This is a code repo for the Scientific PDF entity extractor PoC as a part of the coding challenge. I developed this repo using Windows Machine (WSL) on VSCode.
You are more than welcome to check the hugging face spaces however the development work is captured in this github repo (on development branch to be precise, but I merged it with main via PR).
You can find the deployed version of this app on https://huggingface.co/spaces/piotrkan/scipdf_extract_app. The deployed application is using the same codebase as present in this repo, with some hacks to facilitate deployment & acquisition of pretrained models from online urls. While the application is live and running and I have manually tested it without breaking, it is possible that the app will break if Hugging face server is down - therefore, if the application is down, you can let me know via github.
If you would like to use the API endpoint from the hosted app, run:
#enfothelial_dysfunction is an example pdf paper from data/ dir
api_call.py --name enfothelial_dysfunction --live
The application can be also reproduced locally by following this github, see below. Please note that the app is a prototype as it is far from perfect in terms of both performance and deployment.
To reproduce the repo locally, run the following
#clone
git clone https://github.com/piotrkan/scipdf_extractor.git
#create venv, can also create conda but venv is more universal
python -m venv .venv_name
#windows activation
.\.venv_name\Scripts\activate
#linux or mac activation
source ./.venv_name/bin/activate
#once activated, install dependencies
pip install -r requirements.txt
# if you wish to reproduce the notebook
python -m ipykernel install --user --name=venv_name
To start the gradio app locally, run:
python app.py
Once that is achieved, you should be able to access your application using the following link on your machine: http://127.0.0.1:7861
If you wish to use the app's local API, you can run:
#enfothelial_dysfunction is an example pdf paper from data/ dir
api_call.py --name enfothelial_dysfunction
This will then create the results directory where json files are being saved.
To examine the exploratory work which I used for decision making, see explore.ipynb
Initially I was planning to only focus on the ML part of the project and not work on API, as I am not a software developer. However I realized that there is no clear model, output or data specifications/requirements, and that there is no labelled data (pdfs were provided but not associated entities; of course in real-world scenario we often don't have labels but then we should have domain experts / more information about what is actually needed) - therefore I decided to attempt completing the whole challenge.
My solution is far from perfect or production-ready software and I am aware of it - but as I needed to complete the entire challenge in a relatively short timeframe, I decided to do more of a 'prototype' app rather than a final product. Some areas for improvement: both preprocessing of the data and post-processing of the output still need to be done; API and the APP should be more standalone, secure and optimized; NER system would also require fine-tuning, however as the time
ml_challenge/
├── data/ # Data directory; normally would be empty or not present in git
│ # as its bad practice to put data in git; for the purpose of this task, will be present
├── results/ # Results directory; created upon running api_call
├── flagged/ # errors log, errors flagged by app users; created upon running the gradio app
├── src/ # Source code
│ ├── models.py # nlp related utilities
│ └── data.py # data-related utilities
│
├── notebooks/
│ └── data.py # exploratory notebook with benchmarking and pdf parsing solutions
├── ANSWERS.md # answers to the questions
└── README.md # Project README