Register a model

Once you are satisfied with a trained model, you can register it in the Snowflake Model Registry under a database, schema, and name of your choice. Registered models are versioned and can be retrieved later for inference.

There are two ways to register a model:

• Register a model immediately after training it in the current workflow.
• Load a model that was trained in an earlier workflow and register it.

Note

A trained model can only be registered once. You cannot register the same model under different names.

Register a model you just trained

After calling gnn.fit(), call gnn.register_model() on the same GNN instance:

gnn = GNN(
    exp_database="EXPERIMENTS_DB",
    exp_schema="EXPERIMENTS_SC",
    graph=gnn_graph,
    property_transformer=pt,
    train=Train,
    validation=Val,
    task_type="regression",
)

gnn.fit()

gnn.register_model(
    model_database="PRODUCTION_DB",
    model_schema="MODEL_REGISTRY",
    model_name="my_dataset_model",
    version_name="v1",
    comment="First baseline model",
)

The arguments to register_model() are:

• model_database and model_schema — the Snowflake database and schema where the model will be registered.
• model_name — the model name under which the model will be registered in.
• version_name — the version label for this registration. Use it to track multiple versions of the same model_name.
• comment — a free-text description stored alongside the registered model.

The RelationalAI Native App must have permission to create models in model_database.model_schema. As a reminder, to grant the required permissions, run:

-- Grant access to required database and schema
GRANT USAGE ON DATABASE <DATABASE> TO APPLICATION RELATIONALAI;
GRANT USAGE ON SCHEMA <DATABASE>.<SCHEMA> TO APPLICATION RELATIONALAI;
-- Allow registering trained models
GRANT CREATE MODEL ON SCHEMA <DATABASE>.<SCHEMA> TO APPLICATION RELATIONALAI;

Register a previously trained model

To register a model from an earlier training run, instantiate a GNN with the run’s model_run_id, call gnn.load() to retrieve the model, and then call gnn.register_model():

gnn = GNN(
    exp_database="EXPERIMENTS_DB",
    exp_schema="EXPERIMENTS_SC",
    model_run_id="<model_run_id>",
    source_concept=SourceConcept,
    task_type="regression",
)

gnn.load()

gnn.register_model(
    model_database="PRODUCTION_DB",
    model_schema="MODEL_REGISTRY",
    model_name="my_dataset_model",
    version_name="v2",
    comment="Ready for production",
)

The model_run_id value is the job ID of the original training run — the same identifier printed in the training logs and shown in the run in Snowflake Experiment Tracking (as the suffix in RUN_<job_id>, or under model_run_id in the run details). Also see Monitor training.

For the full list of valid and invalid GNN method-call sequences, see Understand GNN workflows.