Optim

FederatedOptimizer

:class:FederatedOptimizer wrapper an :class:torch.optim.Optimizer, and provide some necessary functions for federated learning. The simplest way to generate an federated optimizer is to use this wrapper like:

sgd = optim.SGD(...)
# For aggregator
fed_sgd = FederatedOptimizer(sgd, role=openfed.aggregator)
# For collaborator
fed_sgd = FederatedOptimizer(sgd, role=openfed.collaborator)

FederatedOptimizer usually has different behaviors when it plays different roles. It has two special functions, namely :func:acg_step and :func:round.

• acg_step: If you want to calculate some statistic metric of dataset with the downloaded model, you can implement here. This function will be called before the training phase.

• round: If you need to calculate some statistic metric of dataset with the trained model, you can implement here. This function will be called after the training phase.