- Rewrite faces.py to use cv2.FaceDetectorYN instead of raw ONNX
(handles multi-scale anchor decoding and NMS internally)
- Load original photo files at up to 4000px for face detection instead
of 240px thumbnails — faces were too small to detect at thumbnail res
- Falls back to thumbnail if original is unavailable
Tested: 33 faces extracted from 13 photos, clustered into 1 person.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Rewrite all vision tasks to use sync psycopg2 sessions instead of
asyncpg — fixes 'another operation in progress' and event loop errors
when Celery forks workers sharing the async connection pool
- Letterbox-pad images to exactly 640x640 for YuNet face detector
(was crashing on non-square thumbnails)
- Deduplicate object detections per label per photo — keep highest
confidence only to avoid photo_tags PK violation on multiple
detections of the same class
- Add all queues (-Q default,high,low,vision) to worker command
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add export_models.py for OpenCLIP ViT-B/32 and YOLOv8n ONNX export
- Fix ArgMax(13) ORT ARM64 incompatibility by passing eot_indices as a
separate ONNX input (computed outside the graph in embed.py)
- Use legacy TorchScript exporter (dynamo=False) for IR version 9 compat
- Upgrade onnxruntime to 1.18.1
- Rewrite bootstrap_models.py with clear separation of auto-downloadable
models (YuNet, SFace) vs manually-exported ones (OpenCLIP, YOLOv8n)
- Wire bootstrap into worker CMD (runs before Celery)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
