Consolidator: Mergeable Adapter with Grouped Connections For Visual Adaptation (ICLR-2023)

Official implementation for ICLR 2023 paper Consolidator: Mergeable Adapter with Grouped Connections for Visual Adaptation

TL;DR

To enrich model capacity with limited storage budget, we design a module consolidator enabling two-stage consolidation process: 1. between adaptation and storage and 2. between loading and inference. Such paradigm can reach better performance using quite few parameters, and bring no extra inference cost.

Environment Setup

conda create -n consolidator python=3.8
conda activate consolidator
pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install mmcv==1.3.11 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.8/index.html
pip install -r requirements.txt

Data Preparation

cd data/vtab-source
python get_vtab1k.py

Quick Start

1. Downloading the pre-trained checkpoints

Download ViT-B and MoCo v3 ViT-B to ckpts/.

2. Training & Evaluation

For in21k_vit-b:

# training
bash configs/Consolidator/VTAB/train_consolidator_vtab.sh

# collecting results
python collect_res.py --logs-path saves/vit-b

For mocov3_vit-b:

# training
bash configs/Consolidator/VTAB/mocov3_train_consolidator_vtab.sh

# collecting results
python collect_res.py --logs-path saves/mocov3_vit-b

3. Converting

When you want to convert a consolidator to a normal linear layer, the equivalent weight and bias are:

m = Consolidator(768, 768, fc_groups=(384,), rep_drop=0.5, with_channel_shuffle=True)
weight, bias = m.get_actual_params()

You can verify the result by

import torch
import torch.nn.functional as F
x = torch.randn(64, 197, 768)
m.eval()
out1 = m(x)
out2 = F.linear(x, weight, bias)
print('Difference:')
print(((out2 - out1) ** 2).sum())

4. Main Results

Citation

@inproceedings{
    hao2023consolidator,
    title={Consolidator: Mergable Adapter with Group Connections for Visual Adaptation},
    author={Tianxiang Hao and Hui Chen and Yuchen Guo and Guiguang Ding},
    booktitle={The Eleventh International Conference on Learning Representations },
    year={2023},
    url={https://openreview.net/forum?id=J_Cja7cpgW}
}

Contact

If you have any question, please contact [email protected].

Acknowledgments

The code of this repository is based on NOAH and timm.

Thanks for their great works.