Overview

A framework for continuous and self-training(without training dataset) of custom image classifier so that it can describe the image appearing on digital documents such as ebooks, comics etc.

How to run

• Deploy the framework

  • Setup an Azure Machine Learning workspace
  • Run this notebook. Before running,
    • Configure Azure Machine Learning Development Environment.
    • Install the required dependencies
    • Download the config.json of your AML Workspaceto the root. It is needed for the authorization of your AML workspace.
    • If you want, you can change the step scripts with your training and scrapping logic. Learn about Azure Machine Learning Pipelines
    • The current logic uses Custom Vision API image classifier, which is trained by the image-describer framework, and Bing Image Search, for web scrapping.
    • If you want to use the same logic, replace the required subscription keys and endpoints in the scripts of train and scrapper step.

• Train using the framework

  • You can use the framework to train your classifier using REST endpoint, through Azure Portal or by creating a schedule for self training, as shown in the later sections of this notebook.

• Use\Test your image classifier

  • Run comic_reader.py. Before running,
    • If you are using the current logic, Put your subscription keys and endnpoints in keys.json.
  • A .wmv file will be generated in the root folder
  • Play the voice transcript generated by the comic_reader.py