Struct QuantizedTensor 

pub struct QuantizedTensor<T> { /* private fields */ }

Quantized tensor with INT8 data and metadata.

Stores quantized values along with the information needed to dequantize them back to FP32.

Invariants (Enforced at Construction)

• INV-1: data.len() == metadata.numel()
• INV-2: metadata.scale > 0.0
• INV-3: All values in data are in range [qmin, qmax]

Example

use ruvector_cnn::quantize::{QuantizedTensor, QuantizationParams, QuantizationMode};

let fp32_data = vec![1.0, 2.0, -1.0, 0.5];
let shape = vec![4];
let params = QuantizationParams::from_minmax(-2.0, 2.0, QuantizationMode::Symmetric)?;

// Quantize
let quantized = QuantizedTensor::<i8>::quantize(&fp32_data, &shape, &params)?;

// Dequantize
let dequantized = quantized.dequantize()?;

Implementations

impl QuantizedTensor<i8>

pub fn new(data: Vec<i8>, metadata: QuantizationMetadata) -> CnnResult<Self>

Create a new quantized tensor with validation.

Arguments

• data - Quantized INT8 values
• metadata - Quantization metadata (scale, zero_point, shape)

Errors

• If data.len() != metadata.numel() (INV-1)
• If metadata is invalid (INV-2)

pub fn quantize( fp32_data: &[f32], shape: &[usize], params: &QuantizationParams, ) -> CnnResult<Self>

Quantize FP32 data to INT8.

Arguments

• fp32_data - Input FP32 values
• shape - Tensor shape
• params - Quantization parameters

Returns

Quantized INT8 tensor.

Example

let fp32 = vec![1.0, 2.0, -1.0];
let shape = vec![3];
let params = QuantizationParams::from_minmax(-2.0, 2.0, QuantizationMode::Symmetric)?;
let quantized = QuantizedTensor::quantize(&fp32, &shape, &params)?;

pub fn dequantize(&self) -> CnnResult<Vec<f32>>

Dequantize INT8 data to FP32.

Returns

FP32 values with the same shape.

Example

let dequantized = quantized.dequantize()?;
assert_eq!(dequantized.len(), quantized.data().len());

pub fn data(&self) -> &[i8]

Get reference to quantized data.

pub fn data_mut(&mut self) -> &mut [i8]

Get mutable reference to quantized data.

pub fn metadata(&self) -> &QuantizationMetadata

Get reference to metadata.

pub fn shape(&self) -> &[usize]

Get tensor shape.

pub fn scale(&self) -> f32

Get scale factor.

pub fn zero_point(&self) -> i32

Get zero point.

pub fn check_bounds(&self, qmin: i8, qmax: i8) -> bool

Check bounds invariant: all values in [qmin, qmax].

This is a sanity check to ensure data hasn’t been corrupted. Should always return true for properly constructed tensors.

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

Validate all invariants.

Invariants

• INV-1: data.len() == metadata.numel()
• INV-2: metadata.scale > 0.0
• INV-3: All values in [-127, 127]

pub fn reshape(&mut self, new_shape: Vec<usize>) -> CnnResult<()>

Reshape the tensor to a new shape.

Arguments

• new_shape - New shape (must have same total elements)

Errors

If new_shape.iter().product() != self.data.len().

Trait Implementations

impl<T: Clone> Clone for QuantizedTensor<T>

fn clone(&self) -> QuantizedTensor<T>

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<T: Debug> Debug for QuantizedTensor<T>

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

Formats the value using the given formatter.

impl<'de, T> Deserialize<'de> for QuantizedTensor<T>

where T: Deserialize<'de>,

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

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl<T> Serialize for QuantizedTensor<T>

where T: Serialize,

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

where __S: Serializer,

Serialize this value into the given Serde serializer.