LArctan-SKAN

English / 简体中文

Introduction

This repository contains the experimental code for the paper "LArctan-SKAN: Simple and Efficient Single-Parameterized Kolmogorov-Arnold Networks using Learnable Trigonometric Function" [1] (paper at this link). If you are looking for a Python library to quickly build SKANs, please click here to visit the SKAN GitHub repository.

Usage

This code runs on Python 3.12.3. To use the library, ensure that the following Python packages are installed:

numpy==2.1.2
pandas==2.2.3
scikit_learn==1.5.2
single_kan==0.2.0
torch==2.4.1+cu121
torchvision==0.19.1+cu121
tqdm==4.66.4

To execute the experiment in the paper, run the following script:

python LArctan_SKAN_30epoch_lr000101.py

This will run the experimental setup described in the paper. Note that this repository only includes code for the LSS-SKAN, LSin-SKAN, LCos-SKAN, and LArctan-SKAN networks. For other networks, refer to the GitHub repository LSS-SKAN and the file LSS_SKAN_30epoch_lr000101.py.

References

[1] Z. Chen and X. Zhang, “LSS-SKAN: Efficient Kolmogorov-Arnold Networks based on Single-Parameterized Function,” Oct. 19, 2024, arXiv: arXiv:2410.14951. doi: 10.48550/arXiv.2410.14951.