COSMOS - Efficient Multi-Objective Optimization for Deep Learning

This is the official implementation for COSMOS: a method to learn Pareto fronts that scales to large datasets and deep models.

For details see paper.

Usage

1. Download the dataset as described in readme.md in the respective data folder.
2. Run the code:

python multi_objective/main.py --dataset mm --method cosmos

For the logs and results see newly created folder results.

For the settings see settings.py

---

Available datasets:

command-line option	Description
-d mm	Multi-MNIST dataset
-d mf	Multi-Fashion dataset
-d mfm	Multi-Fashion+MNIST dataset
-d adult	Adult dataset
-d compass	Compass dataset
-d default	Default dataset
-d celeba	CelebA dataset

---

Available algorithms:

command-line option	Description
-m cosmos	COSMOS algorithm
-m hyper_ln	PHN (Linear Scalarization) algorithm [1]
-m hyper_epo	PHN (EPO) algorithm [1]
-m pmtl	ParetoMTL algorithm [2]
-m single_task	Treat each objective as single task
-m uniform	Uniform scaling of all objectives

[1] Navon, A., Shamsian, A., Chechik, G. and Fetaya, E., 2020. Learning the Pareto Front with Hypernetworks. arXiv preprint arXiv:2010.04104.

[2] Lin, X., Zhen, H.L., Li, Z., Zhang, Q. and Kwong, S., 2019. Pareto multi-task learning. arXiv preprint arXiv:1912.12854.

Installation

Requirements:

1. CUDA capable GPU (I will port it for CPU usage in future)
2. Only tested on Ubuntu 20.04

Create a venv:

python3 -m venv cosmos
source cosmos/bin/activate

Clone repository:

git clone ...
cd moo

Install requirements:

pip install -r requirements.txt

The large number of dependencies is partly due to the baselines, available in this repository as well.

Acknowledgments

To be done.