mvp18 / ucdr

[ICCV 2021] PyTorch implementation of "Universal Cross-Domain Retrieval: Generalizing across Classes and Domains"

UCDR

Universal Cross-Domain Retrieval: Generalizing across Classes and Domains | ICCV 2021

Requirements and Setup

Python - 3.7.6, PyTorch - 1.1.0, CUDA - 9.0, cuDNN - 7.5.1, NVIDIA Driver Version >= 384.13

conda create --name torch11 --file requirements.txt
conda activate torch11
conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch

Download datasets

Check downloads folder for scripts. Change path_dataset in download_sketchy.sh.

Pretrained Models

Download from here.

Reproducing our Results

Check reproduce_runs.sh per folder in src/algos.

TODOs

• Add trainer and test files for SAKE, Doodle2Search, Base NW, EISNet, CuMix

🎓 Cite

If this code was helpful for your research, consider citing:

@article{paul2021universal,
  title={Universal Cross-Domain Retrieval: Generalizing Across Classes and Domains},
  author={Paul, Soumava and Dutta, Titir and Biswas, Soma},
  journal={arXiv preprint arXiv:2108.08356},
  year={2021}
}

🙏 Acknowledgements

Parts of our code have been borrowed from the following repositories:

• SEM-PCYC, CVPR'19
• Doodle to Search, CVPR'19
• SAKE, ICCV'19
• CuMix, ECCV'20
• EISNet, ECCV'20

ucdr's Issues

About args.trainvalid

Thank you for your excellent work.
I notice that args.trainvalid exists in the domainnet.py, sketchy_extended.py and other datasets' script, but this argument does't exist in options_xxx.py.
Should I remove this argument, or set it True?