Handling ML Predictions in a Flask App

created by Austin Poor

This repo demonstrates a technique for handling long-running processes (like ML model predictions) inside a Flask app, by passing them off to Celery.

While this is just a simple example, the same workflow could help in cases where you're performing slow, complicated tasks like making ML predictions, building recommendations, or updating a database.

Demo

The app is composed of Redis (the broker and backend for Celery), a Celery worker, and then the Flask app.

The Flask app is composed of a simple HTML landing page and two API endpoints.

The HTML page has 4 range sliders, allowing you to choose values for the width and height of the iris's sepal and petal, for prediction.

When you click the Submit button, a javascript function sends the parameters in a POST request to /api/predict, which schedules the prediction with the Celery worker, and returns the taskid that will store the result.

Then, after making the request, the browser polls the second API endpoint (/api/get-result/<taskid>) until the status changes from pending to SUCCESS (or to an error). It then updates the status and the prediction table.

Instructions to Run

Requirements

• Python 3.6+
• scikit-learn
• Docker & docker-compose

1. Download the repo
2. Train the ml model with the train.py script
3. Build and run the containers with docker-compose up --build
4. Then go to localhost:5000