IDog

Summary

The web application “IDog” allows users to input/capture images of dogs and get a prediction of its breed. IDog then presents profiles of individual dogs that are available for adoption/rescue from a nearby shelter based on zip code. We used transfer deep learning and trained a ML model on top of Microsoft’s ResNet with 18 layers, which is trained on ImageNet. Our training images were scraped from ImageNet; 800 total images split 80%-20% into training and validation, respectively. We achieved a validation accuracy of 92% for 10-breed classification, compared to ResNet’s baseline accuracy of 95%. Our web application utilizes Flask on the back-end and Bootstrap/Jinja on the front-end. We use PostgreSQL for our database. IDog is deployed on a Digital Ocean droplet using a Ubuntu Virtual Machine. We use Nginx for SSL termination and GUnicorn as a web server, creating multiple workers for parallelization. We make calls to the Petfinder API for dog profile information, MaxMind GeoLite2 API for converting user IP to a zip code, and the Google Maps API for shelter location map integration on individual dog profile pages. The overall project was implemented in roughly 6 weeks. The app is live at www.idog.tech.

Prerequisites

Quickstart