Source-free Domain Adaptation via Avatar Prototype Generation and Adaptation

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

Paper

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.

Getting Started

Installation

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

Data Preparation

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.

Training

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

Testing

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

Results

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

Method	Plane	Bicycle	Bus	Car	Horse	Knife	Mcycl	Person	Plant	Sktbrd	Train	Truck	Per-class	Pre-training
ResNet-101	55.1	53.3	61.9	59.1	80.6	17.9	79.7	31.2	81.0	26.5	73.5	8.5	52.4	-
PrDA	86.9	81.7	84.6	63.9	93.1	91.4	86.6	71.9	84.5	58.2	74.5	42.7	76.6	-
SHOT	92.6	81.1	80.1	58.5	89.7	86.1	81.5	77.8	89.5	84.9	84.3	49.3	79.6	-
MA	94.8	73.4	68.8	74.8	93.1	95.4	88.6	84.7	89.1	84.7	83.5	48.1	81.6	-
BAIT	93.7	83.2	84.5	65.0	92.9	95.4	88.1	80.8	90.0	89.0	84.0	45.3	82.7	-
CPGA (ours)	95.6	89.0	75.4	64.9	91.7	97.5	89.7	83.8	93.9	93.4	87.7	69.0	86.0	Download

Visualization

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

Citation

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}
}

Contact

For any question, please file an issue or contact

Zhen Qiu: seqiuzhen@mail.scut.edu.cn
Hongbin Lin: sehongbinlin@mail.scut.edu.cn

DanJun6737/CPGA