Bi-ViT: Pushing the Limit of Vision Transformer Quantization
Yanjing Li, Sheng Xu, Mingbao Lin, Xianbin Cao, Chuanjian Liu, Xiao Sun, Baochang Zhang
Abstract
Vision transformers (ViTs) quantization offers a promising prospect to facilitate deploying large pre-trained networks on resource-limited devices. Fully-binarized ViTs (Bi-ViT) that pushes the quantization of ViTs to its limit remain largely unexplored and a very challenging task yet, due to their unacceptable performance. Through extensive empirical analyses, we identify the severe drop in ViT binarization is caused by attention distortion in self-attention, which technically stems from the gradient vanishing and ranking disorder. To address these issues, we first introduce a learnable scaling factor to reactivate the vanished gradients and illustrate its effectiveness through theoretical and experimental analyses. We then propose a ranking-aware distillation method to rectify the disordered ranking in a teacher-student framework. Bi-ViT achieves significant improvements over popular DeiT and Swin backbones in terms of Top-1 accuracy and FLOPs. For example, with DeiT-Tiny and Swin-Tiny, our method significantly outperforms baselines by 22.1% and 21.4% respectively, while 61.5x and 56.1x theoretical acceleration in terms of FLOPs compared with real-valued counterparts on ImageNet. Our codes and models are attached on https://github.com/YanjingLi0202/Bi-ViT/ .
BibTeX
@article{Li_Xu_Lin_Cao_Liu_Sun_Zhang_2024, title={Bi-ViT: Pushing the Limit of Vision Transformer Quantization}, volume={38}, url={https://ojs.aaai.org/index.php/AAAI/article/view/28109}, DOI={10.1609/aaai.v38i4.28109}, abstractNote={Vision transformers (ViTs) quantization offers a promising prospect to facilitate deploying large pre-trained networks on resource-limited devices. Fully-binarized ViTs (Bi-ViT) that pushes the quantization of ViTs to its limit remain largely unexplored and a very challenging task yet, due to their unacceptable performance. Through extensive empirical analyses, we identify the severe drop in ViT binarization is caused by attention distortion in self-attention, which technically stems from the gradient vanishing and ranking disorder. To address these issues, we first introduce a learnable scaling factor to reactivate the vanished gradients and illustrate its effectiveness through theoretical and experimental analyses. We then propose a ranking-aware distillation method to rectify the disordered ranking in a teacher-student framework. Bi-ViT achieves significant improvements over popular DeiT and Swin backbones in terms of Top-1 accuracy and FLOPs. For example, with DeiT-Tiny and Swin-Tiny, our method significantly outperforms baselines by 22.1% and 21.4% respectively, while 61.5x and 56.1x theoretical acceleration in terms of FLOPs compared with real-valued counterparts on ImageNet. Our codes and models are attached on https://github.com/YanjingLi0202/Bi-ViT/ .}, number={4}, journal={Proceedings of the AAAI Conference on Artificial Intelligence}, author={Li, Yanjing and Xu, Sheng and Lin, Mingbao and Cao, Xianbin and Liu, Chuanjian and Sun, Xiao and Zhang, Baochang}, year={2024}, month={Mar.}, pages={3243-3251} }