Snaq on Eschernode

Presentation by Anil Karaka at the Open Food Data Hackdays 2018

Transcript

1. Snaq on Eschernode.
2. Introduction • Anil Karaka. • Winner of Crowdai's Learning How to Walk challenge. • Founder of eschernode.com Business Innovation food.opendata | 2018
3. Problem Statement. • Estimate food portion size, nutritional value, contents automatically. Business Innovation food.opendata | 2018 Use Cases • People will know the nutritional composition instantly. • Identify the dish, especially helpful when people are travelling, so that the can know what they are eating. eg. Simply take a picture at a buffet and know what it is. • Helpful to people following a diet.
4. First steps. • The data set we were provided is a 700 pictures each of three different classes. Identify these classes automatically. • Available Classes. 1. Bowl plate. 2. Regular plate. 3. Soup plate. Business Innovation food.opendata | 2018
5. Training using Eschernode. • Trained a deep learning classifier on Resnet18 architecture. Business Innovation food.opendata | 2018
6. Infrastructure. • Trained on a 60GB machine for 5 hours. Business Innovation food.opendata | 2018
7. Results. • Best Cross entropy error of about 0.863 after training for 42 epochs Business Innovation food.opendata | 2018