Skip to content

Latest commit

 

History

History
53 lines (31 loc) · 2.42 KB

README.md

File metadata and controls

53 lines (31 loc) · 2.42 KB

PYRA

The official implementation of our ECCV 2024 publication, PYRA (Parallel Yielding Re-Activation).

News

  • 2024-08-19: We have released our source code for PYRA!

  • 2024-07-17: The arXiv version of our ECCV final submission is now released!

  • 2024-07-03: The code coming soon. We promise that it will be available before the main conference date.

Preparation

Datasets

We use the VTAB-1k dataset to evaluate our proposed PYRA. Use instructions in directory data/vtab-source to build VTAB-1k dataset locally (Internet access is demanded).

Environment

We use Anaconda or Miniconda to maintain the fine-tuning environment of PYRA. Simply follow the following instructions to prepare the environment:

conda create -n PYRA python=3.8
conda activate PYRA
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
conda install timm==0.5.4 jupyter
pip install scikit-image ptflops easydict PyYAML pillow opencv-python scipy mmcv==1.7.1 -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install -U fvcore

You can use other mirrors beside tuna.tsinghua.edu.cn as long as everything is installed successfully.

Checkpoints

Use this link to download all pre-trained model weights used for task adaptation in PYRA. After downloading all model checkpoints, unzip them to the weight/ directory under the main directory (PYRA/).

Training & Evaluation

We provide training scripts for the experiments reported in the article. In the default pipeline, evaluation is executed in between training epochs. All training details are saved in the logs/ directory. You can browse the results and training details in the folders of corresponding experiments.

To conduct fine-tuning, simply run the scripts under the scripts/ directory.

WARNING: Simply training PYRA with prompt tuning leads to problems, as prompt tokens might be merged!

FAQs

If you have any questions, please submit the issues describing your question as detailed as possible. If you don't see our reply in a fairly long time, please email me at: xiongyizhe2001@163.com.

Acknowledgments

This codebase is built upon NOAH, ToMe, and timm.

Many thanks to their great work!