inference crate

Burn-backed (or heuristic) detector factory and inference plugin for Bevy apps.

Details

• Backend: defaults to backend-ndarray; enable --features backend-wgpu for WGPU. Needs burn features enabled in the root build if you want GPU.
• Model: loads TinyDet (default) or BigDet from the shared models crate via BinFileRecorder (full precision). Pass a weights path to the factory to load a checkpoint; otherwise it falls back to a heuristic detector.
• Use: app orchestrators insert the detector built by inference::InferenceFactory when mode==Inference. Ensure the checkpoint exists and matches the model config.
• Smoke: unit test ensures fallback when no weights are provided. Add an integration test pointing at a real checkpoint once available.

License

Apache-2.0 (see LICENSE in the repo root).