A Concise Handbook of TensorFlow 2

简体中文版

繁體中文版

English Version

This is a concise handbook of TensorFlow 2 based on Keras and Eager Execution mode, aiming to help developers with some basic machine learning and Python knowledge to get started with TensorFlow 2 quickly.

The code of this handbook is based on TensorFlow 2.0 and 2.1 stable version. All sample code in this handbook can be found here . This handbook is a project of Google Summer of Code 2019 .

As of April 2020, a discussion board has been added at the end of each chapter. You are welcome to discuss, exchange ideas and provide suggestions at the end of the chapters.

This handbook is still in a beta version. Your suggestions are important!

GitHub： https://github.com/snowkylin/tensorflow-handbook

Q&A and discussion board: https://discuss.tf.wiki/c/share/handbook-comment-en

Preface

• Preface
  • Target readers
  • Usage
  • Acknowledgement
• TensorFlow Overview
  • For students and researchers: To build and train models
  • For developers and engineers: To call and deploy models
  • What can TensorFlow do for us?

Basic

• Installation and Environment Configuration
  • General steps for installation
  • Guide for TensorFlow GPU version installation
    • Preperations for GPU hardwares
    • Installation of NVIDIA drivers
    • Installation of CUDA Toolkit and cuDNN
  • Your first program
  • IDE configuration
  • The hardware configuration for TensorFlow *
• TensorFlow Basic
  • TensorFlow 1+1
  • Automatic differentiation mechanism
  • A basic example: Linear regression
    • Linear regression under NumPy
    • Linear regression under TensorFlow
• Model Construction and Training
  • Models and layers
  • Basic example: multi-layer perceptron (MLP)
    • Data acquisition and pre-processing with tf.keras.datasets
    • Model construction with tf.keras.Model and tf.keras.layers
    • Model training with tf.keras.losses and tf.keras.optimizer
    • Model Evaluation with tf.keras.metrics
  • Convolutional Neural Network (CNN)
    • Implementing Convolutional Neural Networks with Keras
    • Using predefined classical CNN structures in Keras
  • Recurrent Neural Network (RNN)
  • Deep Reinforcement Learning (DRL)
  • Keras Pipeline *
    • Use Keras Sequential/Functional API to build models
    • Train and evaluate models using the compile, fit and evaluate methods of Keras
  • Custom layers, losses and metrics *
    • Custom layers
    • Custom loss functions and metrics
• Common Modules in TensorFlow
  • Variable saving and restore: tf.train.Checkpoint
  • Visualization of training process: TensorBoard
    • Real-time monitoring of indicator change
    • Visualize Graph and Profile Information
    • Example: visualize the training process of MLP
  • Dataset construction and preprocessing: tf.data
    • Dataset construction
    • Dataset preprocessing
    • Increase the efficiency using the parallelization strategy of tf.data
    • Fetching elements from datasets
    • Example: cats_vs_dogs image classification
  • TFRecord: Dataset format of TensorFlow
    • Convert the dataset into a TFRecord file
    • Read the TFRecord file
  • Graph execution mode: @tf.function *
    • Basic usage of tf.function
    • Internal mechanism of tf.function
    • AutoGraph: Converting Python control flows into TensorFlow graphs
    • Using traditional tf.Session
  • TensorFlow dynamic array: tf.TensorArray *
  • Setting and allocating GPUs: tf.config *
    • Allocate GPUs for current program
    • Setting GPU memory usage policy
    • Simulating a multi-GPU environment with a single GPU

Deployment

• TensorFlow Model Export
  • Export models by SavedModel
• TensorFlow Serving
  • Installation of TensorFlow Serving
  • TensorFlow Serving models deployment
    • Keras Sequential mode models deployment
    • Custom Keras models deployment
  • Calling models deployed by TensorFlow Serving on client
    • An example of Python client
    • An example of Node.js client

Large-scale Training

• Distributed training with TensorFlow
  • Training on a single machine with multiple GPUs: MirroredStrategy
  • Training on multiple machines: MultiWorkerMirroredStrategy

Extensions

• TensorFlow Datasets: Ready-to-use Datasets
• TensorFlow Quantum: Hybrid Quantum-classical Machine Learning *
  • Basic concepts of quantum computing
    • Quantum bit
    • Quantum logic gate
    • Quantum circuit
    • Example: Create a simple circuit circuit using Cirq
  • Hybrid Quantum - Classical Machine Learning
    • Quantum datasets and quantum gates with parameters
    • Parametric quantum circuit (PQC)
    • Embedding parametric quantum circuits into machine learning models
    • Example: binary classification of quantum datasets

Indices and tables

• 索引

• 模块索引

• 搜索页面