Hyperparameter Tuning Using Kubeflow
summary
Kubeflow’s Katib is a Kubernetes-native hyperparameter tuner and NAS engine that runs experiments as CRDs, spawns trials as containerized jobs, and stores metrics in a MySQL backend. It runs parallel trials (configurable) and supports TensorFlow, PyTorch, MXNet, plus pluggable suggestion services like Bayesian optimization and grid/random search. A demo shows about 100 trials improving digit-recognition accuracy from ~96% to 98–99% and a UI/CLI for inspecting per-trial parameters and metrics.
AI-written summary · about
chapters (20)
- 0:00Intro
- 1:09An Example: Digits Recognition with MNist
- 2:00What is Hyperparameter Tuning?
- 2:38Why is Hyperparameter Tuning Hard?
- 4:28How does Kubernetes Help?
- 7:06Introducing Kubeflow
- 9:24Concepts: Experiment
- 11:04Concepts: Suggestion
- 11:51Concepts: Trial
- 12:36Workflow for Hyperparameter Tuning
- 13:23System Architecture
- 14:22Demo: Setting Up an Experiment
- 16:08Demo: Configuring Search Space
- 18:35Demo: Viewing Trial Metrics
- 19:15Classical vs Automated Machine Learning
- 20:09Landscape of Automated Machine Learning
- 22:09Workflow for Neural Architecture Search
- 23:20What's Coming?
- 25:54How to Contribute?
- 32:49Demo: Viewing Experiment Results
related talks
21mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
36mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
51mMachine Learning using Kubeflow and Kubernetes
27mManaging Thousands of Automatic Machine Learning Experiments with...
29mKubeflow Trainer Observability: Real-Time Progress Tracking for Reliab...
49mTutorial: From Notebook to Kubeflow Pipelines to KFServing: the Data Science Odyssey
14mKubeflow: Where ML Meets Kubernetes
54mScaling Machine Learning Pipelines in Cloud