WEB-API for Pytorch NER model

This project seeks to facilitate the exchange and diffusion of NER models built with different architectures. We strongly used the code from NCRF++ which allow to build various NER neural models in pytorch.

The tag_serve repo adapts the ner model from NCRF++ and wrap it in a flask API to allow live demo via a web page and a deployment for medium scale production.

Installing from source

You can install tag_serve from source by cloning the git repository:

git clone https://github.com/strayMat/tag_serve.git

Create a Python 3.6 virtual environment, and install the necessary requirements by running:

./src/install_requirements.sh

Call the API

Launch API and web demo locally

• Launch the API: python app.py
• Open in the client in browser: firefox client/pred_client.html

Launch API and send multiple files

• Launch the API: python app.py
• Launch call function: python client/call.py -i decoding/ins/ -o decoding/outs/ (add -v to get visualization .html: python client/call.py -i decoding/ins/ -o decoding/outs/ -v)

With a curl command

In your terminal, run :

curl -H 'Content-type:application/json' -d '{"file":"Paris is wonderful!"}' localhost:5000/predict

Change the trained model used by the API

You can either give a specific model to the api, when launching the python code app.py or replace the default model in the pretrained directory.

• Specify a model to app.py: Launch the api with the -m option and specify your new_model name, python app.py -m myModel/new_model where the folder myModel should contain new_model.xpt and new_model.model (the architecture and the weights of the model).

• Replace the baseline model: Replace directly the baseline files in the pretrained directory: put new baseline.xpt and baseline.model in the pretrained/ folder (you can check that the default model of the app is pretrained/baseline by typing python app.py --help)

Use docker

You can deploy the model with docker. Go on docker website to install docker and docker-compose. Then build the docker with:

sudo docker build --build-arg http_proxy=$yourProxy -t yourTag .

Run the docker with :

sudo docker run -d -p 5000:5000 --name tagger yourTag python3 /app/app.py

You can now access the docker API with the call methods of your choice (client/call.py, client/predict_client.html or curl)

Train your own model

Go see the demonstration notebook: train_decode_template.ipynb

References:

• NCRF++ : An Open-source Neural Sequence Labeling Toolkit, Yang et Zhang, 2018

• Many thanks to Guillaume Genthial for the excellent blog post on python web-api with flask.