[CVPR2021] Domain Consensus Clustering for Universal Domain Adaptation

[Paper]

Prerequisites

To install requirements:

pip install -r requirements.txt

• Python 3.6
• GPU Memory: 10GB
• Pytorch 1.4.0

Getting Started

Download the dataset: Office-31, OfficeHome, VisDA, DomainNet.

Data Folder structure:

Your dataset DIR:
|-Office/domain_adaptation_images
| |-amazon
| |-webcam
| |-dslr
|-OfficeHome
| |-Art
| |-Product
| |-...
|-VisDA
| |-train
| |-validataion
|-DomainNet
| |-clipart
| |-painting
| |-...

You need you modify the data_path in config files, i.e., config.root

Training

Train on one transfer of Office:

CUDA_VISIBLE_DEVICES=0 python office_run.py note=EXP_NAME setting=uda/osda/pda source=amazon target=dslr

To train on six transfers of Office:

CUDA_VISIBLE_DEVICES=0 python office_run.py note=EXP_NAME setting=uda/osda/pda transfer_all=1

Train on OfficeHome:

CUDA_VISIBLE_DEVICES=0 python officehome_run.py note=EXP_NAME setting=uda/osda/pda source=Art target=Product

or

CUDA_VISIBLE_DEVICES=0 python officehome_run.py note=EXP_NAME setting=uda/osda/pda transfer_all=1 

The final results (including the best and the last) will be saved in the ./snapshot/EXP_NAME/result.txt.

Notably, transfer_all wil consumes more shared memory.

Citation

If you find it helpful, please consider citing:

@inproceedings{li2021DCC,
  title={Domain Consensus Clustering for Universal Domain Adaptation},
  author={Li, Guangrui and Kang, Guoliang and Zhu, Yi and Wei, Yunchao and Yang, Yi},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2021}
}