Hyperparameter tuning can quickly become a bottleneck in real-world machine learning operations, especially when multiple models, data sets, and iterations are involved. In this talk, we discuss how our data science team automated and sped up hyperparameter optimization, using Kubeflow Katib as a key part of our MLOps.
Before, experimentation in PepsiCo was slow, inconsistent, and not reproducible. When we standardized on Kubeflow Pipelines and Katib, we provided the ability for data scientists to spin-up repeatable, automated and resource-aware experiments right in Kubernetes without manual tuning or worrying about infrastructure.
We share how our PepsiCo DS team has implemented different search strategies in practice. In addition, we will address experiment tracking, search-space design, use of Katib in pipelines, and operational lessons learned when running Kubeflow Katib.