search

wenh06 / fl-sim

A Simple Simulation Framework for Federated Learning Based on PyTorch
MIT License
5 stars 0 forks source link

For non-iid secenario? #5

Open CaesarGo opened 2 months ago

CaesarGo commented 2 months ago

Thanks for the very helpful repo. I would like to know if there any interface for non-iid data setting for each client?

I don't find it in the src code.

Thanks.

wenh06 commented 2 months ago

non-iid of the clients comes from the data. Most of the federated datasets are distributed in some (fixed) non-iid manner. Because they are inherited from previous research works that already fixed the distributions (perhaps for the ability to reproduce their results). In a few of the federated datasets, one can control how they are distributed across the clients (in such a non-iid manner that one wants). Ideally one should be able to distribute the data across the clients as one wants but this is not implemented yet.