katib/examples/v1beta1

Katib Examples

Katib is an open source project which uses Kubernetes CRD to run Automated Machine Learning (AutoML) tasks. To know more about Katib follow the official guides.

This directory contains examples of Katib Experiments in action. To install Katib on your Kubernetes cluster check the setup guide. You can use various Katib interfaces to run these examples.

For a complete description of the Katib Experiment specification follow the configuration guide

Local Cluster Example

Get started with Katib Experiments from your local laptop and Kind cluster by following this example.

AutoML Algorithms

The following examples show various AutoML algorithms in Katib.

Hyperparameter Tuning

Check the Hyperparameter Tuning Experiments for the following algorithms:

• Random Search

• Grid Search

• Bayesian Optimization

• Tree of Parzen Estimators (TPE)

• Multivariate TPE

• Covariance Matrix Adaptation Evaluation Strategy (CMA-ES)

• Sobol's Quasirandom Sequence

• HyperBand

• PBT

Neural Architecture Search

Check the Neural Architecture Search Experiments for the following algorithms:

• Efficient Neural Architecture Search (ENAS)

• Differentiable Architecture Search (DARTS)

Early Stopping

Improve your Hyperparameter Tuning Experiments with the following Early Stopping algorithms:

• Median Stopping Rule

Katib Python SDK Examples

To learn more about Katib Python SDK check this directory.

Resume Katib Experiments

You can use different resume policies in Katib Experiments. Follow this guide to know more about it. Check the following examples:

• Resume From Volume

• Resume Long Running Experiment

Metrics Collector

Katib supports the various metrics collectors and metrics strategies. Check the official guide to know more about it. In this directory you can find the following examples:

• File Metrics Collector

• Custom Metrics Collector

• Metrics Collection Strategy

Trial Template

You can specify different settings for your Trial template. To know more about it follow this guide. Check the following examples:

• Trial with ConfigMap Source

• Trial with Metadata Substitution

Trial Images

Check the following images for the Trial containers:

• Tensorflow MNIST with summaries

• PyTorch MNIST

• ENAS Keras CNN CIFAR-10

• DARTS PyTorch CNN CIFAR-10

• PBT proof of concept

Katib with Kubeflow Training Operator

Katib has out of the box support for the Kubeflow Training Operators to perform Trial's Worker job. Check the following examples for the various distributed operators:

• TFJob MNIST with summaries

• PyTorchJob MNIST

• XGBoostJob LightGBM

• MPIJob Horovod

Katib with Kubeflow Pipelines

To run Katib with Kubeflow Pipelines check these examples.

Katib with Argo Workflows

To know more about using Argo Workflows in Katib check this directory.

Katib with Tekton Pipelines

To know more about using Tekton Pipelines in Katib check this directory.

FPGA Support in Katib Experiments

You can run Katib Experiments on FPGA based instances. For more information check these examples.