kind-cluster

Katib Example with Kind Cluster

Follow this example to run Katib Experiment on your local laptop with Kind cluster. This example doesn't require any public or private cloud to run Katib Experiments.

Prerequisites

Install the following tools to run the example:

• Docker >= 20.10
• Kind >= 0.13
• kubectl >= 1.27

Installation

Run the following command to create Kind cluster with the Katib components:

./deploy.sh

If the above script was successful, Katib components will be running:

$ kubectl get pods -n kubeflow

NAME                                READY   STATUS      RESTARTS   AGE
katib-controller-566595bdd8-x7z6w   1/1     Running     0          67s
katib-db-manager-57cd769cdb-x4lnz   1/1     Running     0          67s
katib-mysql-7894994f88-7l8nd        1/1     Running     0          67s
katib-ui-5767cfccdc-nt6mz           1/1     Running     0          67s

Run Katib Experiment

You can use various Katib interfaces to run your first Katib Experiment.

For example, create Hyperparameter Tuning Katib Experiment with random search algorithm using kubectl:

kubectl create -f https://raw.githubusercontent.com/kubeflow/katib/master/examples/v1beta1/hp-tuning/random.yaml

This example uses a PyTorch neural network to train an image classification model using the MNIST dataset. You can check the training container source code here. The Experiment runs twelve training jobs (Trials) and tunes the following hyperparameters:

• Learning Rate (lr).
• Momentum (momentum).

After creating above example, check the Experiment status:

$ kubectl get experiment random -n kubeflow

NAME     TYPE      STATUS   AGE
random   Running   True     6m19s

Check the Suggestion status:

$ kubectl get suggestion -n kubeflow

NAME     TYPE      STATUS   REQUESTED   ASSIGNED   AGE
random   Running   True     4           4          6m21s

Check the Trials statuses:

$ kubectl get trial -n kubeflow

NAME              TYPE        STATUS   AGE
random-9hmdjqk9   Running     True     99s
random-cf7tfss2   Succeeded   True     5m21s
random-fr5lfn2x   Running     True     5m21s
random-z9wqm7xh   Running     True     5m21s

You can get the best hyperparameters with the following command:

$ kubectl get experiment random -n kubeflow -o jsonpath='{range .status.currentOptimalTrial.parameterAssignments[*]}{.name}: {.value}{"\n"}{end}'

lr: 0.028162244250364066
momentum: 0.583672196492823

To view created Experiment in Katib UI, follow this guide.

Cleanup

To cleanup Kind cluster run:

kind delete cluster