SViG: A Similarity-thresholded Approach for Vision Graph Neural Networks

Accepted in IEEE Access, 2025

Link to Paper: https://ieeexplore.ieee.org/document/10845790

Requirements

• Pytorch 1.7.1
• timm 0.3.2
• torchprofile 0.0.4
• apex
• torch_scatter 2.0.7

Please, use the following commands to install the requirements:

conda env create -f environment.yml

pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio===0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-1.7.1+cu110.html

SViG-Ti models pretrained on ImageNet

Each checkpoint has two checkpoints, one for the model updated without using EMA, and an EMA-model. If the EMA column is checked, please use --model-ema when evaluating.

Starting Threshold	Decrement	Top-1	EMA
0.86	0.03	74.6	✗
0.89	No Decrement	74.1	✓
0.89	0.02	74.2	✓
0.89	0.03	74.6	✗
0.89	0.04	74.0	✓
0.91	0.03	74.5	✓
0.93	0.03	74.4	✓
0.96	0.03	74.2	✓

Evaluation

• Evaluate example:

python -m torch.distributed.launch --nproc_per_node=8 train.py <ImageNet_path> --model vig_ti_224_gelu -j 8 --amp --drop 0 --drop-path .1 -b 128 --num-classes 1000 --pretrain_path <checkpoint_path> --evaluate --start-thresh <starting_threshold> --dec <decrement_per_layer> --model-ema(use only if you want to use the EMA checkpoint)

Training

• Training SViG on 8 GPUs:

python -m torch.distributed.launch --nproc_per_node=8 train.py <path_to_imagenet> --model vig_ti_224_gelu --sched cosine --epochs 300 --opt adamw -j 8 --warmup-lr 1e-6 --mixup .8 --cutmix 1.0 --model-ema --model-ema-decay 0.99996 --aa rand-m9-mstd0.5-inc1 --color-jitter 0.4 --warmup-epochs 20 --opt-eps 1e-8 --repeated-aug --remode pixel --reprob 0.25 --amp --lr 2e-3 --weight-decay .05 --drop 0 --drop-path .1 -b 128 --output <path_to_save_models> --start-thresh <starting_threshold> --dec <decrement_per_layer>

To cite this work, please use the following bibTex

@ARTICLE{10845790,
  author={Elsharkawi, Ismael and Sharara, Hossam and Rafea, Ahmed},
  journal={IEEE Access}, 
  title={SViG: A Similarity-Thresholded Approach for Vision Graph Neural Networks}, 
  year={2025},
  volume={13},
  number={},
  pages={19379-19387},
  keywords={Computer vision;Computer architecture;Vectors;Graph neural networks;Transformers;Image edge detection;Point cloud compression;Image resolution;Image representation;Image classification;Graph Neural Networks;Vision Graph Neural Networks;Image Classification},
  doi={10.1109/ACCESS.2025.3531691}}

Acknowledgement

This repo partially uses code from Vision Graph Neural Networks, deep_gcns_torch and timm.