/comment-filter-finest-bert-engee

Comment Filtering model trained on English and Estonian data using EMBEDDIA/finest-bert

Primary LanguagePythonMIT LicenseMIT

Dockerized REST API for news comment filtering

This repository is based on flask-rest-docker which provides the basic dockerized Flask REST API skeleton.

Requirements

  • docker
  • docker-compose

How to use

The code is ready to be tested without additional configuration. However, it is strongly recommended that you modify .env.prod and change Postgres and Flower user and password data before moving the code into production.

The functions that define the API reside in services/web/project/__init__.py. They call functions services/web/project/api_functions.py to do the computation. Asynchronous functions which are handled by Celery reside in services/celery-queue/tasks.py.

The skeleton provides few examples of synchronous and asynchronous functions which can serve as templates for implementing your own. In short, you need to:

  1. Define the API call in services/web/project/__init__.py by writing a Resource class with the appropriate methods (e.g., get and post), decorate the class and the methods to get automated documentation and testing.
  2. Implement the actual function in services/web/project/hate_speech_classifier.py. If the function is asynchronous, write also a handler in services/celery-queue/tasks.py. Results of asynchronous functions are stored in Redis which is configured to be persistent.

Development

The following command

$ docker-compose up -d --build

will build the images and run the containers. If you go to http://localhost:5000 you will see a web interface where you can check and test your REST API. Flower monitor for Celery is running on http://localhost:5555. Note that the web folder is mounted into the container and the Flask development server reloads your code automatically so any changes will be visible immediately.

Production

The following command

$ docker-compose -f docker-compose.prod.yml up -d --build

will build the images and run the containers. The web interface is now available at http://localhost and the Flower monitor at http://localhost:5555. If you change the source code, you will have to do a rebuild for changes to take effect.

Pre-trained classifier models

The classifiers require pre-trained models. On running the containers, a script will check for their existence and download them if missing. If you have trouble with this, try running this process manually:

sh ./services/web/project/models/model_download.sh