node_example

Example

In example.ts we give a brief example of discojs, in it we run two clients training jointly via federated learning. It trains on a few examples of the face task, the samples are already stored in the repo and so it is not necessary to download any additional data.

The example is self-contained, it runs a Disco server that is usually run externally.

To run the example do

nvm use
./get_training_data
cd discojs && npm ci
cd discojs-node && npm run build
cd ../..
cd server && npm ci
cd ..
cd example && npm ci && npm start

Disco.js API

In example.ts (line 100) the following function gives an example of the Disco.js API.

To run the clients with decentralized training or local training, set the training scheme to: TrainingSchemes.DECENTRALIZED or TrainingSchemes.LOCAL, respectively.

async function runUser (url: URL): Promise<void> {
  // Load the data, the dataset must be of type data.Data, see discojs import above.
  const data = await loadData(TASK)

  // Start training
  const disco = new Disco(TASK, { url })
  await disco.fit(data)

  // Stop training and cleanly disconnect from the remote server
  await disco.close()
}

To simulate more or less users, simply add more in line 70:

// Add more users to the list to simulate more clients
await Promise.all([
  runUser(url),
  runUser(url)
])