Activation functions cannot be visualized

[XHX-HUST]

While such improvements can indeed enhance the computational efficiency of the original KAN, they also have a fatal drawback: the most prominent advantage of KAN over MLP lies in the symbolic regression analysis of its activation functions. However, E-KAN uses F.linear() for direct matrix multiplication, which can only obtain the output of each layer but cannot get the values of edge-wise activation functions.