Initial implementation: aesthetic image scorer #1

Merged

isaac merged 1 commit from feature/aesthetic-scorer into main

2026-05-09 22:19:56 +00:00

isaac commented

2026-05-09 22:19:52 +00:00

Owner

Single Python script (aesthetic_scorer.py) with CLI interface:
extract, train-head, fine-tune, score, rank, test commands
DINOv2 ViT-G/14 + SigLIP2 ViT-gopt/16 384px ensemble (3072-dim)
Letterbox preprocessing at native resolutions (518 and 384)
Phase-gated training: frozen baseline → shallow head → fine-tune
Multi-head attention scoring head (4×128, QKV→512, FFN 512→2048→512)
Pairwise ranking loss with importance sampling (hard pairs preferred)
VRAM-adaptive cascade: sequential → checkpointing → LoRA fallback
Self-test command (letterbox shape, centroid stability, score range)
Unit tests: 21 passing tests for all non-GPU ML components

- Single Python script (aesthetic_scorer.py) with CLI interface: extract, train-head, fine-tune, score, rank, test commands - DINOv2 ViT-G/14 + SigLIP2 ViT-gopt/16 384px ensemble (3072-dim) - Letterbox preprocessing at native resolutions (518 and 384) - Phase-gated training: frozen baseline → shallow head → fine-tune - Multi-head attention scoring head (4×128, QKV→512, FFN 512→2048→512) - Pairwise ranking loss with importance sampling (hard pairs preferred) - VRAM-adaptive cascade: sequential → checkpointing → LoRA fallback - Self-test command (letterbox shape, centroid stability, score range) - Unit tests: 21 passing tests for all non-GPU ML components

isaac added 1 commit

2026-05-09 22:19:52 +00:00

Initial implementation: aesthetic image scorer ca0300101c

- Single Python script (aesthetic_scorer.py) with CLI interface:
  extract, train-head, fine-tune, score, rank, test commands
- DINOv2 ViT-G/14 + SigLIP2 ViT-gopt/16 384px ensemble (3072-dim)
- Letterbox preprocessing at native resolutions (518 and 384)
- Phase-gated training: frozen baseline → shallow head → fine-tune
- Multi-head attention scoring head (4×128, QKV→512, FFN 512→2048→512)
- Pairwise ranking loss with importance sampling (hard pairs preferred)
- VRAM-adaptive cascade: sequential → checkpointing → LoRA fallback
- Self-test command (letterbox shape, centroid stability, score range)
- Unit tests: 21 passing tests for all non-GPU ML components

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