A real world application based on KAN

[yli-uos]

Very interesting work!

I want to make a real world application accelerator based on KAN (on FPGA or GPU) and trained the KAN neural network on MNIST dataset. After training I would like to get all the activation functions using "model.auto_symbolic(lib=lib)" as shown in the attachment but failed. I randomly pick on node to check the function and the shape looks like the graph below.

My questions are: Can all the functions be expressed as the functions in the lib? How to get the exact expression of the activation like the graph above?

The attachments include the training file and the .pth model after training.

happy to discuss! I really appreciate it~
KAN_MNIST.zip