L3S / PrototypeSound

Prototype Learning for Interpretable Respiratory Sound Analysis

This is a Python and PyTorch code for the prototype learning framework in our paper:

Zhao Ren, Thanh Tam Nguyen, and Wolfgang Nejdl. Prototype Learning for Interpretable Respiratory Sound Analysis. In ICASSP 2022. https://arxiv.org/abs/2110.03536

Citation

@inproceedings{ren2022prototype,
  title={Prototype learning for interpretable respiratory sound analysis},
  author={Ren, Zhao and Nguyen, Thanh Tam and Nejdl, Wolfgang},
  booktitle={ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={9087--9091},
  year={2022},
  organization={IEEE}
}

Abstract

The prototype learning framework aims to generate prototypes of audio singnals for a respiratory sound classification task (normal/crackle/wheeze/both).

The experiments are based on the ICBHI 2017 challenge database.

Preprocessing

python3 preprocessing/preprocessing.py

python3 preprocessing/data_split.py

Training and Test

1. For the basic CNN model without generating prototypes:

sh run.sh

2. For CNN models with generating prototypes:

sh run_prototype.sh

Note:The variant of 'PROTO_FORM' can be adapted for different prototype learning approaches:

Prototype-1D: PROTO_FORM='vector1d'

Prototype-2D-EleSim-Van: PROTO_FORM='vector2d'

Prototype-2D-EleSim-Att: PROTO_FORM='vector2d_att'

Prototype-2D-AvgSim-Van: PROTO_FORM='vector2d_avgp'

Prototype-2D-MaxSim-Van: PROTO_FORM='vector2d_maxp'

Prototype-2D-MaxSim-Att: PROTO_FORM='vector2d_maxp_att'