hitit-cuneiform-ocr / code /BEYOND_95_IMPLEMENTATION.md
savastakan's picture
Initial upload: code + 5 record checkpoints + fuse
f211247 verified
|
Raw
History Blame Contribute Delete
6.08 kB

BEYOND_95 Implementation Status

5 quick-win strategy için kod + config + scaffolding tamamlandı. Hepsi opt-in, bağımsız, config-driven.

Implementation özet

# Strategy Status Files
1 End-to-end seq2seq paradigm ✅ Code + 83 tablet pair src/seq2seq/ + configs/seq2seq_hitit.yaml
2 3D illumination augmentation ✅ Code + test (5 tablet × 4 angle) src/illum_aug/relighting.py
3 TLHdig photo-transliteration ✅ 19.5K tablet ID + email template src/tlhdig_integration/
4 Image retrieval + exemplar ✅ Code + template src/retrieval/exemplar_matching.py
5 Active learning loop ✅ BUAL + Label Studio/Prodigy src/active_learning/al_loop.py

Kritik kazanç noktaları

1. Seq2seq paradigm (en büyük potansiyel +%3-7)

Sonuç: 83 Hitit tablet için image-transliteration pair hazır.

  • datasets/processed/seq2seq_pairs.jsonl
  • Avg 344 sign/tablet, toplam 28,556 sign
  • ViT encoder + ByT5 decoder model template: src/seq2seq/model.py
  • Config: configs/seq2seq_hitit.yaml

Next:

  • Training SLURM job (GPU gerekli)
  • python3 src/seq2seq/train.py --config configs/seq2seq_hitit.yaml
  • SumTablets baseline chrF 0.9755

2. 3D Illumination augmentation (+%2-4 paper-validated)

Sonuç: 2D Lambertian relighting pipeline (3D mesh gerekmeden).

  • Depth-from-shading proxy + Sobel normal map
  • Per tablet 8 azimuth angle → 8× augment
  • Test edildi: 5 tablet × 4 angle = 20 image
  • Tam run: 307 × 8 = 2456 augmented image

Next:

  • python3 src/illum_aug/relighting.py --n-angles 8 (background)

3. TLHdig entegrasyon

Sonuç: 19,549 unique Hitit tablet ID tespit edildi.

  • datasets/processed/tlhdig_tablet_ids.jsonl
  • Top tabletler: Privat 25, KUB 5.1+, StBoTB 1, Bo 86_299
  • Email template: datasets/processed/hpm_request_template.json

Next (kurumsal):

  • Schwemer (Würzburg) ile contact
  • CC-BY lisansı anlaşması
  • HPM foto download pipeline aktivasyonu
  • Tahmini 5K-20K tablet foto

4. Image retrieval + exemplar

Sonuç: FAISS-based ExemplarRetriever class.

  • Gallery: hitit_local 21K crop
  • Query → top-k neighbors + soft vote
  • Leakage protection: same-tablet exemplars filtrelenir
  • λ_exemplar=0.4 classifier fusion

Next:

  • DINOv3-L ile 21K embedding (GPU job)
  • Build gallery index
  • Fuse with 198-class classifier at inference

5. Active learning loop

Sonuç: BUAL-based weekly iteration pipeline.

  • BUAL score = αentropy + βdisagreement + γ*open_score
  • Diverse sampling (KMeans cluster-wise top)
  • Prodigy + Label Studio export formats
  • 50 sample/hafta → 6 ayda 2000+ yeni

Next (kurumsal):

  • Label Studio setup (ücretsiz) veya Prodigy (600€)
  • Paleograph engagement (1 saat/hafta)
  • Schedule weekly retrain

Cumulative gain projection (dürüst)

v4 baseline (full enhancements):                   93-95%
+ Seq2seq paradigm shift:              +3-7%      → 96-98%  ⭐ en büyük
+ 3D illumination augment:             +2-4%      → 97-98%
+ Exemplar retrieval (rare class):     +1-3%      → 97.5-98.5%
+ Active learning (6 ay birikimli):    +1-3%      → 98-99%
+ TLHdig photo integration:            +1-3%      → 98-99%

Theoretical ceiling: %98-99 (198 class Hitit). %99+ için: yıllar boyunca sustained data collection + expert loop.

Aktivasyon sıralaması

Hemen başlatılabilir (bugün)

# Illumination augmentation (CPU, 30-60 dk)
./hitit_ocr/scripts/beyond95_orchestration.sh

GPU job (SLURM)

# Exemplar gallery build + seq2seq training
sbatch hitit_ocr/scripts/pipeline/train_seq2seq.slurm      # TODO
sbatch hitit_ocr/scripts/pipeline/build_exemplar_gallery.slurm  # TODO

Kurumsal aksiyon (haftalar/aylar)

  1. Schwemer'e email (hpm_request_template.json) → TLHdig photo access
  2. Label Studio kurulum → active learning
  3. Paleograph angajman (ODTÜ/Ankara Üniversitesi)
  4. T.C. Kültür Bakanlığı başvurusu → müze envanter

Dosya haritası

hitit_ocr/
├── BEYOND_95.md                      # Strateji (13 yol)
├── BEYOND_95_IMPLEMENTATION.md       # BU — status
├── configs/
│   └── seq2seq_hitit.yaml             # Paradigm shift config
├── src/
│   ├── seq2seq/
│   │   ├── model.py                   # ViT + ByT5
│   │   └── build_pairs.py             # image-transliteration extractor
│   ├── illum_aug/
│   │   └── relighting.py              # 2D Lambertian multi-azimuth
│   ├── tlhdig_integration/
│   │   ├── scrape_photos.py           # XML scanner
│   │   └── hpm_web_scrape.py          # HPM URL generator
│   ├── retrieval/
│   │   └── exemplar_matching.py       # FAISS gallery
│   └── active_learning/
│       └── al_loop.py                 # BUAL + export
└── scripts/
    └── beyond95_orchestration.sh       # Master runner

datasets/processed/
├── seq2seq_pairs.jsonl                # 83 tablet
├── tlhdig_tablet_ids.jsonl            # 19,549 Hittite tablet ID
├── tlhdig_photo_pairs.jsonl           # 0 (TLHdig 0.2 yok)
├── illum_aug_map.json                 # augment log
├── hpm_request_template.json          # Schwemer contact
├── hpm_scrape_targets.jsonl           # (requires permission)
├── active_learning_recipe.json
└── tlhdig_download_recipe.json

Bilinen sınırlar

  • TLHdig 0.2 fotoğraf yok: TLHdig 1.0 (Kasım 2025) veya HPM web scrape gerek
  • Seq2seq training: 83 Hitit tablet yeter değil, pretraining (SumTablets 91K) lazım
  • Exemplar retrieval: DINOv3 embedding çıkarımı GPU job gerek
  • Active learning: Paleograph angajman organizasyonu

Referanslar

  • SumTablets ACL ML4AL 2024 (chrF 97.55)
  • Stötzner ICCVW 2023 (+%4 illumination aug)
  • Kriege ICFHR 2021 (embedded attributes retrieval)
  • Deng BUAL arXiv:2402.15198 (2024)
  • Hethitologie-Portal Mainz (Schwemer, Würzburg)