Effortless model deployment with MLflow — Models with multiple pieces

How to save your Machine Learning models in an open-source format with MLflow to unlock effortless deployment experience later. Saving a recommendation system model.

Welcome back to the series Effortless model deployment with MLflow! If you just join the party, check out the other post of the series:

• MLFlow: Introduction to the MLModel specification.
• Customizing inference with MLFlow: deploying a Computer Vision model with fast.ai.
• Packaging models with multiple pieces (this post).
• Packaging models with multiple assets: deploying a HuggingFace NLP model for classification.
• Packaging stratified models (many models): deploying a partitioned model for demand forecasting.

Introduction: Your model has many pieces

In the first post, we saw an introduction to the MLModel format and why, by persisting your model in an open-source specification format you can achieve great flexibility in deploying models. The only thing it took was to log the model like:

mlflow.keras.log_model(your_model, "classifier", signature)

In the second one, we saw how you can customize the way you run inference for a given model while continuing to keep the…