lluni/JavaNN
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

Added support for choosing weight and bias initializers

Browse source
This commit is contained in:
lluni 2022-05-28 03:19:49 +02:00
parent e7de925373
commit faa547564c
5 changed files with 36 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
src/main/java/de/lluni/javann/Initializer.java Normal file
View file

@ -0,0 +1,8 @@
package de.lluni.javann;
public enum Initializer {
ZEROS,
ONES,
GAUSSIAN,
RANDOM
}

7
src/main/java/de/lluni/javann/examples/ExampleSine.java
View file

@ -1,5 +1,6 @@
package de.lluni.javann.examples;
import de.lluni.javann.Initializer;
import de.lluni.javann.Network;
import de.lluni.javann.functions.ActivationFunctions;
import de.lluni.javann.functions.LossFunctions;
@ -46,11 +47,11 @@ public class ExampleSine {
// create network and add layers
Network network = new Network();
network.addLayer(new FCLayer(8));
network.addLayer(new FCLayer(8, Initializer.GAUSSIAN, Initializer.ONES));
network.addLayer(new ActivationLayer(ActivationFunctions::LeakyReLu, ActivationFunctions::LeakyReLuPrime));
network.addLayer(new FCLayer(8));
network.addLayer(new FCLayer(8, Initializer.GAUSSIAN, Initializer.ONES));
network.addLayer(new ActivationLayer(ActivationFunctions::LeakyReLu, ActivationFunctions::LeakyReLuPrime));
network.addLayer(new FCLayer(1));
network.addLayer(new FCLayer(1, Initializer.GAUSSIAN, Initializer.ONES));
// configure loss function for the network
network.use(LossFunctions::MSE, LossFunctions::MSEPrime);

5
src/main/java/de/lluni/javann/examples/ExampleXOR.java
View file

@ -1,5 +1,6 @@
package de.lluni.javann.examples;
import de.lluni.javann.Initializer;
import de.lluni.javann.Network;
import de.lluni.javann.functions.ActivationFunctions;
import de.lluni.javann.functions.LossFunctions;
@ -19,9 +20,9 @@ public class ExampleXOR {
new SimpleMatrix(new double[][]{{0}})};
Network network = new Network();
network.addLayer(new FCLayer(3));
network.addLayer(new FCLayer(3, Initializer.RANDOM, Initializer.RANDOM));
network.addLayer(new ActivationLayer(ActivationFunctions::tanh, ActivationFunctions::tanhPrime));
network.addLayer(new FCLayer(1));
network.addLayer(new FCLayer(1, Initializer.RANDOM, Initializer.RANDOM));
network.addLayer(new ActivationLayer(ActivationFunctions::tanh, ActivationFunctions::tanhPrime));
network.use(LossFunctions::MSE, LossFunctions::MSEPrime);

5
src/main/java/de/lluni/javann/examples/ExampleXORAddedNeurons.java
View file

@ -1,5 +1,6 @@
package de.lluni.javann.examples;
import de.lluni.javann.Initializer;
import de.lluni.javann.Network;
import de.lluni.javann.functions.ActivationFunctions;
import de.lluni.javann.functions.LossFunctions;
@ -19,9 +20,9 @@ public class ExampleXORAddedNeurons {
new SimpleMatrix(new double[][]{{0}})};
Network network = new Network();
network.addLayer(new FCLayer(1));
network.addLayer(new FCLayer(1, Initializer.RANDOM, Initializer.RANDOM));
network.addLayer(new ActivationLayer(ActivationFunctions::tanh, ActivationFunctions::tanhPrime));
network.addLayer(new FCLayer(1));
network.addLayer(new FCLayer(1, Initializer.RANDOM, Initializer.RANDOM));
network.addLayer(new ActivationLayer(ActivationFunctions::tanh, ActivationFunctions::tanhPrime));
network.addNeuron(0, 2);

21
src/main/java/de/lluni/javann/layers/FCLayer.java
View file

@ -1,5 +1,6 @@
package de.lluni.javann.layers;
import de.lluni.javann.Initializer;
import de.lluni.javann.util.Utilities;
import org.ejml.simple.SimpleMatrix;
@ -9,6 +10,8 @@ import java.util.Random;
* Fully connected layer with n Neurons
*/
public class FCLayer extends Layer {
private final Initializer weightInit;
private final Initializer biasInit;
private SimpleMatrix weights;
private SimpleMatrix biases;
private int numNeurons;
@ -18,14 +21,26 @@ public class FCLayer extends Layer {
* Creates a fully connected layer with numNeurons neurons
* @param numNeurons amount of neurons in this layer
*/
public FCLayer(int numNeurons) {
public FCLayer(int numNeurons, Initializer weightInit, Initializer biasInit) {
this.numNeurons = numNeurons;
this.weightInit = weightInit;
this.biasInit = biasInit;
isInitialized = false;
}
private void initialize(int inputSize) {
this.weights = Utilities.gaussianMatrix(inputSize, numNeurons, 0, 1, 0.1d);
this.biases = Utilities.ones(1, numNeurons);
switch (weightInit) {
case ZEROS -> this.weights = new SimpleMatrix(inputSize, numNeurons);
case ONES -> this.weights = Utilities.ones(inputSize, numNeurons);
case GAUSSIAN -> this.weights = Utilities.gaussianMatrix(inputSize, numNeurons, 0, 1, 0.1d);
case RANDOM -> this.weights = Utilities.randomMatrix(inputSize, numNeurons);
}
switch (biasInit) {
case ZEROS -> this.biases = new SimpleMatrix(1, numNeurons);
case ONES -> this.biases = Utilities.ones(1, numNeurons);
case GAUSSIAN -> this.biases = Utilities.gaussianMatrix(1, numNeurons, 0, 1, 0.1d);
case RANDOM -> this.biases = Utilities.randomMatrix(1, numNeurons);
}
this.isInitialized = true;
}