README.markdown

Official Repository of the ICML 2024 Paper

A Diffusion Model Framework for Unsupervised Neural Combinatorial Optimization (DiffUCO)

Authors:

Sebastian Sanokowski , Sepp Hochreiter , Sebastian Lehner

and of the ICLR 2025 Paper

Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics (SDDS)

Installing the environment

To install the environment fist run

conda env create -f environment.yml

Continue isntalling all missing packages by following the instructions below:

conda activate DiffUCO
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

Optional step:

TSP (Travelor Salesman Person) results never managed it into the paper but if anyone is interested in that problem type pyconcorde has to be installed. For that follow the instructions on: https://github.com/jvkersch/pyconcorde

Ising Theory Baselines:

To compute the Ising Theory Baselines we provide a simple script in /IsingTheoryBaselines/IsingTheory.py

Getting started

• get started by creating a dataset with the DatasetCreator

cd /DatasetCreator

and run for example

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

For more details read

DatasetCreator/README.md

Running experiments

To run an experiment on the created dataset (see above) do the following:

python argparse_ray_main.py --lrs 0.002 --GPUs 0 --n_GNN_layers 8 --temps 0.6 --IsingMode RB_iid_100 --EnergyFunction MIS --N_anneal 2000 --n_diffusion_steps 3 --batch_size 20 --n_basis_states 10 --noise_potential bernoulli --project_name FirstRuns --seed 123 

parameter explanation

--IsingMode Dataset to train on. In this case the RB_iid_100dataset
--EnergyFunction CO problem to train on. In this case the MISproblem
--noise_potential Noise Distribution that is used during training
--batch_size Number of different CO Instances in each batch
--n_basis_sates Number of sampled Diffusion Trajectories in each CO Instance
--temps Starting temperature for annealing

DiffUCO

You can run experiments as in the DiffUCO paper by setting --train_mode REINFORCE when running python argparse_ray_main.py.

SDDS: rKL w/ RL

You can run DiffUCO with more steps by setting --train_mode PPO when running python argparse_ray_main.py. Then, DiffUCO is combined with RL to reduce memory requirements.
Here, you have to specify the mini-batch size for the inner loops steps within PPO.
When running DiffUCO with --n_diffusion_steps A and --n_basis_states B you have to set --minib_diff_steps X and --minib_basis_states Y so that A/X and B/Y are integers.

Example:

python argparse_ray_main.py --lrs 0.002 --GPUs 0 --n_GNN_layers 8 --temps 0.6 --IsingMode BA_small --EnergyFunction MIS --N_anneal 2000 --n_diffusion_steps 30 --batch_size 20 --n_basis_states 10 --noise_potential bernoulli --project_name FirstRuns --seed 123 --train_mode PPO --minib_diff_steps 10 --minib_basis_states 5

SDDS: fKL w/ MC

Alternatively, you can run DiffUCO with more steps by setting --train_mode Forward_KL when running python argparse_ray_main.py. Example:

python argparse_ray_main.py --lrs 0.002 --GPUs 0 --n_GNN_layers 8 --temps 0.6 --IsingMode BA_small --EnergyFunction MIS --N_anneal 2000 --n_diffusion_steps 30 --batch_size 20 --n_basis_states 10 --noise_potential bernoulli --project_name FirstRuns --seed 123 --train_mode Forward_KL --minib_diff_steps 10 --minib_basis_states 5

To evaluate the model use "ConditionalExpectation.py".

After training, you can evaluate the model on the test set with:

python ConditionalExpectation.py --wandb_id <WANDB_ID> --dataset <DATASET_NAME> --GPU <GPU_ID> --evaluation_factor <EVAL_FACTOR> --n_samples <N_SAMPLES>

In the papers <EVAL_FACTOR> is set to 3, i.e. the model uses 3 times more diffusion steps during evaluation than during training.

parameter explanation

--wandb_id is the wandb run id
--dataset is the dataset that will be used for evaluation
--GPU is the GPu that will beused for evaluation
--n_samples is the numer of samples that will be obtained for each graph
--evaluation_factor is the factor by which the number of diffusion steps is increased compared to the number of diffusion steps that are used during training. So for example if the model is trained with 5 diffusion steps and --evaluation_factor 3, then the model will be evaluated with 15 diffusion steps

configs:

All configs from the paper SDDS can be found in argparse/experiments/Ising for experiments in unbiased sampling and in argparse/experiments/UCO for experiments in combinatorial optimization.

model weights:

The following model weights are made available:

Weights on Combinatorial Optimization Problems:

• Weights for DiffUCO from the SDDS paper

CO Problem Type	Dataset	Seed 1	Seed 2	Seed 3
MaxCl	RB_small	k1zc0zgg	yxyr9urj	l3s6eybg
MIS	RB_small	m3h9mz5g	olqaqfnl	08i3m2dl
MIS	RB_large	cvv1wla0	fuu10c4p	00qoqw0s
MDS	BA_small	ydq3mn05	xzc9mds3	gk5s4nar
MDS	BA_large	64dnrg5p	107hsfqv	0liz28ec
MaxCut	BA_small	114mqmhk	t2ud5ttf	icuxbpll
MaxCut	BA_large	ubti92kx	c11rjsun	c6yoqwmp

• Weights for SDDS: rKL w/ RL from the SDDS paper

CO Problem Type	Dataset	Seed 1	Seed 2	Seed 3
MaxCl	RB_small	5uvh7t41	l3ricjby	flbniwwf
MIS	RB_small	zk3wkaap	91icd2vu	fj1lym7o
MIS	RB_large	rj161hwt	eh5td2pi	c8d4u3uo
MDS	BA_small	qfrf58hx	2ll1wcw1	4rgz2hck
MDS	BA_large	t95aukxn	ukunkwov	x2z3k811
MaxCut	BA_small	9wstv9tl	m5l9z4j6	nka5ez0d
MaxCut	BA_large	oi3fyq7w	4qmwye2w	6irzwfyk

• Weights for SDDS: fKL w/ MC from the SDDS paper

CO Problem Type	Dataset	Seed 1	Seed 2	Seed 3
MaxCl	RB_small	ud133dhi	yp999f1m	thpyvtjx
MIS	RB_small	otpu58r3	9xoy68e6	w3u4cer6
MIS	RB_large	6rrd7m5b	rjcm5oto	lgo8u2aq
MDS	BA_small	a7zogxmh	df9rgk6a	yjwopr68
MDS	BA_large	x3mdgetb	cpg13tch	05juku5c
MaxCut	BA_small	8ah3bsvm	c1l3z0d4	s2ug6f2y
MaxCut	BA_large	r3ils8y0	qidpkk4j	96u1z4mu

Weights on Statistical Physics Problems:

Ising Model 24x24	Weight Id
fKL w/ MC	qkfzunur
rKL w/ RL	ewmsen06 or sw5qr5e6

Spin Glasses 15x15	Weight Id
fKL w/ MC	4hl3jr35
rKL w/ RL	sw5qr5e6

How to evaluate models in unbiased sampling:

evauation can be ran with the following command:

python evaluate_unbiased_sampling.py --wandb_id <WANDB_ID>  --GPU 0 --n_sampling_rounds 400 --seeds 9  --n_test_basis_states 1200

parameter explanation

--wandb_id wandb_id of the model weights
--GPU ID of the GPU
--n_sampling_rounds number of sampling repetitions
--n_test_basis_states the number of samples in each sampling round
--seeds Number of seeds over which the results will be averaged\

For NIS overall <n_sampling_rounds> x <n_test_basis_states> samples will be used to compute the observables. For NMCMC <n_test_basis_states> are the number of states in the buffer and MCMC updates are performed over <n_sampling_rounds>.

The script in playground/Autocorrelation/eval_unbiased.py can then be used to load the results after "running evaluate_unbiased_sampling.py".