greek-htr-calamari

Greek HTR (calamari)

Handwritten text recognition checkpoint for Byzantine Greek manuscript line crops. This is the finetuned stable checkpoint published by nomikos-project.

The Hub file is best.pt (calamari-pytorch-v1): CNN + 2 bidirectional LSTMs, line height 48, CTC charset size 259. Some training YAML files originally said lstm_layers: 3; that setting was not applied, and the saved weights are 2 BiLSTMs.

Intended use

• Transcribe already-segmented Greek line images (grayscale or RGB; resized to height 48).
• Expert-in-the-loop HTR: the model drafts, a reader should correct.

Out of scope: full-page OCR, other scripts, or modern typeset Greek.

How to use

Resolve through the inference registry:

weights_source: hf://nomikos-project/greek-htr-calamari@stable

Prefetch:

PYTHONPATH=. python scripts/hf/fetch_model.py greek-calamari-v1 --registry-tag stable

Or load best.pt with this repo’s predictor after downloading the file:

from huggingface_hub import hf_hub_download
from src.inference.calamari import CalamariPredictor

path = hf_hub_download("nomikos-project/greek-htr-calamari", "best.pt")
CalamariPredictor([path]).predict_files(["line.png"], output_dir="preds")

Training

Two-stage CTC training on cleaned Greek GT (seed 1111, batch 32, 5 augmented copies + 1 original):

1. Pretrain 150 epochs, LR 0.001 on data/processed/greek/pretraining (18,239 train / 2,298 val / 2,274 test lines). W&B: zy1praog.
2. Finetune 100 epochs, LR 0.0001 from that pretrain on data/processed/greek/finetuning (164 / 19 / 21 lines). W&B: c7m8jnqe.

best.pt is selected on validation CER, not test.

Evaluation

Scored from this Hub best.pt with python -m src.evaluate.calamari on data/processed/greek/finetuning.

best.pt is selected on val CER (0.156). W&B eval/* during training is that same val split, not test. The test pack is very small, so test CER is noisier and higher.

Limitations

• 21-line test pack; one or two hard lines move CER a lot.
• Two very long Bessarion lines exist in pretraining GT; this model is trained as a line recognizer, not a page model.
• Predictions still need human correction, especially abbreviations, numerals, and damaged ink.

Links

• Org: nomikos-project
• Sibling models: syriac-htr-calamari, armenian-htr-calamari