spd_learn.functional.log_tokens_to_spd#

spd_learn.functional.log_tokens_to_spd(tokens: Tensor, *, preserve_norm: bool = False, matrix_scale: float = 1.0) → Tensor[source]#

Reconstruct matrices from symmetric log coordinates.

Parameters:

  • tokens (torch.Tensor) – Log coordinates of shape (…, n * (n + 1) // 2).

  • preserve_norm (bool, default=False) – Scale off-diagonal coordinates by sqrt(2).

  • matrix_scale (float, default=1.0) – Positive matrix multiplier before the logarithm.

Returns:

Reconstructed matrices of shape (…, n, n), after undoing matrix_scale.

Return type:

torch.Tensor

Notes

Use the same basis and scale as the forward map. Previously floored eigenvalues cannot be recovered.