H2O World - Python Pipelines - Spencer Aiello
Download as PPTX, PDF2 likes844 views
This document discusses moving machine learning models from prototyping to production. It notes that while code may be well-written, different environments can make translation difficult. The author proposes using H2O, which allows users to export feature preprocessing and models to new environments. An example is given of preprocessing loan data with H2OAssembly, building a gradient boosted machine model, and exporting the entire workflow into a Storm topology for production.
Related topics:1 of 10
Downloaded 57 times
H2O World - Python Pipelines - Spencer Aiello
- 1. To Production and Beyond Spencer Aiello
- 2. The Problem • Goal: o Move from prototype to production • Road block: o Prototyping Environment Cages Your: • Feature preprocessing • Models • Ideas
- 3. The Problem • Even if your code is beautiful:
- 4. The Problem • You cannot drag-n-drop into a new environment. • Translation may be difficult; humans make mistakes
- 5. A Solution H2O gives you wings: • Export Preprocessing • Export Models
- 6. H2OAssembly o Build Rich Feature Preprocessing Assembly Lines • Clean, reduce, and expand datasets by composing any of the 100s of primitives available in H2O • Build hygenic processing assembly lines that can be applied to new batches of data • Export your feature preprocessing steps as a plain old java object and apply to streaming tuples
- 7. H2OAssembly
- 9. Live Demo • Lending Club Data: Predict Interest Rate o Four-part dataset of loan data o 500K rows, 52 columns o Preprocess 5 columns within a 16 step assembly o Build a simple GBM to predict interest rate o Export everything into a Storm topology