Overview

In this exercise, you will use a Katib experiment to manage and track your hyperparameter tunning for automated machine learning. Katib experiments help data scientist optimize hyperparameters when training machine learning models by integrating directly with Kubeflow.

---

The Challenge

Different sets of hyperparameters are needed in order to train a machine learning model. Running these different sets of hyperparemters through a model is referred to as hyperparemeter tunning. Hyperparameter tunning can be manual and time consuming. Katib simplifies hyperparameter tuning by automating and tracking the results of each trial run that uses a different set of hyperparameters.

Watch this video showing a demo of this exercise.

---

Experiment Description

The objective of this experiment is to maximize the output of the function by increasing the value of a and decreasing the value of b.

You will use a Jupyter notebook in Kubeflow to perform the setup and initiate the Katib experiment.

The notebook will first define the objective function,$F(a,b)=4a-b^2$, which will be used for the hyperparameter tunning.

The notebook will then define the valid range of values for both hyperparameters a and b

Next, the notebook will create the Katib experiment in Kubeflow and run the hyperparemeters through the function, keeping track of each trial's results.

Finally, the most optimal hyperparameters will be selected and the results of all trials will be viewable within the Katib UI.

---