Help Wanted: Example code for StudyJob

[opus111]

This is not a bug, but perhaps missing documentation

I want to use do basic HyperParameter optimization with Katib. I did the tutorials, and read all the docs. Unfortunately, the tutorials just load existing Docker images. They do not explain how to implement your own StudyJob.

I could not find any example code that explains how a StudyJob interacts with Katib. How does it get its parameters? How does it report its results? What is the contract? I guess you would call this the StudyJob API

Any pointers to code or docs would be appreciated

Thank you
Peter

[johnugeorge]

Please follow the doc at https://www.kubeflow.org/docs/components/hyperparameter/

[johnugeorge]

One sample example that is using random algorithm: https://github.com/kubeflow/katib/blob/master/examples/random-example.yaml

[opus111]

Thank you Johnu. Yes, those are the documents I read, and the instructions I followed.

Perhaps I am missing it, but nowhere in those docs are instructions for how to write your own StudyJob. They show how to install Katib, and then run several pre-built examples. It does not have any links to the code that created those examples, or documentation on how the code works.

There must be the equivalent of a StudyJob API which the Containers implement. Where do I find that?

[opus111]

OK. I found this...

https://github.com/kubeflow/katib/blob/master/docs/developer-guide.md

In which is says this...

Implement new suggestion algorithm
Suggestion API is defined as GRPC service at API/api.proto. You can attach new algorithm easily.

implement suggestion API

make k8s service named vizier-suggestion-{ algorithm-name } and expose port 6789
And to add new suggestion service, you don't need to stop components ( vizier-core, modeldb, and 
 anything) that are already running.

Unfortunately, there is no API/api.proto in the repository.

[YujiOshima]

@opus111 The api.proto is this
Thank you for pointing it. We should fix API/api.proto to pkg/api/api.proto.