Model Registry Tutorial

api.models is a facade over Datamint’s MLflow-backed model registry. It wraps MLflow’s RegisteredModel / ModelVersion objects in plain Python objects (Model, ModelVersion), so you can register, list, and inspect models without learning MLflow’s object model.

This notebook covers:

  • Registering a model with api.models.create()

  • Listing and finding models with api.models.get_list() / get_by_name()

  • Filtering to deployed models with only_deployed=True

  • Inspecting a model’s versions with model.get_versions() / get_latest_version()

  • Reading what a version was trained for: get_task_type(), get_supported_modes(), get_annotation_specs()

  • Reading training metrics with get_metrics()

  • Checking deployment status with is_deployed()

  • Finding which projects a model is associated with, with model.get_projects()

Setup

Update datamint first if needed, then configure your API key.

[ ]:

%pip install -U datamint --quiet

[ ]:

from datamint import Api

api = Api()

Register A Model

Registering ahead of training gives you a stable name to reference later. create() returns the existing model instead of raising when one with this name is already registered (exists_ok=True by default).

[ ]:

MODEL_NAME = "tutorial_model_registry_demo"

model = api.models.create(MODEL_NAME, description="Model created for the model registry tutorial")
model.name, model.description

List And Find Models

get_list() returns every registered model; get_by_name() returns a single one, or None if it doesn’t exist.

[ ]:

all_models = api.models.get_list()
print(f"Registered models: {[m.name for m in all_models]}")

found = api.models.get_by_name(MODEL_NAME)
missing = api.models.get_by_name("does-not-exist")
found.name, missing

Filter To Deployed Models

Pass only_deployed=True to skip models that don’t have a deployed image yet. See 05_deployment/01_deploy_registered_model.ipynb for how to deploy one.

[ ]:

deployed_models = api.models.get_list(only_deployed=True)
print(f"Deployed models: {[m.name for m in deployed_models]}")

model.is_deployed()

Inspect Versions Of A Trained Model

The rest of this notebook needs a model with at least one version behind it, typically one registered by a Datamint trainer (see 06_end_to_end) or by 04_experiment_tracking/01_mlflow_manual_logging.ipynb. Replace MODEL_NAME below with one you’ve already trained.

[ ]:

MODEL_NAME = "FracAtlas_adapted"  # replace with a model you've already registered/trained

model = api.models.get_by_name(MODEL_NAME)
if model is None:
    raise ValueError(f"Model '{MODEL_NAME}' was not found. Train one first, e.g. via the 06_end_to_end notebooks.")

versions = model.get_versions()
print(f"{MODEL_NAME} has {len(versions)} version(s)")

latest = model.get_latest_version()
latest.version, latest.run_id

What A Version Was Trained For

If the version was logged with the datamint MLflow flavor (true for anything trained through a Datamint trainer), you can read the task type, supported prediction modes, and annotation specs straight from the model artifact, without inspecting the training run.

[ ]:

print("Task type:", latest.get_task_type())
print("Supported modes:", latest.get_supported_modes())
print("Annotation specs:", latest.get_annotation_specs())

# Model.get_supported_modes() is a shortcut that delegates to the latest version
model.get_supported_modes() == latest.get_supported_modes()

Training Metrics

get_metrics() reads the metrics logged for this version’s training run. It returns {} instead of raising when there’s no training run behind the version, for example a model registered from outside Datamint.

[ ]:

latest.get_metrics()

[ ]:

model.get_metrics()  # shortcut, uses the latest version

Find Which Projects A Model Belongs To

A registered model can be associated with one or more Datamint projects. model.get_projects() returns the full Project objects, resolved via api.projects.

[ ]:

projects = model.get_projects()
print(f"{MODEL_NAME} belongs to {len(projects)} project(s)")

for project in projects:
    print(project.id, project.name)

Next Steps

Once you have a model version you’re happy with, alias it and deploy it, see 05_deployment/01_deploy_registered_model.ipynb. Registered models are also created automatically when you pass --ai-model <name> to datamint upload with a name that doesn’t exist yet.