README.md

DISCO examples

This folder contains examples of how to use DISCO via Node.js scripts. Before running an example, make sure you have followed all the installation instructions described in DEV.md.

Using discojs-node for federated training

training.ts imports discojs-node (rather than the alternative discojs-web) to create two clients jointly training on the Titanic dataset in a federated manner. The example is self-contained and illustrates key uses of discojs-node:

• How to train a federated model on pre-defined tasks
• How to create multiple clients
• How to start a server instance from a script
• How to load and preprocess local data

You can run the training example as follows:

cd docs/example
npm run train # compiles TypeScript and runs training.ts

As you can see in training.ts a client is represented by a Disco object:

const disco = new Disco(task, { url, scheme: "federated" });
for await (const round of disco.fit(dataset))
  for await (const epoch of round)
    for await (const batch of epoch);
await disco.close();

To simulate more or less users change the number of calls to runUser:

// The Promise creates two clients and waits for their training to complete
await Promise.all([runUser(url), runUser(url)]);

Adding a new custom task

The server handles the pre-defined machine learning tasks and sends them to users when queried. custom_task.ts illustrates how to add a custom task to the server such that it can be queried by users. More details on custom tasks can be found in TASK.md.

You can run the custom task example with:

cd docs/examples
npm run custom_task # compiles TypeScript and runs custom_task.ts

Creating a CSV file to connect a dataset in DISCO

DISCO allows connecting data through a CSV file that maps filenames to labels. The python notebook dataset_csv_creation.ipynb shows how to create such a CSV and the scin_dataset.ipynb shows how to download the SCIN dataset and how to preprocess it in a format accepted by DISCO.