Skip to content

A deep learning based web-app developed using the MERN stack and Tensorflow.js.

License

Notifications You must be signed in to change notification settings

theairbend3r/poke-zoo

Repository files navigation

Node.js CI GitHub license GitHub issues

PokeZoo

PokeZoo is webapp that enables users to create their own Pokemon collections by identifying Pokemon from images using deep learning.

App Flow

You can either create a new username/password or use the following -

username: ramos
password: ramos1234

Tech Stack

This webapp is hosted on Heroku. All features are added using a CI-CD pipeline through Github.

  • Backend - Node.js, Express.js
  • Frontend - React.js, Emotion.js, TailwindCSS, Twin.macro
  • Database - MongoDB
  • Authentication - JsonWebToken
  • Pokemon Data API - pokeAPI
  • Deep learning Inference - Tensorflow.js
  • Deep learning training - Docker + Tensorflow + Python

Application Features

The application features are as follows -

  1. Once a user signs-up/logs-in, they will be redirected to the home page that displays their current collection.

  2. On the home page, they can create new collections.

  3. Once there is at least one collection, Pokemon can be added to it from the explore page.

  4. The user can also edit the collection name as well as the Pokemon in it.

  5. All collection names have to be unique and duplicate Pokemon are not allowed in the same collection.

  6. Images can be used as in input for Pokemon recognition in 2 ways - webcam or upload.

  7. The output shows the 6 closest matches for a given Pokemon image.

Deep Learning

App Flow

PokeZoo has a feature to identify Pokemon from images. It is trained to identify 150 Pokemon. The model used is a pre-trained Mobilenet fine-tuned using Tensorflow (using Python and Docker) on a Pokemon images dataset. This model was then converted to Tensorflow.js which runs on the browser.

The inference happens on the browser. WebGL acceleration is used where available by Tensorflow.js.

Here are the performance-metrics for the model.

precision recall f1-score support
accuracy 0.81 1355
macro avg 0.82 0.81 0.80 1355
weighted avg 0.82 0.81 0.81 1355

Screenshots

Landing Page

Landing Page

Home Page

Home Page

Explore Page

Explore Page

Find Page

Find Page