Boston Housing Example

In the video below we will look at how we can set up a hyperparameter tuning job using the high level approach in SageMaker. To following along, open up the Boston Housing - XGBoost (Hyperparameter Tuning) - High Level.ipynb notebook in the Tutorials directory.

Generally speaking, the way to think about hyperparameter tuning inside of SageMaker is that we start with a base collection of hyperparameters which describe a default model. We then give some additional set of hyperparameters ranges. These ranges tell SageMaker which hyperparameters can be varied, with the goal being to improve the default model.

We then describe how to compare models, which in our instance is just by way of specifying a metric. Then we describe how many total models we want SageMaker to train.

Note: In addition to creating a tuned model in this notebook, we also saw how the attach method can be used to create an Estimator object which is attached to an already completed training job. This method is useful in other situations as well.

You will notice that throughout this module we train the same model multiple times. In most of the Boston Housing notebooks, for example, we train an XGBoost model with the same hyperparameters. The reason for this is so that each notebook is self contained and can be run even if you haven't run the other notebooks.

In your case however, you've probably already created an XGBoost model on the Boston Housing training set with the standard hyperparameters. If you wanted to, you could use the attach method to avoid having to re-train the model.