Custom TrOCR (Not used in actual inference)

A custom made TrOCR model. This model underperformed and is for that reason not used.

---

Scripts

ocr/data_loader.py

• Defines dataset classes:
  • Bible text
  • N-grams
  • Bible with noise
  • N-grams with noise
  • Mixed datasets
• Uses ocr/image_creator.py to render text as images.

ocr/image_creator.py

• Renders text into image format.
• Supports padding, grayscale conversion, and font selection.

ocr/tokenizer.py

• Tokenizer class for:
  • Encoding text to token IDs
  • Decoding token IDs to text
  • Vocabulary management

ocr/ocr_model.py

• Custom TrOCR-based model.
• Uses a ViT encoder and an autoregressive text decoder.

ocr/train.py

• Trains the OCR model on synthetic data.
• Handles model saving, loss logging, and evaluation.

ocr/inference.py

• Loads a trained model and runs inference on new images.
• Outputs predicted text.

---