Yolov5 object detection with FastAPI

Credit: originally a fork of OurNemanja/YOLOv5-fastapi-celery-redis-rabbitm

Application to expose Yolov5 model using FastAPI. Inferencing requests are submitted to a Celery task queue, and an asynchronous API is available for polling for results. This approach is described in detail in the excellent article Serving ML Models in Production with FastAPI and Celery by @jonathanreadshaw

• FastAPI: python framework for building APIs
• Celery: task queue with focus on real-time processing
• RabbitMQ: message broker used to route messages between API and the workers from Celery
• Redis: in-memory database to store results and process status from the tasks

Usage

Run all containers:

docker-compose up

This will start:

• rabbitmq: message broker
• redis: in-memory database on port 6379, simply used by rabbitmq
• worker: application logic (Yolo model, FastAPI and Celery) on port 8000

Perform some API requests using the integrated Swagger UI http://localhost:8000/docs

Development

• python3 -m venv venv
• source venv/bin/activate
• pip install -r requirements.txt
• isort .
• black .
• pytest . # none implemented

Overview of the code

• api/app.py: expose the endpoints and send the request task to celery.
• celery_tasks/tasks.py: receive the task and send (enqueue) to workers process.
• celery_tasks/yolo.py: code to initialize and expose a method receive a picture and return the predictions.