Question

A retail company retrains its SageMaker image-classification model every week as new product photos arrive. Full training from scratch on the entire growing dataset takes over 10 hours per run and costs are climbing. The data science team wants to continue improving the existing model using only the newly labeled images each week, without discarding what the model has already learned, while minimizing training time and cost. Which approach best meets these requirements?

Accepted Answer

Use SageMaker incremental training by passing the previous model artifacts as a channel input and training on the new data plus a sample of prior data.

Answer

Launch a new SageMaker Automatic Model Tuning job each week with a warm start of type IDENTICAL_DATA_AND_ALGORITHM.

Answer

Enable SageMaker managed spot training so the full retraining job costs less each week.

Answer

Store the dataset in the SageMaker Feature Store offline store to speed up each full retraining run.