On hyperparameters for Table 3.

[honggudrnjs]

Thanks for sharing the code of this valuable work.
Following the results of your work, I wonder hyperparameters for reproduce of Table 3.
Currently, the hyperparameter setting in 'args.py' does not fit to Table , because E is set to 1 and server_valid_ratio is different from what is explained in the paper.
I followed the details in the paper like the following:
CIFAR-10, fmnist vs CIFAR-100
LR 0.08 vs 0.01
wd 5e-4 vs 5e-5
server_valid_ratio 0.01 vs 0.1 (10*(# of classes) / (# of test data))

However, we found that running FedLAW with CNN (w/ Dirichlet alpha 0.1, CIFAR-100) produces random prediction.
Thus, we wonder if we missed some details. Is client_valid_ratio 0.3 right? or server_valid_ratio or server_epochs should be changed?

[ZexiLee]

Thanks for this raised issue.

I guess "produces random prediction" means that FedLAW works well in the early but suddenly crashes into the random prediction model.
I met these situations before and not only happened to FedLAW but also FedAvg and other methods. I think this is due to severe model drifts.
I suggest you slightly tune server_valid_ratio, server_epochs, server_lr, or other hyperparameters for FedLAW. Sometimes, changing another random seed is just fine. Also, reducing the local epochs and learning rates may also help.

Hoping my reply can help you.

[honggudrnjs]

Thanks for your reply. Could you check dirichlet alpha one more time? I guess dirichlet alpha for non-IID is actually 1.0, not 0.1 as in Table 3. In CNN and CIFAR-100 setting, FedLAW is not working for alpha=0.1, but is working for alpha=1.0 with similar performance in Table 3.