The
ARTIficial Intelligence INfrastructure
Workshop
Build your own highly scalable and robust data storage systems that can support a variety of cutting-edge AI applications
Chinmay Arankalle, Gareth Dwyer, Bas Geerdink, Kunal Gera, Kevin Liao, and Anand N.S.
The Artificial Intelligence Infrastructure Workshop
Copyright © 2020 Packt Publishing
All rights reserved. No part of this course may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, without the prior written permission of the publisher, except in the case of brief quotations embedded in critical articles or reviews.
Every effort has been made in the preparation of this course to ensure the accuracy of the information presented. However, the information contained in this course is sold without warranty, either express or implied. Neither the authors, nor Packt Publishing, and its dealers and distributors will be held liable for any damages caused or alleged to be caused directly or indirectly by this course.
Packt Publishing has endeavored to provide trademark information about all of the companies and products mentioned in this course by the appropriate use of capitals. However, Packt Publishing cannot guarantee the accuracy of this information.
Authors: Chinmay Arankalle, Gareth Dwyer, Bas Geerdink, Kunal Gera, Kevin Liao, and Anand N.S.
Reviewers: Brent Broadnax, John Wesley Doyle, Tim Hoolihan, Rochit Jain, Sasikanth Kotti, Asheesh Mehta, Arunkumar Nair, Madhav Pandya, Ashish Patel, Shovon Sengupta, and Ashish Tulsankar
Managing Editor: Ashish James
Acquisitions Editors: Manuraj Nair, Royluis Rodrigues, Kunal Sawant, Anindya Sil, Archie Vankar, and Karan Wadekar
Production Editor: Shantanu Zagade
Editorial Board: Megan Carlisle, Samuel Christa, Mahesh Dhyani, Heather Gopsill, Manasa Kumar, Alex Mazonowicz, Monesh Mirpuri, Bridget Neale, Dominic Pereira, Shiny Poojary, Abhishek Rane, Brendan Rodrigues, Erol Staveley, Ankita Thakur, Nitesh Thakur, and Jonathan Wray
First published: August 2020
Production reference: 1130820
ISBN: 978-1-80020-984-8
Published by Packt Publishing Ltd.
Livery Place, 35 Livery Street
Birmingham B3 2PB, UK
Table of Contents
Preface i
1. Data Storage Fundamentals 1
Introduction 2
Problems Solved by Machine Learning 3
Image Processing – Detecting Cancer in Mammograms with Computer Vision 3
Text and Language Processing – Google Translate 4
Audio Processing – Automatically Generated Subtitles 6
Time Series Analysis 7
Optimizing the Storing and Processing of Data for Machine Learning Problems 7
Diving into Text Classification 8
Looking at TF-IDF Vectorization 9
Looking at Terminology in Text Classification Tasks 11
Exercise 1.01: Training a Machine Learning Model to Identify Clickbait Headlines 12
Designing for Scale – Choosing the Right Architecture and Hardware 21
Optimizing Hardware – Processing Power, Volatile Memory, and Persistent Storage 21
Optimizing Volatile Memory 23
Optimizing Persistent Storage 24
Optimizing Cloud Costs – Spot Instances and Reserved Instances 25
Using Vectorized Operations to Analyze Data Fast 26
Exercise 1.02: Applying Vectorized Operations to Entire Matrices 27
Activity 1.01: Creating a Text Classifier for Movie Reviews 33
Summary 36
2. Artificial Intelligence Storage Requirements 39
Introduction 40
Storage Requirements 41
The Three Stages of Digital Data 43
Data Layers 44
From Data Warehouse to Data Lake 45
Exercise 2.01: Designing a Layered Architecture for an AI System 47
Requirements per Infrastructure Layer 49
Raw Data 50
Security 50
Basic Protection 50
The AIC Rating 51
Role-Based Access 52
Encryption 53
Exercise 2.02: Defining the Security Requirements for Storing Raw Data 54
Scalability 55
Time Travel 56
Retention 57
Metadata and Lineage 59
Historical Data 60
Security 60
Scalability 61
Availability 61
Exercise 2.03: Analyzing the Availability of a Data Store 62
Availability Consequences 63
Time Travel 65
Locality of Data 67
Metadata and Lineage 67
Streaming Data 68
Security 68
Performance 69
Availability 69
Retention 70
Exercise 2.04: Setting the Requirements for Data Retention 71
Analytics Data 72
Performance 72
Cost-Efficiency 73
Quality 73
Model Development and Training 74
Security 75
Availability 75
Retention 75
Activity 2.01: Requirements Engineering for a Data-Driven Application 76
Summary 77
3. Data Preparation 79
Introduction 80
ETL 80
Data Processing Techniques 82
Exercise 3.01: Creating a Simple ETL Bash Script 83
Traditional ETL with Dedicated Tooling 93
Distributed, Parallel Processing with Apache Spark 94
Exercise 3.02: Building an ETL Job Using Spark 95
Activity 3.01: Using PySpark for a Simple ETL Job to Find Netflix Shows for All Ages 102
Source to Raw: Importing Data from Source Systems 104
Raw to Historical: Cleaning Data 105
Raw to Historical: Modeling Data 106
Historical to Analytics: Filtering and Aggregating Data 107
Historical to Analytics: Flattening Data 107
Analytics to Model: Feature Engineering 107
Analytics to Model: Splitting Data 109
Streaming Data 110
Windows 110
Event Time 112
Late Events and Watermarks 113
Exercise 3.03: Streaming Data Processing with Spark 114
Activity 3.02: Counting the Words in a Twitter Data Stream to Determine the Trending Topics 123
Summary 125
4. The Ethics of AI Data Storage 127
Introduction 128
Case Study 1: Cambridge Analytica 130
Summary and Takeaways 135
Case Study 2: Amazon's AI Recruiting Tool 136
Imbalanced Training Sets 136
Summary and Takeaways 139
Case Study 3: COMPAS Software 140
Summary and Takeaways 142
Finding Built-In Bias in Machine Learning Models 143
Exercise 4.01: Observing Prejudices and Biases in Word Embeddings 145
Exercise 4.02: Testing Our Sentiment Classifier on Movie Reviews 151
Activity 4.01: Finding More Latent Prejudices 158
Summary 160
5. Data Stores: SQL and NoSQL Databases 163
Introduction 164
Database Components 165
SQL Databases 166
MySQL 167
Advantages of MySQL 167
Disadvantages of MySQL 167
Query Language 167
Terminology 168
Data Definition Language (DDL) 168
Data Manipulation Language (DML) 169
Data Control Language (DCL) 170
Transaction Control Language (TCL) 171
Data Retrieval 172
SQL Constraints 176
Exercise 5.01: Building a Relational Database for the FashionMart Store 179
Data Modeling 186
Normalization 187
Dimensional Data Modeling 190
Performance Tuning and Best Practices 193
Activity 5.01: Managing the Inventory of an E-Commerce Website Using a MySQL Query 194
NoSQL Databases 198
Need for NoSQL 199
Consistency Availability Partitioning (CAP) Theorem 200
MongoDB 201
Advantages of MongoDB 201
Disadvantages of MongoDB 202
Query Language 202
Terminology 202
Exercise 5.02: Managing the Inventory of an E-Commerce Website Using a MongoDB Query 210
Data Modeling 217
Lack of Joins 217
Joins 219
Performance Tuning and Best Practices 221
Activity 5.02: Data Model to Capture User Information 221
Cassandra 226
Advantages of Cassandra 226
Disadvantages of Cassandra 227
Dealing with Denormalizations in Cassandra 227
Query Language 228
Terminology 228
Exercise 5.03: Managing Visitors of an E-Commerce Site Using Cassandra 231
Data Modeling 237
Column Family Design 238
Distributing Data Evenly across Clusters 239
Considering Write-Heavy Scenarios 239
Performance Tuning and Best Practices 240
Activity 5.03: Managing Customer Feedback Using Cassandra 240
Exploring the Collective Knowledge of Databases 242
Summary 245
6. Big Data File Formats 247
Introduction 248
Common Input Files 248
CSV – Comma-Separated Values 249
JSON – JavaScript Object Notation 249
Choosing the Right Format for Your Data 250
Orientation – Row-Based or Column-Based 251
Row-Based 251
Column-Based 252
Partitions 253
Schema Evolution 254
Compression 254
Introduction to File Formats 255
Parquet 255
Exercise 6.01: Converting CSV and JSON Files into the Parquet Format 260
Avro 266
Exercise 6.02: Converting CSV and JSON Files into the Avro Format 267
ORC 274
Exercise 6.03: Converting CSV and JSON Files into the ORC Format 276
Query Performance 282
Activity 6.01: Selecting an Appropriate Big Data File Format for Game Logs 284
Summary 285
7. Introduction to Analytics Engine (Spark) for Big Data 287
Introduction 288
Apache Spark 289
Fundamentals and Terminology 290
How Does Spark Work? 294
Apache Spark and Databricks 295
Exercise 7.01: Creating Your Databricks Notebook 296
Understanding Various Spark Transformations 304
Exercise 7.02: Applying Spark Transformations to Analyze the Temperature in California 306
Understanding Various Spark Actions 311
Spark Pipeline 312
Exercise 7.03: Applying Spark Actions to the Gettysburg Address 313
Activity 7.01: Exploring and Processing a Movie Locations Database Using Transformations and Actions 319
Best Practices 321
Summary 322
8. Data System Design Examples 325
Introduction 326
The Importance of System Design 327
Components to Consider in System Design 328
Features 328
Hardware 329
Data 329
Architecture 330
Security 330
Scaling 331
Examining a Pipeline Design for an AI System 331
Reproducibility – How Pipelines Can Help Us Keep Track of Each Component 334
Exercise 8.01: Designing an Automatic Trading System 334
Making a Pipeline System Highly Available 342
Exercise 8.02: Adding Queues to a System to Make It Highly Available 344
Activity 8.01: Building the Complete System with Pipelines and Queues 348
Summary 350
9. Workflow Management for AI 353
Introduction 354
Creating Your Data Pipeline 355
Exercise 9.01: Implementing a Linear Pipeline to Get the Top 10 Trending Videos 357
Exercise 9.02: Creating a Nonlinear Pipeline to Get the Daily Top 10 Trending Video Categories 363
Challenges in Managing Processes in the Real World 372
Automation 372
Failure Handling 373
Retry Mechanism 373
Exercise 9.03: Creating a Multi-Stage Data Pipeline 375
Automating a Data Pipeline 382
Exercise 9.04: Automating a Multi-Stage Data Pipeline Using a Bash Script 382
Automating Asynchronous Data Pipelines 385
Exercise 9.05: Automating an Asynchronous Data Pipeline 388
Workflow Management with Airflow 392
Exercise 9.06: Creating a DAG for Our Data Pipeline Using Airflow 394
Activity 9.01: Creating a DAG in Airflow to Calculate the Ratio of Likes-Dislikes for Each Category 405
Summary 408
10. Introduction to Data Storage on Cloud Services (AWS) 411
Introduction 412
Interacting with Cloud Storage 413
Exercise 10.01: Uploading a File to an AWS S3 Bucket Using AWS CLI 416
Exercise 10.02: Copying Data from One Bucket to Another Bucket 421
Exercise 10.03: Downloading Data from Your S3 Bucket 423
Exercise 10.04: Creating a Pipeline Using AWS SDK Boto3 and Uploading the Result to S3 425
Getting Started with Cloud Relational Databases 430
Exercise 10.05: Creating an AWS RDS Instance via the AWS Console 431
Exercise 10.06: Accessing and Managing the AWS RDS Instance 442
Introduction to NoSQL Data Stores on the Cloud 450
Key-Value Data Stores 452
Document Data Stores 452
Columnar Data Store 453
Graph Data Store 454
Data in Document Format 455
Activity 10.01: Transforming a Table Schema into Document Format and Uploading It to Cloud Storage 456
Summary 458
11. Building an Artificial Intelligence Algorithm 461
Introduction 462
Machine Learning Algorithms 462
Model Training 463
Closed-Form Solution 463
Non-Closed-Form Solutions 464
Gradient Descent 464
Exercise 11.01: Implementing a Gradient Descent Algorithm in NumPy 467
Getting Started with PyTorch 478
Exercise 11.02: Gradient Descent with PyTorch 481
Mini-Batch SGD with PyTorch 488
Exercise 11.03: Implementing Mini-Batch SGD with PyTorch 492
Building a Reinforcement Learning Algorithm to Play a Game 500
Exercise 11.04: Implementing a Deep Q-Learning Algorithm in PyTorch to Solve the Classic Cart Pole Problem 506
Activity 11.01: Implementing a Double Deep Q-Learning Algorithm to Solve the Cart Pole Problem 513
Summary 516
12. Productionizing Your AI Applications 519
Introduction 520
pickle and Flask 521
Exercise 12.01: Creating a Machine Learning Model API with pickle and Flask That Predicts Survivors of the Titanic 522
Activity 12.01: Predicting the Class of a Passenger on the Titanic 536
Deploying Models to Production 537
Docker 538
Kubernetes 539
Exercise 12.02: Deploying a Dockerized Machine Learning API to a Kubernetes Cluster 541
Activity 12.02: Deploying a Machine Learning Model to a Kubernetes Cluster to Predict the Class of Titanic Passengers 555
Model Execution in Streaming Data Applications 557
PMML 558
Apache Flink 559
Exercise 12.03: Exporting a Model to PMML and Loading it in the Flink Stream Processing Engine for Real-time Execution 559
Activity 12.03: Predicting the Class of Titanic Passengers in Real Time 572
Summary 575
Appendix 579
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.