Merge pull request #295 from BrainJS/294-hidden-size-fix

fix: Resolve issue with different size hidden layers for recurrent ne…
BrainJS · Nov 4, 2018 · 0f98db9 · 0f98db9
2 parents 71864ac + 772eb8c
commit 0f98db9
Show file tree

Hide file tree

Showing 7 changed files with 50 additions and 27 deletions.
diff --git a/browser.js b/browser.js
@@ -6,7 +6,7 @@
  *   license: MIT (http://opensource.org/licenses/MIT)
  *   author: Heather Arthur <fayearthur@gmail.com>
  *   homepage: https://github.com/brainjs/brain.js#readme
- *   version: 1.4.3
+ *   version: 1.4.4
  *
  * acorn:
  *   license: MIT (http://opensource.org/licenses/MIT)
@@ -4236,13 +4236,11 @@ var RNN = function () {
   }, {
     key: 'mapModel',
     value: function mapModel() {
-      var _this = this;
-
       var model = this.model;
       var hiddenLayers = model.hiddenLayers;
       var allMatrices = model.allMatrices;
       this.initialLayerInputs = this.hiddenLayers.map(function (size) {
-        return new _matrix2.default(_this.hiddenLayers[0], 1);
+        return new _matrix2.default(size, 1);
       });
 
       this.createInputMatrix();
@@ -4572,8 +4570,6 @@ var RNN = function () {
   }, {
     key: 'fromJSON',
     value: function fromJSON(json) {
-      var _this2 = this;
-
       var defaults = this.constructor.defaults;
       var options = json.options;
       this.model = null;
@@ -4619,7 +4615,7 @@ var RNN = function () {
         equationConnections: []
       };
       this.initialLayerInputs = this.hiddenLayers.map(function (size) {
-        return new _matrix2.default(_this2.hiddenLayers[0], 1);
+        return new _matrix2.default(size, 1);
       });
       this.bindEquation();
     }

diff --git a/browser.min.js b/browser.min.js
diff --git a/dist/recurrent/rnn.js b/dist/recurrent/rnn.js
diff --git a/dist/recurrent/rnn.js.map b/dist/recurrent/rnn.js.map
diff --git a/package.json b/package.json
@@ -1,7 +1,7 @@
 {
   "name": "brain.js",
   "description": "Neural network library",
-  "version": "1.4.3",
+  "version": "1.4.4",
   "author": "Heather Arthur <fayearthur@gmail.com>",
   "repository": {
     "type": "git",

diff --git a/src/recurrent/rnn.js b/src/recurrent/rnn.js
@@ -152,7 +152,7 @@ export default class RNN {
     let model = this.model;
     let hiddenLayers = model.hiddenLayers;
     let allMatrices = model.allMatrices;
-    this.initialLayerInputs = this.hiddenLayers.map((size) => new Matrix(this.hiddenLayers[0], 1));
+    this.initialLayerInputs = this.hiddenLayers.map((size) => new Matrix(size, 1));
 
     this.createInputMatrix();
     if (!model.input) throw new Error('net.model.input not set');
@@ -506,7 +506,7 @@ export default class RNN {
       equations: [],
       equationConnections: [],
     };
-    this.initialLayerInputs = this.hiddenLayers.map((size) => new Matrix(this.hiddenLayers[0], 1));
+    this.initialLayerInputs = this.hiddenLayers.map((size) => new Matrix(size, 1));
     this.bindEquation();
   }
 

diff --git a/test/recurrent/rnn.js b/test/recurrent/rnn.js
@@ -20,6 +20,28 @@ describe('rnn', () => {
       net.initialize();
       assert.notEqual(net.model, null);
     });
+    it('can setup different size hiddenLayers', () => {
+      const inputSize = 2;
+      const hiddenLayers = [5,4,3];
+      const networkOptions = {
+        learningRate: 0.001,
+        decayRate: 0.75,
+        inputSize: inputSize,
+        hiddenLayers,
+        outputSize: inputSize
+      };
+
+      const net = new RNN(networkOptions);
+      net.initialize();
+      net.bindEquation();
+      assert.equal(net.model.hiddenLayers.length, 3);
+      assert.equal(net.model.hiddenLayers[0].weight.columns, inputSize);
+      assert.equal(net.model.hiddenLayers[0].weight.rows, hiddenLayers[0]);
+      assert.equal(net.model.hiddenLayers[1].weight.columns, hiddenLayers[0]);
+      assert.equal(net.model.hiddenLayers[1].weight.rows, hiddenLayers[1]);
+      assert.equal(net.model.hiddenLayers[2].weight.columns, hiddenLayers[1]);
+      assert.equal(net.model.hiddenLayers[2].weight.rows, hiddenLayers[2]);
+    });
   });
   describe('basic operations', () => {
     it('starts with zeros in input.deltas', () => {
@@ -354,9 +376,12 @@ describe('rnn', () => {
 
     describe('.fromJSON', () => {
       it('can import model from json', () => {
-        let dataFormatter = new DataFormatter('abcdef'.split(''));
-        let jsonString = JSON.stringify(new RNN({
-          inputSize: 6, //<- length
+        const inputSize = 6;
+        const hiddenLayers = [10, 20];
+        const dataFormatter = new DataFormatter('abcdef'.split(''));
+        const jsonString = JSON.stringify(new RNN({
+          inputSize, //<- length
+          hiddenLayers,
           inputRange: dataFormatter.characters.length,
           outputSize: dataFormatter.characters.length //<- length
         }).toJSON(), null, 2);
@@ -368,6 +393,12 @@ describe('rnn', () => {
         assert.equal(clone.inputSize, 6);
         assert.equal(clone.inputRange, dataFormatter.characters.length);
         assert.equal(clone.outputSize, dataFormatter.characters.length);
+
+        assert.equal(clone.model.hiddenLayers.length, 2);
+        assert.equal(clone.model.hiddenLayers[0].weight.columns, inputSize);
+        assert.equal(clone.model.hiddenLayers[0].weight.rows, hiddenLayers[0]);
+        assert.equal(clone.model.hiddenLayers[1].weight.columns, hiddenLayers[0]);
+        assert.equal(clone.model.hiddenLayers[1].weight.rows, hiddenLayers[1]);
       });
 
       it('can import model from json using .fromJSON()', () => {