CST

Code release for “Cycle Self-Training for Domain Adaptation” (NeurIPS 2021)

Prerequisites

  • torch>=1.7.0
  • torchvision
  • qpsolvers
  • numpy
  • prettytable
  • tqdm
  • scikit-learn
  • webcolors
  • matplotlib

Training

VisDA-2017

CUDA_VISIBLE_DEVICES=0 python run_cst.py data/visda-2017 -d VisDA2017 -s Synthetic -t Real -a resnet101 \
--epochs 30 --early 12 --lr 0.002 --per-class-eval --temperature 3.0 --center-crop --log logs/cst/VisDA2017 \
--trade-off 0.08 trade-off1 2.0 --trade-off3 0.5 --threshold 0.97 -b 28

Office Home

CUDA_VISIBLE_DEVICES=0 python run_cst.py data/office-home -d OfficeHome -s Pr -t Rw -a resnet50 \
--epochs 30 --early 30 --temperature 2.5 --bottleneck-dim 2048 --log logs/cst/OfficeHome_Pr2Rw \
--trade-off1 2.0 --trade-off3 0.5 --threshold 0.97 --trade-off 0.015

Acknowledgement

This code is implemented based on the Transfer Learning Library, and it is our pleasure to acknowledge their contributions.

The SAM code is adapted from https://github.com/davda54/sam.

Citation

If you use this code for your research, please consider citing:

@article{liu2021cycle,
  title={Cycle Self-Training for Domain Adaptation},
  author={Liu, Hong and Wang, Jianmin and Long, Mingsheng},
  journal={arXiv preprint arXiv:2103.03571},
  year={2021}
}

Contact

If you have any problem about our code, feel free to contact

GitHub

View Github