pub struct CompressionMetrics {
pub n_parameters: u64,
pub original_bits_per_param: f32,
pub effective_bits_per_param: f32,
pub sparsity: f32,
pub quantization_mse: f32,
}Expand description
Compression statistics for a single layer or entire model.
Fields§
§n_parameters: u64Number of parameters.
original_bits_per_param: f32Bit-width of the original (full-precision) representation.
effective_bits_per_param: f32Effective bit-width used in the compressed representation.
For quantized weights this equals quant_bits; for pruned weights
the effective bits = quant_bits × (1 − sparsity).
sparsity: f32Fraction of weights that are pruned ∈ [0, 1].
quantization_mse: f32Mean squared quantization error (0 if no quantization was applied).
Implementations§
Source§impl CompressionMetrics
impl CompressionMetrics
Sourcepub fn quantized_only(
n_parameters: u64,
quant_bits: u32,
quant_mse: f32,
) -> Self
pub fn quantized_only( n_parameters: u64, quant_bits: u32, quant_mse: f32, ) -> Self
Create metrics for a layer quantized to quant_bits bits with no pruning.
Sourcepub fn pruned_only(n_parameters: u64, sparsity: f32) -> Self
pub fn pruned_only(n_parameters: u64, sparsity: f32) -> Self
Create metrics for a layer pruned to sparsity with FP32 weights.
Sourcepub fn quantized_and_pruned(
n_parameters: u64,
quant_bits: u32,
sparsity: f32,
quant_mse: f32,
) -> Self
pub fn quantized_and_pruned( n_parameters: u64, quant_bits: u32, sparsity: f32, quant_mse: f32, ) -> Self
Create metrics for a layer that is both quantized and pruned.
Sourcepub fn compression_ratio(&self) -> f32
pub fn compression_ratio(&self) -> f32
Ratio of original to compressed storage: original_bits / effective_bits.
Returns f32::INFINITY if the effective bits per param is 0.
Sourcepub fn total_original_bits(&self) -> f64
pub fn total_original_bits(&self) -> f64
Total original bits for this layer.
Sourcepub fn total_compressed_bits(&self) -> f64
pub fn total_compressed_bits(&self) -> f64
Total compressed bits for this layer.
Trait Implementations§
Source§impl Clone for CompressionMetrics
impl Clone for CompressionMetrics
Source§fn clone(&self) -> CompressionMetrics
fn clone(&self) -> CompressionMetrics
1.0.0 · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
source. Read more