MLCOST-18: Use warm-start and checkpointing hyperparameter tuning
Where feasible, use warm start hyperparameter tuning. Warm start can consist of using a parent job for a model trained previously or using transfer learning. Warm start of hyperparameter tuning jobs eliminates the need to start a tuning job from scratch. Create a new hyperparameter tuning job that is based on selected parent jobs or pre-trained models. Use checkpointing capabilities to restart a training job from the last saved checkpoint. Reuse previous trainings as prior knowledge, or use checkpointing to accelerate the tuning process and reduce the cost.
Implementation plan
-
Use warm-start hyperparameter tuning - Use warmstart to start a hyperparameter tuning job using one or more previous tuning jobs as a starting point. The results of previous tuning jobs are used to inform which combinations of hyperparameters to search over in the new tuning job. Hyperparameter tuning uses Bayesian or random search to choose combinations of hyperparameter values from ranges that you specify.
-
Use checkpointing hyperparameter tuning - Use checkpoints in Amazon SageMaker AI to save the state of ML models during training. Checkpoints are snapshots of the model and can be configured by the callback functions of ML frameworks. You can use the saved checkpoints to restart a training job from the last saved checkpoint.
Blogs
Videos
Examples
