search

StarostinV / convkan

Convolutional layer for Kolmogorov-Arnold Network (KAN)
MIT License
76 stars 8 forks source link

ConvKan don't work well with pytorch-summary #6

Closed maayange17 closed 3 months ago

maayange17 commented 3 months ago

Using pytorch-summary, I saw that the parameters number is counted as zero. I also found a solution.

See the next issue: https://github.com/sksq96/pytorch-summary/issues/65

StarostinV commented 3 months ago

Honestly, pytorch-summary implementation is a bit strange. A usual way to get the total number of parameters is

total_num_params = sum(p.numel() for p in model.parameters())
trainable_num_params = sum(p.numel() for p in model.parameters() if p.requires_grad)

Here numel() method returns exactly what you want: the total number of elements in the tensor.

This works perfectly with convkan.

EDIT: by the way, you can probably check out the successor of pytorch-summary - torchinfo