Posted On: Aug 16, 2022

Amazon Rekognition Custom Labels is an automated machine learning (AutoML) service that allows customers to build custom computer vision models to detect objects and scenes specific to their business without in-depth machine learning expertise. Starting today, Custom Labels can automatically scale inference units of a trained model based on customer workload. This reduces model inference cost as customers no longer need to over-provision inference units to support spiky or fluctuating image volumes.

Previously, Custom Labels customers with unpredictable workloads had to set minimum inference units to support the peak volume of images they expected to process. This resulted in higher costs as the minimum inference units were consumed even if volumes were lower or absent. With autoscaling support, customers can now set both minimum and maximum inference units. Custom Labels dynamically adjusts inference units up or down within the specified minimum and maximum inference units based on image volumes. Customers are only charged for the inference units they consume. Note that the minimum inference unit allowed is 1. As an example, say a customer specifies a minimum inference unit of 1 and maximum inference units of 5. If the customer workload consumed 5 inference units for 5 hours and 1 inference unit for the rest of the day, customer is now only charged $176 (19 hours x $4 per hour x 1 Inference unit + 5 hours x $4 per hour x 5 Inference units). Without autoscaling, the customer would have been charged $480 (24 hours x $4 per hour x 5 Inference units).

This feature is available in all supported Amazon Rekognition Custom Labels regions. To learn more about Amazon Rekognition Custom Labels, visit feature documentation. To get started, visit the Custom Labels console today.