PetscRegressorLinearSetUseKSP#

Set a flag to indicate that a KSP object, instead of a Tao one, should be used to fit the linear regressor

Synopsis#

#include "petscregressor.h" 
PetscErrorCode PetscRegressorLinearSetUseKSP(PetscRegressor regressor, PetscBool flg)

Logically Collective

Input Parameters#

regressor - the PetscRegressor context
flg - PETSC_TRUE to use a KSP, PETSC_FALSE to use a Tao object (default is false)

Options Database Key#

regressor_linear_use_ksp <true,false> - use KSP

Notes#

KSPLSQR with no preconditioner is used to solve the normal equations by default.

For sequential MATSEQAIJ sparse matrices QR factorization a PCType of PCQR can be used to solve the least-squares system with a MatSolverType of MATSOLVERSPQR, using, for example,

  -ksp_type none -pc_type qr -pc_factor_mat_solver_type sp

if centering, PetscRegressorLinearSetFitIntercept(), is not used.

Developer Notes#

It should be possible to use Cholesky (and any other preconditioners) to solve the normal equations.

It should be possible to use QR if centering is used. See ml/regressor/ex1.c and ex2.c

It should be possible to use dense SVD PCSVD and dense qr directly on the rectangular matrix to solve the least squares problem.

Adding the above support seems to require a refactorization of how least squares problems are solved with PETSc in KSPLSQR

Level#

intermediate

Location#

src/ml/regressor/impls/linear/linear.c

Implementations#

PetscRegressorLinearSetUseKSP_Linear() in src/ml/regressor/impls/linear/linear.c

Index of all PetscRegressor routines
Table of Contents for all manual pages
Index of all manual pages