Eight tracks, 339 lessons, one scaffold: a PyTorch self-study lab

What it does

DL-Hub is a Chinese-language PyTorch curriculum organized into eight tracks, from tensor basics up through LLMs and federated learning. It packages 339 lessons behind a single, unified training scaffold so students modify architectures instead of debugging boilerplate. Every lesson supports offline smoke testing with synthetic data, meaning you can verify the code runs before touching a real dataset.

The interesting bit

The author treats syllabus coverage like a regression suite: dlhub/topic_coverage.py maps every topic to concrete artifacts, and pytest checks that the registry matches reality. It is the rare tutorial repo that appears to worry about curriculum drift.

Key highlights

• Eight tracks spanning Vision, NLP, GNNs, Point Cloud, Generative models, Multimodal, LLMs, and Federated Learning
• Model zoos cataloging hundreds of architectures per domain, though the README quotes slightly different totals in different sections (Vision at 736 versus 791, NLP at 813 versus 814)
• Learning paths from a 6-lesson weekend sprint to a 319-lesson full curriculum
• Each lesson includes its own README with prerequisites and acceptance criteria
• NumPy implementations of 31 classic ML algorithms for fundamentals practice

Caveats

• The architecture counts in the Model Zoo section are internally inconsistent with the overview tables, so treat the per-domain totals as approximate.
• With 339 lessons and thousands of catalogued architectures, this is a syllabus, not a production framework; expect educational scaffolding rather than battle-hardened libraries.

Verdict

Worth bookmarking if you want a structured, runnable path from LeNet to Mamba and prefer Chinese-language guidance. Look elsewhere if you need a single-purpose research tool or a production serving layer.