MOML

This repository contains the source code of our papers "Multi-Objective Meta Learning" (NeurIPS 2021) and "Multi-Objective Meta-Learning" (AIJ 2024).

Environment

Python 3.7.10
torch 1.8.0+cu111
torchvision 0.9.0+cu111

Dataset

Office-31: Download from https://www.cc.gatech.edu/~jhoffman/domainadapt/#datasets_code
Office-Home: Download from https://www.hemanthdv.org/officeHomeDataset.html
NYUv2: Download the pre-processed data from https://github.com/lorenmt/mtan#image-to-image-predictions-one-to-many

Experiments

MTL

Training on the NYUv2 dataset with the MOML method, you can run the code below (default option is training without data augmentation)

cd ./MTL
python moml_nyu.py --gpu_id [GPU_ID] --model [DMTL, MTAN] --MGDA --dataset_path [ROOT]

Training on the Office-31 or Office-Home dataset with the MOML method, you can run the code below

cd ./MTL
python moml_office.py --gpu_id [GPU_ID] --model [DMTL, MTAN] --dataset [office-31, office-home] --batchsize 64 --MGDA --dataroot [ROOT]

SSDA

Training on the Office-31 dataset with the MOML+MME method, you can run the code below

cd ./SSDA
python moml_MME.py --gpu_id [GPU_ID] --source [SOURCE] --taeget [TARGET] --MGDA

Citation

If you found this code/work to be useful in your own research, please consider citing the following:

@inproceedings{ye2021moml,
  title={Multi-Objective Meta Learning},
  author={Ye, Feiyang and Lin, Baijiong and Yue, Zhixiong and Guo, Pengxin and Xiao, Qiao and Zhang, Yu},
  booktitle={Conference on Neural Information Processing Systems},
  year={2021}
}

@article{ye2024moml,
  title={Multi-Objective Meta-Learning},
  author={Ye, Feiyang and Lin, Baijiong and Yue, Zhixiong and Zhang, Yu and Tsang, Ivor},
  journal={Artificial Intelligence},
  volume={335},
  pages={104184},
  year={2024}
}

Acknowledgement

Thanks for the public code base https://github.com/lorenmt/mtan, https://github.com/VisionLearningGroup/SSDA_MME, and https://github.com/isl-org/MultiObjectiveOptimization.

Contact

If you have any questions, please contact bj.lin.email@gmail.com.

Baijiong-Lin/MOML