NeuralSEP: A Neural Separation Algorithm for the Rounded Capacity Inequalities

This repository provides implemented codes for the paper, NeuralSEP.

A Neural Separation Algorithm for the Rounded Capacity Inequalities

https://arxiv.org/abs/2306.17283

Python requirement

Clone project and create an environment with conda:

conda create -n neuralsep python=3.8
conda activate neuralsep

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

Note

If you use the different cuda version, please modify the URL for torch-scatter in requirements.txt before running it; see here.
If you have any trouble with dgl, please refer here.
To run cutting plane embedded NeuralSEP, other installations are required.

Usage

Training

python train_coarsening_model.py

Training data can be downloaded here.

Evaluating with cutting plane methods

cd src/jl
julia experiment_with_random_instances.jl

To run the autoregressive model, the function learned_rounded_capacity_cuts in src/jl/cvrp_cutting.jl needs to be modified.
You can change pre-trained model directories in julia_main.py for each model.

Others

Julia requirement

Julia >= 1.8.3
JuMP
Gurobi
CPLEX
TSPLIB
CVRPLIB
CVRPSEP
Graphs
PyCall
Pickle

Note: to use the created python env, activate the env and re-configure PyCall when installing PyCall in julia.

using Pkg
ENV["PYTHON"] = Sys.which("python")
ENV["PYCALL_JL_RUNTIME_PYTHON"] = Sys.which("python")
Pkg.add("PyCall")
Pkg.build("PyCall")

CPLEX (1.2.10)

A license is required (the student license is free).

Gurobi (10.0.1)