gradgraph 0.0.2

Core library for graph-structured parameter optimization via gradient methods.

GradGraph

GradGraph is a Python library for gradient-based parameter optimization on graph-structured data.
It provides tools to preprocess temporal graphs into optimization-ready arrays and TensorFlow templates for simulating ODE/PDE-based dynamical systems.

---

Features

• Graph preprocessing

  • Extract linear paths from networkx graphs
  • Apply sliding-window transformations to generate equal-length arrays
  • Handle temporal node attributes (e.g. position, time of appearance)

• Differentiable dynamical models

  • Define custom ODE/PDE-inspired layers (BasePDESystemLayer)
  • Use local vs. global trainable parameters with constraints
  • Built on tf.keras for automatic differentiation and training

• Training utilities

  • BasePDESystemTrainer integrates with the full Keras training API
  • Gradient accumulation and optimizer splitting (global vs. local)
  • Early stopping, learning rate scheduling, and checkpointing callbacks

• Applications

  • Modeling biological growth (e.g. fungal networks)
  • Simulating transport, epidemiological spread, and other temporal graph processes
  • General gradient-based parameter estimation on irregular domains

---

Installation

From PyPI (recommended)

The package is published on PyPI, so you can install it directly with:

pip install gradgraph

From source

git clone https://github.com/Fricker95/gradgraph.git
cd gradgraph
pip install -e .

Dependencies include: numpy, scipy, networkx, tensorflow>=2.0, matplotlib.

---

Examples

See the notebooks in examples/ (e.g., example_linear_growth_graph.ipynb) for end-to-end workflows: loading graphs, feature extraction, interpolation, model setup, training, and evaluation.

---

Documentation

Full API docs and guides are available on ReadTheDocs:
https://gradgraph.readthedocs.io/en/latest/

If you find something missing or unclear, please open an issue or PR.

---

Community

Who is this for?
Researchers and practitioners in network science, computational biology, applied mathematics, and machine learning who fit dynamical models on graphs and need efficient, gradient-based parameter estimation.

How to contribute

• Open an issue for bugs or feature requests: https://github.com/Fricker95/gradgraph/issues
• Submit a pull request for improvements (tests appreciated!).
• Follow a clear commit message style and include concise docstrings/type hints where possible.

If you plan a larger contribution, feel free to start a discussion in an issue to align on design.

---

License

This project is released under the MIT License.
See the LICENSE file for details.

---

Citation

This work was supported by EUR SPECTRUM at Université Côte d’Azur (50%) and by the French National Research Agency (ANR) through the project NEMATIC (50%), grant number ANR-21ANR08Z6RCHX.

If you use GradGraph in your research, please cite the associated JOSS paper (pending):

@article{Fricker2025gradgraph,
  title   = {GradGraph: Gradient-based Parameter Optimization on Graph-Structured Data in Python},
  author  = {Fricker, Nicolas E. and Monasse, Laurent and Guerrier, Claire},
  journal = {Journal of Open Source Software},
  year    = {2025},
  note    = {In review},
}