40oleg/neurochain

Neuro

NeuroChain.js

Simple neural-library which let you create simple fully-connected neural networks. Wrote on TypeScript and you can use it directly in Node.js.

Install

npm install neurochan

Connect

let { Net } = require('./node_modules/neurochan/Net');
let { Layer } = require('./node_modules/neurochan/Layer');

Using

You just have to create instance of Net and add some layers to it. The first argument is count of neurons in a layer you added and the second is an activation function. The first layer doesn't have any activation function because it isn't necessary, but you have to state this explicity. Next you train net and can check quality, but now you can't save weights, therefore you have to learn your net every time you want to use it.

let net: Net = new Net();
net.AddLayer(new Layer(10, "none"));
net.AddLayer(new Layer(12, "LeakyReLU"));
net.AddLayer(new Layer(2, "sigmoid"));

Activation Functions

There are several activation functions at this moment.

• sigmoid
• ReLU
• Leaky ReLU
• hyperbolic tangent
• softplus
• softsign
• argtg

Example of simplest neural network

//dataset
let train = [
  [[0,0],[0]],
  [[0,1],[1]],
  [[1,0],[1]],
  [[1,1],[0]]
];

//making instance of dataset
let net: Net = new Net();

//adding layers
net.AddLayer(new Layer(2, "none", 0));
net.AddLayer(new Layer(3, "tanh", 0));
net.AddLayer(new Layer(1, "tanh", 0));

//setting train and test sample
net.SetTrainSample(train);

//training net
net.Train(2500000, 1, 0.1);

//check result
console.log(net.Run([0,0])); //0.008190680297392377
console.log(net.Run([0,1])); //0.9673620088435199
console.log(net.Run([1,0])); //0.9673148671177659
console.log(net.Run([1,1])); //0.008177935725508057

Performance

You should use NeuroChan only educational purposes because of speed of learning and lack of usefull technologies.

Methods

net.Normalize(sample, min_value, max_value); // normalization input values
// sample - your dataset in the standart format (3D-array)
// min_value - minimal value in the dataset
// max_value - maximal value in the dataset

net.Augmentaion(sample, coef); //sample augmentaion (DOESNT'T SUPPORT)
// sample - your dataset in the standart format (3D-array)
// coef - augmentaion coefficient

net.Train(count_of_repetitions, starting_learing_rate, ending_learning_rate);
// count_of_repetitions - value which have to be more then train-sample 3-5 times

net.Test(repetitions); // evaluation method which prints quality of learning
// before using this method you have to set test sample with method .SetTestSample(test_sample)
// repetitions - count of examples used to define quality

net.SetTrainSample(sample); // set train dataset

net.SetTestSample(sample); // set test dataset

net.AddLayer(Layer); // adding layer
// Layer is an object you import from library