
This folder contains the code, including both CUDA kernels and fitting scripts, along with configurational YAML files, which contain all the details on our fitting protocols.

=================================================================

The ldKAN folder contains a Python package. It can be installed as 

pip install .

from the root (code_configs) directory

=================================================================

The folder experiments contains our training scripts along with yaml configurations we used. 

Subfolders cifar10 and imagenet contain the experiments with MLP- and lmKAN- based CNNs. 

The subfolder cifar10_fastkan_compare contains the comparison with FastKANs on the CIFAR-10 in a fully-connected setup.

The subfolder methane contains the experiments on the methane dataset. 

The mlp_teacher_experiments subfolder contains experiments with general function approximation. 


=================================================================

The code follows slightly different naming conventions compared to the final paper. 

ldKAN, or DKAN are the old names for lmKANs (lookup d-dimensional KAN or just D-dimensional KAN). 

Hessian regularization in the code is referred to as Frobenius regularization (as it is the Frobenius norm of the Hessian)


n_chunks in the code refers to the number of intervals, N_i, in the paper. 