Source code
Errata
Book Forum
Source code on GitHub
Slideshare: Put Deep Learning to good use!
PyTorch Crash Course, Part 1
PyTorch Crash Course, Part 2
PyTorch Crash Course, Part 3
A Pre-Trained Model that Fakes It until It Makes It
🎙️ Eli Stevens, Luca Antiga, and Thomas Viehmann interviewed
Register your pBook for a free eBook
FREE
You can see this entire book for free.Click the table of contents to start reading.
Every other day we hear about new ways to put deep learning to good use: improved medical imaging, accurate credit card fraud detection, long range weather forecasting, and more. PyTorch puts these superpowers in your hands, providing a comfortable Python experience that gets you started quickly and then grows with you as you—and your deep learning skills—become more sophisticated. Deep Learning with PyTorch will make that journey engaging and fun.With this publication, we finally have a definitive treatise on PyTorch. It covers the basics and abstractions in great detail.
About the Technology
Although many deep learning tools use Python, the PyTorch library is truly Pythonic. Instantly familiar to anyone who knows PyData tools like NumPy and scikit-learn, PyTorch simplifies deep learning without sacrificing advanced features. It’s excellent for building quick models, and it scales smoothly from laptop to enterprise. Because companies like Apple, Facebook, and JPMorgan Chase rely on PyTorch, it’s a great skill to have as you expand your career options. It’s easy to get started with PyTorch. It minimizes cognitive overhead without sacrificing the access to advanced features, meaning you can focus on what matters the most - building and training the latest and greatest deep learning models and contribute to making a dent in the world. PyTorch is also a snap to scale and extend, and it partners well with other Python tooling. PyTorch has been adopted by hundreds of deep learning practitioners and several first-class players like FAIR, OpenAI, FastAI and Purdue.About the book
Deep Learning with PyTorch teaches you to create neural networks and deep learning systems with PyTorch. This practical book quickly gets you to work building a real-world example from scratch: a tumor image classifier. Along the way, it covers best practices for the entire DL pipeline, including the PyTorch Tensor API, loading data in Python, monitoring training, and visualizing results. After covering the basics, the book will take you on a journey through larger projects. The centerpiece of the book is a neural network designed for cancer detection. You'll discover ways for training networks with limited inputs and start processing data to get some results. You'll sift through the unreliable initial results and focus on how to diagnose and fix the problems in your neural network. Finally, you'll look at ways to improve your results by training with augmented data, make improvements to the model architecture, and perform other fine tuning.
Table of Contents takes you straight to the bookdetailed table of contents
Part 1: Core PyTorch
1 Introducing deep learning and the PyTorch Library
1.1 The deep learning revolution
1.2 PyTorch for deep learning
1.3 Why PyTorch?
1.3.1 The deep learning competitive landscape
1.4 An overview of how PyTorch supports deep learning projects
1.5 Hardware and software requirements
1.5.1 Using Jupyter Notebooks
1.6 Exercises
1.7 Summary
2 Pretrained networks
2.1 A pretrained network that recognizes the subject of an image
2.1.1 Obtaining a pretrained network for image recognition
2.1.2 AlexNet
2.1.3 ResNet
2.1.4 Ready, set, almost run
2.1.5 Run!
2.2 A pretrained model that fakes it until it makes it
2.2.1 The GAN game
2.2.2 CycleGAN
2.2.3 A network that turns horses into zebras
2.3 A pretrained network that describes scenes
2.3.1 NeuralTalk2
2.4 Torch Hub
2.5 Conclusion
2.6 Exercises
2.7 Summary
3 It starts with a tensor
3.1 The world as floating-point numbers
3.2 Tensors: Multidimensional arrays
3.2.1 From Python lists to PyTorch tensors
3.2.2 Constructing our first tensors
3.2.3 The essence of tensors
3.3 Indexing tensors
3.4 Named tensors
3.5 Tensor element types
3.5.1 Specifying the numeric type with dtype
3.5.2 A dtype for every occasion
3.5.3 Managing a tensor’s dtype attribute
3.6 The tensor API
3.7 Tensors: Scenic views of storage
3.7.1 Indexing into storage
3.7.2 Modifying stored values: In-place operations
3.8 Tensor metadata: Size, offset, and stride
3.8.1 Views of another tensor’s storage
3.8.2 Transposing without copying
3.8.3 Transposing in higher dimensions
3.8.4 Contiguous tensors
3.9 Moving tensors to the GPU
3.9.1 Managing a tensor’s device attribute
3.10 NumPy interoperability
3.11 Generalized tensors are tensors, too
3.12 Serializing tensors
3.12.1 Serializing to HDF5 with h5py
3.13 Conclusion
3.14 Exercises
3.15 Summary
4 Real-world data representation using tensors
4.1 Working with images
4.1.1 Adding color channels
4.1.2 Loading an image file
4.1.3 Changing the layout
4.1.4 Normalizing the data
4.2 3D images: Volumetric data
4.2.1 Loading a specialized format
4.3 Representing tabular data
4.3.1 Using a real-world dataset
4.3.2 Loading a wine data tensor
4.3.3 Representing scores
4.3.4 One-hot encoding
4.3.5 When to categorize
4.3.6 Finding thresholds
4.4 Working with time series
4.4.1 Adding a time dimension
4.4.2 Shaping the data by time period
4.4.3 Ready for training
4.5 Representing text
4.5.1 Converting text to numbers
4.5.2 One-hot-encoding characters
4.5.3 One-hot encoding whole words
4.5.4 Text embeddings
4.5.5 Text embeddings as a blueprint
4.6 Conclusion
4.7 Exercises
4.8 Summary
5 The mechanics of learning
5.1 A timeless lesson in modeling
5.2 Learning is just parameter estimation
5.2.1 A hot problem
5.2.2 Gathering some data
5.2.3 Visualizing the data
5.2.4 Choosing a linear model as a first try
5.3 Less loss is what we want
5.3.1 From problem back to PyTorch
5.4 Down along the gradient
5.4.1 Decreasing loss
5.4.2 Getting analytical
5.4.3 Iterating to fit the model
5.4.4 Normalizing inputs
5.4.5 Visualizing (again)
5.5 PyTorch’s autograd: Backpropagating all things
5.5.1 Computing the gradient automatically
5.5.2 Optimizers a la carte
5.5.3 Training, validation, and overfitting
5.5.4 Autograd nits and switching it off
5.6 Conclusion
5.7 Exercise
5.8 Summary
6 Using a neural network to fit the data
6.1 Artificial neurons
6.1.1 Composing a multilayer network
6.1.2 Understanding the error function
6.1.3 All we need is activation
6.1.4 More activation functions
6.1.5 Choosing the best activation function
6.1.6 What learning means for a neural network
6.2 The PyTorch nn module
6.2.1 Using call rather than forward
6.2.2 Returning to the linear model
6.3 Finally a neural network
6.3.1 Replacing the linear model
6.3.2 Inspecting the parameters
6.3.3 Comparing to the linear model
6.4 Conclusion
6.5 Exercises
6.6 Summary
7 Telling birds from airplanes: Learning from images
7.1 A dataset of tiny images
7.1.1 Downloading CIFAR-10
7.1.2 The Dataset class
7.1.3 Dataset transforms
7.1.4 Normalizing data
7.2 Distinguishing birds from airplanes
7.2.1 Building the dataset
7.2.2 A fully connected model
7.2.3 Output of a classifier
7.2.4 Representing the output as probabilities
7.2.5 A loss for classifying
7.2.6 Training the classifier
7.2.7 The limits of going fully connected
7.3 Conclusion
7.4 Exercises
7.5 Summary
8 Using convolutions to generalize
8.1 The case for convolutions
8.1.1 What convolutions do
8.2 Convolutions in action
8.2.1 Padding the boundary
8.2.2 Detecting features with convolutions
8.2.3 Looking further with depth and pooling
8.2.4 Putting it all together for our network
8.3 Subclassing nn.Module
8.3.1 Our network as an nn.Module
8.3.2 How PyTorch keeps track of parameters and submodules
8.3.3 The functional API
8.4 Training our convnet
8.4.1 Measuring accuracy
8.4.2 Saving and loading our model
8.4.3 Training on the GPU
8.5 Model design
8.5.1 Adding memory capacity: Width
8.5.2 Helping our model to converge and generalize: Regularization
8.5.3 Going deeper to learn more complex structures: Depth
8.5.4 Comparing the designs from this section
8.5.5 It’s already outdated
8.6 Conclusion
8.7 Exercises
8.8 Summary
Part 2: Learning from images in the real world: Early detection of lung cancer
9 Using PyTorch to fight cancer
9.1 Introduction to the use case
9.2 Preparing for a large-scale project
9.3 What is a CT scan, exactly?
9.4 The project: An end-to-end detector for lung cancer
9.4.1 Why can’t we just throw data at a neural network until it works?
9.4.2 What is a nodule?
9.4.3 Our data source: The LUNA Grand Challenge
9.4.4 Downloading the LUNA data
9.5 Conclusion
9.6 Summary
10 Combining data sources into a unified dataset
10.1 Raw CT data files
10.2 Parsing LUNA’s annotation data
10.2.1 Training and validation sets
10.2.2 Unifying our annotation and candidate data
10.3 Loading individual CT scans
10.3.1 Hounsfield Units
10.4 Locating a nodule using the patient coordinate system
10.4.1 The patient coordinate system
10.4.2 CT scan shape and voxel sizes
10.4.3 Converting between millimeters and voxel addresses
10.4.4 Extracting a nodule from a CT scan
10.5 A straightforward dataset implementation
10.5.1 Caching candidate arrays with the getCtRawCandidate function
10.5.2 Constructing our dataset in LunaDataset.init
10.5.3 A training/validation split
10.5.4 Rendering the data
10.6 Conclusion
10.7 Exercises
10.8 Summary
11 Training a classification model to detect suspected tumors
11.1 A foundational model and training loop
11.2 The main entry point for our application
11.3 Pretraining setup and initialization
11.3.1 Initializing the model and optimizer
11.3.2 Care and feeding of data loaders
11.4 Our first-pass neural network design
11.4.1 The core convolutions
11.4.2 The full model
11.5 Training and validating the model
11.5.1 The computeBatchLoss function
11.5.2 The validation loop is similar
11.6 Outputting performance metrics
11.6.1 The logMetrics function
11.7 Running the training script
11.7.1 Needed data for training
11.7.2 Interlude: The enumerateWithEstimate function
11.8 Evaluating the model: Getting 99.7% correct means we’re done, right?
11.9 Graphing training metrics with TensorBoard
11.9.1 Running TensorBoard
11.9.2 Adding TensorBoard support to the metrics logging function
11.10 Why isn’t the model learning to detect nodules?
11.11 Conclusion
11.12 Exercises
11.13 Summary
12 Improving training with metrics and augmentation
12.1 High-level plan for improvement
12.2 Good dogs vs. bad guys: False positives and false negatives
12.3 Graphing the positives and negatives
12.3.1 Recall is Roxie’s strength
12.3.2 Precision is Preston’s forte
12.3.3 Implementing precision and recall in logMetrics
12.3.4 Our ultimate performance metric: The F1 score
12.3.5 How does our model perform with our new metrics?
12.4 What does an ideal dataset look like?
12.4.1 Making the data look less like the actual and more like the “ideal”
12.4.2 Contrasting training with a balanced LunaDataset to previous runs
12.4.3 Recognizing the symptoms of overfitting
12.5 Revisiting the problem of overfitting
12.5.1 An overfit face-to-age prediction model
12.6 Preventing overfitting with data augmentation
12.6.1 Specific data augmentation techniques
12.6.2 Seeing the improvement from data augmentation
12.7 Conclusion
12.8 Exercises
12.9 Summary
13 Using segmentationto find suspected nodules
13.1 Adding a second model to our project
13.2 Various types of segmentation
13.3 Semantic segmentation: Per-pixel classification
13.3.1 The U-Net architecture
13.4 Updating the model for segmentation
13.4.1 Adapting an off-the-shelf model to our project
13.5 Updating the dataset for segmentation
13.5.1 U-Net has very specific input size requirements
13.5.2 U-Net trade-offs for 3D vs. 2D data
13.5.3 Building the ground truth data
13.5.4 Implementing Luna2dSegmentationDataset
13.5.5 Designing our training and validation data
13.5.6 Implementing TrainingLuna2dSegmentationDataset
13.5.7 Augmenting on the GPU
13.6 Updating the training script for segmentation
13.6.1 Initializing our segmentation and augmentation models
13.6.2 Using the Adam optimizer
13.6.3 Dice loss
13.6.4 Getting images into TensorBoard
13.6.5 Updating our metrics logging
13.6.6 Saving our model
13.7 Results
13.8 Conclusion
13.9 Exercises
13.10 Summary
14 End-to-end nodule analysis, and where to go next
14.1 Towards the finish line
14.2 Independence of the validation set
14.3 Bridging CT segmentation and nodule candidate classification
14.3.1 Segmentation
14.3.2 Grouping voxels into nodule candidates
14.3.3 Did we find a nodule? Classification to reduce false positives
14.4 Quantitative validation
14.5 Predicting malignancy
14.5.1 Getting malignancy information
14.5.2 An area under the curve baseline: Classifying by diameter
14.5.3 Reusing preexisting weights: Fine-tuning
14.5.4 More output in TensorBoard
14.6 What we see when we diagnose
14.6.1 Training, validation, and test sets
14.7 What next? Additional sources of inspiration (and data)
14.7.1 Preventing overfitting: Better regularization
14.7.2 Refined training data
14.7.3 Competition results and research papers
14.8 Conclusion
14.8.1 Behind the curtain
14.9 Exercises
14.10 Summary
Part 3: Deployment
15 Deploying to production
15.1 Serving PyTorch models
15.1.1 Our model behind a Flask server
15.1.2 What we want from deployment
15.1.3 Request batching
15.2 Exporting models
15.2.1 Interoperability beyond PyTorch with ONNX
15.2.2 PyTorch’s own export: Tracing
15.2.3 Our server with a traced model
15.3 Interacting with the PyTorch JIT
15.3.1 What to expect from moving beyond classic Python/PyTorch
15.3.2 The dual nature of PyTorch as interface and backend
15.3.3 TorchScript
15.3.4 Scripting the gaps of traceability
15.4 LibTorch: PyTorch in C++
15.4.1 Running JITed models from C++
15.4.2 C from the start: The C API
15.5 Going mobile
15.5.1 Improving efficiency: Model design and quantization
15.6 Emerging technology: Enterprise serving of PyTorch models
15.7 Conclusion
15.8 Exercises
15.9 Summary
What's inside
- Training deep neural networks
- Implementing modules and loss functions
- Utilizing pretrained models from PyTorch Hub
- Exploring code samples in Jupyter Notebooks
About the authors
Eli Stevens had roles from software engineer to CTO, and is currently working on machine learning in the self-driving-car industry. Luca Antiga is cofounder of an AI engineering company and an AI tech startup, as well as a former PyTorch contributor. Thomas Viehmann is a PyTorch core developer and machine learning trainer and consultant. consultant based in Munich, Germany and a PyTorch core developer.
Deep Learning with PyTorch authors Luca Antiga (L) and Eli Stevenson (R) eating dessert in San Francisco's Mission District with the book's editor Frances Lefkowitz. Luca is from Bergamo, Italy, Eli lives in San Jose, and Frances hails from San Francisco.
