Question

A data scientist is training a deep learning model in a SageMaker training job and wants to systematically compare multiple runs that use different learning rates. For each run, they need to capture per-epoch training and validation loss so the runs can be visualized side by side in SageMaker Studio, and they want each run grouped under a single logical project. Which approach best meets these requirements with the least custom infrastructure?

Accepted Answer

Create a SageMaker Experiment and log each training job as a run using the SageMaker experiments SDK, calling log_metric to record loss values at each epoch.

Answer

Write per-epoch loss values to a CSV file in Amazon S3 for each job and build a QuickSight dashboard to compare the runs.

Answer

Enable SageMaker Model Monitor on the training jobs to capture loss metrics and compare them in the Model Registry.

Answer

Print loss values to stdout so they appear in CloudWatch Logs, then manually export and chart them in a spreadsheet.