Skip to main content

GCP Vertex AI: the unified training and deployment platform on Google Cloud

Vertex AI is Google Cloud's managed platform for the full machine-learning lifecycle: notebooks, data, custom training, tuning, model registry, online and batch prediction, pipelines, and foundation models — under one project, one IAM model, and one billing surface.

Course duration: 7h​

What you will learn​

  • Set up a Vertex AI project the right way: enabled APIs, regions, service accounts, buckets, and per-second billing you can read
  • Work in Vertex AI Workbench and Colab Enterprise notebooks with idle shutdown and direct BigQuery access
  • Read training data from BigQuery and Cloud Storage into Vertex managed datasets or DataFrames, understanding query costs
  • Run a CustomTrainingJob in a prebuilt or custom container, on the machine type and accelerators your model actually needs
  • Tune hyperparameters at scale with Vertex AI Vizier: metrics reported by the code, parallel trials, early stopping
  • Upload models to the Model Registry with versions, default aliases, attached evaluation and lineage back to the training job
  • Deploy two model versions on one endpoint with 90/10 traffic split, min and max replicas, latency budgets and request logging
  • Run nightly batch predictions from BigQuery to BigQuery, and know when batch beats online
  • Build a Vertex AI Pipeline with KFP components, capture artifacts and lineage, and gate deployment on a metric
  • Call a foundation model from Model Garden and fine-tune it with managed tuning, reasoning about cost per token

Prerequisites​

  • Course 20 — MLOps
  • Notions of Google Cloud: projects, IAM, Cloud Storage

Course modules​

  1. Vertex AI: components and vocabulary
  2. Managed notebooks and development environments
  3. Data on Cloud Storage and BigQuery
  4. Custom training and containers
  5. Hyperparameter tuning at scale
  6. Model Registry and versions
  7. Endpoints, traffic split and scaling
  8. Batch prediction
  9. Vertex AI Pipelines
  10. Foundation models and Model Garden

The red thread​

Every module works on the same problem: a payment-transaction fraud-detection model whose data lives in BigQuery — a public-style table of tens of millions of rows with a strong class imbalance (roughly 0.2% fraud). Module 2 opens a Workbench notebook, module 3 pulls a training slice from BigQuery, module 4 trains it in a custom container, module 5 tunes it with Vizier, module 6 registers it, module 7 exposes two versions behind one endpoint with a 90/10 traffic split, module 8 runs a nightly batch scoring job from BigQuery to BigQuery, and module 9 wires the whole thing into a KFP pipeline with lineage and a deployment gate. Module 10 changes register: a foundation model from the Model Garden classifies free-text dispute comments attached to the same transactions.

The Python SDK google-cloud-aiplatform and the gcloud CLI carry the examples. Everything runs on a laptop with one project and a modest budget: expect a couple of dollars per full run through, not hundreds.

Assessment and certificate​

The course ends with a 40-question exam covering every module. On success, a certificate of completion is issued; its number can be verified by anyone on the platform.

Free courses, by contrast, end with a 5-question quiz and a certificate preview, without certification.