spd_learn.modules.TimeFrequencyCovariances#

class spd_learn.modules.TimeFrequencyCovariances(segments: Sequence[tuple[int, int, int]], *, center: bool = False, shrinkage: float = 0.0, ridge: float = 1e-05)[source]#

Bases: Module

Estimate covariance matrices from ordered filter-bank windows.

See [Ju and Guan, 2024].

Parameters:

  • segments (sequence of tuple) – Ordered (band, start, stop) windows in samples; stop is exclusive.

  • center (bool, default=False) – Subtract each channel mean within its window.

  • shrinkage (float, default=0.0) – Mixture weight of the trace-scaled identity, in [0, 1].

  • ridge (float, default=1e-05) – Positive absolute loading in squared signal units.

Notes

Without centering this computes a second moment, using divisor window length. Leading trials and channel order are preserved; filtering is external.

forward(input: Tensor) → Tensor[source]#

Estimate one covariance matrix per window.

Parameters:

input (torch.Tensor) – Filtered trials of shape (batch, bands, channels, samples).

Returns:

Covariances of shape (batch, nodes, channels, channels).

Return type:

torch.Tensor