Skip to content

Latest commit

 

History

History
103 lines (71 loc) · 4.49 KB

ROADMAP.md

File metadata and controls

103 lines (71 loc) · 4.49 KB

Katib 2022/2023 Roadmap

AutoML Features

  • Support advance HyperParameter tuning algorithms:

    • Population Based Training (PBT) - #1382
    • Tree of Parzen Estimators (TPE)
    • Multivariate TPE
    • Sobol’s Quasirandom Sequence
    • Asynchronous Successive Halving - ASHA
  • Support multi-objective optimization - #1549

  • Support various HP distributions (log-uniform, uniform, normal) - #1207

  • Support Auto Model Compression - #460

  • Support Auto Feature Engineering - #475

  • Improve Neural Architecture Search design

Backend and API Enhancements

  • Conformance tests for Katib - #2044
  • Support push-based metrics collection in Katib - #577
  • Support PostgreSQL as a Katib DB - #915
  • Improve Katib scalability - #1847
  • Promote Katib APIs to the v1 version
  • Support multiple CRD versions (v1beta1, v1) with conversion webhook

Improve Katib User Experience

  • Simplify Katib Experiment creation with Katib SDK - #1951
  • Fully migrate to a new Katib UI - Project 1
  • Expose Trial logs in Katib UI - #971
  • Enhance Katib UI visualization metrics for AutoML Experiments
  • Improve Katib Config UX - #2150

Integration with Kubeflow Components

  • Kubeflow Pipeline as a Katib Trial target - #1914
  • Improve data passing when Katib Experiment is part of Kubeflow Pipeline - #1846

History

Katib 2021 Roadmap

New Features

AutoML

  • Support Population Based Training #1382
  • Support ASHA
  • Support Auto Model Compression #460
  • Support Auto Feature Engineering #475
  • Various CRDs for HP, NAS and other AutoML techniques.

UI

  • Migrate to the new Katib UI Project 1
  • Hyperparameter importances visualization with fANOVA algorithm

Enhancements

  • Finish AWS CI/CD migration
  • Support various parameter distribution #1207
  • Finish validation for Algorithms #1126
  • Refactor Hyperband #1389
  • Support multiple CRD version with conversion webhook
  • MLMD integration with Katib Experiments

Katib 2020 Roadmap

New Features

Hyperparameter Tuning

  • Support Early Stopping #692

Neural Architecture Search

  • Support Advanced NAS Algorithms like DARTs, ProxylessNAS #461

Other Features

  • Support Auto Model Compression #460
  • Support Auto Feature Engineering #475

Enhancements

Common

  • Delete Suggestion deployment after Experiment is finished #1061
  • Save Suggestion state after deployment is deleted #1062
  • Reconsider the design of Trial Template #906
  • Design an extensible model for integrating with custom resources.
  • Add validation for algorithms (a.k.a suggestions) #1126
  • Katib UI fixes and enhancements
  • Investigate Kubeflow Metadata integration
  • Investigate the alignment with concept and implementation of "experiments" and "jobs/runs" in KFP #4955

Neural Architecture Search

  • Refactor structure for NAS algorithms #1125
  • Refactor the design for NAS model constructor #1127
  • ENAS enhancements such as micro mode, RNN support