// the find
d2l-ai/d2l-en
Interactive deep learning book with multi-framework code, math, and discussions. Adopted at 500 universities from 70 countries including Stanford, MIT, Harvard, and Cambridge.
Dive into Deep Learning is a full deep learning textbook written as runnable Jupyter notebooks, with the same material implemented in parallel across PyTorch, TensorFlow, MXNet, and (partially) JAX so the math and the code sit next to each other. It's for someone learning DL from the ground up who wants working code tied to derivations, not a pure theory text or a single-framework tutorial.
Framework parity is real, not token — every major chapter has a from-scratch version and a 'concise' framework-API version, and CI actually builds and runs all four framework variants on push. Topic coverage goes well beyond the usual CNN/RNN intro into recommender systems, Gaussian processes, and hyperparameter optimization, which most competing books skip entirely. It's been battle-tested in classrooms at 500 universities, so the exercises and explanations have had years of student feedback baked in.
Last commit is August 2024, so there's nothing on the LLM/diffusion-era material that's now standard in any 2026 DL curriculum. MXNet is a dead framework at this point but still gets equal billing and maintenance weight in every chapter, which is dead weight for new readers. The repo is bloated with hundreds of .graffle diagram source files (Mac-only OmniGraffle format) checked into the tree alongside the rendered images, and the `d2l` PyPI package is just book-specific helper functions, not something worth depending on outside the book itself.