brivangl/vgg_kagn_bn11_v4_opt

How to use

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]),
    ])