Open Indithem opened 2 months ago
also, in our approach All clients send the updates model parameters (w) instead of gradient changes ($\Delta$ w)
We need to find weighted average of changes, and add this change to global parameters $$W{global} \leftarrow W{global} + \sum \dfrac{n_k}{n} \Delta w_k$$
ok, it seems doing as previously said, and as how we did are equivalent,
but only when all $n_k$ are 1. lets verify how same they will be in other cases
(approach till now $W{global} \leftarrow \dfrac{1}{n} \sum W{global} + \Delta w_k$)
In that paper, for fed-avg algo, they used weights for each client, corresponding to the difference in training sets for each client in each round....
Simulate this also...