Machine Learning at Scale with Kubeflow

En esta sesión revisaremos Kubeflow, con un caso práctico para que puedas entender como Kubeflow te puede ayudar a gestionar tu modelo de Machine Learning.
Durante la sesión veremos el ciclo completo del modelo, la edición con Jupyter, la creación de los contenedores de entrenamiento y predicción, la gestión del un A/B testing y el uso de herramientas como Tensorboard y la modificación de los parámetros del modelo.

Transcript

1. Machine Learning at Scale with Kubernetes @iolalla Israel Olalla
2. ML is great!
3. ML is hard!
4. Magical AI goodness Lots of pain Most folks
5. Ok, but why? • Set up from scratch • Integrate with existing legacy systems • Eventual need to migrate based on business needs DIY HOSTED* • First five minutes? YES. • Next five years … all of the above • … plus single cloud only! * #NotAllHostedSolutions
6. Haven’t we heard this story before?
7. Containers & Kubernetes
8. Cloud Native Apps
9. So...
10. We need Cloud Native ML!
11. Composability Portability Scalability
12. Building a model Building a model Composability
13. Building a model Data ingestion Data analysis Data transformation Data validation Data splitting Trainer Model validation Training at scale Logging Roll-out Serving Monitoring Composability
14. Portability
15. Model UX Tooling Framework Storage Runtime Drivers OS Accelerator HW Experimentation Portability
16. Model UX Tooling Framework Storage Runtime Drivers OS Accelerator HW Experimentation Portability
17. Experimentation Portability
18. Experimentation Portability
19. Training Experimentation Portability
20. Cloud Training Experimentation Portability
21. Pfft. That doesn’t matter to me.
22. Wrong.
26. Your laptop counts.
27. Cloud Training Experimentation Portability
28. Cloud Training Experimentation Portability
29. Demo
30. What’s in the box? Data ingestion Data analysis Data transformation Data validation Data splitting Logging Roll-out Building a model Trainer Model validation Training at scale Serving Monitoring
31. Katib
32. PyTorch # Initialize a ksonnet APP APP_NAME=my-kubeflow ks init ${APP_NAME} cd ${APP_NAME} # Install Kubeflow components ks registry add kubeflow github.com/kubeflow/kubeflow/tree/master/kubeflow ks pkg install kubeflow/core ks pkg install kubeflow/tf-serving ks pkg install kubeflow/tf-job # Deploy Kubeflow NAMESPACE=kubeflow kubectl create namespace ${NAMESPACE} ks generate core kubeflow-core --name=kubeflow-core --namespace=${NAMESPACE} ks apply default -c kubeflow-core ks pkg install kubeflow/pytorch
33. Platform Integrations • Cloud Native Security ◦ Data Exfiltration prevention ◦ Identity Aware Proxy • Cloud-specific Infrastructure ◦ Accelerators (GPUs, TPUs) ◦ Preemptible VMs • Kubernetes extensions ◦ Node Pools ◦ Auto scaling clusters
34. TPU Acceleration Number of Images (M) Images / Second Data Courtesy of RiseML Non-TPU 0.92 TPU 1.71 (Higher is better) Total Dollars Cost to Reach 75% Accuracy TPU $55 Non-TPU $129 (Lower is better)
35. Scalability • More accelerators (GPU, TPU) • More CPUs • More disk/networking • More skillsets (SWEs, data scientists) • More teams • More experiments
36. Composability Portability Scalability
37. You know what’s really good at composability, portability and scalability?
38. Containers & Kubernetes
39. Containers & Kubernetes except...
40. • Containers • Packaging • Kubernetes service endpoints • Persistent volumes • Scaling • Immutable deployments • GPUs, Drivers & the GPL • Cloud APIs • DevOps • ... First, become an expert in: Oh, you want to use ML on Kubernetes?
41. Cloud Training Experimentation Portability
42. Cloud Training Experimentation Portability
43. Cloud Training Experimentation Portability
44. Cloud Training Experimentation Kubeflow Portability
45. Bespoke solutions Cloud-specific, non-portable tech Forking of Kubernetes APIs Reminder… What you DIDN’T see
46. • 1000+ commits • 100+ Community contributors • 20+ Companies contributing, including: Momentum!
47. • Data science will touch EVERY industry. • We can’t ask people to become a PhD in statistics though. • How do WE help everyone take advantage of this transformation? It’s a whole new world
48. Kubeflow is open! Open comm- unity Open design Open source Open to ideas
49. website: https://kubeflow.org github: https://github.com/kubeflow/kubeflow slack: kubeflow (http://kubeflow.slack.com) twitter: @kubeflow iolalla@google.com Come Help!
50. Thank you!