Amazon Rekognition Custom Labels now guides customers to fix dataset related errors, enabling faster creation of a high quality custom inference API
Amazon Rekognition Custom Labels is an automated machine learning (AutoML) feature that allows customers to find objects and scenes in images, unique to their business needs, with a simple inference API. Customers can create a custom ML model simply by uploading labeled images. No ML expertise is required.
Amazon Rekognition Custom Labels is an automated machine learning (AutoML) feature that allows customers to find objects and scenes in images, unique to their business needs, with a simple inference API. Customers can create a custom ML model simply by uploading labeled images. No ML expertise is required.
Today, we are pleased to announce new tools that guide users on how to fix dataset related errors which cause training failures. Determining the root cause of a training failure is often tedious and time consuming. For example, customers with errors deep within their datasets may get an error such as “The manifest file contains too many invalid rows”. To find and fix the invalid rows in their training or testing dataset manifest file, a customer would need to manually review thousands of individual rows. With this update, Amazon Rekognition Custom Labels provides tools such as root cause analysis, granular error logs, and recommended fixes, so that customers can quickly pinpoint and fix dataset related errors, and successfully train their model. Customers can review errors and access detailed reports in the console and in the response from DescribeProjectVersions. New reports generated include a summary validation file, which details any aggregate errors and unmet dataset validations, and training and testing dataset manifests validation files, which include detailed row level errors. In the training and testing manifest validation files, error message are placed within specific rows where the error occurs. This removes any guesswork and saves valuable time that otherwise is spent painstakingly parsing datasets manifest files. Now, customers have the tools and insights to pinpoint the exact cause, location and required action to debug their datasets and successfully train a production-ready custom ML model.
To learn more about improved debugging tools, visit our documentation. Improved debugging tools and insights are available in all Amazon Rekognition Custom Labels regions. You can find the list of supported regions in the region table.
Get started with Amazon Rekognition Custom Labels today.