Amazon SageMaker

Amazon SageMaker

Amazon SageMaker is a fully managed, end-to-end machine learning platform that provides every tool needed to build, train, and deploy ML models. It is the most comprehensive ML service on AWS.

SageMaker Studio

Unified Interface

• What: A single web-based IDE for all ML development activities
• Features:

  • Integrated development environment for ML
  • Unified access to ALL SageMaker features
  • Collaboration features for teams
  • Git integration
  • Visual experiment tracking

Notebooks

• JupyterLab notebooks: Interactive coding environment for data exploration and model prototyping
• Pre-configured environments: Python, R, and ML libraries pre-installed
• Managed infrastructure: Automatically provisions and scales compute
• Kernel management: Switch between different compute instances and ML frameworks

SageMaker Data Wrangler

Data Preparation

• What: Visual tool for data preparation and transformation — no coding required
• Capabilities:

  • Import data from S3, Athena, Redshift, Snowflake, Databricks, and more
  • 300+ built-in data transformations
  • Visual data flow builder (drag-and-drop)
  • Export to SageMaker Pipelines, notebooks, or Python scripts
  • Quick model training to validate data quality

Data Quality Tool

• Automatically assess data quality
• Identify missing values, duplicates, outliers
• Generate data quality reports
• Recommend transformations for common data issues

SageMaker Ground Truth

Data Labeling

• What: Managed service for creating labeled training datasets using human annotators
• How:

  • Define labeling task (image classification, text classification, bounding boxes, etc.)
  • Use Amazon Mechanical Turk, private workforce, or third-party vendors
  • Built-in workflow management for labeling tasks
• Supported Tasks:

  • Image classification, object detection, semantic segmentation
  • Text classification, named entity recognition
  • Video frame labeling
  • 3D point cloud labeling
  • Audio classification

Ground Truth Plus

• What: A premium, turnkey data labeling service where AWS expert workforce handles the entire labeling process
• Difference from Ground Truth: With Plus, you provide the data and AWS manages the entire labeling workflow (project setup, workforce management, quality control)
• Best For: Organizations that don't want to manage labeling workflows

SageMaker Feature Store

• What: A centralized, managed repository for storing, sharing, and managing ML features
• Features:

  • Online Store: Low-latency feature retrieval for real-time inference
  • Offline Store: Historical feature data for training (stored in S3)
  • Feature versioning and lineage tracking
  • Feature sharing across teams and models
  • Batch or streaming ingestion
• Use Case: Ensure consistent features between training and inference (avoid training-serving skew)

SageMaker Training

Training Jobs

• What: Managed infrastructure for training ML models at any scale
• How:

  • Specify training algorithm (built-in or custom container)
  • Point to training data in S3
  • Choose instance type and count
  • SageMaker provisions infrastructure, runs training, stores model artifacts in S3
• Instance Types: CPU, GPU (P3, P4, G4, G5), and ML-optimized instances
• Spot Training: Use Spot Instances for up to 90% cost savings on training
• Managed Warm Pools: Keep instances warm between training jobs to reduce startup time

Distributed Training

• What: Train large models across multiple instances
• Strategies:

  • Data Parallelism: Split training data across instances; each instance has a copy of the model
  • Model Parallelism: Split the model across instances when the model is too large for a single instance
• Supported Frameworks: TensorFlow, PyTorch, MXNet, Hugging Face

SageMaker Autopilot

AutoML

• What: Automatically builds, trains, and tunes ML models with no ML expertise required
• How:

  • Provide a tabular dataset (CSV)
  • Specify the target column
  • Autopilot automatically:

    1. Analyzes data and identifies problem type
    2. Generates multiple ML pipelines (feature engineering + algorithms)
    3. Trains and tunes multiple models
    4. Ranks models by performance
    5. Provides full visibility into the generated code and models
• Transparency: Unlike black-box AutoML, Autopilot provides all generated notebooks and code for review
• Problem Types: Classification, regression, time-series forecasting

SageMaker JumpStart

Pre-trained Models

• What: A model hub with hundreds of pre-trained models ready for deployment or fine-tuning
• Sources: Hugging Face, PyTorch Hub, TensorFlow Hub, and Amazon models
• Capabilities:

  • One-click deployment of pre-trained models
  • Fine-tuning with your data (transfer learning)
  • Solution templates for common ML tasks

Foundation Models

• Access to foundation models (similar to Bedrock but within SageMaker)
• Deploy FMs on SageMaker infrastructure
• Fine-tune FMs with SageMaker training jobs
• Full control over model hosting infrastructure

SageMaker Canvas

No-Code ML

• What: Build ML models using a visual, no-code interface — designed for business analysts
• No ML expertise required
• How:

  • Import data
  • Select target column
  • Canvas automatically builds and trains models
  • Make predictions through a visual interface
• Built-in Integrations: Connect to data in S3, Redshift, and local files
• Model Types: Classification, regression, time-series forecasting, text and image analysis
• Sharing: Share models with data scientists for review via SageMaker Studio

SageMaker Clarify

Bias Detection

• Detect bias in training data and model predictions
• Pre-training bias metrics: Check data for imbalances before training
• Post-training bias metrics: Check model predictions for unfair bias
• Bias Types Detected: Class imbalance, label bias, feature bias, prediction bias

Explainability

• Explain model predictions using SHAP (SHapley Additive exPlanations) values
• Show which features contributed most to each prediction
• Generate feature importance reports
• Supports tabular, NLP, and computer vision models

SageMaker Model Monitor

Model Quality Monitoring

• Continuously monitor deployed models for quality degradation
• Detect when model accuracy drops below acceptable thresholds
• Compare current predictions against baseline
• Alert when intervention is needed

Data Drift Detection

• Detect when input data distribution changes over time (data drift)
• Types of drift monitored:

  • Data Quality Drift: Changes in data statistics (mean, std, missing values)
  • Model Quality Drift: Changes in prediction accuracy
  • Bias Drift: Changes in fairness metrics over time
  • Feature Attribution Drift: Changes in feature importance

SageMaker Debugger

• What: Debug and profile ML training jobs in real-time
• Capabilities:

  • Monitor training metrics (loss, accuracy) in real-time
  • Detect training issues (vanishing gradients, overfitting, exploding gradients)
  • Profile hardware utilization (CPU, GPU, memory, I/O)
  • Automatic issue detection with built-in rules
  • Generate profiling reports

SageMaker Experiments

• What: Track, organize, and compare ML experiments
• Capabilities:

  • Log hyperparameters, metrics, and artifacts for each training run
  • Compare experiments side-by-side
  • Visualize training curves
  • Organize experiments into groups
  • Reproduce past experiments

SageMaker Pipelines

CI/CD for ML

• What: Build automated, repeatable ML workflows (MLOps)
• How: Define a Directed Acyclic Graph (DAG) of ML steps
• Steps Include:

  • Data processing
  • Model training
  • Model evaluation
  • Conditional logic (deploy if accuracy > threshold)
  • Model registration

Pipeline Steps

SageMaker Model Registry

Model Versioning

• What: Central catalog for managing ML model versions
• Capabilities:

  • Track model versions with metadata
  • Model approval workflow (Pending → Approved → Rejected)
  • Deploy approved models to endpoints
  • Integration with SageMaker Pipelines for automated registration
  • Track model lineage (training data, algorithm, hyperparameters)

SageMaker Inference

Real-time Inference

• Deploy models as persistent endpoints for low-latency, synchronous predictions
• Auto-scaling based on traffic
• A/B testing with traffic splitting

Batch Transform

• Process large datasets asynchronously (no persistent endpoint needed)
• Pay only for compute time used
• Results stored in S3

Serverless Inference

• What: Auto-scaling inference endpoints that scale to zero when not in use
• When to Use: Intermittent or unpredictable traffic
• Benefits: Pay only when requests are made (no idle costs)
• Limitation: Cold start latency when scaling from zero

SageMaker Edge Manager

• Deploy and manage ML models on edge devices
• Monitor model performance on edge
• Over-the-air model updates
• Supports devices like IoT sensors, cameras, industrial equipment

SageMaker Model Cards

Model Documentation

• What: Create standardized documentation for ML models
• Contents:

  • Model purpose and intended use
  • Training details (algorithm, data, hyperparameters)
  • Evaluation results and metrics
  • Ethical considerations
  • Known limitations
• Purpose: Responsible AI — ensure transparency and accountability

SageMaker Model Dashboard

• Centralized view of all deployed models
• Monitor model health, performance, and compliance
• Track model lineage and versions
• Identify models that need attention

SageMaker Role Manager

• Simplify creation of IAM roles for SageMaker with pre-built permissions
• Role templates for common ML personas (data scientist, ML engineer, ML operations)
• Least-privilege access by default

MLFlow on SageMaker

• What: Run open-source MLflow on SageMaker managed infrastructure
• Track experiments, package models, and deploy using the familiar MLflow interface
• Managed servers (no infrastructure to manage)
• Integration with SageMaker training and deployment

Network Isolation Mode

• Run SageMaker training and inference in complete network isolation
• No outbound network access from training containers
• Data only accessible from S3 via VPC endpoints
• Prevents data exfiltration during training
• Ensures sensitive training data stays within the VPC
• Required for some compliance scenarios

Quick Reference Table