Enum ConvStep 

pub enum ConvStep<'a> {
    GnSilu {
        w: &'a [f32],
        b: &'a [f32],
    },
    Conv {
        w: &'a [f32],
        b: &'a [f32],
        out_ch: usize,
        in_ch: usize,
        k: usize,
        stride: usize,
        pad: usize,
        far: bool,
    },
    Save,
    AddSaved {
        via: Option<(&'a [f32], &'a [f32])>,
    },
}

One stage of a convolution chain over an activation [c][h][w].

Variants

GnSilu

GroupNorm(32) with its affine weights, then SiLU, in place.

Fields

w: &'a [f32]

b: &'a [f32]

Conv

A k×k convolution [out, in, k, k] and its bias. pad: rows and columns of reflect padding on every side; far: one reflected row and column on the far side instead (a downsample’s padding).

Fields

w: &'a [f32]

b: &'a [f32]

out_ch: usize

in_ch: usize

k: usize

stride: usize

pad: usize

far: bool

Save

Keep the current activation as the residual branch’s input.

AddSaved

Add the kept activation — through a 1×1 convolution (weight, bias) when the branch changes the channel count.

Fields

via: Option<(&'a [f32], &'a [f32])>