How To Use AutoML on Google Cloud’s Vertex AI

Alternative to custom training: AutoML — build a great model from scratch

Vertex AI (Source: Google Cloud)

Vertex AI Tutorial Series

1. A Step-by-Step Guide to Training a Model on Google Cloud’s Vertex AI
2. A Step-by-Step Guide to Tuning a Model on Google Cloud’s Vertex AI
3. How To Operationalize a Model on Google Cloud’s Vertex AI
4. How To Use AutoML on Google Cloud’s Vertex AI (this article)
5. How To Use BigQuery ML on Google Cloud’s Vertex AI
6. How to Use Pipeline on Google Cloud’s Vertex AI

Background

In part 1, part 2, and part 3 of this series, we started from scratch, trained a custom model, optimized it, and deployed it on Vertex AI to serve online predictions. It was quite a journey but still relatively straightforward. However, in reality, we may be stuck at the model training and optimization stage for months because finding the optimal model structure and training scheme for a real-world problem can be very challenging.

AutoML is Google Cloud’s attempt to address this challenge. AutoML allows users to bootstrap model creation in a code-free manner. It builds on pre-trained models of typical data types and machine learning tasks, such as image classification, text entity extraction, video object tracking, tabular regression, etc. Users only need to provide the training data and direct AutoML to fine-tune the pre-trained models to “fit” them.

In this article, we’ll explore using AutoML to solve the same image classification problem we’ve been solving in previous articles. Just to recap, the dataset we’ve been using is CIFAR10, which contains 60,000 32x32 images of ten classes.

Prepare the Data

First of all, we need to prepare the data. As in the first article, we’ll need to get the data in a shape that can be easily consumed downstream. We also need to put it on Google Cloud Storage (GCS) so that Vertex AI can access it. For AutoML, plain JPEG images are fine. But we need a metadata file to tell AutoML the GCS locations of those images, their labels, and their splits. One extra wrinkle here is that…