PyDGN: a research library for Deep Graph Networks

Documentation

❗❗ 2025 Update: PyDGN is now considered legacy; users are encouraged to switch to MLWiz❗❗

This is a Python library to easily experiment with Deep Graph Networks (DGNs). It provides automatic management of data splitting, loading and common experimental settings. It also handles both model selection and risk assessment procedures, by trying many different configurations in parallel (CPU or GPU).

Citing this work

If you used this library for your project, please consider citing us:

@article{pydgn,
  author = {Errica, Federico and Bacciu, Davide and Micheli, Alessio},
  doi = {10.21105/joss.05713},
  journal = {Journal of Open Source Software},
  month = oct,
  number = {90},
  pages = {5713},
  title = {{PyDGN: a Python Library for Flexible and Reproducible Research on Deep Learning for Graphs}},
  url = {https://joss.theoj.org/papers/10.21105/joss.05713},
  volume = {8},
  year = {2023}
}

Installation:

Automated tests passing on Windows, Linux, and MacOS. Requires at least Python 3.8. Simply run

pip install pydgn

Quickstart:

Build dataset and data splits

pydgn-dataset --config-file examples/DATA_CONFIGS/config_NCI1.yml

Train

pydgn-train  --config-file examples/MODEL_CONFIGS/config_SupToyDGN.yml 

And we are up and running!

To debug your code you can add --debug to the command above, but the "GUI" will be disabled.

To stop the computation, use CTRL-C to send a SIGINT signal, and consider using the command ray stop to stop all Ray processes. Warning: ray stop stops all ray processes you have launched, including those of other experiments in progress, if any.

Using the Trained Models

It's very easy to load the model from the experiments (see also the Tutorial):

from pydgn.evaluation.util import *

config = retrieve_best_configuration('RESULTS/supervised_grid_search_toy_NCI1/MODEL_ASSESSMENT/OUTER_FOLD_1/MODEL_SELECTION/')
splits_filepath = 'examples/DATA_SPLITS/CHEMICAL/NCI1/NCI1_outer10_inner1.splits'
device = 'cpu'

# instantiate dataset
dataset = instantiate_dataset_from_config(config)

# instantiate model
model = instantiate_model_from_config(config, dataset, config_type="supervised_config")

# load model's checkpoint, assuming the best configuration has been loaded
checkpoint_location = 'RESULTS/supervised_grid_search_toy_NCI1/MODEL_ASSESSMENT/OUTER_FOLD_1/final_run1/best_checkpoint.pth'
load_checkpoint(checkpoint_location, model, device=device)

# you can now call the forward method of your model
y, embeddings = model(dataset[0])

Projects using PyDGN

• Infinite Contextual Graph Markov Model (ICML 2022)
• Graph Mixture Density Networks (ICML 2021)
• Contextual Graph Markov Model (ICML 2018, JMLR 2020)
• Extended Contextual Graph Markov Model (IJCNN 2021)
• Continual Learning Benchmark for Graphs (WWW Workshop 2021, spotlight)

Data Splits

We provide the data splits taken from

Errica Federico, Podda Marco, Bacciu Davide, Micheli Alessio: A Fair Comparison of Graph Neural Networks for Graph Classification. 8th International Conference on Learning Representations (ICLR 2020). Code

in the examples/DATA_SPLITS folder.

License:

PyDGN >= 1.0.0 is BSD 3-Clause licensed, as written in the LICENSE file.