name-undecided/lib/NeuralNetwork.js

A toy Neural Network library.

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
exports.NeuralNetwork = void 0;
var Matrix_1 = require("./Matrix");
// Declare map functions
function random(_x) {
    return Math.random();
}
function dsigmoid(x) {
    return x * (1 - x);
}
function sigmoid(x) {
    return 1 / (1 + Math.exp(-x));
}
function relu(x) {
    return x > 0 ? x : 0;
}
function squared(x) {
    return x * x;
}
var NeuralNetwork = /** @class */ (function () {
    function NeuralNetwork(inNum, hidNum, outNum, lr) {
        this.inputNum = inNum;
        this.hiddenNum = hidNum;
        this.outputNum = outNum;
        this.weightsIH = new Matrix_1.Matrix(this.hiddenNum, this.inputNum);
        this.weightsIH.map(random);
        this.weightsHO = new Matrix_1.Matrix(this.outputNum, this.hiddenNum);
        this.weightsHO.map(random);
        this.biasIH = new Matrix_1.Matrix(this.hiddenNum, 1);
        this.biasIH.map(random);
        this.biasHO = new Matrix_1.Matrix(this.outputNum, 1);
        this.biasHO.map(random);
        this.lr = lr;
    }
    NeuralNetwork.prototype.feedForward = function (inputs) {
        // Convert inputs into a Matrix object
        var inputsMatrix = Matrix_1.Matrix.fromArray(inputs, this.inputNum, 1);
        // Calculate the hidden layer output
        var hiddenInputs = Matrix_1.Matrix.mult(this.weightsIH, inputsMatrix);
        var hiddenOutputs = Matrix_1.Matrix.addMatrix(hiddenInputs, this.biasIH);
        var hiddenOutputsSigmoid = Matrix_1.Matrix.map(hiddenOutputs, sigmoid);
        // Calculate the output layer output
        var outputInputs = Matrix_1.Matrix.mult(this.weightsHO, hiddenOutputsSigmoid);
        var outputOutputs = Matrix_1.Matrix.addMatrix(outputInputs, this.biasHO);
        var outputOutputsSigmoid = Matrix_1.Matrix.map(outputOutputs, sigmoid);
        return outputOutputsSigmoid.toArray().reduce(function (accumulator, value) { return accumulator.concat(value); }, []);
    };
    NeuralNetwork.prototype.train = function (inputs, targetsArr) {
        // Convert inputs into a Matrix object
        var inputsMatrix = Matrix_1.Matrix.fromArray(inputs, this.inputNum, 1);
        // Calculate the hidden layer output
        var hiddenInputs = Matrix_1.Matrix.mult(this.weightsIH, inputsMatrix);
        var hiddenOutputs = Matrix_1.Matrix.addMatrix(hiddenInputs, this.biasIH);
        var hiddenOutputsSigmoid = Matrix_1.Matrix.map(hiddenOutputs, sigmoid);
        // Calculate the output layer output
        var outputInputs = Matrix_1.Matrix.mult(this.weightsHO, hiddenOutputsSigmoid);
        var outputOutputs = Matrix_1.Matrix.addMatrix(outputInputs, this.biasHO);
        var outputOutputsSigmoid = Matrix_1.Matrix.map(outputOutputs, sigmoid);
        // Convert arrays to Matrix objects
        var outputs = outputOutputsSigmoid;
        var targets = Matrix_1.Matrix.fromArray(targetsArr, this.outputNum, 1);
        // Calculate the output errors
        var outputErrors = Matrix_1.Matrix.subMatrix(targets, outputs);
        // Calculate the gradient of the output errors
        var gradients = Matrix_1.Matrix.map(outputs, dsigmoid);
        gradients.hadamard(outputErrors);
        gradients.scale(this.lr);
        var hiddenTransposed = Matrix_1.Matrix.transpose(hiddenOutputsSigmoid);
        var weightsHODeltas = Matrix_1.Matrix.mult(gradients, hiddenTransposed);
        // Update the weights of the output layer
        this.weightsHO.addMatrix(weightsHODeltas);
        // Update the biases of the outpu layer
        this.biasHO.addMatrix(gradients);
        // Calculate the hidden errors
        var weightsHOTransposed = Matrix_1.Matrix.transpose(this.weightsHO);
        var hiddenErrors = Matrix_1.Matrix.mult(weightsHOTransposed, outputErrors);
        // Calculate the gradient of the hidden errors
        var hiddenGradients = Matrix_1.Matrix.map(hiddenOutputsSigmoid, dsigmoid);
        hiddenGradients.hadamard(hiddenErrors);
        hiddenGradients.scale(this.lr);
        var inputsTransposed = Matrix_1.Matrix.transpose(inputsMatrix);
        var weightsIHDeltas = Matrix_1.Matrix.mult(hiddenGradients, inputsTransposed);
        // Update the weights of the hidden layer
        this.weightsIH.addMatrix(weightsIHDeltas);
        // Update the biases of the hidden layer
        this.biasIH.addMatrix(hiddenGradients);
    };
    return NeuralNetwork;
}());
exports.NeuralNetwork = NeuralNetwork;