机器学习生产系统（影印版）
机器学习生产系统（影印版）
Robert Crowe, Hannes Hapke, Emily Caveness, Di Zhu 著
出版时间：2025年04月
页数：445
“一本全面详实的书籍，为你提供了在生产环境中构建、部署、管理机器学习系统的完整流程。本书涵盖了你需要了解的方方面面，从充分利用数据，到训练模型，再到将这些模型部署到可扩展的基础设施上，以及管理细节，以确保系统的平稳运行。”
——Laurence Moroney
AI顾问、教师、作者

机器学习（ML）和人工智能（AI）领域正在蓬勃发展，几乎每天都有新的研究、模型和技术出现。面对如此丰富的选择，数据科学家、机器学习工程师和软件开发人员很容易迷失在将AI/ML模型从实验阶段推向生产的众多步骤中。
这本实用书籍专注于生产环境机器学习，指导你将ML模型转化为可行的产品和应用。生产环境机器学习涵盖了ML的所有领域，不仅限于简单的模型训练。本书特别强调了ML流水线，帮助你为ML生产系统奠定基础。
你即将开启探索之旅，学习将ML应用投入生产所需的广泛技术，以及需要考虑的问题和方法。关键的ML工程主题包括：
● 数据收集、验证、存储、特征工程
● 模型分析、服务、监控、日志记录
● 使用TensorFlow Extended（TFX）和其他工具编排机器学习流水线
本书提供了深入的实例，包括适用于自然语言处理（NLP）和计算机视觉模型的端到端机器学习流水线。
  1. Foreword
  2. Preface
  3. 1. Introduction to Machine Learning Production Systems
  4. What Is Production Machine Learning?
  5. Benefits of Machine Learning Pipelines
  6. When to Use Machine Learning Pipelines
  7. Steps in a Machine Learning Pipeline
  8. Looking Ahead
  9. 2. Collecting, Labeling, and Validating Data
  10. Important Considerations in Data Collection
  11. Responsible Data Collection
  12. Labeling Data: Data Changes and Drift in Production ML
  13. Labeling Data: Direct Labeling and Human Labeling
  14. Validating Data: Detecting Data Issues
  15. Validating Data: TensorFlow Data Validation
  16. Example: Spotting Imbalanced Datasets with TensorFlow Data Validation
  17. Conclusion
  18. 3. Feature Engineering and Feature Selection
  19. Introduction to Feature Engineering
  20. Preprocessing Operations
  21. Feature Engineering Techniques
  22. Feature Transformation at Scale
  23. Using TensorFlow Transform
  24. Feature Selection
  25. Example: Using TF Transform to Tokenize Text
  26. Conclusion
  27. 4. Data Journey and Data Storage
  28. Data Journey
  29. ML Metadata
  30. Using a Schema
  31. Enterprise Data Storage
  32. Conclusion
  33. 5. Advanced Labeling, Augmentation, and Data Preprocessing
  34. Advanced Labeling
  35. Data Augmentation
  36. Preprocessing Time Series Data: An Example
  37. Conclusion
  38. 6. Model Resource Management Techniques
  39. Dimensionality Reduction: Dimensionality Effect on Performance
  40. Quantization and Pruning
  41. Knowledge Distillation
  42. Conclusion
  43. 7. High-Performance Modeling
  44. Distributed Training
  45. Efficient Input Pipelines
  46. Training Large Models: The Rise of Giant Neural Nets and Parallelism
  47. Conclusion
  48. 8. Model Analysis
  49. Analyzing Model Performance
  50. Advanced Model Analysis
  51. Advanced Model Debugging
  52. Model Remediation
  53. Discrimination Remediation
  54. Fairness
  55. Continuous Evaluation and Monitoring
  56. Conclusion
  57. 9. Interpretability
  58. Explainable AI
  59. Model Interpretation Methods
  60. Example: Exploring Model Sensitivity with SHAP
  61. Conclusion
  62. 10. Neural Architecture Search
  63. Hyperparameter Tuning
  64. Introduction to AutoML
  65. Key Components of NAS
  66. AutoML in the Cloud
  67. Using AutoML
  68. Generative AI and AutoML
  69. Conclusion
  70. 11. Introduction to Model Serving
  71. Model Training
  72. Model Prediction
  73. Latency
  74. Throughput
  75. Cost
  76. Resources and Requirements for Serving Models
  77. Model Deployments
  78. Model Servers
  79. Managed Services
  80. Conclusion
  81. 12. Model Serving Patterns
  82. Batch Inference
  83. Introduction to Real-Time Inference
  84. Real-Time Inference Use Cases
  85. Serving Model Ensembles
  86. Data Preprocessing and Postprocessing in Real Time
  87. Inference at the Edge and at the Browser
  88. Conclusion
  89. 13. Model Serving Infrastructure
  90. Model Servers
  91. Building Scalable Infrastructure
  92. Containerization
  93. Reliability and Availability Through Redundancy
  94. Hardware Accelerators
  95. Conclusion
  96. 14. Model Serving Examples
  97. Example: Deploying TensorFlow Models with TensorFlow Serving
  98. Example: Profiling TF Serving Inferences with TF Profiler
  99. Example: Basic TorchServe Setup
  100. Conclusion
  101. 15. Model Management and Delivery
  102. Experiment Tracking
  103. Introduction to MLOps
  104. MLOps Methodology
  105. Three Types of Custom Components
  106. TFX Deep Dive
  107. Managing Model Versions
  108. Continuous Integration and Continuous Deployment
  109. Progressive Delivery
  110. Conclusion
  111. 16. Model Monitoring and Logging
  112. The Importance of Monitoring
  113. Observability in Machine Learning
  114. Logging
  115. Distributed Tracing
  116. Monitoring for Model Decay
  117. Retraining Your Model
  118. Conclusion
  119. 17. Privacy and Legal Requirements
  120. Why Is Data Privacy Important?
  121. Legal Requirements
  122. Pseudonymization and Anonymization
  123. Differential Privacy
  124. Federated Learning
  125. Encrypted ML
  126. Conclusion
  127. 18. Orchestrating Machine Learning Pipelines
  128. An Introduction to Pipeline Orchestration
  129. Pipeline Orchestration with TFX
  130. Orchestrating TFX Pipelines with Apache Beam
  131. Orchestrating TFX Pipelines with Kubeflow Pipelines
  132. Google Cloud Vertex Pipelines
  133. Choosing Your Orchestrator
  134. Alternatives to TFX
  135. Conclusion
  136. 19. Advanced TFX
  137. Advanced Pipeline Practices
  138. Custom TFX Components: Architecture and Use Cases
  139. Using Function-Based Custom Components
  140. Writing a Custom Component from Scratch
  141. Implementation Review
  142. Reusing Existing Components
  143. Creating Container-Based Custom Components
  144. Which Custom Component Is Right for You?
  145. TFX-Addons
  146. Conclusion
  147. 20. ML Pipelines for Computer Vision Problems
  148. Our Data
  149. Our Model
  150. Custom Ingestion Component
  151. Data Preprocessing
  152. Exporting the Model
  153. Our Pipeline
  154. Executing on Apache Beam
  155. Executing on Vertex Pipelines
  156. Model Deployment with TensorFlow Serving
  157. Conclusion
  158. 21. ML Pipelines for Natural Language Processing
  159. Our Data
  160. Our Model
  161. Ingestion Component
  162. Data Preprocessing
  163. Putting the Pipeline Together
  164. Executing the Pipeline
  165. Model Deployment with Google Cloud Vertex
  166. Conclusion
  167. 22. Generative AI
  168. Generative Models
  169. GenAI Model Types
  170. Agents and Copilots
  171. Pretraining
  172. Fine-Tuning
  173. Parameter-Efficient Fine-Tuning
  174. Human Alignment
  175. Prompting
  176. Chaining
  177. Retrieval Augmented Generation
  178. ReAct
  179. Evaluation
  180. LMOps
  181. GenAI Attacks
  182. Responsible GenAI
  183. Conclusion
  184. 23. The Future of Machine Learning Production Systems and Next Steps
  185. Let’s Think in Terms of ML Systems, Not ML Models
  186. Bringing ML Systems Closer to Domain Experts
  187. Privacy Has Never Been More Important
  188. Conclusion
  189. Index
书名：机器学习生产系统（影印版）
国内出版社：东南大学出版社
出版时间：2025年04月
页数：445
书号：978-7-5766-2005-4
原版书书名：Machine Learning Production Systems
原版书出版商：O'Reilly Media
Robert Crowe
 
Robert Crowe是Google JAX和生成式AI产品经理，致力于帮助开发人员快速学习提高工作效率所需的知识。
 
 
Hannes Hapke
 
数据科学家，善于解决各行各业的机器学习问题，与他人合著有《自然语言处理实战》。

Hannes Hapke是Digits公司的首席机器学习工程师，也是多部机器学习相关出版物的合著者。
 
 
Emily Caveness
 
Emily Caveness是Google软件工程师，目前从事机器学习数据分析和验证工作。
 
 
Di Zhu
 
Di Zhu是Google软件工程师，参与了包括MLOps基础设施和实用机器学习解决方案在内的多个项目。
 
 
The animal on the cover of Machine Learning Production Systems is a black-throated magpie-jay (Calocitta colliei), a striking, intelligent bird species in the Corvidae, or crow, family native to Mexico’s Baja Peninsula. Black-throated magpie-jays have long, blue tail feathers. An average adult grows to about 26.6 inches in total length and 8.8 ounces. These birds can live up to 20 years.
In 2019, the IUCN Red List found that the population of black-throated magpie-jays is decreasing, but not threatened. Many of the animals on O’Reilly covers are endangered; all of them are important to the world.
购买选项
定价：167.00元
书号：978-7-5766-2005-4
出版社：东南大学出版社
联系出版社邮购

© 2026，奥莱利技术咨询(北京)有限公司。版权所有。
传真: +86-10-88097463
 京ICP备05003502号