refactor: use alexnet to replace efficientnet

[helin0815]

Issue

Description

Proposal

Explanation

Changelog entry

Any other comments?

[jafermarq]

Hi @helin0815, I think instead of removing efficientnet what we can do is have both models and then let the user decide which one to use. For that, you can introduce a new argument to argparse (both for server and clients) and then do if/else and load the appropriate model. Does it sound good?

[helin0815]

@jafermarq good.The problem I am currently facing is that if using Efficientnet, adding a little bit of noise during differential privacy noise addition can cause loss to become nan. When I try using Alexnet, I can add noise normally and I will add this part to Readme

[helin0815]

@jafermarq hello,please review again.

[jafermarq]

Hi @helin0815,

I updated your branch with the latest main. I had to fix some merge conflicts because we recently updated the advanced-pytorch example. I also did some small modifications to how --model used.

Currently your example with --model alexnet doesn't work because the CIFAR-10 batches are of shape 224x224 but your model expects them to be 32x32... I'll get back to you about this early in the week. maybe it's a good idea to update the example to user 32x32.

In the mean time, could you update the README and show how people using the example can change the model?

[helin0815]

lgtm

[helin0815]

@jafermarq hello bro,I just deleted the differential privacy related section in readme. What should I do next

[jafermarq]

@jafermarq hello bro,I just deleted the differential privacy related section in readme. What should I do next

I pushed with some changes. The main one is replacing alexnet.py with the model that torchvision.models.AlexNet provides. The reason for this is that this example is ready for 224x224 inputs (as opposed to 32x32, which is what the model in alexnet.py expected).

I added a couple of sentences at the end of the reamde to indicate how people can switch between efficientnet/alexent. Please take a look. Something else to include?

[jafermarq]

@helin0815 , Thanks for the additions to this example. This will be useful for users that want to see how easy it is to switch between ML models for Flower clients. 💯

[helin0815]

@jafermarq Thank you for submitting the code and modifying the readme so that this PR can be merged smoothly. When you last replied to me, I was already asleep. In China, this is my sleeping time. I hope this PR can be helpful to other flower users. Thank you for your efforts