Struct QuantizationParams 

pub struct QuantizationParams {
    pub scale: f32,
    pub zero_point: i32,
    pub qmin: i8,
    pub qmax: i8,
}

Quantization parameters for a tensor or tensor slice.

Defines the mapping between floating-point values and quantized integers:

• Symmetric: x_q = round(x / scale)
• Asymmetric: x_q = round(x / scale) + zero_point

Invariants

• scale > 0.0 (enforced at construction)
• qmin <= zero_point <= qmax
• For symmetric mode: zero_point == 0

Fields

scale: f32

Scale factor for quantization. Maps FP32 range to INT8 range.

zero_point: i32

Zero point for asymmetric quantization. Always 0 for symmetric quantization.

qmin: i8

Minimum quantized value (typically -128 for i8).

qmax: i8

Maximum quantized value (typically 127 for i8).

Implementations

impl QuantizationParams

pub fn from_minmax( min_val: f32, max_val: f32, mode: QuantizationMode, ) -> CnnResult<Self>

Create symmetric quantization parameters from min/max values.

Symmetric quantization uses zero_point = 0 and maps the range [-max_abs, max_abs] to [-127, 127].

Arguments

• min_val - Minimum value in the FP32 tensor
• max_val - Maximum value in the FP32 tensor

Returns

Quantization parameters with zero_point = 0.

Example

let params = QuantizationParams::from_minmax(-10.0, 10.0, QuantizationMode::Symmetric);
assert_eq!(params.zero_point, 0);
assert!(params.scale > 0.0);

pub fn from_percentile( percentile_min: f32, percentile_max: f32, mode: QuantizationMode, ) -> CnnResult<Self>

Create quantization parameters from percentile values.

Used during calibration to exclude outliers that would skew the quantization range. Typically uses 0.001 and 0.999 percentiles.

Arguments

• percentile_min - Lower percentile value (e.g., -10.0)
• percentile_max - Upper percentile value (e.g., 10.0)
• mode - Quantization mode (symmetric or asymmetric)

Example

// Use 99.8% of the data range, excluding outliers
let params = QuantizationParams::from_percentile(
    -9.5, 9.5, QuantizationMode::Symmetric
);

pub fn validate(&self) -> CnnResult<()>

Validate that the parameters satisfy invariants.

Invariants

• scale > 0.0
• qmin <= qmax
• qmin <= zero_point <= qmax
• For symmetric mode: zero_point == 0

pub fn quantize_value(&self, value: f32) -> i8

Quantize a single FP32 value to INT8.

Formula:

• Symmetric: x_q = round(x / scale)
• Asymmetric: x_q = round(x / scale) + zero_point

Result is clamped to [qmin, qmax].

pub fn dequantize_value(&self, value: i8) -> f32

Dequantize a single INT8 value to FP32.

Formula:

• Symmetric: x = x_q * scale
• Asymmetric: x = (x_q - zero_point) * scale

Trait Implementations

impl Clone for QuantizationParams

fn clone(&self) -> QuantizationParams

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for QuantizationParams

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for QuantizationParams

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for QuantizationParams

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.