Phase 2 of ADR-103: trained count head on the existing 1,077 paired
samples (the same data that produced pose_v1 yesterday).
Honest result: 65.1% eval accuracy / 100% within ±1 / MAE 0.349 on
the held-out time-window. Per-class: 100% on "empty room" / 0% on
"1 person". The model overfit by epoch 100 (train_acc → 1.0,
eval_loss climbed 0.67 → 7.8) and the "best" checkpoint is the
snapshot that happened to predict the eval window's class
distribution (140/215 = 65.1%, matches eval_acc exactly). Confidence
head Spearman = 0.023 ⇒ uncalibrated. Same data-bound failure mode
as pose_v1 (#645), bounded by single-session training data; same
fix path (multi-room).
What v0.0.1 still validates end-to-end:
* PyTorch → safetensors → Candle Rust loads cleanly on first try.
`cog-person-count health` reports `backend: candle-cpu` and emits
real per-frame predictions instead of the stub backend's hard-coded
{1 person, 0 confidence}. Architecture parity between train-count.py
and src/inference.rs::CountNet is bit-exact.
* ONNX export bit-clean (16 KB, opset 18, dynamic batch axis).
* Training wall time: 5.6 s for 400 epochs on RTX 5080.
* Binary size unchanged (2.36 MB stripped), model loads via mmap at
runtime.
This commit ships:
* scripts/align-ground-truth.js: extended to emit n_persons_mode +
n_persons_max per window so the training pipeline has count
labels. Backwards-compatible (additive fields).
* scripts/train-count.py: new — mirrors CountNet architecture
exactly, loads paired.jsonl, trains 400 epochs with
CE+BCE+Brier loss, exports safetensors + ONNX + per-epoch JSON.
* v2/.../cog/artifacts/{count_v1.safetensors,count_v1.onnx,
count_train_results.json}: the trained artifacts.
* v2/.../cog/README.md: Status table updated with the v0.0.1 numbers
+ an Honest Caveat section explaining the data-bound result.
* docs/benchmarks/person-count-cog.md: new — full v0.0.1 benchmark
log mirroring the format docs/benchmarks/pose-estimation-cog.md
established. Includes comparison to ADR-103 v0.1.0 acceptance
gates and per-class breakdown.
Still pending:
* `run` subcommand wiring (long-running polling loop, same as pose)
* Cross-compile + sign + GCS upload (mirror of pose cog pipeline)
* Live install on cognitum-v0
* v0.2.0: re-train on multi-room data, LoRA per-room adapters,
Stoer-Wagner min-cut clip in fusion stage
Two blockers discovered while running ADR-079 P7→P8 end-to-end against
a 30-minute paired session (39,088 GT frames + 45,625 CSI frames):
1. `readFileSync(_, 'utf8').split('\n')` hit Node's `String.MaxLength`
(~512 MB) on the 750 MB CSI recording. Result:
Error: Cannot create a string longer than 0x1fffffe8 characters
Replaced loadJsonl with a 1 MiB byte-buffer streaming reader that
decodes line-by-line, so memory use stays bounded by the largest
single record.
2. The sensing-server has long since switched from the legacy `raw_csi`
/ `feature` typed records to a single `sensing_update` record per
tick (with nodes[].amplitude and top-level features). The aligner
filtered on the old types and produced 0 frames every time. Added a
`sensing_update` branch that projects each tick into rawCsi/features
entries the existing windowing code can consume, and updated
extractCsiMatrix to use already-extracted amplitudes when iqHex is
absent. timestamp is now accepted as either ISO string (legacy) or
numeric float-seconds (current).
End-to-end verified: produces 1,077 paired samples at
`--min-confidence 0.3 --window-frames 20` from the full 30-min
recording; downstream `train-wiflow-supervised.js` runs to completion.
See follow-up #640 for the PCK gap (data + GPU needed) — those are
training concerns, not aligner concerns.
rUv