MLOps

[YunchaoYang]

Outline

This post will elaborate on how to effectively use Machine Learning infrastructure. Namely, the best practice in building, maintaining and scaling production-ready deep learning systems.

0) Build a production ready deep learning pipeline

Tensoflow extended (TFX) pytorch ? Nvidia ?

1) Kubernetes with Google Cloud: Deploy your Deep Learning model effortlessly.

2) Scalability in ML

3) Docker containers and Docker Compost

4) uWSGI Nginx serving a Tensorflow model to users with Flask, uWSGI as a web server and Nginx as a reverse proxy.

5) Deploy a Deep Learning model as a web application with Flask

references:

1. https://theaisummer.com/topics/mlops/

[YunchaoYang]

The MLOps projects are roughly categorized into the following:

1. Training Orchestration
2. Model Monitoring
3. Model Testing
4. Model Serving
5. Data Versioning
6. Feature Store
7. Experiment Tracking
8. Explainability

Another different categorization strategy:

1. Automates ML workflow
2. CI/CD for ML
3. cron job scheduler: Tools for monitoring cron jobs, a command line job scheduler on Unix like operating system
4. Data Catalog
5. Data Enrichment
6. Data Exploration
7. Data Management
8. Data Processing
9. Data Validation
10. Data Visualization
11. Feature Engineering
12. Feature Store
13. Hyperparameter Tuning
14. Knowledge Sharing
15. Machine Learning Platform
16. and so on

[YunchaoYang]

develop ML products and rapidly bring them into production.

automate and operationalize ML products

concept drift/data drift

[YunchaoYang]

key ideas:

1. Model metadata storage and management 2. Data and pipeline versioning 3. Hyperparameter tuning 4. Run orchestration and workflow pipelines 5. Model deployment and serving 6. Production model monitoring