MLA-C01 Practice Question 5: Hyperband Tuning Strategy
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 1 · question 5 of 10
An ML engineer is tuning a PyTorch image classifier that emits a validation:accuracy metric after every epoch. A full training run takes about four hours on the chosen GPU instance, and the current tuning job launches 100 such runs to completion. Finance has asked for a large reduction in tuning cost, and the engineer must not narrow the hyperparameter ranges being explored.
Which change to the SageMaker AI automatic model tuning job best achieves this?
Show answer and explanation
Answer:
Hyperband is a multi-fidelity strategy: it uses the intermediate results that a training job publishes at each resource level, gives more epochs to configurations that look strong, and stops the ones that clearly will not win. It is designed for exactly this kind of iterative algorithm (one that reports an objective metric after every epoch), and AWS notes it can significantly outperform random search and Bayesian optimization on speed. Because most trials die long before four hours, total GPU-hours drop sharply without shrinking the search space. Note that when you use Hyperband you leave TrainingJobEarlyStoppingType set to OFF; Hyperband applies its own advanced early stopping internally.
Option A makes cost worse, not better. Serializing to one training job at a time maximizes how much each trial learns from prior results, but every one of the 100 trials still runs to completion, so the GPU-hours are unchanged and the job takes far longer in wall-clock time.
Option C is not even valid here. Grid search supports only categorical hyperparameter ranges, and exhaustively covering a continuous range such as a learning rate would explode the number of training jobs rather than reduce it.
Option D shortens the calendar, not the bill. Running more trials concurrently finishes sooner, but you still pay for 100 complete four-hour runs, and with random search the extra parallelism buys no search quality either.
Reference: Understand the hyperparameter tuning strategies available in Amazon SageMaker AI