FLAME Integration with Kubeflow Katib

[akramIOT]

Since FLAME is supporting workloads to run in a de-coupled manner with the data plane workload running in edge nodes, its important to support Hyper Parameter tuning for DL in an optimum manner. In this specific context, is there a plan to support any API level integration with MLOps tools like 1) Kubeflow and 2) If support for Kubeflow is there/planned in roadmap then can we have an API from FLAME control plane/API server to tweak/parametrize Kubelfow Katib component for HP tuning to improve model efficacy ?

[myungjin]

@akramIOT Thanks for sharing your ideas. Hyperparameter tuning and NAS are indeed important topics for DL. Having said that, integration with Kubeflow is not our priority in our roadmap. Also, we need to evaluate Katib more carefully before we set up a roadmap for Katib. If we have the roadmap, we will publicize it through this thread.

[akramIOT]

@akramIOT Thanks for sharing your ideas. Hyperparameter tuning and NAS are indeed important topics for DL. Having said that, integration with Kubeflow is not our priority in our roadmap. Also, we need to evaluate Katib more carefully before we set up a roadmap for Katib. If we have the roadmap, we will publicize it through this thread.

Thanks @myungjin