SpaRTAN-S

SpaRTAN is a lightweight architectural design which shows consistent efficiency and competitive performance when benchmarked against ImageNet and COCO dataset. It was introduced in the paper SpaRTAN and released in this repo. SpaRTAN-S is a scaled-up version of SpaRTAN-T.

Model Description

SpaRTAN-S shares the same configurations as SpaRTAN-T presented in the paper, SpaRTAN, except the number of channels at each stage, as outlined below.

Stage	Channel
S1	64
S2	128
S3	320
S4	512

Intended Uses & Limitations

You can use the raw model for image classification. Using as a feature extractor, SpaRTAN-S can be fine-tuned on various downstream tasks including object detection.

Training Procedure

Same training procedure as outlined in the paper, SpaRTAN, is used to train this model.

Evaluation Result

Model	Resolution	Params (M)	FLOPs (G)	Top-1 (%)	top-5 (%)
SpaRTAN-S	224x224	18.51	3.86	82.35	96.14

Implementation

Please refer to this repo for full implementation.

Citation

@inproceedings{
    title={SpaRTAN: Spatial Reinforcement Token-based Aggregation Network for Visual Recognition},
    author={Pay, Quan Bi and Baskaran, Vishnu Monn and Loo, Junn Yong and Wong, KokSheik and See, Simon},
    booktitle={2025 International Joint Conference on Neural Networks (IJCNN)},
    pages={to appear},
    year={2025},
    organization={IEEE},
    note={Accepted}
}