eso 0.1.0.dev50

Evolutionary Spectrogram Optimisation : optimise the selection and size of frequency bands of spectrograms to maximise deep learning classification accuracy while minimising the model's size using genetic algorithms

ESO: Evolutionary Spectrogram Optimisation 🧬🔊

Optimising spectrogram inputs for efficient passive acoustic monitoring using genetic algorithms. Based on our paper presented at the ICLR 2024 ML4RS Workshop.

---

The Challenge 🤔

Traditional bioacoustic monitoring with Convolutional Neural Networks (CNNs) often requires:

• Large, complex models.
• Computationally expensive preprocessing (like filtering and downsampling).
• This makes real-time analysis on low-resource edge devices difficult. 🐢⏳

Our Solution: ESO ✨

ESO (Evolutionary Spectrogram Optimisation) tackles this by using a Genetic Algorithm (GA) to intelligently select the most informative frequency bands directly from the original spectrogram.

Instead of processing the entire spectrogram, ESO:

1. 🧬 Encodes horizontal spectrogram bands (defined by position Pₜ and height h) as "genes".
2. 📜 Combines genes into "chromosomes", representing a specific selection of bands.
3. 💪 Evolves a population of chromosomes using selection, crossover, and mutation.
4. 📈 Optimizes for a fitness function balancing classification performance (F1-score) and model simplicity (trainable parameters).
5. 📉 Outputs the best chromosome, which defines narrow bands to be extracted and stacked, creating a highly compressed input for a much simpler CNN.

Concept from Figure 1: Genes define bands, chromosomes collect genes, bands are stacked for the CNN.

Key Benefits & Features 🚀

• ✅ Drastically Reduced Model Size: ~91% fewer trainable parameters compared to the baseline.
• ✅ Faster Inference: ~70% reduction in processing time for raw audio.
• ✅ Efficient: Minimizes need for heavy preprocessing like downsampling.
• ✅ Effective: Achieves comparable performance with only a minor (~4%) trade-off in F1-score.
• 🐍 Usable as a Python package.
• 🖥️ Includes an easy-to-use Graphical User Interface (GUI).
• 📊 TensorBoard integration for monitoring training progress.

Getting Started 🛠️

1. Clone the repository

   git clone https://github.com/ufuk-cakir/ESO.git
   cd ESO
   

2. Set up a virtual environment (recommended)

   python -m venv venv
   # On Linux/macOS
   source venv/bin/activate
   # On Windows
   venv\Scripts\activate
   

3. Install PyTorch based on your system configuration
   (see PyTorch to choose the correct version for your machine)

   pip install torch --index-url https://download.pytorch.org/whl/cu126
   

4. Install other dependencies

   pip install -r requirements.txt
   

Running ESO 🏃

• Using the GUI:

  python path/to/your/repository/eso_app.py 
  

  The GUI provides options to select data, configure hyperparameters, run ESO, and monitor progress (including TensorBoard).

• As a Python Package: Import the necessary modules from the eso package in your Python scripts. (Refer to the documentation or example scripts within the repository for specific usage details).

Citation ✍️

If you use ESO in your research or work, please cite our paper:

@inproceedings{cakir2024eso,
  title={{ESO: Evolutionary Spectrogram Optimisation for Passive Acoustic Monitoring}},
  author={Ufuk {\c{C}}ak{\i}r and Lor{\`e}ne Jeantet and Joel Lontsi and Emmanuel Dufourq},
  booktitle={ICLR 2024 Machine Learning for Remote Sensing (ML4RS) Workshop},
  year={2024},
  url={https://ml-for-rs.github.io/iclr2024/camera_ready/papers/51.pdf}
}