MLA-C01 Practice Question 26: Tuning Warm Start Parent Job
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 3 · question 6 of 10
A completed SageMaker AI hyperparameter tuning job ran 100 training jobs and found a best validation RMSE the team wants to improve on. The next tuning job will use the same training data and the same training image, will widen the ranges of the two hyperparameters it already tunes (the count of static plus tunable hyperparameters is unchanged), and will run 100 more training jobs. The team wants the new job's search to be informed by the first job's results, and wants SageMaker AI to report the best training job across both the parent and the new job.
Show answer and explanation
Answer:
Warm start seeds a new tuning job with the results of up to five completed parent jobs, so the search starts from what the parent already learned instead of from scratch. IDENTICAL_DATA_AND_ALGORITHM is the type for exactly these conditions: same input data and same training image, with freedom to change hyperparameter ranges and the maximum number of training jobs. It is also the only warm start type for which DescribeHyperParameterTuningJob returns OverallBestTrainingJob, the best training job across the new job and all its parents (option C).
- A is wrong because an independent job, whatever its strategy, ignores the 100 results already paid for. Hyperband controls how budget is spent within one job; it does not carry knowledge across jobs. - B is wrong because re-submitting the definition starts a brand new search with no parent. It also cannot widen the ranges and still reuse the earlier results, and it re-pays for ground already covered. - D is wrong because TRANSFER_LEARNING is the type for when the data or the algorithm image changes between jobs. Here both are identical, and this type does not report the OverallBestTrainingJob across parent and child.
Reference: Run a Warm Start Hyperparameter Tuning Job