Official implementation of Fair Resource Allocation in Multi-Task Learning.

The performance is evaluated under 3 scenarios:

• Image-level Classification. The CelebA dataset contains 40 tasks.
• Regression. The QM9 dataset contains 11 tasks, which can be downloaded automatically from Pytorch Geometric.
• Dense Prediction. The NYU-v2 dataset contains 3 tasks and the Cityscapes dataset (UPDATE: the small version) contains 2 tasks.

Following Nash-MTL and FAMO, we implement our method with the MTL library.

First, create the virtual environment:

conda create -n mtl python=3.9.7
conda activate mtl
python -m pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113

Then, install the repo:

git clone https://github.com/OptMN-Lab/fairgrad.git
cd fairgrad
python -m pip install -e .

The dataset by default should be put under experiments/EXP_NAME/dataset/ folder where EXP_NAME is chosen from {celeba, cityscapes, nyuv2, quantum_chemistry}. To run the experiment:

cd experiments/EXP_NAME
sh run.sh

Cityscapes, NYU-v2, QM9. Please refer to Table 2,3,8 of our paper for more details.

CelebA. For detailed results of our method, please refer to issue1. For single-task results and other baselines including FAMO, CAGrad, etc, please refer to issue2 for more details.

The experiments are conducted on Meta-World benchmark. To run the experiments on MT10 and MT50 (the instructions below are partly borrowed from CAGrad):

1. Create python3.6 virtual environment.
2. Install the MTRL codebase.
3. Install the Meta-World environment with commit id d9a75c451a15b0ba39d8b7a8b6d18d883b8655d8.
4. Copy the mtrl_files folder to the mtrl folder in the installed mtrl repo, then

cd PATH_TO_MTRL/mtrl_files/ && chmod +x mv.sh && ./mv.sh

5. Follow the run.sh to run the experiments.