search

KindXiaoming / pykan

Kolmogorov Arnold Networks
MIT License
14.95k stars 1.37k forks source link

the experiment code about fitting LANs to images #388

Open yeruoforever opened 2 months ago

yeruoforever commented 2 months ago

This is an extremely exciting work, which allows us to fit functions with interpretable symbols, which is bound to bring a new perspective to image processing and representation, such as representing vivid images with very few parameters.

However, when I was exploring, I encountered problems: the network based on KAN could not fit the image like SIREN, and its performance did not even reach the MLP network that does not use positional encoding.

Is the author convenient to release the network code in Appendix B.3 for me to check my own mistakes?

I would be very grateful.

KindXiaoming commented 2 months ago

Thank you for your kind words! it would take a while to clean up the codes, but we indeed need to carefully choose initializations like in SIREN. In fact, the LAN network is exactly SIREN when splines are turned off, and splines are only learning minor corrections to sines. I never tried KANs for images since it was too slow back then.