What's changed: In-scope coverage: model dev/AI services
2.1Model Selection and Training
Understand three model-building approaches on SageMaker (built-in algorithms, script mode/bring-your-own container, AutoML/JumpStart) and training-job basics. The starting point for "ML Model Development" in MLA-C01.
Once data is ready, you select and train a model. SageMaker offers several approaches with different effort and flexibility.
2.1.1Build approaches
- Built-in algorithms: pass data to AWS-provided algorithms like XGBoost; least effort.
- Script mode / BYO container: train with your own code (PyTorch/TensorFlow/scikit-learn); most flexibility.
- AutoML (Autopilot) / JumpStart: auto-search the best model, or use pretrained/foundation models; fast start.
- Training jobs: run training by specifying instance type, data input, and hyperparameters.
Common on MLA: least effort, strong on tabular = built-in XGBoost, your own framework = script mode/BYO container, auto model search = Autopilot, pretrained/foundation models = JumpStart.
Choose a build approach by the "effort-vs-flexibility trade-off." Built-in algorithms (XGBoost, linear learner, k-means, etc.) just take data and are especially strong on tabular data with minimal effort. Script mode / BYO container trains with your own code (PyTorch/TensorFlow/scikit-learn) for maximum flexibility (good for migrating assets). AutoML (Autopilot) auto-searches models and hyperparameters; JumpStart quickly uses pretrained/foundation models. Training runs as a training job, specifying instance type, input (File/Pipe/FastFile input modes), and hyperparameters, with outputs saved as model artifacts in S3. Scale with distributed training (data/model parallel) and cut cost with managed spot training (up to ~90% off, with checkpoints for interruptions). Match the algorithm to the problem type (classification/regression/forecasting/recommendation/vision/NLP), and start from a strong baseline like XGBoost.
| Situation | Best approach |
|---|---|
| Tabular, least effort | Built-in (XGBoost, etc.) |
| Train with your own framework | Script mode / BYO container |
| Auto model search | AutoML (Autopilot) |
| Use pretrained/foundation models | JumpStart |
Scenario: quickly build a tabular churn-prediction model. Start with built-in XGBoost for a strong baseline (data in S3, a training job). Cut training cost with managed spot + checkpoints. If you need more accuracy, try custom features or another framework via script mode, or hand off the search to Autopilot.
Q. Tabular, least effort? Built-in XGBoost. Q. Own PyTorch code? Script mode / BYO container. Q. Auto model search? Autopilot. Q. Pretrained/foundation models? JumpStart. Q. Cut training cost? Managed spot + checkpoints.
Watch the mix-ups: (1) Autopilot = auto model search / JumpStart = pretrained/foundation models—don’t swap. (2) Managed spot can be interrupted—checkpoints are essential. (3) Built-in algorithms still have input-format (RecordIO/CSV) and hyperparameter requirements. (4) Full-loading huge data every time is slow—consider Pipe/FastFile modes.
For large training, distributed training and Spot (managed spot training) cut cost; use checkpoints to handle interruptions.
2.1.2Section summary
- Choose built-in / script·BYO / AutoML·JumpStart
- Cut cost with managed spot + checkpoints
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. You want a high-performing classification/regression model on tabular data with least effort. Which fits best?
Q2. You want to train your own PyTorch code on SageMaker. Which fits best?
Q3. You want to auto-search the best model and hyperparameters with little code. What do you use?