Table Transformer (TATR)

A deep learning model based on object detection for extracting tables from PDFs and images.

First proposed in "PubTables-1M: Towards comprehensive table extraction from unstructured documents".

This repository also contains the official code for these papers:

• "GriTS: Grid table similarity metric for table structure recognition"
• "Aligning benchmark datasets for table structure recognition"

Note: If you are looking to use Table Transformer to extract your own tables, here are some helpful things to know:

• TATR can be trained to work well across many document domains and everything needed to train your own model is included here. But at the moment pre-trained model weights are only available for TATR trained on the PubTables-1M dataset. (See the additional documentation for how to train your own multi-domain model.)
• TATR is an object detection model that recognizes tables from image input. The inference code built on TATR needs text extraction (from OCR or directly from PDF) as a separate input in order to include text in its HTML or CSV output.

Additional information about this project for both users and researchers, including data, training, evaluation, and inference code is provided below.

News

08/22/2023: We have released 3 new pre-trained models for TATR-v1.1 (trained on 1. PubTables-1M, 2. FinTabNet.c, and 3. both datasets combined) according to the details in our paper.
04/19/2023: Our latest papers (link and link) have been accepted at ICDAR 2023.
03/09/2023: We have added more image cropping to the official training script (like we do in our most recent paper) and updated the code and environment.yml to use Python 3.10.9, PyTorch 1.13.1, and Torchvision 0.14.1, among others.
03/07/2023: We have released a new simple inference pipeline for TATR. Now you can easily detect and recognize tables from images and convert them to HTML or CSV.
03/07/2023: We have released a collection of scripts to create training data for TATR and to canonicalize pre-existing datasets, such as FinTabNet and SciTSR.
03/01/2023: New paper "Aligning benchmark datasets for table structure recognition" is now available on arXiv.
11/25/2022: We have made the full PubTables-1M dataset alternatively available for download from Hugging Face.
05/05/2022: We have released the pre-trained weights for the table structure recognition model trained on PubTables-1M.
03/23/2022: Our paper "GriTS: Grid table similarity metric for table structure recognition" is now available on arXiv
03/04/2022: We have released the pre-trained weights for the table detection model trained on PubTables-1M.
03/03/2022: "PubTables-1M: Towards comprehensive table extraction from unstructured documents" has been accepted at CVPR 2022.
11/21/2021: Our updated paper "PubTables-1M: Towards comprehensive table extraction from unstructured documents" is available on arXiv.
10/21/2021: The full PubTables-1M dataset has been officially released on . : Initial version of the Table Transformer (TATR) project is released.

PubTables-1M

The goal of PubTables-1M is to create a large, detailed, high-quality dataset for training and evaluating a wide variety of models for the tasks of table detection, table structure recognition, and functional analysis.

It contains:

• 575,305 annotated document pages containing tables for table detection.
• 947,642 fully annotated tables including text content and complete location (bounding box) information for table structure recognition and functional analysis.
• Full bounding boxes in both image and PDF coordinates for all table rows, columns, and cells (including blank cells), as well as other annotated structures such as column headers and projected row headers.
• Rendered images of all tables and pages.
• Bounding boxes and text for all words appearing in each table and page image.
• Additional cell properties not used in the current model training.

Additionally, cells in the headers are canonicalized and we implement multiple quality control steps to ensure the annotations are as free of noise as possible. For more details, please see our paper.

Pre-trained Model Weights

We provide different pre-trained models for table detection and table structure recognition.

Table Detection:

Table Structure Recognition:

Evaluation Metrics

Table Detection:

Table Structure Recognition:

Training and Evaluation Data

PubTables-1M is available for download from Microsoft Research Open Data.

We have also uploaded the full set of archives to Hugging Face.

The dataset on Microsoft Research Open Data comes in 5 tar.gz files:

• PubTables-1M-Image_Page_Detection_PASCAL_VOC.tar.gz: Training and evaluation data for the detection model
  • /images: 575,305 JPG files; one file for each page image
  • /train: 460,589 XML files containing bounding boxes in PASCAL VOC format
  • /test: 57,125 XML files containing bounding boxes in PASCAL VOC format
  • /val: 57,591 XML files containing bounding boxes in PASCAL VOC format
• PubTables-1M-Image_Page_Words_JSON.tar.gz: Bounding boxes and text content for all of the words in each page image
  • One JSON file per page image (plus some extra unused files)
• PubTables-1M-Image_Table_Structure_PASCAL_VOC.tar.gz: Training and evaluation data for the structure (and functional analysis) model
  • /images: 947,642 JPG files; one file for each page image
  • /train: 758,849 XML files containing bounding boxes in PASCAL VOC format
  • /test: 93,834 XML files containing bounding boxes in PASCAL VOC format
  • /val: 94,959 XML files containing bounding boxes in PASCAL VOC format
• PubTables-1M-Image_Table_Words_JSON.tar.gz: Bounding boxes and text content for all of the words in each cropped table image
  • One JSON file per cropped table image (plus some extra unused files)
• PubTables-1M-PDF_Annotations_JSON.tar.gz: Detailed annotations for all of the tables appearing in the source PubMed PDFs. All annotations are in PDF coordinates.
  • 401,733 JSON files; one file per source PDF

To download from the command line:

1. Visit the dataset home page with a web browser and click Download in the top left corner. This will create a link to download the dataset from Azure with a unique access token for you that looks like https://msropendataset01.blob.core.windows.net/pubtables1m?[SAS_TOKEN_HERE].
2. You can then use the command line tool azcopy to download all of the files with the following command:

azcopy copy "https://msropendataset01.blob.core.windows.net/pubtables1m?[SAS_TOKEN_HERE]" "/path/to/your/download/folder/" --recursive

Then unzip each of the archives from the command line using:

tar -xzvf yourfile.tar.gz

Code Installation

Create a conda environment from the yml file and activate it as follows

conda env create -f environment.yml
conda activate tables-detr

Model Training

The code trains models for 2 different sets of table extraction tasks:

1. Table Detection
2. Table Structure Recognition + Functional Analysis

For a detailed description of these tasks and the models, please refer to the paper.

To train, you need to cd to the src directory and specify: 1. the path to the dataset, 2. the task (detection or structure), and 3. the path to the config file, which contains the hyperparameters for the architecture and training.

To train the detection model:

python main.py --data_type detection --config_file detection_config.json --data_root_dir /path/to/detection_data