Table Of ContentProduction-Ready Applied
Deep Learning
Learn how to construct and deploy complex models in PyTorch
and TensorFlow deep learning frameworks
Tomasz Palczewski
Jaejun (Brandon) Lee
Lenin Mookiah
BIRMINGHAM—MUMBAI
Production-Ready Applied Deep Learning
Copyright © 2022 Packt Publishing
All rights reserved. No part of this book 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 book to ensure the accuracy of the information
presented. However, the information contained in this book is sold without warranty, either express or
implied. Neither the authors, nor Packt Publishing or its dealers and distributors, will be held liable for any
damages caused or alleged to have been caused directly or indirectly by this book.
Packt Publishing has endeavored to provide trademark information about all of the companies and products
mentioned in this book by the appropriate use of capitals. However, Packt Publishing cannot guarantee
the accuracy of this information.
Publishing Product Manager: Ali Abidi
Senior Editor: Nazia Shaikh
Content Development Editor: Shreya Moharir
Technical Editor: Rahul Limbachiya
Copy Editor: Safis Editing
Project Coordinator: Farheen Fathima
Proofreader: Safis Editing
Indexer: Rekha Nair
Production Designer: Aparna Bhagat
Marketing Coordinators: Shifa Ansari and Abeer Dawe
First published: September 2022
Production reference: 1260822
Published by Packt Publishing Ltd.
Livery Place
35 Livery Street
Birmingham
B3 2PB, UK.
ISBN 978-1-80324-366-5
www.packt.com
To Sylwia, Anna, and Matt – my loves, my life.
To my Mom, my brother Piotr, and my family.
- Tomasz
To my parents, Changhee and Kyung Ja, for always loving and supporting me.
- Jaejun
To my mom, Chendurkani, for her unconditional support and encouragement.
- Lenin
Finally, we would like to dedicate this book to self-motivated and
value-driven individuals who put their time into learning
new technologies to make the world more exciting.
Contributors
About the authors
Tomasz Palczewski is currently working as a staff software engineer at Samsung Research America
(SRA). He has a Ph.D. in physics and an eMBA degree from Quantic. His zeal for getting insights out of
large datasets using cutting-edge techniques led him to work across the globe at CERN (Switzerland),
LBNL (Italy), J-PARC (Japan), University of Alabama (US), and the University of California, Berkeley
(US). In 2016, he was deployed to the South Pole to calibrate the world’s largest neutrino telescope.
At some point, he decided to pivot his career and focus on applying his skills in industry. Currently,
Dr. Palczewski works on modeling user behavior and creating value for advertising and marketing
verticals by deploying machine learning (ML), deep learning, and statistical models at scale.
I had the idea of writing a book that my younger self would appreciate. The book would show different
aspects of production-ready deep learning. I am grateful that Jaejun and Lenin were excited about
the idea and joined the project. Without their help, this would not have turned out as it did. Finally, I
would like to thank my wife for all her support.
Jaejun (Brandon) Lee is currently working as an AI research lead at RoboEye.ai, integrating cutting-
edge algorithms in computer vision and AI into industrial automation solutions. He obtained his
master’s degree from the University of Waterloo with research focused on natural language processing
(NLP), specifically speech recognition. He has spent many years developing a fully productionized yet
open source wake word detection toolkit with a web browser deployment target, Howl. Luckily, his
effort has been picked up by Mozilla’s Firefox Voice and it is actively providing a completely hands-
free experience to many users all over the world.
I would like to thank Tomasz for offering this remarkable opportunity to become an author. Next, I am
really grateful to Lenin for sharing his knowledge of data engineering throughout our journey. Lastly, I
would like to thank Erica for her encouragement.
Lenin Mookiah is a machine learning engineer who has worked with reputed tech companies –
Samsung Research America, eBay Inc., and Adobe R&D. He has worked in the technology industry
for over 11 years in various domains: banking, retail, eDiscovery, and media. He has played various
roles in the end-to-end productization of large-scale machine learning systems. He mainly employs
the big data ecosystem to build reliable feature pipelines that data scientists consume. Apart from his
industrial experience, he researched anomaly detection in his Ph.D. at Tennessee Tech University
(US) using a novel graph-based approach. He studied entity resolution on social networks during
his master’s at Tsinghua University, China.
Working with Tomasz and Jaejun was very exciting. I sincerely thank both for the collaboration on this
book. I have learned many aspects of data science from both.
About the reviewers
Utkarsh Srivastava is an AI/ML professional, trainer, YouTuber, and blogger. He loves to tackle and
develop ML, NLP, and computer vision algorithms to solve complex problems. He started his data science
career as a blogger of his own blog (datamahadev.com) and YouTube channel (datamahadev),
followed by working as a senior data science trainer in an institute in Gujarat. Additionally, he has
trained and counseled 1,000+ working professionals and students in AI/ML. Utkarsh has successfully
completed 40+ freelance training and development work/projects in data science and analytics, AI/
ML, Python development, and SQL. He hails from Lucknow and is currently settled in Bangalore,
India, as an analyst at Deloitte USI Consulting.
I would like to thank my mother, Mrs. Rupam Srivastava, for her continuous guidance and support
throughout my hardships and struggles. Thanks also to the Supreme Para-Brahman.
Neeraj Jhaveri is a cloud solution architect at Microsoft with expertise in providing data and AI
solutions. He has around 20 years of IT experience. Over the last decade, working on data and
analytics, he has provided AI architect solutions on Azure. Using Azure ML and Cognitive Services,
he has helped customers move to Azure using the latest technologies. He received a master’s degree
in computer science from NYIT. He provides frequent tech talks for the fast-tracking implementation
of AI solutions in Azure.
Pooya Rezaei is an ML software engineer at Google using machine learning to estimate offline conversions
from Google Ads. Previously, he was an ML engineer at Iterable for two years optimizing their email
marketing automation platform to maximize reach. He received a B.Sc. from the University of Tehran,
an M.Sc. from the Sharif University of Technology, and a Ph.D. from the University of Vermont, all
in electrical and computer engineering.
Table of Contents
Preface xiii
Part 1 – Building a Minimum Viable Product
1
Effective Planning of Deep Learning-Driven Projects 3
Technical requirements 3 Planning a DL project 9
What is DL? 3 Defining goal and evaluation metrics 9
Understanding the role of DL in our Stakeholder identification 11
daily lives 5 Task organization 12
Resource allocation 13
Overview of DL projects 7
Defining a timeline 14
Project planning 7
Managing a project 15
Building minimum viable products 7
Building fully featured products 8 Summary 16
Deployment and maintenance 8 Further reading 17
Project evaluation 8
2
Data Preparation for Deep Learning Projects 19
Technical requirements 20 Collecting data 24
Setting up notebook environments 20 Cleaning data 27
Data preprocessing 31
Setting up a Python environment 20
Installing Anaconda 20 Extracting features from data 35
Setting up a DL project using Anaconda 21 Converting text using bag-of-words 35
Data collection, data cleaning, and Applying term frequency-inverse document
frequency (TF-IDF) transformation 36
data preprocessing 23
viii Table of Contents
Creating one-hot encoding (one-of-k) 37 Performing data visualization 42
Creating ordinal encoding 38 Performing basic visualizations using Matplotlib 43
Converting a colored image into a grayscale Drawing statistical graphs using Seaborn 45
image 39
Introduction to Docker 47
Performing dimensionality reduction 39
Introduction to Dockerfiles 47
Applying fuzzy matching to handle similarity
between strings 41 Building a custom Docker image 48
Summary 48
3
Developing a Powerful Deep Learning Model 49
Technical requirements 49 PyTorch model training 64
Going through the basic theory of DL 50 Implementing and training a model
How does DL work? 50 in TF 69
DL model training 51 TF data loading logic 69
Components of DL frameworks 53 TF model definition 71
TF model training 77
The data loading logic 53
The model definition 53 Decomposing a complex,
Model training logic 53 state-of-the-art model implementation 83
Implementing and training a model StyleGAN 84
in PyTorch 55 Implementation in PyTorch 86
Implementation in TF 91
PyTorch data loading logic 55
PyTorch model definition 57 Summary 94
4
Experiment Tracking, Model Management, and Dataset Versioning 95
Technical requirements 95 DL project tracking with MLflow and
Overview of DL project tracking 96 DVC 104
Components of DL project tracking 96 Setting up MLflow 104
Tools for DL project tracking 97 Setting up MLflow with DVC 106
DL project tracking with Weights & Dataset versioning – beyond Weights
Biases 99 & Biases, MLflow, and DVC 109
Summary 109
Setting up W&B 100
Table of Contents ix
Part 2 – Building a Fully Featured Product
5
Data Preparation in the Cloud 113
Technical requirements 113 Creating a Glue job for ETL 132
Data processing in the cloud 114 Creating a Glue Data Catalog 133
Introduction to ETL 114 Setting up a Glue context 134
Data processing system architecture 114 Reading data 135
Defining the data processing logic 136
Introduction to Apache Spark 119
Writing data 136
Resilient distributed datasets and DataFrames 120
Utilizing SageMaker for ETL 137
Loading data 121
Processing data using Spark operations 122 Creating a SageMaker notebook 140
Processing data using user-defined functions 126 Running a Spark job through a SageMaker
notebook 141
Exporting data 127
Running a job from a custom container
Setting up a single-node EC2
through a SageMaker notebook 142
instance for ETL 128
Comparing the ETL solutions in AWS 144
Setting up an EMR cluster for ETL 130
Summary 145
6
Efficient Model Training 147
Technical requirements 147 Training a TensorFlow model using SageMaker 160
Training a model on a single machine 147 Training a PyTorch model using SageMaker 161
Training a model in a distributed fashion
Utilizing multiple devices for training in
using SageMaker 162
TensorFlow 148
SageMaker with Horovod 163
Utilizing multiple devices for training in
PyTorch 150 Training a model using Horovod 164
Training a model on a cluster 151 Setting up a Horovod cluster 165
Configuring a TensorFlow training script
Model parallelism 152
for Horovod 166
Data parallelism 155
Configuring a PyTorch training script
Training a model using SageMaker 158 for Horovod 169
Setting up model training for SageMaker 159 Training a DL model on a Horovod cluster 170
