Title Page
Copyright and Credits
Hands-On Convolutional Neural Networks with TensorFlow
Packt Upsell
Why subscribe?
PacktPub.com
Contributors
About the authors
Packt is searching for authors like you
Preface
Who this book is for
What this book covers
To get the most out of this book
Download the example code files
Conventions used
Get in touch
Reviews
Setup and Introduction to TensorFlow
The TensorFlow way of thinking
Setting up and installing TensorFlow
Conda environments
Checking whether your installation works
TensorFlow API levels
Eager execution
Building your first TensorFlow model
One-hot vectors
Splitting into training and test sets
Creating TensorFlow graphs
Variables
Operations
Feeding data with placeholders
Initializing variables
Training our model
Loss functions
Optimization
Evaluating a trained model
The session
Summary
Deep Learning and Convolutional Neural Networks
AI and ML
Types of ML
Old versus new ML
Artificial neural networks
Activation functions
The XOR problem
Training neural networks
Backpropagation and the chain rule
Batches
Loss functions
The optimizer and its hyperparameters
Underfitting versus overfitting
Feature scaling
Fully connected layers
A TensorFlow example for the XOR problem
Convolutional neural networks
Convolution
Input padding
Calculating the number of parameters (weights)
Calculating the number of operations
Converting convolution layers into fully connected layers
The pooling layer
1x1 Convolution
Calculating the receptive field
Building a CNN model in TensorFlow
TensorBoard
Other types of convolutions
Summary
Image Classification in TensorFlow
CNN model architecture
Cross-entropy loss (log loss)
Multi-class cross entropy loss
The train/test dataset split
Datasets
ImageNet
CIFAR
Loading CIFAR
Image classification with TensorFlow
Building the CNN graph
Learning rate scheduling
Introduction to the tf.data API
The main training loop
Model Initialization
Do not initialize all weights with zeros
Initializing with a mean zero distribution
Xavier-Bengio and the Initializer
Improving generalization by regularizing
L2 and L1 regularization
Dropout
The batch norm layer
Summary
Object Detection and Segmentation
Image classification with localization
Localization as regression
TensorFlow implementation
Other applications of localization
Object detection as classification – Sliding window
Using heuristics to guide us (R-CNN)
Problems
Fast R-CNN
Faster R-CNN
Region Proposal Network
RoI Pooling layer
Conversion from traditional CNN to Fully Convnets
Single Shot Detectors – You Only Look Once
Creating training set for Yolo object detection
Evaluating detection (Intersection Over Union)
Filtering output
Anchor Box
Testing/Predicting in Yolo
Detector Loss function (YOLO loss)
Loss Part 1
Loss Part 2
Loss Part 3
Semantic segmentation
Max Unpooling
Deconvolution layer (Transposed convolution)
The loss function
Labels
Improving results
Instance segmentation
Mask R-CNN
Summary
VGG, Inception Modules, Residuals, and MobileNets
Substituting big convolutions
Substituting the 3x3 convolution
VGGNet
Architecture
Parameters and memory calculation
Code
More about VGG
GoogLeNet
Inception module
More about GoogLeNet
Residual Networks
MobileNets
Depthwise separable convolution
Control parameters
More about MobileNets
Summary
Autoencoders, Variational Autoencoders, and Generative Adversarial Networks
Why generative models
Autoencoders
Convolutional autoencoder example
Uses and limitations of autoencoders
Variational autoencoders
Parameters to define a normal distribution
VAE loss function
Kullback-Leibler divergence
Training the VAE
The reparameterization trick
Convolutional Variational Autoencoder code
Generating new data
Generative adversarial networks
The discriminator
The generator
GAN loss function
Generator loss
Discriminator loss
Putting the losses together
Training the GAN
Deep convolutional GAN
WGAN
BEGAN
Conditional GANs
Problems with GANs
Loss interpretability
Mode collapse
Techniques to improve GANs' trainability
Minibatch discriminator
Summary
Transfer Learning
When?
How? An overview
How? Code example
TensorFlow useful elements
An autoencoder without the decoder
Selecting layers
Training only some layers
Complete source
Summary
Machine Learning Best Practices and Troubleshooting
Building Machine Learning Systems
Data Preparation
Split of Train/Development/Test set
Mismatch of the Dev and Test set
When to Change Dev/Test Set
Bias and Variance
Data Imbalance
Collecting more data
Look at your performance metric
Data synthesis/Augmentation
Resample Data
Loss function Weighting
Evaluation Metrics
Code Structure best Practice
Singleton Pattern
Recipe for CNN creation
Summary
Training at Scale
Storing data in TFRecords
Making a TFRecord
Storing encoded images
Sharding
Making efficient pipelines
Parallel calls for map transformations
Getting a batch
Prefetching
Tracing your graph
Distributed computing in TensorFlow
Model/data parallelism
Synchronous/asynchronous SGD
When data does not fit on one computer
The advantages of NoSQL systems
Installing Cassandra (Ubuntu 16.04)
The CQLSH tool
Creating databases, tables, and indexes
Doing queries in Python
Populating tables in Python
Doing backups
Scaling computation in the cloud
EC2
AMI
Storage (S3)
SageMaker
Summary
References
Chapter 1
Chapter 2
Chapter 3
Chapter 4
Chapter 5
Chapter 7
Chapter 9
Other Books You May Enjoy
Leave a review - let other readers know what you think
· · · · · · (收起)