Tensorflow Federated in a hierarchical fashion

121 views Asked by At

I'm trying to use tensorflow federated to simulate a federated learning algorithm in a hierarchical topology. For what I've read, only a few strategies are implemented (such as tff.learning.build_federated_averaging_process), which relies on a single aggregation server. Is there any way to have networks that are more complex than the standard n-clients 1-server?

1

There are 1 answers

0
Keith Rush On

TFF has native support for intermediate aggregation; a binary, for example, which runs code analogous to that implemented for testing here and serves analogously to the remote executor service could connect to workers and aggregate in a hierarchical manner.

You may, however, mean something more general than this.