modelzone-sdk 0.3.0

Modelzone SDK – a slim model training and serving toolkit

Modelzone SDK

A slim Python SDK for defining, training, and serving machine-learning models.

Installation

pip install modelzone-sdk

For local training with Delta Lake data access:

pip install modelzone-sdk[training]

For AzureML experiment tracking:

pip install modelzone-sdk[azureml]

Quick start

Define a model by subclassing ModelDefinition and implementing train and predict:

from modelzone.core import ModelDefinition, PredictContext, ModelArtifact
from modelzone.training import TrainingContext


class MyModel(ModelDefinition):
    def train(self, ctx: TrainingContext) -> ModelArtifact:
        ctx.print("Training started")

        # … your training logic …
        fitted = train_something(seed=ctx.seed)

        ctx.log_metric("accuracy", 0.95)
        return ModelArtifact(model=fitted, features=["feature_a", "feature_b"])

    def predict(self, ctx: PredictContext):
        df = ctx.db.query("input_table", ctx.time_interval)
        return ctx.model.predict(df[ctx.features])

Local training

from modelzone.training import LocalBackend

backend = LocalBackend(root="./runs")
result = backend.run(MyModel(), seed=42, params={"lr": 0.01})
print(result.run_id)

AzureML training

from modelzone.azureml import AzureMLBackend

backend = AzureMLBackend(
    subscription_id="...",
    resource_group="...",
    workspace_name="...",
    experiment_name="my_experiment",
)
result = backend.run(MyModel(), seed=42)

Loading a trained model for prediction

from modelzone.predict import load_model_artifact

model_artifact = load_model_artifact(".model")
print(model_artifact.features)

CLI

modelzone train my_model_package            # Local training
modelzone train my_model_package --azureml   # AzureML training
modelzone register my_model_package <run_id>
modelzone fetch my_model_package --output-dir .model

Development

poetry install --with dev,test
pytest