Better reproducibility with SageMaker Automatic Model Tuning
Amazon SageMaker Automatic Model Tuning now gives you the option to set the seed to generate random hyperparameters for more reproducible tuning results. This enables use cases where you need to be able to reproduce your tuning job results, such as for compliance or regulatory reasons.
Amazon SageMaker Automatic Model Tuning now gives you the option to set the seed to generate random hyperparameters for more reproducible tuning results. This enables use cases where you need to be able to reproduce your tuning job results, such as for compliance or regulatory reasons.
SageMaker Automatic Model Tuning allows you to find the most accurate version of your machine learning model by searching for the optimal set of hyperparameter configurations. Previously, running the same tuning job more than once could lead to different recommended hyperparameter configurations due to the stochastic nature of the search strategies. This meant that you would not always be able to reproduce your previous tuning results even when running a tuning job on the same algorithm, dataset and with the same configurations.
Starting today, you can specify an integer as a random seed for hyperparameter tuning to generate hyperparameters. When running the same tuning job again, you can use the same seed to produce hyperparameter configurations that are more consistent with your previous results. For the Random and Hyperband strategies, using the same random seed can give you up to 100% reproducibility of the previous hyperparameter configuration for the same tuning job. For the Bayesian strategy, using the same random seed will significantly improve reproducibility for the same tuning job.
The ability to specify a random seed is now available for Amazon SageMaker Automatic Model Tuning in all commercial AWS Regions. To learn more, review the technical documentation or visit SageMaker Automatic Model Tuning web page.