// the find
MingchaoZhu/DeepLearning
Python for《Deep Learning》，该书为《深度学习》(花书) 数学推导、原理剖析与源码级别代码实现
A companion project to the Goodfellow/Bengio/Courville 'Deep Learning' textbook that re-derives the math and implements every algorithm from scratch in NumPy — no PyTorch, no sklearn. It's aimed at students and self-learners who find the book's lack of code frustrating and want to see the equations turn into running Python.
From-scratch NumPy implementations mean you actually see what backprop, BFGS, and XGBoost are doing at the operation level, not behind a framework abstraction. Coverage is broader than most similar projects — it goes past CNNs into Bayesian optimization, Gaussian process regression, and quasi-Newton methods (DFP, BFGS, L-BFGS). The PDF companion document pairs derivations with code side-by-side, which is a more useful format than notebooks scattered across a repo. The chapter-7 additions covering L1 sparsity proofs and the full boosting family (Adaboost, GBDT, XGBoost) are the kind of thing that's hard to find explained well anywhere.
The repo is effectively abandoned — last commit was June 2020, and about half the chapters (10, 12–20) have no code or PDF at all, so the second half of the book is just a table of empty rows. The code is entirely in Chinese comments, which cuts off a significant portion of potential readers. There are no tests anywhere, so correctness of the implementations is take-it-on-faith. The file organization is messy — PDFs dumped in the root, a shallow `code/` folder with flat chapter files — making it hard to navigate without reading the README first.