Table Structure Recognition

Running docker container

Preparing the environment

To run docker first install sudo apt install nvidia-container-runtime

Make sure you have nvidia-docker2 installed too!

Then edit/create /etc/docker/daemon.json with content:

{
    "runtimes": {
        "nvidia": {
            "path": "/usr/bin/nvidia-container-runtime",
            "runtimeArgs": []
         } 
    },
    "default-runtime": "nvidia" 
}

And restart docker daemon with sudo systemctl restart docker

All of this allows docker to use GPU during build time, as it is needed to compile mmdetection library.

Alternatively, you may run all the commands from Dockerfile by yourself, after running a container, because the GPU will be available then.

Building and running

Firstly, download the model from here and save it at the path server/model/models/model.pth

Server will use port 8444, you can specify another one in the docker-compose.yml

Run docker-compose up to start all the containers

Troubleshooting

Sometimes EasyOCR may have a problem with downloading a model, you may see this in logs. To fix it you simply need to restart docker daemon with sudo systemctl restart docker

Running client

To install a package simply run pip install . in the main directory, this will install tsr package, which provides all the functionalities you need.

1. Use FileReader class to load document.
2. Read data to DataInstance class using FileReader read method.
3. Construct model using desirable parameters:
   • method for recognizing split tables: "model" or "heuristic"
   • OCR method: "tesseract" or "easyocr"
   • language for OCR: "pl", "en", "uk", "ru"
   • threshold - to determine the level of certainty - float between 0.0 and 1.0
4. Set address for your Docker server (and chosen port)
5. Use Service predict method to get the result.

Running scripts

We have created some useful scripts for you to try out! You can check how to use them via python <script>.py -h