Jo-wang / Cal-SFDA

The official repo of the paper "Cal-SFDA: Source-Free Domain-adaptive Semantic Segmentation with Differentiable Expected Calibration Error", accepted as oral paper in ACM Multimedia 2023.

Cal-SFDA

The official repo of the paper "Cal-SFDA: Source-Free Domain-adaptive Semantic Segmentation with Differentiable Expected Calibration Error", published in ACM Multimedia 2023.

Training & Testing

Train the source-only model

• Train with a pretrained ResNet model:

python so_run.py

where all the configs are in the ./config/so_config.yml file.

Train the value net model

• Train with selected source checkpoint:

python rl_run.py

where all the configs are in the ./config/rl_config.yml file.

Target adaptation

• Train with seg model + value net checkpoint:

python adaptive_target_run.py

where all the configs are in the ./config/adaptive_target_config.yml file.

To use multiple GPU:

Modify your GPU id in xx_run.py file. (replace xx with the corresponding training process)

e.g., os.environ["CUDA_VISIBLE_DEVICES"] = '0,1,2'

some important config in xx_config.yml:

• init_weight: initial weight from ResNet101 pretrained model
• restore_from: restore from a specific epoch address, e.g., epoch1.pth
• snapshot: the address you want to save your checkpoint.
• rl_restore_from: restore the seg model + value net checkpoint.
• plabel: your pseudo label path.

Note: run this experiment requires Weights & Biases to log the performance. Please install it in your own environment: pip install wandb

Acknowledgement

• https://github.com/fumyou13/LDBE

Citation

  @article{DBLP:journals/corr/abs-2308-03003,
  author       = {Zixin Wang and
              Yadan Luo and
              Zhi Chen and
              Sen Wang and
              Zi Huang},
  title        = {Cal-SFDA: Source-Free Domain-adaptive Semantic Segmentation with Differentiable
              Expected Calibration Error},
  journal      = {CoRR},
  volume       = {abs/2308.03003},
  year         = {2023},
  url          = {https://doi.org/10.48550/arXiv.2308.03003},
  doi          = {10.48550/ARXIV.2308.03003},
  eprinttype    = {arXiv},
  eprint       = {2308.03003},
  timestamp    = {Mon, 21 Aug 2023 17:38:10 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2308-03003.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
  }