Amazon SageMaker

Overview

• Managed service for training, evaluate and deploy models
• Provides EC2 infrastructure
  • Including GPU
• Compare: older Amazon Machine Learning

Training

• Built-in algorithms
  • Supervised, unsupervised
  • General-purpose, image processing, extracting topics from text
• Using Apache Spark (MLLib)
• Deep Learning - custom code
  • TensorFlow, MXNet
• Custom algorithm (Docker Image)
• Compute resources
  • Managed by AWS

Evaluation

• Online
  • Production variants
• Offline
  • holdout set
  • k-fold 

Deployment

• Auto Scaling Group of EC2 
• Exposes HTTPS endpoint
• Can be used for real-time queries
• Multiple models can be deployed at the same time
  • aka production variants
  • useful for testing