Python Deep Learning Projects

9 projects demystifying neural network and deep learning models for building intelligent systems

Matthew Lamons, Rahul Kumar, Abhishek Nagaraja

• Éditeur: Packt Publishing
• Année: 2018
• ISBN:
  9781789134759
• Édition: 1
• Sujet: Computer Science
• Sous-sujet: Artificial Intelligence (AI) & Semantics

Insightful projects to master deep learning and neural network architectures using Python and Keras

Key Features

• Explore deep learning across computer vision, natural language processing (NLP), and image processing
• Discover best practices for the training of deep neural networks and their deployment
• Access popular deep learning models as well as widely used neural network architectures

Book Description

Deep learning has been gradually revolutionizing every field of artificial intelligence, making application development easier.

Python Deep Learning Projects imparts all the knowledge needed to implement complex deep learning projects in the field of computational linguistics and computer vision. Each of these projects is unique, helping you progressively master the subject. You'll learn how to implement a text classifier system using a recurrent neural network (RNN) model and optimize it to understand the shortcomings you might experience while implementing a simple deep learning system.

Similarly, you'll discover how to develop various projects, including word vector representation, open domain question answering, and building chatbots using seq-to-seq models and language modeling. In addition to this, you'll cover advanced concepts, such as regularization, gradient clipping, gradient normalization, and bidirectional RNNs, through a series of engaging projects.

By the end of this book, you will have gained knowledge to develop your own deep learning systems in a straightforward way and in an efficient way

What you will learn

• Set up a deep learning development environment on Amazon Web Services (AWS)
• Apply GPU-powered instances as well as the deep learning AMI
• Implement seq-to-seq networks for modeling natural language processing (NLP)
• Develop an end-to-end speech recognition system
• Build a system for pixel-wise semantic labeling of an image
• Create a system that generates images and their regions

Who this book is for

Python Deep Learning Projects is for you if you want to get insights into deep learning, data science, and artificial intelligence. This book is also for those who want to break into deep learning and develop their own AI projects.

It is assumed that you have sound knowledge of Python programming

Object Detection Using OpenCV and TensorFlow

Welcome to the second chapter focusing on computer vision in Python Deep Learning Projects (a data science pun to kick us off!). Let's think about what we accomplished in Chapter 8, Handwritten Digits Classification Using ConvNets, where we were able to train an image classifier with a convolutional neural network (CNN) to accurately classify handwritten digits in an image. What was a key characteristic of the raw data, and what was our business objective? The data was less complicated than it could have been because each image only had one handwritten digit in it and our goal was to accurately assign a digital label to the image.

What would have happened if each image had multiple handwritten digits in it? What would have happened if we had a video of the digits? What if we want to identify where the digits are in the image? These questions represent challenges that real-world data embodies, and they drive our data science innovation to new models and capabilities.

Let's expand our line of questions and imagination to the next (hypothetical) business use case for our Python deep learning project, where we're looking to build, train, and test an object detection and classification model to be used by an automobile manufacturer in their new line of self-driving cars. Autonomous vehicles need to have fundamental computer vision capabilities that you and I have organically by way of our physiology and experiential learning. We as humans can examine our field of vision and report whether or not a specific item is present and where in relation to other objects that item (if present) is located. So, if I were to ask you if you see a chicken, you'd likely say no, unless you live on a farm and are looking out your window. But if I ask you if you see a keyboard, you'd likely say yes, and could even say that the keyboard is different from other objects and is in front of the wall before you.

This is no trivial task for a computer. As Deep Learning Engineers, you are going to learn the intuition and model architecture that empowers you to build a powerful object detection and classification engine that we can envision being tested for use in autonomous vehicles. The data inputs that we're going to be working with in this chapter will be much more informationally complex than what we've had in previous projects, and the outcomes when we get them right will be that much more impressive.

So, let's get started!

Object detection intuition

When you need your application to find and name things in an image, you need to build a deep neural network for object detection. The visual field is very complex, and a camera for still images and video captures frames with many, many objects in them. Object detection is used in manufacturing for process automation in production lines; autonomous vehicles sensing pedestrians, other cars, the road, and signs, for example; and, of course, facial recognition. Computer vision solutions based on machine learning and deep learning require you, the Data Scientist, to build, train, and evaluate models that can differentiate one object from another and then accurately classify those detected objects.

As you've seen in other projects we've worked on, CNNs are very powerful models for image data. We need to look at expansions on the basic architecture that has performed so well on a single (still) image with simple information to see what works best for complex images and video.

Progress recently has been made with these networks: Faster R-CNN, region-based fully convolutional network (R-FCN), MultiBox, solid-state drive (SSD), and you only look once (YOLO). We've seen the value of these models in common consumer applications such as Google Photos and Pinterest Visual Search. We are even seeing some of these that are lightweight and fast enough to perform well on mobile devices.

Recent progress in the field can be researched with the following list of references:

• PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection, arXiv:1608.08021
• R-CNN: Rich feature hierarchies for accurate object detectio...

Table des matières

• Title Page
• Copyright and Credits
• Dedication
• Packt Upsell
• Contributors
• Preface
• Building Deep Learning Environments
• Training NN for Prediction Using Regression
• Word Representation Using word2vec
• Building an NLP Pipeline for Building Chatbots
• Sequence-to-Sequence Models for Building Chatbots
• Generative Language Model for Content Creation
• Building Speech Recognition with DeepSpeech2
• Handwritten Digits Classification Using ConvNets
• Object Detection Using OpenCV and TensorFlow
• Building Face Recognition Using FaceNet
• Automated Image Captioning
• Pose Estimation on 3D models Using ConvNets
• Image Translation Using GANs for Style Transfer
• Develop an Autonomous Agent with Deep R Learning
• Summary and Next Steps in Your Deep Learning Career
• Other Books You May Enjoy