How-to Guides
doc2data is currently structured along two top-level modules and the subpackage experimental .
The top-level modules contain functionality to create PDF collections and to access contents of individual files. The modules are:
| Module | Description |
|---|---|
| Reading PDF files and creating collections | |
| utilities | Utilities |
The subpackage doc2data.experimental contains modules for feature creation and model training:
| Module | Description |
|---|---|
| preprocessing | Feature extractors for images, tokens and embeddings |
| ocr | Wrapper for the dotTR OCR package |
| base_processors | Data pipelines for model training with TensorFlow & PyTorch |
| task_processes | Task-specific processors |
| trainers | Generic model training |
| utils | Additional utilities |
Colab Notebooks
The following notebooks showcase typical applications of the above modules. The starting point is a number of PDF files with annotations. The output is a neural network for a specific document processing task.
| Colab | Notebook | Description |
|---|---|---|
 |
Document classification | Train a custom Keras model to classify pages as images on a subset of the RVL-CDIP dataset. |
|
Token classification | Fine-tune the LayoutXLM model in PyTorch to classify each token on a document page. |