Nash-MTL

Official implementation of "Multi-Task Learning as a Bargaining Game".

Setup environment

conda create -n nashmtl python=3.9.7
conda activate nashmtl
conda install pytorch==1.9.0 torchvision==0.10.0 cudatoolkit=10.2 -c pytorch
conda install pyg -c pyg -c conda-forge

Install the repo:

git clone https://github.com/AvivNavon/nash-mtl.git
cd nash-mtl
pip install -e .

Run experiment

To run experiments:

cd experiment/<experiment name>
python trainer.py --method=nashmtl

Follow instruction on the experiment README file for more information regarding, e.g., datasets.

Here <experiment name> is one of [toy, quantum_chemistry, nyuv2]. You can also replace nashmtl with on of the following MTL methods.

We also support experiment tracking with Weights & Biases with two additional parameters:

python trainer.py --method=nashmtl --wandb_project=<project-name> --wandb_entity=<entity-name>

MTL methods

We support the following MTL methods with a unified API. To run experiment with MTL method X simply run:

python trainer.py --method=X

Method (code name)	Paper (notes)
Nash-MTL (nashmtl)	Multi-Task Learning as a Bargaining Game
CAGrad (cagrad)	Conflict-Averse Gradient Descent for Multi-task Learning
PCGrad (pcgrad)	Gradient Surgery for Multi-Task Learning
IMTL-G (imtl)	Towards Impartial Multi-task Learning
MGDA (mgda)	Multi-Task Learning as Multi-Objective Optimization
DWA (dwa)	End-to-End Multi-Task Learning with Attention
Uncertainty weighting (uw)	Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics
Linear scalarization (ls)	- (equal weighting)
Scale-invariant baseline (scaleinvls)	- (see Nash-MTL paper for details)
Random Loss Weighting (rlw)	A Closer Look at Loss Weighting in Multi-Task Learning

Citation

If you find Nash-MTL to be useful in your own research, please consider citing the following paper:

@article{navon2022multi,
  title={Multi-Task Learning as a Bargaining Game},
  author={Navon, Aviv and Shamsian, Aviv and Achituve, Idan and Maron, Haggai and Kawaguchi, Kenji and Chechik, Gal and Fetaya, Ethan},
  journal={arXiv preprint arXiv:2202.01017},
  year={2022}
}