Sources of ML Models

Four ways to get a modelCopy Anchor Link

The story: You need a cake for a party. You can:

1. Buy a finished cake, and maybe add your own decorations.
2. Use the bakery's own tested cake mixes, made to scale up to huge batches.
3. Bring your own recipe and bake it in the bakery's kitchen, with their standard ovens.
4. Bring your own recipe and your own custom oven.

Each step is more work, but gives you more control.

In AI/AWS terms: SageMaker AI supports four ways, from least to most effort:

1. Pre-trained models in SageMaker JumpStart: ready to deploy, or fine-tune and deploy. Least effort and lowest operational overhead.
2. Built-in algorithms: SageMaker's own algorithms, designed to scale to large datasets and significant compute.
3. Script mode with supported frameworks: your own training code on pre-made images for scikit-learn, TensorFlow, PyTorch, MXNet, or Chainer (the bakery's ovens).
4. Custom Docker images: bring your own container with any packages you need (your own oven). Most effort.

SageMaker JumpStartCopy Anchor Link

The story: The bakery's display counter: famous cakes from well-known bakers, which you can buy as is, adjust, or taste first. Next to it are party kits with plates and decorations already set up, and recipe cards you can follow step by step.

In AI/AWS terms:

• Pre-trained open-source models from popular model hubs for many problem types
• Deploy, fine-tune, and evaluate them, with incremental training before deployment
• Solution templates that set up infrastructure for common use cases (the party kits)
• Runnable example notebooks (the recipe cards)

What the built-in algorithms coverCopy Anchor Link

The story: The bakery's own cake mixes come in many flavors, one for each common kind of occasion.

In AI/AWS terms: Built-in algorithms cover supervised learning (classification and regression), unsupervised learning (clustering, dimension reduction, anomaly detection), image processing, time series, and text analysis.