ToyNN Examples

XOR Truth Table

X	Y	Output
0	0	0
0	1	1
1	0	1
1	1	0

import toynn from 'toynn';

const X = [
    new toynn.NArray([0, 0]).reshape(1, 2),
    new toynn.NArray([0, 1]).reshape(1, 2),
    new toynn.NArray([1, 0]).reshape(1, 2),
    new toynn.NArray([1, 1]).reshape(1, 2),
];

const y = [
    new toynn.NArray([0]),
    new toynn.NArray([1]),
    new toynn.NArray([1]),
    new toynn.NArray([0]),
];

const model = new toynn.NN('xor');

const layer1 = new toynn.Layer(2, 1);
layer1.use(toynn.functions.sigmoid);

model.add(layer1);

model.train({
    x: X,
    y,
    epochs: 1000,
    alpha: 0.005,
    loss: toynn.errors.MSE,
    verbose: true,
    optimizer: new toynn.optimizers.RMSProp(),
});

let newData = new toynn.NArray([1, 0]).reshape(1, 2);
// make prediction
console.log(model.forward(newData).flatten());

Table of 4

import toynn from 'toynn';

const X = [
    new toynn.NArray([1, 2, 3, 4]).reshape(1, -1),
    new toynn.NArray([5, 6, 7, 8]).reshape(1, -1),
];

const y = [
    new toynn.NArray([4, 8, 12, 16]),
    new toynn.NArray([20, 24, 28, 32]),
];

const model = new toynn.NN('tableOf4');

const layer1 = new toynn.Layer(4, 4);
layer1.use(toynn.functions.linear);

model.add(layer1);

model.train({
    x: X,
    y: y,
    epochs: 500,
    alpha: 0.005,
    loss: toynn.errors.MSE,
    verbose: true,
});

const newData = new toynn.NArray([9, 10, 11, 12]).reshape(1, -1);
const prediction = model
    .forward(newData)
    .map((e) => Math.floor(e))
    .flatten();
console.log(`Prediction: ${prediction}`);

IRIS Classifier

import toynn from 'toynn';

const classes = {
    'Iris-setosa': 0,
    'Iris-versicolor': 1,
    'Iris-virginica': 2,
};

const datasetURL =
    'https://ocw.mit.edu/courses/15-097-prediction-machine-learning-and-statistics-spring-2012/89d88c5528513adc4002a1618ce2efb0_iris.csv';

const irisDataset = await toynn.Dataset.from(datasetURL);

const X = irisDataset.slice(0, -1); // every column except for the last one
const y = irisDataset.slice(-1); // last column

X.onGet = (element) => {
    return element.reshape(1, -1);
};

y.onGet = (element) => {
    element = element.flatten()[0];
    element = toynn.utils.onehotEncode(classes[element], 3);
    return new toynn.NArray(element).reshape(-1);
};

const [trainX, testX, trainY, testY] = toynn.utils.trainTestSplit(X, y, {
    testSize: 0.2,
    shuffle: true,
});

const model = new toynn.NN('irisClassifier');

const layer1 = new toynn.Layer(4, 5);
layer1.use(toynn.functions.sigmoid);
const layer2 = new toynn.Layer(5, 3);
layer2.use(toynn.functions.softmax);

model.add(layer1);
model.add(layer2);
model.train({
    x: trainX,
    y: trainY,
    epochs: 500,
    alpha: 0.005,
    verbose: true,
    loss: toynn.errors.MSE,
});

// get accuracy
let accuracy = 0,
    prediction;

for (let i = 0; i < testX.length; i++) {
    prediction = model.forward(testX.get(i)).max().index;

    if (prediction === testY.get(i).max().index) {
        accuracy += 1;
    }
}

console.log(`Accuracy: ${accuracy / testX.length}%`);