modelzone-sdk 1.0.0

Modelzone SDK – a slim model training and serving toolkit

ModelZone SDK

A Python package for managing, training, and installing machine learning models. The package consists of:

• A Python SDK used for writing the code for training and inference of a machine learning model.
• A CLI used for tracking and registration of models in Azure ML.

Installation

pip install modelzone-sdk

Overview

Taking a look at the various Python machine learning frameworks out there, you will come across many different workflows used for training and prediction. ModelZone SDK uses the following definition:

A model project is a Python package with a training function, which is responsible for (1) tracking metrics and graphs used for comparing different model experiments and (2) generating model artifacts (such as model parameters and hyper-parameters) that need to be shipped with the package.

ModelZone SDK assumes your machine learning workflow looks more or less like the following:

1. Experimentation by training different models (and hyper-parameters).
2. Release of the best suited model for production usage.
3. Using the released model for inference in operations.

---

Quick guide

We have made an example of how a training and inference project should look like:

• Training project: examples/training_example
• Inference project: examples/predict_example
• Test script running the full workflow: examples/test.sh

---

Detailed guide

Model project structure

Training function

Your training function must be defined within the Python package:

# my-model/model.py
def train():
    X, y = ...
    lin_reg = LinearRegression()
    lin_reg.fit(X, y)

Tracking

If you have experience with the built-in logging module of Python, tracking experiment data with ModelZone SDK will feel much the same. Tracking is done using a global tracker, which needs to be configured before running your code:

# my-model/model.py
import modelzone as mz

tracker = mz.get_tracker()

def train():
    X, y = ...
    lin_reg = LinearRegression()
    tracker.log_tag("model_type", "LinearRegression")
    lin_reg.fit(X, y)
    y_hat = lin_reg.predict(X)
    mae = (y - y_hat).abs().mean()
    tracker.log_metric("MAE", mae)

When using the CLI tool, this takes care of configuring your tracker to log to the correct place (local or Azure AI Workspace).

Artifacts

You are free to store model artifacts as you would any data to be included in a Python package. However, you can also use the Artifact class from ModelZone SDK, which helps saving and loading artifacts by name (behind the scenes, it uses the pickle module):

# my-model/model.py
import modelzone as mz

class Predictor(mz.Artifact):

    def __init__(self, model: Estimator):
        self.model = model

    def predict(self, ...):
        ...

def train():
    X, y = ...
    lin_reg = LinearRegression()
    lin_reg.fit(X, y)
    predictor = Predictor(lin_reg)
    predictor.save(name="my-artifact")

When installing this package elsewhere, you can now get back the artifact as follows:

from mypackage import Predictor

predictor = Predictor.load(name="my-artifact")