Efficient Continuous Pareto Exploration in Multi-Task Learning

Pingchuan Ma*, Tao Du*, and Wojciech Matusik

ICML 2020 [Project Page] [Paper] [Video] [Slides]

@inproceedings{ma2020efficient,
    title={Efficient Continuous Pareto Exploration in Multi-Task Learning},
    author={Ma, Pingchuan and Du, Tao and Matusik, Wojciech},
    booktitle={International Conference on Machine Learning},
    pages={6522--6531},
    year={2020},
    organization={PMLR}
}

Quick Start

Online demos for MultiMNIST and UCI-Census are available in Google Colab! Try them now!

Prerequisites

• Ubuntu 16.04 or higher;
• conda 4.8 or higher.

Installation

We will use $ROOT to refer to the root folder where you want to put this project in. We compiled continuous pareto MTL into a package pareto for easier deployment and application.

cd $ROOT
git clone https://github.com/mit-gfx/ContinuousParetoMTL.git
cd ContinuousParetoMTL
conda env create -f environment.yml
conda activate cpmtl
python setup.py install

Example for MultiMNIST

After pareto is installed, we are free to call any primitive functions and classes which are useful for Pareto-related tasks, including continuous Pareto exploration. We provide an example for MultiMNIST dataset, which can be found by:

cd multi_mnist

First, we run weighted sum method for initial Pareto solutions:

python weighted_sum.py

The output should be like:

0: loss [2.313036/2.304537] top@1 [7.65%/10.65%]
0: 1/30: loss [1.463346/0.909529] top@1 [51.52%/69.72%]
0: 2/30: loss [0.889257/0.638646] top@1 [71.29%/78.55%]
0: 3/30: loss [0.703745/0.534612] top@1 [77.77%/81.86%]
0: 4/30: loss [0.622291/0.491764] top@1 [80.13%/83.02%]

Based on these starting solutions, we can run our continuous Pareto exploration by:

python cpmtl.py

The output should be like:

0: 1/10: loss [0.397692/0.350267] top@1 [86.57%/88.11%]
    86.37% 86.57% Δ=0.20% absΔ=0.20%
    88.10% 88.11% Δ=0.01% absΔ=0.01%

0: 2/10: loss [0.392314/0.351280] top@1 [86.85%/88.07%]
    86.37% 86.57% 86.85% Δ=0.28% absΔ=0.48%
    88.10% 88.11% 88.07% Δ=-0.04% absΔ=-0.03%

0: 3/10: loss [0.387585/0.352643] top@1 [86.92%/88.03%]
    86.37% 86.57% 86.85% 86.92% Δ=0.07% absΔ=0.55%
    88.10% 88.11% 88.07% 88.03% Δ=-0.04% absΔ=-0.07%

Now you can play it on your own dataset and network architecture!

Jupyter Notebooks for Submission

Open up a terminal to launch Jupyter:

cd submission
jupyter notebook

You can run the following Jupyter script to reproduce figures in the paper:

fig2.ipynb
fig3.ipynb
fig4.ipynb
fig5_multimnist.ipynb
fig5_uci.ipynb

Contact

If you have any questions about the paper or the codebase, please feel free to contact pcma@csail.mit.edu or taodu@csail.mit.edu.