Directo: Generating Directed Graphs with Dual Attention and Asymmetric Encoding

A Pytorch implementation to generated directed graphs in the discrete flow matching framework.

Paper: https://arxiv.org/abs/2506.16404

💻 Environment installation

There are two options for environment installation: a conda environment or using pixi.

🐍 To create the conda environment:

conda env create -f environment.yaml
conda activate directo

✨ To use pixi, install it and run the desired command preceeded by:

pixi run ...

👟 Training

The script is runnable from main.py (inside src) which uses Hydra overrides.

Debug mode

python main.py +experiment=debug

Full training

An example of a training run would be:

python main.py +experiment=visual_genome dataset=visual_genome dataset.acyclic=False model.extra_features=rrwp-ppr

⚙️ Parameters

The following parameters can be configured:

• dataset, experiment:
  Choose one of: synthetic | tpu_tile | visual_genome

• model.extra_features:
  These are the positional encodings it is possible to choose from. Extra features to include: null | eigenvalues | magnetic-eigenvalues | rrwp | rrwp-ppr | mult-magnetic-eigenvalues (Q = 5) | mult10-magnetic-eigenvalues (Q = 10)

• dataset.acyclic:
  Set to true or false depending on the dataset (only relevant for some synthetic datasets, otherwise the default is already set in the corresponding Hydra config).

Additional options are available for the synthetic dataset. If you select dataset=synthetic, you can further configure:

• dataset.graph_type:
  Graph generation model: er | sbm | ba

• dataset.acyclic:
  Whether to generate directed acyclic graphs: true (for ER, BA) | false (for ER, SBM)

To run discrete diffusion set model.backbone=diffusion

🎰 Sampling

Sampling optimization

To search for the best sampling configuration, you can run:

python main.py +experiment=visual_genome dataset=visual_genome dataset.acyclic=False \
    model.extra_features=rrwp-ppr general.num_sample_fold=5 general.test_only=your_checkpoint \
    sample.search=all

⚙️ Available parameters

• all — searches through the three parameters
• stochasticity — searches over stochastic level
• distortion — searches over distortion functions
• target_guidance — searches over targuet guidance levels

✅ Final sampling

Once the best sampling parameters are identified, update the configuration accordingly and run:

python main.py +experiment=visual_genome dataset=visual_genome dataset.acyclic=False \
    model.extra_features=rrwp-ppr general.num_sample_fold=5 general.test_only=your_checkpoint \
    sample.eta=10 sample.omega=0.1 sample.time_distortion=polydec

🧪 Baselines

Baseline experiments follow a similar procedure and their adaptation can be found in the src/baselines directory.

📚 Citation

@article{carballo2025directo,
  title     = {Generating Directed Graphs with Dual Attention and Asymmetric Encoding},
  author    = {Carballo-Castro, Alba and Madeira, Manuel and Qin, Yiming and Thanou, Dorina and Frossard, Pascal},
  year      = {2025},
  url       = {https://arxiv.org/abs/2506.16404}
}