Struct ConvNextBlock 

pub struct ConvNextBlock { /* private fields */ }

A single ConvNeXt block.

All weight tensors are flat row-major Vec<f32>.

Implementations

impl ConvNextBlock

pub fn new(cfg: ConvNextConfig, rng: &mut LcgRng) -> VisionResult<Self>

Construct a new block with random depthwise / pointwise weights, zeroed depthwise bias, identity LayerNorm affine, and layer-scale gamma set to cfg.layer_scale_init.

Kernels and pointwise weights use a Kaiming-ish scaled normal initialisation drawn from the deterministic LCG RNG.

Errors

Propagates configuration validation from ConvNextConfig::new.

pub fn config(&self) -> &ConvNextConfig

Read-only access to the block configuration.

pub fn dw_kernel_mut(&mut self) -> &mut [f32]

Mutable access to the depthwise kernel (used for tests that install an identity / delta kernel).

pub fn dw_bias_mut(&mut self) -> &mut [f32]

Mutable access to the depthwise bias.

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

Depthwise convolution: each channel c is convolved with its own kernel × kernel filter, zero-padded “same” (pad = (k-1)/2), stride 1, plus per-channel bias.

Input / output are (C, H, W) flat.

Errors

DimensionMismatch if x.len() != C·H·W.

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

Channel LayerNorm: at each spatial position (h, w), gather the C channel values, normalise to zero-mean / unit-variance (ε = 1e-6), then apply per-channel affine γ · x̂ + β.

Input / output are (C, H, W) flat.

Errors

DimensionMismatch if x.len() != C·H·W.

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

Forward pass: (C·H·W) → (C·H·W).

See the module docs for the full pipeline. When layer_scale_init is 0.0, the residual branch is multiplied by zero so forward(x) == x.

Errors

• DimensionMismatch if x.len() != C·H·W.
• NonFinite if the output contains non-finite values.

pub fn n_params(&self) -> usize

Total number of learnable parameters in this block.

dw_kernel  : C · k²
dw_bias    : C
ln_gamma   : C
ln_beta    : C
pw1_weight : hidden · C
pw1_bias   : hidden
pw2_weight : C · hidden
pw2_bias   : C
layer_scale: C