Smoothing certification and telemetry#
After fit_reml(), three attributes say how
automatic smoothing ended; after a plain fit(),
which selects no smoothing, all three are None.
smoothing_convergence_reason_ names why it
stopped (fixed_only when every penalty was fixed by policy).
smoothing_certified_ is True only for a
strict interior stop with every check passed: the fit converged for any
reason other than practical_plateau, the coefficient solve converged on
the curvature it was asked for with no fallback, no term was left pushing
against the finite cap (and, after a stationary stop, none rested on it),
and no term ended on an exact face. Any one of those voids it, whatever the
family, so read the other attributes before treating False as a defect: a
term selected away or shrunk to its null space is listed in
exact_face_components_, a term still pushing
against the cap in
smoothing_unresolved_upper_bound_, and a
plateau stop, which the default practical_reml=True allows, in the
convergence reason. exact_face_components_ is
always a tuple: the terms whose penalty was accepted as infinite. What such
a term keeps depends on its penalty and is named in the fit diagnostics:
nothing at all, a straight line, or the rest of the penalty’s null space.
coefficient_curvature reports the curvature the
coefficient solve was asked to use, observed or Fisher; the curvature used
at the end, after any fallback, is training_telemetry().curvature.actual_source.
training_telemetry() returns the immutable audit
record of the accepted fit.
Return strict matched-certification status, or |
|
Return how automatic smoothing stopped, or |
|
Return smoothing components with unresolved pressure at the finite cap. |
|
Return smoothing components accepted at the exact infinity face. |
|
Return the curvature the dense coefficient solve is asked to use. |
|
Return immutable audit metadata for the accepted fitted revision. |