How dali92002/docentr works
Project type and purpose
• DocEnTR is a Python-based, open-source machine-learning tool that implements an end-to-end document image enhancement transformer (DocEnTr) on top of the vit-pytorch library.
• Its primary purpose is to ingest degraded document images, preprocess them into patches, train/fine-tune a Vision Transformer model for binarization, and produce enhanced (binarized) outputs.
File structure
• Top level:
• config.py, cog.yaml — configuration and inference spec
• load_data.py, process_dibco.py — prepare/split datasets (DIBCO, H-DIBCO, PALM)
• train.py — training pipeline, logging, saving weights, visualizations
• test.py / predict.py — model evaluation and inference scripts
• models/ — model definitions (binae.py defines the DocEnTr architecture)
• utils.py — helper functions (e.g., data loaders, transforms)
• demo/ + demo.ipynb — interactive notebook for single-image inference
• git_images/ — sample input/output illustrations
• There are no separate “frontend” or “backend” services—this is a monolithic ML codebase.
README for additional insights
• The README describes data preparation, training, testing, demo usage, and links to pretrained model weights.
• It highlights usage patterns (CLI args), expected directory layouts, and output artifacts (visualizations, weights folder).
Main components & interactions (
a. Data Storage
• Raw Datasets (DIBCOSETS on disk)
• Configuration (config.py, CLI args)
b. Data Processing
• process_dibco.py → splits images into patches, organizes train/val/test sets
• load_data.py + utils.py → PyTorch Dataset/DataLoader abstraction, transforms
c. Model Core
• models/binae.py → DocEnTr architecture (Vision Transformer backbone + decoder head)
• vit-pytorch library (external dependency)
d. Training Pipeline
• train.py → consumes DataLoader → forward/backward passes → logs metrics → saves best weights → produces “vis…” folder with sample outputs
e. Evaluation / Testing
• test.py / predict.py → loads saved weights → runs inference on test set → writes binarized images to “vis…”
f. Demo / Notebook
• demo.ipynb → sample single-image inference using pretrained weights
g. External Services
• Pretrained model zoo (Google Drive links)
• Cog.yaml defines API or containerized deployment (Replicate badge)
Interactions:
Data Storage → Data Processing → Model Core
↳ Training Pipeline → Saved Weights & Visualizations
↳ Evaluation → Binarized Outputs
Model Core + Saved Weights → Demo Notebook
Architectural patterns & design principles
• Clear separation of concerns: data prep vs. model definition vs. training vs. inference
• Modular code: models/ for architecture, utils/ for helpers, scripts for each major pipeline step
• CLI-driven workflows: adjustable hyperparameters without code changes
• Reproducibility: config.py + cog.yaml for environment and deployment
Be as detailed as possible
• Scripts as separate components if desired (process_dibco.py, train.py, test.py, demo notebook)
• Configuration injection (config.py, command-line args) feeding into each pipeline stage
• Feedback loops (e.g., hyperparameter tweaking → retraining)