximum kernels to find.
indices	Matrix to store resulting indices of max-kernel search in.
products	Matrix to store resulting max-kernel values in.

◆ serialize()

void serialize	(	Archive &	ar,
		const uint32_t
	)

Serialize the model.

Referenced by FastMKS< mlpack::kernel::CosineDistance >::Naive().

◆ SingleMode() [1/2]

bool SingleMode ( ) const

inline

Get whether or not single-tree search is used.

Definition at line 296 of file fastmks.hpp.

◆ SingleMode() [2/2]

bool& SingleMode ( )

inline

Modify whether or not single-tree search is used.

Definition at line 298 of file fastmks.hpp.

◆ Train() [1/5]

void Train ( const MatType & referenceSet )

"Train" the FastMKS model on the given reference set (this will just build a tree, if the current search mode is not naive mode).

Parameters

referenceSet Set of reference points.

◆ Train() [2/5]

void Train	(	const MatType &	referenceSet,
		KernelType &	kernel
	)

"Train" the FastMKS model on the given reference set and use the given kernel.

This will just build a tree and replace the metric, if the current search mode is not naive mode.

Parameters

referenceSet	Set of reference points.
kernel	Kernel to use for search.

◆ Train() [3/5]

void Train ( MatType && referenceSet )

"Train" the FastMKS model on the given reference set (this will just build a tree, if the current search mode is not naive mode).

This takes ownership of the reference set.

Parameters

referenceSet Set of reference points.

◆ Train() [4/5]

void Train	(	MatType &&	referenceSet,
		KernelType &	kernel
	)

"Train" the FastMKS model on the given reference set and use the given kernel.

This will just build a tree and replace the metric, if the current search mode is not naive mode. This takes ownership of the reference set.

Parameters

referenceSet	Set of reference points.
kernel	Kernel to use for search.

◆ Train() [5/5]

void Train ( Tree * referenceTree )

Train the FastMKS model on the given reference tree.

This takes ownership of the tree, so you do not need to delete it! This will throw an exception if the model is searching in naive mode (i.e. if Naive() == true).

Parameters

referenceTree Tree to use as reference data.

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

/home/ryan/src/mlpack.org/_src/mlpack-git/src/mlpack/methods/fastmks/fastmks.hpp

Public Types

Public Member Functions

Detailed Description

template<typenameKernelType,typenameMatType=arma::mat,template<typenameTreeMetricType,typenameTreeStatType,typenameTreeMatType>classTreeType=tree::StandardCoverTree> class mlpack::fastmks::FastMKS< KernelType, MatType, TreeType >

Member Typedef Documentation

◆ Tree

Constructor & Destructor Documentation

◆ FastMKS() [1/8]

◆ FastMKS() [2/8]

◆ FastMKS() [3/8]

◆ FastMKS() [4/8]

◆ FastMKS() [5/8]

◆ FastMKS() [6/8]

◆ FastMKS() [7/8]

◆ FastMKS() [8/8]

◆ ~FastMKS()

Member Function Documentation

◆ Metric() [1/2]

◆ Metric() [2/2]

◆ Naive() [1/2]

◆ Naive() [2/2]

◆ operator=() [1/2]

◆ operator=() [2/2]

◆ Search() [1/3]

◆ Search() [2/3]

◆ Search() [3/3]

◆ serialize()

◆ SingleMode() [1/2]

◆ SingleMode() [2/2]

◆ Train() [1/5]

◆ Train() [2/5]

◆ Train() [3/5]

◆ Train() [4/5]

◆ Train() [5/5]

template
<
typename
KernelType
,
typename
MatType
=
arma::mat
,
template
<
typename
TreeMetricType
,
typename
TreeStatType
,
typename
TreeMatType
>
class
TreeType
=
tree::StandardCoverTree
>

class mlpack::fastmks::FastMKS< KernelType, MatType, TreeType >