AWS Lab: Choosing the Optimal ML Modeling Approach

In this lab, you will navigate the decision-making process for selecting the right machine learning approach on AWS. You will explore the three tiers of AWS ML services: AI Services (pre-built), ML Services (SageMaker built-in algorithms), and ML Frameworks (custom scripts), determining which fits a specific business use case based on data type and required customization.

[!WARNING] Remember to run the teardown commands at the end of this lab to avoid ongoing charges for SageMaker resources.

Prerequisites

An active AWS Account.
IAM user with AmazonSageMakerFullAccess, AmazonRekognitionFullAccess, and CloudWatchLogsFullAccess.
AWS CLI installed and configured with your credentials.
Basic familiarity with Python and the AWS Management Console.
A target region (e.g., us-east-1).

Learning Objectives

Categorize AWS ML offerings into the Three Tiers of Customization.
Invoke an AI Service (Amazon Rekognition) for rapid prototyping without model training.
Configure a SageMaker Built-in Algorithm (XGBoost) for structured data classification.
Evaluate which modeling approach satisfies business requirements for latency, cost, and accuracy.

Architecture Overview

This lab demonstrates two paths: the "No-Code/Pre-trained" path via AI Services and the "Managed Algorithm" path via SageMaker.

Loading Diagram...

Step-by-Step Instructions

Step 1: Analyze the Scenario and Select Tier

We have two tasks:

Task A: Identify objects in images for a security app (Rapid deployment needed).
Task B: Predict customer churn based on CSV files (Custom logic required).

Decision: We will use Amazon Rekognition for Task A and SageMaker XGBoost for Task B.

Step 2: Use an AI Service for Task A (Image Analysis)

For Task A, we utilize a pre-trained model. No training code is required.

bash

# Identify labels in a sample image (replace with your bucket/key or use a local file)
aws rekognition detect-labels \
    --image '{"S3Object":{"Bucket":"brainybee-public-assets","Name":"security-sample.jpg"}}' \
    --region <YOUR_REGION>

▶Console alternative

Search for Rekognition in the AWS Console.
In the left sidebar, click Label Detection.
Upload an image of a street or office.
View the "Results" panel for detected objects and confidence scores.

Step 3: Prepare SageMaker Studio for Task B

For tabular data prediction, we need a managed environment.

Navigate to Amazon SageMaker > Domains.
Click Create Domain > Quick setup.
Once the status is "InService", launch Studio.

Step 4: Configure a SageMaker Built-in Algorithm

We will use the SageMaker Python SDK to point to the XGBoost container image. This avoids writing custom model architecture code.

python

import sagemaker
from sagemaker import image_uris

region = sagemaker.Session().boto_region_name
# Get the URI for the built-in XGBoost container
container = image_uris.retrieve("xgboost", region, version="1.5-1")

print(f"Target Container: {container}")

[!TIP] Using image_uris.retrieve ensures you are using a Docker image optimized for AWS infrastructure.

Step 5: Define the Modeling Strategy

Use the following TikZ diagram to visualize where these choices sit on the complexity vs. effort scale:

\begin{tikzpicture}[scale=0.8] \draw[->, thick] (0,0) -- (8,0) node[right] {Customization}; \draw[->, thick] (0,0) -- (0,6) node[above] {Development Effort};

code

% AI Services
\filldraw[blue!20, draw=blue!50] (0.5,0.5) rectangle (2.5,1.5) node[midway, black] {AI Services};
% Built-in Algos
\filldraw[green!20, draw=green!50] (3,2.5) rectangle (5,3.5) node[midway, black] {Built-in};
% Custom Frameworks
\filldraw[orange!20, draw=orange!50] (5.5,4.5) rectangle (7.5,5.5) node[midway, black] {Frameworks};

\node at (1.5, -0.5) {\scriptsize Low};
\node at (6.5, -0.5) {\scriptsize High};

\end{tikzpicture}

Checkpoints

Step	Action	Expected Result
1	Rekognition CLI	A JSON response containing labels like "Person" or "Building" with confidence scores > 80%.
2	SageMaker Studio	The domain status shows a green "InService" checkmark.
3	Image URI	The notebook prints a string like `683313688378.dkr.ecr.us-east-1.amazonaws.com/sagemaker-xgboost:1.5-1`.

Concept Review

Feature	AI Services (Rekognition/Comprehend)	SageMaker Built-in Algorithms	Custom Frameworks (Script Mode)
Expertise	Low (App Developer)	Medium (Data Scientist)	High (ML Engineer)
Model Control	None (Black box)	High (Hyperparameters)	Total (Full Architecture)
Training Data	Not required	Required	Required
Use Case	Common (OCR, Sentiment)	Standard (Regression, XGBoost)	Research/Specialized

Troubleshooting

Error	Cause	Fix
`AccessDeniedException`	Missing IAM permissions	Add `AmazonRekognitionFullAccess` to your IAM user/role.
`ResourceLimitExceeded`	SageMaker quota reached	Check Service Quotas in the console; ensure no old Studio apps are running.
`InvalidImageUri`	Incorrect region or version	Ensure the `region` in your Python script matches your CLI configuration.

Challenge

Scenario: A client wants to detect if customers are happy or sad in the security photos from Task A. Question: Should you use a Custom SageMaker model or an AI Service?

▶Show solution

Use Amazon Rekognition (Face Analysis). It provides sentiment/emotion detection out-of-the-box, saving the cost and time of labeling a custom dataset of emotional faces.

Cost Estimate

Amazon Rekognition: First 5,000 images per month are Free Tier eligible. Otherwise, ~$0.001 per image.
SageMaker Studio: ~$0.05/hour for ml.t3.medium (Free Tier covers 250 hours/month for first 2 months).
S3 Storage: Negligible for small lab datasets (<$0.01).

Teardown

[!IMPORTANT] Failure to delete the SageMaker Domain App can result in unexpected hourly charges.

bash

# 1. List and stop any running SageMaker Studio apps
aws sagemaker list-apps --domain-id <YOUR_DOMAIN_ID>

# 2. Delete the app (example for a JupyterServer app)
aws sagemaker delete-app \
    --domain-id <YOUR_DOMAIN_ID> \
    --app-type JupyterServer \
    --app-name default \
    --user-profile-name <YOUR_USER_NAME>

# 3. Delete any S3 buckets created during the lab
aws s3 rb s3://brainybee-lab-data --force