NeuralN

Powerful Neural Network for Node.js

NeuralN is a C++ Neural Network library for Node.js with multiple advantages compared to existing solutions:

• Works with extra large datasets (>1Go allowed by nodejs)
• Multi-Threaded training available.

Large datasets

With Node.js and the V8, it is not possible to work with large datasets since the maximum allowed memory is around 512MB for 32-bits machines and 1GB for 64-bits machines. When you are working with datasets of several gigabytes, it quickly becomes difficult to train you network with all your data.

NeuralN allows you to use datasets as big as your memory can contain.

Multi-Threaded

Working with large datasets increases the performances of your final network, but the learning phase can sometimes take up to several days or even weeks to obtain good results.

With the multi-threaded training method of NeuralN, you can significantly reduce the duration of the learning phase, by training your network simultaneously on different parts of your dataset. The results of each iteration are then combined.

Install

npm install neuraln

How it works

var NeuralN = require('neuraln');
 
/* Create a neural network with 4 layers (2 hidden layers) */
var network = new NeuralN([ 1, 4, 3, 1 ]);
 
/* Add points to the training set */
for(var i = -1; i < 1; i+=0.1) {
  network.train_set_add([ i ], [ Math.abs(Math.sin(i)) ]);
}
 
/* Train the network with one of the two available methods */
/* monothread (blocking) vs multithread (non-blocking)     */
network.train({
  target_error: 0.01,
  iterations: 20000,
 
  multithread: true,
  /* Relevant only when multithread is true: */
  step_size: 100,
  threads: 4
}, function(err) {
 
});
 
/* Run */
var result = network.run([ (Math.random() * 2) - 1 ]);
 
/* Retrieve the network's string representation */
var string = network.to_string();
 
/* Retrieve the network's state string */
var state = network.get_state();

Instantiation & Methods

var network = new NeuralN(layers, momentum, learning_rate, bias);
var network = new NeuralN(network_string);

Instantiate a new network with the following parameters:

• layers is an array representing the layers of the network
• momentum is a number between 0 and 1. This parameter is optional and defaults to 0.3
• learning_rate is a number. This parameter is optional and defaults to 0.1
• bias is a number. This parameter is optional and defaults to -1

Or

• network_string a string from a previous network (using to_string)

network.train_set_add(input, output);

Add a training data point with input and output being arrays of numbers. input and output must contain as many values as the number of neurons of the first and last layers

network.train([options, ]callback);

Train the network with the training set until the target_error or the max_iterations has been reached. The options are optional parameters:

• target_error: 0.01
• iterations: 20000
• multithread is a boolean, which defaults to false.
• step_size represents the number of points of the training set to use by thread at each iteration. Default to 100 (only with multithread: true)
• threads represents the number of threads to be used for the training. Default to 4 (only with multithread: true)
• callback(err) is called once the training is done.

All these parameters are optional except for the callback

network.run(input)

Runs the given input throught the network and returns its output

network.to_string()

Returns a string representation of the network in order to save and reload it later

network.get_state()

Returns a string representation of each neuron of the network. It allows you to understand which entrance neurons most impacted the final result.

network.to_json();
 
// Example:
{ layers: [ 1, 4, 3, 1 ],
  momentum: 0.3,
  learning_rate: 0.1,
  bias: -1,
  biases:
   [ [],
     [ -0.00000901958, -0.00000414136, 0.00000156238, -0.00000275219 ],
     [ 0.000125352, 0.000145129, 0.000285706 ],
     [ -0.00914877 ] ],
  weights:
   [ [],
     [ [ 0.218714 ], [ 0.285424 ], [ 0.236087 ], [ 0.329174 ] ],
     [ [ 0.0541952, -0.057953, -0.0293854, 0.030311 ],
       [ -0.106412, -0.0125738, 0.0167244, -0.117874 ],
       [ -0.0977025, -0.0275803, 0.0262269, 0.00674729 ] ],
     [ [ -0.0480921, -0.0574143, -0.118449 ] ] ] }

Returns a json representation of the network. This is not recommended when the network structure gets big.

network.get_state_json();
 
// Example:
{ layers: [ 1, 4, 3, 1 ],
  values:
   [ ,
     [ [ 0.978927 ], [ 1.10844 ], [ 0.926947 ], [ 0.974797 ] ],
     [ [ 2.39933, -2.49017, -5.58942, 3.32711 ],
       [ -2.09446, -0.193907, 1.53176, -2.44019 ],
       [ -2.62491, -0.0477072, 2.40908, -3.03165 ] ],
     [ [ 1.38492, -0.294276, -0.362479 ] ] ] }

Returns a json representation of the network's state. This is not recommended when the network structure gets big.

Contact us

Feel free to contact us at [e-mail]

License

Distributed under the MIT License.

Copyright Teleportd Ltd. and other Contributors

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.