May I know which part of your codes to reduce the nodes like the pictures below?

[sunrainmao]

[ZPZhou-lab]

Unfortunately, the current implementation is only for the KAN infrastructure to test the actual performance of this new network structure in various tasks. The author of KAN has added many fancy designs to the interpretability of the model, such as sparse regularization(similar to L1) you show. These functions have not been implemented yet, so the feature of reducing link nodes cannot be achieved in the current training process.

If you think it is necessary, I can consider spending some time to implement it.

Best! 😊