Package deepnetts.net.weights

Class RandomWeights

java.lang.Object

deepnetts.net.weights.RandomWeights

public class RandomWeights
extends Object

Weights randomization utility methods.

Constructor Summary

Constructors

Constructor	Description
RandomWeights()

Method Summary

Modifier and Type	Method	Description
static void	gaussian(float[] weights, float mean, float std)
static void	he(float[] weights, int numInputs)	He initialization, used for relu activations.zero-mean.
static void	initSeed(long seed)	Initializes random number generator with specified seed.
static void	normal(float[] weights)
static void	randomize(float[] array)	Initialize the elements of specified array with random numbers with uniform distribution in range [-0.5, 0.5].
static void	uniform(float[] weights, float min, float max)
static void	uniform(float[] weights, int numInputs)	Uniform U[-a,a] where a=1/sqrt(in).
static void	widrowHoff(float[] array, float input, float hidden)
static void	xavier(float[] weights, int numIn, int numOut)	Normalized uniform initialization U[-a,a] with a = sqrt(6/(in + out)).

Methods inherited from class java.lang.Object

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

Constructor Details

RandomWeights

public RandomWeights()

Method Details

initSeed

public static void initSeed(long seed)

Initializes random number generator with specified seed.

Parameters:

seed - init value for random generator

randomize

public static void randomize(float[] array)

Initialize the elements of specified array with random numbers with uniform distribution in range [-0.5, 0.5].

Parameters:

array - array to initialize

widrowHoff

public static void widrowHoff(float[] array,
 float input,
 float hidden)

uniform

public static void uniform(float[] weights,
 int numInputs)

Uniform U[-a,a] where a=1/sqrt(in). Understanding the difficulty of training deep feedforward neural networks Xavier Glorot, Yoshua Bengio, 2010, Commonly used "heuristic" , Eq. 1 pg 251 http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf http://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf This method is reffered to as commonly used heuristics in paper above.

Parameters:

weights - an array of weights to randomize

numInputs - a number of inputs from previous layer

uniform

public static void uniform(float[] weights,
 float min,
 float max)

he

public static void he(float[] weights,
 int numInputs)

He initialization, used for relu activations.zero-mean. Zero-mean Gaussian distribution whose standard deviation (std) is√2/nl. (eq 10 in paper below) Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification https://arxiv.org/pdf/1502.01852.pdf Koristi za Relu i leaky relu

Parameters:

weights - weights to initialize

numInputs - number of inputs

gaussian

public static void gaussian(float[] weights,
 float mean,
 float std)

normal

public static void normal(float[] weights)

xavier

public static void xavier(float[] weights,
 int numIn,
 int numOut)

Normalized uniform initialization U[-a,a] with a = sqrt(6/(in + out)). Properly scaled uniform distribution (normalized initialization). Sta treba u brojiocu 6 ili 1 u=i da li u imeniocu treba zbir in + out ili samo in ili da aimam obe verzije i pitanje je da li je ispod uniformna ili gausova distribucija Xavier Glorot, Yoshua Bengio, 2010, Understanding the difficulty of training deep feedforward neural networks http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf , pg. 253, eq 16. http://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf Use for tanh

Parameters:

weights -

numIn - size of the previous layer (number of inputs)

numOut - size of initialized layer (number of outputs) https://towardsdatascience.com/weight-initialization-techniques-in-neural-networks-26c649eb3b78