EC2 Training → SageMaker Endpoint → Lambda Inference (AWS)

Author: Ibrahim Malik

This project demonstrates a practical AWS machine learning deployment workflow:

• Train a PyTorch image classifier on a GPU-backed EC2 instance
• Deploy the trained model to an Amazon SageMaker real-time endpoint
• Invoke the endpoint via an AWS Lambda function
• Configure provisioned concurrency (Lambda) and auto-scaling (endpoint)

The repository is structured as a portfolio-quality project with production-style code, examples, and supporting artefacts.

---

Architecture Overview

1. Dataset stored in S3
2. Training runs on EC2 (GPU) and produces model artefacts
3. Model is deployed to a SageMaker real-time endpoint
4. Lambda receives requests and forwards payloads to the endpoint
5. Endpoint scaling and Lambda concurrency are configured for latency and cost control

---

Repository Structure

.
├── examples/
│   ├── invoke_endpoint_example.py    # Local endpoint invocation script
│   └── lambda_event.json             # Example Lambda invocation payload
├── figures/
│   └── *.png                         # EC2, SageMaker, Lambda, and IAM screenshots
├── notebooks/
│   └── train_and_deploy.ipynb        # End-to-end orchestration notebook
├── reports/
│   └── writeup.md                    # Technical write-up
├── src/
│   ├── ec2_training/
│   │   └── train.py                  # EC2 training entrypoint
│   └── lambda_handler/
│       └── handler.py                # Lambda inference handler
├── legacy/
│   ├── ec2train1.py                  # Original submission files (archived)
│   ├── lambdafunction.py
│   └── test-output.txt
├── .gitignore
├── LICENSE
├── pyproject.toml
└── requirements.txt

---

How to Run

Option 1 — View results (no AWS required)

Open notebooks/train_and_deploy.ipynb to inspect the end-to-end workflow, review screenshots in figures/ for EC2, SageMaker, Lambda, and IAM setup evidence, and read the technical write-up in reports/writeup.md.

requirements.txt is provided for optional local development and notebook viewing. Training and deployment are intended to run in AWS.

---

Option 2 — Run end-to-end (AWS required)

⚠️ Requires valid AWS credentials and will incur charges.

1. Upload the dataset to S3
2. Train on EC2 using src/ec2_training/train.py
3. Run the orchestration notebook: notebooks/train_and_deploy.ipynb
4. Deploy a SageMaker endpoint
5. Configure Lambda with the following environment variable:

SAGEMAKER_ENDPOINT_NAME=<your-endpoint-name>

6. Invoke the endpoint locally via examples/invoke_endpoint_example.py, or via Lambda using examples/lambda_event.json

---

Key Concepts Demonstrated

• GPU-backed model training on EC2
• SageMaker real-time endpoint deployment
• Lambda-based inference integration using boto3 invoke_endpoint
• Environment-based configuration with no hard-coded infrastructure values
• Least-privilege IAM design and security review
• Provisioned concurrency and endpoint auto-scaling considerations
• Production-style repository organisation and modular code structure

---

Attribution

Originally completed as part of the Udacity AWS Machine Learning Engineer Nanodegree. Refactored and documented for professional portfolio presentation.

---

Licence

MIT Licence.