about the implementation of FedAvg

[haoyu0408]

Why does the FedAvg use a simple average without weight?

[Grassyue]

Why does the FedAvg use a simple average without weight?

Hi, I have the same question with you. Have you solve it ?

[GuoJingtao-1997]

I think maybe the num of samples (train + test) for each client are the same and therefore the weight for each client are the same. So we can directly average them. This is my viewpoint.

[haoyu0408]

Why does the FedAvg use a simple average without weight?

Hi, I have the same question with you. Have you solve it ?

For FedAvg, loss=p1L1+...+pkLk, where pi=ni/n, Li=li/ni. I think this implementation changes it as nloss=L1+...+Lk. And for this loss function, it can use this implementation. For loss=p1L1+...+pk*Lk, you need to divide ni for each loss function in local clients and then weighted sum their parameters.