CIANLabxBROWNUniv/fin_whale_1

Paper Informations

Self-Supervised vs Supervised Representation Learning for Fin Whale Vocalization Detection by Adam Chareyre, Haodong Zhang, Shuwen Ge, Randall Balestriero, Sébastien Paris, Hervé Glotin OpenReview link: https://openreview.net/forum?id=fCe4l0T320

Requirements

Update path files

Depending of the method you want to use (SSL or Supervised), you will need to fill some path. .env file, SSL_PRETRAINED/DP_run.sh file, SSL_PRETRAINED/fine_tune.sh file

Install dependencies

Certain modules are required to run this code.

The Python version used is Python 3.11.2.

To install all the necessary modules, run the following command:

pip install -r requirements.txt

How to start an experiment

Supervised experiement

After updating the .env file, you can execute the following command to reproduce experiments :

python -m fin_whale_1.EXPERIMENTATIONS.tsne_supervised
python -m fin_whale_1.EXPERIMENTATIONS.sem_snr_10_seeds
python -m fin_whale_1.EXPERIMENTATIONS.sem_size_10_seeds
python -m fin_whale_1.EXPERIMENTATIONS.sem_classic_10_seeds

Please refer to documentation in the top of each file to get more description.

SSL experiement

If users want to do pretrain, they should use

SSL_PRETRAINED/DDP_run.sh

It supports multi-gpu runing.

For a finetuning, the user must run the following command :

python -m fine_whale_1.SUPERVISED.SSL.fine_tune_model

To reproduce SSL experiments, you can execute the following command :

python -m fin_whale_1.EXPERIMENTATIONS.fineTune_classic_10_seeds
python -m fin_whale_1.EXPERIMENTATIONS.fineTune_size_10_seeds
python -m fin_whale_1.EXPERIMENTATIONS.fineTune_snr_10_seeds