PyNAS is a Python framework for Neural Architecture Search (NAS) experiments focused on resource-constrained, edge-deployable deep learning models. The project is developed by ESA Phi-lab with Little Place Lab and targets workflows where model accuracy, memory footprint, and deployment cost must be considered together.
PyNAS provides building blocks for evolutionary NAS experiments, with an emphasis on compact neural networks for onboard and edge AI use cases.
| Area | What is included |
|---|---|
| Architecture generation | Utilities for generating, parsing, and rebuilding architecture codes |
| Search operators | Genetic mutation and single-point crossover primitives |
| Model blocks | Convolutional, pooling, activation, residual, classifier, and U-Net components |
| Training utilities | PyTorch Lightning modules, segmentation losses, metrics, and early stopping helpers |
| Data utilities | Hugging Face dataset download helper with retry handling |
| Item | Status |
|---|---|
| Package name | esa-pynas |
| Version | 0.1.2 |
| Python support | >=3.11,<3.15 |
| Distribution | Local checkout; PyPI publication planned |
| License | Apache 2.0 |
| Development status | Alpha / research preview |
PyNAS uses uv for reproducible dependency management. From a local checkout:
uv sync --lockedFor development and verification:
uv sync --locked --dev
uv run pytest
uv run ruff format --check src scripts data tests
uv run ruff check src scripts data testsThe default lock configuration uses CPU PyTorch wheels for deterministic Windows, Linux, and macOS CI environments. For CUDA-enabled training workloads, install the PyTorch wheel index that matches your driver and hardware before running experiments.
Generate and inspect a candidate architecture code:
frompynas.core.architecture_builderimport (
generate_code_from_parsed_architecture,
generate_random_architecture_code,
parse_architecture_code,
)
architecture_code=generate_random_architecture_code()
parsed_architecture=parse_architecture_code(architecture_code)
round_tripped_code=generate_code_from_parsed_architecture(parsed_architecture)
print(architecture_code)
print(round_tripped_code)Use the public package interface for training-related utilities:
frompynasimport (
CategoricalCrossEntropyLoss,
GenericLightningSegmentationNetwork,
Individual,
calculate_iou,
)The burned-area segmentation dataset referenced by this project is hosted on Hugging Face:
Install the optional data dependencies and run the downloader:
uv sync --locked --extra data
uv run python data/download_hf_datasets.pyThe downloader supports dataset and model repositories from the Hugging Face Hub, retry handling for transient network failures, progress reporting, and custom local output directories. Edit the repo_ids list in data/download_hf_datasets.py to change the default download targets.
.
├── data/ # Dataset download utilities and data notes
├── docs/ # Documentation assets and static documentation files
├── examples/ # Usage examples and demos
├── notebooks/ # Experiment and walkthrough notebooks
├── papers/ # Research manuscript assets
├── scripts/ # Training and data-loading scripts
├── src/pynas/ # PyNAS Python package
│ ├── blocks/ # Neural network building blocks
│ ├── core/ # Architecture, population, configuration, and model logic
│ ├── opt/ # Evolutionary optimization operators
│ └── train/ # Losses, metrics, and training helpers
└── tests/ # Unit and integration tests
Spaceborne edge computing enables AI-capable CubeSats to process data onboard, reduce downlink pressure, and operate with greater autonomy. These systems face strict memory, power, and latency constraints, so model design must account for both predictive performance and deployment cost.
PyNAS explores evolutionary Neural Architecture Search for this setting. The framework is designed around compact segmentation architectures and hardware-aware optimization for CubeSat-class platforms such as NVIDIA Jetson AGX Orin and Intel Movidius Myriad X targets.
The research motivation and validation details are described in the accompanying Scientific Reports article, "Optimizing deep learning models for on-orbit deployment through neural architecture search" (DOI: 10.1038/s41598-025-21467-8).
- Project documentation: sirbastiano.github.io/pynas-docs
- Source repository: github.com/ESA-PhiLab/pynas
- Scientific Reports paper: Optimizing deep learning models for on-orbit deployment through neural architecture search
- ESA Phi-lab: philab.esa.int
- Little Place Lab: littleplace.com
If you use PyNAS in academic work, cite the Scientific Reports paper below. Add the Zenodo software citation as well once the project DOI is published.
@article{delprete2025pynas,
title = {Optimizing deep learning models for on-orbit deployment through neural architecture search},
author = {Del Prete, Roberto and Thind, Parampuneet Kaur and Mazzeo, Andrea and Whitley, Matthew and Papa, Lorenzo and Long{\'e}p{\'e}, Nicolas and Meoni, Gabriele},
journal = {Scientific Reports},
volume = {15},
pages = {37783},
year = {2025},
doi = {10.1038/s41598-025-21467-8},
url = {https://doi.org/10.1038/s41598-025-21467-8}
}- Roberto Del Prete (Google Scholar)
- Parampuneet Thind (Google Scholar)
- Andrea Mazzeo
- Lorenzo Papa (Google Scholar)
- Matthew Whitley
- Gabriele Meoni (Google Scholar)
- Nicolas Longepe (Google Scholar)
This project is licensed under the Apache License 2.0. See LICENSE for the full license text.