A Document AI Package - Jupyter notebook tutorials

Breaking changes

With the latest release of deepdoctection v.0.33.0 the package has been refactored and is not compatible with previous releases. If you are on a previous version, please update to the latest version or choose the repo version that is tagged with v.0.32.0

Jupyter Notebooks for deepdoctection

In this repo you will find jupyter notebooks that used to be in the main repo deepdoctection. If you encouter problems, feel free to open an issue in the deepdoctection repository.

In addition, the repo contains a folder with examples that are used in the notebooks.git

Get_Started.ipynb:

• Introduction to deepdoctection
• Analyzer
• Output structure: Page, Layouts, Tables
• Saving and reading a parsed document

Pipelines.ipynb:

• Pipelines
• Analyzer configuration
• Pipeline components
• Layout detection models
• OCR matching and reading order

Analyzer_Configuration.ipynb

• Analyzer Configuration
• How to change configuration
• High level Configuration
• Layout models
• Table transformer
• Custom model
• Table segmentation
• Text extraction
• PDFPlumber
• Tesseract
• DocTr
• AWS Textract
• Word matching
• Text ordering

Analyzer_with_Table_Transformer.ipynb:

• Analyzer configuration for running Table Transformer
• General configuration
• Table segmentation

Doclaynet_with_YOLO.ipynb

• Writing a predictor from a third party library
• Adding the model wrapper for YOLO
• Adding the model to the ModelCatalog
• Modifying the factory class to build the Analyzer
• Running the Analyzer with the YoloDetector

Doclaynet_Analyzer_Config.ipynb

• Advanced Analyzer Configuration
• Adding the model wrapper for YOLO
• Configuration to parse the page with respect to granular layout segments
• Extracting figures
• Relating captions to figures and tables

Custom_Pipeline.ipynb:

• Model catalog and registries
• Predictors
• Instantiating Pipeline backbones
• Instantiating Pipelines

Datasets_and_Eval.ipynb:

• Creation of custom datasets
• Evaluation
• Fine tuning models

Data_structure.ipynb:

• Diving deeper into the data structure
• Page and Image
• ObjectTypes
• ImageAnnotation and sub categories
• Adding an ImageAnnotation
• Adding a ContainerAnnotation to an ImageAnnotation
• Sub images from given ImageAnnotation

Using_LayoutLM_for_sequence_classification.ipynb:

• Fine tuning LayoutLM for sequence classification on a custom dataset
• Evaluation
• Building and running a production pipeline

Running_pre_trained_models_from_other_libraries.ipynb

• Installing and running pre-trained models provided by Layout-Parser
• Adding new categories

The next three notebooks are experiments on a custom dataset for token classification that has been made available through Huggingface. It shows, how to train and evaluate each model of the LayoutLM family and how to track experiments with W&B.

Layoutlm_v1_on_custom_token_classification.ipynb

• LayoutLMv1 for financial report NER
• Defining object types
• Visualization and display of ground truth
• Defining Dataflow and Dataset
• Defining a split and saving the split distribution as W&B artifact
• LayoutLMv1 training
• Further exploration of evaluation
• Evaluation with confusion matrix
• Visualizing predictions and ground truth
• Evaluation on test set
• Changing training parameters and settings

Layoutlm_v2_on_custom_token_classification.ipynb

• LayoutLMv2 for financial report NER
• Defining ObjectTypes, Dataset and Dataflow
• Loading W&B artifact and building dataset split
• Exporing the language distribustion across the split
• Evaluation
• LayoutXLM for financial report NER
• Training XLM models on separate languages

Layoutlm_v3_on_custom_token_classification.ipynb

• LayoutLMv3 for financial report NER
• Evaluation
• Conclusion

To use the notebooks deepdoctection must be installed.