Last week we trained models using SageMaker's built-in algorithms. But in practice, you often need to train models written directly in PyTorch or TensorFlow, or train in your own environment bundling company standard libraries. SageMaker offers three custom training modes depending on "how much of my code do I bring".
In the MLA-C01 exam, this topic appears in scenarios like "I want to bring framework code with minimal changes", "I need special dependencies·system packages". Today we clarify the boundaries of three axes: script mode, BYOC, and framework containers.
SageMaker training divides by "how much of AWS's provided container you keep as-is":