Kubeflow: TensorFlow on Kubernetes
summary
Kubeflow v0.3 is deployed on Kubernetes using ksonnet to generate manifests and Ambassador plus Istio for routing, providing a JupyterHub notebook environment. Distributed TensorFlow runs via TFJob CustomResources that launch GPU-backed Docker containers with master/worker/ps topology, persisting models to PVC/blob storage and serving them with TF Serving; hyperparameter sweeps spawn multiple TFJobs (e.g., a 3×3 matrix creating nine jobs).
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:55ML lifecycle challenges
- 3:20What is Kubeflow
- 5:46Kubeflow architecture
- 7:09Deploy Kubeflow with KSonnet
- 8:37Kubeflow components in action
- 10:03Jupyter Notebooks
- 11:49Training models with TFJobs
- 13:33Model management with volumes
- 14:35Distributed model training
- 16:19Distributed jobs with TFConfig
- 18:27Hyperparameter optimization
- 20:31Deploy and test models
- 21:55TensorFlow Serving
- 24:01Model operations
- 26:09Monitoring tools
- 27:15Conclusion
related talks
36mKubeflow: Tensorflow on Kubernetes
35mKubeflow: Data Scientist ready?
46mKubeflow, TensorFlow on Kubernetes
21mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
14mKubeflow: Where ML Meets Kubernetes
22mThe Containerization of Machine Learning: TensorFlow: Kubernetes, and Kubeflow
31mKubeflow - Data Science on Steroids
31mKubeflow Intro – Michał Jastrzębski & Ala Raddaoui, Intel
