OptimizationJob CRD for HPO: The Evolution of the Katib Project

[akshaychitneni]

This issue tracks work on adding OptimizationJob CRD support as discussed in https://github.com/kubeflow/sdk/tree/main/docs/proposals/46-hyperparameter-optimization#potential-api-for-optimizationjob-crd that focuses on hyperparameter optimization (HPO) for TrainJobs

It should include -

• CRD specifically for hyperparameter optimization of TrainJobs
• Integration with TrainJob API
• Support for model/dataset initialization shared across trials
• Support push-based metrics collection via SDK
• Integration with SDK's OptimizerClient API

Design Document

https://docs.google.com/document/d/1Y8IJ-UdZ7VCEAlax_xEFbbqEi7EB6SfIX4D7ua-xn4M/edit

cc @andreyvelich @kramaranya

[github-actions]

🎉 Welcome to the Kubeflow Katib repo! 🎉

Thanks for opening your first issue! We're happy to have you as part of our community 🚀

Next steps:

• Our team will review your issue soon! cc @kubeflow/wg-automl-leads
• Check out the Contributing Guide and the Kubeflow Contributor Guide
• Join the Kubeflow Slack channels: https://www.kubeflow.org/docs/about/community/#kubeflow-slack-channels
• Join the AutoML & Training WG meetings: https://bit.ly/2PWVCkV

Feel free to ask questions in the comments. Thanks again for contributing! 🙏

[Krishna-kg732]

Hi @andreyvelich i would like to take a go at this ,
I will start with an initial proposal.
/assign

[aniket2405]

Hi @akshaychitneni and @andreyvelich,

I see that SDK PR (kubeflow/sdk#124) was successfully merged into kubeflow/sdk, which establishes the OptimizerClient and currently orchestrates the jobs via Katib Experiments. That's great.

I can help tackle the OptimizationJob CRD backend implementation. To kick things off on the controller side, I’ve drafted the initial Go API types (OptimizationJob, OptimizationJobSpec, Objective, etc.) mapped to the KEP-46 YAML proposal.

I'd like to open a PR to establish these v1alpha1 Go types.

[aniket2405]

@akshaychitneni @andreyvelich
I've created a draft PR for you to go through once.
We can further discuss the requirements ahead.

[andreyvelich]

Thank you for your help, that sounds great!
We've been discussing with @akshaychitneni how we can move forward, and what features we want to support in the OptimizationJob. Maybe we can try to talk about it on Wed if you are available @aniket2405 (Kubeflow SDK call at 8am PST).

https://docs.google.com/document/d/1jH2WAX2ePxOfI4JuiVK9nPlesDMiyg67xzLwhpR7wTQ/edit?tab=t.0

[aniket2405]

Sure, @andreyvelich
That sounds great.
I'll join the Kubeflow SDK call, and we can discuss how to move forward.

[Rahuldrabit]

Hi @akshaychitneni and @andreyvelich,
I have work on network architecture search and hyperparameter tunning and optimisation of neural network .
I just wanted to ask that , is this project is Deep Learning or Machine Learning oriented ? is that true i am very interested in this project .
If you want I can share result and project for review .

[aniket2405]

Design doc for this effort: http://bit.ly/4aGAWsu
Will be converted to KEP once the features are finalized.
Let me know how do we want to discuss about the features here.
cc: @akshaychitneni @andreyvelich