This repository provides the official implementation for "Source-free Domain Adaptation via Avatar Prototype Generation and Adaptation". (IJCAI2021)

Source-free Domain Adaptation via Avatar Prototype Generation and Adaptation To handle source-free domain adaptation task, we propose a Contrastive Prototype Generation and Adaptation (CPGA) method. Specifically, CPGA consists of two stage: (1) Prototype generation: under the guidance of the fixed classifier, a generator is trained to generate avatar feature prototypes via and . (2) Protype adaptation: in each training batch, we use the learned prototype generator to generate one prototype for each class. Based on the generated prototypes and pseudo labels obtained by clustering, we align each pseudo-labeled target feature to the corresponding class prototype by training a domain-invariant feature extractor via , and . Note that the classifier is fixed during the whole training phase.

• Clone this repository:

git clone https://github.com/SCUT-AILab/CPGA.git
cd CPGA

• Install the requirements by runing the following command:

pip install -r requirements.txt

• The .pkl files of data list and its corresponding labels have been put in the directory ./data.

• Please manually download the VisDA-C benchmark from the official websites and put it in the directory ./dataset.

• First, to obtain the pre-trained model on the source domain:

python train_source --gpu 0 --data_root ./dataset/VISDA-C/train --label_file ./data/visda_synthesis_9_1_split.pkl

• Second, to train CPGA on the target domain (please assign a source-trained model path):

python main --gpu 0,1 --max_epoch 1400 --source_model_path ./model_source/synthesis_resnet101_best.pkl --data_path ./dataset/VISDA-C/validation --label_file ./data/visda_real_train.pkl

To test CPGA on the target domain using the trained model (please assign a trained model path)

python test --gpu 0 --model_path ./model_VISDA-C/best.pkl --data_path ./dataset/VISDA-C/validation --label_file ./data/visda_real_train.pkl

Classification accuracies (%) on the large-scale VisDA dataset (ResNet-101).

We have released the code in the directory ./visualization.

If you find our work useful in your research, please cite the following paper:

@inproceedings{Qiu2021CPGA,
  title={Source-free Domain Adaptation via Avatar Prototype Generation and Adaptation},
  author={Zhen Qiu and Yifan Zhang and Hongbin Lin and Shuaicheng Niu and Yanxia Liu and Qing Du and Mingkui Tan},
  booktitle={International Joint Conference on Artificial Intelligence},
  year={2021}
}

For any question, please file an issue or contact

Zhen Qiu: [email protected]
Hongbin Lin: [email protected]