torch-conv-kan

Why we included this project

Kolmogorov-Arnold networks swap fixed activation functions for learnable ones, and this repository collects convolutional versions of them in one place. You get 1D, 2D, and 3D KAN conv layers built on several basis-function families, from B-splines and Legendre to Chebyshev, Gram, wavelet, and Bernstein, so it is easy to compare how each behaves inside otherwise familiar architectures. The repo also bundles ResNet-, DenseNet-, VGG-, and U-Net-style models with accelerate-based training scripts, which makes reproducing the paper's CIFAR and Tiny ImageNet experiments fairly straightforward. Pretrained ImageNet-1k checkpoints are included, so you can see what these layers actually do at scale instead of only on small datasets. If you are weighing whether KAN convolutions belong in a vision pipeline, this is one of the more complete places to start.