Something different from the official results for KAN. about efficient-kan HOT 1 OPEN

Hi ASCIIJK! I was very interested in your comment and tried to play a bit with the EfficientKAN code.

When I replicated the with EfficientKAN, I got results similar to yours.

However, as you said, the sb_trainable and sp_trainable parameters seem to be missing in the constructor, so I decided to add them, just like PyKAN does.

As you can see, I use them to disable the trainability of the base_weight and splin_weight parameters.

After training with the configuration

KAN([1, 1],
        grid_size=200,
        grid_eps=1,
        scale_base=0,
        sp_trainable=False,
        sb_trainable=False,
)

with 100 epochs for each peak as shown in the , I got the following result:

You can read my training code here: ,
and how I edited the EfficientKAN class here:

from efficient-kan.