Hyperparameter tuning should not be part of your ML code

Original article was published by Dimitris Poulopoulos on Artificial Intelligence on Medium

From Jupyter Notebooks to HP Tuning

In this article, I follow a different approach. Instead of showing you how to set up your workspace first, we dive directly into the task at hand: we have a Jupyter Notebook that’s training an ML model, and we want to find the optimal set of the learning algorithm’s hyperparameters automatically, without writing a single line of code. Later, I will provide you with the details about how to do all these yourself.

That been said, we use Kubeflow and a tool called Katib. Katib is a Kubernetes-based system for HP tuning and Neural Architecture Search. Katib supports several ML frameworks, including TensorFlow, Apache MXNet, PyTorch, XGBoost, and others.

Also, for this example, we use the Titanic dataset and Kale (Kubeflow Automated pipeLines Engine). Kale is an open-source project that aims at simplifying the Data Science experience of deploying Kubeflow Pipelines workflows. So, let’s see how simple it is to run an HP tuning experiment from your Jupyter environment using Kale and Katib.

The project example is available on GitHub, but as we discussed setting up your environment comes later. For now, assuming we have a Notebook kernel running, all we have to do is open Kale from the left panel, enable it and set up a Katib job.

HP Tuning with Katib and Kale — Image by Author

Moreover, we see that, when enabled, Kale annotates the cells of the Notebook. Each annotation is a step in the ML Pipeline. You can play around and create your own steps, but don’t forget to set the dependencies of each cell also. In any case, the annotations that we care about are two:

  • The pipeline-parameters cell, that defines which are the hyperparameters in this Notebook
  • The pipeline-metrics cell in the end, which defines the metric that we try to optimize by changing the hyperparameters

The Katib UI will pick up the parameters automatically from the pipeline-parameters. All we got to do is choose the values that we would like to try for each parameter, the optimization algorithm and the metric we want to optimize. After, pressing the COMPILE AND RUN KATIB JOB button on the bottom, we can follow the link to watch the experiment live or wait to get the results via the Kale user interface. This is what the Katib experiment looks in the end.

Katib Experiment — Image by Author

Kale will return the best hyperparameter combination for our model. This is as easy as it gets!

How Can I Get All That?

So, let’s see how to set up our environment now. To get an instance of Kubeflow running, we will use MiniKF. To install it is pretty simple; all we need is a GCP account and the ability to deploy applications from the Marketplace.

MiniKF deployment — Image by Author
  1. Visit the MiniKF on GCP page
  2. Choose Launch, set the VM configuration and click deploy

For best performance, it is recommended to keep the default VM configuration

That’s it! The deployment takes up to ten minutes, and you can watch the progress by following the on-screen instructions; ssh into the machine, run minikf on the terminal and wait until you have your endpoint and credentials ready.

Provision of MiniKF completed — Image by Author

Now, we are ready to visit the Kubeflow Dashboard. Click on the URL, enter your credentials, and you’re ready to go!

Kubeflow dashboard — Image by Author

Running a Jupyter Server

Now, we need a Jupyter Notebook instance. Creating a Jupyter Notebook is relatively easy in MiniKF, using the provided Jupyter Web App:

  1. Choose notebooks from the left panel
  2. Choose the New Server button
  3. Fill in a name for the server and request the amount of CPU and RAM you need and click LAUNCH

To follow this tutorial leave the Jupyter Notebook image as is (jupyter-kale:v0.5.0-47-g2427cc9 — Note that the image tag may differ)

Creating a Jupyter Server — Image by Author

After completing these four steps, wait for the Notebook Server to get ready and connect. You’ll be transferred to your familiar JupyterLab workspace. To get the example of the story, create a new terminal in the JupyterLab environment and clone the following repo:

git clone https://github.com/dpoulopoulos/medium.git

You can find the Titanic example in medium > minikf > titanic-katib.ipynb. If you want to see more things you can do with MiniKF and Kale, check out the following stories.


In this story, we saw how to decouple our model training pipelines from the HP Tuning process. We demonstrated how we can use MiniKF to create a single node Kubeflow instance. Then, we turn to Kale and Katib to automate the process of running HP Tuning experiments from a Jupyter Notebook.

We are now ready to let Katib handle HP Tuning for us and use state of the art techniques, like Bayesian Optimization and Neural Architecture Search (NAS).

About the Author

My name is Dimitris Poulopoulos and I’m a machine learning engineer working for Arrikto. I have worked on designing and implementing AI and software solutions for major clients such as the European Commission, Eurostat, IMF, the European Central Bank, OECD, and IKEA.

If you are interested in reading more posts about Machine Learning, Deep Learning, Data Science and DataOps follow me on Medium, LinkedIn or @james2pl on twitter.

Opinions expressed are solely my own and do not express the views or opinions of my employer.