Document Segmentation Assemble - DOSA

Installation and requirements

Tested for Ubuntu 18.04/20.04.

Use of a GPU significantly speeds up generation of detection outputs, but it is possible to run the inference demo code on CPU.

Python Virtual Environment

1. Set up python = 3.7.x environment: pyenv install 3.7.12 pyenv virtualenv 3.7.12 dosa-env

2. Activate the environment pyenv shell dosa-env

3. Update pip & setuptools python -m pip install --upgrade pip setuptools

Models required

5. Install requirements pip install -r requirements.txt
   • (for GPU-enabled installation: pip install -r requirements_gpu.txt)

Mask R-CNN & DocParser

6. Install Mask R-CNN pip install -e ./Mask_RCNN

7. Install DocParser

• pip install -e ./DocParser
• Download model weights follow instruction in DocParser/docparser/default_models/README.md

PaddleOCR

8. Insall PaddlePaddle pip install paddlepaddle==2.1.3
   • (for GPU-enabled installation: pip install paddlepaddle-gpu==2.1.3)

Installing paddlepaddle will raise warning error about dependency of gast==0.2.2 in tensorflow==1.15.5 vs. gast==0.4.0 in paddlepaddle==2.1.3. Just ignore it!

9. Insall PaddleOCR pip install -e ./PaddleOCR

fastAPI server

10. Install poetry following instruction https://github.com/python-poetry/poetry#osx--linux--bashonwindows-install-instructions

11. Install server dependencies poetry install

Run Server and Demo

Try each model with script in ./demos, or running API server in ./server and ./demos/server_api.py