To reproduce the experiments, please first download the official release of the dataset files into the specific folders. Please make sure the folder structure keeps the same with configure.py. If you want to use different folders, please modify the configure.py and dloader.py to point the path accordingly.
cd Exp_digits
python3 train_digits.py --re_weighting 1 --initial_lr 1e-3 --train_samples 3000 # You can change number of training samples as you wish (3000, 5000, or 8000)
cd Exp_vision
python3 train.py --dataset pacs --initial_lr 2e-4 --ratio 0.1 # you can modify the ratio as you wish to run the experiments, you can replace pacs with office or office-home etc.. to run the exps.
cd Exp_vision
python3 train.py --dataset office31 --initial_lr 2e-4 --ratio 0.2 --drift_ratio 0.1 # You can modify the drift_ratio as you wish (we report the results of drift ratio from 0.1 to 0.8)
python3 train.py --dataset office_home --initial_lr 2e-4 --ratio 0.2 --drift_ratio 0.5 # You can modify the drift_ratio as you wish (we report the results of drift ratio from 0.1 to 0.8)
If you think our work is helpful to your research, please consider to cite as,
@article{Zhou_Chaib-draa_Wang_2021,
title={Multi-task Learning by Leveraging the Semantic Information},
volume={35},
url={https://ojs.aaai.org/index.php/AAAI/article/view/17323},
number={12},
journal={Proceedings of the AAAI Conference on Artificial Intelligence},
author={Zhou, Fan and Chaib-draa, Brahim and Wang, Boyu}, year={2021}, month={May}, pages={11088-11096} }