BeefaceData/amharic-ocr-recognizer

A lightweight PP-OCRv6-inspired Amharic/Fidel line OCR recognizer with teacher–student distillation. # Amharic Line Recognizer for Fidel A from-scratch, ~24.5M-parameter CTC recognizer for Amharic sentence/line OCR, trained on the Fidel dataset (CMU Africa / Upanzi Network). Architecturally inspired by PP-OCRv6's recognition module (lightweight CNN backbone + local/global attention neck), rebuilt in PyTorch with a custom Amharic character vocabulary, since PP-OCRv6's own pretrained weights don't cover Ethiopic script. Current published SOTA on Fidel is **2.64% CER / 7.29% WER** (fine-tuned SuryaOCR, per the Fidel paper) — that's the number this project is trying to approach or beat at a fraction of the parameter count. ## Files | File | What it does | |---|---| | `audit_fidel.py` | Scans the label CSVs (+ images, if present) and reports type balance, character vocabulary, text/image stats, writer overlap, and corrupt files. Run this first against any new copy of the dataset. | | `vocab.json` | The final, cleaned 355-class character vocabulary (354 real characters + 1 CTC blank token), built from the actual audit output. This is the model's output alphabet — don't regenerate it casually, since a vocab change requires retraining from scratch. | | `model.py` | The recognizer architecture: `AmharicRecognizer`. CTC head is the only one used at inference; an auxiliary NRTR (encoder-decoder) head is trained jointly for regularization, then discarded. | | `dataset.py` | `FidelLineDataset` (reads the CSVs + images, resizes preserving aspect ratio, encodes text via `vocab.json`) and `make_writer_disjoint_split` (see **Known Issues** below). | | `train.py` | The training loop: combined CTC + NRTR loss, a 4-phase curriculum (synthetic → typed → handwritten → balanced mix), CER tracking, checkpointing. Run with `--self_test` to sanity-check the whole pipeline without real images. | | `char_vocab.txt`, `rare_chars.csv`, `summary.json` | Raw outputs from the most recent audit run (on `train_labels.csv` / `test_labels.csv` only — image-level stats not yet run, see below). | # …