katib

Why we included this project

Katib is a Kubernetes-native AutoML tool that runs hyperparameter tuning, early stopping, and neural architecture search as ordinary cluster workloads. It stays agnostic to the ML framework, so a trial can be a PyTorch or TensorFlow model or a job defined as an Argo Workflow or Tekton pipeline, and it supports custom resources out of the box. Teams that already standardize on Kubernetes for ML get experiment tuning inside the cluster rather than bolting on a separate tool. Katib also bundles several search algorithms, so you are not stuck with one tuning strategy when different models call for different approaches.