Official implementation for DINE

DINE: Domain Adaptation from Single and Multiple Black-box Predictors

Problem setting (Black-box UDA):

Our framework (Distill-and-Finetune):

Prerequisites:

• python == 3.6.8
• cudatoolkit == 11.0
• pytorch ==1.7.1
• torchvision == 0.8.2
• numpy, scipy, sklearn, PIL, argparse, tqdm

Demo:

• Configure the pytorch environment

• Download the dataset and write the correspinding text files via https://github.com/tim-learn/Generate_list (check the file path in the 'data/office-home/**_list.txt' folder)

• Run the following code (reproduce the results for ResNet-50 (source backbone) -> ResNet-50 (target backbone) in upper Table 2)

  • training the source model first

    python DINE_dist.py --gpu_id 0 --seed 2021 --output_src ./ckps/src --dset office-home --s 0 --da uda --net_src resnet50 --max_epoch 50

  • the first step (Distill)

    python DINE_dist.py --gpu_id 0 --seed 2021 --output_src ./ckps/src --dset office-home --s 0 --da uda --net_src resnet50 --max_epoch 30 --net resnet50 --output ./ckps/tar --distill --topk 1

  • the second step (Finetune)

    python DINE_ft.py --gpu_id 0 --seed 2021 --dset office-home --s 0 --da uda --net_src resnet50 --max_epoch 30 --net resnet50 --lr 1e-2 --output ./ckps/tar

Results:

Citation

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

@inproceedings{liang2022dine,
    title={DINE: Domain Adaptation from Single and Multiple Black-box Predictors},
    author={Liang, Jian and Hu, Dapeng and Feng, Jiashi and He, Ran},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    year={2022}
}

Contact

• liangjian92@gmail.com