omarrayyann / KAN-Conv2D

Drop-in convolutional KAN replacement of Conv2d

Convolutional 2D KAN Implementation

This repositry contains 3 drop-in convolutional KAN replacements. Each work on top of a different KAN implementation:

1. Efficient implementation of Kolmogorov-Arnold Network (KAN)
2. Original KAN implementation
3. Fast KAN implementation

Installation

git clone [e-mail]/omarrayyann/KAN-Conv2D
cd KAN-Conv2D
pip install -r requirements.txt

Usage

You should be able to just replace torch.nn.Conv2D() with ConvKAN()

from ConvKAN import ConvKAN

# Implementation built on the efficient KAN Implementation (https://github.com/Blealtan/efficient-kan)
conv = ConvKAN(in_channels=3, out_channels=4, kernel_size=3, stride=1, padding=1, version="Efficient")

# Implementation built on the original KAN Implementation (https://github.com/KindXiaoming/pykan)
conv = ConvKAN(in_channels=3, out_channels=4, kernel_size=3, stride=1, padding=1, version="Original")

# Implementation built on the fast KAN Implementation (https://github.com/ZiyaoLi/fast-kan)
conv = ConvKAN(in_channels=3, out_channels=4, kernel_size=3, stride=1, padding=1, version="Fast")