A notebook per chapter, and outputs that actually match

What it does

This repository holds one Jupyter notebook for every chapter of Deep Learning with PyTorch Step-by-Step. Each notebook contains the full code from its chapter, and the author insists that running cells sequentially should reproduce the exact outputs printed in the book. It is structured coursework rather than a reusable library: four parts covering fundamentals, computer vision, sequences, and NLP.

The interesting bit

The author recently revised the materials for PyTorch 2.x, publishing PDFs that highlight changes in Torchvision, HuggingFace, and other dependencies in red. That level of maintenance for a printed companion repo is unusual; most book repos fossilize shortly after publication.

Key highlights

• One notebook per chapter, from gradient descent basics to Transformers and NLP
• Explicitly designed for reproducibility: same outputs as the printed text
• Covers four progressive tracks: fundamentals, CNNs, RNNs/sequences, and NLP
• Revised for PyTorch 2.x with published diffs for Volume I, II, and III
• Ready-to-run Colab links for every chapter; Binder and local setups also supported

Caveats

• This is curriculum, not a library: you clone it to study, not to import it
• The README leads with a promo for a separate LLM fine-tuning book before getting to the actual content
• Chapters 4 and 11 needed significant revisions for PyTorch 2.x, so older printed editions may not match the current code

Verdict

Grab this if you are learning PyTorch from the ground up and want a guided, reproducible path from regression to Transformers. Skip it if you are hunting for drop-in model implementations or a utility package.