Question

A data scientist is configuring a SageMaker automatic model tuning job for a deep learning model. They want the tuner to explore learning rate values that span several orders of magnitude, from 0.0001 to 0.1. They notice that with a default linear parameter scaling, the tuner spends most of its trials sampling large learning rate values and rarely tests the small ones. How should they configure the learning rate hyperparameter range to sample more evenly across all magnitudes?

Accepted Answer

Define the learning rate as a ContinuousParameter and set its scaling type to Logarithmic.

Answer

Define the learning rate as an IntegerParameter with a linear scaling type.

Answer

Define the learning rate as a ContinuousParameter with the default Auto scaling and increase the maximum number of training jobs.

Answer

Define the learning rate as a CategoricalParameter listing only 0.0001 and 0.1.