Mixture of Mixups for Multi-label Classification of Rare Anuran Sounds

Authors: Ilyass Moummad, Nicolas Farrugia, Romain Serizel, Jeremy Froidevaux, Vincent Lostanlen

---

Presented at EUSIPCO 2024 - Special Session on Signal Analysis for Biodiversity. Access the full paper here.

This project introduces a framework utilizing mixing regularization methods—Mixup, Manifold Mixup, and MultiMix—to address challenges in multi-label classification and class imbalance within the Anuraset dataset.

The implementation is based on the official AnuraSet baseline, available on GitHub. You can also download the dataset directly from Nature's publication.

Required Python Libraries

pip install -r requirements

Repository Structure

• main.py: Main script for training and evaluation.
• dataset.py: Dataset class for handling data operations.
• models.py: Contains the MobileNetV3 model implementation.
• train.py: Utility functions for model training.
• val.py: Utility functions for model validation.
• transforms.py: Data transformation classes.
• args.py: Argument parsing.

Training and Evaluation

• Original AnuraSet: python3 main.py --dataset anuraset --rootdir <dataset_path> --mix mix2 --device cuda --sr 16000 --workers 16 --save

• AnuraSet-36N (removing non-overlapping classes between training and testing): --dataset anuraset36n

• AnuraSet-36 (removing non-overlapping classes as well as silence examples): --dataset anuraset36

Replace <dataset_path> with the actual path to your dataset.

Citation

If you find this work useful, please cite it:

@misc{2403.09598,
  Author = {Ilyass Moummad and Nicolas Farrugia and Romain Serizel and Jeremy Froidevaux and Vincent Lostanlen},
  Title = {Mixture of Mixups for Multi-label Classification of Rare Anuran Sounds},
  Year = {2024},
  Eprint = {arXiv:2403.09598},
}