Does openfl collaborators train sequentially?

[chirag2796]

Hi, I wanted to know if the openfl collaborators train one by one in a round before aggregation, or does it happen at the same time in parallel like a real FL setting, and the aggregator waits for the last collaborator to complete the round training?

I think it is the former, but just asking for my sanity sake. If it does happen one by one, is there any way to make it (or simulate it) parallel. I am doing some baseline experiments for a research and need to count parallel time. Any unconventional/hack way of doing it is also appreciated.

Thanks.

[psfoley]

Welcome to the community, @chirag2796, and thanks for using OpenFL!

Yes, using the Aggregator-based workflow or Director-based workflow the collaborators will all train in parallel. The only interface where collaborators train sequentially is through our python API (example here). If you are currently using that interface, it should be pretty straightforward to migrate your code (model, dataloader, etc.) to the director-based workflow.