gpr-opthyprms

From PMFLib Wiki
Jump to navigation Jump to search

Navigation: Documentation / Utilities / gpr-opthyprms


gpr-opthyprms: version 6.1090.a7cf688 (2026-06-18 21:15)

Name

gpr-opthyprms

Synopsis

gpr-opthyprms [OPTIONS] [--] ACCU REALM HYPRMS

Description

The program finds optimal GPR hyperparameters, which maximize logarithm of marginal likelihood. The optimization is performed by the L-BFGS optimizer employing either analytical or numerical gradients of logML w.r.t. hyperparameters.

Arguments

ACCU
Name of the file containing the input PMF accumulator.
REALM
Realm for GPR hyperparameter optimization. The list of supported realms can be obtained by --listrealms.
HYPRMS
Name of the file containing the optimized hyperparameters.

Options

--listrealms
List supported realms.
-t, --target NAME
Specify the optimized target, which can be either logml (log of marginal likelihood) or logpl (log of pseudo-likelihood from leave-one-out cross-validation, LOO-CV).
(Default: logml)
-l, --limit LIMIT
Only bins containing more samples than NUMBER are considered properly sampled.
(Default: 1000)
-a, --lmethod NAME
Linear algebra method for matrix inversion. Supported algorithms: default, svd (SVD – singular value decomposition, divide-and-conquer driver), svd2 (SVD – singular value decomposition, simple driver), lu (LU factorization), and ll (LL – Cholesky factorization).
(Default: default)
-r, --rcond NUMBER
Rank condition for SVD. The chosen value must be carefully tested. Calculation at computer precision is requested with -1 (not recommended).
(Default: 1e-06)
-s, --sigmaf2 SPEC
Variance of the reconstructed free-energy surface (signal variance) in the form SigmaF2(1)[xSigmaF2(2)[x...]]. The last value pads the rest.
(Default: 15.0)
--minsigmaf2 NUMBER
Minimum value of SigmaF2.
(Default: 0.1)
-o, --covar SPEC
Covariances between GPR tasks in the form CoVar1[xCoVar2[x...]]. The last value pads the rest.
(Default: 0.0)
--mincovar NUMBER
Minimum value of CoVar.
(Default: -1000)
-w, --wfac SPEC
Characteristic scale of collective variables in the form WFac1[xWFac2[x...]]. The last value pads the rest.
(Default: 3.0)
--minwfac NUMBER
Minimum value of WFac.
(Default: 0.1)
-n, --sigman2 SPEC
Noise variances (σ²) for each CV in the form SigmaN2(1)[xSigmaN2(2)[x...]]. The last value pads the rest.
(Default: 1e-5)
--minsigman2 NUMBER
Minimum value of SigmaN2.
(Default: 0)
--enablesigmaf2 SPEC
Enable optimization of SigmaF2 hyperparameters. Flags are specified in the form SigmaF2(1)Enabled[xSigmaF2(2)Enabled[x...]] with F and T for disabled and enabled, respectively. The last value pads the rest.
(Default: F)
--enablecovar SPEC
Enable optimization of CoVar hyperparameters. Flags are specified in the form CoVar1Enabled[xCoVar2Enabled[x...]] with F and T for disabled and enabled, respectively. The last value pads the rest.
(Default: F)
--enablewfac SPEC
Enable optimization of WFac hyperparameters. Flags are specified in the form WFac1Enabled[xWFac2Enabled[x...]] with F and T for disabled and enabled, respectively. The last value pads the rest.
(Default: F)
--enablesigman2 SPEC
Enable optimization of SigmaN2 hyperparameters. Flags are specified in the form SigmaN2(1)Enabled[xSigmaN2(2)Enabled[x...]] with F and T for disabled and enabled, respectively. The last value pads the rest.
(Default: F)
--numeric
Use numerical gradients.
--noptsteps NUMBER
Maximum number of optimization steps.
(Default: 100)
--nresets NUMBER
Maximum number of resets due to insufficient optimization progress.
(Default: 3)
--termeps NUMBER
Termination criterion for the L-BFGS optimizer (see L-BFGS code).
(Default: 1e-05)
--termval NUMBER
Termination criterion for the L-BFGS optimizer. Minimum change of the optimized property.
(Default: 1e-06)
--nlbfgscorr NUMBER
Number of corrections used in an L-BFGS update.
(Default: 10)
--test
Compare analytical and numerical gradients for input hyperparameters.
--cd5
Use the 5-point stencil for numerical differentiation.
--loadhyprms NAME
Name of the file containing the GPR hyperparameters.
--stat
Calculate detailed GPR statistics using the optimized hyperparameters.
--sptype
Determine the type of stationary point.
--kernel NAME
GPR kernel type. Supported types: ardse (ARD squared exponential), ardmc52 (ARD Matérn class 5/2), default (=ardse).
(Default: default)
--fdkernel
Use the first derivatives of the GPR kernel (only for SmootherGPR).
--useinv
Use the matrix inversion pathway (for testing only).
--calclogpl
Calculate logPL for --stat if --target is not logpl.
--globalmin SPEC
Position of the global minimum provided as a single string in the form CV1xCV2x...xCVn (relevant for error determination). If not set, the position is determined automatically.
-v, --verbose
Increase output verbosity.
--version
Output version information and exit.
-h, --help
Display this help and exit.