|
| 1 | +import numpy as np |
| 2 | +from joblib import Parallel, delayed |
| 3 | +from skglm.datafits import Logistic, QuadraticSVC |
| 4 | +from skglm.estimators import GeneralizedLinearEstimator |
| 5 | +from sklearn.model_selection import KFold, StratifiedKFold |
| 6 | +from sklearn.metrics import accuracy_score, mean_squared_error |
| 7 | + |
| 8 | + |
| 9 | +class GeneralizedLinearEstimatorCV(GeneralizedLinearEstimator): |
| 10 | + """Cross-validated wrapper for GeneralizedLinearEstimator. |
| 11 | +
|
| 12 | + This class performs cross-validated selection of the regularization parameter(s) |
| 13 | + for a generalized linear estimator, supporting both L1 and elastic-net penalties. |
| 14 | +
|
| 15 | + Parameters |
| 16 | + ---------- |
| 17 | + datafit : object |
| 18 | + Datafit (loss) function instance (e.g., Logistic, Quadratic). |
| 19 | + penalty : object |
| 20 | + Penalty instance with an 'alpha' parameter (and optionally 'l1_ratio'). |
| 21 | + solver : object |
| 22 | + Solver instance to use for optimization. |
| 23 | + alphas : array-like of shape (n_alphas,), optional |
| 24 | + List of alpha values to try. If None, they are set automatically. |
| 25 | + l1_ratio : float or array-like, optional |
| 26 | + The ElasticNet mixing parameter(s), with 0 <= l1_ratio <= 1. |
| 27 | + Only used if the penalty supports 'l1_ratio'. If None, defaults to 1.0 (Lasso). |
| 28 | + cv : int, default=4 |
| 29 | + Number of cross-validation folds. |
| 30 | + n_jobs : int, default=1 |
| 31 | + Number of jobs to run in parallel for cross-validation. |
| 32 | + random_state : int or None, default=None |
| 33 | + Random seed for cross-validation splitting. |
| 34 | + eps : float, default=1e-3 |
| 35 | + Ratio of minimum to maximum alpha if alphas are set automatically. |
| 36 | + n_alphas : int, default=100 |
| 37 | + Number of alphas along the regularization path if alphas are set automatically. |
| 38 | +
|
| 39 | + Attributes |
| 40 | + ---------- |
| 41 | + alpha_ : float |
| 42 | + Best alpha found by cross-validation. |
| 43 | + l1_ratio_ : float or None |
| 44 | + Best l1_ratio found by cross-validation (if applicable). |
| 45 | + best_estimator_ : GeneralizedLinearEstimator |
| 46 | + Estimator fitted on the full data with the best parameters. |
| 47 | + coef_ : ndarray |
| 48 | + Coefficients of the fitted model. |
| 49 | + intercept_ : float or ndarray |
| 50 | + Intercept of the fitted model. |
| 51 | + alphas_ : ndarray |
| 52 | + Array of alphas used in the search. |
| 53 | + scores_path_ : ndarray |
| 54 | + Cross-validation scores for each parameter combination. |
| 55 | + n_iter_ : int or None |
| 56 | + Number of iterations run by the solver (if available). |
| 57 | + n_features_in_ : int or None |
| 58 | + Number of features seen during fit. |
| 59 | + feature_names_in_ : ndarray or None |
| 60 | + Names of features seen during fit. |
| 61 | + """ |
| 62 | + |
| 63 | + def __init__(self, datafit, penalty, solver, alphas=None, l1_ratio=None, |
| 64 | + cv=4, n_jobs=1, random_state=None, |
| 65 | + eps=1e-3, n_alphas=100): |
| 66 | + super().__init__(datafit=datafit, penalty=penalty, solver=solver) |
| 67 | + self.alphas = alphas |
| 68 | + self.l1_ratio = l1_ratio |
| 69 | + self.cv = cv |
| 70 | + self.n_jobs = n_jobs |
| 71 | + self.random_state = random_state |
| 72 | + self.eps = eps |
| 73 | + self.n_alphas = n_alphas |
| 74 | + |
| 75 | + def _score(self, y_true, y_pred): |
| 76 | + """Compute the performance score (higher is better).""" |
| 77 | + if isinstance(self.datafit, (Logistic, QuadraticSVC)): |
| 78 | + return accuracy_score(y_true, y_pred) |
| 79 | + return -mean_squared_error(y_true, y_pred) |
| 80 | + |
| 81 | + def fit(self, X, y): |
| 82 | + """Fit the model using cross-validation.""" |
| 83 | + if not hasattr(self.penalty, "alpha"): |
| 84 | + raise ValueError( |
| 85 | + "GeneralizedLinearEstimatorCV only supports penalties which " |
| 86 | + "expose an 'alpha' parameter." |
| 87 | + ) |
| 88 | + n_samples, n_features = X.shape |
| 89 | + |
| 90 | + if self.alphas is not None: |
| 91 | + alphas = np.sort(self.alphas)[::-1] |
| 92 | + else: |
| 93 | + alpha_max = np.max(np.abs(X.T @ y)) / n_samples |
| 94 | + alphas = np.geomspace( |
| 95 | + alpha_max, |
| 96 | + alpha_max * self.eps, |
| 97 | + self.n_alphas |
| 98 | + ) |
| 99 | + has_l1_ratio = hasattr(self.penalty, "l1_ratio") |
| 100 | + l1_ratios = [1.] if not has_l1_ratio else np.atleast_1d( |
| 101 | + self.l1_ratio if self.l1_ratio is not None else [1.]) |
| 102 | + |
| 103 | + scores_path = np.empty((len(l1_ratios), len(alphas), self.cv)) |
| 104 | + best_loss = -np.inf |
| 105 | + |
| 106 | + def _solve_fold(k, train, test, alpha, l1, w_init): |
| 107 | + pen_kwargs = {k: v for k, v in self.penalty.__dict__.items() |
| 108 | + if k not in ("alpha", "l1_ratio")} |
| 109 | + if has_l1_ratio: |
| 110 | + pen_kwargs['l1_ratio'] = l1 |
| 111 | + pen = type(self.penalty)(alpha=alpha, **pen_kwargs) |
| 112 | + |
| 113 | + est = GeneralizedLinearEstimator( |
| 114 | + datafit=self.datafit, penalty=pen, solver=self.solver |
| 115 | + ) |
| 116 | + if w_init is not None: |
| 117 | + est.coef_ = w_init[0] |
| 118 | + est.intercept_ = w_init[1] |
| 119 | + est.fit(X[train], y[train]) |
| 120 | + y_pred = est.predict(X[test]) |
| 121 | + return est.coef_, est.intercept_, self._score(y[test], y_pred) |
| 122 | + |
| 123 | + for idx_ratio, l1_ratio in enumerate(l1_ratios): |
| 124 | + warm_start = [None] * self.cv |
| 125 | + |
| 126 | + for idx_alpha, alpha in enumerate(alphas): |
| 127 | + if isinstance(self.datafit, (Logistic, QuadraticSVC)): |
| 128 | + kf = StratifiedKFold(n_splits=self.cv, shuffle=True, |
| 129 | + random_state=self.random_state) |
| 130 | + split_iter = kf.split(np.arange(n_samples), y) |
| 131 | + else: |
| 132 | + kf = KFold(n_splits=self.cv, shuffle=True, |
| 133 | + random_state=self.random_state) |
| 134 | + split_iter = kf.split(np.arange(n_samples)) |
| 135 | + fold_result = Parallel(self.n_jobs)( |
| 136 | + delayed(_solve_fold)(k, tr, te, alpha, l1_ratio, warm_start[k]) |
| 137 | + for k, (tr, te) in enumerate(split_iter) |
| 138 | + ) |
| 139 | + |
| 140 | + for k, (coef_fold, intercept_fold, loss_fold) in enumerate(fold_result): |
| 141 | + warm_start[k] = (coef_fold, intercept_fold) |
| 142 | + scores_path[idx_ratio, idx_alpha, k] = loss_fold |
| 143 | + |
| 144 | + mean_loss = np.mean(scores_path[idx_ratio, idx_alpha]) |
| 145 | + if mean_loss > best_loss: |
| 146 | + best_loss = mean_loss |
| 147 | + self.alpha_ = float(alpha) |
| 148 | + self.l1_ratio_ = float(l1_ratio) if has_l1_ratio else None |
| 149 | + |
| 150 | + # Refit on full dataset |
| 151 | + pen_kwargs = {k: v for k, v in self.penalty.__dict__.items() |
| 152 | + if k not in ("alpha", "l1_ratio")} |
| 153 | + if has_l1_ratio: |
| 154 | + pen_kwargs["l1_ratio"] = self.l1_ratio_ |
| 155 | + best_penalty = type(self.penalty)( |
| 156 | + alpha=self.alpha_, **pen_kwargs |
| 157 | + ) |
| 158 | + best_estimator = GeneralizedLinearEstimator( |
| 159 | + datafit=self.datafit, |
| 160 | + penalty=best_penalty, |
| 161 | + solver=self.solver |
| 162 | + ) |
| 163 | + best_estimator.fit(X, y) |
| 164 | + self.best_estimator_ = best_estimator |
| 165 | + self.coef_ = best_estimator.coef_ |
| 166 | + self.intercept_ = best_estimator.intercept_ |
| 167 | + self.n_iter_ = getattr(best_estimator, "n_iter_", None) |
| 168 | + self.n_features_in_ = getattr(best_estimator, "n_features_in_", None) |
| 169 | + self.feature_names_in_ = getattr(best_estimator, "feature_names_in_", None) |
| 170 | + self.alphas_ = alphas |
| 171 | + self.scores_path_ = np.squeeze(scores_path) |
| 172 | + return self |
| 173 | + |
| 174 | + def predict(self, X): |
| 175 | + return self.best_estimator_.predict(X) |
| 176 | + |
| 177 | + def predict_proba(self, X): |
| 178 | + return self.best_estimator_.predict_proba(X) |
| 179 | + |
| 180 | + def score(self, X, y): |
| 181 | + return self.best_estimator_.score(X, y) |
0 commit comments