This the official implementation of the DiffAlign model as seen in Equivariant Denoisers Cannot Copy Graphs: Align Your Graph Diffusion Models
The processed data can be downloaded from this link. The checkpoint for our best model (aligned with absorbing transition) can be found here.
To train our best model, run the following command:
python3 scripts/train.py +experiment=align_absorbing
Run the following script to generate samples similar to the ones used in the results (Table ...). ´your_experiment_name´ is the name of the experiment where you want to save the samples. Alternatively, you can obtain the exact samples used in the paper in this link.
python3 src/sample_array_job.py
+experiment=7ck
general.wandb.mode=offline
general.wandb.run_id=7ckmnkvc
diffusion.edge_conditional_set=test
general.wandb.checkpoint_epochs=[720]
test.condition_first=0
test.condition_index=0
test.n_conditions=5000
test.n_samples_per_condition=100
dataset.shuffle=False
dataset.dataset_nb=uspto50k
general.wandb.load_run_config=True
hydra.run.dir=../experiments/your_experiment_name/
test.total_cond_eval=5000
train.seed=329
diffusion.diffusion_steps=100
diffusion.diffusion_steps_eval=100
dataset.add_supernode_edges=True
dataset.num_workers=0
To evaluate samples, run the following command. Make sure ´your_experiment_name´ is the name of the experiment where you saved the samples earlier.
python3 src/evaluate_array_job.py
+experiment=7ck
general.wandb.mode=offline
general.wandb.run_id=7ckmnkvc
diffusion.edge_conditional_set=test
general.wandb.checkpoint_epochs=[720]
test.condition_first=0
test.condition_index=0
test.n_conditions=5000
test.n_samples_per_condition=100
dataset.shuffle=False
dataset.dataset_nb=uspto50k
general.wandb.load_run_config=True
hydra.run.dir=../experiments/your_experiment_name/
test.total_cond_eval=5000
train.seed=329
diffusion.diffusion_steps=100
diffusion.diffusion_steps_eval=100
dataset.add_supernode_edges=True
dataset.num_workers=0
@inproceedings{
laabid2025equivariant,
title={Equivariant Denoisers Cannot Copy Graphs: Align Your Graph Diffusion Models},
author={Najwa Laabid and Severi Rissanen and Markus Heinonen and Arno Solin and Vikas Garg},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=onIro14tHv}
}