How kartikgill/easter2 works
Project type and purpose
• Easter2.0 is a research‐oriented machine‐learning codebase implementing a convolutional architecture for handwritten text recognition (HTR). It provides data‐loading, model training, evaluation, and inference scripts, along with sample Jupyter notebooks.
File structure
• Top‐level directories:
• data/ ─ raw and preprocessed datasets (e.g., IAM)
• logs/ ─ training/evaluation log outputs
• notebooks/ ─ exploratory & tutorial notebooks (data download, training, testing)
• src/ ─ core code: configuration, data loader, model definition, training, and prediction
• weights/ ─ pre‐trained model checkpoints
• Key files in src/:
• config.py ─ paths, hyperparameters, dataset settings
• data_loader.py ─ data ingestion and augmentation (TACo)
• easter_model.py ─ model architecture (1D conv, residual connections, SE module) and train() entry point
• predict.py ─ batch prediction for full lines/images
• predict_line.py ─ single‐line inference CLI
• Supporting files:
• requirements.txt ─ Python dependencies (Keras/TensorFlow, numpy, etc.)
• notebooks/*.ipynb ─ examples of dataset download, training loops, evaluation metrics
README for additional insights
• Confirms the focus on 1D convolutional HTR, dense residual blocks, SE module, and TACo data augmentation.
• Explains setup steps: download weights, configure data paths, install dependencies, run train() or predict_line.py.
Details
• That Data Loader incorporates TACo augmentation
• That config.py governs all paths & hyperparameters
• Explicit separation of training vs. inference workflows