What it does

onnx-tensorflow converts ONNX models into TensorFlow format so you can run inference through TensorFlow’s runtime. Load an ONNX model, call prepare(), feed it input — the output comes back from TensorFlow. There’s also a CLI (onnx-tf convert) and a programmatic API for batch conversions.

The interesting bit

The project sits in an awkward but useful spot in the ONNX ecosystem: it’s the inverse of the more actively maintained tf2onnx, which goes TensorFlow → ONNX instead. The README is admirably blunt about this — the repo “is not actively maintained and will be deprecated,” and the steering committee is openly recruiting a new owner.

Key highlights

• Supports TensorFlow 2.8.0 exactly; TF 1.x is dead
• Tests against the full ONNX Model Zoo, though inference on converted models isn’t verified — only that conversion succeeds
• Full test suite runs 15–45 minutes; model zoo tests take 1+ hour
• Requires tensorflow-probability and tensorflow-addons alongside core TF
• Code style enforced via yapf and a TensorFlow-borrowed pylint config

Caveats

• Pinning to TF 2.8.0 specifically is a fragility signal; newer TF releases may break things
• The deprecation notice is prominent — new production dependencies should be weighed carefully

Verdict

Worth a look if you’re stuck with ONNX models and a TensorFlow-only deployment environment, but treat it as maintenance-mode infrastructure. If you need the other direction (TF → ONNX), tf2onnx is where the active development lives.