ML runtime (ONNX Runtime minimal build)

Both detectors (docs/guided-view.md, docs/speech-bubbles.md) run through ONNX Runtime 1.20.0. The Maven onnxruntime-android AAR ships a 17.6 MB libonnxruntime.so with every kernel and execution provider; the app instead bundles a minimal build with only the operators the two YOLO26n graphs use, built by tools/ort/build_aar.sh into app/libs/onnxruntime-minimal-1.20.0.aar (2026-08-28).

Models

A minimal build loads only ORT-format models, and converting with the stock convert_onnx_models_to_ort constant-folds the fp16 Cast that keeps the weights small, so tools/ort/make_ort_models.py does it in three steps:

1. optimise the fp32 ONNX offline at ORT_ENABLE_EXTENDED (operator fusions such as QuickGelu for SiLU and FusedMatMul; not ALL, whose NCHWc layout transform is x86-only),
2. store every Conv weight as fp16 behind a Cast node (the fp16 weights give the same boxes as fp32 on the 147-page ground truth; weight-only int8 lost up to 0.04 F1 on single series and was rejected),
3. save as .ort with optimisations disabled, so the casts stay in the graph and run once per inference (~2 ms).

assets/models/{panels,bubbles}.ort are 5.2 MB each (9.8 MB fp32 ONNX before), stored uncompressed and copied to filesDir on first use by OnnxBoxDetector. tools/ort/required_operators.config is generated from the two .ort files (onnxruntime.tools.ort_format_model.create_config_from_models).

Rebuilding the AAR

tools/ort/build_aar.sh <work dir> creates a venv (onnxruntime 1.20.0 tools, cmake 3.31, ninja), clones v1.20.0 with submodules, refreshes the Eigen archive hash in cmake/deps.txt (GitLab re-packs the zip, so the pinned sha1 no longer matches), and runs build_aar_package.py with tools/ort/aar_build_settings.json: arm64-v8a only, minSdk 30, --minimal_build --disable_rtti --enable_lto --build_java, no NNAPI/XNNPACK (XNNPACK made no difference on these models), MinSizeRel. About 10 minutes on an M-series Mac. The JNI layer resolves ai.onnxruntime.* by name, so app/proguard-rules.pro keeps that package.

When a model changes: re-run make_ort_models.py on the fp32 exports, and if required_operators.config changes, rebuild the AAR. Bump DETECTIONS_VERSION in PanelDetector.kt either way.