Local Epochs Inefficiency Caused by Device Heterogeneity in Federated Learning

Algorithm 2

Dynamically set number of local epochs.

Input: The clients are indexed by ; B local minibatch size, T Coomunication time window, M time prediction model.

1: Server executes:

2: initialize

3: for each round t, t =1,2,…,Ndo

4:   

5:   

6:   for each client in parallel: do

7:      ClientUpdate

8:   end for

9:   

10: end for

11: ClientUpdate

12: // Client receives the global model

13: (number of batches per epoch divided by B)

14:

15: GetTrainingTime(M, f)

16:

17: for each epoch from e to do

18:   for each batch from k to do

19:    

20:    

21:    

22:   end for

23: end for

24: return to the Server