julian
f4b806a538
The reviewer host's collector.sqlite was created by an earlier build, before the `kind` column. CREATE TABLE IF NOT EXISTS shapes only a new database, so every query naming the column failed: the collector's /health (container unhealthy), every booth ingest, and the trainer's readiness — whose stdlib server printed the traceback and dropped the socket, which the collector could only render as "trainer not reachable: fetch failed". Nine days like that. - CollectorDb.#migrate(): PRAGMA table_info against the list of columns added since the first deploy; ALTER TABLE ADD COLUMN for each missing one (all nullable or defaulted). Append to that list whenever a column joins the CREATE. Test replays the original schema: health, ingest, stats, a legacy row reads back with the defaults. - Trainer Handler._guarded(): any unexpected exception → 500 JSON naming it, never a dropped connection; /health keeps answering. Test drives readiness against an old-schema DB. - The collector's training status proxy includes the trainer's error text. Wiki: the incident and the schema rule (vision-review-outbox), what the message means (bodytype-classifier-training), log. Deploy: the new collector migrates on start; nothing manual. Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
parking-trainer
The phase-B body-type classifier job. Reads the wash collector's volume
(collector.sqlite + crops/), trains a classifier on the reviewer's labels, and writes a
versioned model folder the vision image bakes in — or refuses when validation is below the
floor. Design and decisions: wiki/decisions/bodytype-classifier-training.md.
parking-trainer inspect --data /data # what a run would train on
parking-trainer train --data /data --out /out # features mode (minutes)
parking-trainer train --mode finetune --epochs 12 ... # full fine-tune (about an hour on 4 cores)
parking-trainer evaluate --model /out/<version>/bodytype.onnx --data /data
parking-trainer publish /out/<version> --url https://git.infra.msai.al/api/packages/mca/generic/parking-bodytype
Exit codes: 0 model written · 2 not enough labels · 3 below the floor (report written,
no model) · 1 other.
A passing run writes <out>/<version>/:
| file | what |
|---|---|
bodytype.onnx |
the classifier; input image = RGB float32 0–255 [N,3,S,S], output logits [N,K]; normalisation is inside the graph |
bodytype.json |
sidecar: version, class list (in vocabulary order), input size, crop margin, backbone, mode, label counts, validation metrics |
report.md |
the human report: accuracy, per-class recall/precision, confusion matrix, dropped classes, loss weights |
metrics.json |
the same numbers, machine-readable |
On the reviewer's host the image runs serve as the trainer service of the
wash-collector stack: a job API (/health, /readiness, /versions, /jobs) on the compose
network that the collector's Training section (/review) drives — readiness, Train /
Evaluate / Publish, reports and logs. Jobs run as subprocesses of the CLI, one at a time; state
and logs persist under /out/jobs/. The CLI stays for debugging:
docker compose -f docker-compose.collector.yml exec trainer parking-trainer inspect.
Local dev: uv sync --extra train (CPU torch, ~200 MB), uv run pytest -q. The test suite
runs without the extra (torch tests skip), matching CI.