xudonmao / VMT

VMT

Code for Virtual Mixup Training for Unsupervised Domain Adaptation.

Acknowledgments

This code is based on dirt-t.

Dependencies

numpy==1.14.1
scikit_image==0.13.1
scipy==1.0.0
tensorflow_gpu==1.6.0
tensorbayes==0.4.0

Train

1. Run VMT

python -u run_dirtt.py --datadir data --src mnist --trg svhn --inorm 1 --run 0 --dirt 0 --dw 0.01 --svw 1 --tvw 0.06 --tcw 0.06 --smw 1 --tmw 0.06

2. Run DIRT-T

python -u run_dirtt.py --datadir data --src mnist --trg svhn --inorm 1 --run 0 --dirt 5000 --dw 0.01 --svw 1 --tvw 0.06 --tcw 0.06 --smw 1 --tmw 0.06

Citation

If you use this work in your research, please cite:

@article{arxiv1905.04215,
  author = {Xudong Mao and Yun Ma and Zhenguo Yang and Yangbin Chen and Qing Li},
  title = {Virtual Mixup Training for Unsupervised Domain Adaptation},
  journal = {arXiv preprint arXiv:1905.04215},
  year = {2019}
}