f*ck jsperf's crappy sigmoid implementation

2 files changed, 10 insertions, 8 deletions

diff --git a/code/weightvis.js b/code/weightvis.js
index e3cec42..f61ec39 100644
--- a/code/weightvis.js
+++ b/code/weightvis.js
@@ -504,7 +504,7 @@ var NetVisualizer = (function() {
 
     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); }
+        function activation(x) { return 1 / (1 + Math.exp(-x)); }
         var weights = this.data[this.layout.data.batches-1][this.layout.data.epochs-1];
         var lastLayerOut = inputs;
         var hasBias = false;
@@ -525,6 +525,7 @@ var NetVisualizer = (function() {
                 thisLayerOut[tar] = activation(thisLayerOut[tar]);
             }
             lastLayerOut = thisLayerOut;
+            //console.log("layer out", layer, lastLayerOut);
         }
         return lastLayerOut;
     };
diff --git a/index.html b/index.html
index 7836a44..2e17c5c 100644
--- a/index.html
+++ b/index.html
@@ -138,14 +138,15 @@
             <script type="text/javascript">
                 new VisIntegrater(function(c) {
                     weightsWrapper.afterLoad(function() {
-                        // TODO visualization
-                        // TODO shhht… find rows with diff<5 first
                         weightsWrapper.openIn(c);
-                        var row = Math.floor(Math.random()*bostonset.length());
-                        console.log(row);
-                        var inputs = bostonset.inputs(row);
-                        console.log(bostonset.asOutput(weightsWrapper.activate(inputs)[0]));
-                        console.log(bostonset.expectedOutput(row));
+                        var testRows = [400, 175, 454, 481, 476, 364, 377, 75, 103, 84, 458, 483];
+                        for(c=0; c<10; c++){
+                            //var row = testRows[Math.floor(Math.random()*testRows.length)];
+                            var row = Math.floor(Math.random()*bostonset.length());
+                            var inputs = bostonset.inputs(row);
+                            console.log("calculated", bostonset.asOutput(weightsWrapper.activate(inputs)[0]));
+                            console.log("real", bostonset.expectedOutput(row));
+                        }
                     });
                 }, function(c) {
                     weightsWrapper.remove(c);