Skip to content

Model Registry

The model registry of AIoD contains information about the models, enabling us to pull and use this information across the front-ends and Nextflow pipeline. Through this, we can allow for anyone to add models to the registry and propagate these changes through the framework with minimal user/developer effort.

Each manifest in the registry defines a model family (a group of models that can be run in the same environment, e.g. Cellpose 1/2/3) that is comprised of an arbitrary number of model versions (a specific model within a family, e.g. cyto3).

Schema

The schema for model manifests in AIoD is intended to be lightweight, as similar efforts in the past have created too much of a burden on users and thus inhibited adoption. Our schema is defined by a Pydantic model, allowing us to not only validate manifests, but to ensure that when read into Python all inputs are of the appropriate type (this is useful for automatic ingest).

As mentioned, each manifest itself represents a model family. In short, the schema defines:

  • Meta information about the model (associated websites, publications, authors, repositories etc.)
  • Information about each model version:
    • It's task (what it is doing, e.g. segmenting mitochondria)
    • It's parameters (and what they're for)
    • It's location (where to find the model)

Note that the parameters, metadata etc. can be written at the top level and apply to all models in a model family, but each version can override this as needed for specific differences.

Automatic Ingest

As it would be a monumental pain to e.g. add elements to the Napari plugin dropdowns every time we added a new model family/version, the registry contains simple functionality to allow easy ingestion of all manifests. This can then be used to automatically construct the UI at runtime, allowing for the interface to always reflect all possible options within the registry.

Expanding the Registry

To add new models to the registry, see our model contribution section. Using Pydantic, validation will be confirmed via pytest upon a pull request.