source-free-domain-adaptation

This is an source-free domain adaptation repository based on PyTorch. It was developed by Wenxin Su. If you encounter any issues or have questions, please don't hesitate to contact Wenxin at [email protected]. It is also the official repository for the following works:

• [ARXIV'24]Unified Source-Free Domain Adaptation (LCFD)
• [CVPR'24]Source-Free Domain Adaptation with Frozen Multimodal Foundation Model (DIFO)
• [IJCV'23]Source-Free Domain Adaptation via Target Prediction Distribution Searching (TPDS)
• [NN'22]Semantic consistency learning on manifold for source data-free unsupervised domain adaptation (SCLM)
• [IROS'21]Model Adaptation through Hypothesis Transfer with Gradual Knowledge Distillation (GKD)

This repository is also supports the following methods:

• Source, SHOT, NRC, COWA, AdaContrast, PLUE

We encourage contributions! Pull requests to add methods are very welcome and appreciated.

Our Publications

• [ARXIV'24]Proxy Denoising for Source-Free Domain Adaptation (ProDe) and the code will release soon.

• [ARXIV'24]Unified Source-Free Domain Adaptation (LCFD), and Code

• [CVPR'24]Source-Free Domain Adaptation with Frozen Multimodal Foundation Model, Code, and Chinese version

• [IJCV'23]Source-Free Domain Adaptation via Target Prediction Distribution Searching and Code

• [TMM'23]Progressive Source-Aware Transformer for Generalized Source-Free Domain Adaptation and Code

• [CAAI TRIT'22]Model adaptation via credible local context representation and Code

• [NN'22]Semantic consistency learning on manifold for source data-free unsupervised domain adaptation and Code

• [IROS'21]Model Adaptation through Hypothesis Transfer with Gradual Knowledge Distillation and Code

Preliminary

To use the repository, we provide a conda environment.

conda update conda
conda env create -f environment.yml
conda activate sfa

• Datasets
  • office-31 Office-31
  • office-home Office-Home
  • VISDA-C VISDA-C
  • domainnet126 DomainNet (cleaned)
  • imagenet_a ImageNet-A
  • imagenet_r ImageNet-R
  • imagenet_v2 ImageNet-V2
  • imagenet_k ImageNet-Sketch

You need to download the above dataset,modify the path of images in each '.txt' under the folder './data/'.In addition, class name files for each dataset also under the folder './data/'.The prepared directory would look like:

├── data
    ├── office-home
        ├── amazon_list.txt
        ├── classname.txt
        ├── dslr_list.txt
        ├── webcam_list.txt
    ├── office-home
        ├── Art_list.txt
        ├── classname.txt
        ├── Clipart_list.txt
        ├── Product_list.txt
        ├── RealWorld_list.txt
    ...  ...

For the ImageNet variations, modify the ${DATA_DIR} in the conf.py to your data directory where stores the ImageNet variations datasets.

Training

We provide config files for experiments.

Source

• For office-31, office-home and VISDA-C, there is an example to training a source model :

CUDA_VISIBLE_DEVICES=0 python image_target_of_oh_vs.py --cfg "cfgs/office-home/source.yaml" SETTING.S 0

• For domainnet126, we follow AdaContrast to train the source model.

• For adapting to ImageNet variations, all pre-trained models available in Torchvision or timm can be used.

• We also provide the pre-trained source models which can be downloaded from here.

Target

After obtaining the source models, modify the ${CKPT_DIR} in the conf.py to your source model directory. For office-31, office-home and VISDA-C, simply run the following Python file with the corresponding config file to execute source-free domain adaptation.

CUDA_VISIBLE_DEVICES=0 python image_target_of_oh_vs.py --cfg "cfgs/office-home/difo.yaml" SETTING.S 0 SETTING.T 1

For domainnet126 and ImageNet variations.

CUDA_VISIBLE_DEVICES=0 python image_target_in_126.py --cfg "cfgs/domainnet126/difo.yaml" SETTING.S 0 SETTING.T 1

Acknowledgements

• SHOT official
• NRC official
• COWA official
• AdaContrast official
• PLUE official
• CoOp official
• RMT official