SegmentedGossipProtocol<T>
Models & Types · AiDotNet.FederatedLearning.Decentralized
Implements Segmented Gossip — communication-efficient gossip that exchanges model segments.
Inheritance: FederatedLearningComponentBase<T> → SegmentedGossipProtocol<T>
For Beginners
In standard gossip protocols, each pair of communicating nodes exchanges the entire model. For large models, this is very expensive. Segmented gossip splits the model into segments and only exchanges one segment per communication round. Over multiple rounds, all segments get exchanged, achieving the same convergence but with much less per-round communication.
How It Works
Algorithm:
Reference: Bellet, A., et al. (2024). “Segmented Gossip for Communication-Efficient Decentralized Learning.” 2024.
Constructors
| Constructor | Summary |
|---|---|
SegmentedGossipProtocol(int numSegments) |
Creates a new Segmented Gossip protocol. |
Properties
| Property | Summary |
|---|---|
double CompressionRatio { get; } |
Gets the communication compression ratio. |
int CurrentRound { get; } |
Gets the current round number. |
int NumSegments { get; } |
Gets the number of segments. |
Methods
| Method | Summary |
|---|---|
Dictionary<string, T[]> GossipExchange(Dictionary<string, T[]> myModel, Dictionary<string, T[]> peerModel) |
Performs one round of segmented gossip between two peers. |