Added support for adding new neurons
This commit is contained in:
parent
7738781bb5
commit
8c82838c54
3 changed files with 136 additions and 9 deletions
88
src/main/java/BlankLayer.java
Normal file
88
src/main/java/BlankLayer.java
Normal file
|
@ -0,0 +1,88 @@
|
||||||
|
import org.ejml.simple.SimpleMatrix;
|
||||||
|
|
||||||
|
import java.util.Random;
|
||||||
|
|
||||||
|
/**
|
||||||
|
* Layer initialized with 1 neuron.
|
||||||
|
* Assumes that each new neuron is fully connected to every previous neuron (this will be changed in the future).
|
||||||
|
*/
|
||||||
|
public class BlankLayer extends Layer {
|
||||||
|
SimpleMatrix weights;
|
||||||
|
SimpleMatrix biases;
|
||||||
|
|
||||||
|
public BlankLayer(int inputSize) {
|
||||||
|
Random random = new Random();
|
||||||
|
this.weights = new SimpleMatrix(inputSize, 1, true,
|
||||||
|
random.doubles(inputSize, -1, 1).toArray());
|
||||||
|
this.biases = new SimpleMatrix(1, 1, true,
|
||||||
|
random.doubles(1, -1, 1).toArray());
|
||||||
|
}
|
||||||
|
|
||||||
|
/**
|
||||||
|
* Updates input size when previous layer has newly added neurons.
|
||||||
|
* @param n amount of new neurons in previous layer
|
||||||
|
*/
|
||||||
|
public void updateInputSize(int n) {
|
||||||
|
Random random = new Random();
|
||||||
|
|
||||||
|
// add new weights
|
||||||
|
SimpleMatrix newWeights = new SimpleMatrix(this.weights.numRows() + n, this.weights.numCols());
|
||||||
|
for (int i = 0; i < this.weights.numRows(); i++) {
|
||||||
|
for (int j = 0; j < this.weights.numCols(); j++) {
|
||||||
|
newWeights.set(i, j, this.weights.get(i, j));
|
||||||
|
}
|
||||||
|
}
|
||||||
|
for (int i = 0; i < newWeights.getNumElements(); i++) {
|
||||||
|
if (newWeights.get(i) == 0) {
|
||||||
|
newWeights.set(i, random.nextDouble(-1, 1));
|
||||||
|
}
|
||||||
|
}
|
||||||
|
this.weights = newWeights;
|
||||||
|
}
|
||||||
|
|
||||||
|
/**
|
||||||
|
* Adds new neurons at the end of the layer
|
||||||
|
* @param n amount how many new neurons should be added
|
||||||
|
*/
|
||||||
|
public void addNeuron(int n) {
|
||||||
|
Random random = new Random();
|
||||||
|
|
||||||
|
// add new weights
|
||||||
|
SimpleMatrix newWeights = new SimpleMatrix(this.weights.numRows(), this.weights.numCols() + n);
|
||||||
|
for (int i = 0; i < this.weights.numRows(); i++) {
|
||||||
|
for (int j = 0; j < this.weights.numCols(); j++) {
|
||||||
|
newWeights.set(i, j, this.weights.get(i, j));
|
||||||
|
}
|
||||||
|
}
|
||||||
|
for (int i = 0; i < newWeights.getNumElements(); i++) {
|
||||||
|
if (newWeights.get(i) == 0) {
|
||||||
|
newWeights.set(i, random.nextDouble(-1, 1));
|
||||||
|
}
|
||||||
|
}
|
||||||
|
this.weights = newWeights;
|
||||||
|
|
||||||
|
// add new biases
|
||||||
|
SimpleMatrix newBiases = new SimpleMatrix(1, this.biases.numCols() + n);
|
||||||
|
double[] newBiasValues = random.doubles(n, -1, 1).toArray();
|
||||||
|
System.arraycopy(this.biases.getDDRM().data, 0, newBiases.getDDRM().data, 0, this.biases.numCols());
|
||||||
|
System.arraycopy(newBiasValues, 0, newBiases.getDDRM().data, this.biases.numCols(), n);
|
||||||
|
this.biases = newBiases;
|
||||||
|
}
|
||||||
|
|
||||||
|
@Override
|
||||||
|
public SimpleMatrix forwardPropagation(SimpleMatrix inputs) {
|
||||||
|
this.input = inputs;
|
||||||
|
this.output = this.input.mult(this.weights).plus(this.biases);
|
||||||
|
return this.output;
|
||||||
|
}
|
||||||
|
|
||||||
|
@Override
|
||||||
|
public SimpleMatrix backwardPropagation(SimpleMatrix outputError, double learningRate) {
|
||||||
|
SimpleMatrix inputError = outputError.mult(this.weights.transpose());
|
||||||
|
SimpleMatrix weightsError = this.input.transpose().mult(outputError);
|
||||||
|
|
||||||
|
this.weights = this.weights.plus(learningRate, weightsError);
|
||||||
|
this.biases = this.biases.plus(learningRate, outputError);
|
||||||
|
return inputError;
|
||||||
|
}
|
||||||
|
}
|
|
@ -12,10 +12,10 @@ public class ExampleXOR {
|
||||||
new SimpleMatrix(new double[][]{{0}})};
|
new SimpleMatrix(new double[][]{{0}})};
|
||||||
|
|
||||||
Network network = new Network();
|
Network network = new Network();
|
||||||
network.add(new FCLayer(2, 3));
|
network.addLayer(new FCLayer(2, 3));
|
||||||
network.add(new ActivationLayer(ActivationFunctions::tanh, ActivationFunctions::tanhPrime));
|
network.addLayer(new ActivationLayer(ActivationFunctions::tanh, ActivationFunctions::tanhPrime));
|
||||||
network.add(new FCLayer(3, 1));
|
network.addLayer(new FCLayer(3, 1));
|
||||||
network.add(new ActivationLayer(ActivationFunctions::tanh, ActivationFunctions::tanhPrime));
|
network.addLayer(new ActivationLayer(ActivationFunctions::tanh, ActivationFunctions::tanhPrime));
|
||||||
|
|
||||||
network.use(LossFunctions::MSE, LossFunctions::MSEPrime);
|
network.use(LossFunctions::MSE, LossFunctions::MSEPrime);
|
||||||
network.fit(X_train, y_train, 1000, 0.1d);
|
network.fit(X_train, y_train, 1000, 0.1d);
|
||||||
|
|
|
@ -4,19 +4,34 @@ import java.util.ArrayList;
|
||||||
import java.util.function.BiFunction;
|
import java.util.function.BiFunction;
|
||||||
|
|
||||||
public class Network {
|
public class Network {
|
||||||
|
private ArrayList<Layer> layers;
|
||||||
ArrayList<Layer> layers;
|
private BiFunction<SimpleMatrix, SimpleMatrix, Double> loss;
|
||||||
BiFunction<SimpleMatrix, SimpleMatrix, Double> loss;
|
private BiFunction<SimpleMatrix, SimpleMatrix, SimpleMatrix> lossPrime;
|
||||||
BiFunction<SimpleMatrix, SimpleMatrix, SimpleMatrix> lossPrime;
|
|
||||||
|
|
||||||
public Network() {
|
public Network() {
|
||||||
layers = new ArrayList<>();
|
layers = new ArrayList<>();
|
||||||
}
|
}
|
||||||
|
|
||||||
public void add(Layer layer) {
|
public void addLayer(Layer layer) {
|
||||||
layers.add(layer);
|
layers.add(layer);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
/**
|
||||||
|
* Adds n neurons to a specific layer and also updates this and the next layer's weights and biases.
|
||||||
|
* Only works if there are two successive BlankLayers.
|
||||||
|
* @param layer index of layer in the ArrayList layers
|
||||||
|
* @param n amount how many new neurons should be added
|
||||||
|
*/
|
||||||
|
public void addNeuron(int layer, int n) {
|
||||||
|
if (!(this.layers.get(layer) instanceof BlankLayer)) {
|
||||||
|
System.out.println("This layer is not a BlankLayer");
|
||||||
|
} else if (!(this.layers.get(layer + 2) instanceof BlankLayer)) {
|
||||||
|
System.out.println("The next layer is not a BlankLayer");
|
||||||
|
}
|
||||||
|
((BlankLayer) this.layers.get(layer)).addNeuron(n);
|
||||||
|
((BlankLayer) this.layers.get(layer + 2)).updateInputSize(n);
|
||||||
|
}
|
||||||
|
|
||||||
public void use(BiFunction<SimpleMatrix, SimpleMatrix, Double> loss, BiFunction<SimpleMatrix, SimpleMatrix, SimpleMatrix> lossPrime) {
|
public void use(BiFunction<SimpleMatrix, SimpleMatrix, Double> loss, BiFunction<SimpleMatrix, SimpleMatrix, SimpleMatrix> lossPrime) {
|
||||||
this.loss = loss;
|
this.loss = loss;
|
||||||
this.lossPrime = lossPrime;
|
this.lossPrime = lossPrime;
|
||||||
|
@ -65,4 +80,28 @@ public class Network {
|
||||||
System.out.println("epoch " + (i+1) + "/" + epochs + " error=" + err);
|
System.out.println("epoch " + (i+1) + "/" + epochs + " error=" + err);
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
|
public ArrayList<Layer> getLayers() {
|
||||||
|
return layers;
|
||||||
|
}
|
||||||
|
|
||||||
|
public void setLayers(ArrayList<Layer> layers) {
|
||||||
|
this.layers = layers;
|
||||||
|
}
|
||||||
|
|
||||||
|
public BiFunction<SimpleMatrix, SimpleMatrix, Double> getLoss() {
|
||||||
|
return loss;
|
||||||
|
}
|
||||||
|
|
||||||
|
public void setLoss(BiFunction<SimpleMatrix, SimpleMatrix, Double> loss) {
|
||||||
|
this.loss = loss;
|
||||||
|
}
|
||||||
|
|
||||||
|
public BiFunction<SimpleMatrix, SimpleMatrix, SimpleMatrix> getLossPrime() {
|
||||||
|
return lossPrime;
|
||||||
|
}
|
||||||
|
|
||||||
|
public void setLossPrime(BiFunction<SimpleMatrix, SimpleMatrix, SimpleMatrix> lossPrime) {
|
||||||
|
this.lossPrime = lossPrime;
|
||||||
|
}
|
||||||
}
|
}
|
Loading…
Reference in a new issue