Struct MbConvBlock 

pub struct MbConvBlock { /* private fields */ }

EfficientNet MBConv inverted residual block.

Weight layout (row-major):

• expand_w: [exp_ch × in_ch] (1×1 expand conv, bias: expand_b [exp_ch])
• dw_w : [exp_ch × k × k] (depthwise weights, bias: dw_b [exp_ch])
• se_fc1_w: [se_ch × exp_ch] (SE squeeze, bias: se_fc1_b [se_ch])
• se_fc2_w: [exp_ch × se_ch] (SE excite, bias: se_fc2_b [exp_ch])
• proj_w : [out_ch × exp_ch] (1×1 project conv, bias: proj_b [out_ch])

Implementations

impl MbConvBlock

pub fn new(config: MbConvConfig, rng: &mut VisionRng) -> VisionResult<Self>

Construct a new MBConv block with Xavier-uniform weight initialization.

Errors

• VisionError::InvalidImageSize if in_channels or out_channels is 0.
• VisionError::InvalidEmbedDim if expand_ratio is 0.
• VisionError::NonPositiveTemperature if se_ratio ≤ 0.

pub fn has_skip(&self) -> bool

Whether this block uses a skip (residual) connection.

pub fn forward(&self, x: &[f32], batch_size: usize) -> VisionResult<Vec<f32>>

Forward pass.

Input / output layout

x: [batch_size × in_channels] row-major. Returns: [batch_size × out_channels] row-major.

Pipeline (per sample)

1. Expand : h = ReLU6(W_exp · x + b_exp) output: [exp_ch]
2. Depthwise : channel-wise multiplication by dw_w[c, :] mean (1D proxy), then bias + ReLU6. Full 2-D depthwise conv would require H×W spatial input; here we use dw_w[c] as a per-channel scale (mean of k×k filter weights).
3. SE : global pool → FC1+ReLU → FC2+sigmoid → broadcast multiply.
4. Project : out = W_proj · h_se + b_proj (no activation).
5. Skip : if has_skip, out += x.

Errors

• VisionError::DimensionMismatch if x.len() != batch_size * in_channels.