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
First implementation PR for ADR-103. Same incremental shape that
ADR-101 used: scaffold the cog crate, ship a stub-backend release
that satisfies the runtime contract + 15 tests + measured cold-start,
then follow up with the trained count_v1.safetensors in a separate PR.
What ships:
* v2/crates/cog-person-count/ — new workspace member.
- Cargo.toml: candle-core/candle-nn 0.9 (cpu default, cuda feature
opt-in), safetensors, ureq, sha2 — same dep shape as the pose cog
but minus wifi-densepose-train (this cog has no training-side
consumer, so the dep tree is materially smaller → 2.36 MB
binary vs the pose cog's 4.5 MB).
- src/inference.rs: CountNet (Conv1d 56→64→128→128 encoder + count
head Linear(128→64→8)+softmax + confidence head
Linear(128→32→1)+sigmoid). Stub backend returns
`{1-person, 0-confidence}` honestly when no safetensors present.
- src/fusion.rs: fuse_confidence_weighted() — Bayesian product of
per-node distributions with confidence-weighted log-sum, plus
fuse_with_mincut_clip() hook for the v0.2.0 Stoer-Wagner
upper-bound (`ruvector-mincut` dep lands when min-cut graph
builder is ready). Confidences floored at 1e-3 and probs floored
at 1e-9 before logs — no NaN propagation.
- src/publisher.rs: emits {count, confidence, count_p95_low,
count_p95_high, n_nodes, probs} per ADR-103 §"Output".
- src/main.rs: full ADR-100 four-verb CLI (version|manifest|health
|run). The `run` subcommand explicitly returns "wiring pending
v0.0.1" so the in-process library API is the v0.0.1-clean
integration path.
- tests/smoke.rs (8 tests) + fusion::tests (7 tests, in-lib) — 15
total, all green. Cover stub-backend behaviour, wrong-shape
rejection, fusion math (empty / single / agreement / high-conf
override / normalisation), p95-range correctness, and min-cut
clip semantics.
- cog/{manifest.template.json, config.schema.json, README.md} +
cog/artifacts/ placeholder dir.
* v2/Cargo.toml: registers the new workspace member.
Verified locally:
cargo check -p cog-person-count --no-default-features → clean
cargo test -p cog-person-count --no-default-features → 8/8 pass
cargo test -p cog-person-count --lib → 7/7 pass
cargo build -p cog-person-count --release → 2.36 MB binary
./cog-person-count version → "person-count 0.3.0"
./cog-person-count manifest → JSON skeleton
./cog-person-count health → backend:stub,
count:1, conf:0,
p95:[1,1]
Cold-start: 30 sequential `health` invocations → 53.3 ms/invocation
(vs cog-pose-estimation's 76.2 ms — smaller dep tree)
cog/README.md adds:
* Security section — six-row threat table covering safetensor mmap
trust, non-finite outputs, sensing fetch failures, fusion
divide-by-zero / log-of-zero, min-cut degenerate cases, and stdout
spoofing.
* Performance / optimization section — binary size, release profile
(already opt-level=3 / lto=fat / codegen-units=1 / strip=true at
workspace level), cold-start comparison table, projected warm-path
latency budget.
Still pending (separate PRs, ADR-103 §"Migration"):
* Train count_v1.safetensors on the existing 1,077 paired samples
with `n_persons` labels (Candle on RTX 5080, same script that
produced pose_v1.safetensors yesterday).
* `run` subcommand wiring (long-running polling loop, same shape as
cog-pose-estimation::runtime).
* Cross-compile + sign + GCS upload (mirror of cog-pose-estimation
release pipeline).
* Server-side `csi.rs::score_to_person_count` call-site rewire to
consume this cog when installed; falls back to PR #491's heuristic
when not.
rUv