Table extractions for COSMOS

pip install tableextractions==0.5



Interface for applying Cosmos to document segmentation

Current milestone (with demo links):

Running the standalone images

We provide a separate repo ( describing how to use our canonical docker images, which include everything necessary to run the model.

Building + running the model from scratch

It is also possible to build the model image yourself. To do so:

  1. Switch to the cosmos directory
  2. Run, specifying the PDF input and desired output directories with the INPUT_DIR and OUTPUT_DIR environment variables, respectively
OUTPUT_DIR=./output/ INPUT_DIR=/path/to/input/docs DEVICE=cpu docker-compose up

Layout of the model

Documentation can be viewed at

The entry points for the program is cosmos/

The procedure of the program is laid out generally as follows (docs correspond to paths)

  1. Preprocessing -- cosmos/preprocessing
    • Turn PDFs into PNGs so that they can be fed to a computer vision pipeline.
  2. Create proposals -- cosmos/connected_components
    • Generate region proposals within document pages, this segments each page.
  3. Ingesting data -- cosmos/ingestion
    • Prepare region proposals to be classified by a Neural Network as Body Text, Equation, Figure, etc.
  4. Model inference -- Inference runner: cosmos/infer || Model definition: cosmos/model
    • Run the Neural Network on each region proposal.
  5. Convert to HTML/XML -- cosmos/converters
    • Results are converted to HTML/XML and class specific information extraction modules are run.
  6. Postprocessing -- cosmos/postprocessing
    • Update class labels in light of extracted information.
  7. Equation specific OCR -- cosmos/latex_ocr
    • Custom extraction pipeline for equations.
  8. Create knowledge base of figures and tables -- cosmos/construct_caption_tables
  9. Create knowledge base of equations -- cosmos/UnicodeParser

License and Acknowledgements

All development work supported by DAPRA ASKE HR00111990013 and UW-Madison.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this repo except in compliance with the License. You may obtain a copy of the License at:

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.