The KA-Conv repository offers implementations of Kolmogorov-Arnold Convolutional Networks (KA-Conv) with different basis functions. This project aims to extend and refine the ConvKAN framework by integrating various activation functions and providing comparative performance metrics.

pip install kaconv
pip install -r requirements.txt

git clone https://github.com/XiangboGaoBarry/KA-Conv.git
python -m build
pip install -r requirements.txt

The following table presents the comparative results of different KA-Conv implementations using various activation functions. Key metrics include accuracy, parameter count, and throughput.

We compare the our results with

• EfficientKANLinear: Implemented as per EfficientKANLinear
• FastKANLinear: Implemented as per FastKANLinear

Currently, with the same hidden layer setups, KA-Conv with RBF and BSpline activations outperform the original nn.Conv2d. However, KA-Conv also adds extra complexity, leading to more parameters and lower throughput. When reducing the number of parameters of the model to the same level as that of the model implemented with nn.Conv2d, the performance of the model implemented with KA-Conv is lower.

KA-Conv has lower throughput than nn.Conv despite that our implementation has +93% acceleration over other implementations.

We are comparing the performance of the model on larger datasets and larger models, such as ResNet on ImageNet. The results will be released soon.

This model is built upon FastKAN. We extend our gratitude to the creators of the original KAN for their pioneering work in this field.