ispc-lab / BMD

[ECCV 2022] The official repository of our paper "BMD: A General Class-balanced Multicentric Dynamic Prototype Strategy for Source-free Domain Adaptation"

BMD: Class-balanced Multicentric Dynamic Prototype Strategy for SFDA [ECCV 2022].

Attention: Our new work on source-free universal domain adaptation has been accepted by CVPR-2023! The paper "Upcycling Models under Domain and Category Shift" is available at https://arxiv.org/abs/2303.07110. The code also has been made public at https://github.com/ispc-lab/GLC.

The official repository of our paper "BMD: A General Class-balanced Multicentric Dynamic Prototype Strategy for Source-free Domain Adaptation". Here, we present the demo implementation on VisDA-C dataset.

Prerequisites

• python3
• pytorch >= 1.7.0
• torchvision
• numpy, scipy, sklearn, PIL, argparse, tqdm, wandb

Step

1. Please first prepare the pytorch enviroment.
2. Please download the VisDA-C dataset from the official website, and then unzip the dataset to the ./data folder.
3. Prepare the source model by running following command

   sh ./scripts/train_soruce.sh

4. Perform the target model adaptation by running following command. Please note that you need to first assign the source model checkpoint path in the ./scripts/train_target.sh script.

   sh ./scripts/train_target.sh

Acknowledgement

This codebase is based on SHOT-ICML2020.

Citation

If you find it helpful, please consider citing:

@inproceedings{sanqing2022BMD,
  title={BMD: A General Class-balanced Multicentric Dynamic Prototype Strategy for Source-free Domain Adaptation},
  author={Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, Dacheng Tao},
  booktitle={European conference on computer vision},
  year={2022}
}