Awesome

AdaViT: Adaptive Vision Transformers for Efficient Image Recognition

Lingchen Meng*¹, Hengduo Li*², Bor-Chun Chen³, Shiyi Lan², Zuxuan Wu¹, Yu-Gang Jiang¹, Ser-Nam Lim³<br> ¹Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan Univeristy, ²University of Maryland, ³Meta AI<br> * Equal contribution

This repository is an official implementation of the AdaViT.
Our codes are based on the pytorch-image-models and T2T-ViT

Abstract

Built on top of self-attention mechanisms, vision transformers have demonstrated remarkable performance on a variety of tasks recently. While achieving excellent performance, they still require relatively intensive computational cost that scales up drastically as the numbers of patches, self-attention heads and transformer blocks increase. In this paper, we argue that due to the large variations among images, their need for modeling long-range dependencies between patches differ. To this end, we introduce AdaViT, an adaptive computation framework that learns to derive usage policies on which patches, self-attention heads and transformer blocks to use throughout the backbone on a per-input basis, aiming to improve inference efficiency of vision transformers with a minimal drop of accuracy for image recognition. Optimized jointly with a transformer backbone in an end-to-end manner, a light-weight decision network is attached to the backbone to produce decisions on-the-fly. Extensive experiments on ImageNet demonstrate that our method obtains more than $2\times$ improvement on efficiency compared to state-of-the-art vision transformers with only $0.8%$ drop of accuracy, achieving good efficiency/accuracy trade-offs conditioned on different computational budgets. We further conduct quantitative and qualitative analysis on learned usage polices and provide more insights on the redundancy in vision transformers.

Model Zoo

We have put our model checkpoints here.

Eval our model

Download our AdaViT with T2T-ViT-19 from google drive and perform the command below. You can expect to get the Acc about 81.1 with 3.9 GFLOPS.

python3 ada_main.py /path/to/your/imagenet  \
  --model ada_step_t2t_vit_19_lnorm \
  --ada-head --ada-layer --ada-token-with-mlp \
  --flops-dict adavit_ckpt/t2t-19-h-l-tmlp_flops_dict.pth \
  --eval_checkpoint /path/to/your/checkpoint

python3 ada_main.py /path/to/your/imagenet  \
  --model ada_step_deit_small_patch16_224 \
  --ada-head --ada-layer --ada-token-with-mlp \
  --flops-dict adavit_ckpt/deit-s-h-l-tmlp_flops_dict.pth \
  --eval_checkpoint /path/to/your/checkpoint

Citation

@inproceedings{meng2022adavit,
  title={AdaViT: Adaptive Vision Transformers for Efficient Image Recognition},
  author={Meng, Lingchen and Li, Hengduo and Chen, Bor-Chun and Lan, Shiyi and Wu, Zuxuan and Jiang, Yu-Gang and Lim, Ser-Nam},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={12309--12318},
  year={2022}
}