Posted On: Nov 2, 2022

Amazon SageMaker Autopilot experiments using hyperparameter training are up to 2x faster to generate ML models on datasets greater than 100 MB running 100 or more trials. Amazon SageMaker Autopilot automatically builds, trains, and tunes the best ML models based on your data while allowing you to maintain full control and visibility. 

SageMaker Autopilot offers two training modes - Hyperparameter optimization (HPO) and Ensemble. In the HPO mode, SageMaker Autopilot selects the algorithms that are most relevant to your dataset and selects the best range of hyperparameters to tune your models using Bayesian optimization. However for larger datasets (> 100MB), the tuning time with Bayesian optimization can be longer. Starting today, SageMaker Autopilot will use a new multi-fidelity hyperparameter optimization (HPO) strategy that employs the state-of-the-art hyperband tuning algorithm on datasets that are greater than 100 MB with 100 or more trials while continuing to leverage the Bayesian optimization strategy for data sets lesser than 100MB. With the multi-fidelity optimization strategy, trials that are performing poorly against a selected objective metric are stopped early thereby freeing up resources for well performing trials. This in turn reduces the tuning time for HPO training mode SageMaker Autopilot experiments on large data sets. 

With this release the model training and tuning time is up to 2X faster than before enabling customers to deliver the best performing ML model sooner. To evaluate the performance improvements, we used multiple OpenML benchmark datasets with varying sizes ranging from 100 MB to 10 GB. Based on our results, moderately large datasets (100MB - 1 GB) saw 41% (from average 345 to 203 mins) and very large datasets (> 1 GB) saw a 48% improvement (from average 2010 to 1053 mins) runtime improvements respectively. With this enhancement, you can run your SageMaker Autopilot experiments faster without making any changes to existing job configurations.

For more information, see the documentation and to learn more about SageMaker Autopilot, visit the product page.