search

Blealtan / efficient-kan

An efficient pure-PyTorch implementation of Kolmogorov-Arnold Network (KAN).
MIT License
3.49k stars 306 forks source link

Contributing? #5

Closed dschonholtz closed 1 month ago

dschonholtz commented 1 month ago

Hey I have a fork, with some not useful stuff on it yet. (Mostly just profiling showing forward passes suck due to b-splines and some comparisons to MLPs.)

Do you want folks to contribute to this? Are you interested in making b-splines more efficient with something like: https://github.com/GistNoesis/FourierKAN/blob/main/fftKAN.py

Let me know what you think.

dschonholtz commented 1 month ago

Hey @Blealtan I am gonna roll my own version of this repo since it seems like you aren't actively replying. I'll be pulling in the conv kan, fourier kan and doing comparisons on all of them.

Also, hoping to get into really doing break downs of why things are happening the way they are happening.

minhhnguyen0312 commented 1 month ago

Hi there @dschonholtz. Have you tried to measure the performance of ur algorithm compare to this one and the original?

You may refer to this tutorial

I'm looking forward to your results. Have a nice day!

Blealtan commented 1 month ago

Yeah the Fourier variant definitely sounds attractive.

I do have little time maintaining this right now (working on my PhD viva). Also, this repo mostly serves as a playground to show the idea that KAN (in its current shape) can be reformulated into sth like MLP and become computationally efficient. If you are going down on real experiments and have performance issues around all of this, please let me know and I'm willing to help.

In short: for now, I don't think I have the time to accept contributions; please go ahead fork the repo or copy code snippets under MIT license.