Struct CompressionManager 

pub struct CompressionManager<T: Float + Debug + ScalarOperand + Send + Sync + 'static> { /* private fields */ }

Compression manager

Implementations

impl<T: Float + Debug + ScalarOperand + Send + Sync + 'static> CompressionManager<T>

pub fn new() -> Result<Self>

A quantizing compression manager.

There is no configurable target ratio: Self::compress always quantizes to 8-bit codes, so the achieved ratio is fixed by size_of::<T>() and reported per payload by CompressedData::compression_ratio. The compression_ratio this used to accept was stored and never consulted by the quantizer, which made it a knob that silently did nothing.

pub fn compress(&self, tensor: &Array2<T>) -> Result<CompressedData<T>>

Compress a tensor by affine 8-bit quantization.

Every element is mapped to a u8 code by code = round((x - min) / step) with step = (max - min) / 255, and the payload is those codes plus the two f64 parameters min and step. compressed_size is therefore the real byte count of what is stored: codes.len() + 2·size_of::<f64>() + shape metadata.

This replaces a version that stored the input verbatim as Vec<T> and reported compressed_size = bytes / 2 — a fabricated 50% ratio for a payload that had not shrunk at all.

Quantization is lossy; the reconstruction error is bounded by step / 2 per element, which CompressedData::max_reconstruction_error reports. A constant tensor (max == min) has step = 0 and round-trips exactly.

Errors

Returns Err when the tensor is empty or contains a non-finite value (there is no finite quantization range for NaN/inf).

pub fn decompress(&self, compressed: &CompressedData<T>) -> Result<Array2<T>>

Reconstruct a quantized tensor: x̂ = min + code · step.

Errors

Returns Err when the stored shape is not two-dimensional or does not match the number of stored codes.

pub fn store( &mut self, key: String, compressed: CompressedData<T>, ) -> Result<()>

pub fn retrieve(&self, key: &str) -> Result<Option<CompressedData<T>>>

pub fn remove(&mut self, key: &str) -> Result<bool>

pub fn clear(&mut self) -> Result<()>

pub fn get_memory_usage(&self) -> usize

pub fn optimize_compression_ratios( &mut self, _patterns: &AccessPatterns, ) -> Result<()>