flower Server and client on flower framework

What is your question?

In the first round, can we make the server wait for all clients to connect it "to get initial model parameters" instead of gets initial model parameters from one of the connected clients.

Sep 29 '22 05:09 AlesaRLife

Thanks for reaching out @AlesaRLife.

Could you give a bit more background, what would be the use case for this?

The server cannot wait for all clients since the server does not know how many clients there are in total. What we could do though is to add a parameter that tells the server to wait for at least n clients. Before we do so, it'd be good to understand the motivation for this feature request better.

Sep 29 '22 11:09 danieljanes

thank you for replying @danieljanes I want this, so that I can collect the parameters for example weights from all the clients and see if there are common factors between them before go to the server in the first round.

Sep 29 '22 12:09 AlesaRLife

@AlesaRLife could you perhaps do this using a custom strategy? Make the strategy wait for all the clients, collect what you need using get_properties, then start the actual training?

Oct 17 '22 21:10 danieljanes