Deep Learning Reproducibility Project

By Kaan Yilmaz, Beyza Hizli and Manisha Sethia

Blog post:

https://www.notion.so/Combinatorial-Optimization-with-Graph-Convolutional-Neural-Networks-88565eafc8834c7190330b0b3cc1083a#9569125c63c44077a9dd041334c4cdfa

Introduction

This repository contains the source code we used to reproduce GCNN model for setcover instances, extra hyperparameter search and an evaluation of an additional dataset (mik) of the paper:

Exact Combinatorial Optimization with Graph Convolutional Neural Networks by Gasse et al.

Requirements

• Miniconda (https://docs.conda.io/en/latest/miniconda.html)
• (Optional) Data folder with some setcover samples (~155 MB) (https://drive.google.com/open?id=1c7-kEHPj4eoZUXwjzDOcCS9C_EIVz3gT)
• Data folder with all setcover samples (~11 GB) (https://drive.google.com/open?id=1Gd09zytoGK_vhqLpPIUGiHccjRnA1IbB)
• Extra data folder with all mik samples (~3 GB) (https://drive.google.com/open?id=1KZeRtykYhHDUpVW4N1F-Nq_BsqwXX0cJ)

Installation

• Download and install miniconda
• Add miniconda3/bin to your PATH
  • Verify with the following command in your terminal: conda -V
• Run the following command in your terminal from the root of this project:

conda env create -f environment.yml

• Download and extract the data archive in the root of the project
  • Verify that the data folder exists
  • The folder structure for setcover should be data/samples/setcover/500r_1000c_0.05d/{train|test|valid}/*.pkl
  • The folder structure for mik should be data/samples/mik/{train|test|valid}/*.pkl

Training

We have provided the trained models ourselves, see the folder trained_models. If you want to train it anyway, then please rename this folder to something else.

Original Setcover instances

Run the following command in the terminal from the root of this project:

source activate deep-learning-project
python train_gcnn.py --seed 0
python train_gcnn.py --seed 1
python train_gcnn.py --seed 2
python train_gcnn.py --seed 3
python train_gcnn.py --seed 4

Extra hyperparameter search

Run the following command in the terminal from the root of this project:

source activate deep-learning-project
python train_gcnn.py --lr 0.0001  
python train_gcnn.py --lr 0.01  
python train_gcnn.py --optimizer RMSprop

Additional dataset

Run the following command in the terminal from the root of this project:

source activate deep-learning-project
python train_gcnn.py --problem mik --samples_path data/samples/mik

Results

The log and trained model parameters are stored in:

trained_models/{problem}/baseline/{seed}/{lr-high|lr-low|lr-normal}/{optimizer}/

Testing

Original Setcover instances

Run the following command in the terminal from the root of this project:

source activate deep-learning-project
python test.py

Extra hyperparameter search

Run the following command in the terminal from the root of this project:

source activate deep-learning-project
python test.py --lr 0.0001  
python test.py --lr 0.01  
python test.py --optimizer RMSprop

Additional dataset

source activate deep-learning-project
python test.py --problem mik --samples_path data/samples/mik

Results

The results are stored in the folder results/{problem}_test_{date}.csv

Help: running out of (GPU) memory

Try to reduce the value of valid_batch_size in config.json