Self-Supervised Learning for Few-Shot Bird Sound Classification

Authors: Ilyass Moummad, Romain Serizel, Nicolas Farrugia

Update: Our paper is accepted to ICASSPW SASB 2024

Preprint: https://arxiv.org/abs/2312.15824

We train a feature extractor on the training split of BirdCLEF2020 using self-supervised learning and evaluate it on the validation/test splits following MetaAudio Benchmark

To Download BirdCLEF 2020 data, go to Aircrowd and make an account to be able to access LifeCLEF 2020 Bird Challenge ressources.
Go to Ressources and download the file that contains "Download Links" for Train (Train data from the challenge is split into new train/val/test sets for the few-shot benchmark, we refer to MetaAudio for more details).

Data Prepration

Put CNN14 PANN checkpoint "Cnn14_map=0.431.pth" in util/ folder
pann_selection.py: from the dataset stored in --datapath, this script creates .pt files with the highest PANN activation of birds in ---targetpath

For all the training/evaluation scripts, specify --datapath for the stored data: ps stands for PANN Selection, and tp stands for Temporal Proximity

Training

train_ps.py: train on the PANN selected segments using --loss loss function
train_tp.py: train on the whole training set using temporal proximity for sampling two views if --tprox otherwise it uses two random crops from each audio file

Supported training losses : Barlow Twins bt, SimCLR simclr, FroSSL fro, and SupCon supcon

Evaluation

eval_ps.py: evalute the model trained using --loss on the --split split on the segments selected using PANN
eval_tp.py: evalute the model trained using --loss on the --split split on each chunk of the file, predictions are aggregated using mean \

args.py: contains all the arguments beside the one to be specified above, set to default values, as well as their description

To cite this work:

@misc{moummad2023selfsupervised,
      title={Self-Supervised Learning for Few-Shot Bird Sound Classification}, 
      author={Ilyass Moummad and Romain Serizel and Nicolas Farrugia},
      year={2023},
      eprint={2312.15824},
      archivePrefix={arXiv},
      primaryClass={cs.SD}
}