tools/quantise-models.sh now writes each model's Hexagon form from a
per-model table: the form its role takes on the NPU (int8, A16W8 or
A16W16), the exact graph rewrites it needs, and the nodes that must stay
float. Ranges are min/max over photographs fed exactly as the app feeds
each model -- the detector and segmenter letterboxes with their own pads
and normalisation, landmark crops from the detector's boxes, MI-GAN with a
panorama-like border, XFeat's grey proxy. The old tool used an
antialiased resize, YOLO's pad of 128 and /255 for every model that was
not a face model, none of which is what the app does.
tools/htp_graph.py holds the rewrites, each checked against the input
graph before use: the denoiser's 6-D Bayer pack and XFeat's 224-slice
unfold as SpaceToDepth (QNN stops at rank 5), computed reshape targets
folded, and bilinear Resize as two MatMuls (the HTP refuses
ResizeBilinear at XFeat's sizes). The denoiser takes ranges computed by
darkroom-denoise's gate on a smaller tile of the same network.
The first int8 files found no faces at all, and for two reasons the
tool now guards against. The calibration set was landscape photographs
with no faces in them, so the score head's ranges had never seen the
face regime; the set is now proxies from the library itself. And ONNX
Runtime's strided and moving-average calibration modes both degrade
these graphs measurably (a quarter of the faces at eight images, none
at ninety-six), while driving the calibrator in chunks by hand gives
ranges identical to a single pass — so the tool does that, four images
at a time, and feeds quantize_static through its range cache.
Measured against f32 over 400 proxies (docs/inference.md §10.1): the
10g form finds every face above 32 px the f32 form finds; 500m and
2.5g find 96%, and what they lose sits at a median confidence of 0.52
against the 0.50 threshold. Shipped with the number on record.
The Android unpack list gains the three int8 files; without that the
tablet never saw them. D13's runtime half records the reopening.
tools/quantise-models.sh writes the QDQ form QNN's HTP backend takes
whole: opset 17, per-channel int8 weights, uint8 activations, ranges
from running the f32 graph over photographs fed exactly as the app
feeds them. The calibration is strided, four images at a time, because
every ONNX Runtime calibrator holds each image's whole set of
activations until it folds them — a gigabyte an image on the 10g
detector, and an OOM kill with no message when folded once at the end.
Release-time, never on the device (docs/inference.md §5): it needs
real photographs and a person reading the recall measurement that
gates whether each file is offered.
dtourolle