Offline handwritten digit recognition

This small web app aims to demonstrate how machine learning models can be used completely offline.

While the web resources (HTML, CSS and JS) are cached using a service worker, the model and its weights are saved into IndexedDB using TensorFlow.js. They are loaded from there wherever possible, while falling back to the network if not found (for example, when the user clears the site data).

This way, after the first successful load on a modern browser, all resources are available even without a network connection.

The model has been trained using the mnist-node example code in tfjs-examples repository.

Running the code

You need to have Node.js installed. In the project root directory, run these commands:

npm install
npm start

Creating a minified build

Assuming you have run npm install previously:

npm run build