Skip to content

SageMaker Service Overview

How is AWS SageMaker different from other ML frameworks? (Lesson 49)

Overview of SageMaker

  • SageMaker is a cloud service that leverages AWS resources for training and hosting machine learning models.

Key Differences from Direct Framework Usage

  1. Framework Installation:
  2. Traditional ML frameworks like SKLearn, TensorFlow, or PyTorch can be installed and used locally on a laptop.

  3. SageMaker, being cloud-based, requires the use of AWS resources.

  4. Production Application:

  5. Converting local ML models into production applications often requires additional tools or cloud services.

  6. SageMaker simplifies this process, making it more straightforward to transition models to production.

  7. Cost and Trial Period:

  8. Using SageMaker involves some cost, but it offers a two-month free trial for getting started.

  9. Containers and Built-in Algorithms:

  10. SageMaker utilizes containers to encapsulate algorithms and frameworks.

  11. Provides containers for built-in algorithms (XGBoost, DeepAR, FM, etc.) and popular frameworks (PyTorch, SKLearn, TensorFlow).

  12. Training Parameters:

  13. Requires specifying training parameters like data location (S3 or EFS), hyperparameters, server type for training, and storage location for trained artifacts.

  14. Script File for Frameworks:

  15. In addition to training parameters, frameworks on SageMaker require a script file.

  16. This script includes the model-building code specific to PyTorch, TensorFlow, or SKLearn.

  17. Hosting Models:

  18. Post-training, models can be hosted on SageMaker for deployment.
  19. SageMaker's containerized approach offers a standardized interface for building and deploying models across different algorithms and frameworks.

Conclusion

  • SageMaker stands out in its ability to streamline the transition from model training to production deployment, backed by AWS's robust cloud infrastructure.

Introduction to SageMaker (Lesson 50)

Key Capabilities

  1. Jupyter Notebook Environment:
  2. Managed Jupyter notebook instances for development and data preparation.
  3. Automated Python installation and patch application.

  4. Custom Python packages installation supported.

  5. Machine Learning Algorithm Support:

  6. Wide variety of optimized ML algorithms for AWS Cloud.
  7. Optimized environments for frameworks like TensorFlow, Apache MxNet.

  8. Custom algorithm deployment capability.

  9. Training Infrastructure:

  10. Scalable training on one or multiple compute instances.

  11. Handles large datasets, with trained model artifacts stored in S3.

  12. Model Deployment:

  13. Real-time prediction support with load-balanced compute instances.
  14. Auto-scaling for instance management and workload adaptation.
  15. Batch transform for non-interactive, large-scale inference tasks.

Deployment Options

  • Real-time Prediction:
  • Persistent endpoint for client application model invocations.
  • Auto-replacement of instances in case of hardware/system issues.

  • Customizable scaling policies based on workload.

  • Batch Transform:

  • Ideal for non-interactive, large dataset inference.
  • Manages resource allocation, dataset processing, and results storage.
  • Terminates instances post-processing, enhancing cost-effectiveness.

Summary

  • SageMaker offers a comprehensive suite of services for building, training, and deploying ML models in the cloud.
  • Features include a fully managed development environment, extensive algorithm and framework support, scalable training infrastructure, and flexible deployment options for real-time and batch processing.

Instance Type and Pricing (Lesson 51)

Instance Families

  1. Standard Instances:
  2. Low-cost, balanced performance and memory.

  3. T2, T3 (occasional burst utilization), M5 (sustained load) instances.

  4. Compute Optimized Instances:

  5. High-performance CPUs for CPU-intensive tasks.

  6. C4, C5 instance types.

  7. Accelerated Computing Instances:

  8. Powerful GPUs for GPU-optimized algorithms.

  9. P2, P3 instances for faster training and GPU-enabled hosting.

  10. Inference Acceleration:

  11. Fractional GPU capabilities as an add-on to other instances.
  12. Suitable for models needing partial GPU support in inference.

Choosing Instances

  • Based on algorithm requirements: CPU-bound or GPU-bound.
  • Experiment with instance sizes for optimal performance and cost.
  • Instance naming: ML.<instance type>.<generation>.<size> (e.g., ML.C5.2xlarge).

Pricing Components

  • Instance Cost: Varies by type and size.
  • Variable Costs: Depends on fractional GPUs, storage allocated, and data transfer.
  • Region-Specific Pricing: Slight variations between AWS regions.

Free-Tier Offer (for New Users)

  • Development: T2/T3 medium notebook instances - 250 hours/month.
  • Training: M4/M5 extra-large instances - 50 hours/month.
  • Hosting: M4/M5 extra-large instances - 125 hours/month.

Beyond Free-Tier

  • Development (Example):
  • T3 extra-large instance: $0.23/hour.
  • Storage: $0.14/GB/month.
  • Data Transfer: $0.016/GB (in/out).
  • Training Costs: Based on instance runtime and storage during training.
  • Hosting Options:
  • Real-time Hosting: Cost depends on instance type, GPUs, storage, data transfer.
  • Batch Transform: Charges for job duration, suitable for non-interactive tasks.

Summary

SageMaker provides a range of instance options catering to various ML requirements, with pricing involving multiple components. Users can choose instances based on their algorithm's needs and manage costs effectively by right-sizing and using the free-tier benefits.

Save Money on SageMaker Usage (Lesson 52)

1. SageMaker Savings Plan

  • Overview: Offers up to 64% discount over on-demand pricing for consistent usage.
  • Commitment: Choose a 1-year or 3-year term with a fixed $/hour usage.
  • Applicability: Applies to SageMaker Notebook, Studio, Training, Real-time Inference, Batch Transform, Data Wrangler, etc.
  • Global Benefit: Discounts apply across regions, instance families, and sizes.
  • Challenges:
  • Determining steady-state usage: Use AWS Cost Explorer to analyze and recommend plans.
  • Fluctuating Usage: Billing based on hourly usage, not monthly aggregate. Unused hours may not carry over.
  • Example: Convert a $10/hour steady-state usage to a $4/hour savings plan to reduce monthly charges from $7,200 to $2,880.
  • Resource: SageMaker Savings Plan Announcement

2. Managed Spot Training

  • Discounts: Up to 90% off on AWS's unused capacity.
  • Use Case: Ideal for training and automatic model tuning jobs (not for inference).
  • Handling Interruptions: SageMaker automatically manages spot instance interruptions and resumes training when capacity is available.
  • No Long-term Commitment: Flexible and cost-effective for training and tuning.
  • Resource: Managed Spot Training Documentation

3. Learning with SageMaker

  • Two-Month Trial: Ideal for learners, providing free-tier instance types.
  • Beyond Free-Tier: Managed spot training is recommended for cost savings.
  • Multiple Accounts for Extended Trial: Sign up with different email IDs for additional two-month trial periods (albeit cumbersome).

By utilizing these methods, users can significantly reduce their AWS SageMaker costs while maintaining efficient machine learning operations.

Also refer to the CloudPractitionerReview-InfraPricingSupport.pdf. This review material provides a quick overview of important concepts related to infrastructure, pricing, support plans, and shared responsibility model in the cloud

Data Format (Lesson 53)

Supported Data Formats

  • Common Formats: CSV and RecordIO (protobuf)
  • Other Formats: Some SageMaker algorithms also support LibSVM, JSON, and Parquet formats.

Storage and Retrieval

  • Storage Location: Data for training and validation must be stored in an Amazon S3 bucket.
  • Data Access:
  • Single File: If a single file in S3 is specified, its contents are used for training.
  • Folder: If a folder is specified, all data within that folder is used for training.

Usage of Channels

  • Train Channel: For the training dataset.
  • Test Channel: For the evaluation dataset.

Data Transfer Modes

  1. File Mode:
  2. Entire data from S3 is copied to training instance volumes.
  3. Requires enough disk space on training instances for the full dataset and model artifacts.
  4. Pipe Mode:
  5. Streams data from S3 to training instances.
  6. Faster start times and better throughput.
  7. Reduces storage needs on training instances.

Hands-On Lab: Data Format Exploration

  • Objective: Create a dataset in CSV and RecordIO formats, upload to S3, and download from S3.
  • Tools Used: Pandas, NumPy, Boto3, SageMaker library.
  • Process:
  • Create a dataset with 3 input features (X1, X2, X3) and a binary target variable (Y).
  • Convert the dataset to CSV using Pandas to_csv method.
  • Upload the CSV file to S3 using the write_to_s3 helper function.
  • Use write_numpy_to_dense_tensor method from SageMaker library to create a RecordIO formatted file.
  • Upload and download RecordIO file to/from S3.

Summary

  • SageMaker supports multiple data formats for efficient training and prediction.
  • The choice of data format (CSV, RecordIO, others) depends on algorithm compatibility and dataset size.
  • Data must be stored in S3, with training and validation data distinguished via channels.
  • File and Pipe modes offer different approaches to data transfer from S3 to training instances.
  • Practically, users can create, upload, and manage these data formats using SageMaker's integrated tools and AWS libraries.

SageMaker Built-in Algorithms (Lesson 54)

AWS SageMaker provides a range of built-in algorithms optimized for cloud-based machine learning. These algorithms can be broadly categorized based on their applications:

Text Data Algorithms

  1. BlazingText:
  2. Modes: Unsupervised (Word2Vec) and supervised.

  3. Use: Converts text to vector, grouping semantically similar words. Useful for text classification.

  4. Object2Vec:

  5. Type: Supervised.
  6. Use: Converts text to vector while capturing sentence structure. Suitable for associating customers with products, movies with ratings, etc.

Recommender Systems and Collaborative Filtering

  1. Factorization Machines:
  2. Type: Suitable for high-dimensional sparse datasets.
  3. Use: Popular for building recommender systems and collaborative filtering.

Classification and Regression

  1. K-Nearest Neighbor (KNN):

  2. Use: Simple, effective for classification (majority class of k-nearest neighbors) and regression (average value of k-nearest neighbors).

  3. Linear Models:

  4. Types: Linear regression, logistic regression, multinomial logistic regression.

  5. XGBoost:

  6. Use: Gradient boosted tree algorithm for both regression and classification.

Time Series Forecasting

  1. DeepAR:
  2. Use: Trains on multiple time series, predicts new similar series.

Image Analysis

  1. Object Detection:

  2. Use: Detects and classifies objects in images, provides bounding boxes.

  3. ImageClassification:

  4. Use: Multi-label classification of images.

  5. Semantic Segmentation:

  6. Use: Tags each pixel in an image with a class label, useful in computer vision.

Language Processing

  1. Sequence to Sequence:
  2. Use: Useful for text summarization, language translation, speech to text.

Clustering and Topic Modeling

  1. K Means:

  2. Type: Unsupervised clustering algorithm.

  3. LDA (Latent Dirichlet Allocation):

  4. Use: Groups documents based on topics.

  5. Neural Topic Modeling:

  6. Use: Similar to LDA, for document grouping by topics.

Dimensionality Reduction

  1. PCA (Principal Component Analysis):
  2. Use: Reduces dataset dimensions while retaining information.

Anomaly Detection

  1. Random Cut Forest:

  2. Use: Detects anomalies in data, assigns scores to points.

  3. IP Insights:

  4. Use: Detects unusual network activity, useful for security applications.

Summary

  • SageMaker's built-in algorithms cover a wide range of ML applications.
  • They are optimized for AWS cloud, offering scalability and GPU training advantages.
  • Users can choose the appropriate algorithm based on their data and use case requirements.

SageMaker Ground Truth

  • Purpose: Assists in labeling data for machine learning.
  • Capabilities:
  • Automatic Labeling: Uses provided examples to label remaining data.
  • Human Labeling: Distributes labeling tasks to human labelers via integration with Mechanical Turk, managing the workflow.

SageMaker Neo

  • Functionality: Optimizes machine learning models for deployment across cloud and edge locations.
  • Key Features:
  • Cross-Compilation: Converts and optimizes models for various hardware like Intel, Nvidia, ARM.
  • Use Case: Suitable for applications where low latency is crucial, allowing deployment at edge locations.

Developing Models with Other Frameworks

  • Container-Based Architecture: Utilizes Docker containers for training and deploying models.
  • Pre-Built Docker Images: Available for built-in algorithms and popular frameworks like TensorFlow, Apache MxNet, etc.
  • Integration with Apache Spark:
  • Provides a library for Python and Scala.
  • Directly consumes data frames in Spark clusters, converting them to protobuf and uploading to S3.
  • Framework Support: Offers support for TensorFlow, Apache MxNet, Scikit-learn, PyTorch, Chainer, etc.
  • Custom Algorithm Development:
  • Option to extend SageMaker Docker images or use custom container images.
  • Suitable for unique runtime or language requirements.
  • Advanced Use with Deep Learning AMIs:
  • For modifying or extending frameworks.
  • Ideal for users contributing to code or debugging at the framework level.

Summary

SageMaker provides a flexible environment for machine learning, offering: - Built-in algorithms for ease of use and scalability. - Support for popular frameworks via pre-built container images. - Options to extend or bring custom container images for unique requirements. - Ground Truth for data labeling and Neo for optimizing model deployment.