[1] | 1 | package geniusweb.blingbling.Ranknet;
|
---|
| 2 |
|
---|
| 3 | //import neuralnet.activationfunction.IActivationFunction;
|
---|
| 4 | import org.nd4j.linalg.api.ndarray.INDArray;
|
---|
| 5 | import org.nd4j.linalg.factory.Nd4j;
|
---|
| 6 | //import org.nd4j.shade.guava.collect.Lists;
|
---|
| 7 |
|
---|
| 8 | import java.util.LinkedList;
|
---|
| 9 | import java.util.List;
|
---|
| 10 |
|
---|
| 11 | public class Layer {
|
---|
| 12 | private INDArray weights;
|
---|
| 13 | private INDArray biases;
|
---|
| 14 | private IActivationFunction activationFunction;
|
---|
| 15 |
|
---|
| 16 | // Iteration instance variables.
|
---|
| 17 | private INDArray input;
|
---|
| 18 | private INDArray z;
|
---|
| 19 | private INDArray activation;
|
---|
| 20 | private INDArray activationDerivative;
|
---|
| 21 |
|
---|
| 22 | public static LayerBuilder Builder() {
|
---|
| 23 | return new LayerBuilder();
|
---|
| 24 | }
|
---|
| 25 |
|
---|
| 26 | protected Layer(INDArray weights, INDArray biases, IActivationFunction activationFunction) {
|
---|
| 27 | this.weights = weights;
|
---|
| 28 | this.biases = biases;
|
---|
| 29 | this.activationFunction = activationFunction;
|
---|
| 30 | }
|
---|
| 31 |
|
---|
| 32 | public void activate(INDArray input) {
|
---|
| 33 | this.input = input;
|
---|
| 34 | z = calculateZ(input);
|
---|
| 35 | activation = calculateActivation(z);
|
---|
| 36 | activationDerivative = calculateActivationDerivative(z);
|
---|
| 37 | }
|
---|
| 38 |
|
---|
| 39 | public INDArray getInput() {
|
---|
| 40 | return input;
|
---|
| 41 | }
|
---|
| 42 |
|
---|
| 43 | public INDArray getZ() {
|
---|
| 44 | return z;
|
---|
| 45 | }
|
---|
| 46 |
|
---|
| 47 | public INDArray getActivation() {
|
---|
| 48 | return activation;
|
---|
| 49 | }
|
---|
| 50 |
|
---|
| 51 | public INDArray getActivationDerivative() {
|
---|
| 52 | return activationDerivative;
|
---|
| 53 | }
|
---|
| 54 |
|
---|
| 55 | public INDArray getErrorGradient(INDArray error, INDArray prevActivationDerivative) {
|
---|
| 56 | return error.mmul(weights).mul(prevActivationDerivative);
|
---|
| 57 | }
|
---|
| 58 |
|
---|
| 59 | public void updateWeights(INDArray gradients) {
|
---|
| 60 | List<INDArray> rows = new LinkedList<INDArray>();
|
---|
| 61 | int stride = activation.rows() * weights.rows();
|
---|
| 62 |
|
---|
| 63 | for (int row = 0; row < gradients.rows(); row = row + stride) {
|
---|
| 64 | int[] extractRows = new int[stride];
|
---|
| 65 | for (int i = 0; i < stride; i++) {
|
---|
| 66 | extractRows[i] = (row * stride) + i;
|
---|
| 67 | }
|
---|
| 68 | rows.add(gradients.getRows(extractRows));
|
---|
| 69 | }
|
---|
| 70 | weights = weights.add(Nd4j.averageAndPropagate(rows));
|
---|
| 71 | }
|
---|
| 72 |
|
---|
| 73 | public void updateBiases(INDArray gradients) {
|
---|
| 74 | List<INDArray> rows = new LinkedList<INDArray>();
|
---|
| 75 | int stride = activation.rows();
|
---|
| 76 |
|
---|
| 77 | for (int row = 0; row < gradients.rows(); row = row + stride) {
|
---|
| 78 | int[] extractRows = new int[stride];
|
---|
| 79 | for (int i = 0; i < stride; i++) {
|
---|
| 80 | extractRows[i] = (row * stride) + i;
|
---|
| 81 | }
|
---|
| 82 | rows.add(gradients.getRows(extractRows));
|
---|
| 83 | }
|
---|
| 84 | biases = biases.add(Nd4j.averageAndPropagate(rows));
|
---|
| 85 | }
|
---|
| 86 |
|
---|
| 87 | public INDArray calculateZ(INDArray input) {
|
---|
| 88 | INDArray biasMatrix = biases;
|
---|
| 89 | for (int i = 1; i < input.rows(); i++) {
|
---|
| 90 | biasMatrix = Nd4j.hstack(biasMatrix, biases);
|
---|
| 91 | }
|
---|
| 92 | return input.mmul(weights.transpose()).add(biasMatrix);
|
---|
| 93 | }
|
---|
| 94 |
|
---|
| 95 | public INDArray calculateActivation(INDArray z) {
|
---|
| 96 | return activationFunction.output(z);
|
---|
| 97 | }
|
---|
| 98 |
|
---|
| 99 | public INDArray calculateActivationDerivative(INDArray z) {
|
---|
| 100 | return activationFunction.derivative(z);
|
---|
| 101 | }
|
---|
| 102 | } |
---|