jeandut

One should be able to log average loss across clients and to provide metrics list for logging

FLamby's strategies should accept optimizer arguments

Using `Pip install` without extra arguments (and with or without -e) should work to install the full FLamby suite. This would help both integration within FL-frameworks and possible upload to...

#176 is an attempt in improving the speed of loading but this as this might have numerous side-effects, we'll hold for now.

Optimizations to be done on fed_benchmark

1

Hi, Thanks for starting the refactoring. Here are a few comments to go even further. I will not mark anything as compulsory given the time constraints, but it would be...

Factorize GPU util

1

The util to either choose a GPU or purposefully choose to disable GPU use althogether should be factorized as it is in every benchmarks.

calc_aggregated_delta_weights should be implemented in fed_avg and used there and in fed_opt.

It would be nice if one client only testing could be used with `allclose` statements to check perfect equality between federated versions and their `torch.optim counterparts`. This is done for...

The sampling bricks of LIDC use complex logic and could be tested properly on some simple examples. In addition the general flow of the code could be improved (less ambiguous...

Make benchmarks more homogeneous across datasets.