Implementation of a solver operating on the blocks of the Hessian. More...

Inheritance diagram for BlockSolver< Traits >:

Collaboration diagram for BlockSolver< Traits >:

Public Member Functions
	BlockSolver ()
	Default constructor.
bool	buildStructure (bool zeroBlocks, IndexScratch &scratch)
	Builds the sparse block matrix structure from the graph.
bool	buildSystem ()
	Builds the linear system (Hessian blocks and gradient) from the current graph.
void	clear ()
	Clears all solver state, resetting sizes to zero.
bool	init (bool online=false)
	Initializes the block solver.
void	multiplyHessian (g_type dest, const g_type src) const
	Multiplies the pose Hessian by a vector: dest = Hpp * src.
SparseOptimizer &	optimizer ()
	Returns a mutable reference to the underlying optimizer.
const SparseOptimizer &	optimizer () const
	Returns a const reference to the underlying optimizer.
void	restoreDiagonal ()
	Restores the Hessian diagonal from the backup.
bool	schur ()
	Returns whether the Schur complement is currently enabled.
bool	setLambda (g_type lambda, bool backup=false)
	Adds a damping factor to the diagonal of the Hessian.
void	setSchur (bool s)
	Enables or disables the Schur complement.
bool	solve ()
	Solves the linear system, optionally using the Schur complement.
bool	supportsSchur ()
	Returns true; this solver supports the Schur complement.
bool	updateStructure (const Buffer< HyperGraph::HGVertex * > &vset, const Set< HyperGraph::HGVertex * > &edges)
	Updates the structure after vertices or edges have changed.
Public Member Functions inherited from Solver
	Solver ()
	Default constructor.
	~Solver ()
	Destructor.
g_type *	b ()
	return b, the right hand side of the system
uint04	vectorSize () const
	return the size of the solution vector (x) and b
g_type *	x ()
	return x, the solution vector

Public Attributes
LinearSolverType	solver
	The underlying linear solver.

Protected Member Functions
void	resize (IndexScratch &scratch, int totalDim)
	Resizes all internal matrices and vectors for the given problem size.
Protected Member Functions inherited from Solver
void	resizeVector (uint04 sx)
	Resizes the solution and right-hand side vectors.

Protected Attributes
Buffer< g_type >	_bschur
	Right-hand side for the Schur complement system.
Buffer< g_type >	_coefficients
	Coefficient buffer for Schur elimination.
Buffer< LandmarkVectorType >	_diagonalBackupLandmark
	Backup of landmark Hessian diagonal.
Buffer< PoseVectorType >	_diagonalBackupPose
	Backup of pose Hessian diagonal.
SparseBlockMatrixDiagonal< LandmarkMatrixType >	_DInvSchur
	Inverse diagonal of the landmark Hessian.
bool	_doSchur
	Whether to use the Schur complement.
SparseBlockMatrix< LandmarkMatrixType >	_Hll
	Landmark-landmark Hessian block.
SparseBlockMatrix< PoseLandmarkMatrixType >	_Hpl
	Pose-landmark Hessian block.
SparseBlockMatrixCCS< PoseLandmarkMatrixType >	_HplCCS
	CCS view of the pose-landmark Hessian.
SparseBlockMatrix< PoseMatrixType >	_Hpp
	Pose-pose Hessian block.
SparseBlockMatrix< PoseMatrixType >	_Hschur
	Schur complement of the Hessian.
SparseBlockMatrixCCS< PoseMatrixType >	_HschurTransposedCCS
	Transposed CCS view of the Schur complement.
uint04	_numLandmarks
	Number of pose and landmark blocks.
uint04	_sizeLandmarks
	Total scalar dimensions of poses and landmarks.
Protected Attributes inherited from Solver
Buffer< g_type >	_b
	The right-hand side vector.
Buffer< g_type >	_x
	The solution vector.

Detailed Description

template<typename Traits>
class BlockSolver< Traits >

Implementation of a solver operating on the blocks of the Hessian.

Definition at line 66 of file block_solver.h.

Member Function Documentation

◆ buildStructure()

template<typename Traits>

bool BlockSolver< Traits >::buildStructure	(	bool	zeroBlocks,
		IndexScratch &	scratch )

Builds the sparse block matrix structure from the graph.

Parameters

[in]	zeroBlocks	Whether to zero the blocks after allocation.
[in]	scratch	Scratch buffers for index assignments.

Returns: True on success.

Definition at line 76 of file block_solver.hpp.

References _DInvSchur, _doSchur, _Hll, _Hpl, _HplCCS, _Hpp, _Hschur, _HschurTransposedCCS, _numLandmarks, _sizeLandmarks, Buffer< t_type, t_memory_manager >::add(), SparseBlockMatrixHashMap< MatrixType >::addBlock(), IndexScratch::block_landmark_indices, IndexScratch::block_pose_indices, SparseBlockMatrixHashMap< MatrixType >::columns(), OptimizableGraph::OGVertex::dimension(), HyperGraph::HGVertex::edges(), OptimizableGraph::OGVertex::hessianIndex(), OptimizableGraph::OGEdge::mapHessianMemory(), OptimizableGraph::OGVertex::mapHessianMemory(), OptimizableGraph::OGVertex::marginalized(), resize(), OptimizableGraph::OGVertex::setColInHessian(), solver, HyperGraph::HGEdge::vertex(), and HyperGraph::HGEdge::vertexCount().

◆ buildSystem()

template<typename Traits>

bool BlockSolver< Traits >::buildSystem ( )

inline

Builds the linear system (Hessian blocks and gradient) from the current graph.

Returns: True on success.

Definition at line 124 of file block_solver.h.

◆ init()

template<typename Traits>

bool BlockSolver< Traits >::init ( bool online = false )

Initializes the block solver.

Parameters

[in] online Whether this is an online (incremental) session.

Returns: True on success.

Definition at line 364 of file block_solver.hpp.

References _Hll, _Hpl, _Hpp, and solver.

◆ multiplyHessian()

template<typename Traits>

void BlockSolver< Traits >::multiplyHessian	(	g_type *	dest,
		const g_type *	src ) const

inline

Multiplies the pose Hessian by a vector: dest = Hpp * src.

Parameters

[out]	dest	The destination vector.
[in]	src	The source vector.

Definition at line 292 of file block_solver.h.

◆ resize()

template<typename Traits>

void BlockSolver< Traits >::resize	(	IndexScratch &	scratch,
		int	totalDim )

protected

Resizes all internal matrices and vectors for the given problem size.

Parameters

[in]	scratch	Scratch buffers for index assignments.
[in]	totalDim	The total dimension of the problem.

Definition at line 49 of file block_solver.hpp.

References _bschur, _coefficients, _DInvSchur, _doSchur, _Hll, _Hpl, _HplCCS, _Hpp, _Hschur, _HschurTransposedCCS, IndexScratch::block_landmark_indices, IndexScratch::block_pose_indices, and Solver::resizeVector().

Referenced by buildStructure().

◆ setLambda()

template<typename Traits>

bool BlockSolver< Traits >::setLambda	(	g_type	lambda,
		bool	backup = false )

Adds a damping factor to the diagonal of the Hessian.

Parameters

[in]	lambda	The damping factor.
[in]	backup	Whether to backup the diagonal before modification.

Returns: True on success.

Definition at line 319 of file block_solver.hpp.

References _diagonalBackupLandmark, _diagonalBackupPose, _Hll, _Hpp, _numLandmarks, and Solver::b().

◆ setSchur()

template<typename Traits>

void BlockSolver< Traits >::setSchur ( bool s )

inline

Enables or disables the Schur complement.

Parameters

[in] s True to enable, false to disable.

Definition at line 288 of file block_solver.h.

◆ solve()

template<typename Traits>

bool BlockSolver< Traits >::solve ( )

inline

Solves the linear system, optionally using the Schur complement.

Returns: True on success.

Definition at line 162 of file block_solver.h.

◆ updateStructure()

template<typename Traits>

bool BlockSolver< Traits >::updateStructure	(	const Buffer< HyperGraph::HGVertex * > &	vset,
		const Set< HyperGraph::HGVertex * > &	edges )

Updates the structure after vertices or edges have changed.

Parameters

[in]	vset	The updated vertices.
[in]	edges	The updated edges.

Returns: True on success.

Definition at line 245 of file block_solver.hpp.

References _Hpp, _sizeLandmarks, OptimizableGraph::OGVertex::dimension(), OptimizableGraph::OGVertex::hessianIndex(), OptimizableGraph::OGEdge::mapHessianMemory(), OptimizableGraph::OGVertex::mapHessianMemory(), OptimizableGraph::OGVertex::marginalized(), Solver::resizeVector(), OptimizableGraph::OGVertex::setColInHessian(), HyperGraph::HGEdge::vertex(), and HyperGraph::HGEdge::vertexCount().

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

GraphOptimization/Headers/block_solver.h
GraphOptimization/Headers/block_solver.hpp

Public Member Functions

Public Attributes

Protected Member Functions

Protected Attributes

Detailed Description

Member Function Documentation

◆ buildStructure()

◆ buildSystem()

◆ init()

◆ multiplyHessian()

◆ resize()

◆ setLambda()

◆ setSchur()

◆ solve()

◆ updateStructure()