pixifier

A picture classifier using CNN (convolutional neural networks)

This repository's goal is to create a way to classify pictures in categories.

It's composed of three docker containers :

1. rebuild-pics.yml : Get pictures used for training in a distant database, download them and resize them. Threre are 3 modes to download pics :

• api : download JSON data for all categories (1 request)
• request : download JSON data for each category (1 request = 1 category)
• elasticsearch : request an Elasticsearch API for each category

2. classify-pics.yml : Train a CNN with theses pictures.

3. report.yml : Start a web server in order to easily see the results on a webpage.

Requirements :

• docker
• docker-compose
• git

How to install :

1. Clone this repository :

"git clone https://github.com/anthill/pixifier.git"

2. Mount the 3 docker containers :

"docker-compose -f rebuild-pics.yml build"

"docker-compose -f classify-pics.yml build"

"docker-compose -f report.yml build"

How to configure :

The data for the training (pictures and labels) should be served by an api descibed by the file param-ini.json in app. This is a sample content:

{
  "urlMode":"api" OR "request" OR "elasticsearch,
  "urlAPI": "http://_MY_WEBSITE/_MY_API_",
  "urlPics": "http://_MY_WEBSITE/_MY_PICS_DIR",
  "maxPics": 100,
  "classes": [
    {
      "active": true,
      "name": "velo",
      "categories": [
        "bicyclettes"
      ],
      "keywords": [
        "velo",
        "vtt",
        "bicyclette"
      ]
    },
    {
      "active": true,
      "name": "lit",
      "categories": [
        "mobilier",
        "literie"
      ],
      "keywords": [
        "lit"
      ]
    }
  ]
}

• urlAPI : URL of the training API sending you every pictures, and their associated categories

  here's what the API should look like :

  [{"category":"CATEGORY","title":"KEYWORD1 KEYWORD2 ","pics":"pic1.jpg;pic2.jpg;"},{"category":"sport","title":"red bicycle","pics":"bicycle(1).jpg;"}]

• urlPics : URL where pics are located

• maxPics : How many pics should it download for each object

• classes : An array which determines what objects you want to download pics for :

  • active : boolean : allows to disable some objects (useful when testing)

  • name : String : name of the object (ex : bicycle)

  • categories : Array : categories where it will search for this item in the API (ex: sport)

  • keywords : Array : keywords related to this item in the API (ex: "bicycle", "mountain bike", "bike")

How to use :

Start the three containers :

"docker-compose -f rebuild-pics.yml up"

"docker-compose -f classify-pics.yml up"

"docker-compose -f report.yml up"

Other informations :

Special thanks to kaparthy, who created the library ConvNetJS, that we use.