rlab/plugin/neural/hiddenLayer.js

Javascript scientific library like R

/**
 * Created by joonkukang on 2014. 1. 12..
 */
var sigmoid = R.NN.sigmoid;

HiddenLayer = module.exports = function (settings) {
	this.input = settings['input'];
	var a = 1. / settings['nIn'];
	settings['W'] = settings['W'] || R.randomM(settings['nIn'],settings['nOut'],-a,a);
	settings['b'] = settings['b'] || R.newV(settings['nOut']);
	settings['activation'] = settings['activation'] || sigmoid;
	this.W = settings['W'];
	this.b = settings['b'];
	this.activation = settings['activation'];
}

HiddenLayer.prototype.output = function(input) {
  this.input = input || this.input;
	return this.linearOutput(this.input).map1(this.activation);
};

HiddenLayer.prototype.linearOutput = function(input) { // before activation
  this.input = input || this.input;
  var linearOutput = R.addMV(R.mdot(this.input, this.W), this.b);
  return linearOutput;
}

HiddenLayer.prototype.backPropagate = function (input) { // example+num * nOut matrix
    if(typeof input === 'undefined') throw new Error("No BackPropagation Input.")
		var linearOutput = input.mdot(this.W.tr());
    return linearOutput;
}

HiddenLayer.prototype.sampleHgivenV = function(input) {
	this.input = input || this.input;
	var hMean = this.output();
	var hSample = R.NN.binarySample(hMean); // Gibb's random sampling
	return hSample;
}