Convolutional KANs
Collection
Collection of pretrained Conv KANs
•
5 items
•
Updated
•
1
First, clone the repository:
git clone https://github.com/IvanDrokin/torch-conv-kan.git
cd torch-conv-kan
pip install -r requirements.txt
Then you can initialize the model and load weights.
import torch
from models import VGGKAGN_BN
model = VGGKAGN_BN.from_pretrained('brivangl/vgg_kagn_bn11_v4_opt',
groups=1,
degree=3,
dropout=0.05,
l1_decay=0,
width_scale=3,
affine=True,
norm_layer=nn.BatchNorm2d,
expected_feature_shape=(1, 1),
vgg_type='VGG11v4')
Transforms, used for validation on Imagenet1k:
from torchvision.transforms import v2
transforms_val = v2.Compose([
v2.ToImage(),
v2.Resize(256, antialias=True),
v2.CenterCrop(224),
v2.ToDtype(torch.float32, scale=True),
v2.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])