AdaptiveCompressor<T>
Models & Types ยท AiDotNet.FederatedLearning.Compression
Adaptive compressor: dynamically adjusts compression ratio per client based on bandwidth, gradient importance, and staleness.
Inheritance: FederatedLearningComponentBase<T> โ AdaptiveCompressor<T>
For Beginners
In a real federation, clients have different network speeds. A phone on 5G can send more data than one on 3G. Rather than using the same compression for everyone, the adaptive compressor gives faster clients less compression (better accuracy) and slower clients more compression (saves bandwidth). It also prioritizes clients whose gradients carry more information.
How It Works
How it works:
Constructors
| Constructor | Summary |
|---|---|
AdaptiveCompressor(AdvancedCompressionOptions options) |
Initializes a new instance of AdaptiveCompressor. |
Methods
| Method | Summary |
|---|---|
ValueTuple<Tensor<T>, double> Compress(Tensor<T> gradient, int clientId, int currentRound, double roundTripMs) |
Compresses a gradient using an adaptively determined top-k ratio. |
double ComputeAdaptiveRatio(int clientId, Tensor<T> gradient, int currentRound) |
Gets the current adaptive compression ratio for a client. |
void RecordBandwidth(int clientId, double roundTripMs) |
Records a bandwidth measurement for a client. |