apps/trainer (parking-trainer): inspect / train / evaluate / publish. Reads the wash
collector's SQLite + crops read-only off its volume; time split (validation = newest
slice); thin classes dropped; damped class weights; `features` mode (frozen ImageNet
backbone, on-disk feature cache, seconds to retrain) and `finetune` mode (light
augmentation). CPU-only torch from PyTorch's wheel index. ONNX export checked against
the torch model; NO model file below the validation floor (exit 3, report still written);
exit 2 = not enough labels. `evaluate` scores a shipped model on labels reviewed after
training + the unlabelled pile; `publish` PUTs a version folder to a Gitea generic package.
Light core deps; the `train` extra is heavy — CI syncs without it, torch tests skip.
apps/vision: BodyTypeClassifier (bodytype.onnx + sidecar = the preprocessing contract:
crop margin, input size, RGB 0-255, normalisation inside the graph) and
RefinedVehicleDetector over YOLOX — refines only `car` or a class the classifier trained
on, min-confidence, `detector_class` on the result; path set but no file = phase B off
without an error; a broken file is a health detail. models/bodytype.version (tracked,
empty) pins the published version the Dockerfile fetches at build (BuildKit secret;
a pin that cannot be fetched fails the build). Verified: a trainer model gives identical
probabilities inside the vision service; both images built and smoke-tested.
Delivery: parking-trainer image in build-images.yml, the `trainer` compose profile on the
collector stack (CPU, read-only data, TRAINER_OUT), commented TRAINER_OUT/PUBLISH_TOKEN in
the wash-collector stack, .dockerignore for both Python contexts, trainer deps synced in CI.
Wiki: bodytype-classifier-training rewritten as built (+ one fleet model not per site,
secrets/access, where the crops live), opencv-anpr-service §Phase B, vision-review-outbox,
vision-service-packaging, fleet-deployment-komodo, index, log.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
The operator's category choice is a hypothesis, not truth (user, 2026-09-06): each wash
order with a vehicle read queues a package for a trusted reviewer over the private overlay
(Netbird); the verdict becomes the phase-B training label and the per-operator error rate.
wiki/concepts/vision-review-outbox.md.
- Boxes: the vision service returns the vehicle bbox; snapshot.ts stores the vehicle and
plate boxes on the read as FRACTIONS of the analysed frame (the stored snapshot is a
downscaled copy); vehicleForIdentity() returns them.
- carwash_review_outbox (migration 0031) + review-outbox.ts: crop = detector box + 8 %
margin, ≤ 640 px, plate blurred in place from the plate box; payload carries a
pseudonymous booth id and a keyed operator hash — no site name, no plate, no OSD, no
bystanders; multipart POST with a per-booth bearer; 2xx → sent (image dropped);
400/404/413/415/422 → abandoned; anything else → backoff 1 min·2^n capped 6 h; voided
orders and items older than 14 days abandoned unsent. Nothing queued while unconfigured.
- Enqueue is fire-and-forget off the intake path in createOrder; the loop runs every
CARWASH_REVIEW_INTERVAL_SEC (60) and stops on close.
- GET /api/carwash/review/status (site:read) + a "Remote review" line in Setup → Car wash.
- Env CARWASH_REVIEW_URL / _TOKEN / _BOOTH_ID (all three or off) documented in
.env.example and forwarded by compose.
- Tests: review-outbox.test.ts (crop + blur on a synthetic frame, config/pseudonyms,
queue/drain/backoff/abandon, through the app). Wiki: new concept page, index,
venue-modules As built, log. The collector is not built.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
Fills /analyze vehicle.body_type + confidence (car / motorcycle / bus / truck from COCO,
mapped to the shared vocabulary) for the Car Wash desk's category suggestion
(venue-modules.md §Vehicle category from vision). Advisory: the operator decides, a
confident downgrade is flagged, nothing is gated on it.
- vision_service/vehicle.py: pure numpy/cv2 letterbox (pad 114, raw BGR), stride-grid
decode, class-agnostic NMS, one vehicle per frame (the box holding the plate's centre,
else the largest); YoloxVehicleDetector on onnxruntime CPU, 2 intra-op threads.
- recognizer.py: WithVehicle composes the stage over any plate recognizer (stub included);
a failing stage yields vehicle=null + a "vehicle: …" note in /health.detail — never
costs the plate read. model_version reads "<plate>+yolox:yolox_s.onnx@640".
- settings: VISION_VEHICLE_MODEL_PATH (unset = off), _INPUT_SIZE (640), _MIN_CONFIDENCE
(0.4, the detector's floor; the flag threshold is site config).
- Dockerfile bakes yolox_s.onnx (best-effort curl at build; no network → stage off) and
sets the path; compose forwards it (empty = off); .env.example documents it.
- Measured on four real dev entry frames (DS-2CD1047G3H, 2560×1440): car at 0.83–0.88 in
~240–330 ms; empty lane with a person → none.
- tests/test_vehicle.py: decode/NMS/pick/letterbox on synthetic tensors, the composition,
and a missing-model /health. Wiki: opencv-anpr-service, venue-modules, log.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
julian