// the find
d2l-ai/d2l-zh
《动手学深度学习》：面向中文读者、能运行、可讨论。中英文版被70多个国家的500多所大学用于教学。
D2L-zh is the Chinese edition of 'Dive into Deep Learning', a full textbook covering ML fundamentals through transformers with runnable PyTorch/TensorFlow/MXNet/PaddlePaddle code. It's aimed at Chinese-speaking practitioners and students who want theory paired with working code, not just API tutorials. Used at 500+ universities across 70+ countries.
1. Four framework implementations (PyTorch, TensorFlow, MXNet, PaddlePaddle) maintained in parallel — same concepts, same chapter structure, so you can follow along in whichever you use day-to-day. 2. Notebooks are genuinely runnable, not illustration-only: each concept is implemented from scratch first, then shown with high-level APIs, which actually builds understanding rather than hiding it. 3. Coverage depth is real — goes from linear regression all the way through BERT pretraining, attention mechanisms, and multi-GPU training, not just the toy examples most tutorials stop at. 4. The d2l utility library is minimal and honest — it's just shared helpers for the book, not a framework, so you're never confused about what's 'real' code versus scaffolding.
1. Last pushed July 2024 and the content largely predates the LLM/diffusion era — there's no coverage of transformer fine-tuning workflows, LoRA, or anything post-BERT, which is where most practitioners are working now. 2. MXNet chapters are effectively dead weight — Apache MXNet is retired and the MXNet code paths in this repo have no practical value for anyone starting today. 3. The multi-framework parallelism is a maintenance burden that shows: PaddlePaddle and TensorFlow chapters lag behind the PyTorch versions and some sections are clearly less polished. 4. Chinese-only for the primary content means non-Chinese readers get the English repo (d2l-en) instead, so this repo's audience is inherently limited and contribution friction is higher for the broader open-source community.