Struct SkipTranscoderRemlInputs 

pub struct SkipTranscoderRemlInputs<'a> {
    pub y_out: ArrayView2<'a, f64>,
    pub y_hat: ArrayView2<'a, f64>,
    pub z: ArrayView2<'a, f64>,
    pub w_dec: ArrayView2<'a, f64>,
    pub lambda_sparse: f64,
    pub skip_proj: Option<ArrayView2<'a, f64>>,
    pub skip_u: Option<ArrayView2<'a, f64>>,
}

Inputs to the closed-form Gaussian REML/Laplace score of a trained skip-transcoder.

Every feature of the effective design is a rank-1 outer product of a per-observation activation column with an output-space loading vector, flattened over the (observation, output) pair:

• sparse atom a: D_a = vec(z[:, a] · W_dec[a, :]^T),
• skip component r: D_r = vec((XV)[:, r] · U[:, r]^T),

where XV = x_in · skip_V is the skip’s data-dependent projection onto its rank_skip-dimensional input subspace and U = skip_U is its output loading. Because the skip map enters the prediction only through the products XV and U (the prediction is (x_in · V) · U^T), passing those two products — rather than V and U separately — makes the score invariant to the unidentifiable balancing gauge U -> cU, V -> V/c, which leaves the represented function unchanged.

Fields

y_out: ArrayView2<'a, f64>

y_hat: ArrayView2<'a, f64>

z: ArrayView2<'a, f64>

w_dec: ArrayView2<'a, f64>

lambda_sparse: f64

skip_proj: Option<ArrayView2<'a, f64>>

Skip input projection XV = x_in · skip_V, shape (n_obs, rank_skip).

skip_u: Option<ArrayView2<'a, f64>>

Skip output loading U = skip_U, shape (out_dim, rank_skip).