AdvancedCompressionStrategy
Enums · AiDotNet.Models.Options
Advanced gradient compression strategies for federated learning communication efficiency.
For Beginners
These strategies go beyond basic top-k or quantization to achieve 100-1000x compression while maintaining model quality. They are designed for production FL deployments with bandwidth constraints (mobile, satellite, edge devices).
Fields
| Field | Summary |
|---|---|
AdvancedCompressionStrategy Adaptive |
Adaptive compression: dynamically adjusts compression ratio per client based on estimated bandwidth, gradient importance, and staleness. |
AdvancedCompressionStrategy FedDT |
FedDT: Decision-tree-based compression for heterogeneous architectures. |
AdvancedCompressionStrategy FedKD |
FedKD: Knowledge-distillation-based compression. |
AdvancedCompressionStrategy FetchSGD |
FetchSGD: Count-sketch + top-k hybrid compression for massive models. |
AdvancedCompressionStrategy GradientSketch |
Gradient sketching: Count Sketch-based compression that maps gradients into a compact sketch using hash functions. |
AdvancedCompressionStrategy OneBitSGD |
1-bit SGD: extreme compression where each gradient component is encoded as a single bit (sign only). |
AdvancedCompressionStrategy PowerSGD |
PowerSGD: low-rank gradient approximation using SVD-like factorization. |
AdvancedCompressionStrategy SignSGD |
SignSGD: transmit only the sign of each gradient element (1-bit per parameter). |
int value__ |