org.deeplearning4j.optimize.solvers

Class BackTrackLineSearch

java.lang.Object

org.deeplearning4j.optimize.solvers.BackTrackLineSearch

All Implemented Interfaces:

Serializable, LineOptimizer

public class BackTrackLineSearch
extends Object
implements LineOptimizer

Author:

Aron Culotta [email protected] Adapted from mallet with original authors above. Modified to be a vectorized version that uses jblas matrices for computation rather than the mallet ops. Numerical Recipes in C: p.385. lnsrch. A simple backtracking line search. No attempt at accurately finding the true minimum is made. The goal is only to ensure that BackTrackLineSearch will return a position of higher value., Adam Gibson

See Also:

Serialized Form

Field Summary

Fields

Modifier and Type	Field and Description
protected double	ALF

Constructor Summary

Constructors

Constructor and Description
BackTrackLineSearch(Model optimizable, ConvexOptimizer optimizer)
BackTrackLineSearch(Model layer, StepFunction stepFunction, ConvexOptimizer optimizer)

Method Summary

Modifier and Type	Method and Description
int	getMaxIterations()
double	getStepMax()
double	optimize(org.nd4j.linalg.api.ndarray.INDArray parameters, org.nd4j.linalg.api.ndarray.INDArray gradients, org.nd4j.linalg.api.ndarray.INDArray searchDirection) Line optimizer
void	setAbsTolx(double tolx) Sets the tolerance of absolute diff in function value.
void	setMaxIterations(int maxIterations)
void	setRelTolx(double tolx) Sets the tolerance of relative diff in function value.
double	setScoreFor(org.nd4j.linalg.api.ndarray.INDArray parameters)
void	setStepMax(double stepMax)

Methods inherited from class java.lang.Object

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

Field Detail

ALF

protected final double ALF

See Also:

Constant Field Values

Constructor Detail

BackTrackLineSearch

public BackTrackLineSearch(Model layer,
                           StepFunction stepFunction,
                           ConvexOptimizer optimizer)

Parameters:

layer -

stepFunction -

optimizer -

BackTrackLineSearch

public BackTrackLineSearch(Model optimizable,
                           ConvexOptimizer optimizer)

Parameters:

optimizable -

optimizer -

Method Detail

setStepMax

public void setStepMax(double stepMax)

getStepMax

public double getStepMax()

setRelTolx

public void setRelTolx(double tolx)

Sets the tolerance of relative diff in function value. Line search converges if abs(delta x / x) < tolx for all coordinates.

setAbsTolx

public void setAbsTolx(double tolx)

Sets the tolerance of absolute diff in function value. Line search converges if abs(delta x) < tolx for all coordinates.

getMaxIterations

public int getMaxIterations()

setMaxIterations

public void setMaxIterations(int maxIterations)

setScoreFor

public double setScoreFor(org.nd4j.linalg.api.ndarray.INDArray parameters)

optimize

public double optimize(org.nd4j.linalg.api.ndarray.INDArray parameters,
                       org.nd4j.linalg.api.ndarray.INDArray gradients,
                       org.nd4j.linalg.api.ndarray.INDArray searchDirection)
                throws InvalidStepException

Description copied from interface: LineOptimizer

Line optimizer

Specified by:

optimize in interface LineOptimizer

Parameters:

parameters - the parameters to optimize

gradients - the line/rate of change

searchDirection - the point for the line search to go in

Returns:

the next step size

Throws:

InvalidStepException