Struct MbConvBlock 

pub struct MbConvBlock {
    pub expand_conv: Option<Conv2dLayer>,
    pub expand_bn: Option<BatchNorm2dLayer>,
    pub depthwise_w: Tensor,
    pub depthwise_bn: BatchNorm2dLayer,
    pub se: Option<SqueezeExciteBlock>,
    pub project_conv: Conv2dLayer,
    pub project_bn: BatchNorm2dLayer,
    pub use_residual: bool,
    pub expanded_ch: usize,
}

MBConv block (EfficientNet / MobileNetV2 inverted residual, inference-mode).

Structure: expand 1x1 -> BN -> activation -> depthwise 3x3 -> BN -> activation -> SE -> project 1x1 -> BN. Supports optional skip connection when stride=1 and in_channels=out_channels.

Fields

expand_conv: Option<Conv2dLayer>

expand_bn: Option<BatchNorm2dLayer>

depthwise_w: Tensor

depthwise_bn: BatchNorm2dLayer

se: Option<SqueezeExciteBlock>

project_conv: Conv2dLayer

project_bn: BatchNorm2dLayer

use_residual: bool

expanded_ch: usize

Implementations

impl MbConvBlock

pub fn new( in_channels: usize, out_channels: usize, expand_ratio: usize, kernel_size: usize, stride: usize, se_ratio: Option<usize>, epsilon: f32, ) -> Result<Self, ModelError>

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

Forward inference: input [N,H,W,C_in] -> [N,H',W',C_out].