sport_classification_service

Problem statement

We aim to address the challenge of accurate and efficient sport image classification within our service. The goal is to enhance the precision of categorizing sports based on user-provided images by leveraging a Convolutional Neural Network (CNN) model trained on a diverse sport-specific dataset.

Kaggle data set used - https://www.kaggle.com/datasets/sidharkal/sports-image-classification

Solution

As a part of this application, a pre-trained model - Inceptionv3 was further trained with a dataset containing images of different sports classes- cricket, wrestling, tennis, badminton, soccer, swimming, and karate and wrapped as tf-serving(based on the above kaggle dataset). The trained model has an accuracy of 94% in classifying the sport. Please note the training dataset contained only 7 sport classes [cricket, wrestling, tennis, badminton, soccer, swimming, and karate]

Solution Architecture

Note

Images, model are checked into the repo making it a large repo interms of size. Hence git-lfs was used to store file greater than 100MB. Follow the installation instructions for git-lfs here.

Project structure

Folder	Description
dataset	Directory containing training and test data - image files.
gateway	Directory containing flask application which acts as a gateway for tf-serving
model	Directory containing '.h5' model which was trianed and tensorflow generated model which can be used in tf-serving.
trainer	Directory containing notebooks and script which were used for training the model.
kube-config	Configuration related to kubernetes deployments.

Pre-requisties

• python 3.10 or above.
• docker, docker-compose, Kind, kubectl
• pip3, pipenv
• git-lfs

Installing dependencies

Use pipenv install to install dependencies from respective directories, Only if you want to train model and build images yourself.

• Folder gateway - contains dependencies related to flask and tf-serving.
• Folder trainer - contains dependencies related to tensorflow and other libraries required for training the model.

How to run locally with Docker compose

1. Clone this repo
2. Install git-lfs.
3. git lfs pull.(The trained model is saved in the model directory already)
4. docker-compose up
5. check if the containers is up docker ps

How to run on local kubernetes

1. Ensure kind kubernetes in installed
2. run ./deploy-local-kube.sh <cluster-name>
3. kubectl port-forward services/sports-gateway-service 9696:9696

Testing with python script

1. cd into ./gateway directory
2. pipenv install
3. pipenv shell
4. python3 predict_test.py

Testing with curl

run ./predict_test.sh

or

  curl --request POST \
  --url http://localhost:9696/predict \
  --header 'Content-Type: application/json' \
  --data '{"url": <<image url>>}'

<<image url>> - any image url from internet

Sample Images for testing

http://bit.ly/3RtIOTq
https://bit.ly/3GMDW6T
https://bit.ly/48nRO3l

AWS Cloud Deployment

Both the Gateway and the Model is deployed on AWS EKS cluster.

The gateway is available for testing hostname:a1192352e2f8a43ca8923f350baf8983-1337907194.us-east-2.elb.amazonaws.com

you can test the application deployed on AWS Cluster running

./predict_test.sh a1192352e2f8a43ca8923f350baf8983-1337907194.us-east-2.elb.amazonaws.com <<image_url>>

gateway Service API

API	Description	Response	Response-type
/ping	ping api to check the status	pong	String
/predict	Return the prediction of the classification	check below sample	Json

Sample Response for '/predict'

{
  "prediction": "Karate",
  "probability_percentage": {
    "Badminton": 1.887187517368716e-09,
    "Cricket": 8.943498869484756e-06,
    "Karate": 0.9999909400939941,
    "Soccer": 1.0174683495733916e-07,
    "Swimming": 3.659216159246359e-12,
    "Tennis": 2.4010535781826547e-09,
    "Wrestling": 1.1995511250972868e-09
  }
}

Training the model from scratch

The trained model is already checked into the repos under model directory along with the notebooks used for training and testing. Incase you want to do from scratch follow below steps

1. cd into trainer directory
2. pipenv install
3. pipenv shell
4. python3 ./train_model.py - Running this code saves the .h5 model with highest accuracy.
5. python3 ./convert_model.py - converts to tensor saved model format.
6. Note: verify h5_model_name & tensor_model_name in convert_model.py before running.