Designing Production-Grade Deep Learning Pipelines

About this course

Most deep learning courses stop when the model reaches a good score in a notebook. Production starts there. This course takes one model all the way: from local experiments to an automated pipeline on Google Cloud that validates new data, retrains, judges the new model against the old one, and deploys it only if it is better.

You work on a single real problem throughout: predicting the compressive strength of concrete from its mix design, using the UCI Concrete Compressive Strength dataset. Because the case study never changes, every new tool solves a problem you have already felt.

What you will build, step by step

Local foundations. Split data into time-ordered spans, engineer features without leakage, build and train a Keras DNN, and track every run in MLflow.

TFX pipelines. Rebuild the workflow from ExampleGen to Pusher: statistics, schema, data validation, Transform, Bayesian hyperparameter tuning with Keras Tuner, training, evaluation gates and the latest-blessed-model resolver.

Local orchestration. Assemble the components into a DAG, run it with Apache Beam, and use artifact caching and ML Metadata.

Google Cloud. Set up the GCP project, IAM and Cloud Storage, containerize the pipeline, and run it on Vertex AI Pipelines.

Deployment. Register models in the Vertex AI Model Registry, serve them from Vertex AI Endpoints, and deploy automatically with Cloud Run functions.

Monitoring and retraining. Understand drift and training–serving skew, configure Vertex AI Model Monitoring, and retrain automatically when new data arrives.

CI/CD. Ship pipeline changes from GitHub with GitHub Actions and keyless Workload Identity Federation.

How the course is taught

Read the full description on Udemy on Udemy (opens in a new tab)

Prices, discounts and availability are set by Udemy. We may earn a commission when you purchase through links on this site.