FedStar

[Raj-Parekh24]

Issue

#2042

Description

Implementation of FedStar in flower.

Related issues/PRs

Proposal

Explanation

Checklist

• Implement proposed change
• Write tests
• Update documentation
• Update changelog
• Make CI checks pass
• Ping maintainers on Slack (channel #contributions)

Any other comments?

[Raj-Parekh24]

Hi @jafermarq ,
I have replicated all codes and suggestions in this PR and also checked the working of the code. Please review the code.

Thanks for the support!

[jafermarq]

Hi @Raj-Parekh24,

I have now completed a first pass through your code. I have highlighted a few formatting issues in the README.md. In general the main points are:

• It's better not to rely on tensorflow_addons since TF warns to not use.
• I ran your code for 100 rounds with the default config (i.e. ambient_context dataset) but I only saw ~40% accuracy on the server side.
• I'm not entirely sure if the logic around your GPU selection for clients is the best. On my system (with just one big GPU) only the server made use of it even though less than 5% of the VRAM was used. Wouldn't it be better to use Flower's simulation engine (i.e. use start_simulation as oppose to start_server+start_client?

Ping me again over here or Slack once you had a chance to look into the suggestions mentioned above and also those highlighted below. It would be good to start seeing some results on the README.md (even if still aren't fully reproducing those in the FedStar paper).

[jafermarq]

Hi @Raj-Parekh24 , just a couple of points based on our last sync on slack.

[jafermarq]

is it correct reporting only the number of labelled examples?

[Raj-Parekh24]

I thought of putting on labeled samples because our semi-supervised architecture is run on the client side and only the unlabelled samples are available there. The server only knows about labeled samples.

[jafermarq]

Amazing job @Raj-Parekh24 reproducing FedStar, a FL method for semi-supervised audio recognition! It's great including a new TensorFlow baseline too.

[Raj-Parekh24]

Thanks a lot @jafermarq,

For your guidance and support throughout the project!