💡 Notes - Deployment of Machine Learning Models

05 May 2021 - 2 mins read time
Tags: Notes Learning



01. Course Content

  1. Deployment Overview
    • Model deployment and why it matters
    • Research and production environments
  2. Machine Learning System Architecture
    • Machine Learning system architecture
      • architecture components
      • challenges and principles of creating a suitable system architecture
      • different architecture approaches
    • Building reproducible pipelines
      • challenges and how to mitigate them
  3. Research Environment
    • Machine Learning pipeline
      • data gathering
      • feature engineering and selection
      • model building and assessment
    • Starting point for deploying a machine learning model
  4. Production Code
    • Transform the notebooks to production code
      • code and project structure
      • OOP
      • logging, error handling and unit tests
      • use of environments and testing tools
      • production tools
  5. Serving Model via REST API
    • Create an API - FastAPI
    • Send data and obtain predictions
    • Schema validation
    • Exception handling
  6. Packaging a Model
    • Create a package with the model
    • Deploy to private repository
    • Tooling
  7. CI/CD
    • Introduction to CI/CD
    • CircleCI
    • Publish model to a private repository
    • Test the CI pipeline
  8. Deploy to Platform as a Service (Paas)
    • What is PaaS
    • Tooling
    • Deploying with CI/CD
  9. Running Apps in Containers
    • Intro to containers and Docker
    • Installing Docker
    • Creating an API App Dockerfile
    • Build and run a Docker container
    • Release to PaaS utilising Docker
  10. Assessing Reproducibility
    • Compare the results between production and research models
    • Assess deployed model performance



02. Overview of Model Deployment

  1. Deployment of Machine Learning Models
    • What is Machine learning deployment?
      • the deployment of machine learning models is the process for making models available in production environments, where they can provide predictions to other software systems
      • one of the last stages in the Machine Learning lifecycle
      • potentially the most challenging stage
    • Why is model deployment challenging?
      • challenges of traditional software: reliability, reusability, maintainability, flexibility
      • additional challenges specific to Machine Learning: reproducibility
      • needs coordination of data scientists, IT teams, software developers and business professionals: ensure model works reliably, ensure model delivers the intended results
      • potential discrepancy between programming language in which the model is developed and the production system language - re-coding the model extends the project timeline and risks lack of reproducibility
    • Why is model deployment important?
      • to start using a Machine Learning model, it needs to be effectively deployed into production, so that they can provide predictions to other software systems
      • to maximize the value of Machine Learning models, we need to be able to reliably extract the predictions and share them with other systems