Struct GradientCompressor 

pub struct GradientCompressor { /* private fields */ }

Gradient compressor using random projection.

Uses Johnson-Lindenstrauss style random projection for communication-efficient distributed training.

Implementations

impl GradientCompressor

pub fn new(config: TrainingConfig) -> GradientCompressor

Create a new gradient compressor.

pub fn compress( &self, gradients: &[f32], ratio: Option<f32>, ) -> Result<CompressedGradient, TrainingError>

Compress gradients using random projection.

Arguments

• gradients - Original gradient vector
• ratio - Compression ratio (0 < ratio <= 1), None uses default

Returns

Compressed gradient representation.

pub fn decompress( &self, compressed: &CompressedGradient, ) -> Result<Vec<f32>, TrainingError>

Decompress gradients.

Arguments

• compressed - Compressed gradient

Returns

Reconstructed gradient vector (approximate).

pub fn compress_ternary( &self, gradients: &[f32], ratio: Option<f32>, ) -> Result<TernaryCompressedGradient, TrainingError>

Compress and immediately quantize to ternary for maximum compression.

This achieves ~300x compression: 10x from projection + ~30x from ternary.

pub fn decompress_ternary( &self, compressed: &TernaryCompressedGradient, ) -> Result<Vec<f32>, TrainingError>

Decompress ternary compressed gradients.

Trait Implementations

impl Clone for GradientCompressor

fn clone(&self) -> GradientCompressor

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for GradientCompressor

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.