Jensssen / Image-Classification-App

This repository contains my Rock Paper Scissor Classification App Project. It has been programmed using Kotlin and it uses AWS Sagemaker as a backend service to perform inference.

Rock-Paper-Scissor-App

This repository contains my Rock Paper Scissor Classification App Project. It has been programmed using Kotlin and it uses AWS Sagemaker as a backend service to perform inference. The classification network is base on MobileNet V2 and fine tuned with manually recorded images using Keras.

General Idea and Motivation behind the Project

The general motivation behind this project was, to learn more about different AWS Services. In addition, I wanted to try out, writing a new App using Kotlin. To this point in time, I only developed apps in native Java code (eg. SpotFoxx).

Since I am a Machine Learning Engineer and ML is awesome, I decided to build an image based classification app. However, the ML part of this project was keeped as minimal as possible, so that I have more time to focus on the AWS backend part and a little bit on the Kotlin app.

Tech stack

The following list shows most of the different programming languages, frameworks and tools that have been used for this project.

• Kotlin -> Used to programm the app
• Python -> Used inside of the Lambda Function, Sagemaker and NN training
• Keras -> Used for the NN training
• Git -> Version Control
• AWS Amplify -> Great help to develop the App backend
• AWS Services
  • IAM -> Identity management inside of AWS
  • Cognito -> App user identity management
  • S3 -> Storage of training data and NN Model
  • Lambda -> Used to access Sagemaker Endpoint
  • Sagemaker -> Used to host endpoint that performs inference

Final Result

The following Link leads to a YouTube Video that demonstrates the App.

VIDEO

Architecture

The following image show the underlying project architecture. A small remark: The architecture has not been developed in a way that it is very cost efficient. For example the kotlin app could call the Sagemaker Endpoint directly and no lambda function is needed in between. Instead, I wanted to try out many different things eg. how lambda functions work. Therefore, I implemented it like this.

Machine Learning Related Code

The ML related code can be found in an extra repository which can be found here. As already mentioned, the main focus of this project was not ML. Therefore, the ML part is keeped very minimal.