js/som.js

// Javascript implementation pf Kohonen's Self Organizing Map
// Based on http://www.ai-junkie.com/ann/som/som1.html

var mapWidth = 800,
    mapHeight = 800;

function getDistance(weight, inputVector) {
	var distance = 0;
	for (var i = 0; i <weight.length; i++) {
		distance += (inputVector[i] - weight[i]) * (inputVector[i] - weight[i]);
	}
	return Math.sqrt(distance);
}

function makeRandomWeights(vSize, eSize) {
	var weights = [];
  if(typeof ArrayBuffer === 'undefined') {
    // lacking browser support
	  while (weights.length < vSize) {
      var arr = new Array(eSize);
      for(var i = 0; i < eSize; i++) { arr[i]= Math.random(); }
      weights.push(arr);
    }
  } else {
	  while (weights.length < vSize) {
      var arr = new Float64Array(eSize);
      for(var i = 0; i < eSize; i++) { arr[i]= Math.random(); }
      weights.push(arr);
    }
  }
  return weights;
}

function getBMUIndex(weights, target) {
	var BMUIndex = 0;
	var bestScore = 99999;
	
	for (i=0; i < weights.length; i++) {
		distance = getDistance(weights[i], target);
		if (distance < bestScore) {
			bestScore = distance;
			BMUIndex = i;
		}
	}
	return BMUIndex;
}

function convertIndexToXY(idx, dimW) {
	var x = parseInt(idx % dimW,10);
	var y = parseInt((idx / dimW),10);
	return [x,y];
}


function getEucledianDistance(coord1, coord2) {
    return (coord1[0] - coord2[0]) * (coord1[0] - coord2[0]) + (coord1[1] - coord2[1]) * (coord1[1] - coord2[1]);
}

// utilitity that creates contiguous vector of zeros of size n
  var zeros = function(n) {
    if(typeof(n)==='undefined' || isNaN(n)) { return []; }
    if(typeof ArrayBuffer === 'undefined') {
      // lacking browser support
      var arr = new Array(n);
      for(var i=0;i<n;i++) { arr[i]= 0; }
      return arr;
    } else {
      return new Float64Array(n); // typed arrays are faster
    }
  }

// compute L2 distance between two vectors
var L2 = function(x1, x2) {
  var D = x1.length;
  var d = 0;
  for(var i=0;i<D;i++) { 
    var x1i = x1[i];
    var x2i = x2[i];
    d += (x1i-x2i)*(x1i-x2i);
  }
  return d;
}

// compute pairwise distance in all vectors in X
var xtod = function(X) {
    var N = X.length;
  var dist = zeros(N * N); // allocate contiguous array
  for(var i=0;i<N;i++) {
    for(var j=i+1;j<N;j++) {
      var d = L2(X[i], X[j]);
      dist[i*N+j] = d;
      dist[j*N+i] = d;
      }
  }
  return dist;
}

function dotproduct(a,b) {
    var n = 0, lim = Math.min(a.length,b.length);
    for (var i = 0; i < lim; i++) n += a[i] * b[i];
    return n;
 }

function vecsum(a,b) {
  var lim = a.length;
  var sum = new Array(lim);
  for (var i = 0; i < lim; i++) sum[i] = a[i] + b[i];
  return sum;
 }

function norm2(a) {var sumsqr = 0; for (var i = 0; i < a.length; i++) sumsqr += a[i]*a[i]; return Math.sqrt(sumsqr);}

function cosine_sim(x, y) {
    xnorm = norm2(x);
    if(!xnorm) return 0;
    ynorm = norm2(y);
    if(!ynorm) return 0;
    return dotproduct(x, y) / (xnorm * ynorm);
}

function makeSOM(data, training_iterations) {
	var dimW = 6;
	var dimH = 6;
  
	var radius = (dimW * dimH) / 2;
	var learning_rate = 1;
	var time_constant = training_iterations / Math.log(radius);
	var inputs = xtod(data.vecs);
  var dimI = data.vecs.length;
	var weights = makeRandomWeights(dimW * dimH, dimI);
	var radius_decaying = 0;
	var learning_rate_decaying = 0;
	var svg;
  var no_targets = (data.target.match(/.[ |]+./g) || []).length+1;
//  var avg, avgsim1, avgsim2, minsim;
  var refIndex;
  var colorScale;
  var urlprefix =  new URLSearchParams(window.location.search);
  urlprefix.delete("word");
  urlprefix.append("word","");

  if(no_targets > 1) {
    refIndex=1;
    colorScale = d3.scale.linear()
      .range(['green', 'yellow', 'red']) // or use hex values
      .domain([-1, 0, 1]);
    
    // avg = vecsum(inputs.slice(0, dimI), inputs.slice(dimI, 2*dimI));
    // avgsim1 = cosine_sim(inputs.slice(0, dimI), avg);
    // avgsim2 = cosine_sim(inputs.slice(dimI, 2*dimI), avg);

    $("#somcolor2").css("background-color", colorScale(0));
    $("#somcolor1").css("background-color", colorScale(-1));
    $("#somcolor3").css("background-color", colorScale(1));
  } else {
    refIndex = data.words.length-1;
    colorScale = d3.scale.linear()
      .range(['white', 'red'])
      .domain([-1, 1]);
    $("#somcolor1").css("background-color", colorScale(1));
    $("#somcolor3").css("background-color", colorScale(-1));
  }
  
  $("#somword1").html(data.words[0]);
  $("#somword2").html(data.words[refIndex]);
  minsim = cosine_sim(inputs.slice(0, dimI), inputs.slice(refIndex*dimI, (refIndex+1)*dimI));

  var itdom = document.getElementById("iterations");

  var div = d3.select("#som2");

	data.coords = [];
  for(var i=0; i< data.words.length; i++) {
    data.coords[i] = [Math.floor(dimW/2), Math.floor(dimH/2)];
  }
  
  svg = div.append("svg")
					 .attr("width", mapWidth)
					 .attr("height", mapHeight);

  var rects = svg.selectAll(".r")
      .data(weights)
      .enter().append("rect")
			.attr("class", "r")
      .attr("width", mapWidth/dimW)
      .attr("height", mapHeight/dimH)
      .attr("fill", "white")
      .attr("z-index", "-1")
      .attr("x", function(d, i) { return (i % dimW) * (mapWidth/dimW);})
      .attr("y", function(d, i) { return (Math.floor(i / dimW) * (mapWidth/dimW)); })
            

  var g = svg.selectAll(".b")
			.data(data.words)
			.enter().append("g")
			.attr("class", "u");
	g.append("a")
	  .attr("xlink:href", function(word) {return "?"+urlprefix+word;})
	  .attr("title", function(d, i) {
		  return "rank: "+i +"  "+"freq. rank: "+data.ranks[i].toString().replace(/\B(?=(\d{3})+(?!\d))/g, ",");
	  })
		.append("text")
    .attr("text-anchor", "bottom")
    .attr("font-size", 12)
    .attr("fill", function(d) {
			if(data.target.indexOf(" "+d+" ") >= 0) {
				return "blue";
			} else {
				return "#333"
			}
	  })
    .text(function(d) { return d; });
  
  var som_interval = setInterval(somStep, 0);
  var it=0;
  
  function updateSOM() {
    var oc = [];
    for(var x = 0; x < dimW; x++) {
      for(var y = 0; y < dimH; y++) {
        oc[y*dimW+x]=1;
      }
    }
    svg.selectAll('.u')
      .data(data.coords)
      .transition()
      .attr("transform", function(d, i) { 
			  return "translate(" +
          (d[0]*(mapWidth/dimW)+4) + "," +
          (d[1]*(mapHeight/dimH)+oc[d[1]*dimW+d[0]]++*14+4) + ")"; });

    var colorFun = function(d, i) {
      var sim1=cosine_sim(d, inputs.slice(0, dimI));
      var sim2=cosine_sim(d, inputs.slice(dimI, 2*dimI));
      var col;
//      col = (sim1-avgsim1)/(1-avgsim1)-(sim2-avgsim2)/(1-avgsim2);
      col = (sim2-sim1)/(1-minsim);
//      console.log(Math.floor(i/dimW)+","+i%dimW+":"+(sim1-minsim)/(1-minsim)+ " " + (sim2-minsim)/(1-minsim) + "--> "+ col);
      if(col > 1) col=1;
      if(col < -1) col=-1;
      return colorScale(col);
    };
    
    if(it>training_iterations*.6) {
      svg.selectAll(".r")
        .data(weights)
        .transition()
        .attr("fill", colorFun);
    }
  }
  
  function somStep() {
	  if(it++ >= training_iterations) {
      updateSOM();
      clearInterval(som_interval);
      return;
    }
	  itdom.innerHTML = it;
	  radius_decaying = radius * Math.exp(-it/time_constant);
    learning_rate_decaying = learning_rate * Math.exp(-it/time_constant);
    //learning_rate_decaying = learning_rate * Math.exp(-it/training_iterations);
    
    //pick a random input to train
    var current=Math.floor(Math.random()*dimI)
    var iv = inputs.slice(current*dimI, (current+1)*dimI);
    // Determine the BMU
    BMUIdx = getBMUIndex(weights, iv);
    var coord1 = convertIndexToXY(BMUIdx, dimW);
    data.coords[current] = coord1;
    var widthSq = radius_decaying * radius_decaying;
    for (var v in weights) {
      var coord2 = convertIndexToXY(v, dimW);
      var dist = getEucledianDistance(coord1, coord2);
      // Determine if the weight is within the training radius
      if (dist < widthSq) {
        // console.log(dist, learning_rate_decaying, radius_decaying, it);
        influence = Math.exp(-dist/(2*widthSq));
        for (vidx = 0;vidx<weights[v].length;vidx++) {
          weights[v][vidx] += influence * learning_rate_decaying * (iv[vidx] - weights[v][vidx]);
        }
      }
    }
//  }
    if(it % 10 == 0) {
      updateSOM();
    }
	}

}