More Related Content Similar to Kruize-HPO-meetup-Aug2022.pdf (20) Kruize-HPO-meetup-Aug2022.pdf2. 2
Agenda
➢ Challenges / Motivation
➢ Introduction to Bayesian Optimization / HPO
➢ HPO Use cases
➢ HPO Demo
➢ Experiment results
5. 5
Building an ML Model?
Fix the Hyperparameters first!
⚫ Model architecture
⚫ Learning rate
⚫ Number of epochs
⚫ Number of branches in a decision tree
⚫ Number of clusters in a clustering algorithm
6. 6
Fixing Hyperparameters - Options
⚫ Manual Search
⚫ Random Search
⚫ Grid Search
⚫ Automated Hyperparameter Tuning - Bayesian Optimization
⚫ ...
12. Kruize HPO vs Default Config
[Obj Fn = maximize(1.25*Throughput / 1.5*Response_time / 0.25*Max_response_time),
Fixed Resources]
Default
14.21 ms
2.39 ms
12
13. Kruize HPO vs Default Config
[Obj Fn = maximize(1.25*Throughput / 1.5*Response_time / 0.25*Max_Response_Time),
Fixed Resources]
83% better response time 1.3% better throughput
13
14. Interested in HPO?
⚫ HPO - https://github.com/kruize/hpo
⚫ Kruize Demos - https://github.com/kruize/kruize-demos
Call for collaboration !
Kruize Slack
kruize@redhat.com
14