How factodeeplearning/meta-dan works
Meta-DAN is a PyTorch handwritten-text-recognition research package supporting page-level and line-level models. Users format a dataset or supply an image, choose a model through the CLI, then train, evaluate, or predict; training can use synthetic samples and pretrained weights, and recognition output can include attention visualizations. The graph preserves the README’s model families and dataset-formatting capability. Some subsystem links are documented only by README or the SOURCE INDEX; unsampled implementation details are not asserted.
Main entry points: run.py Model runner · training_manager.py Training engine · ocr_dataset_manager.py OCR datasets · cli.py CLI
Components 22
Entry and formatting
- CLI
dar/cli.py - Dataset formatters
dar/dataset_format - Model runner
dar/text_recognition/run.py
Datasets and synthesis
- OCR datasets
dar/text_recognition/ocr_dataset_manager.py - Dataset loading
dar/generic/dataset_manager.py - Synthetic generation
dar/text_recognition/synthetic/synthetic_manager.py - Synthetic content
dar/text_recognition/synthetic
Recognition models
- Tokenizer
dar/text_recognition/tokenizer.py - Line CTC
dar/text_recognition/line/ctc - DAN
dar/text_recognition/document/dan - Faster DAN
dar/text_recognition/document/faster_dan - W-DAN
dar/text_recognition/document/w_dan - MT-DAN
dar/text_recognition/document/mt_dan - Meta-DAN
dar/text_recognition/document/meta_dan
Training and evaluation
- OCR manager
dar/text_recognition/ocr_manager.py - Training engine
dar/generic/training_manager.py - Metrics
dar/generic/metric_manager.py - Schedulers
dar/generic/scheduler.py - Pretrained weights
dar/text_recognition/download_pretrained_weights.py - Predictions
Other components
- User
- Hugging Face
Connections 22
- User → CLI: invokes
- CLI → Model runner: dispatches model
dar/cli.py - CLI → Dataset formatters: dispatches format
dar/cli.py - Dataset formatters → OCR datasets: formats dataset for
- Model runner → Line CTC: selects model
dar/text_recognition/run.py - Model runner → DAN: selects model
dar/text_recognition/run.py - Model runner → Faster DAN: selects model
dar/text_recognition/run.py - Model runner → W-DAN: selects model
dar/text_recognition/run.py - Model runner → MT-DAN: selects model
dar/text_recognition/run.py - Model runner → Meta-DAN: selects model
dar/text_recognition/run.py - Model runner → OCR manager: creates manager
dar/text_recognition/run.py - OCR manager → Training engine: extends
dar/text_recognition/ocr_manager.py - Training engine → OCR datasets: uses dataset
dar/text_recognition/ocr_manager.py - OCR datasets → Dataset loading: extends dataset
dar/text_recognition/ocr_dataset_manager.py - OCR datasets → Synthetic generation: requests samples
dar/text_recognition/ocr_dataset_manager.py - Synthetic generation → Synthetic content: uses generators
dar/text_recognition/synthetic/synthetic_manager.py - Training engine → Metrics: uses metrics
dar/generic/training_manager.py - Training engine → Schedulers: uses schedulers
dar/generic/training_manager.py - Training engine → Pretrained weights: downloads weights
dar/generic/training_manager.py - Pretrained weights → Hugging Face: retrieves checkpoints
- OCR manager → Predictions: writes results
dar/text_recognition/ocr_manager.py - User → Predictions: requests prediction