Pumpkin Book: Formula Notes for the Watermelon Book
/pumpkin-book
南瓜书 (Pumpkin Book) is worth keeping open in a second tab the moment a Watermelon Book formula stops making sense — it earns that role by doing one narrow thing, filling in derivation steps, instead of trying to re-teach machine learning from scratch. It's not a substitute for the textbook itself, and treating it as a standalone course means skipping the framing that makes the derivations make sense.
南瓜书是什么
南瓜书 (Pumpkin Book) is a companion document to Zhou Zhihua's Machine Learning textbook, known in Chinese as the Watermelon Book (西瓜书), written by self-described 'math-weak' self-learners who wanted the missing derivation steps spelled out. It doesn't teach machine learning on its own — every section maps to a formula Watermelon Book states without full proof, and the two are meant to be read side by side.
南瓜书的核心特点
- ✓Derivations written from an undergraduate math starting point, with anything beyond that level pushed into appendices and references instead of explained inline.
- ✓Coverage keyed to the January 2016 first edition of Watermelon Book, so each note maps to a specific formula in a specific chapter.
- ✓A Bilibili video tutorial series that walks through the same derivations.
- ✓Free PDF releases on GitHub for offline reading.
- ✓A public errata page for corrections found after publication.
- ✓Explicit guidance to skim, not deep-dive, the Chapter 1 and 2 formulas on a first pass.
谁适合读南瓜书
Pumpkin Book is for people already reading, or about to read, Watermelon Book who have roughly undergraduate-level calculus, linear algebra, and probability under their belt — the README pitches it at readers who can follow but not necessarily reconstruct that kind of math. Complete beginners are explicitly told to skim the Chapter 1 and 2 formulas rather than dig in, and come back to them later once more of the book makes sense.
南瓜书的正确使用姿势
- •You're reading Watermelon Book chapter by chapter and hit a formula you can't re-derive yourself.
- •You're self-studying without a TA or classmate to ask, and want a written proof rather than a video-only explanation.
- •You want a structured pace instead of studying alone — Datawhale's study-group cohorts follow the book on a schedule.
- •You already know the result of a formula but want to understand why it's true, not just how to apply it.
Strengths
- ✓Targets the exact chapters and formulas Watermelon Book skips, instead of re-explaining machine learning generally.
- ✓Free to read and download as a PDF, with a separately edited print edition for readers who want a physical copy.
- ✓Backed by more than one format — written derivations, a Bilibili video series, and a scheduled study group — not just a static text.
- ✓Maintainers say they typically reply to Issues within 24 hours, with a public errata page for known mistakes.
南瓜书的不足之处
- △Requires Watermelon Book itself to make sense of — it's built as a companion, not a standalone machine learning course.
- △Written by self-described 'math-weak' self-learners rather than professional educators, so anything past undergraduate math is deferred to outside references rather than derived in the text.
- △The README gives no indication of an English edition; readers who don't read Chinese have no documented way in.
相关学习资源推荐
Frequently asked questions
Pumpkin Book is a community-written companion to Zhou Zhihua's Machine Learning textbook, nicknamed the Watermelon Book in Chinese. It doesn't replace that textbook — it works through the same chapters and fills in the formula derivations the original leaves for readers to work out on their own.
Pumpkin Book's open-source content is free to read on GitHub and download as a PDF from the releases page. The print edition sold through JD, Dangdang, and Tmall is the same manuscript after further editing by People's Posts and Telecommunications Press, the publisher.
The README says derivations are written from an undergraduate math perspective, with anything beyond that level pushed into appendices and references. Complete beginners are told not to dig into the Chapter 1 and 2 formulas yet — skim them and come back once you've learned more.
Pumpkin Book starts wherever Zhou Zhihua's Watermelon Book states a formula without showing the steps behind it. Readers are meant to read Watermelon Book as the main text and open Pumpkin Book only when a specific derivation doesn't click.
Yes — Pumpkin Book has a companion video tutorial series hosted on Bilibili, walking through the same material as the written derivations. Datawhale also runs an organized study-group program alongside the book for readers who want a structured pace.
Not clearly documented — the README and the book's content are written in Chinese, with no English edition mentioned. Readers who don't read Chinese would need to translate the text themselves or rely on separate English-language resources for the Watermelon Book.
Related repositories
Curious whether pumpkin-book is right for you?
Let ChatGPT, Claude, or Perplexity look into it — click below and see what AI actually says about pumpkin-book.