Inventory Bin Classification — CV on SageMaker.

Overview

An end-to-end computer vision pipeline classifying warehouse bin item counts from 10,000+ images using ResNet-50 transfer learning, with deterministic stratified splits stored as S3 ImageFolder channels. SageMaker Bayesian hyperparameter tuning (learning rate, weight decay, batch size, backbone-freeze) with 2-instance distributed training on managed spot instances delivered a ~22% macro-F1 gain over baseline. Served through a real-time SageMaker endpoint with a custom PyTorch inference handler and AWS Lambda integration.

Highlights

• +~22% macro-F1 gain from SageMaker Bayesian HPO with 2-instance distributed training on managed spot instances
• +Deterministic stratified splits as S3 ImageFolder channels for reproducible training
• +Real-time SageMaker endpoint with custom PyTorch inference handler + Lambda (image URL / S3 URI → JSON label, confidence, probabilities)
• +Full resource cleanup — no orphaned endpoints or storage

Stack

Python · PyTorch · ResNet-50 · AWS SageMaker · AWS Lambda · AWS S3 · Docker