[BUG] in HeteroFL, the accuracy of the global model does not change about plato HOT 2 CLOSED

Why is this a bug? It appears to me that this may be due to fact that the design of the HeteroFL itself is poor.

from plato.

@wangxm233 In the implementation of HeteroFL, it contains HeteroFL algorithm according to their paper including the statistical batch normalization. This example shows that statistical batch normalization is a poor design (e.g. may not converge) when the structures of sub models are randomly sampled.

On the other hand, in the example of FedRolex, where we give an example of removing the sBN module. You can disable the sBN module in the HeteroFL by set parameter parameters.model.track to false in the configuration file and annotate function weights_aggregated and clients_processed in heterofl_server.py or refer to the implementation in FedRolex where sBN module is disabled.

Additional Context

More time spent during mode aggregation is caused by the process of doing statistical batch normalization operations.

from plato.