This project demonstrates the end-to-end process of building, training, deploying, and orchestrating an image classification model using AWS SageMaker, Lambda, and Step Functions.
The model classifies delivery vehicles for Scones Unlimited, a logistics company, into bicycles and motorcycles to optimize delivery routing and logistics operations.
Build a scalable image classification pipeline using AWS SageMaker.
Train a CNN model using the CIFAR-100 dataset to distinguish bicycles and motorcycles.
Deploy the trained model as a real-time inference endpoint.
Enable data capture and model monitoring.
Build a Step Function workflow with Lambda functions to:
Fetch and serialize image data from S3.
Invoke the deployed SageMaker endpoint.
Filter predictions based on a confidence threshold.
- Data Staging
Used the CIFAR-100 dataset (Python version).
Extracted, transformed, and filtered only the bicycle and motorcycle images.
Saved the images as .png files in local train/ and test/ folders.
Uploaded data and manifest files (train.lst, test.lst) to an S3 bucket.
- Model Training & Deployment
Retrieved the prebuilt SageMaker image-classification container.
Trained the model using SageMaker’s built-in algorithm on an ml.p3.2xlarge instance.
Achieved ~81% validation accuracy.
Deployed the model on a ml.m5.xlarge endpoint.
Configured data capture for model monitoring.
- Lambda Functions
Three Lambda functions were built to support the Step Function workflow:
serializeImageData – Downloads and base64-encodes an image from S3.
classifyImageData – Invokes the SageMaker endpoint and returns prediction probabilities.
filterInferences – Checks if the model confidence exceeds the defined threshold (e.g., 0.94).
- Step Function Workflow
Orchestrated the above three Lambdas in sequence:
Serialize → Classify → Filter
Configured transitions between steps and ensured outputs were properly mapped.
Tested with multiple executions to validate successful and failure paths.
- Model Monitoring
Verified data captured in S3 under /data_capture/AllTraffic/...
Visualized recent inference confidence trends over time using matplotlib.
- Visualization
Sampled 50 training images and ran inference against the deployed endpoint.
Compared predictions with true labels.
Built a scatter plot visualization of prediction confidence vs. true labels:
Blue dots = bicycles
Red dots = motorcycles
Green dashed line = confidence threshold (0.94)
Validation Accuracy: ~81.7%
Sample Evaluation Accuracy (50 images): ~85–90%
Model confidence consistent across both classes (no major bias).