kubeflow/katib

What it solves

Kubeflow Katib is a Kubernetes-native project for automated machine learning (AutoML), designed to automate the process of finding the best hyperparameters, neural architectures, and early stopping rules for machine learning models. It removes the manual effort of tuning models to achieve optimal performance.

How it works

Katib operates as a control plane on Kubernetes, acting as an agnostic layer that can tune hyperparameters for applications written in any language. It integrates with various ML frameworks like TensorFlow, PyTorch, and XGBoost, and can execute training jobs using Kubernetes Custom Resources, including the Kubeflow Training Operator, Argo Workflows, and Tekton Pipelines.

Who it’s for

It is primarily for data scientists and ML engineers who need to scalable, cloud-native hyperparameter tuning and architecture search on Kubernetes.

Highlights

• Broad Algorithm Support: Supports a wide range of search algorithms including Random Search, Grid Search, Bayesian Optimization, TPE, CMA-ES, and HyperBand.
• HPO and NAS: Provides capabilities for Hyperparameter Tuning, Neural Architecture Search (ENAS, DARTS), and Early Stopping (Median Stop).
• Framework Agnostic: Works with any ML framework and any language, leveraging existing tools like Optuna, Hyperopt, and Scikit Optimize.
• Python SDK: Includes a Python SDK to simplify the creation and management of tuning jobs.