mahi97 / evofl

[NIPS'23] EvoFed: Leveraging Evolutionary Strategies for Efficient and Privacy-Preserving Federated Learning

evofl's Issues

Description

Could you share the description of the programming environment? It is important for us to reproduce the results.

What's the dimension of the average pooling?

Hi,

i have a question about the right dimensions stated in the paper according to the current code available here.

Given these networks in the supplementary material:

The number of parameters is neither 11k and not 2.3 milions as stated in the paper for the networks if we use the average pooling in your code.
The number of parameters should be:

8 X 5 X 5+8+8 X 16 X 5 X 5+16+weights and biases of last fully conencted layer.
Now, in the table's paper is stated that the fully connected is 16 X 10 but is not true because they are not the input neurons instead they are the input channels. Now, since the input images are 28 by 28 and you kernel rows and columns are 5 by 5 and from your code you used an average pooling with size 2 by 2 with stride 2 we have for both dimensions: ((((28 - 5)/1 + 1 - 2)/2 + 1 - 5)/1 + 1 - 2/2 + 1) = 4.
So the output from the last average pooling should be 4 X 4 X 16 so the weights of the last fully connected layers should be 16 X 16 X 10 so the final number of parameters should be:

8 X 5 X 5+8+8 X 16 X 5 X 5+16+16 X 16 X 10+10 = 5994 parameters not 11k.

Unless i'm missing something.

and

same for the cifar-10 dataset.

with the only difference that the input images are 3 X 32 X 32 so being 32 by 32 with the same previous computations (considering the average pooling dimensions and stride you have set in the code) the first fully connected layer after the average pooling should have 5 X 5 X 128 X 256 + 256 parameters and with all the right computations it pops out the the actual number of parameters is more or less 1 M not 2.3M

Is possible to provide the actual average pooling dimensions and stride? i'm trying to reproduce on my own the paper and compare with another algorithm but i need the same network architecture if possible. Thanks in advance