Struct FakeQuantize 

pub struct FakeQuantize {
    pub bits: u32,
    pub symmetric: bool,
    pub scale: f32,
    pub zero_point: i32,
    pub enabled: bool,
}

Fake quantization operator for quantization-aware training (QAT).

Maintains the current scale and zero-point that are updated during calibration / training via an associated observer.

Fields

bits: u32

Quantization bit-width.

symmetric: bool

Whether to use symmetric quantization (zp = 0).

scale: f32

Current quantization scale (must be > 0).

zero_point: i32

Current zero-point.

enabled: bool

Whether fake quantization is enabled. When disabled, forward returns the input unchanged.

Implementations

impl FakeQuantize

pub fn new( bits: u32, symmetric: bool, scale: f32, zero_point: i32, ) -> QuantResult<Self>

Create a new fake quantizer with the given scale and zero-point.

Errors

• QuantError::InvalidBitWidth — bits is 0 or > 16.
• QuantError::InvalidScale — scale is ≤ 0 or non-finite.

pub fn with_defaults(bits: u32, symmetric: bool) -> QuantResult<Self>

Create with default scale=1.0, zp=0 for the given bit-width.

Errors

• QuantError::InvalidBitWidth — bits is 0 or > 16.

pub fn update_params(&mut self, scale: f32, zero_point: i32) -> QuantResult<()>

Update scale and zero-point (e.g., from an observer).

Errors

• QuantError::InvalidScale — scale is ≤ 0 or non-finite.

pub fn quant_range(&self) -> (i32, i32)

Integer quantization bounds [q_min, q_max].

pub fn float_range(&self) -> (f32, f32)

Float clipping bounds [x_min, x_max] corresponding to the integer range.

pub fn forward(&self, x: &[f32]) -> Vec<f32>

Forward pass: quantize-then-dequantize.

If enabled = false, returns the input unchanged.

pub fn backward(&self, grad_output: &[f32], x: &[f32]) -> QuantResult<Vec<f32>>

Backward pass (Straight-Through Estimator).

Passes grad_output through where x is inside the representable float range; zeros the gradient where x is clipped.

Errors

• QuantError::DimensionMismatch — grad_output and x lengths differ.

pub fn quantization_noise(&self, x: &[f32]) -> f32

Estimate quantization noise (MSE between input and fake-quantized output).

Useful for measuring quantization error at the current scale/zp.

Trait Implementations

impl Clone for FakeQuantize

fn clone(&self) -> FakeQuantize

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for FakeQuantize

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

Formats the value using the given formatter.