VGG-like Kolmogorov-Arnold Convolutional network with Gram polynomials

This model is a Convolutional version of Kolmogorov-Arnold Network with VGG-11 like architecture, pretrained on Imagenet1k dataset. KANs were originally presented in [1, 2]. Gram version of KAN originally presented in [3]. For more details visit our torch-conv-kan repository on GitHub.

Model description

The model consists of consecutive 10 Gram ConvKAN Layers with InstanceNorm2d, polynomial degree equal to 5, GlobalAveragePooling and Linear classification head:

KAGN Convolution, 32 filters, 3x3
Max pooling, 2x2
KAGN Convolution, 64 filters, 3x3
Max pooling, 2x2
KAGN Convolution, 128 filters, 3x3
KAGN Convolution, 128 filters, 3x3
Max pooling, 2x2
KAGN Convolution, 256 filters, 3x3
KAGN Convolution, 256 filters, 3x3 10 Max pooling, 2x2
KAGN Convolution, 256 filters, 3x3
KAGN Convolution, 256 filters, 3x3
Max pooling, 2x2
KAGN Convolution, 512 filters, 3x3
KAGN Convolution, 512 filters, 3x3
Global Average pooling
Output layer, 1000 nodes.

Intended uses & limitations

You can use the raw model for image classification or use it as pretrained model for further finetuning.

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
model = vggkagn(3,
                1000,
                groups=1,
                degree=5,
                dropout=0.15,
                l1_decay=0,
                dropout_linear=0.25,
                width_scale=2,
                vgg_type='VGG11v4',
                expected_feature_shape=(1, 1),
                affine=True
                )
model.from_pretrained('brivangl/vgg_kagn11_v4')

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]),
    ])

Training data

This model trained on Imagenet1k dataset (1281167 images in train set)

Training procedure

Model was trained during 200 full epochs with AdamW optimizer, with following parameters:

{'learning_rate': 0.0009, 'adam_beta1': 0.9, 'adam_beta2': 0.999, 'adam_weight_decay': 5e-06,
'adam_epsilon': 1e-08, 'lr_warmup_steps': 7500, 'lr_power': 0.3, 'lr_end': 1e-07, 'set_grads_to_none': False}

And this augmnetations:

transforms_train = v2.Compose([
    v2.ToImage(),
    v2.RandomHorizontalFlip(p=0.5),
    v2.RandomResizedCrop(224, antialias=True),
    v2.RandomChoice([v2.AutoAugment(AutoAugmentPolicy.CIFAR10),
                     v2.AutoAugment(AutoAugmentPolicy.IMAGENET)
                     ]),
    v2.ToDtype(torch.float32, scale=True),
    v2.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

Evaluation results

On Imagenet1k Validation:

Accuracy, top1	Accuracy, top5	AUC (ovo)	AUC (ovr)
61.17	83.26	99.42	99.43

On Imagenet1k Test: Coming soon

BibTeX entry and citation info

If you use this project in your research or wish to refer to the baseline results, please use the following BibTeX entry.

@misc{torch-conv-kan,
  author = {Ivan Drokin},
  title = {Torch Conv KAN},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/IvanDrokin/torch-conv-kan}}
}

References

[1] Ziming Liu et al., "KAN: Kolmogorov-Arnold Networks", 2024, arXiv. https://arxiv.org/abs/2404.19756
[2] https://github.com/KindXiaoming/pykan
[3] https://github.com/Khochawongwat/GRAMKAN

brivangl
/

vgg_kagn11_v4