encog

https://www.npmjs.com/package/encog

Encog is a NodeJs ES6 framework based on the Encog Machine Learning Framework by Jeff Heaton.

All credits of the framework should go to Jeff Heaton - http://www.heatonresearch.com/encog/

Based on the encog-java-core v3.4 - https://github.com/encog/encog-java-core

Full documentation and source code - https://github.com/redsoul/encog

Installation

npm install encog --save

Usage

Just require the library and all of Encog namespace will be available to you:

const Encog = require('encog');

Unit Tests

npm install --only=dev
npm test

Implemented algorithms

• Networks
  • Basic Network
  • Hopfield Network
  • BAM (Bidirectional associative memory) Network
  • Freeform Network
• Training
  • Back Propagation
  • Manhattan Propagation
  • Resilient Propagation
  • Stochastic Gradient Descent
    • Momentum
    • Nesterov
    • RMS Prop
    • AdaGrad
    • Adam
  • Levenberg Marquardt
  • Neural Simulated Annealing
• Patterns
  • ADALINE
  • Feed Forward (Perceptron)
  • Elman Network
  • Jordan Network
  • Hopfield Network
  • BAM Network
• Activation Functions
  • Elliott
  • Symmetric Elliott
  • Gaussian
  • Linear
  • Ramp
  • ReLu
  • Sigmoid
  • Softmax
  • Steepened Sigmoid
  • Hyperbolic tangent
• Error Functions
  • Arctangent
  • Cross Entropy
  • Linear
  • Output

Examples

Back Propagation example using XOR Data Set

const Encog = require('encog');
const XORdataset = Encog.Utils.Datasets.getXORDataSet();

//adjust the log level
Encog.Log.options.logLevel = 'info';

// create a neural network
const network = new Encog.Networks.Basic();
network.addLayer(new Encog.Layers.Basic(null, true, 2));
network.addLayer(new Encog.Layers.Basic(new Encog.ActivationFunctions.Sigmoid(), true, 4));
network.addLayer(new Encog.Layers.Basic(new Encog.ActivationFunctions.Sigmoid(), false, 1));
network.randomize();

const train = new Encog.Training.Propagation.Back(network, XORdataset.input, XORdataset.output);

Encog.Utils.Network.trainNetwork(train, {maxIterations: 250});
const accuracy = Encog.Utils.Network.validateNetwork(network, XORdataset.input, XORdataset.output);
console.log('Accuracy:', accuracy);

Resilient Propagation example using Iris Flower Data Set (https://en.wikipedia.org/wiki/Iris_flower_data_set)

const Encog = require('encog');
const _ = require('lodash');

//adjust the log level
Encog.Log.options.logLevel = 'info';

const dataEncoder = new Encog.Preprocessing.DataEncoder();
let irisDataset = Encog.Utils.Datasets.getIrisDataSet();
irisDataset = _.shuffle(irisDataset);
irisDataset = Encog.Preprocessing.DataToolbox.trainTestSplit(irisDataset);

/******************/
//data normalization
/******************/

//apply a specific mapping to each column
const mappings = {
    'Sepal.Length': new Encog.Preprocessing.DataMappers.MinMaxScaller(),
    'Sepal.Width': new Encog.Preprocessing.DataMappers.MinMaxScaller(),
    'Petal.Length': new Encog.Preprocessing.DataMappers.MinMaxScaller(),
    'Petal.Width': new Encog.Preprocessing.DataMappers.MinMaxScaller(),
    'Species': new Encog.Preprocessing.DataMappers.OneHot(),
};

//Fit to data, then transform it.
let trainData = dataEncoder.fit_transform(irisDataset.train, mappings);
//transform the test data based on the train data
let testData = dataEncoder.transform(irisDataset.test, mappings);

//slice the data in input and output
trainData = Encog.Preprocessing.DataToolbox.sliceOutput(trainData.values, 3);
testData = Encog.Preprocessing.DataToolbox.sliceOutput(testData.values, 3);

// create a neural network
const network = new Encog.Networks.Basic();
network.addLayer(new Encog.Layers.Basic(null, true, 4));
network.addLayer(new Encog.Layers.Basic(new Encog.ActivationFunctions.Sigmoid(), true, 10));
network.addLayer(new Encog.Layers.Basic(new Encog.ActivationFunctions.Sigmoid(), true, 5));
network.addLayer(new Encog.Layers.Basic(new Encog.ActivationFunctions.Sigmoid(), false, 3));
network.randomize();

// train the neural network
const train = new Encog.Training.Propagation.Resilient(network, trainData.input, trainData.output);
Encog.Utils.Network.trainNetwork(train, {minError: 0.01, minIterations: 5});

//validate the neural network
let accuracy = Encog.Utils.Network.validateNetwork(network, testData.input, testData.output);
console.log('Accuracy:', accuracy);

//save the trained network
Encog.Utils.File.saveNetwork(network, 'iris.dat');

//load a pretrained network
const newNetwork = Encog.Utils.File.loadNetwork('iris.dat');

//validate the neural network
accuracy = Encog.Utils.Network.validateNetwork(newNetwork, testData.input, testData.output);
console.log('accuracy: ', accuracy);

Stochastic Gradient Descent with Adam update example using the bank note authentication dataset

const Encog = require('encog');
const _ = require('lodash');
const dataEncoder = new Encog.Preprocessing.DataEncoder();

//adjust the log level
Encog.Log.options.logLevel = 'info';

(async () => {
    const dataset = await Encog.Preprocessing.DataToolbox.readTrainingCSV(
        './node_modules/encog/examples/data/data_banknote_authentication.csv'
    );
    const shuffledDataset = _.shuffle(dataset);

    const splittedDataset = Encog.Preprocessing.DataToolbox.trainTestSplit(shuffledDataset);

    /******************/
    //data normalization
    /******************/
    //apply a specific mapping to each column
    const mappings = {
        'variance': new Encog.Preprocessing.DataMappers.MinMaxScaller(),
        'skewness': new Encog.Preprocessing.DataMappers.MinMaxScaller(),
        'curtosis': new Encog.Preprocessing.DataMappers.MinMaxScaller(),
        'entropy': new Encog.Preprocessing.DataMappers.MinMaxScaller(),
        'class': new Encog.Preprocessing.DataMappers.IntegerParser()
    };
    //Fit to data, then transform it.
    let trainData = dataEncoder.fit_transform(splittedDataset.train, mappings);
    //transform the test data based on the train data
    let testData = dataEncoder.transform(splittedDataset.test, mappings);

    //slice the data in input and output
    trainData = Encog.Preprocessing.DataToolbox.sliceOutput(trainData.values, 1);
    testData = Encog.Preprocessing.DataToolbox.sliceOutput(testData.values, 1);

    // create a neural network
    const network = new Encog.Networks.Basic();
    network.addLayer(new Encog.Layers.Basic(null, true, 4));
    network.addLayer(new Encog.Layers.Basic(new Encog.ActivationFunctions.Sigmoid(), true, 40));
    network.addLayer(new Encog.Layers.Basic(new Encog.ActivationFunctions.Sigmoid(), true, 40));
    network.addLayer(new Encog.Layers.Basic(new Encog.ActivationFunctions.Sigmoid(), true, 40));
    network.addLayer(new Encog.Layers.Basic(new Encog.ActivationFunctions.Sigmoid(), false, 1));
    network.randomize();

    // train the neural network
    const train = new Encog.Training.SGD.StochasticGradientDescent(network, trainData.input, trainData.output, new Encog.Training.SGD.Update.Adam());
    Encog.Utils.Network.trainNetwork(train, {minError: 0.01, minIterations: 50, maxIterations: 200});

    //validate the neural network
    let accuracy = Encog.Utils.Network.validateNetwork(network, testData.input, testData.output);
    console.log('Accuracy:', accuracy);

    //save the trained network
    Encog.Utils.File.saveNetwork(network, 'banknote_authentication.dat');

    //load a pretrained network
    const newNetwork = Encog.Utils.File.loadNetwork('banknote_authentication.dat');

    //validate the neural network
    accuracy = Encog.Utils.Network.validateNetwork(newNetwork, testData.input, testData.output);
    console.log('accuracy: ', accuracy);
})();

Hopfield Network example custom binary dataset

const Encog = require('encog');
const _ = require('lodash');
const hopfieldPatterns = Encog.Utils.Datasets.getHopfieldPatterns();
const HopfieldPattern = new Encog.Patterns.Hopfield();

//adjust the log level
Encog.Log.options.logLevel = 'info';

HopfieldPattern.setInputLayer(35);
const network = HopfieldPattern.generate();

_.each(hopfieldPatterns, function (pattern) {
    network.addPattern(pattern);
});

network.runUntilStable(10);
const input = [
    0, 0, 0, 0, 0,
    0, 1, 1, 1, 0,
    0, 0, 0, 0, 0,
    0, 1, 1, 0, 0,
    0, 0, 0, 0, 0,
    0, 1, 1, 1, 0,
    0, 0, 0, 0, 0
];
const result = network.compute(input);
console.log('Result:', result);

/*
Output:

0, 0, 0, 0, 0,
0, 1, 1, 1, 0,
0, 1, 0, 0, 0,
0, 1, 1, 0, 0,
0, 1, 0, 0, 0,
0, 1, 1, 1, 0,
0, 0, 0, 0, 0
*/

Node.js version compatibility

8.0.0 or higher