Posted On: Jul 15, 2022
Amazon SageMaker Automatic Model Tuning enables you to find the most accurate version of a machine learning (ML) model by finding the optimal set of hyperparameter configurations for your dataset. SageMaker Automatic Model Tuning now supports increased limits for two service quotas, with up to 50% higher number of total training jobs that can be run per tuning job and maximum number of hyperparameters that can be searched per tuning job.
Starting today, you can now run up to 750 training jobs in total as part of single tuning job, which is 1.5 times the previous default limit of 500 when either the “Bayesian” or “random” search method is used. The ability to run an increased number of total training jobs per tuning job enables exploring more hyperparameter combinations, and it helps to optimize the tradeoff between wall-clock time, predictive performance, and overall cost. When more combinations are explored, the chances of finding high quality hyperparameter configurations are increased, and thus, it can improve the quality of ML model. Additionally, through a limit increase request using the AWS Support Center, for the “random search” strategy SageMaker Automatic Model Tuning will continue to support exploring up to 10,000 hyperparameter configurations.
Furthermore, up to 30 hyperparameters can now be tuned for any search strategy, 1.5 times the previous limit of 20. Such higher number of hyperparameters allows using SageMaker Automatic Model Tuning for use cases such as Neural Architecture Search, which typically requires a larger number of hyperparameters to be tuned.
Increased limits for SageMaker Automatic Model Tuning are now available in all commercial AWS Regions and applicable to all tuning jobs. You can find the new limits in the resource limits page and the list of Amazon SageMaker default quotas in the service quotas page. You can launch SageMaker Automatic Model Tuning jobs with higher limits in AWS Console, using AWS SDK or Sagemaker SDK. To learn more, please visit SageMaker Automatic Model Tuning technical documentation.