2.0.0

v2.0.0 — Solver Pattern Abstraction

Breaking changes

• None — full backward compatibility preserved. All existing and free-function (, , etc.) APIs work unchanged.

New features

• Solver Pattern abstraction (issue #12): abstract base class with 5 concrete subclasses:
  • — SVD-based ridge regression (the main workhorse)
  • — ordinary least squares
  • — CG on normal equations (10–50x faster than SVD with identical results)
  • — robust Cauchy-loss IRLS (downweights outliers)
  • — SciPy nonlinear Cauchy (reference validator)
• Dependency injection: accepts any instance
• Compatibility guards: warns when robust loss is used with a non-robust solver; raises when non-SVD solver is used with multi-alpha arrays
• 22 new tests for the Solver API (177 total, zero regressions)

Example

from pyeeg.solvers import ConjugateGradientSolver, IRLSSolver
from pyeeg import TRFEstimator

# Fast iterative solver
trf = TRFEstimator(alpha=100.0, solver=ConjugateGradientSolver())

# Robust fitting
trf = TRFEstimator(alpha=100.0, solver=IRLSSolver(max_iter=50))

See scripts/examples/solver_showcase.py for a full comparison.

PyPI

https://pypi.org/project/natmeeg/2.0.0/