Accelerate deep learning in the cloud with custom ML environments
summary
Shows how to go from an EC2 Jupyter notebook to reproducible containers by using Deep Learning AMIs as bases, writing a Dockerfile that copies a CIFAR training script, building and pushing images to ECR, then running them on Amazon EKS and SageMaker. Demonstrations include GPU utilization around 60–80%, EC2 instances with up to 8 GPUs and 16 TB local storage, kubectl apply job submission, and SageMaker script-mode and custom-container estimators that successfully train and report billable seconds equal to training seconds.
AI-written summary · about
chapters (15)
- 0:00Introduction to custom ML environments
- 1:49Benefits of EC2 for ML development
- 3:17Demo of Jupyter on EC2 instance
- 6:11Challenges with single EC2 setup
- 8:40Why containers solve ML portability
- 12:17Demo of building custom containers
- 16:40Scaling ML with Kubernetes
- 19:54Kubernetes demo for training jobs
- 23:56Limitations of Kubernetes for ML
- 24:59Introducing Kubeflow on EKS
- 26:10Amazon SageMaker fully managed ML
- 28:18Demo of SageMaker script mode
- 30:50Demo of custom container in SageMaker
- 31:55Combining Kubernetes and SageMaker
- 33:24Recap and resources
related talks
44mScaling ML from 0 to millions of users
19mUsing Amazon SageMaker to develop and deploy ML models
12mLeveraging AWS and Kubernetes for Scalable AI/ML Workloads in the Cloud
19mDeveloping large-scale machine learning algorithms on Amazon SageMaker
34mDeploying a Machine Learning System That Only Runs Locally
18mKeynote: Containerizing Deep Learning Using Kubeflow on AWS
38mAll You Need to Know to Build Your GPU Machine Learning Cloud [B]
30mBuild consistent and portable ML environments with containers