Skip to content

Katib Output

Katib Execution

During execution Katib will run an experiment to identify the best hyperparameter values as per the objective and associated function (maximize or minimize) specified when setting up the Katib job. The output of each of this experiment as well as the performance details and the overall best run are all available via the Kubeflow UI. Since Katib is being facilitated by Kale you can confirm the pipeline created by clicking VIEW.


Experiments (KFP)

The Experiments (KFP) page shows the experiments that have been run in KubeFlow pipelines with the names based on the Kale and Katib inputs from the JupyterLabs Notebook. Selecting an individual experiment will provide a list of the individual runs that comprise the experiments, keeping in mind that this number is limited based on values provided in the Run Parameters section. The page also displays at the end of each run the value achieved for the Search Objective.


Experiments (AutoML)

The Experiments (AutoML) page shows the list of experiments.


Selecting an expirement will show greater detail about the trial runs and the associated hyperparameters that Katib is attempting to optimize. The page opens to a graph showing the experiments and the hyperparameter values considered.


Scrolling down shows the number of runs and the breakdown of successful and failed runs as well as the achieved value of the optimal Search Objective.


Ideal Model

The ideal model hyperparameters based on the Katib tuning using the prescribed Search Algorithm and the associated Search Metrics are available through either Experiments UI. As a good practice we recommend taking the optimized hyperparameters and using them in the production models that you will serve based on the pipelines that are built with Kale & Katib.