final touches

README.md

@@ -41,7 +41,7 @@ console.log('prediction:', prediction.indexOf(Math.max.apply(null, prediction)))
 
 ## Performance
 
-All runs below were performed with an i5 MacBook Air 13" Mid 2012 with 4GB RAM.
+All runs below were performed with a MacBook Pro Retina 13" Early 2015 with 8GB RAM.
 
 ### [MNIST [0-9]](https://yann.lecun.com/exdb/mnist/)
 
@@ -61,12 +61,12 @@ To test with the MNIST dataset: click on the title above, download the 4 data fi
 * **Neurons**
  * ```400``` input
  * ```160``` hidden
- * ```4``` output
+ * ```10``` output
 * **Learning rate:** ```0.03```
 * **Training set:** ```60000``` digits
 * **Testing set:** ```10000``` digits
-* **Training time:** ```36 min 22 s 50 ms```
-* **Success rate:** ```90.14%```
+* **Training time:** ```21 min 53 s 753 ms```
+* **Success rate:** ```95.16%```
 
 ### [A-Za-z0-9]
 
@@ -84,23 +84,24 @@ To test with the MNIST dataset: click on the title above, download the 4 data fi
  "image_size": 16,
  "threshold": 400,
  "network": {
- "hidden": 40,
- "learning_rate": 0.03
+ "hidden": 60,
+ "learning_rate": 0.1,
+ "output": 62
  }
 }
 ```
 
 * **Neurons**
  * ```256``` input
- * ```40``` hidden
- * ```8``` output
+ * ```60``` hidden
+ * ```62``` output
 * **Learning rate:** ```0.03```
 * **Training set**
  * **Size:** ```124000``` characters
  * **Sample:** ![abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789](https://raw.github.com/mateogianolio/mlp-character-recognition/master/examples/abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789.png)
 * **Testing set:** ```62000``` characters
-* **Training time:** ```4 min 8 s 2 ms```
-* **Success rate:** ```78.60322580645162%```
+* **Training time:** ```8 min 18 s 560 ms```
+* **Success rate:** ```93.58225806451614%```
 
 ### [a-z]
 
@@ -119,22 +120,23 @@ To test with the MNIST dataset: click on the title above, download the 4 data fi
  "threshold": 400,
  "network": {
  "hidden": 40,
- "learning_rate": 0.1
+ "learning_rate": 0.1,
+ "output": 26
  }
 }
 ```
 
 * **Neurons**
  * ```256``` input
  * ```40``` hidden
- * ```8``` output
+ * ```26``` output
 * **Learning rate:** ```0.1```
 * **Training set**
  * **Size:** ```52000``` characters
  * **Sample:** ![abcdefghijklmnopqrstuvwxyz](https://raw.github.com/mateogianolio/mlp-character-recognition/master/examples/abcdefghijklmnopqrstuvwxyz.png)
 * **Testing set:** ```26000``` characters
-* **Training time:** ```2 min 10 s 752 ms```
-* **Success rate:** ```91.77692307692308%```
+* **Training time:** ```1 min 55 s 414 ms```
+* **Success rate:** ```93.83846153846153%```
 
 ### [0-9]
 
@@ -161,14 +163,14 @@ To test with the MNIST dataset: click on the title above, download the 4 data fi
 * **Neurons**
  * ```256``` input
  * ```40``` hidden
- * ```8``` output
+ * ```10``` output
 * **Learning rate:** ```0.1```
 * **Training set**
  * **Size:** ```20000``` digits
  * **Sample:** ![0123456789](https://raw.github.com/mateogianolio/mlp-character-recognition/master/examples/0123456789.png)
 * **Testing set:** ```10000``` digits
-* **Training time:** ```1 min 6 s 620 ms```
-* **Success rate:** ```99.22%```
+* **Training time:** ```0 min 44 s 363 ms```
+* **Success rate:** ```99.59%```
 
 ## Configuration
 
@@ -178,20 +180,21 @@ Tweak the network for your needs by editing the ```config.json``` file located i
 // config.json
 {
  "mnist": false,
- "text": "abcdefghijklmnopqrstuvwxyz",
+ "text": "0123456789",
  "fonts": [
  "sans-serif",
  "serif"
  ],
  "training_set": 2000,
- "testing_set": 500,
- "image_size": 20,
+ "testing_set": 1000,
+ "image_size": 16,
  "threshold": 400,
  "network": {
  "hidden": 40,
  "learning_rate": 0.1
  }
 }
+
 ```
 
 * **```mnist```**
@@ -216,15 +219,16 @@ Tweak the network for your needs by editing the ```config.json``` file located i
 
 ## Usage
 
-Clone this repository. The script is using [canvas](https://www.npmjs.com/package/canvas), so you'll need to install the **Cairo** rendering engine. On OSX, this can be done with the following one-liner (copied from canvas README):
+Clone this repository. The script is using [canvas](https://www.npmjs.com/package/canvas), so you'll need to install the **Cairo** rendering engine. On OS X, assuming you have Homebrew installed, this can be done with the following (copied from canvas README):
 
 ```bash
-$ wget https://raw.githubusercontent.com/LearnBoost/node-canvas/master/install -O - | sh
+$ brew install cairo jpeg giflib
 ```
 
 Then install npm dependencies and test it:
 
 ```bash
 $ npm install
 $ node main.js
+$ node test.js
 ```

config.json

@@ -1,12 +1,16 @@
 {
+ "mnist": false,
  "text": "0123456789",
+ "fonts": [
+ "sans-serif",
+ "serif"
+ ],
  "training_set": 2000,
- "testing_set": 500,
+ "testing_set": 1000,
  "image_size": 16,
  "threshold": 400,
  "network": {
  "hidden": 40,
- "output": 10,
  "learning_rate": 0.1
  }
 }

main.js

@@ -22,7 +22,7 @@
  config.fonts = config.fonts || ['serif', 'sans-serif'];
  config.distortion = config.distortion === undefined ? true : config.distortion;
  config.network.hidden = config.network.hidden || 40;
- config.network.output = config.network.output || 8;
+ config.network.output = config.network.output || 10;
  config.network.learning_rate = config.network.learning_rate || 0.1;
 
  console.log('... done');
@@ -47,75 +47,72 @@
  };
 
  // captcha callback
- function generate(index) {
- return function(text, data) {
- var png = new PNG({ filterType: 4 });
-
- png.parse(data, function (text, index) {
- return function(error, data) {
- if(error)
- throw error;
-
- var position,
- chunk = [],
- pixel = [],
- i, j, k, x, y;
-
- for(i = 0; i < config.text.length; i++) {
- for(y = 0; y < data.height; y++) {
- for(x = i * config.image_size; x < (i * config.image_size + config.image_size); x++) {
- position = (data.width * y + x) << 2;
-
- for(j = 0; j < 3; j++)
- pixel.push(data.data[position + j]);
-
- chunk.push(
- pixel.reduce(function(previous, current) {
- return previous + current;
- }) > config.threshold ? 0 : 1
- );
- pixel = [];
- }
- }
-
- chunk = tools.center(chunk);
-
- var output = Array.apply(null, new Array(config.network.output)).map(Number.prototype.valueOf, 0);
- output[i] = 1;
-
- if(index < config.training_set) {
- training.push({
- input: chunk,
- output: output
- });
- } else {
- testing.push({
- input: chunk,
- output: output
- });
- }
-
- chunk = [];
+ var k = 0;
+ function generate(text, data) {
+ if (k === 0)
+ fs.writeFileSync('./examples/' + text + '.png', data, 'base64');
+
+ var png = new PNG({ filterType: 4 });
+
+ png.parse(data, function(error, data) {
+ if(error)
+ throw error;
+
+ var position,
+ chunk = [],
+ pixel = [],
+ i, j, x, y;
+
+ for(i = 0; i < config.text.length; i++) {
+ for(y = 0; y < data.height; y++) {
+ for(x = i * config.image_size; x < (i * config.image_size + config.image_size); x++) {
+ position = (data.width * y + x) << 2;
+
+ for(j = 0; j < 3; j++)
+ pixel.push(data.data[position + j]);
+
+ chunk.push(
+ pixel.reduce(function(previous, current) {
+ return previous + current;
+ }) > config.threshold ? 0 : 1
+ );
+ pixel = [];
  }
-
- if(index === samples - 1) {
- console.log('... done');
- console.log();
-
- network.train(perceptron, training, config.network.learning_rate);
- fs.writeFileSync('./ocr.js', 'module.exports = ' + perceptron.standalone().toString());
- network.test(perceptron, testing);
- }
- };
- });
-
- if(index === 0)
- fs.writeFileSync('./examples/' + text + '.png', data, 'base64');
- };
+ }
+
+ chunk = tools.center(chunk);
+
+ var output = Array.apply(null, new Array(config.network.output)).map(Number.prototype.valueOf, 0);
+ output[i] = 1;
+
+ if(k < config.training_set) {
+ training.push({
+ input: chunk,
+ output: output
+ });
+ } else {
+ testing.push({
+ input: chunk,
+ output: output
+ });
+ }
+
+ chunk = [];
+ }
+
+ if(k++ === samples - 1) {
+ console.log('... done');
+ console.log();
+
+ network.train(perceptron, training, config.network.learning_rate);
+ fs.writeFileSync('./ocr.js', 'module.exports = ' + perceptron.standalone().toString());
+ network.test(perceptron, testing);
+ }
+ });
  }
 
  console.log('generating images ...');
 
  for(index = 0; index < samples; index++)
- captcha.generate(settings, generate(index));
+ captcha.generate(settings, generate);
 })();