3net.js - npm Package Compare versions

Comparing version 0.0.3 to 0.0.4

index.js

@@ -21,14 +21,3 @@ var math = require('mathjs');
 
-neuralNetwork.prototype.train = function (input, label, learning_rate, iters, regularization) {
-    if (!iters) iters = 500;
-    if (!learning_rate) learning_rate = 0.1;
-    if (!regularization) regularization = 1;
-    if (label.length !== this.layer3) return false;
-    if (input.length !== this.layer1) return false;
-
-    [this.theta1, this.theta2] = neuralMath.gradientDescent(this.backpropagation, this.theta1, this.theta2, this.theta1gradient, this.theta2gradient, iters, learning_rate);
-    return true;
-}
-
-neuralNetwork.prototype.backpropagation = function () {
+neuralNetwork.prototype.backpropagation = function (input, label, options) {
     //Forward
@@ -43,10 +32,56 @@ this.predict(input);
     //Get gradients
-    this.theta2gradient = math.add(this.theta2gradient, neuralMath.outerProduct(d3, this.a2));
-    this.theta1gradient = math.add(this.theta1gradient, neuralMath.outerProduct(d2, this.a1));
+    this.theta2gradient = neuralMath.outerProduct(d3, this.a2);
+    this.theta1gradient = neuralMath.outerProduct(d2, this.a1);
 
-    return [this.theta1gradient, this.theta2gradient];
+    if (options.regularization) {
+        for (var i = 0; i < math.size(this.theta1gradient)[0]; i += 1) {
+            for (var j = 1; j < math.size(this.theta1gradient)[1]; j += 1) {
+                this.theta1gradient[i][j] += options.regularization * this.theta1gradient[i][j];
+            }
+        }
+        for (var i = 0; i < math.size(this.theta2gradient)[0]; i += 1) {
+            for (var j = 1; j < math.size(this.theta2gradient)[1]; j += 1) {
+                this.theta2gradient[i][j] += options.regularization * this.theta2gradient[i][j];
+            }
+        }
+    }
+}
 
+neuralNetwork.prototype.gradientDescent = function (input, label, options) {
+    var lastCost = neuralMath.getCost(this.a3, label, [this.theta1, this.theta2], options);
+    for (var i = 0; i < options.iters; i += 1) {
+        this.backpropagation(input, label, options);
+
+        var currentCost = neuralMath.getCost(this.a3, label, [this.theta1, this.theta2], options);
+        if (i > 0 && math.abs(currentCost - lastCost) <= 0.001) return;
+        lastCost = currentCost;
+
+        this.theta1 = math.subtract(this.theta1, math.dotMultiply(options.learning_rate, this.theta1gradient));
+        this.theta2 = math.subtract(this.theta2, math.dotMultiply(options.learning_rate, this.theta2gradient));
+    }
 }
 
-function neuralNetwork(layer1, layer2, layer3) {
+neuralNetwork.prototype.train = function (input, label, options) {
+    if (!options) options = {};
+    if (!options.iters) options.iters = 500;
+    if (!options.learning_rate) options.learning_rate = 0.5;
+    if (!options.regularization) options.regularization = 0.1;
+    if (label.length !== this.layer3) return false;
+    if (input.length !== this.layer1) return false;
+
+    this.gradientDescent(input, label, options);
+    return true;
+}
+
+neuralNetwork.prototype.exportNet = function () {
+    var data = {};
+    data.layer1 = this.layer1;
+    data.layer2 = this.layer2
+    data.layer3 = this.layer3;
+    data.theta1 = this.theta1;
+    data.theta2 = this.theta2;
+    return data;
+}
+
+function neuralNetwork(layer1, layer2, layer3, theta) {
     this.layer1 = layer1;
@@ -56,4 +91,9 @@ this.layer2 = layer2;
 
-    this.theta1 = math.random([layer2, layer1 + 1], -1, 1);
-    this.theta2 = math.random([layer3, layer2 + 1], -1, 1);
+    if (!theta) {
+        this.theta1 = math.random([layer2, layer1 + 1], -5, 5);
+        this.theta2 = math.random([layer3, layer2 + 1], -5, 5);
+    } else {
+        this.theta1 = theta[0];
+        this.theta2 = theta[1];
+    }
 
@@ -74,2 +114,7 @@ this.theta1gradient = math.zeros([layer2, layer1 + 1]);
 
-exports.createNet = createNet;
+function importNet(data) {
+    return new neuralNetwork(data.layer1, data.layer2, data.layer3, [data.theta1, data.theta2]);
+}
+
+exports.createNet = createNet;
+exports.importNet = importNet;

neuralMath.js

@@ -22,3 +22,3 @@ var math = require('mathjs');
 
-function getCost(pred, actual) {
+function getCost(pred, actual, theta, options) {
     if (pred.length !== actual.length) return false;
@@ -30,2 +30,16 @@ var cost = 0;
     }
+    if (options.regularization) {
+        var regCost = 0;
+        for (var i = 0; i < math.size(theta[0])[0]; i += 1) {
+            for (var j = 1; j < math.size(theta[0])[1]; j += 1) {
+                regCost += theta[0][i][j] ^ 2;
+            }
+        }
+        for (var i = 0; i < math.size(theta[1])[0]; i += 1) {
+            for (var j = 1; j < math.size(theta[1])[1]; j += 1) {
+                regCost += theta[1][i][j] ^ 2;
+            }
+        }
+        cost += regCost * (options.regularization / 2);
+    }
     return cost;
@@ -32,0 +46,0 @@ }

package.json

 {
     "name": "3net.js",
-    "version": "0.0.3",
+    "version": "0.0.4",
     "description": "A simple library for implementing 3 layer neural networks",
     "main": "index.js",
     "keywords": [
-    "neural", "network", "machine", "learning", "3netjs", "3net"
+        "neural", "network", "machine", "learning", "3netjs", "3net", "data", "multilayer", "perceptron"
     ],
@@ -24,7 +24,3 @@ "repository": {
         "mathjs": "^2.1.1"
-    },
-    "gitHead": "fd2f0bc5d74bb032e4365b5c6181557820921f8b",
-    "_id": "3net.js@0.0.1",
-    "_shasum": "c594ad9cbdae82a140c5de8f1015e07219f07c35",
-    "_from": "3net.js@*"
+    }
 }

README.md

 # 3net.js
 
+A simple library for implementing 3 layer neural networks
+
 [![NPM](https://nodei.co/npm/3net.js.png)](https://npmjs.org/package/3net.js)
 
-var three_net = require('3net.js');
-var net = three_net.createNet(inputLayer, hiddenLayer, outputLayer);
-net.train(data, label);
-net.predict(data);
-
-A simple library for implementing 3 layer neural networks
+#### Example
+    var three_net = require('3net.js');  //Install with 'npm install 3net.js'
+    
+    //Initialization
+    var inputLayer = 400;
+    var hiddenLayer = 25;
+    var outputLayer = 10;
+    var net = three_net.createNet(inputLayer, hiddenLayer, outputLayer);  
+    
+    //Training
+    options = {"iters": 500, "learning_rate": 0.5, "regularization": 0.1};
+    net.train(data, label, options);  
+    
+    //Predicting
+    net.predict(data);  //Returns a 10 dimensional array of the output neuron activations
+    
+    //Importing and exporting
+    var savedNet = net.exportNet(); //Exports as JSON
+    var copiedNet = three_net.importNet(savedNet); //Imports as JSON

New alerts

License Policy Violation

License

This package is not allowed per your license policy. Review the package's license to ensure compliance.

Found 1 instance in 1 package

Fixed alerts

License Policy Violation

License

This package is not allowed per your license policy. Review the package's license to ensure compliance.

Found 1 instance in 1 package

Improved metrics

Total package byte prevSize

8184

increased by47.59%

Lines of code

136

increased by67.9%

Number of lines in readme file

26

increased by136.36%

No dependency changes