Issue #, if available: N/A - but hope to tackle #26
Description of changes:
As the MLOps maturity of the average business grows, more users are keen to be introduced to more advanced features like SageMaker Feature Store and SageMaker Model Registry even in their first steps with SageMaker.
Built-in algorithms have also moved on, with AutoGluon-Tabular ensembling showing compelling accuracy gains vs XGBoost+HPO on benchmark datasets.
This PR attempts to incorporate these features into our initial built-in algo exercise.
Known limitations:
This increases the complexity of the first hands-on, gearing it more towards demonstrating "art of the possible" and less towards tightly focussed enablement on a specific topic. Not yet clear whether it'd significantly affect timing of the workshop.
AutoGluon model deployment requires re-packing the trained model.tar.gz to add the inference.py script, which unfortunately introduces an extra ~5min wait on both .deploy() and .register() calls
Need to wait for data to propagate from online Feature Store to offline, which can in theory take up to 15mins but is practically often ~5min. Currently this is handled by exploring the Feature Store and building the Autopilot model (from raw CSV) during that period.
By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.
Issue #, if available: N/A - but hope to tackle #26
Description of changes:
As the MLOps maturity of the average business grows, more users are keen to be introduced to more advanced features like SageMaker Feature Store and SageMaker Model Registry even in their first steps with SageMaker.
Built-in algorithms have also moved on, with AutoGluon-Tabular ensembling showing compelling accuracy gains vs XGBoost+HPO on benchmark datasets.
This PR attempts to incorporate these features into our initial built-in algo exercise.
Known limitations:
model.tar.gzto add theinference.pyscript, which unfortunately introduces an extra ~5min wait on both.deploy()and.register()callsBy submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.