KAN feels hard to learn "Addition"

[ycq091044]

Hello, thanks for the interesting method. I have a fail trial and need some help here.

Using the hellokan.ipynb template, I changed the complex formula into "Addition". With all other parts fixed, it seems the model training loss does not go down and cannot discovers the "Addition" operation.

[Flecart]

I think you should do as in hellokan you should prune the model, then retrain it, and prune it, until you can't prune it anymore

[KindXiaoming]

Hi Sorry, the API can be a bit confusing, but instead of

f = lambda x: x[:,0] + x[:,1]

it should be

f = lambda x: x[:,[0]] + x[:,[1]]

instead. If that's still not working, try reduce lamb. Let me know how this goes!

[CoryKornowicz]

I was also a bit confused about the definition of that lambda, and I redefined it locally as
f = lambda x: x[:, 0, np.newaxis] + x[:, 1, np.newaxis]
to make the return dimension more explicit. You can also replace np.newaxis with None and the effect is the same.

[ycq091044]

Thanks, change into f = lambda x: x[:,[0]] + x[:,[1]] works.