Struct QuantizedDepthwiseConv2d 

pub struct QuantizedDepthwiseConv2d { /* private fields */ }

Quantized Depthwise Convolution Layer

Performs depthwise separable convolution in INT8:

• Each input channel is convolved with a single kernel
• No cross-channel mixing (unlike standard convolution)
• Memory efficient for mobile architectures

Implementations

impl QuantizedDepthwiseConv2d

pub fn from_fp32( channels: usize, kernel_size: usize, weights: &[f32], bias: Option<&[f32]>, stride: usize, padding: usize, input_scale: f32, ) -> Self

Create from FP32 depthwise convolution

Arguments

• channels - Number of input/output channels
• kernel_size - Kernel size (assumed square)
• weights - FP32 weights [channels, kh, kw]
• bias - Optional FP32 bias [channels]
• stride - Stride
• padding - Padding
• input_scale - Expected input activation scale

pub fn forward_int8( &self, input: &[u8], input_shape: &[usize], input_scale: f32, input_zero_point: u8, ) -> CnnResult<Tensor>

Forward pass with INT8 computation

Arguments

• input - Quantized u8 input tensor (NHWC layout)
• input_shape - Input shape [N, H, W, C]
• input_scale - Input quantization scale
• input_zero_point - Input quantization zero point

Trait Implementations

impl Clone for QuantizedDepthwiseConv2d

fn clone(&self) -> QuantizedDepthwiseConv2d

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for QuantizedDepthwiseConv2d

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

Formats the value using the given formatter.