Skip to content

Latest commit

 

History

History
110 lines (77 loc) · 3.19 KB

README.md

File metadata and controls

110 lines (77 loc) · 3.19 KB

Neural Prompt Search

S-Lab, Nanyang Technological University

TL;DR

The idea is simple: we view existing parameter-efficient tuning modules, including Adapter, LoRA and VPT, as prompt modules and propose to search the optimal configuration via neural architecture search. Our approach is named NOAH (Neural prOmpt seArcH).


[arXiv][project page]

Updatas

[05/2022] arXiv paper has been released.

Environment Setup

conda create -n NOAH python=3.8
conda activate NOAH
pip install -r requirements.txt

Data Preparation

1. Visual Task Adaptation Benchmark (VTAB)

cd data/vtab-source
python get_vtab1k.py

2. Few-Shot and Domain Generation

  • Images

    Please refer to DATASETS.md to download the datasets.

  • Train/Val/Test splits

    Please refer to files under data/XXX/XXX/annotations for the detail information.

Quick Start For NOAH

We use the VTAB experiments as examples.

1. Downloading the Pre-trained Model

Model Link
ViT B/16 link

2. Supernet Training

sh configs/NOAH/VTAB/supernet/slurm_train_vtab.sh PATH-TO-YOUR-PRETRAINED-MODEL

3. Subnet Search

sh configs/NOAH/VTAB/search/slurm_search_vtab.sh PARAMETERS-LIMITES

4. Subnet Retraining

sh configs/NOAH/VTAB/subnet/slurm_retrain_vtab.sh PATH-TO-YOUR-PRETRAINED-MODEL

We add the optimal subnet architecture of each dataset in the experiments/NOAH/subnet/VTAB.

5. Performance

fig1

Citation

If you use this code in your research, please kindly cite this work.

@misc{zhang2022neural,
      title={Neural Prompt Search}, 
      author={Yuanhan Zhang and Kaiyang Zhou and Ziwei Liu},
      year={2022},
      eprint={2206.04673},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknoledgments

Part of the code is borrowed from CoOp, AutoFormer, timm and mmcv.

Thanks to Chong Zhou (https://chongzhou96.github.io/) for the code of downloading the VTAB-1k.

Hits