MLA-C01 Practice Question 86: Autopilot Training Mode
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 9 · question 6 of 10
A marketing team uses the SageMaker Autopilot API to build a baseline model for a tabular lead-conversion dataset of 45 MB. An engineer explicitly set the job's training mode to hyperparameter optimization (HPO), and the job ran up to 100 trials and took many hours. The team wants Autopilot to produce a strong model in much less time, and it has no requirement to tune individual algorithms' hyperparameters.
What should the engineer change?
Show answer and explanation
Answer:
Autopilot has two training modes. *Ensembling* uses the AutoGluon library to train several base models, runs 10 trials with different model and meta parameter settings, and combines the models with a stacking ensemble. *HPO* tunes hyperparameters and runs up to 100 trials by default. In Auto mode, Autopilot picks ensembling for datasets of 100 MB or less, and AWS recommends ensembling for datasets smaller than 100 MB for the best runtime and performance. At 45 MB, forcing HPO was the wrong choice (option A).
- B is wrong because cutting HPO to 10 trials does shorten the job, but it gives up search quality, while ensembling is the mode AWS recommends for this dataset size for both speed and performance. - C is wrong because duplicating rows distorts the data and leaks duplicate records between training and validation. It also makes every trial slower, since each trial trains on twice as much data. - D is wrong because faster instances trim each trial without changing the approach. It still runs 100 trials, at higher cost per hour, for a dataset that suits ensembling.
Reference: Training modes and algorithm support - Amazon SageMaker Autopilot