NEWS · MODELS · #2599
Fleet-wide edge computer-vision pipeline sample using Amazon SageMaker Python SDK v3 and Neo
AWS published a sample (aws-samples GitHub) updating its 2023 visual-inspection pattern to the Amazon SageMaker Python SDK v3, replacing the deprecated SageMaker Edge Manager with SageMaker Neo. The dual-branch SageMaker Pipeline trains a per-camera anomaly autoencoder and a ResNet18 classifier, evaluates models through quality gates, compiles approved artifacts with SageMaker Neo to produce device-optimized builds (ONNX + Neo-compiled library), and uses Step Functions, an SQS FIFO queue, a DynamoDB semaphore, Lambda, SNS, and EventBridge for fleet orchestration and paced retraining; deployment is via AWS CDK and a full synthetic run costs about $0.07 in compute.
KEY POINTS
- AWS published a sample (aws-samples GitHub) updating its 2023 visual-inspection pattern to the Amazon SageMaker Python SDK v3, replacing the deprecated SageMaker Edge Manager with SageMaker Neo.
- The dual-branch SageMaker Pipeline trains a per-camera anomaly autoencoder and a ResNet18 classifier, evaluates models through quality gates, compiles approved artifacts with SageMaker Neo to produce device-optimized builds (ONNX + Neo-compiled library), and uses Step Functions, an SQS FIFO queue, a DynamoDB semaphore, Lambda, SNS, and EventBridge for fleet orchestration and paced retraining; deployment is via AWS CDK and a full synthetic run costs about $0.07 in compute.
- This provides a reproducible pattern to scale per-camera CV models across fleets while ensuring each registered model is device-optimized and quality-gated, reducing manual operations.
WHY IT MATTERS
This provides a reproducible pattern to scale per-camera CV models across fleets while ensuring each registered model is device-optimized and quality-gated, reducing manual operations.