This repository provides code for the paper Differentiation of Blackbox Combinatorial Solvers.

By Marin Vlastelica*, Anselm Paulus*, Vít Musil, Georg Martius and Michal Rolínek.

Autonomous Learning Group, Max Planck Institute for Intelligent Systems.

For a condensed version containing only the modules (along with additional solvers) see this repository.

0. Introduction
1. Installation
2. Usage
3. Notes

This repository provides a visualization for all the datasets used in Differentiation of Blackbox Combinatorial Solvers. Additionally, the training code for the Warcraft Shortest Path experiment is provided.

Disclaimer: This code is a PROTOTYPE. It should work fine but use at your own risk.

First install the requirements with via one of the following options:

• Option 1 (requires pipenv):

  pipenv install (use --skip-lock flag for speedup at your own risk)

  pipenv shell

  If the installation causes problems, the following could be a fix: sudo apt install python3.6-dev

• Option 2 (requires python 3.6):

  pip3 install -r requirements.txt

Next, download our datasets and extract each dataset to the data/ directory in this repository.

• Dataset visualization:

  For an easy overview over the datasets use the data/data_visualization.ipynb jupyter notebook.

• Warcraft shortest path experiment:

  Change directory to project root folder.

  Run Warcraft shortest path experiment with gradients through Dijkstra:

  python main.py settings/warcraft_shortest_path/12x12_combresnet.json

  Run Warcraft shortest path baseline ResNet18 experiment:

  python main.py settings/warcraft_shortest_path/12x12_baseline.json

  The results are stored in the results directory in the project root folder.

Contribute: If you spot a bug or some incompatibility, raise an issue or contribute via a pull request! Thank you!