dspk.ops.linalg

Models fitted on the device: PCA and probabilistic PCA, scikit-learn style.

Data is (channels, samples): channels are the features, samples the observations. FastICA follows once its iterations run on the device.

PCA	Principal component analysis, fitted on the device.
PPCA	Probabilistic PCA (Tipping & Bishop).

Classes

PCA

Principal component analysis, fitted on the device.

Channels are the features and samples the observations. Make one with fit.

Examples

>>> import numpy as np
>>> import dspk
>>> x = np.random.default_rng(0).standard_normal((8, 5_000)).astype(np.float32)
>>> pca = dspk.ops.linalg.PCA.fit(x, 3)
>>> pca.transform(x).shape
(3, 5000)

Properties

Name	Type	Description
components	np.ndarray	The components (channels, n_components), one per column.
mean	np.ndarray	Mean of each channel (channels,), removed before projecting.
explained_variance	np.ndarray	Variance along each component (n_components,).
explained_variance_ratio	np.ndarray	Each component’s share of the total variance (n_components,).

PCA.fit

fit(
    data,
    n_components: int,
    *,
    runtime: str | None = None,
) -> PCA

Fits n_components principal components.

Parameters

• data : array_like
  Signal (channels, samples), converted to float32 on the host first (these models read host data).
• n_components : int
  Components to keep, 1 to channels.
• runtime : str, default None
  A runtime from runtime.available ("cuda", "hip", "wgpu", "cpu") for this call. By default: the one chosen with runtime.set, else DSPK_RUNTIME, else the first available (GPUs first).

Returns

• PCA

Raises

• ValueError: For n_components outside 1..channels, or fewer than two samples.
• RuntimeError: When the runtime has no device.

PCA.transform

transform(data, *, runtime: str | None = None) -> np.ndarray

Projects data onto the components.

Parameters

• data : array_like
  Signal (channels, samples), converted to float32 on the host first (these models read host data).
• runtime : str, default None
  A runtime from runtime.available ("cuda", "hip", "wgpu", "cpu") for this call. By default: the one chosen with runtime.set, else DSPK_RUNTIME, else the first available (GPUs first).

Returns

• np.ndarray — (n_components, samples).

Raises

• ValueError: When data has another channel count than the fit.

PPCA

Probabilistic PCA (Tipping & Bishop): principal components plus an isotropic noise level.

Closed-form maximum likelihood, fitted on the device. Make one with fit.

Properties

Name	Type	Description
weights	np.ndarray	The loadings W (channels, n_components).
mean	np.ndarray	Mean of each channel (channels,).
noise_variance	float	The isotropic noise variance, sigma squared.

PPCA.fit

fit(
    data,
    n_components: int,
    *,
    runtime: str | None = None,
) -> PPCA

Fits n_components latent dimensions.

Parameters

• data : array_like
  Signal (channels, samples), converted to float32 on the host first (these models read host data).
• n_components : int
  Latent dimensions, 1 to channels.
• runtime : str, default None
  A runtime from runtime.available ("cuda", "hip", "wgpu", "cpu") for this call. By default: the one chosen with runtime.set, else DSPK_RUNTIME, else the first available (GPUs first).

Returns

• PPCA

Raises

• ValueError: For n_components outside 1..channels, or fewer than two samples.
• RuntimeError: When the runtime has no device.

PPCA.transform

transform(data, *, runtime: str | None = None) -> np.ndarray

The posterior mean of the latent variables, E[z | x], for data.

Parameters

• data : array_like
  Signal (channels, samples), converted to float32 on the host first (these models read host data).
• runtime : str, default None
  A runtime from runtime.available ("cuda", "hip", "wgpu", "cpu") for this call. By default: the one chosen with runtime.set, else DSPK_RUNTIME, else the first available (GPUs first).

Returns

• np.ndarray — (n_components, samples).

Raises

• ValueError: When data has another channel count than the fit.