Interpretability 11: sparse initialization

from kan import *

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(device)

model = KAN([5,5,5,1], sparse_init=False, device=device)
x = torch.rand(100,5).to(device)
model.get_act(x)
model.plot()

cuda
checkpoint directory created: ./model
saving model version 0.0
../_images/Interp_11_sparse_init_1_1.png 
from kan import *

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(device)

model = KAN([5,5,5,1], sparse_init=True, device=device)
x = torch.rand(100,5).to(device)
model.get_act(x)
model.plot()

cuda
checkpoint directory created: ./model
saving model version 0.0
../_images/Interp_11_sparse_init_2_1.png