Skip to main content

Automation

Updated May 13, 2023 ·

Overview​

Automation improve efficiency and reduce human error in ML pipelines. The four main principles are:

  • Continuous Integration (CI): Regularly merge code changes into a shared repository.
  • Continuous Delivery (CD): Automatically build, test, and deploy code.
  • Continuous Training (CT): Automatically update models as new data arrives.
  • Continuous Monitoring (CM): Continuously track model performance.

Continuous Integration and Delivery​

Continuous Integration (CI) means frequently integrating code changes and testing them automatically.

  • Changes are tested as soon as they are committed.
  • CI tests each code change to ensure it works.

On the other hand, Continuous Deployment (CD) automates the release of validated code after testing.

  • After testing, the new code is deployed automatically.

For more information, please see CICD Overview.

info

Tools like Git, AWS CodePipeline, Jenkins, and Travis CI are commonly used to implement CI/CD.

Continuous Training​

Continuous Training involves automatically retraining models as new data becomes available, keeping models accurate.

  • Reduce the risk of model decay
  • Reduce the time required to retrain models

Continuous Monitoring​

Continuous Monitoring is the practice of continuously monitoring performance, identifying issues early, and triggering retraining if necessary.

  • Also reduces the risk of model decay
  • Improves overall accuracy ad performance
  • Access to consistent and reliable ML metrics

Example: ML Automation at Scale​

Here’s how automation works in a typical ML pipeline:

  1. Code Commit: Commit model code to Git.
  2. CI/CD: Automatically build and test with Jenkins.
  3. Deploy: If tests pass, deploy to a test environment.
  4. Model Deployment: Package model in Docker, deploy to cloud/local.
  5. Monitoring: Use Prometheus or Grafana for performance tracking.
  6. Retraining: Trigger retraining if performance dips.