We created PredicTable, a web application for restaurants to predict future sales, to help reduce overproduction and create a menu plan tailored for the customers and with this help make the world a bit more sustainable. PredicTable uses DISCO, to run federated learning without sharing any private data from the restaurants, so that they can keep their privacy.
This project and MVP showcase that Disco can also work outside of a research setting and provide real value in the real world.
❓ WHY DISCO?
- To build deep learning models across private datasets without compromising data privacy, ownership, sovereignty, or model performance
- To create an easy-to-use platform that allows non-specialists to participate in collaborative learning
We provide a run_server.sh and a run_client.sh to help launch the project.
Firstly run run_server.sh and then run run_client.sh when you want to create a new client. Note that each time you need to open a new terminal window to run a bash script.
The detailed commands for running are as below:
cd disco
nvm use
npm ci
cd discojs/
npm ci
./build.sh
cd ../server/
npm ci
npm link ../discojs/discojs-node
npm run dev
cd ../exa
npm ci --legacy-peer-deps
npm link ../discojs/discojs-node
npm run dev
Please ensure to have nvm and npm installed. It can happen that you get errors and need to install additional packages.
In the folder exa you can find the implementation of our project. For every resaurant, a Disco client is created, which can communicate with the running server. To train the model, visit the website and use Train Model button or use url path /train. Select the dummy CSV from this repository and let the magic happen. Currently it's set to support federated learning.
We took some part of the web-client to make this project possible. The training is working and is tested with multiple inputs for a simple linear regression model. The frontend is MVP to show what can be possible with our project and how we can use Disco for non-traditional and real-life related tasks. We used Vue and TypeScript to have maximal compatibility with Disco. Also the model is here a MVP with dummy data, to prove that Disco could work in this setting and is capable of solving tasks outside of the scientific environment and provide value to economy and sustainability.