pyOLO: Automated YOLOv3 training tool using Python

Introduction

pyOLO is a tool written in Python that helps users train their data with YOLO easily. All you need is some initial setup and your images for training! This project cotains a darknet model cloned from: https://github.com/AlexeyAB/darknet.git

Currently pyOLO mainly consists of 5 steps:

• Step 1: Preparing Object Images
• Step 2: Generating Dataset
• Step 3: Setting Up YOLO Environment
• Step 4: Run YOLO Training
• Step 5: Run YOLO Testing

Requirements

• Linux (Tested on: Ubuntu 18.04)
• pip, python virtual environment for basic environment setup
  • Run sudo apt install python-pip python-venv on your terminal to install.
  • You might have to install a specific version of venv, depending on your python version. e.g if you are using python3.9, you might have to run sudo apt install python-pip python3.9-venv instead.
• darknet Requirements check https://github.com/AlexeyAB/darknet#requirements-for-windows-linux-and-macos

How To Use

0. Initial Setup

• You only have go get through this step once for each device.
• clone this repository to your workspace, and move to the pyOLO directory.

git clone https://github.com/mushypeas/pyOLO.git
cd pyOLO

• Make & activate a virtual environment. If you don't know how, just run the following lines:

python3 -m venv ~/venv/pyOLO
echo alias pyolo='source ~/venv/pyOLO/bin/activate' >> ~/.bashrc
pyolov

Now, simply running pyolov on your terminal will activate your virtual environment. If you want to exit the virtual environment, run deactivate on your terminal.

• Install the required python libraries by running

pip install -r requirements.txt

1. Preparing Images

• Take Pictures of Background Images (TODO)
• Take Pictures of Object Images (TODO)
• Import Background Images to Project Move the background images to pyOLO/backgrounds.
• Import Object Images to Project Move the images of an object in a directory named as the object's YOLO class name, and move those directories into pyOLO/objects. The name of the images doesn't matter. For example, if you want to train YOLO to recognize a dog and a tuna can, the hierarchy would be as follows:

pyOLO
├── objects
│   ├── dog
│   │   ├── firstdog.png
│   │   ├── second_dog.jpg
│   │   ├── ...
│   │   └── lastdog.png
│   └── tuna can
│       ├── tuna01.png
│       ├── tuunatwo.jpg
│       ├── ...
│       └── finalTunaCan.png
├── ...

... And your all set!

2. Running pyOLO

*Dont forget to activate the virtual environment! Before runnung pyOLO, you'll have to edit pyOLO/settings.json.

# settings.json
{
    "extensions": ["png", "PNG", "jpg", "JPG", "jpeg", "JPEG", "bmp", "BMP"],
    "bg_size": [640, 480],
    "dataset_size": 10,
    "threads": 2,
    "compile_options": {
        "GPU": 0,
        "CUDNN": 0,
        "CUDNN_HALF": 0,
        "OPENCV": 0
    }
}

• extensions: [ext1, ext2, ...]The allowed extensions of background/object images. Add more extensions if necessary - it will work as long as the Python Pillow library supports it.
• bg_size: [width, height] Resolution of the background image in pixels.
• dataset_size: Generated YOLO dataset per background. YOLO would train & test with <dataset_size> * <number of background images> images.
• threads: Number of threads to run Step 1 and Step 2.
• compile_options: Compilation options for YOLO darknet. Use only if you satisfied the requirements.

If you don't want settings.json to be on git changes, run

git update-index --skip-worktree settings.json

from the project directory.

If it's the first time running pyOLO, just run

python pyOLO.py

from the project root directory. You could also run the process step by step by giving the step number as a argument:

python pyOLO.py <1-5>

• Step 1: Preparing Object Images

• Step 2: Generating Dataset

• Step 3: Setting Up YOLO Environment

• Step 4: Run YOLO Training

• Step 5: Run YOLO Testing