''Hot Dogs or Not" - At Scale with Kubernetes [I]
summary
Portable ML pipeline that runs unchanged from laptop minikube to GKE with GPUs, packaged as cube flow components deployable via kubectl. Includes JupyterHub, a Tensorflow operator and Tensorflow Serving wired together; demo trains ResNet on synthetic data: local laptop ~0.5 images/sec, GPU cluster ~240 images/sec. Declarative training spec creates master/parameter servers/workers, autoscaling and priority eviction are supported, model serving points at a GCS-hosted model, and the demo mostly succeeds.
AI-written summary · about
chapters (14)
- 0:00Introduction and audio check
- 0:35Introducing ML on Kubernetes
- 1:45What is machine learning?
- 3:19ML use case: Data centers
- 4:29The need for ML DevOps
- 6:55Kubernetes as an infrastructure
- 14:34Introducing Kubeflow
- 15:29Kubeflow components
- 16:42Live demo: Local JupyterHub
- 20:24Demo: Portability to cloud
- 22:03TensorFlow operator demo
- 24:39Distributed training demo
- 26:36Serving and conclusion
- 31:33Q&A session
related talks
01h 24mTensorFlow: Hotdog or Not with Toppings #bbuzz
26mMachine Learning Meets K8s
42mKubernetes for Machine Learning: Managing and Scaling AI Workloads
54mEconomics and Best Practices of Running AI/ML Workloads on Kubernetes
46mDeploying scalable Machine Learning models with Kubernetes
32mGPU Machine Learning From Laptop to Cloud
31mKubeflow Intro – Michał Jastrzębski & Ala Raddaoui, Intel
36mAccelerating High-Performance Machine Learning at Scale i...