DAC-Net (BMVC 2021 Oral Presentation)

Pytorch implementation of DAC-Net ("Zhongying Deng, Kaiyang Zhou, Yongxin Yang, Tao Xiang. Domain Attention Consistency for Multi-Source Domain Adaptation. BMVC 2021")

Installation

Please install the Dassl codebase, and then copy the files in this repository to Dassl. When asked to overwrite, say yes (some __init__.py files may be overwritten, these files is modified to include the backbone or trainer used for DAC-Net). Then you will install pytorch 1.7.1 + cuda 10.1, python 3.7.

Training

Create a folder like output/dacnet_pacs (under the Dassl root path) where checkpoint and log can be saved.

Then

bash train_dacnet.sh /path/to/your/dataset

Run the above bash script, then the experiments on PACS will be running. In the script, $DATA denotes the location where datasets are installed. For experiments on Digit-Five and DomainNet, modify the --source-domains, --target-domains and related config files, such as --dataset-config-file configs/datasets/da/digit5_ca.yaml and --config-file configs/trainers/da/dacnet/digit5.yaml.

The detailed training settings are in the folder named configs, such as datasets and backbone name used for DAC-Net (see configs/datasets/da), and lr, optimizer etc. (see configs/trainers/da/dacnet)

Some important files are under the folder of dassl:

Implementation of DAC-Net can be found in dassl/engine/da/dacnet.py;
The backbone CNN model of our DAC-Net can be found in dassl/modeling/backbone/resnet_ca.py (for PACS and DomainNet where ResNet is adopted as backbone) and dassl/modeling/backbone/cnn_digit5_m3sda_ca.py (for Digit-Five);
Config definition for DAC-Net can be found in dassl/config/defaults.py (see last 5 lines);

Trained model on Sketch domain of PACS can be found here. This model gives 84.88% on the Sketch domain.

Test

Similar to train_dacnet.sh, testing can be done like this:

DATA=/root_path/to/your/dataset
CUDA_VISIBLE_DEVICES=0 python tools/train.py --root $DATA --trainer DACNet \
 --source-domains cartoon art_painting photo --target-domains sketch \
 --dataset-config-file configs/datasets/da/pacs_ca.yaml --config-file configs/trainers/da/dacnet/pacs.yaml \
 --output-dir output/dacnet_pacs/sketch \
 --eval-only \
 --model-dir output/dacnet_pacs/sketch \
 --load-epoch 30 \
 MODEL.INIT_HEAD_WEIGHTS output/dacnet_pacs/sketch/classifier/model.pth.tar-30

Citation

If you find this code useful, please consider citing the following paper:

@article{deng2021domain,
  title={Domain Attention Consistency for Multi-Source Domain Adaptation},
  author={Deng, Zhongying and Zhou, Kaiyang and Yang, Yongxin and Xiang, Tao},
  journal={arXiv preprint arXiv:2111.03911},
  year={2021}
}