Switch face detection to CNN model with parallel batch processing

- faces.py: use model="cnn" (more accurate, better at angles/small faces/poor
  lighting) instead of HOG; model comment explains the trade-off clearly
- faces.py: accept multiple image paths; process with ThreadPoolExecutor
  (dlib releases GIL during C++ inference → genuine thread parallelism);
  output JSON dict {path: [faces]} for batch calls
- update.rb: batch_detect_faces() collects all unprocessed images per
  directory and calls faces.py once per directory rather than once per image,
  avoiding repeated model load overhead
- update.rb: FACES_WORKERS read from config.yml faces.workers (default 4;
  set to 20 in this install's config.yml on a 64-core Xeon)
- update.rb: process_dir() now takes idx/total and prints [N/total] prefix
  on every Scanning/Skipping line for progress monitoring

To monitor a long run:
  nohup ruby /opt/albumen/scripts/update.rb > /tmp/faces_update.log 2>&1 &
  tail -f /tmp/faces_update.log

Resume/restart is fully safe: sentinel files are only written after
atomic_write_json, so an aborted directory reruns cleanly from scratch.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

0 files changed, 0 insertions, 0 deletions