edu.stanford.nlp.loglinear.learning

Class BacktrackingAdaGradOptimizer

java.lang.Object

edu.stanford.nlp.loglinear.learning.AbstractBatchOptimizer

edu.stanford.nlp.loglinear.learning.BacktrackingAdaGradOptimizer

public class BacktrackingAdaGradOptimizer
extends AbstractBatchOptimizer

Created on 8/26/15.

Author:

keenon

Handles optimizing an AbstractDifferentiableFunction through AdaGrad guarded by backtracking.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
protected class	BacktrackingAdaGradOptimizer.AdaGradOptimizationState

Nested classes/interfaces inherited from class edu.stanford.nlp.loglinear.learning.AbstractBatchOptimizer

AbstractBatchOptimizer.OptimizationState

Constructor Summary

Constructors

Constructor and Description
BacktrackingAdaGradOptimizer()

Method Summary

Modifier and Type	Method and Description
protected AbstractBatchOptimizer.OptimizationState	getFreshOptimizationState(ConcatVector initialWeights) This is called at the beginning of each batch optimization.
boolean	updateWeights(ConcatVector weights, ConcatVector gradient, double logLikelihood, AbstractBatchOptimizer.OptimizationState optimizationState, boolean quiet) This is the hook for subclassing batch optimizers to override in order to have their optimizer work.

Methods inherited from class edu.stanford.nlp.loglinear.learning.AbstractBatchOptimizer

addDenseConstraint, addSparseConstraint, optimize, optimize

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

BacktrackingAdaGradOptimizer

public BacktrackingAdaGradOptimizer()

Method Detail

updateWeights

public boolean updateWeights(ConcatVector weights,
                             ConcatVector gradient,
                             double logLikelihood,
                             AbstractBatchOptimizer.OptimizationState optimizationState,
                             boolean quiet)

Description copied from class: AbstractBatchOptimizer

This is the hook for subclassing batch optimizers to override in order to have their optimizer work.

Specified by:

updateWeights in class AbstractBatchOptimizer

Parameters:

weights - the current weights (update these in place)

gradient - the gradient at these weights

logLikelihood - the log likelihood at these weights

optimizationState - any saved state the optimizer wants to keep and pass around during each optimization run

quiet - whether or not to dump output about progress to the console

Returns:

whether or not we've converged

getFreshOptimizationState

protected AbstractBatchOptimizer.OptimizationState getFreshOptimizationState(ConcatVector initialWeights)

Description copied from class: AbstractBatchOptimizer

This is called at the beginning of each batch optimization. It should return a fresh OptimizationState object that will then be handed to updateWeights() on each update.

Specified by:

getFreshOptimizationState in class AbstractBatchOptimizer

Parameters:

initialWeights - the initial weights for the optimizer to use

Returns:

a fresh OptimizationState