Zhongying-Deng / DIDA

Pytorch implementation for "Dynamic Instance Domain Adaptation" (DIDA-Net, accepted to IEEE T-IP).

DIDA-Net

Pytorch implementation for DIDA-Net ("Zhongying Deng, Kaiyang Zhou, Da Li, Junjun He, Yi-Zhe Song, Tao Xiang. Dynamic Instance Domain Adaptation. arXiv:2203.05028"). This paper has been accepted to IEEE T-IP (the T-IP version).

Installation

1. If you want to reproduce the results in DIDA-Net paper, please first install the Dassl codebase, then copy the files in this repository to Dassl. When asked to overwrite, please say yes (some __init__.py files may be overwritten. These files are modified to include the backbone used for DIDA-Net. Other files may also be different from the latest Dassl codebase, e.g. dassl/optim/optimizer.py is updated to set a different learning rate for new layers, and dassl/data/transforms/transforms.py is from an old version of Dassl codebase which does not use center crop in testing).

   After that, pytorch 1.7.1 + cuda 10.1, python 3.7 should be installed. Note that installing Dassl is a must.

2. If you only want to see the implementatation of DIDA-Net backbone, please directly refer to dassl/modeling/backbone/resnet_dida.py and dassl/modeling/backbone/dida_module.py. You can modify these files to accommodate your own codebase.

Training

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

2. Prepare the datasets (PACS, Digit-Five and DomainNet) as in Dassl. Then run the bash script as

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

3. The experiments on PACS will start running. In train_didanet.sh, $DATA denotes the directory where datasets are located. For example, there should be a folder named pacs under such directory, i.e., /path/to/your/dataset/pacs.

   For the experiments on Digit-Five and DomainNet, specify the source and target domains using --source-domains, --target-domains and modify related config files in train_didanet.sh, e.g., --dataset-config-file configs/datasets/da/digit5.yaml and --config-file configs/trainers/da/didanet/digit5_dida_net.yaml.

   Note that the detailed training settings are under the folder named configs, such as datasets (see configs/datasets/da), lr, optimizer etc. (see configs/trainers/da/didanet)

The most important files are under the folder of dassl:

• Implementation of DIDA-Net can be found in dassl/modeling/backbone/resnet_dida.py (for PACS and DomainNet where ResNet is used as backbone) anddassl/modeling/backbone/cnn_digit5_m3sda_dida.py (for Digit-Five).

The trained DIDA-Net model on Sketch domain of PACS can be found here. This model gives 86.00% on the Sketch domain. You can uncompress the trained model to Dassl root path, then do testing as below.

Test

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

DATA=/root_path/to/your/dataset
CUDA_VISIBLE_DEVICES=0 python tools/train.py --root $DATA --trainer FixMatch \
 --source-domains cartoon art_painting photo --target-domains sketch \
 --dataset-config-file configs/datasets/da/pacs.yaml --config-file configs/trainers/da/didanet/pacs_dida_net.yaml \
 --output-dir output/didanet_pacs/sketch \
 --eval-only \
 --model-dir output/didanet_pacs/sketch \
 --load-epoch 20 \
 MODEL.BACKBONE.NAME resnet18_dida

Citation

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

@ARTICLE{9813442,
  author={Deng, Zhongying and Zhou, Kaiyang and Li, Da and He, Junjun and Song, Yi-Zhe and Xiang, Tao},
  journal={IEEE Transactions on Image Processing}, 
  title={Dynamic Instance Domain Adaptation}, 
  year={2022},
  volume={31},
  number={},
  pages={4585-4597},
  doi={10.1109/TIP.2022.3186531}}

or the arxiv version:

@article{deng2022dynamic,
  title={Dynamic Instance Domain Adaptation},
  author={Deng, Zhongying and Zhou, Kaiyang and Li, Da and He, Junjun and Song, Yi-Zhe and Xiang, Tao},
  journal={arXiv preprint arXiv:2203.05028},
  year={2022}
}