Nidavellir Tools

Nidavellir Tools packages trained PyTorch models with reproducibility metadata, portable test tensors, and BioImage.IO-compatible artifacts. It also stages and loads packaged models for transfer-learning runs.

The library is intentionally independent of any model architecture, training framework, microscopy modality, or dataset. Consuming projects remain responsible for their model, DataLoader, augmentation, model RDF, and model card.

Status: the initial API may change before version 1.0.

Installation

Install the core package from PyPI:

python -m pip install nidavellir-tools

Optional integrations are installed explicitly, for example:

python -m pip install "nidavellir-tools[bioimageio,huggingface,mlflow]"

For development from a checkout, use python -m pip install -e ".[dev]".

Command line

nidavellir --help
nidavellir build --help
nidavellir samples --help
nidavellir inspect --help
nidavellir validate --help

The existing command names remain available during migration:

nidavellir-build --help
nidavellir-registry --help
nidavellir-samples --help

Image axes

Sample TIFF creation supports model tensors described by explicit BioImage.IO axes. Standard 2D and 3D layouts include BCYX and BCZYX. The exact NPY tensors retain the complete model input/output boundary, including the batch dimension; TIFFs are presentation samples derived from one selected batch item.

Containerized development

docker build -t nidavellir-tools:dev .
docker run --rm nidavellir-tools:dev --help

Run tests without modifying the host Python environment:

docker run --rm \
  -v "$PWD:/workspace" \
  -w /workspace \
  python:3.12-slim \
  bash -lc 'python -m pip install -e ".[dev]" && pytest -q'

Origin

The project was extracted from luiskuhn/nuxnet-training, where the workflow was first exercised for 3D nuclei segmentation and parent-to-child fine-tuning. The standalone package is derived from that implementation while removing NuxNet-specific assumptions.