Struct FakeQuantize 

pub struct FakeQuantize {
    pub config: FakeQuantConfig,
    pub scale: f32,
    pub zero_point: i32,
    pub initialized: bool,
}

Fake quantization operation with Straight-Through Estimator (STE)

This struct holds the state for fake quantization including learned or calibrated scale and zero_point parameters.

Fields

config: FakeQuantConfig

Quantization configuration

scale: f32

Scale factor for quantization

zero_point: i32

Zero point for asymmetric quantization

initialized: bool

Whether scale has been initialized

Implementations

impl FakeQuantize

pub fn new(config: FakeQuantConfig) -> Self

Create new fake quantization operation

pub fn q4() -> Self

Create with 4-bit symmetric quantization

pub fn q8() -> Self

Create with 8-bit symmetric quantization

pub fn calibrate(&mut self, data: &[f32])

Initialize scale from data (min-max calibration)

For symmetric: scale = max(|min|, |max|) / qmax For asymmetric: scale = (max - min) / (qmax - qmin)

pub fn forward(&self, input: &Tensor) -> Tensor

Forward pass: fake quantize (quantize → dequantize)

Simulates quantization effects while keeping values in floating point. Output = dequantize(quantize(input))

pub fn forward_with_calibration(&mut self, input: &Tensor) -> Tensor

Forward pass with auto-calibration

If not initialized, calibrates from input data first.

pub fn backward(&self, grad_output: &Tensor) -> Tensor

Backward pass: Straight-Through Estimator (STE)

The gradient passes through unchanged: ∂L/∂x = ∂L/∂y (where y = fake_quantize(x))

This allows gradients to flow during training despite the non-differentiable quantization operation.

pub fn backward_clamped(&self, grad_output: &Tensor, input: &Tensor) -> Tensor

Backward pass with gradient clipping (clamped STE)

Clips gradients to zero outside the quantization range. This can improve training stability.

pub fn scale(&self) -> f32

Get the quantization scale

pub fn zero_point(&self) -> i32

Get the zero point

pub fn is_initialized(&self) -> bool

Check if calibrated

pub fn num_levels(&self) -> usize

Get number of quantization levels

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.