Regularized Contrastive Learning for Few-shot Bioacoustic Sound Event Detection

Authors: Ilyass Moummad, Romain Serizel, Nicolas Farrugia

This is the implementation of our paper accepted at ICASSP 2024
We improve over our previous work that ranked 2nd in the DCASE 2023 Challenge Task5 by adding a regularization term to the training loss, and by improving the inference strategy

args.py: contains default values for arguments

Data

Download the dataset from Zenodo: https://zenodo.org/records/6482837 To create the spectrograms of the training set:
create_train.py: with argument --traindir for the folder containing the training datasets

Training

To train the feature extractor:
train.py: with arguments --traindir (the same as above) --method: optional argument for the pretraining method scl for SupCon and ssl for SimCLR

Evaluation

To evaluate using finetuning:
eval_finetune: with arguments --valdir for the folder containing the validation datasets
To evaluate without finetuning (for faster inference):
eval_nofinetune: with arguments --valdir for the folder containing the validation datasets

To get the scores:
evaluation.py : with arguments --pred_file for the predictions csv file created by the eval script (the file is in : traindir/../../outputs/eval.csv'), --ref_files for the path of validation datasets (same as --valdir), and --save_path for the folder where to save the json file containing the scores

If you any question or a problem with the code/results do not hesitate to mail me on : ilyass.moummad@imt-atlantique.fr or open an issue on this repository, I am very responsive.

---

To cite our paper

@misc{moummad2023regularized,
      title={Regularized Contrastive Pre-training for Few-shot Bioacoustic Sound Detection}, 
      author={Ilyass Moummad and Romain Serizel and Nicolas Farrugia},
      year={2023},
      eprint={2309.08971},
      archivePrefix={arXiv},
      primaryClass={cs.SD}
}