Skip to content

Latest commit

 

History

History
177 lines (108 loc) · 5.35 KB

README.md

File metadata and controls

177 lines (108 loc) · 5.35 KB

🟢 FrugalAI

Answers to the challenge FrugalAI task — Deforestation Detection This repository contains code, notebooks, and results related to the Frugal AI Challenge — an ML competition encouraging efficient and environmentally sustainable AI models (e.g., minimizing energy, CO₂ impact) while maintaining strong performance. ([GitHub][2])

📌 Project Overview

The goal of this repository is to build a model to detect illegal deforestation (e.g., chainsaw sounds) in audio recordings while prioritizing frugality: low compute, low energy, and high accuracy.

This includes:

  • A Jupyter notebook pipeline for model development
  • A RandomForest.ipynb showing a lightweight classical model
  • Data preprocessing artifacts in spectrograms/
  • Quantified results (accuracy and environmental metrics)

📁 Repository Structure

FrugalAI/
├── spectrograms/            # (likely) audio preprocessing outputs or utilities
├── FrugalAI.ipynb           # Main notebook: model development & evaluation
├── RandomForest.ipynb       # Notebook exploring Random Forest model baseline
├── README.md                # This comprehensive README

🧠 Challenge Context

This project was built as part of the Frugal AI Challenge — an initiative promoting efficient AI solutions that balance performance with environmental sustainability. Participants were evaluated not only on accuracy but also on energy consumption during training & inference. ([GitHub][2])

Your repository’s FrugalAI.ipynb likely includes loading audio data, transforming it into spectrograms/features, training a model, validating results, and reporting the energy footprint.

🚀 Quick Start

🛠️ Requirements

General tools you’ll need:

  • Python 3.8+
  • Jupyter Notebook or Jupyter Lab
  • ML packages (e.g., scikit‑learn, librosa, numpy, pandas)
  • Optional: visualization libraries (matplotlib, seaborn)

Install dependencies (example):

pip install numpy scipy pandas scikit-learn librosa matplotlib

You may need to add additional packages based on the notebooks’ imports.

📊 Running the Notebook

  1. Open FrugalAI.ipynb in Jupyter

    • Follow the cells: load data, preprocess audio, extract features.
    • Train and evaluate your model.
    • Record metrics such as accuracy, energy consumption, and CO₂ impact (if measured).

  2. Explore RandomForest.ipynb

    • Compares a Random Forest baseline — often efficient and interpretable.
    • Typically useful for tabular or engineered audio feature data.

  3. Usage of spectrograms/

    • Contains outputs from audio → spectrogram conversion for model input.
    • You can visualize, augment, or use these features for ML pipelines.

📈 Evaluation & Results

According to the repository summary:

  • Your Random Forest baseline achieved ~80% accuracy
  • With extremely low energy usage: ≈ 0.00017 kg CO₂ and ≈ 0.003002 kWh of electricity (measured or estimated) — demonstrating frugality. ([GitHub][3])

This aligns with the “frugal” criteria of the challenge: strong performance with minimal environmental and computational cost.

🧩 How It Works (High Level)

  1. Data Input

    • Audio recordings (e.g., short clips of chainsaw vs. environment)
    • Converted to spectrograms (time‑frequency representations)

  2. Feature Extraction

    • Extract spectral features or embed audio for classification

  3. Model Training

    • Lightweight models (e.g., Random Forest, small neural networks)
    • Emphasis on low training and inference cost

  4. Evaluation

    • Measure accuracy on validation/test sets
    • Optionally log energy impact (e.g., using packages like codecarbon)

  5. Reporting

    • Document results, charts, and insights in notebooks

🧪 Example Code Snippet (Audio to Spectrogram)

Here’s a typical snippet you might see in the notebooks (example for feature extraction):

import librosa
import numpy as np

def extract_spectrogram(audio_path, sr=12000, n_mels=128):
    y, sr = librosa.load(audio_path, sr=sr)
    spect = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=n_mels)
    spect_db = librosa.power_to_db(spect, ref=np.max)
    return spect_db

spectrogram = extract_spectrogram("sample_audio.wav")

Use these spectrograms as features in any ML classifier.

📦 Dependencies (suggested)

Typical Python packages include:

  • numpy
  • pandas
  • scikit-learn
  • librosa
  • matplotlib / seaborn
  • notebook

Add these to a requirements.txt or Conda environment for reproducibility.

🏆 Challenge Impact

Frugal AI encourages responsible and efficient AI — prioritizing both accuracy and sustainability. Models like yours demonstrate how lightweight approaches like Random Forests or simple feature‑based classifiers can be competitive while drastically reducing energy consumption. ([GitHub][2])

❤️ Contributing

If this evolves further, consider adding:

  • Dataset loader scripts
  • A CLI or Python module for reproducible runs
  • Scripts to measure and record environmental metrics
  • A Dockerfile or Binder config for easy sharing
  • More model baselines (light neural nets, boosted trees)

📄 License

If not already specified, consider adding an MIT License or another permissive open‑source license.