Low-impact performance optimizations for all SuStaIn classes - #84

Draft

gdevenyi wants to merge 15 commits into

ucl-pond:masterfrom

gdevenyi:performance-glm-5-1-low-impact

Commits on Apr 21, 2026

Replace np.tile with broadcasting, vectorize loops in AbstractSustain
gdevenyi
committed
Replace np.tile with broadcasting, np.argsort, np.logical_and in ZscoreSustain
gdevenyi
committed
Replace np.tile with broadcasting, np.argsort, dtype=bool in OrdinalSustain
gdevenyi
committed
Replace np.tile with broadcasting, np.argsort, np.argmax in MixtureSustain
gdevenyi
committed
Replace np.tile with broadcasting, np.argsort, np.logical_and in ZScoreSustainMissingData
gdevenyi
committed
Replace np.tile with broadcasting, np.argsort, np.logical_and in MixedTypeSustain
gdevenyi
committed
Fix DataFrame.append removal in pandas 2.0+
gdevenyi
committed
perf: cache invariant constants in ZscoreSustain.__init__
gdevenyi
committed
perf: replace scipy.stats.norm.pdf with manual Gaussian in MixedTypeSustain
gdevenyi
committed
perf: replace O(J*B) product per stage with O(J) running log-sum in OrdinalSustain
gdevenyi
committed
perf: cache NaN mask and use broadcasting in ZScoreSustainMissingData
gdevenyi
committed
perf: eliminate (J,I,K+1) intermediate tensor in ZscoreSustain
gdevenyi
committed
perf: replace multiply-then-sum with einsum in _calculate_likelihood
gdevenyi
committed
perf: replace np.delete+np.concatenate with in-place array moves
gdevenyi
committed
perf: replace NaN pre-allocated EM trace arrays with list.append
gdevenyi
committed