`shrinkage.ProximalFit`

Shrinkage for Python · python · 0.2.0

ProximalFit(_native: shrinkage._shrinkage._ModelFit)

Owned original-scale parameters and full original-loss diagnostics.

Inspect termination before treating the result as converged. Arrays are independent snapshots, and the result retains no training data. The objective, mapping, and dual diagnostics use optimization-scale penalties.

`solver`

Source: packages/python/src/shrinkage/fit.py

solver(self) -> typing.Literal["hybrid", "proximal_gradient"]

Explicitly selected numerical method.

`family`

Source: packages/python/src/shrinkage/fit.py

family(self) -> Family

"gaussian" or "binomial"; binomial predictions are logits.

`coefficients`

Source: packages/python/src/shrinkage/fit.py

coefficients(self) -> FloatArray

Original-scale coefficients, shape (n_features,).

`intercept`

Source: packages/python/src/shrinkage/fit.py

intercept(self) -> float

Original-scale intercept, including the centering correction.

`preprocessing`

Source: packages/python/src/shrinkage/fit.py

preprocessing(self) -> shrinkage.Preprocessing

Training transformations; prediction never estimates new ones.

`termination`

Source: packages/python/src/shrinkage/fit.py

termination(self) -> shrinkage.Termination

Whether the full mapping passed or an iteration budget expired.

`iterations`

Source: packages/python/src/shrinkage/fit.py

iterations(self) -> int

Accepted PG updates, or hybrid PG transitions plus block sweeps.

`objective`

Source: packages/python/src/shrinkage/fit.py

objective(self) -> float

Original averaged loss plus the complete optimization-scale penalty.

`stopping`

Source: packages/python/src/shrinkage/fit.py

stopping(self) -> shrinkage.StoppingCriterion

Resolved full-space proximal-gradient mapping criterion.

`stopping_value`

Source: packages/python/src/shrinkage/fit.py

stopping_value(self) -> float

Full mapping infinity norm, including the free-intercept derivative.

`stopping_threshold`

Source: packages/python/src/shrinkage/fit.py

stopping_threshold(self) -> float

Absolute mapping tolerance used to certify convergence.

`step_size`

Source: packages/python/src/shrinkage/fit.py

step_size(self) -> float

Step used to evaluate the final full original-loss mapping.

`dual_certificate`

Source: packages/python/src/shrinkage/fit.py

dual_certificate(self) -> FloatArray | None

Feasible original-loss dual vector, shape (n_samples,), if available.

`dual_objective`

Source: packages/python/src/shrinkage/fit.py

dual_objective(self) -> float | None

Original-problem dual lower bound, or None without screening.

`duality_gap`

Source: packages/python/src/shrinkage/fit.py

duality_gap(self) -> float | None

Padded original-problem gap; the full mapping controls convergence.

`screening_diagnostics`

Source: packages/python/src/shrinkage/fit.py

screening_diagnostics(self) -> shrinkage.ScreeningDiagnostics

Certified restrictions and work in the original feature space.

`predict`

Source: packages/python/src/shrinkage/fit.py

predict(x: object, *, copy: bool = True) -> FloatArray

Predict means or logits, returning float64 shape (n_samples,).

Binomial predictions are logits, not probabilities or class labels. Input rules match fitting, except zero rows are allowed. Inputs are copied before releasing the GIL and must remain unchanged until return. Prediction uses original-scale coefficients without refitting any preprocessing. copy=False is unsupported.