What it does The TensorFlow Model Garden is a monorepo collecting reference implementations of machine-learning models. It splits into three tiers: official (Google-maintained, TF2, performance-optimized), research (researcher-maintained, TF1 or TF2), and community (curated external repos). There’s also Orbit, a small training-loop library that wraps tf.distribute for CPU/GPU/TPU.

The interesting bit The README treats “state-of-the-art” as a moving target and openly admits the research folder is only as maintained as its original authors feel like being. That honesty is rare in official project docs. The pip package (tf-models-official) lags behind master, so they ship a nightly build for the brave.

Key highlights

• official/ models are kept current with latest TF2 APIs and meant to be readable, not just fast
• Training logs on TensorBoard.dev are published “to the extent possible” — not all models qualify
• Orbit provides a forkable training-loop scaffold if Keras fit loops feel too rigid
• Installable via pip or clone-and-PYTHONPATH; Windows PowerShell instructions included
• Active “help wanted: paper implementation” label for contributors

Caveats

• The research/ directory maintenance level is explicitly “by researchers” — your mileage may vary
• NLP usage requires an extra tensorflow-text-nightly install, easy to miss
• No topics or tags on the repo itself, which makes discovery harder than it should be

Verdict Worth bookmarking if you want canonical TF2 patterns or need a baseline to beat. Skip it if you’re committed to PyTorch or need guarantees that every model is production-ready; the research tier is explicitly caveat emptor.