Extending PhenoMe

Most users start by extracting metadata from paths or CSVs, or by adding a domain-specific property function. Pick the extension point that matches what you need.

Extension map

Metadata extraction

Your file paths or layout differ from the defaults; extract condition, drug, timepoint from paths or a CSV.

• Experiment details - overview.
• Path templates - folder-based parsing.
• CSV lookup.
• Advanced patterns - regex, MetadataBase.

Property computation

Add domain-specific features (morphology, intensity, texture) from images and masks.

• Select properties — Presets - built-in factories.
• Writing property functions.
• Advanced recipes - GLCM, OpenCV, multi-channel.

Vision models

Use a different architecture (for example ResNet or custom transformers) or load ONNX / TensorFlow models natively.

• Vision model wrappers - subclass ModelWrapper.
• load_dinov2_model - reference implementation.

Plugins and registries

Register named property functions, metadata extractors, or report sections so other code can discover them.

• Plugins.
• register_property, register_metadata_extractor, register_report_section.

Rule of thumb

Prefer hooks before forking internals:

1. A metadata_fn if you need different metadata.
2. A property function if you need different features.
3. A ModelWrapper subclass if you need a different backbone.
4. A plugin registration when you want the extension to be discoverable across projects.

Only modify phenome.pipeline or phenome.mixins when none of the above is enough; see Architecture for boundaries.

See also

ArchitecturePackage layout and data flow.

Developer guideDocstrings, Astro build, CI.

TestingRuff, mypy, interrogate before you open a PR.

HDF5 protocolCheckpoint layout for integrators.