Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Vertex AI

Overview

Vertex AI is Google Cloud’s fully managed ML platform that unifies the entire ML workflow. It provides AutoML for no-code model training, custom training for full control, a feature store, model registry, pipelines, and endpoints for serving. Vertex AI integrates deeply with Google Cloud services and provides access to Google’s TPU infrastructure.

Key Components

graph TD
    A[Vertex AI] --> B[AutoML]
    A --> C[Custom Training]
    A --> D[Feature Store]
    A --> E[Model Registry]
    A --> F[Pipelines]
    A --> G[Endpoints]
    A --> H[Matching Engine]
    A --> I[Generative AI]
    B --> B1[Tables, Images, Text, Video]
    C --> C1[Custom containers on GCP infra]
    D --> D1[Online + Offline features]
    E --> E1[Model versioning]
    F --> F1[Kubeflow-based pipelines]
    G --> G1[Managed serving with autoscaling]
    H --> H1[Vector similarity search]
    I --> I1[Vertex AI Studio, Model Garden]

AutoML

from google.cloud import aiplatform

# Train AutoML model
job = aiplatform.AutoMLTabularTrainingJob(
    display_name="fraud-detection",
    optimization_prediction_type="classification",
    column_transformations=[
        {"numeric": {"column_name": "amount"}},
        {"categorical": {"column_name": "merchant_type"}},
    ],
)

model = job.run(
    dataset=dataset,
    target_column="is_fraud",
    training_fraction_split=0.8,
    validation_fraction_split=0.1,
    test_fraction_split=0.1,
)

Custom Training

job = aiplatform.CustomTrainingJob(
    display_name="custom-training",
    script_path="train.py",
    container_uri="us-docker.pkg.dev/vertex-ai/training/pytorch-gpu.1-13:latest",
    requirements=["torch", "scikit-learn"],
    model_serving_container_image_uri="us-docker.pkg.dev/vertex-ai/prediction/pytorch-gpu.1-13:latest",
)

model = job.run(
    replica_count=1,
    machine_type="n1-standard-4",
    accelerator_type="NVIDIA_TESLA_T4",
    accelerator_count=1,
)

Model Deployment

# Deploy to endpoint
endpoint = model.deploy(
    deployed_model_display_name="fraud-detection-v2",
    machine_type="n1-standard-4",
    min_replica_count=1,
    max_replica_count=10,  # Autoscaling
    traffic_split={"0": 100},  # All traffic to new model
)

# Online prediction
response = endpoint.predict(instances=[{"amount": 100, "merchant_type": "online"}])

Vertex AI Pipelines

from kfp.v2 import dsl
from google_cloud_pipeline_components.v1 import automl

@dsl.pipeline(name="training-pipeline")
def pipeline():
    dataset_op = automl.TabularDatasetCreateOp(
        project=PROJECT_ID,
        display_name="dataset",
        gcs_source="gs://bucket/data.csv",
    )
    training_op = automl.AutoMLTabularTrainingJobRunOp(
        project=PROJECT_ID,
        display_name="training",
        dataset=dataset_op.outputs["dataset"],
        target_column="target",
    )

Interview Questions

  1. What is Vertex AI? — Google Cloud’s unified ML platform providing AutoML, custom training, feature store, model registry, pipelines, and managed serving. It integrates with GCP services and TPU infrastructure.

  2. AutoML vs Custom Training? — AutoML: no-code, Google handles architecture/hyperparameters, limited customization. Custom Training: full control over code, architecture, and training process.

  3. How does Vertex AI handle scaling? — Endpoints autoscale based on traffic (min/max replicas). Training jobs can use distributed training with GPU/TPU clusters. Pipelines scale with K8s.

  4. Vertex AI vs SageMaker? — Both provide end-to-end ML platforms. Vertex AI: better TPU access, tighter GCP integration, stronger AutoML. SageMaker: larger ecosystem, more flexible pricing, AWS integration.

  5. What is Vertex AI Matching Engine? — A managed vector similarity search service for building recommendation systems, semantic search, and other nearest-neighbor applications.

Summary

Vertex AI provides a comprehensive, managed ML platform on Google Cloud. Its AutoML capabilities enable rapid prototyping, while custom training offers full control. Deep GCP integration, TPU access, and managed serving make it a strong choice for teams already on Google Cloud.

Cross-References