DFedAvgMProtocol<T>
Models & Types · AiDotNet.FederatedLearning.Decentralized
Implements DFedAvgM — Decentralized FedAvg with Momentum for peer-to-peer FL.
Inheritance: FederatedLearningComponentBase<T> → DFedAvgMProtocol<T>
For Beginners
In decentralized FL, there’s no central server. Clients communicate directly with their neighbors in a network graph. DFedAvgM improves on basic decentralized averaging by adding momentum to the averaging step, which smooths out the oscillations caused by heterogeneous data and sparse communication graphs.
How It Works
Algorithm per round:
Reference: Sun, T., et al. (2023). “Decentralized Federated Averaging with Momentum.” TMLR 2023.
Constructors
| Constructor | Summary |
|---|---|
DFedAvgMProtocol(double momentum) |
Creates a new DFedAvgM protocol. |
Properties
| Property | Summary |
|---|---|
double Momentum { get; } |
Gets the momentum coefficient. |
Methods
| Method | Summary |
|---|---|
Dictionary<string, T[]> AverageWithMomentum(int clientId, Dictionary<string, T[]> currentModel, Dictionary<int, Dictionary<string, T[]>> neighborModels, Dictionary<int, double> mixingWeights) |
Performs one round of decentralized averaging with momentum for a client. |