How iitb-research-code/indic-trocr works
Project type and purpose
• This is an OCR research/training pipeline (no frontend or service layer).
• Purpose: train and evaluate a Transformer-based OCR model (TrOCR) specialized for handwritten Indian languages using image encoder + autoregressive text decoder.
File structure
• Top-level:
• requirements.txt – lists Python deps (Hugging Face Transformers, PyTorch, etc.)
• scripts/
– train.py – main training pipeline
– test.py – evaluation/inference script
• No separate “src” or “backend” directory: training logic and model code live inside scripts/.
README for additional insights
• Describes dataset inputs (IIIT-HW word-level images + label files)
• Explains external dependency: a language-specific RoBERTa decoder model from Hugging Face
• Shows CLI usage to train or test
Elements
• Clear component labels (e.g., “Preprocessing Module”, “Vision Encoder”)
• External services (orange clouds)
Notes on detail
• Critical files and modules (requirements.txt → environment; train.py/test.py → pipelines)
• Checkpointing and logging steps
• Where hyperparameters are configured (training_args in train.py)