Misiker101/ahlr-vt

AHLR-VT: A Hybrid CNN–Vision Transformer Architecture for Offline Amharic Handwritten Text-Line Recognition # AHLR-VT: A Hybrid CNN–Vision Transformer for Offline Amharic Handwritten Text-Line Recognition Official code for **AHLR-VT**, a hybrid CNN + ViT-Base/16 + CTC architecture for line-level offline handwritten text recognition (HTR) in Amharic (Ge'ez script), together with the depth-ablation family used for the parameter-scaled efficiency comparison in the paper. - **Paper:** AHLR-VT: A Hybrid CNN–Vision Transformer Architecture for Offline Amharic Handwritten Text-Line Recognition - **Dataset:** AHLD-29K on Hugging Face — we created a new benchmark dataset, 29,947 writer-independent, line-level handwritten Amharic text-line images from 180 writers, split 80:10:10 (train/validation/test). - **Author:** Misiker Kassahun Zewde --- ## Table of contents 1. Repository structure 2. Prerequisites 3. Step-by-step setup 4. Training 5. Evaluation & statistical validation 6. Citation 7. License --- ## 1. Repository structure ``` ahlr-vt/ ├── README.md ├── requirements.txt ├── LICENSE ├── .gitignore ├── vocab.json ├── src/ │ ├── __init__.py │ ├── dependency_check.py │ ├── dataset.py <- loads AHLD-29K from Hugging Face, builds vocab, PyTorch Dataset │ ├── model.py <- HybridCNNViT architecture + variant registry (depth 2/4/6/8/12) │ ├── train.py <- training loop with real early stopping (CLI entry point) │ ├── evaluate.py <- CTC metrics/decoding, per-line logged evaluation, efficiency profiling │ └── stats.py <- bootstrap CI, paired significance tests, confusion analysis, length robustness, greedy-vs-beam (CLI entry point) ├── ablation_suite/ # Ablation & Depth Comparison/Study │ ├── ahlr_vt_pipeline.py # Consolidated pipeline (Hybrid & Pure-ViT) │ └── README.md # Execution instructions ├── checkpoints/ <- created automatically; holds *.pth files ├── results/ <- created automatically; holds all output CSVs └── runs/ …