Specific training quantity (e.g. energies, forces, charges). More...

Collaboration diagram for nnp::Training::Property:

[legend]

Public Member Functions
	Property (std::string const &property)
	Constructor.

Public Attributes
std::string	property
	Copy of identifier within Property map.

std::string	displayMetric
	Error metric for display.

std::string	tiny
	Tiny abbreviation string for property.

std::string	plural
	Plural string of property;.

SelectionMode	selectionMode
	Selection mode for update candidates.

std::size_t	numTrainPatterns
	Number of training patterns in set.

std::size_t	numTestPatterns
	Number of training patterns in set.

std::size_t	taskBatchSize
	Batch size for each MPI task.

std::size_t	writeCompEvery
	Write comparison every this many epochs.

std::size_t	writeCompAlways
	Up to this epoch comparison is written every epoch.

std::size_t	posUpdateCandidates
	Current position in update candidate list (SM_SORT/_THRESHOLD).

std::size_t	numGroupedSubCand
	Number of subcandidates which are considered before changing the update candidate.

std::size_t	countGroupedSubCand
	Counter for number of used subcandidates belonging to same update candidate.

std::size_t	rmseThresholdTrials
	Maximum trials for SM_THRESHOLD selection mode.

std::size_t	countUpdates
	Counter for updates per epoch.

std::size_t	numUpdates
	Number of desired updates per epoch.

std::size_t	patternsPerUpdate
	Patterns used per update.

std::size_t	patternsPerUpdateGlobal
	Patterns used per update (summed over all MPI tasks).

std::size_t	numErrorsGlobal
	Global number of errors per update.

double	epochFraction
	Desired update fraction per epoch.

double	rmseThreshold
	RMSE threshold for update candidates.

std::vector< int >	errorsPerTask
	Errors per task for each update.

std::vector< int >	offsetPerTask
	Offset for combined error per task.

std::vector< std::string >	errorMetrics
	Error metrics available for this property.

std::map< std::string, double >	errorTrain
	Current error metrics of training patterns.

std::map< std::string, double >	errorTest
	Current error metrics of test patterns.

std::vector< UpdateCandidate >	updateCandidates
	Vector with indices of training patterns.

std::vector< std::vector< int > >	weightsPerTask
	Weights per task per updater.

std::vector< std::vector< int > >	offsetJacobian
	Stride for Jacobians per task per updater.

std::vector< std::vector< double > >	error
	Global error vector (per updater).

std::vector< std::vector< double > >	jacobian
	Global Jacobian (per updater).

std::map< std::size_t, SelectionMode >	selectionModeSchedule
	Schedule for varying selection mode.

Detailed Description

Specific training quantity (e.g. energies, forces, charges).

Definition at line 405 of file Training.h.

Constructor & Destructor Documentation

◆ Property()

Training::Property::Property ( std::string const & property )

Constructor.

Definition at line 4085 of file Training.cpp.

                                                 :
    property               (property ),
    displayMetric          (""       ),
    tiny                   (""       ),
    plural                 (""       ),
    selectionMode          (SM_RANDOM),
    numTrainPatterns       (0        ),
    numTestPatterns        (0        ),
    taskBatchSize          (0        ),
    writeCompEvery         (0        ),
    writeCompAlways        (0        ),
    posUpdateCandidates    (0        ),
    numGroupedSubCand      (0        ),
    countGroupedSubCand    (0        ),
    rmseThresholdTrials    (0        ),
    countUpdates           (0        ),
    numUpdates             (0        ),
    patternsPerUpdate      (0        ),
    patternsPerUpdateGlobal(0        ),
    numErrorsGlobal        (0        ),
    epochFraction          (0.0      ),
    rmseThreshold          (0.0      )
{
    if (property == "energy")
    {
        tiny = "E";
        plural = "energies";
        errorMetrics = {"RMSEpa", "RMSE", "MAEpa", "MAE"};
    }
    else if (property == "force")
    {
        tiny = "F";
        plural = "forces";
        errorMetrics = {"RMSE", "MAE"};
    }
    else if (property == "charge")
    {
        tiny = "Q";
        plural = "charges";
        errorMetrics = {"RMSE", "MAE"};
    }
    else
    {
        throw runtime_error("ERROR: Unknown training property.\n");
    }
 
    // Set up error metrics
    for (auto m : errorMetrics)
    {
        errorTrain[m] = 0.0;
        errorTest[m] = 0.0;
    }
    displayMetric = errorMetrics.at(0);
}

References countGroupedSubCand, countUpdates, displayMetric, epochFraction, errorMetrics, errorTest, errorTrain, numErrorsGlobal, numGroupedSubCand, numTestPatterns, numTrainPatterns, numUpdates, patternsPerUpdate, patternsPerUpdateGlobal, plural, posUpdateCandidates, property, rmseThreshold, rmseThresholdTrials, selectionMode, nnp::Training::SM_RANDOM, taskBatchSize, tiny, writeCompAlways, and writeCompEvery.

Member Data Documentation

◆ property

std::string nnp::Training::Property::property

Copy of identifier within Property map.

Definition at line 411 of file Training.h.

Referenced by Property().

◆ displayMetric

std::string nnp::Training::Property::displayMetric

Error metric for display.

Definition at line 413 of file Training.h.

Referenced by Property().

◆ tiny

std::string nnp::Training::Property::tiny

Tiny abbreviation string for property.

Definition at line 415 of file Training.h.

Referenced by Property().

◆ plural

std::string nnp::Training::Property::plural

Plural string of property;.

Definition at line 417 of file Training.h.

Referenced by Property().

◆ selectionMode

SelectionMode nnp::Training::Property::selectionMode

Selection mode for update candidates.

Definition at line 419 of file Training.h.

Referenced by Property(), and nnp::Training::update().

◆ numTrainPatterns

std::size_t nnp::Training::Property::numTrainPatterns

Number of training patterns in set.

Definition at line 421 of file Training.h.

Referenced by Property(), and nnp::Training::setupUpdatePlan().

◆ numTestPatterns

std::size_t nnp::Training::Property::numTestPatterns

Number of training patterns in set.

Definition at line 423 of file Training.h.

Referenced by Property().

◆ taskBatchSize

std::size_t nnp::Training::Property::taskBatchSize

Batch size for each MPI task.

Definition at line 425 of file Training.h.

Referenced by Property(), nnp::Training::setupUpdatePlan(), and nnp::Training::update().

◆ writeCompEvery

std::size_t nnp::Training::Property::writeCompEvery

Write comparison every this many epochs.

Definition at line 427 of file Training.h.

Referenced by Property().

◆ writeCompAlways

std::size_t nnp::Training::Property::writeCompAlways

Up to this epoch comparison is written every epoch.

Definition at line 429 of file Training.h.

Referenced by Property().

◆ posUpdateCandidates

std::size_t nnp::Training::Property::posUpdateCandidates

Current position in update candidate list (SM_SORT/_THRESHOLD).

Definition at line 431 of file Training.h.

Referenced by Property(), and nnp::Training::update().

◆ numGroupedSubCand

std::size_t nnp::Training::Property::numGroupedSubCand

Number of subcandidates which are considered before changing the update candidate.

Definition at line 434 of file Training.h.

Referenced by Property(), and nnp::Training::update().

◆ countGroupedSubCand

std::size_t nnp::Training::Property::countGroupedSubCand

Counter for number of used subcandidates belonging to same update candidate.

Definition at line 437 of file Training.h.

Referenced by Property(), and nnp::Training::update().

◆ rmseThresholdTrials

std::size_t nnp::Training::Property::rmseThresholdTrials

Maximum trials for SM_THRESHOLD selection mode.

Definition at line 439 of file Training.h.

Referenced by Property(), and nnp::Training::update().

◆ countUpdates

std::size_t nnp::Training::Property::countUpdates

Counter for updates per epoch.

Definition at line 441 of file Training.h.

Referenced by Property(), and nnp::Training::update().

◆ numUpdates

std::size_t nnp::Training::Property::numUpdates

Number of desired updates per epoch.

Definition at line 443 of file Training.h.

Referenced by Property(), nnp::Training::setupTraining(), and nnp::Training::setupUpdatePlan().

◆ patternsPerUpdate

std::size_t nnp::Training::Property::patternsPerUpdate

Patterns used per update.

Definition at line 445 of file Training.h.

Referenced by Property(), nnp::Training::setupUpdatePlan(), and nnp::Training::update().

◆ patternsPerUpdateGlobal

std::size_t nnp::Training::Property::patternsPerUpdateGlobal

Patterns used per update (summed over all MPI tasks).

Definition at line 447 of file Training.h.

Referenced by Property(), nnp::Training::setupTraining(), and nnp::Training::setupUpdatePlan().

◆ numErrorsGlobal

std::size_t nnp::Training::Property::numErrorsGlobal

Global number of errors per update.

Definition at line 449 of file Training.h.

Referenced by nnp::Training::allocateArrays(), Property(), and nnp::Training::setupUpdatePlan().

◆ epochFraction

double nnp::Training::Property::epochFraction

Desired update fraction per epoch.

Definition at line 451 of file Training.h.

Referenced by Property(), nnp::Training::setupUpdatePlan(), and nnp::Training::update().

◆ rmseThreshold

double nnp::Training::Property::rmseThreshold

RMSE threshold for update candidates.

Definition at line 453 of file Training.h.

Referenced by Property(), and nnp::Training::update().

◆ errorsPerTask

std::vector<int> nnp::Training::Property::errorsPerTask

Errors per task for each update.

Definition at line 455 of file Training.h.

Referenced by nnp::Training::allocateArrays(), nnp::Training::setupUpdatePlan(), and nnp::Training::update().

◆ offsetPerTask

std::vector<int> nnp::Training::Property::offsetPerTask

Offset for combined error per task.

Definition at line 457 of file Training.h.

Referenced by nnp::Training::setupUpdatePlan(), and nnp::Training::update().

◆ errorMetrics

std::vector<std::string> nnp::Training::Property::errorMetrics

Error metrics available for this property.

Definition at line 459 of file Training.h.

Referenced by Property().

◆ errorTrain

std::map< std::string, double> nnp::Training::Property::errorTrain

Current error metrics of training patterns.

Definition at line 462 of file Training.h.

Referenced by Property(), and nnp::Training::update().

◆ errorTest

std::map< std::string, double> nnp::Training::Property::errorTest

Current error metrics of test patterns.

Definition at line 465 of file Training.h.

Referenced by Property().

◆ updateCandidates

std::vector<UpdateCandidate> nnp::Training::Property::updateCandidates

Vector with indices of training patterns.

Definition at line 467 of file Training.h.

Referenced by nnp::Training::setupUpdatePlan(), and nnp::Training::update().

◆ weightsPerTask

std::vector< std::vector<int> > nnp::Training::Property::weightsPerTask

Weights per task per updater.

Definition at line 470 of file Training.h.

Referenced by nnp::Training::setupUpdatePlan(), and nnp::Training::update().

◆ offsetJacobian

std::vector< std::vector<int> > nnp::Training::Property::offsetJacobian

Stride for Jacobians per task per updater.

Definition at line 473 of file Training.h.

Referenced by nnp::Training::setupUpdatePlan(), and nnp::Training::update().

◆ error

std::vector< std::vector<double> > nnp::Training::Property::error

Global error vector (per updater).

Definition at line 476 of file Training.h.

Referenced by nnp::Training::allocateArrays(), and nnp::Training::update().

◆ jacobian

std::vector< std::vector<double> > nnp::Training::Property::jacobian

Global Jacobian (per updater).

Definition at line 479 of file Training.h.

Referenced by nnp::Training::allocateArrays(), and nnp::Training::update().

◆ selectionModeSchedule

std::map< std::size_t, SelectionMode> nnp::Training::Property::selectionModeSchedule

Schedule for varying selection mode.

Definition at line 482 of file Training.h.

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

/home/runner/work/n2p2/n2p2/src/libnnptrain/Training.h
/home/runner/work/n2p2/n2p2/src/libnnptrain/Training.cpp

Public Member Functions

Public Attributes

Detailed Description

Constructor & Destructor Documentation

◆ Property()

Member Data Documentation

◆ property

◆ displayMetric

◆ tiny

◆ plural

◆ selectionMode

◆ numTrainPatterns

◆ numTestPatterns

◆ taskBatchSize

◆ writeCompEvery

◆ writeCompAlways

◆ posUpdateCandidates

◆ numGroupedSubCand

◆ countGroupedSubCand

◆ rmseThresholdTrials

◆ countUpdates

◆ numUpdates

◆ patternsPerUpdate

◆ patternsPerUpdateGlobal

◆ numErrorsGlobal

◆ epochFraction

◆ rmseThreshold

◆ errorsPerTask

◆ offsetPerTask

◆ errorMetrics

◆ errorTrain

◆ errorTest

◆ updateCandidates

◆ weightsPerTask

◆ offsetJacobian

◆ error

◆ jacobian

◆ selectionModeSchedule