DDA

Paper

Code release for "Dynamic Domain Adaptation for Efficient Inference" (CVPR2021)

Our work proposes a Dynamic Domain Adaptation (DDA) framework to solve the problem of efficient inference in the context of domain adaptation.

Dependencies

The code runs with Python3 and requires Pytorch of version 1.3.1 or higher. Please pip install the following packages:

numpy
torch
heaq
math
random
datetime

Pre-trained models

Pre-trained models for backbone MSDNet can be downloaded here and change the --pretrain_path argument.

Training

VisDA 2017 dataset can be found here in the classification track.

Run the following command in shell:

visda-2017 anytime

python train_dda.py --gpu_id id --dset visda --s_dset_path ../data/visda-2017/train_list.txt --t_dset_path ../data/visda-2017/validation_list.txt --test_dset_path ../data/visda-2017/validation_list.txt --pattern anytime

visda-2017 budgeted batch

python train_dda.py --gpu_id id --dset visda --s_dset_path ../data/visda-2017/train_list.txt --t_dset_path ../data/visda-2017/validation_list.txt --test_dset_path ../data/visda-2017/validation_list.txt --pattern budget

Change --base 4 --step 4 to --base 7 --step 7 to run DDA(step-7), and change the pretrained model path.
See train_dda.py for details.

DomainNet dataset can be found here

Run the following command in shell:

DomainNet anytime

python train_dda.py --gpu_id id --dset domainnet --s_dset_path ../data/domainnet/clipart_train.txt --t_dset_path ../data/domainnet/infograph_train.txt --test_dset_path ../data/domainnet/infograph_test.txt --pattern anytime

DomainNet budgeted batch

python train_dda.py --gpu_id id --dset domainnet --s_dset_path ../data/domainnet/clipart_train.txt --t_dset_path ../data/domainnet/infograph_train.txt --test_dset_path ../data/domainnet/infograph_test.txt --pattern budget

Same options are available as in visda-2017.

Acknowledgements

Some codes in this project are adapted from CDAN and MSDNet. We thank them for their excellent projects.

Citation

If you find this code useful for your research, please cite our paper:

@inproceedings{li2019dynamic,
author = {Shuang Li and Jinming Zhang and Wenxuan Ma and Chi Harold Liu and Wei Li},
title = {Dynamic Domain Adaptation for Efficient Inference},
year = {2021},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
numpages = {9}
}

Contact

If you have any problem about our code, feel free to contact

or describe your problem in Issues.

yuntaodu / DDA

DDA