Official implementation for SHOT++

[TPAMI-2021] Source Data-absent Unsupervised Domain Adaptation through Hypothesis Transfer and Labeling Transfer

Framework:

1. train on the source domain; (Section 3.1)
2. hypothesis transfer with information maximization and self-supervised learning; (Section 3.2 & Section 3.3) (note that SHOT here means results after step 2, which contains an additional rotation-driven self-supervised objective compared with the original SHOT in ICML 2020)

3. labeling transfer with semi-supervised learning. (Section 3.4) (note that SHOT++ has an extra semi-supervised learning step via MixMatch)

Prerequisites:

• python == 3.6.8
• pytorch ==1.1.0
• torchvision == 0.3.0
• numpy, scipy, sklearn, PIL, argparse, tqdm

Dataset:

• Please manually download the datasets Office, Office-Home, VisDA-C, Office-Caltech from the official websites, and modify the path of images in each '.txt' under the folder './object/data/'. [**How to generate such txt files could be found in https://github.com/tim-learn/Generate_list **]

• Concerning the Digits dsatasets, the code will automatically download three digit datasets (i.e., MNIST, USPS, and SVHN) in './digit/data/'.

Training:

1. Unsupervised Closed-set Domain Adaptation (UDA) on the Digits dataset

   • MNIST -> USPS (m2u)

    cd digit/
    python uda_digit.py --gpu_id 0 --seed 2021 --dset m2u --output ckps_digits --cls_par 0.1 --ssl 0.2 
    python digit_mixmatch.py --gpu_id 0 --seed 2021 --dset m2u --output ckps_mm --output_tar ckps_digits --cls_par 0.1 --ssl 0.2 --alpha 0.1

2. Unsupervised Closed-set Domain Adaptation (UDA) on the Office/ Office-Home dataset

   • Train model on the source domain A (s = 0) [--max_epoch 50 for Office-Home]

   cd uda/
   python image_source.py --gpu_id 0 --seed 2021 --trte val --da uda --output ckps/source/ --dset office --max_epoch 100 --s 0

   • Adaptation to other target domains D and W (hypothesis transfer)

   python image_target.py --gpu_id 0 --seed 2021 --da uda --output ckps/target/ --dset office --s 0 --cls_par 0.3 --ssl 0.6

   • Adaptation to other target domains D and W (following labeling transfer) [--max_epoch 50 for Office-Home]

   python image_mixmatch.py --gpu_id 0 --seed 2021 --da uda --dset office --max_epoch 100 --s 0 --output_tar ckps/target/ --output ckps/mixmatch/ --cls_par 0.3 --ssl 0.6 --choice ent --ps 0.0

3. Unsupervised Closed-set Domain Adaptation (UDA) on the VISDA-C dataset

   • Train model on the Synthetic domain [--max_epoch 10 --lr 1e-3]

   cd uda/
   python image_source.py --gpu_id 0 --seed 2021 --trte val --da uda --output ckps/source/ --dset VISDA-C --net resnet101 --lr 1e-3 --max_epoch 10 --s 0

   • Adaptation to the real domain (hypothesis transfer)

   python image_target.py --gpu_id 0 --seed 2021 --da uda --output ckps/target/ --dset VISDA-C --s 0 --net resnet101 --cls_par 0.3 --ssl 0.6

   • Adaptation to the real domain (following labeling transfer)

   python image_mixmatch.py --gpu_id 0 --seed 2021 --da uda --dset VISDA-C --max_epoch 10 --s 0 --output_tar ckps/target/ --output ckps/mixmatch/ --net resnet101 --cls_par 0.3 --ssl 0.6 --choice ent --ps 0.0

4. Unsupervised Partial-set Domain Adaptation (PDA) on the Office-Home dataset

   • Train model on the source domain A (s = 0)

   cd pda/
   python image_source.py --gpu_id 0 --seed 2021 --trte val --da pda --output ckps/source/ --dset office-home --max_epoch 50 --s 0

   • Adaptation to other target domains (hypothesis transfer)

   python image_target.py --gpu_id 0 --seed 2021 --da pda --dset office-home --s 0 --output_src ckps/source/ --output ckps/target/ --cls_par 0.3 --ssl 0.6

   • Adaptation to the real domain (following labeling transfer)

   python image_mixmatch.py --gpu_id 0 --seed 2021 --da pda --dset office-home --max_epoch 50 --s 0 --output_tar ckps/target/ --output ckps/mixmatch/ --cls_par 0.3 --ssl 0.6 --choice ent --ps 0.0

5. Unsupervised Multi-source Domain Adaptation (MSDA) on the Office-Home dataset

   • Train model on the source domains Ar (s = 0), Cl (s = 1), Pr (s = 2), respectively

   cd msda/
   python image_source.py --gpu_id 0 --seed 2021 --trte val --da uda --dset office-home --output ckps/source/ --net resnet50 --max_epoch 50 --s 0
   python image_source.py --gpu_id 0 --seed 2021 --trte val --da uda --dset office-home --output ckps/source/ --net resnet50 --max_epoch 50 --s 1
   python image_source.py --gpu_id 0 --seed 2021 --trte val --da uda --dset office-home --output ckps/source/ --net resnet50 --max_epoch 50 --s 2

   • Adaptation to the target domain (hypothesis transfer)

   python image_target.py --gpu_id 0 --seed 2021 --cls_par 0.3 --ssl 0.6 --da uda --dset office-home --output_src ckps/source/ --output ckps/target/ --net resnet50 --s 0
   python image_target.py --gpu_id 0 --seed 2021 --cls_par 0.3 --ssl 0.6 --da uda --dset office-home --output_src ckps/source/ --output ckps/target/ --net resnet50 --s 1
   python image_target.py --gpu_id 0 --seed 2021 --cls_par 0.3 --ssl 0.6 --da uda --dset office-home --output_src ckps/source/ --output ckps/target/ --net resnet50 --s 2

   • Adaptation to the target domain (labeling transfer)

   python image_mixmatch.py --gpu_id 0 --seed 2021 --da uda --dset office-home --max_epoch 50 --output_tar ckps/target/ --output ckps/mixmatch/ --cls_par 0.3 --ssl 0.6 --choice ent --ps 0.0 --net resnet50 --s 0
   python image_mixmatch.py --gpu_id 0 --seed 2021 --da uda --dset office-home --max_epoch 50 --output_tar ckps/target/ --output ckps/mixmatch/ --cls_par 0.3 --ssl 0.6 --choice ent --ps 0.0 --net resnet50 --s 1
   python image_mixmatch.py --gpu_id 0 --seed 2021 --da uda --dset office-home --max_epoch 50 --output_tar ckps/target/ --output ckps/mixmatch/ --cls_par 0.3 --ssl 0.6 --choice ent --ps 0.0 --net resnet50 --s 2 

   • Combine domain-spetific scores together

   python image_ms.py --gpu_id 0 --seed 2021 --cls_par 0.3 --ssl 0.6 --da uda --dset office-home --output_src ckps/source/ --output ckps/target/ --output_mm ckps/mixmatch/ --net resnet50 --t 3

6. Semi-supervised Domain Adaptation (SSDA) on the Office-Home dataset

   • Train model on the source domain Ar (s = 0)

   cd ssda/
   python image_source.py --gpu_id 0 --seed 2021 --output ckps/source/ --dset office-home --max_epoch 50 --s 0

   • Adaptation to the target domain Cl (t = 1) [hypothesis transfer]

   python image_target.py --gpu_id 0 --seed 2021 --cls_par 0.1 --ssl 0.2 --output_src ckps/source --output ckps/target --dset office-home --s 0 --t 1 

   • Adaptation to the target domain Cl (t = 1) [labeling transfer]

   python image_mixmatch.py --gpu_id 0 --seed 2021 --ps 0.0 --cls_par 0.1 --ssl 0.2 --output_tar ckps/target --output ckps/mixmatch --dset office-home --max_epoch 50 --s 0 --t 1

Please refer ./xxda/run_xxda.sh for all the settings for different methods and scenarios.

Citation

If you find this code useful for your research, please cite our papers

@article{liang2021source,  
 title={Source Data-absent Unsupervised Domain Adaptation through Hypothesis Transfer and Labeling Transfer}, 
 author={Liang, Jian and Hu, Dapeng and Wang, Yunbo and He, Ran and Feng, Jiashi},   
 journal={IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
 year={2021}, 
 note={In Press}  
}

@inproceedings{liang2020we, 
 title={Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation}, 
 author={Liang, Jian and Hu, Dapeng and Feng, Jiashi}, 
 booktitle={International Conference on Machine Learning (ICML)},  
 pages={6028--6039},
 year={2020}
}

Contact

• liangjian92@gmail.com
• dapeng.hu@u.nus.edu
• elefjia@nus.edu.sg