Deep Learning with Python (2nd Edition)

Written by François Chollet, creator of the deep learning framework Keras, this book is the Chinese translation of the second edition of "Deep Learning with Python," translated by Zhang Liang. The book has 14 chapters and continues the first edition's approach of "explaining principles through code," avoiding heaps of mathematical symbols and instead using runnable Python code and intuitive diagrams to explain the core ideas of deep learning. Chapters 1–2 outline the relationship among artificial intelligence, machine learning, and deep learning, and introduce mathematical foundations such as tensors, tensor operations, gradient descent, and backpropagation. Chapter 3 uses TensorFlow 2 and Keras to introduce the composition of neural networks and the compile, train, and evaluate interfaces. Chapters 4–7 use examples such as news classification and housing price regression to explain classification and regression basics, general machine learning principles (generalization, evaluation, regularization, hyperparameter tuning), the universal machine learning workflow, and advanced Keras usage (model subclassing, custom training loops, tf.function). Chapters 8–9 enter computer vision, covering convolution operations, training on small datasets, feature extraction and fine-tuning with pretrained models, residual connections, depthwise separable convolutions, and visual interpretation of convolutional networks. Chapters 10–11 handle time series and text respectively, explaining RNNs, LSTMs, bidirectional RNNs, bag-of-words and sequence text representations, and the Transformer architecture and sequence-to-sequence learning. Chapter 12 introduces text generation, DeepDream, neural style transfer, variational autoencoders, and generative adversarial networks. Chapter 13 discusses engineering practices such as hyperparameter optimization, model ensembling, mixed precision, and multi-GPU and TPU training. Chapter 14 reviews key concepts and discusses the limitations and future of deep learning. Compared with the first edition, this edition is substantially rewritten and adds a chapter on advanced Keras, the universal machine learning workflow, Transformer and modern network architecture patterns, and practical best practices.