Struct PpcaModel 

pub struct PpcaModel {
    pub num_channels: usize,
    pub num_components: usize,
    pub mean: Vec<f32>,
    pub weights: Vec<f32>,
    pub posterior_projection: Vec<f32>,
    pub noise_variance: f32,
    pub eigenvalues: Vec<f32>,
}

Probabilistic Principal Component Analysis (PPCA) model (Tipping & Bishop, 1999).

Generative Gaussian latent variable model: $$\mathbf{x} = \mathbf{W}\mathbf{z} + \boldsymbol{\mu} + \boldsymbol{\epsilon}, \quad \mathbf{z} \sim \mathcal{N}(\mathbf{0}, \mathbf{I}_q), \quad \boldsymbol{\epsilon} \sim \mathcal{N}(\mathbf{0}, \sigma^2 \mathbf{I}_D)$$

Fields

num_channels: usize

num_components: usize

mean: Vec<f32>

weights: Vec<f32>

Factor loading matrix $\mathbf{W}$ of shape [num_channels, num_components] (row-major).

posterior_projection: Vec<f32>

Posterior projection matrix $\mathbf{P} = \mathbf{W} (\mathbf{W}^\top \mathbf{W} + \sigma^2 \mathbf{I})^{-1}$ of shape [num_channels, num_components] so $\mathbb{E}[\mathbf{z} \mid \mathbf{x}] = \mathbf{P}^\top (\mathbf{x} - \boldsymbol{\mu})$.

noise_variance: f32

Isotropic observation noise variance $\sigma^2$.

eigenvalues: Vec<f32>

Top-$q$ eigenvalues of the sample covariance matrix.

Implementations

impl PpcaModel

pub fn fit<F: DspFloat>( client: &Client, data: &[f32], channels: usize, samples: usize, num_components: usize, ) -> Self

Fits PPCA in closed form (maximum likelihood from the covariance eigendecomposition) on host data [channels, samples]; covariance and eigendecomposition on the device in F.

pub fn fit_em_masked<F: DspFloat>( client: &Client, data: &[f32], observed_mask: &[bool], channels: usize, samples: usize, num_components: usize, max_iters: usize, ) -> Self

Fits PPCA via Expectation-Maximization (EM) with an observation mask ([channels, samples], true = observed, false = missing/outside local neighborhood).

pub fn project_cpu( &self, data: &[f32], channels: usize, samples: usize, ) -> Vec<f32>

Computes posterior latent expectations $\mathbb{E}[\mathbf{z} \mid \mathbf{x}] = \mathbf{M}^{-1}\mathbf{W}^\top(\mathbf{x} - \boldsymbol{\mu})$ of shape [num_components, samples] on the CPU.

pub fn reconstruct_cpu(&self, z: &[f32], samples: usize) -> Vec<f32>

Reconstructs observations $\hat{\mathbf{x}} = \mathbf{W}\mathbf{z} + \boldsymbol{\mu}$ ([channels, samples]) from latent codes z ([num_components, samples]).

pub fn to_device<F: DspFloat>(&self, client: &Client) -> DeviceProjection

The projection with its weights uploaded once as F, for repeated calls.

pub fn project_gpu<F: DspFloat>( &self, client: &Client, input_handle: &Handle, output_handle: &Handle, channels: usize, samples: usize, )

One-off device projection of [channels, samples] into [num_components, samples] (uploads the weights; use Self::to_device for repeated calls).

Trait Implementations

impl Clone for PpcaModel

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for PpcaModel

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

Formats the value using the given formatter.