DeTAGProtocol<T>
Models & Types · AiDotNet.FederatedLearning.Decentralized
Implements DeTAG — Decentralized gradient Tracking for exact convergence.
Inheritance: FederatedLearningComponentBase<T> → DeTAGProtocol<T>
For Beginners
Basic decentralized averaging has a problem: it converges to a “consensus” point that may not be the true global minimum because each client only sees their local gradient. DeTAG (Decentralized gradient Tracking) fixes this by having each client track the difference between their local gradient and the global gradient estimate. This correction term ensures exact convergence even with heterogeneous data.
How It Works
Gradient tracking update:
Reference: Li, H., et al. (2023). “DeTAG: Decentralized Tracking-based Asynchronous Gradient methods.” 2023.
Constructors
| Constructor | Summary |
|---|---|
DeTAGProtocol(double learningRate) |
Creates a new DeTAG protocol. |
Properties
| Property | Summary |
|---|---|
double LearningRate { get; } |
Gets the learning rate. |
Methods
| Method | Summary |
|---|---|
T[] GetTracker(int clientId) |
Gets the current gradient tracker for a client (for sharing with neighbors). |
T[] Step(int clientId, T[] currentParams, T[] newGradient, T[] previousGradient, Dictionary<int, T[]> neighborParams, Dictionary<int, T[]> neighborTrackers, Dictionary<int, double> mixingWeights) |
Performs one DeTAG update step for a client, including neighbor-averaged gradient tracking. |