Yolov4-kubernetes

Scaling yolov4 object detection using kubernetes

Getting started

• Download the pretrained yolov4.weights from the official repository

• Build the Docker image

  docker build -t image_name:version .

• Publish the docker image to dockerhub

  docker push image_name:version

• Modify deployments/yolo.yaml with the apropriate container image

• On a kubernetes cluster apply the yolov4-fastapi deployment along with the expose-service to expose the node-port

  kubectl apply -f deployments/yolo.yaml -f deployments/expose-service.yaml

Testing

• Check the service and deployment status using kubectl commands

  kubectl get pods and kubectl get services

• Try acessing the endpoint using any one of the ips of the nodes on port 31117 as mentioned in the expose-service.yaml file

• You can curl the api to get the response

  curl -X 'POST' \
   'http://ip-of-the-node:31117/files/' \
   -H 'accept: application/json' \
   -H 'Content-Type: multipart/form-data' \
   -F 'file=@/path/to/img.jpg;type=image/jpeg'
  

• A sample response would look something like this

  {"result":"{'person': 0.8518975, 'frisbee': 0.23659469, 'tennis racket': 0.24101968}"}

Adding a nginx load balancer

• On a different machine install nginx

  sudo apt install nginx

• Modify loadbalancer/loadbalancer.conf with the appropriate node ip addresses

• Copy loadbalancer/loadbalancer.conf to /etc/nginx/conf.d/loadbalancer.conf

• Restart nginx service

  sudo systemctl restart nginx

• Curl the api using the command mentioned earlier by using the ip of the load balancer