Benchmark
All reproduceable benchmarks are listed below.
Training progress is shown in wandb project
MLP @ MNIST
Method |
Top-1 Accuracy |
Round |
Local Epoch |
Clients |
Distribution |
Balance |
Batch Size |
|
|
|
|
|
|
|
|
CNN @ MNIST
Method |
Top-1 Accuracy |
Round |
Local Epoch |
Clients |
Distribution |
Balance |
Batch Size |
FedAsync |
93.17% |
400 |
3 |
100 |
non-iid |
balance, 2 classes |
50 |
CNN @ CIFAR10
Method |
Top-1 Accuracy |
Round |
Local Epoch |
Clients |
Distribution |
Balance |
Batch Size |
FedAsync |
42.25% |
2000 |
3 |
100 |
non-iid |
balance, 2 classes |
50 |
FedAsync |
62.61% |
10000 |
5 |
100 |
non-iid |
balance, 2 classes |
50 |
CNN @ FEMNIST
Method |
Top-1 Accuracy |
Round |
Local Epoch |
Clients |
Distribution |
Balance |
Batch Size |
|
|
|
|
|
|
|
|
RNN @ Shakespeare
Method |
Top-1 Accuracy |
Round |
Local Epoch |
Clients |
Distribution |
Balance |
Batch Size |
|
|
|
|
|
|
|
|