rlab/plugin/neural/mlp.js

Javascript scientific library like R

/**
 * Created by joonkukang on 2014. 1. 14..
 */
var sigmoid = R.NN.sigmoid, dSigmoid = R.NN.dSigmoid;

MLP = module.exports = function (settings) {
	this.x = settings['input'];
	this.y = settings['label'];
	this.sigmoidLayers = [];
	this.nLayers = settings['hiddenLayerSizes'].length;
	for(var i=0 ; i<this.nLayers+1 ; i++) {
		var inputSize = (i===0)?settings['nIns']:settings['hiddenLayerSizes'][i-1];
		var layerInput = (i===0)?this.x:this.sigmoidLayers[this.sigmoidLayers.length-1].sampleHgivenV();
		var nOut = (i===this.nLayers)?settings['nOuts']:settings['hiddenLayerSizes'][i];
		var sigmoidLayer = new HiddenLayer({
			input : layerInput,
			nIn : inputSize,
			nOut : nOut,
			activation : sigmoid,
			W : (typeof settings['w_array'] === 'undefined')? undefined : settings['w_array'][i],
			b : (typeof settings['b_array'] === 'undefined')? undefined : settings['b_array'][i]
		});
		this.sigmoidLayers.push(sigmoidLayer);
	}
};

MLP.prototype.train = function(settings) {
    var lr = settings['lr'] || 0.6;
		var epochs = settings['epochs'] || 1000;
    var currentProgress = 1;
    for(var epoch=0 ; epoch < epochs ; epoch++) {
        // Feed Forward
        var i, layerInput = [];
        layerInput.push(this.x);
        for(i=0; i<this.nLayers+1 ; i++) {
            layerInput.push(this.sigmoidLayers[i].output(layerInput[i]));
        }
        var output = layerInput[this.nLayers+1];
        // Back Propagation
        var delta = new Array(this.nLayers + 1);
				var linearOutput=this.sigmoidLayers[this.nLayers].linearOutput(layerInput[this.nLayers]).map1(dSigmoid);
				delta[this.nLayers] = this.y.msub(output).mmul(linearOutput);
        for(i = this.nLayers - 1; i>=0 ; i--) {
					  var o = this.sigmoidLayers[i].linearOutput(layerInput[i]).map1(dSigmoid);
					  delta[i] = this.sigmoidLayers[i+1].backPropagate(delta[i+1]).mmul(o);
        }
        // Update Weight, Bias
        for(var i=0; i<this.nLayers+1 ; i++) {
						var xlen = this.x.length;
						var deltaW = layerInput[i].tr().mdot(delta[i]).map1((x)=>x/xlen);
						var deltaB = delta[i].colMean();
						this.sigmoidLayers[i].W = this.sigmoidLayers[i].W.madd(deltaW);
						this.sigmoidLayers[i].b = this.sigmoidLayers[i].b.vadd(deltaB);
        }
				var progress = (1.*epoch/epochs)*100;
				if(progress > currentProgress) {
					console.log("MLP",progress.toFixed(0),"% Completed.");
					currentProgress+=8;
				}
    }
    console.log("MLP Final Cross Entropy : ",this.getReconstructionCrossEntropy());
};

MLP.prototype.getReconstructionCrossEntropy = function() {
	var reconstructedOutput = this.predict(this.x);
	var a = R.map2(this.y, reconstructedOutput, function(x,y) {
		return x*Math.log(y);
	})
	var b = R.map2(this.y, reconstructedOutput, function(x,y) {
			return (1-x)*Math.log(1-y);			
	});
	var crossEntropy = -a.madd(b).colSum().mean(); // -(a+b).colMean()
	return crossEntropy
}

MLP.prototype.predict = function(x) {
	var output = x;
	for(i=0; i<this.nLayers+1 ; i++) {
		output = this.sigmoidLayers[i].output(output);
	}
	return output;
};