MXNet & TensorFlow Pizza Image Classifier.
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
pizza_scraper
taxonomy
tf_dev
whatsonpizza_backend
whatsonpizza_classifier
.env
.gitignore
README.md
demo.py
finetune.py
imgtool.sh
requirements.txt
synset.txt
testmodels.py

README.md

What's On Pizza Project

What's On Pizza - project that gives you an ability to know pizza name by a photo. This repo contains server-side source code and similar training/validation scripts using MXNet and TensorFlow. You can compare two frameworks that solve one task. Client-side can be found here.

Task

Build an ML model that would look at a picture of a pizza and output a list of possible pizza names. We trained a model for a single-label classification, where the output just the name of the pizza. Currently only ten pizza names.

Dataset

Pizza images were collected with pizza scrapper (two python scripts for parsing pdf and html, examples of collected images can be found inside taxonomy folder). Classifier used for sorting pizza images and cleaning dataset from unrepresentative data. Trained models are used on the backend, which logic can be found here. It can be easily deployed with Docker.

The code

We used Inception-BN model as a pre-trained model and fine-tune as a technique for changing initial model wights using collected data.

After training put output models at models folder otherwise, backend will not be able to work.

Classification

For testing MXNet model without backend deployment please use testmodels.py or demo.py.

Please, write us for more details dmaiboroda@lohika.com