pipeline_tag: image-classification
license: mit
Breaking the Low-Rank Dilemma of Linear Attention: RAVLT Model Card
This model card describes the Rank-Augmented Vision Linear Transformer (RAVLT), introduced in the paper "Breaking the Low-Rank Dilemma of Linear Attention". RAVLT achieves state-of-the-art performance on ImageNet-1k classification while maintaining linear complexity.
Key Features:
- High accuracy: Achieves 84.4% Top-1 accuracy on ImageNet-1k (RAVLT-S).
- Parameter efficiency: Uses only 26M parameters (RAVLT-S).
- Computational efficiency: Achieves 4.6G FLOPs (RAVLT-S).
- Linear complexity.
RAVLT is based on Rank-Augmented Linear Attention (RALA), a novel attention mechanism that addresses the low-rank limitations of standard linear attention.
Model Variants
Several RAVLT variants were trained, offering different tradeoffs between accuracy, parameters, and FLOPs:
| Model | Params (M) | FLOPs (G) | Checkpoint |
|---|---|---|---|
| RAVLT-T | 15 | 2.4 | RAVLT-T |
| RAVLT-S | 26 | 4.6 | RAVLT-S |
| RAVLT-B | 48 | 9.9 | RAVLT-B |
| RAVLT-L | 95 | 16.0 | RAVLT-L |
Note: Accuracy values from the paper have not been transcribed, but should be added once the code is available and the accuracy can be independently verified.
How to use (Placeholder - Awaiting Code Release)
Instructions on how to use the model will be provided once the code repository is available. Code will be available at https://github.com/qhfan/RALA.
Citation
@misc{fan2024breakinglowrank,
title={Breaking the Low-Rank Dilemma of Linear Attention},
author={Fan, Qinghao and Liu, Zheng and Li, Hongsheng and Yang, Yisen and Li, Hang},
year={2024},
eprint={2411.07635},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2411.07635},
}