Version 0.5.0🔗

Enhancements🔗

  • Add high-performance PyTorch CPU backend acceleration for generalized eigendecomposition (GEVD), batched SENSAI artifact scoring, and epoch cleaning, delivering up to 2.2x speedup over NumPy and 1.23x speedup over reference MATLAB (Tomas Ros).

  • Implement chunked streaming MODWT Haar Wavelet transform with boundary-matched overlap-save state caching, reducing peak memory usage to under 15 MB across arbitrarily long continuous recordings (Tomas Ros).

  • Introduce precomputation of generalized eigendecompositions (_precompute_gevd) during SENSAI artifact line search, avoiding redundant decompositions across threshold iterations (Tomas Ros).

  • Implement robust generalized eigenvalue decomposition (robust_cholesky_gevd) with dynamic diagonal regularization for rank-deficient reference covariances (Tomas Ros).

  • Add support for decoupled fit_raw and transform_raw processing workflows with smart broadband signal caching (Tomas Ros).

  • Add configurable sensai_tol optimization tolerance parameter to Gedai and AdaptiveMultibandGedai for fine-grained convergence control (Tomas Ros).

  • Add optional [torch] installation extra (pip install "gedai[torch]") with lightweight CPU-only wheel installation instructions (Tomas Ros).

API and behavior changes🔗

  • Add engine parameter ("auto" | "torch" | "numpy") across Gedai, MultibandGedai, and AdaptiveMultibandGedai.

  • Default engine to "auto", which automatically enables PyTorch CPU acceleration when PyTorch is installed and transparently falls back to NumPy otherwise without raising missing dependency errors.

Authors🔗