Titung/urban-surveillance-mode-classifier-finetuned-v_final

YAMNet fine-tuned end-to-end (not just a frozen-embedding head), for a personal safety wearable. Trained on a merge of two datasets:

• Titung/urban-surveillance-audio-raw (primary source; includes a fix for a crash_clips empty-audio bug -- rebuilt from Titung/car-crash-audio-clips)
• Titung/emergency-surveillance-sounds (added Fire, Car Crash, Screams, Siren, and Background examples; introduced Fire as a new 9th class not present in the original taxonomy)

Thin classes (Tire Skidding, Vehicle Mechanical Sound) were additionally oversampled during training. Classes: ['Background', 'Car Crash', 'Glass Breaking', 'Gunshots', 'Screams', 'Siren', 'Tire Skidding', 'Vehicle Mechanical Sound', 'Fire']

⚠️ Known issues in this checkpoint (not yet resolved)

• Fire had zero positive examples in the validation split (a group-aware split bug), so its alert threshold in thresholds.json/alert_policy.json is an uncalibrated default (0.5), not a real calibration. Treat Fire alerting as unreliable until retrained.
• Car Crash and Fire substantially confuse each other on the test set -- ~26% of true Car Crash examples are misclassified as Fire, dragging Fire test precision to ~0.07 despite good recall. Likely acoustic similarity + Fire's high class weight.
• Tire Skidding shows a large val/test precision gap (val precision 0.86 vs. test precision 0.19 at the same recall) -- consistent with overfitting from 4x oversampling on only ~136 source clips.
• Vehicle Mechanical Sound has only 4 test examples -- its reported metrics are not statistically meaningful at this sample size.

This is the fine-tuned counterpart to the frozen-head baseline. Compare metrics.json against that run before deciding which to ship -- fine-tuning only wins if it improves the thin classes without collapsing generalization (check the train/val macro-F1 gap in Cell 11).

Files

• saved_model/yamnet_finetuned.weights.h5 + saved_model/head.weights.h5 -- weights-only export (SavedModel export hit a known TF export bug -- see metrics.json export_format field; reload pattern is in the training notebook)
• class_names.json -- index-to-class-name mapping (model output order, 9 classes)
• thresholds.json -- per-class calibrated decision thresholds (recalibrated for this model; see known issues re: Fire)
• alert_policy.json -- per-mode (driving / trek / night_walk / standby) class thresholds
• metrics.json -- test macro-F1, per-class test F1, and the LRs/epochs used

Test performance

Macro-F1: 0.6323 (see known issues above -- several classes have specific, understood problems)