about results in the experiment

[13015517713]

Hey, Youngjoon. Nice code!
Did you reproduct the similar result with that paper?

[yjlee22]

Obtained similar values ​​to the results written in the paper.

[13015517713]

It's the result about cifar dataset in paper. Although iid, I can't reproduce this result, more than 80% acc. I just trained the model locally, but I found it's about 70% in global testset.
The cifar model (model cnn_v2 in FedShare), is the same as another project "https://github.com/AshwinRJ/Federated-Learning-PyTorch" which implements the fedavg . I also found this problem occurring in that project. Refer "AshwinRJ/Federated-Learning-PyTorch#31".
So, did you reproduce the result about cifar using model cnn_v2? I'll appreciate it if you can share your parameters.

[yjlee22]

The currently written code like the repository you mentioned cannot achieve the same performance as the paper due to the limitations of the model for the CIFAR-10 dataset. If you increase the number of parameters of cnn_v2 model, you can achieve the performance similar to or higher than that of the paper. The following parameter settings are recommended as follows:
"""
self.conv1 = nn.Conv2d(args.num_channels, 64, kernel_size = 5)
self.conv2 = nn.Conv2d(64, 64, kernel_size = 5)
self.fc1 = nn.Linear(1600, 384)
self.fc2 = nn.Linear(384, 192)
self.fc3 = nn.Linear(192, args.num_classes)
"""

[13015517713]

Ok. Thanks for your reply.