defihook / Kipling

A CNN based dynamic model for classifying different items or types of images.

Kipling

Important Information

This program is built using Python v3.11.2 and has only been tested with that, the program has been tested on both MacOS Ventura 13.2.1 and Windows 11.

Install Dependencies

• If you have pipenv installed
  • Run pipenv install
• If not
  • Run pip3 install -r requirements.txt

Different Kipling Modes

Kipling has built in tools to help make your image classification sequence easy to build:

• Option 1 - The Config Creator
  • This will prompt you through some questions to create a valid config.json file for the network.
• Option 2 - The Training Module
  • This will detect all labels based on your data and then train the model based on your config.
• Option 3 - The Prediction Module
  • This will allow you to have images classified with the model.

Config Creator

The config creator was built to help you create a valid config.json file for the network, the following text will explain what each question is.

• Width - You must specify a width for all images to be resized to. Type: Number
• Height - You must specify a height for all images to be resized to. Type: Number
• Training Data Directory - You must specify a valid path that the model can reference for training data. Type: String
• Saving Model Directory - You must specify a valid path that the network can saved the trained model to. Type: String

Training Module

The training module is completely automatic, you must let it run until the program ends, please ensure the following things before you run the training module.

• You have run the Config Creator and you have a valid config.json file.
• You have loaded in ALL of your training data in the following directory format.

<training_dir>
      |__ label_1
            |__ image1.png
            |__ image2.png
            |__ ...
            |__ image100.png
      |__ label_2
            |__ image1.png
            |__ image2.png
            |__ ...
            |__ image100.png
      |__ label_3
            |__ image1.png
            |__ image2.png
            |__ ...
            |__ image100.png

You MUST have a folder for each label you want, and in that folder you must have .png, .jpg or .jpeg that have that label in them. For Example I want to build a model that can identify a cat, a dog and a bird, view an example of how I would layout my training directory below.

<training_dir>
      |__ dog
            |__ dogimage1.png
            |__ dogimage2.png
            |__ ...
            |__ dogimage100.png
      |__ cat
            |__ catimage1.png
            |__ catimage2.png
            |__ ...
            |__ catimage100.png
      |__ bird
            |__ birdimage1.png
            |__ birdimage2.png
            |__ ...
            |__ birdimage3.png

Note

The more images you have of your labels, the more accurate the model will be in it's predictions.

Prediciton Model

The prediciton model will load in your saved training model, this does mean however that you must run the training model before you can run this model. Once everything is loaded the prediction model will ask you to provide the path to an image you want classified, paste in your path and the model will predict what it is.

Note

Not every model is perfect, you may get the wrong result, if you are continuously getting the wrong results, you may need to add more training data and then retrain your model.