GCP Professional ML Engineer Practice Question 1: Low Code Image Classification Choice
Google Cloud Professional Machine Learning Engineer · Test 1 · question 1 of 10
A retailer needs to sort incoming supplier product photos into its own 40-category internal taxonomy. It has 6,000 images already labelled with those categories. The team consists of backend developers with no ML background, they do not want to write or maintain training code, and they need a deployable model within a few days.
Which approach best meets these requirements?
Show answer and explanation
Answer:
AutoML exists for exactly this shape of problem: a labelled dataset, a custom label space, and a team that wants a trained, deployable model with minimal technical effort. Google's own guidance on choosing a training method positions AutoML as the option for creating and training a model with the least code, while custom training is for teams that need control over the algorithm, loss function, or distribution strategy. 6,000 labelled images across 40 classes is a workable AutoML image classification dataset, and the resulting model deploys to an endpoint from the console.
Option A fails on the taxonomy. The Cloud Vision API's label detection returns labels from Google's own general-purpose vocabulary; it has no knowledge of the retailer's 40 internal categories, and hand-mapping generic labels onto a proprietary taxonomy is brittle and lossy. Pre-trained APIs are the right call only when the pre-trained label space is the label space you actually want.
Option C would work technically and gives maximum control, but it is precisely the effort the stem rules out: this team cannot write or maintain a training application.
Option D is the heaviest option on the list. Reserving a managed training cluster of H100 GPUs is aimed at large-scale, high-performance training runs that need guaranteed capacity, wildly disproportionate to 6,000 images, and it still requires a fine-tuning workflow the team cannot author.
Reference: Choose a training method