ML platform
SageMaker Debugger
Inspects and profiles model training jobs.
Key points
- Captures training tensors, metrics, and system data during SageMaker training jobs.
- Helps identify training problems such as poor convergence, overfitting signals, or resource bottlenecks.
- Supports built-in and custom rules to detect issues while training runs.
- Can profile resource utilization to improve training performance.
- Targets training-time visibility rather than deployed inference monitoring.
When to use it
- Choose Debugger when a training job needs troubleshooting or profiling.
- Use it when training is slow, unstable, or consuming resources inefficiently.
- Use it to inspect model training behavior before deployment.
Exam tips
- Debugger is for training jobs; Model Monitor is for deployed endpoints.
- Automatic Model Tuning searches hyperparameters; Debugger explains what is happening during training.
- Use Debugger when the clue is tensors, gradients, training rules, or profiling.
- Debugger does not replace evaluation on validation data.