A PyTorch textbook that actually walks past hello-world

What it does

《深入浅出PyTorch》 (“Thorough PyTorch”) is a free, open-source Chinese tutorial series from the DataWhale learning community. It covers the full arc: tensor basics and autograd, the standard training pipeline (data loading, model building, loss, optimizer, visualization), then advanced training tricks like fine-tuning with torchvision and timm, mixed-precision training, and finally deployment via ONNX. Content lives as markdown and Jupyter notebooks in the repo, with companion video lectures on Bilibili.

The interesting bit

Most intro courses stop at Fashion-MNIST. This one keeps going into chapter 10 — guided source-code readings of ResNet, Vision Transformer, Swin Transformer, RNN/LSTM, and the YOLO series (in collaboration with MMYOLO). Reading real implementations is where most tutorials quietly give up; here it’s the destination.

Key highlights

• Structured for team-based learning: part one (chapters 1–4) is paced at 10 days, part two (5–8) at 11 days
• Two hands-on projects early on: Fashion-MNIST classification and a fruit/vegetable classification notebook
• Visualization chapter covers TensorBoard, Weights & Biases, and SwanLab
• Ecosystem tour of torchvision, PyTorchVideo, torchtext, and torchaudio
• Written by a four-person team of students from Tsinghua, CAS, HIT, and Xidian

Caveats

• Entirely in Chinese — non-readers are out of luck
• The hosted online docs lag behind the repo; the README itself tells you to read the source folder markdown instead
• Part three (practical case studies) and a long list of planned topics — GANs, BERT, DETR, TorchScript, parallel training — are still on the roadmap with no dates attached
• Licensed CC BY-NC-SA 4.0, so no commercial reuse

Verdict

If you read Chinese and want a structured, exercise-heavy path from PyTorch basics to reading production-grade model code, this is one of the more complete free options. If you need English material or a finished curriculum with no gaps, wait for the later parts to land.