org.deeplearning4j.nn.weights

Class WeightInitUtil

java.lang.Object

org.deeplearning4j.nn.weights.WeightInitUtil

public class WeightInitUtil
extends Object

Weight initialization utility

Author:

Adam Gibson

Field Summary

Fields

Modifier and Type	Field and Description
static char	DEFAULT_WEIGHT_INIT_ORDER Default order for the arrays created by WeightInitUtil.

Method Summary

Modifier and Type	Method and Description
static org.nd4j.linalg.api.ndarray.INDArray	initWeights(int[] shape, float min, float max)
static org.nd4j.linalg.api.ndarray.INDArray	initWeights(int[] shape, WeightInit initScheme, org.nd4j.linalg.api.rng.distribution.Distribution dist, char order, org.nd4j.linalg.api.ndarray.INDArray paramView)
static org.nd4j.linalg.api.ndarray.INDArray	initWeights(int[] shape, WeightInit initScheme, org.nd4j.linalg.api.rng.distribution.Distribution dist, org.nd4j.linalg.api.ndarray.INDArray paramView) Initializes a matrix with the given weight initialization scheme.
static org.nd4j.linalg.api.ndarray.INDArray	initWeights(int nIn, int nOut, WeightInit initScheme, org.nd4j.linalg.api.rng.distribution.Distribution dist, org.nd4j.linalg.api.ndarray.INDArray paramView) Initializes a matrix with the given weight initialization scheme
static org.nd4j.linalg.api.ndarray.INDArray	reshapeWeights(int[] shape, org.nd4j.linalg.api.ndarray.INDArray paramsView) Reshape the parameters view, without modifying the paramsView array values.
static org.nd4j.linalg.api.ndarray.INDArray	reshapeWeights(int[] shape, org.nd4j.linalg.api.ndarray.INDArray paramsView, char flatteningOrder) Reshape the parameters view, without modifying the paramsView array values.
static org.nd4j.linalg.api.ndarray.INDArray	uniformBasedOnInAndOut(int[] shape, int nIn, int nOut) Generate a random matrix with respect to the number of inputs and outputs.

Methods inherited from class java.lang.Object

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

Field Detail

DEFAULT_WEIGHT_INIT_ORDER

public static final char DEFAULT_WEIGHT_INIT_ORDER

Default order for the arrays created by WeightInitUtil.

See Also:

Constant Field Values

Method Detail

uniformBasedOnInAndOut

public static org.nd4j.linalg.api.ndarray.INDArray uniformBasedOnInAndOut(int[] shape,
                                                                          int nIn,
                                                                          int nOut)

Generate a random matrix with respect to the number of inputs and outputs. This is a bound uniform distribution with the specified minimum and maximum

Parameters:

shape - the shape of the matrix

nIn - the number of inputs

nOut - the number of outputs

Returns:

INDArray

initWeights

public static org.nd4j.linalg.api.ndarray.INDArray initWeights(int[] shape,
                                                               float min,
                                                               float max)

initWeights

public static org.nd4j.linalg.api.ndarray.INDArray initWeights(int[] shape,
                                                               WeightInit initScheme,
                                                               org.nd4j.linalg.api.rng.distribution.Distribution dist,
                                                               org.nd4j.linalg.api.ndarray.INDArray paramView)

Initializes a matrix with the given weight initialization scheme. Note: Defaults to fortran ('f') order arrays for the weights. Use initWeights(int[], WeightInit, Distribution, char, INDArray) to control this

Parameters:

shape - the shape of the matrix

initScheme - the scheme to use

Returns:

a matrix of the specified dimensions with the specified distribution based on the initialization scheme

initWeights

public static org.nd4j.linalg.api.ndarray.INDArray initWeights(int[] shape,
                                                               WeightInit initScheme,
                                                               org.nd4j.linalg.api.rng.distribution.Distribution dist,
                                                               char order,
                                                               org.nd4j.linalg.api.ndarray.INDArray paramView)

initWeights

public static org.nd4j.linalg.api.ndarray.INDArray initWeights(int nIn,
                                                               int nOut,
                                                               WeightInit initScheme,
                                                               org.nd4j.linalg.api.rng.distribution.Distribution dist,
                                                               org.nd4j.linalg.api.ndarray.INDArray paramView)

Initializes a matrix with the given weight initialization scheme

Parameters:

nIn - the number of rows in the matrix

nOut - the number of columns in the matrix

initScheme - the scheme to use

Returns:

a matrix of the specified dimensions with the specified distribution based on the initialization scheme

reshapeWeights

public static org.nd4j.linalg.api.ndarray.INDArray reshapeWeights(int[] shape,
                                                                  org.nd4j.linalg.api.ndarray.INDArray paramsView)

Reshape the parameters view, without modifying the paramsView array values. Same reshaping mechanism as initWeights(int[], WeightInit, Distribution, INDArray)

Parameters:

shape - Shape to reshape

paramsView - Parameters array view

reshapeWeights

public static org.nd4j.linalg.api.ndarray.INDArray reshapeWeights(int[] shape,
                                                                  org.nd4j.linalg.api.ndarray.INDArray paramsView,
                                                                  char flatteningOrder)

Reshape the parameters view, without modifying the paramsView array values. Same reshaping mechanism as initWeights(int[], WeightInit, Distribution, char, INDArray)

Parameters:

shape - Shape to reshape

paramsView - Parameters array view

flatteningOrder - Order in which parameters are flattened/reshaped