ViT Kurdish vs Arabic Word-Image Classification

PyTorch project for classifying cropped word images as Kurdish or Arabic with:

• ViT-only classifier
• Projection-only classifier
• ViT + projection late-fusion classifier

Project Layout

• configs/: experiment configs
• src/: training, evaluation, models, datasets, and utilities
• data/: generated split CSVs at runtime
• outputs/: checkpoints, logs, and metrics at runtime
• evaluate_annotated_json.py: evaluation script for COCO-style annotated line images

Setup

Install dependencies:

pip install -r requirements.txt

Train

ViT baseline:

python train.py --config configs/vit_baseline.yaml

ViT + projection fusion:

python train.py --config configs/vit_proj_fusion.yaml

Evaluate Annotated JSON

python evaluate_annotated_json.py --config configs/vit_proj_fusion.yaml --checkpoint outputs/vit_proj_fusion/checkpoints/best.pt --json path\\to\\annotations.json --image-dir path\\to\\images --output outputs\\annotated_eval

Notes

• The provided YAML configs use local Windows dataset paths and should be edited for a new machine.
• Generated CSV splits, checkpoints, and output artifacts are intentionally excluded from Git.