Results, losses & training diagnosis

30 vs 120 epochs

Same six datasets / 1,227 weak label targets, seed, settings and excluded dataset. 30 vs 120 epoch comparison and checkpoint audit.

093828_7c8241b7 · completed_with_weak_training

Scan overlap and projected CAD labels

Automatic weak mappings. Defect names follow CSV row numbers. Poor fits and rejected annotations stay visible; only eligible datasets enter experimental training. Mapping evidence and eligibility JSON · Fit data directory · Not fit data directory

Model checks against weak labels

Training fit and held-out-design agreement are shown separately. These targets are not independently reviewed accuracy. CSV · Evidence JSON

Bold marks the best value in each metric column within its table, including ties. Lower BCE is better; higher AP and IoU are better.

# Training diagnosis: 20261020 / 093828_7c8241b7

State: **completed_with_weak_training**. Trained predictions: 280; scored records: 0; folds with epoch losses: 63.

Reviewed dataset labels: 0.

- gaussian_process: requires_training_labels (14 records): GP requires a fitted training posterior; no random/synthetic prior score emitted
- sparse_unet: unavailable_backend (14 records): RuntimeError: RuntimeError: Sparse U-Net unavailable: install optional MinkowskiEngine; dense tensors are not a sparse substitute.
- sparse_unet: unavailable_backend (28 records): RuntimeError: Sparse U-Net unavailable: install optional MinkowskiEngine; dense tensors are not a sparse substitute.

1. Review CAD-node risk labels against the exact geometry_sha256 and node order; use each dataset's review_template.json.
2. CSV X/Y/W/H are image pixels. Supply a reviewed pixel-to-CAD mapping before treating them as CAD labels.
3. Run main.py --prepared-manifest <saved prepared_manifest.json> --labels-dir <reviewed-label-folder> to retain the exact labeled geometry.
4. sparse_unet requires MinkowskiEngine. Choose available models with --models if that optional backend is absent.