CUDA implementation of Legendre KAN

References:

• pytorch implementation

Note

• There are no optimizations other than memory access coalescing. I a cuda beginner and willing to receive optimization suggestions : )

Start

1. Install

pip install -e .

Make sure the version of nvcc in PATH is compatible with your current PyTorch version (it seems minor version difference is OK).

2. Run

   • Run test on MNIST:

   python cheby_test.py
   

   • Run benchmark (code from KAN-benchmarking):

   python benchmark.py --method all --reps 100 --just-cuda