CLAP-S: Support Set-Based Adaptation for Downstream Fiber-Optic Acoustic Recognition

Jingchen Sun^{1, 2}, Shaobo Han¹, Wataru Kohno¹, Changyou Chen²

¹ NEC Labs America, USA
² University at Buffalo, The State University of New York, USA

Introduction

This repository contains the code and dataset for our paper CLAP-S: Support Set-Based Adaptation for Downstream Fiber-Optic Acoustic Recognition. [Paper]

Environment Installation

• Setup conda environment (recommended).

# Create a conda environment
conda create -n claps python=3.10 -y

# Activate the environment
conda activate clap

# Clone this repository
git clone [email protected]:Jingchensun/clap-s.git
cd clap-s
# Install torch
pip install -r requirement.txt

1 Dataset Preparation

The Gunshot-Firework dataset originates from the paper Deep Learning-based Intrusion Detection and Impulsive Event Classification for Distributed Acoustic Sensing across Telecom Networks. We have restructured and released this dataset to facilitate the reproduction of the results presented in the paper. Below are some sample mel spectrograms generated from the audio recordings in this dataset.

Download the outdoor gunshot-firework fiber acoustic dataset from Google Drive and extract them to $data$ folder.

tar -xvf gunshot_firework_data.tar.gz

The Final data structure should looks like:

data/
|–– coil_gunshot/
|–– coil_gunshot_train_labels.txt
|–– coil_gunshot_test_labels.txt
|–– fiber_gunshot/
|–– fiber_gunshot_train_labels.txt
|–– fiber_gunshot_test_labels.txt

2 Run the CLAP-S

This command runs the CLAP-Support on the recorded ESC50 dataset. You can modify the dataset and random seed.

sh clap_support.sh 

3 Run the CLAP-S⁺

Before running CLAP-S⁺, you need to obtain the weights of the 'Adapter,' which consists of two MLP layers. Use the command below to train a few-shot Adapter and save the checkpoint.

sh clap_adapter.sh

After obtaining the Adapter checkpoint, you can train the trainable audio feature Adapter in CLAP-S. Note that this implicit Adapter is different from the explicit Adapter you trained earlier. This highlights a key insight of our method: leveraging the few-shot data twice.

sh clap_support_plus.sh

4 Experiment Result

Fiber-Gunshot Performance

Model	Training Type	1-shot	2-shot	4-shot	8-shot	Full shot
CLAP-S	Training-Free	49.0	57.0	60.0	71.0	84.0
CLAP-S⁺	Training-Required	58.0	66.0	70.0	78.0	86.0

Coil-Gunshot Performance

Model	Training Type	0-shot	2-shot	4-shot	8-shot	Full shot
CLAP-S	Training-Free	57.0	65.0	67.0	72.0	85.0
CLAP-S⁺	Training-Required	62.0	68.0	67.0	69.0	87.0

Contact

If you have any questions regarding this repo, please contact Jingchen Sun ([email protected]).

Citing CLAP-S

If you find this repository useful, please consider giving a star ⭐ and citation

@article{sun2025clap,
  title={CLAP-S: Support Set Based Adaptation for Downstream Fiber-optic Acoustic Recognition},
  author={Sun, Jingchen and Han, Shaobo and Kohno, Wataru and Chen, Changyou},
  journal={arXiv preprint arXiv:2501.09877},
  year={2025}
}

@article{han2024deep,
  title={Deep Learning-based Intrusion Detection and Impulsive Event Classification for Distributed Acoustic Sensing across Telecom Networks},
  author={Han, Shaobo and Huang, Ming-Fang and Li, Tingfeng and Fang, Jian and Jiang, Zhuocheng and Wang, Ting},
  journal={Journal of Lightwave Technology},
  year={2024},
  publisher={IEEE}
}