integrate boston housing dataset and test values

3 files changed, 133 insertions, 2 deletions

diff --git a/code/bostonset.js b/code/bostonset.js
new file mode 100644
index 0000000..50b9d38
--- /dev/null
+++ b/code/bostonset.js
@@ -0,0 +1,100 @@
+var Bostonset = (function() {
+    var attrs =
+        ["crim", "zn", "indus", "chas", "nox", "rm", "age", "dis", "rad", "tax", "ptratio", "black", "lstat", "medv"];
+
+    function Bostonset() {
+        this.data = [];
+        // fixed constants exported from training data
+        this.dataScaler = new MinMaxScaler();
+        this.dataScaler.setScales([[0.00632, 0.0, 0.46, 0.0, 0.385, 3.561, 2.9, 1.1296, 1.0, 187.0, 12.6, 0.32, 1.73], [88.9762, 100.0, 27.74, 1.0, 0.871, 8.78, 100.0, 12.1265, 24.0, 711.0, 22.0, 396.9, 37.97]]);
+        this.targetScaler = new MinMaxScaler();
+        this.targetScaler.setScales([5.0, 50.0]);
+    }
+
+    Bostonset.prototype.load = function(url) {
+        var self = this;
+        d3.csv(url).get(function(e, data) {
+            if (e) console.log(e);
+            data.forEach(function (el) {
+                var arr = [];
+                for (c=0; c<attrs.length; c++) {
+                    arr.push(parseFloat(el[attrs[c]]));
+                }
+                self.data.push(arr);
+            });
+        });
+    };
+
+    Bostonset.prototype.length = function() {
+        return this.data.length;
+    };
+
+    Bostonset.prototype.rawInputs = function(row) {
+        return this.data[row].slice(0, -1); // leave out MEDV
+    };
+
+    Bostonset.prototype.inputs = function(row) {
+        return this.dataScaler.scale(this.rawInputs(row));
+    };
+
+    Bostonset.prototype.asOutput = function(num) {
+        return this.targetScaler.unscale(num);
+    };
+
+    Bostonset.prototype.expectedOutput = function(row) {
+        return this.data[row][attrs.length-1];
+    };
+
+    return Bostonset;
+})();
+
+// TODO abstract for generic shapes?
+var MinMaxScaler = (function() {
+    function MinMaxScaler() {
+    }
+
+    MinMaxScaler.prototype.fit = function(arr) {
+        this.setScales(d3.extent(arr));
+    };
+
+    MinMaxScaler.prototype.setScales = function(fits) {
+        this.min = fits[0];
+        this.max = fits[1];
+    };
+
+    MinMaxScaler.prototype.scale = function(data) {
+        data = this._arrayDivide(this._arraySubtract(data, this.min), this._arraySubtract(this.max, this.min));
+        return data;
+    };
+
+    MinMaxScaler.prototype.unscale = function(data) {
+        data = this._arrayAdd(this._arrayMultiply(data, this._arraySubtract(this.max, this.min)), this.min);
+        return data;
+    };
+
+    MinMaxScaler.prototype._arraySubtract = function(one, two) {
+        if (one.constructor !== Array) return one - two;
+        for (i=0; i<one.length; one[i]-=two[i++]);
+        return one;
+    };
+
+    MinMaxScaler.prototype._arrayAdd = function(one, two) {
+        if (one.constructor !== Array) return one + two;
+        for (i=0; i<one.length; one[i]+=two[i++]);
+        return one;
+    };
+
+    MinMaxScaler.prototype._arrayMultiply = function(one, two) {
+        if (one.constructor !== Array) return one * two;
+        for (i=0; i<one.length; one[i]*=two[i++]);
+        return one;
+    };
+
+    MinMaxScaler.prototype._arrayDivide = function(one, two) {
+        if (one.constructor !== Array) return one / two;
+        for (i=0; i<one.length; one[i]/=two[i++]);
+        return one;
+    };
+
+    return MinMaxScaler;
+})();
diff --git a/code/weightvis.js b/code/weightvis.js
index 64d398c..e3cec42 100644
--- a/code/weightvis.js
+++ b/code/weightvis.js
@@ -33,9 +33,9 @@ var
     nodeHighlightColor    = accentColor
 ;
 
-
 /* network construction */
 
+// TODO make multiple bias layers at positions > 0 possible?
 var Network = (function() {
     function Network(layout) {
         this.layout = layout;
@@ -499,9 +499,35 @@ var NetVisualizer = (function() {
                     me.attr("data-highlight", "true");
                 }
             });
-
         this.reposition();
     };
 
+    NetVisualizer.prototype.activate = function(inputs) {
+        // assumes the sigmoid function is used as activation
+        function activation(x) { return 0.5 + x / Math.sqrt(1 + 4 * x * x); }
+        var weights = this.data[this.layout.data.batches-1][this.layout.data.epochs-1];
+        var lastLayerOut = inputs;
+        var hasBias = false;
+        if (this.layout.net[0].isBias) {
+            for (bias=0; bias<inputs.length; bias++) {
+                lastLayerOut[bias] += weights[0][0][bias];
+            }
+            hasBias = true;
+        }
+        for (layer=hasBias?1:0; layer<this.layout.net.layers-1; layer++) {
+            // manual vector matrix multiplication
+            var thisLayerOut = [];
+            for (tar=0; tar<this.layout.net[layer+1].size; tar++) {
+                thisLayerOut[tar] = 0;
+                for (src=0; src<this.layout.net[layer].size; src++) {
+                    thisLayerOut[tar] += weights[layer][src][tar] * lastLayerOut[src];
+                }
+                thisLayerOut[tar] = activation(thisLayerOut[tar]);
+            }
+            lastLayerOut = thisLayerOut;
+        }
+        return lastLayerOut;
+    };
+
     return NetVisualizer;
 })();
diff --git a/code/weightviswrapper.js b/code/weightviswrapper.js
index c1586d1..1b3a91a 100644
--- a/code/weightviswrapper.js
+++ b/code/weightviswrapper.js
@@ -169,6 +169,11 @@ var WeightVisWrapper = (function() {
         this.button.select(".pause").style("display", "");
     };
 
+    WeightVisWrapper.prototype.activate = function(inputs) {
+        // TODO visualize this :)
+        return this.netVisualizer.activate(inputs);
+    };
+
     return WeightVisWrapper;
 })();