Struct ConditionalPriorIvae 

pub struct ConditionalPriorIvae {
    pub mean: Array2<f64>,
    pub scale: Array2<f64>,
    pub weight: f64,
}

iVAE-style auxiliary-conditional Gaussian log-prior on the latent block.

Stores per-row conditional means μ of shape (n_rows, latent_dim) and scales σ of shape (n_rows, latent_dim), where (μ_{n,i}, σ_{n,i}) are presumed evaluated by some external Smooth at the auxiliary u_n. The negative log-prior contribution to the latent objective is

½ Σ_n Σ_i [ ((t_{n,i} − μ_{n,i}) / σ_{n,i})² + 2 log σ_{n,i} + log 2π ]

scaled by weight. The gradient w.r.t. t is (t − μ) / σ² (times weight); the gradient w.r.t. μ is its negative. Per-row scales make this strictly more general than a fixed N(0, I), which is recovered by μ ≡ 0, σ ≡ 1.

Fields

mean: Array2<f64>

scale: Array2<f64>

weight: f64

Implementations

impl ConditionalPriorIvae

pub fn new( mean: Array2<f64>, scale: Array2<f64>, weight: f64, ) -> Result<Self, String>

pub fn value_and_grad(&self, t: ArrayView2<'_, f64>) -> (f64, Array2<f64>)

Evaluate negative-log-prior value and gradient w.r.t. latent t.

pub fn value(&self, t: ArrayView2<'_, f64>) -> f64

Evaluate value only — useful when only the loss is needed.

Trait Implementations

impl Clone for ConditionalPriorIvae

fn clone(&self) -> ConditionalPriorIvae

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ConditionalPriorIvae

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.