jellehak/yellowmind

YellowMind is a machine learning library for running Convolutional Neural Network (CNN) in the browser or NodeJs.

YellowMind

YellowMind is a machine learning library for running Convolutional Neural Network (CNN) in the browser or NodeJs.

Goal

Make training and running of Convolutional Neural Network (CNN) as easy as possible.

Big thanks

• Most of the underlaying library is based on the great work of DenseInL2 https://github.com/DenseInL2/webcnn. This library has been extended to allow creating machine learning networks purely from JSON.

Roadmap

• Convert WebCNN ( from DenseInL2 ) to use JSON.
• Documentation + example(s)
• Browers support
• NodeJS support
• CLI
• WebGl Support
• Handle streaming data
• Combine and link networks

Installation

TODO

How it works

Input

The input data consist of training data and a network topology.

Training data

Define your training data like:

[
  {  
    "type":"imagemap",
    "labels": {
      "url":"/datasets/mnist/images/labels_0.json"
    },
    "width": 2800,
    "height": 2800,
    "single": {
      "width":28,
      "height":28
    },
    "path":"/datasets/mnist/images/mnist_training_0.png"
  },
  {  
    "type":"imagemap",
    "labels": {
      "url":"/datasets/mnist/images/labels_1.json"
    },
    "width": 2800,
    "height": 2800,
    "single": {
      "width":28,
      "height":28
    },
    "path":"/datasets/mnist/images/mnist_training_1.png"
  }
]

Types

Current supported types:

• imagemap
• image

Image

{  
  "type":"image",
  "width": 2800,
  "height": 2800,
  "path":"/datasets/mnist/images/mnist_training_1.png"
}

Imagemap

{  
  "type":"imagemap",
  "width": 2800,
  "height": 2800,
  "single": {
    "width":28,
    "height":28
  },
  "path":"/datasets/mnist/images/mnist_training_1.png"
}

Network

Define your network:

[
  { name: 'image', type: 'inputImageLayer', width: 24, height: 24, depth: 1 },
  { name: 'conv1', type: 'convLayer', units: 10, kernelWidth: 5, kernelHeight: 5, strideX: 1, strideY: 1, padding: false },
  { name: 'pool1', type: 'maxPoolLayer', poolWidth: 2, poolHeight: 2, strideX: 2, strideY: 2 },
  { name: 'conv2', type: 'convLayer', units: 20, kernelWidth: 5, kernelHeight: 5, strideX: 1, strideY: 1, padding: false },
  { name: 'pool2', type: 'maxPoolLayer', poolWidth: 2, poolHeight: 2, strideX: 2, strideY: 2 },
  { name: 'out', type: 'FCLayer', units: 10, activation: 'softmax' }
]

The complete machine learning definition:

{
  "version": "0.0.1",
  "author": "Your team",
  "training": [
    // ... your training data
  ],
  "network": [
    // ... your network
  ],
}

Training

Output

Examples

Simple example of learning MNIST. open in new window