Lewis Tunstall, Leandro von Werra, Thomas Wolf
“transformers相关书籍的杰作 —— 清晰易懂!”
——Jeremy Howard
——Christopher Manning
Thomas M. Siebel教授

自2017年推出以来,transformers已迅速成为在各种自然语言处理任务中实现最佳结果的主导架构。如果你是一名数据科学家或程序员,这本实践用书将向你展示如何使用Hugging FaceTransformers(基于Python的深度学习库)训练和扩展这些大型模型。
transformers已经被用来撰写真实的新闻故事、改进Google搜索查询,甚至创建会讲老套笑话的聊天机器人。在这本指南中,作者Lewis Tunstall、Leandro von Werra、Thomas Wolf(Hugging Face Transformers的创建者)通过实践方法来教你如何使用transformers以及如何将它集成到你的应用中。你将快速学习可以由transformers帮助解决的各种任务。
● 为核心NLP任务构建、调试和优化transformers模型,例如文本分类、命名实体识别和问答
● 学习如何使用transformers进行跨语言迁移学习
● 在缺乏标记数据的实际场景中应用transformers
● 使用提取、修剪和量化等技术高效部署transformers模型
● 从头开始训练transformers并学习如何扩展到多个GPU和分布式环境
  1. Foreword
  2. Preface
  3. 1. Hello Transformers
  4. The Encoder-Decoder Framework
  5. Attention Mechanisms
  6. Transfer Learning in NLP
  7. Hugging Face Transformers: Bridging the Gap
  8. A Tour of Transformer Applications
  9. The Hugging Face Ecosystem
  10. Main Challenges with Transformers
  11. Conclusion
  12. 2. Text Classification
  13. The Dataset
  14. From Text to Tokens
  15. Training a Text Classifier
  16. Conclusion
  17. 3. Transformer Anatomy
  18. The Transformer Architecture
  19. The Encoder
  20. The Decoder
  21. Meet the Transformers
  22. Conclusion
  23. 4. Multilingual Named Entity Recognition
  24. The Dataset
  25. Multilingual Transformers
  26. A Closer Look at Tokenization
  27. Transformers for Named Entity Recognition
  28. The Anatomy of the Transformers Model Class
  29. Tokenizing Texts for NER
  30. Performance Measures
  31. Fine-Tuning XLM-RoBERTa
  32. Error Analysis
  33. Cross-Lingual Transfer
  34. Interacting with Model Widgets
  35. Conclusion
  36. 5. Text Generation
  37. The Challenge with Generating Coherent Text
  38. Greedy Search Decoding
  39. Beam Search Decoding
  40. Sampling Methods
  41. Top-k and Nucleus Sampling
  42. Which Decoding Method Is Best?
  43. Conclusion
  44. 6. Summarization
  45. The CNN/DailyMail Dataset
  46. Text Summarization Pipelines
  47. Comparing Different Summaries
  48. Measuring the Quality of Generated Text
  49. Evaluating PEGASUS on the CNN/DailyMail Dataset
  50. Training a Summarization Model
  51. Conclusion
  52. 7. Question Answering
  53. Building a Review-Based QA System
  54. Improving Our QA Pipeline
  55. Going Beyond Extractive QA
  56. Conclusion
  57. 8. Making Transformers Efficient in Production
  58. Intent Detection as a Case Study
  59. Creating a Performance Benchmark
  60. Making Models Smaller via Knowledge Distillation
  61. Making Models Faster with Quantization
  62. Benchmarking Our Quantized Model
  63. Optimizing Inference with ONNX and the ONNX Runtime
  64. Making Models Sparser with Weight Pruning
  65. Conclusion
  66. 9. Dealing with Few to No Labels
  67. Building a GitHub Issues Tagger
  68. Implementing a Naive Bayesline
  69. Working with No Labeled Data
  70. Working with a Few Labels
  71. Leveraging Unlabeled Data
  72. Conclusion
  73. 10. Training Transformers from Scratch
  74. Large Datasets and Where to Find Them
  75. Building a Tokenizer
  76. Training a Model from Scratch
  77. Results and Analysis
  78. Conclusion
  79. 11. Future Directions
  80. Scaling Transformers
  81. Going Beyond Text
  82. Multimodal Transformers
  83. Where to from Here?
  84. Index
原版书书名:Natural Language Processing with Transformers
原版书出版商:O'Reilly Media
Lewis Tunstall
Lewis Tunstall是Hugging Face机器学习工程师,致力于为NLP社区开发实用工具,并帮助人们更好地使用这些工具。
Leandro von Werra
Leandro von Werra是Hugging Face机器学习工程师,致力于代码生成模型的研究与社区推广工作。
Thomas Wolf
Thomas Wolf是Hugging Face首席科学官兼联合创始人,他的团队肩负着促进AI研究和普及的使命。
The bird on the cover of Natural Language Processing with Transformers is a coconut lorikeet (Trichoglossus haematodus), a relative of parakeets and parrots. It is also known as the green-naped lorikeet and is native to Oceania.
The plumage of coconut lorikeets blends into their colorful tropical and subtropical surroundings; their green nape meets a yellow collar beneath a deep dark blue head, which ends in an orange-red bill. Their eyes are orange and the breast feathers are red. Coconut lorikeets have one of the longest, pointed tails of the seven species of lorikeet, which is green from above and yellow underneath. These birds measure 10 to 12 inches long and weigh 3.8 to 4.8 ounces.
Coconut lorikeets have one monogamous partner and lay two matte white eggs at a time. They build nests over 80 feet high in eucalyptus trees and live 15 to 20 years in the wild. This species suffers from habitat loss and capture for the pet trade.