Multi-Task problems are not multi-objective

This is the code for the paper "Multi-Task problems are not multi-objective" in which we show that the commonly used Multi-Fashion-MNIST datasets are not suitable for benchmarking multi-objective methods.

For more details see the paper.

Usage

python multi_objective/main.py --config path/to/config.yaml

Config files can be found in configs.

There is also the option to set options using the command line:

python multi_objective/main.py epochs 100

For reproducing the results of the paper see the jupyter notebooks generate_results. For the HPO see hpo.

Installation

Requirements:

1. Only tested on Ubuntu 20.04.
2. python >= 3.7

Create a venv:

python3 -m venv mtl
source mtl/bin/activate

Clone repository:

git clone https://github.com/ruchtem/moo-mtl.git
cd moo-mtl

Upgrade pip and install requirements:

pip install --upgrade pip
pip install -r requirements.txt

Be patient, this takes a while.

The large number of dependencies is partly due to the baselines, available in this repository as well. If cvxopt or cvxpy give you trouble (e.g. ERROR: Failed building wheel for scs) you can omit them, they are only required for the EPO part of PHN.

Finally install the module in editable mode

pip install -e .

Acknowledgments

I would like to thank Samuel Müller for many helpful discussions and suggestions.

Many thanks also to submitit!