Struct fetish_lib::func_schmear::FuncSchmear

pub struct FuncSchmear {
    pub mean: Array2<f32>,
    pub covariance: FuncScatterTensor,
}

Represents a probability distribution over linear mappings whose covariance structure is separable between input and output. This can be conceptualized as a specialization of Schmear for a distribution over matrices where the covariance of the (vectorized) random variable takes the separable form kron(out_covariance, in_covariance) for out_covariance and in_covariance output coordinate and input coordinate covariances, respectively, and kron referring to crate::linalg_utils::kron See also FuncScatterTensor.

Fields

mean: Array2<f32>

covariance: FuncScatterTensor

Implementations

impl FuncSchmear

pub fn compress(&self, mat: ArrayView2<'_, f32>) -> Schmear

Given a transformation matrix from the full, flattened dimension of this FuncSchmear to a smaller dimension, performs a fused FuncSchmear::flatten and Schmear::transform operation using the specified transformation. This fused operation is written to be much faster than manually performing the aforementioned operations.

pub fn flatten(&self) -> Schmear

Converts this FuncSchmear over linear maps to its corresponding Schmear over vectorized linear mappings.

pub fn apply(&self, x: &Schmear) -> Schmear

Computes the output Schmear of this FuncSchmear applied to a given argument Schmear. Given an input schmear, computes the output schmear which would result from sampling (function, input) pairs, computing function(input) for each of them, and then obtaining the Schmear over those results.