Dormouse Detection

Detecting Dormice with Object Detection and Transfer Learning

This repository contains a machine learning pipeline for training an object detection model to identify hazel dormice (Muscardinus avellanarius) in images. Additionally, the dataset is provided in dormouse-detection/training_data and a pre-trained model is downloadable from Releases.

The model architecture uses transfer learning from a YOLOv8 model pre-trained on the ImageNet dataset.

The Dormouse Detection Dataset in this project was created by downloading images from iNaturalist and annotating them with CVAT (Computer Vision Annotation Tool).

🔧 Built with Keras and TensorFlow.

Prediction Results

Project Structure

dormouse-detection/
├── data_downloader/          # Scripts for downloading data
├── logs/                     # Training logs and metrics
├── plots/                    # Visualisation of results
├── src/                      # Source code for the pipeline
│   ├── __init__.py
│   ├── data_loader.py        # Handles data loading and preprocessing
│   ├── logger.py             # Logging
│   ├── metrics_callback.py   # Model performance metrics computed during training
│   ├── model.py              # Model architecture
│   ├── training_pipeline.py  # Training and evaluation
│   ├── utils.py              # Utility functions
├── tests/                    # Unit testing
├── trained_models/           # Saved trained models
├── training_data/            # Training dataset
├── .gitignore                
├── config.py                 # Configuration for model training and hyperparameters
├── inference.py              # Script for running inference with trained models
├── main.py                   # Main entry point for training
├── poetry.lock               
├── pyproject.toml            
└── README.md

Setup and Installation

Create a virtual environment using Conda.

# Create the environment
conda create --name env python=3.9.6

# Activate the environment
conda activate env

Install Poetry and project dependencies.

# Install Poetry
pip install poetry

# Install dependencies
poetry install

Pre-trained Model

In addition to training your own model using the instructions in the Training section, a pre-trained model for those who wish to directly use it for inference without going through the training process is available.

You can download the .keras model file from the Releases section.

Training

To train the model, use the main.py as the entry point:

python main.py

This will load configuration parameters from config.py and start the training process, saving models and logs in their respective directories. After training completes, the test set predictions are visualised and saved to /plots.

To view the training performance on Tensorboard:

tensorboard --logdir=logs/fit

Inference

For running inference on new images using a trained model:

python inference.py --model_path trained_models/model_name.keras --image_path path/to/image.jpg

Or, to predict all images in a folder:

python inference.py --model_path trained_models/model_name.keras --image_dir path/to/dir