Developing large-scale machine learning algorithms on Amazon SageMaker
summary
Amazon SageMaker built-in algorithms are implemented as Docker containers using MXNet (and other frameworks like TensorFlow and PyTorch), support file mode vs pipe mode streaming for data that won’t fit in memory, and expose tunable hyperparameters wired to SageMaker hyperparameter tuning. Performance levers discussed include mixed precision and larger batch sizes, synchronous ring all‑reduce vs asynchronous parameter server, EC2 P3 instances and EBS limits (16 TB), and a tuning run of six jobs (two concurrent) that finds alpha ≈ 0.67.
AI-written summary · about
chapters (7)
- 0:00Start
- 1:25Amazon SageMaker: Algorithms
- 3:00Interface Design
- 7:51System design: Storage
- 15:16Create Algorithm: EMA example
- 15:48AWS Marketplace for Machine Learning
- 16:43Hyperparameter tuning job: EMA example
related talks
19mUsing Amazon SageMaker to develop and deploy ML models
30mChoose the right machine learning algorithm in Amazon SageMaker
31mTrain and tune ML models to the highest accuracy using Amazon SageMaker
44mPragmatic Machine Learning in the Cloud
39mTrain ML models at scale with Amazon SageMaker Training
37mStreamlining Machine Learning Operations
20mIndustrial-Scale Machine Learning
35mAccelerate deep learning in the cloud with custom ML environments