This repository is the official implementation of "Variational Annealing".

The folder "/VAG-CO" contains code with the VAG-CO implementation and also contains code that solves CO problem instances on various datasets. The folder "/meanfield_annealing" contains code with the Mean Field Approximation implementation.

Within the folder /VAG-CO do the following:

To install requirements:

pip install -r requirements.txt

Get an academic gurobi licence at https://www.gurobi.com/. This is a neccesary step to run prepare_datasets.py, which uses gurobi to obtain optimal results for CO problem instances.

run prepare_datasets.py

e.g.

python prepare_datasets.py --dataset RB_iid_100 --problem MVC

All possible datasets and CO problems are listed within prepare_datasets.py.

After you created the datasets you can train your model with:

e.g.

python PPO_configuration.py --IsingMode RB_iid_100 --temps 0.05 --N_anneal 4000 --GPUs 0 --time_horizon 20 --batch_epochs 2 --AnnealSchedule cosine_frac --num_hidden_neurons 40 --mini_Sb 10 --mini_Nb 10 --mini_Hb 10 --EnergyFunction MVC --n_val_workers 30 --encode_node_features 40 --encode_edge_features 30 --lr 0.0005 --message_nh 30 --seed 123 --n_GNN_layers 3 --n_sample_spins 5 --project_name "myfirstrun"

You can find other parser_args that were used in the paper either in parser_MVC.txt or parser_MIS.txt. The list of all parser_args will be added soon.

You can evaluate on a dataset by running evaluate.py

e.g.

python evaluate.py --GPU 0 --batchsize 2 --Ns 8 --sample_mode OG

This code will run evaluation on the RRG-100 MIS dataset and calculate an average APR. If you want to evaluate on another dataset you will have to change the "overall_path" in evaluate.py to a path that contains the config and model weights of a model that is trained on another dataset.

A pretrained model that is trained on RRG-100 dataset is available. See Evaluation.

see Paper.