Something different from the official results for KAN.

[ASCIIJK]

I try to reproduce the experiments (example 4 in official KAN). With official KAN, I get the results as below (Ground-truth is at the top, and the predication is at the bottom):

But with the efficient-kan, I get the results as below:

It shows that previous peak will be higher when learning new peak.
The official model is create by: "model = KAN(width=[1, 1], grid=200, k=3, noise_scale=0.1, bias_trainable=False, sp_trainable=False, sb_trainable=False)"
The efficient-kan model is created by: "model = KAN([1, 1], grid_size=200)"
It seems to be the same except for "bias_trainable=False, sp_trainable=False, sb_trainable=False".

[MrPio]

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: