#Use Kubeflow Model Registry

The Kubeflow Model Registry is a central repository for managing machine learning models, their versions, and associated metadata. It allows data scientists to publish models, track their lineage, and collaborate on model development.

#Access the Model Registry

1. Open Dashboard: Log in to the Kubeflow central dashboard.
2. Model Registry: Click on Model Registry in the sidebar. This will take you to the list of registered models in your namespace.

   Note: If you do not see the Model Registry, ensure a Model Registry instance has been deployed in your namespace by the platform administrator.

#Register a Model

You can register models either through the user interface or programmatically using the Python client.

#Option 1: Using the UI

1. Create Registered Model:

   • Click Register Model.
   • Model Name: Enter a unique name (e.g., fraud-detection).
   • Description: Add a description.
   • Version details: Optionally add version information, tags, and metadata.
   • Model Location: Provide the S3/URI to the model artifact (e.g., s3://my-bucket/models/fraud-detection/v1/).
   • Click Create.

2. Create Version:

   • Click on drop down menu next to the Registered Model and select Register New Version.
   • Enter version name, description, metadata and artifact URI.
   • Click Register new version.

#Option 2: Using Python Client

You can register models directly from your Jupyter Notebook using the model-registry Python client.

Prerequisites:

• Install the client: python -m pip install model-registry=="0.3.5" kserve=="0.13"
• Ensure you have access to the Model Registry service. If running inside a Kubeflow Notebook, you can use the internal service DNS (e.g. http://model-registry-service.<namespace>.svc:8080).

Sample Code:

The following example demonstrates how to register a model stored in S3.

from model_registry import ModelRegistry

# 1. Connect to the Model Registry
# Replace with your actual Model Registry service host/port
# Inside cluster, typically: "http://model-registry-service.<namespace>.svc.cluster.local:8080"
registry = ModelRegistry(
    server_address="http://model-registry-service.kubeflow.svc.cluster.local",
    port=8080,
    author="your name",
    is_secure=False
)

# 2. Register a new Model
rm = registry.register_model(
    "iris",
    "s3://kfserving-examples/models/sklearn/1.0/model",
    model_format_name="sklearn",
    model_format_version="1",
    version="v1",
    description="Iris scikit-learn model",
    metadata={
        "accuracy": 3.14,
        "license": "BSD 3-Clause License",
    }
)

# 3. Retrieve model information
model = registry.get_registered_model("iris")
print("Registered Model:", model, "with ID", model.id)

version = registry.get_model_version("iris", "v1")
print("Model Version:", version, "with ID", version.id)

art = registry.get_model_artifact("iris", "v1")
print("Model Artifact:", art, "with ID", art.id)

#Deploy a Registered Model

Once a model is registered, you can deploy it as an InferenceService using KServe.

To deploy, you typically need the URI of the model artifact. You can retrieve this from the Registry UI or via the Python API:

from kubernetes import client
import kserve

isvc = kserve.V1beta1InferenceService(
    api_version=kserve.constants.KSERVE_GROUP + "/v1beta1",
    kind=kserve.constants.KSERVE_KIND,
    metadata=client.V1ObjectMeta(
        name="iris-model",
        namespace=kserve.utils.get_default_target_namespace(),
        labels={
            "modelregistry/registered-model-id": model.id,
            "modelregistry/model-version-id": version.id,
        },
    ),
    spec=kserve.V1beta1InferenceServiceSpec(
        predictor=kserve.V1beta1PredictorSpec(
            model=kserve.V1beta1ModelSpec(
                storage_uri=art.uri,
                model_format=kserve.V1beta1ModelFormat(
                    name=art.model_format_name, version=art.model_format_version
                ),
            )
        )
    ),
)
ks_client = kserve.KServeClient()
ks_client.create(isvc)

Once deployed, the KServe controller will pull the model from the specified S3 URI and start the inference server.