p2pfl.learning.aggregators.fedopt package

FedOpt family of aggregators for P2PFL.

This package provides implementations of the FedOpt family of federated optimization algorithms. See individual modules for detailed documentation.

class p2pfl.learning.aggregators.fedopt.FedAdagrad(eta=0.1, beta_1=0.9, tau=1e-09, disable_partial_aggregation=False)[source]

Bases: FedOptBase

FedAdagrad - Adaptive Federated Optimization using Adagrad [Reddi et al., 2020].

FedAdagrad adapts the Adagrad optimizer to federated settings, maintaining adaptive learning rates on the server side based on accumulated squared gradients.

Paper: https://arxiv.org/abs/2003.00295

Parameters:

  • eta (float)

  • beta_1 (float)

  • tau (float)

  • disable_partial_aggregation (bool)

class p2pfl.learning.aggregators.fedopt.FedAdam(eta=0.1, beta_1=0.9, beta_2=0.99, tau=1e-09, disable_partial_aggregation=False)[source]

Bases: FedOptBase

FedAdam - Adaptive Federated Optimization using Adam [Reddi et al., 2020].

FedAdam adapts the Adam optimizer to federated settings, maintaining both momentum and adaptive learning rates on the server side.

Paper: https://arxiv.org/abs/2003.00295

Parameters:

  • eta (float)

  • beta_1 (float)

  • beta_2 (float)

  • tau (float)

  • disable_partial_aggregation (bool)

class p2pfl.learning.aggregators.fedopt.FedOptBase(eta=0.1, beta_1=0.9, tau=1e-09, disable_partial_aggregation=False)[source]

Bases: FedAvg

Base class for Federated Optimization (FedOpt) family [Reddi et al., 2020].

Inherits from FedAvg (which inherits from WeightAggregator) as FedOpt algorithms work with neural network weight tensors.

This class extends FedAvg to provide common functionality for adaptive federated optimization algorithms like FedAdagrad, FedAdam, and FedYogi.

Paper: https://arxiv.org/abs/2003.00295

Parameters:

  • eta (float)

  • beta_1 (float)

  • tau (float)

  • disable_partial_aggregation (bool)

SUPPORTS_PARTIAL_AGGREGATION: bool = False
addr: str
current_weights: list[ndarray]
m_t: list[ndarray]
partial_aggregation: bool
class p2pfl.learning.aggregators.fedopt.FedYogi(eta=0.01, beta_1=0.9, beta_2=0.99, tau=0.001, disable_partial_aggregation=False)[source]

Bases: FedOptBase

FedYogi - Adaptive Federated Optimization using Yogi [Reddi et al., 2020].

FedYogi adapts the Yogi optimizer to federated settings, maintaining adaptive learning rates on the server side to handle heterogeneous data distributions.

Paper: https://arxiv.org/abs/2003.00295

Parameters:

  • eta (float)

  • beta_1 (float)

  • beta_2 (float)

  • tau (float)

  • disable_partial_aggregation (bool)

Submodules