yxdyc / pfedgate Goto Github PK

In Algorithm 1, the authors generate the corresponding gating weights for each batch (line 5). My question is that if the gating weights are different between batches, then the blocks selected during the masking phase may be different. So how to get the final uploaded mask? Is it the intersection or union of the masks of each batch? If is a union, the worst case is uploading the full model, which can not improve the communication efficiency.

Hello! We are interested in your work and thank you for opening the source code. We want to compare your method with ours in our paper. But we ran into some problems running the code. We conduct experiments on cifar-100 under dirichlet non-IID with alpha=0.5. However, the accuracy is very low and almost does not change during training. We also test on pathological noniid with 40 clients. The accuracy has been around 50% (some clients 50%, some clients 0%). We think this is due to the hyperparameter settings. Could you please give us some suggestions? Note that we adopt the hyperparameters specified in your paper. But there are still some hyperparameters we do not know how to set and we just adopt the default value in the code. We would really appreciate your help!