Estimate Face Age and Gender (AuraFace v1)

How it fits

Typical backends

Caveats

• Hard constraintThe weights directory referenced by model is scanned recursively and must contain exactly one ONNX file; zero aborts startup with a no-weights error and two or more with an ambiguous-binding error.
• I/O contractOutput cardinality and order match the input box list, so an empty box list yields an empty output without running the network.
• CostOne inference call runs per box, so per-frame cost grows linearly with the number of faces in the frame.
• LimitationGender is reported as a single character from a binary classifier with no unknown or non-binary option, and the component is single-frame: it does not track identity across ticks.
• Parameter interactionThe reported age is the raw prediction minus age_offset; its default corrects a known systematic over-prediction on the bundled network and should be re-validated against the deployed population before changing.
• FallbackA reported age can go negative when age_offset exceeds the raw prediction, so downstream consumers should clamp or filter.
• State lifetimeage_offset is read once at startup; changing it at runtime has no effect and requires a redeploy.
• Fallbackdevice accepts 'cuda', 'cuda:N', or 'cpu'; a CUDA request on a node with no visible GPU falls back to CPU with a runtime warning.