This repo is the official implementation of our NeurIPS2023 paper "Efficient Adaptation of Large Vision Transformer via Adapter Re-Composing" (arXiv).
The advent of high-capacity pre-trained models has revolutionized problem-solving in computer vision, shifting the focus from training task-specific models to adapting pre-trained models. Consequently, effectively adapting large pre-trained models to downstream tasks in an efficient manner has become a prominent research area. Existing solutions primarily concentrate on designing lightweight adapters and their interaction with pre-trained models, with the goal of minimizing the number of parameters requiring updates. In this study, we propose a novel Adapter Re-Composing (ARC) strategy that addresses efficient pre-trained model adaptation from a fresh perspective. Our approach considers the reusability of adaptation parameters and introduces a parameter-sharing scheme. Specifically, we leverage symmetric down-/up-projections to construct bottleneck operations, which are shared across layers. By learning low-dimensional re-scaling coefficients, we can effectively re-compose layer-adaptive adapters. This parameter-sharing strategy in adapter design allows us to significantly reduce the number of new parameters while maintaining satisfactory performance, thereby offering a promising approach to compress the adaptation cost. We conduct experiments on 24 downstream image classification tasks using various Vision Transformer variants to evaluate our method. The results demonstrate that our approach achieves compelling transfer learning performance with a reduced parameter count.
- FGVC & vtab-1k
You can follow VPT to download them.
Since the original vtab dataset is processed with tensorflow scripts and the processing of some datasets is tricky, we also upload the extracted vtab-1k dataset in onedrive for your convenience. You can download from here and then use them with our vtab.py directly. (Note that the license is in vtab dataset).
- Clone this repo:
git clone https://github.com/DavidYanAnDe/ARC.git
cd ARC- Install requirements:
pip install -r requirements.txt- For pre-trained ViT, Swin-B models on ImageNet-21K. You can also manually download them from ViT,Swin Transformer.
To fine-tune a pre-trained ViT model on vtab, run:
python vtab_ARC_train.pyTo fine-tune a pre-trained ViT model on FGVC, run:
python FGVC_ARC_train.pyTo specify the ARC or ARC_att, change the model setting at:
parser.add_argument("--tuning_mode", choices=["ARC_att", "ARC"], default= "ARC", help="tuning mode,can be ARC_att or ARC")If this project is helpful for you, you can cite our paper:
The code is built upon timm. The processing of the vtab-1k dataset refers to vpt, vtab github repo, and NOAH.