Merge pull request #201 from mandjevant/master

mandjevant · web-flow · commit 4bc286479677 · 2023-12-20T10:18:04.000+01:00
Sklearn 1.3 update
diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
@@ -10,7 +10,7 @@ jobs:
     strategy:
       matrix:
         os: [ubuntu-latest, windows-latest]
-        python-version: ['3.7', '3.8', '3.9']
+        python-version: ['3.9', '3.10', '3.11']
 
     steps:
     - uses: actions/checkout@v2
diff --git a/README.md b/README.md
@@ -5,9 +5,6 @@
 [scikit-learn](http://scikit-learn.org/).
 
 See how to use hyperopt-sklearn through [examples](http://hyperopt.github.io/hyperopt-sklearn/#documentation)
-or older
-[notebooks](http://nbviewer.ipython.org/github/hyperopt/hyperopt-sklearn/tree/master/notebooks)
-
 More examples can be found in the Example Usage section of the SciPy paper
 
 Komer B., Bergstra J., and Eliasmith C. "Hyperopt-Sklearn: automatic hyperparameter configuration for Scikit-learn" Proc. SciPy 2014. http://conference.scipy.org/proceedings/scipy2014/pdfs/komer.pdf
diff --git a/hpsklearn/components/__init__.py b/hpsklearn/components/__init__.py
@@ -189,7 +189,7 @@ def any_classifier(name):
         random_forest_classifier(name + ".random_forest"),
         extra_tree_classifier(name + ".extra_trees"),
         ada_boost_classifier(name + ".ada_boost"),
-        gradient_boosting_classifier(name + ".grad_boosting", loss="deviance"),
+        gradient_boosting_classifier(name + ".grad_boosting"),
         sgd_classifier(name + ".sgd")
     ]
 
@@ -207,7 +207,7 @@ def any_sparse_classifier(name):
     sparse_classifiers = [
         linear_svc(name + ".linear_svc"),
         sgd_classifier(name + ".sgd"),
-        k_neighbors_classifier(name + ".knn", metric="euclidean", p=2),
+        k_neighbors_classifier(name + ".knn", p=2),
         multinomial_nb(name + ".multinomial_nb")
     ]
 
@@ -242,7 +242,7 @@ def any_sparse_regressor(name):
     """
     sparse_regressors = [
         sgd_regressor(name + ".sgd"),
-        k_neighbors_regressor(name + ".knn", metric="euclidean", p=2)
+        k_neighbors_regressor(name + ".knn", p=2)
     ]
 
     return hp.choice(name, sparse_regressors)
diff --git a/hpsklearn/components/cluster/_kmeans.py b/hpsklearn/components/cluster/_kmeans.py
@@ -61,8 +61,8 @@ def _kmeans_hp_space(
 
 
 @validate(params=["algorithm"],
-          validation_test=lambda param: not isinstance(param, str) or param in ["auto", "full", "elkan"],
-          msg="Invalid parameter '%s' with value '%s'. Value must be 'auto', 'full' or 'elkan'")
+          validation_test=lambda param: not isinstance(param, str) or param in ["lloyd", "elkan"],
+          msg="Invalid parameter '%s' with value '%s'. Value must be 'lloyd' or 'elkan'")
 def k_means(name: str,
             n_init: typing.Union[int, Apply] = None,
             max_iter: typing.Union[int, Apply] = None,
@@ -94,7 +94,7 @@ def _name(msg):
     hp_space["max_iter"] = scope.int(hp.uniform(_name("max_iter"), 100, 500)) if max_iter is None else max_iter
     hp_space["tol"] = hp.uniform(_name("tol"), 1e-5, 1e-3) if tol is None else tol
     hp_space["copy_x"] = copy_x
-    hp_space["algorithm"] = hp.choice(_name("algorithm"), ["auto", "full", "elkan"]) if algorithm is None else algorithm
+    hp_space["algorithm"] = hp.choice(_name("algorithm"), ["lloyd", "elkan"]) if algorithm is None else algorithm
 
     return scope.sklearn_KMeans(**hp_space)
 
diff --git a/hpsklearn/components/ensemble/_bagging.py b/hpsklearn/components/ensemble/_bagging.py
@@ -82,7 +82,7 @@ def _bagging_random_state(name: str):
           msg="Invalid parameter '%s' with value '%s'. Parameter value must exceed 1.")
 def _bagging_hp_space(
         name_func,
-        base_estimator=None,
+        estimator=None,
         n_estimators: typing.Union[int, Apply] = None,
         max_samples: typing.Union[float, Apply] = None,
         max_features: typing.Union[float, Apply] = None,
@@ -100,7 +100,7 @@ def _bagging_hp_space(
      bagging regressor
     """
     hp_space = dict(
-        base_estimator=base_estimator,
+        estimator=estimator,
         n_estimators=_bagging_n_estimators(name_func("n_estimators")) if n_estimators is None else n_estimators,
         max_samples=_bagging_max_samples(name_func("max_samples")) if max_samples is None else max_samples,
         max_features=_bagging_max_features(name_func("max_features")) if max_features is None else max_features,
diff --git a/hpsklearn/components/ensemble/_gb.py b/hpsklearn/components/ensemble/_gb.py
@@ -23,7 +23,7 @@ def _gb_clf_loss(name: str):
     """
     Declaration search space 'loss' parameter for _gb classifier
     """
-    return hp.choice(name, ["deviance", "exponential"])
+    return hp.choice(name, ["log_loss", "exponential"])
 
 
 def _gb_reg_loss(name: str):
@@ -211,16 +211,16 @@ def _gb_hp_space(
 
 
 @validate(params=["loss"],
-          validation_test=lambda param: not isinstance(param, str) or param in ("deviance", "exponential"),
-          msg="Invalid parameter '%s' with value '%s'. Choose 'deviance' or 'exponential'.")
+          validation_test=lambda param: not isinstance(param, str) or param in ("log_loss", "exponential"),
+          msg="Invalid parameter '%s' with value '%s'. Choose 'log_loss' or 'exponential'.")
 def gradient_boosting_classifier(name: str, loss: typing.Union[str, Apply] = None, **kwargs):
     """
     Return a pyll graph with hyperparameters that will construct
     a sklearn.ensemble.GradientBoostingClassifier model.
 
     Args:
         name: name | str
-        loss: choose 'deviance' or 'exponential' | str
+        loss: choose 'log_loss' or 'exponential' | str
 
     See help(hpsklearn.components._gb._gb_hp_space) for info on
     additional available GradientBoosting arguments.
diff --git a/hpsklearn/components/ensemble/_hist_gradient_boosting.py b/hpsklearn/components/ensemble/_hist_gradient_boosting.py
@@ -26,7 +26,10 @@ def _hist_gradient_boosting_reg_loss(name: str):
      hist gradient boosting regressor
 
     Parameter 'poisson' is also available. Not implemented since
-     'poisson' is only available for non-negative y data
+     'poisson' is only available for non-zero, non-negative y data
+
+    Parameter 'gamma' is also available. Not implemented since
+     'gamma' is only available for non-negative y data
     """
     return hp.choice(name, ["squared_error", "absolute_error"])
 
@@ -141,14 +144,14 @@ def _hist_gradient_boosting_hp_space(
                                                                                 "categorical_crossentropy"),
           msg="Invalid parameter '%s' with value '%s'. "
               "Choose 'auto', 'binary_crossentropy', 'categorical_crossentropy'")
-def hist_gradient_boosting_classifier(name: str, loss: typing.Union[str, Apply] = "auto", **kwargs):
+def hist_gradient_boosting_classifier(name: str, loss: typing.Union[str, Apply] = "log_loss", **kwargs):
     """
     Return a pyll graph with hyperparameters that will construct
     a sklearn.ensemble.HistGradientBoostingClassifier model.
 
     Args:
         name: name | str
-        loss: choose 'auto', 'binary_crossentropy' or 'categorical_crossentropy' | str
+        loss: 'log_loss' | str
 
     See help(hpsklearn.components._hist_gradient_boosting._hist_gradient_boosting_regressor) for info on
     additional available HistGradientBoosting arguments.
@@ -165,7 +168,7 @@ def _name(msg):
 
 @validate(params=["loss"],
           validation_test=lambda param: not isinstance(param, str) or param in ("squared_error", "absolute_error",
-                                                                                "poisson"),
+                                                                                "poisson", "quantile", "gamma"),
           msg="Invalid parameter '%s' with value '%s'. "
               "Choose 'squared_error', 'absolute_error', 'poisson'")
 def hist_gradient_boosting_regressor(name: str, loss: typing.Union[str, Apply] = None, **kwargs):
diff --git a/hpsklearn/components/ensemble/_weight_boosting.py b/hpsklearn/components/ensemble/_weight_boosting.py
@@ -58,7 +58,7 @@ def _weight_boosting_random_state(name: str):
           msg="Invalid parameter '%s' with value '%s'. Parameter value must be non-negative and greater than 0.")
 def _weight_boosting_hp_space(
         name_func,
-        base_estimator=None,
+        estimator=None,
         n_estimators: typing.Union[int, Apply] = None,
         learning_rate: typing.Union[float, Apply] = None,
         random_state=None
@@ -69,7 +69,7 @@ def _weight_boosting_hp_space(
      AdaBoost regressor
     """
     hp_space = dict(
-        base_estimator=base_estimator,
+        estimator=estimator,
         n_estimators=_weight_boosting_n_estimators(name_func("n_estimators")) if n_estimators is None else n_estimators,
         learning_rate=_weight_boosting_learning_rate(name_func("learning_rate"))
         if learning_rate is None else learning_rate,
diff --git a/hpsklearn/components/lightgbm.py b/hpsklearn/components/lightgbm.py
@@ -141,12 +141,14 @@ def _lightgbm_hp_space(
      lightgbm regressor
     """
     hp_space = dict(
-        max_depth=_lightgbm_max_depth(name_func("max_depth")) if max_depth is None else max_depth,
+        # max_depth=_lightgbm_max_depth(name_func("max_depth")) if max_depth is None else max_depth,
+        max_depth=-1,
         num_leaves=_lightgbm_num_leaves(name_func("num_leaves")) if num_leaves is None else num_leaves,
         learning_rate=_lightgbm_learning_rate(name_func("learning_rate")) if learning_rate is None else learning_rate,
         n_estimators=_lightgbm_n_estimators(name_func("n_estimators")) if n_estimators is None else n_estimators,
         min_child_weight=_lightgbm_min_child_weight(name_func("min_child_weight"))
         if min_child_weight is None else min_child_weight,
+        # min_child_samples=5,
         max_delta_step=max_delta_step,
         subsample=_lightgbm_subsample(name_func("subsample")) if subsample is None else subsample,
         colsample_bytree=_lightgbm_colsample_bytree(name_func("colsample_bytree"))
diff --git a/hpsklearn/components/linear_model/_bayes.py b/hpsklearn/components/linear_model/_bayes.py
@@ -18,9 +18,9 @@ def sklearn_ARDRegression(*args, **kwargs):
     return linear_model.ARDRegression(*args, **kwargs)
 
 
-def _bayes_n_iter(name: str):
+def _bayes_max_iter(name: str):
     """
-    Declaration search space 'n_iter' parameter
+    Declaration search space 'max_iter' parameter
     """
     return scope.int(hp.qloguniform(name, low=np.log(150), high=np.log(450), q=1.0))
 
@@ -40,15 +40,15 @@ def _bayes_alpha_lambda(name: str):
     return hp.lognormal(name, mu=np.log(1e-6), sigma=np.log(10))
 
 
-@validate(params=["n_iter"],
+@validate(params=["max_iter"],
           validation_test=lambda param: not isinstance(param, int) or param > 1,
           msg="Invalid parameter '%s' with value '%s'. Parameter value must exceed 1.")
 @validate(params=["alpha_1", "alpha_2", "lambda_1", "lambda_2"],
           validation_test=lambda param: not isinstance(param, float) or param >= 0,
           msg="Invalid parameter '%s' with value '%s'. Parameter value must be equal to or exceed 0.")
 def _bayes_hp_space(
         name_func,
-        n_iter: typing.Union[int, Apply] = None,
+        max_iter: typing.Union[int, Apply] = None,
         tol: typing.Union[float, Apply] = None,
         alpha_1: typing.Union[float, Apply] = None,
         alpha_2: typing.Union[float, Apply] = None,
@@ -65,7 +65,7 @@ def _bayes_hp_space(
      ard regression
     """
     hp_space = dict(
-        n_iter=_bayes_n_iter(name_func("n_iter")) if n_iter is None else n_iter,
+        max_iter=_bayes_max_iter(name_func("max_iter")) if max_iter is None else max_iter,
         tol=_bayes_tol(name_func("tol")) if tol is None else tol,
         alpha_1=_bayes_alpha_lambda(name_func("alpha_1")) if alpha_1 is None else alpha_1,
         alpha_2=_bayes_alpha_lambda(name_func("alpha_2")) if alpha_2 is None else alpha_2,
diff --git a/hpsklearn/components/linear_model/_least_angle.py b/hpsklearn/components/linear_model/_least_angle.py
@@ -38,7 +38,7 @@ def _least_angle_n_nonzero_coefs(name: str):
     """
     Declaration search space 'n_nonzero_coefs' parameter
     """
-    return hp.qloguniform(name, low=np.log(400), high=np.log(600), q=1.0)
+    return scope.int(hp.qloguniform(name, low=np.log(400), high=np.log(600), q=1.0))
 
 
 def _least_angle_alpha(name: str):
@@ -80,7 +80,7 @@ def _least_angle_max_n_alphas(name: str):
     """
     Declaration search space 'max_n_alphas' parameter
     """
-    return hp.loguniform(name, low=np.log(750), high=np.log(1250))
+    return scope.int(hp.loguniform(name, low=np.log(750), high=np.log(1250)))
 
 
 def _least_angle_random_state(name: str):
diff --git a/hpsklearn/components/linear_model/_logistic.py b/hpsklearn/components/linear_model/_logistic.py
@@ -43,7 +43,7 @@ def _logistic_max_iter(name: str):
     """
     Declaration search space 'max_iter' parameter
     """
-    return scope.int(hp.uniform(name, 250, 750))
+    return scope.int(hp.uniform(name, 500, 1000))
 
 
 def _logistic_C(name: str):
diff --git a/hpsklearn/components/linear_model/_quantile.py b/hpsklearn/components/linear_model/_quantile.py
@@ -14,10 +14,9 @@ def sklearn_QuantileRegressor(*args, **kwargs):
 
 @validate(params=["solver"],
           validation_test=lambda param: not isinstance(param, str) or
-                                        param in ["highs-ds", "highs-ipm", "highs",  # noqa
-                                                  "interior-point", "revised simplex"],
+          param in ["highs-ds", "highs-ipm", "highs", "revised simplex"],
           msg="Invalid parameter '%s' with value '%s'. Value must be in ['highs-ds', 'highs-ipm', 'highs', "
-              "'interior-point', 'revised simplex'].")
+              "'revised simplex'].")
 def quantile_regression(name: str,
                         quantile: typing.Union[float, Apply] = None,
                         alpha: typing.Union[float, Apply] = None,
@@ -44,7 +43,7 @@ def _name(msg):
         quantile=hp.normal(_name("quantile"), 0.5, 0.075) if quantile is None else quantile,
         alpha=hp.normal(_name("alpha"), mu=1.0, sigma=0.1) if alpha is None else alpha,
         fit_intercept=hp.choice(_name("fit_intercept"), [True, False]) if fit_intercept is None else fit_intercept,
-        solver=hp.choice(_name("solver"), ["highs-ds", "highs-ipm", "highs", "interior-point", "revised simplex"])
+        solver=hp.choice(_name("solver"), ["highs-ds", "highs-ipm", "highs", "revised simplex"])
         if solver is None else solver,
         solver_options=solver_options
     )
diff --git a/hpsklearn/components/linear_model/_ransac.py b/hpsklearn/components/linear_model/_ransac.py
@@ -17,7 +17,7 @@ def sklearn_RANSACRegressor(*args, **kwargs):
           validation_test=lambda param: not isinstance(param, str) or param in ["absolute_error", "squared_error"],
           msg="Invalid parameter '%s' with value '%s'. Value must be in ['absolute_error', 'squared_error'].")
 def ransac_regression(name: str,
-                      base_estimator=None,
+                      estimator=None,
                       min_samples: float = None,
                       residual_threshold: float = None,
                       is_data_valid: callable = None,
@@ -35,7 +35,7 @@ def ransac_regression(name: str,
 
     Args:
         name: name | str
-        base_estimator: base estimator object
+        estimator: base estimator object
         min_samples: minimum number of samples chosen | float
         residual_threshold: maximum residual | float
         is_data_valid: function called before model is fitted | callable
@@ -53,12 +53,12 @@ def _name(msg):
         return f"{name}.ransac_regression_{msg}"
 
     hp_space = dict(
-        base_estimator=base_estimator,
+        estimator=estimator,
         min_samples=min_samples,  # default None fits linear model with X.shape[1] + 1
         residual_threshold=residual_threshold,
         is_data_valid=is_data_valid,
         is_model_valid=is_model_valid,
-        max_trials=hp.uniform(_name("max_trials"), 50, 150) if max_trials is None else max_trials,
+        max_trials=scope.int(hp.uniform(_name("max_trials"), 50, 150)) if max_trials is None else max_trials,
         max_skips=np.inf if max_skips is None else max_skips,
         stop_n_inliers=np.inf if stop_n_inliers is None else stop_n_inliers,
         stop_score=np.inf if stop_score is None else stop_score,
diff --git a/hpsklearn/components/linear_model/_stochastic_gradient.py b/hpsklearn/components/linear_model/_stochastic_gradient.py
@@ -30,7 +30,7 @@ def _stochastic_gradient_classifier_loss(name: str):
     """
     return hp.pchoice(name, [
         (0.25, "hinge"),
-        (0.25, "log"),
+        (0.25, "log_loss"),
         (0.25, "modified_huber"),
         (0.05, "squared_hinge"),
         (0.05, "perceptron"),
@@ -218,10 +218,10 @@ def _stochastic_gradient_hp_space(
 
 @validate(params=["loss"],
           validation_test=lambda param: not isinstance(param, str) or
-                                        param in ["hinge", "log", "modified_huber", "squared_hinge", "perceptron",   # noqa
+                                        param in ["hinge", "log_loss", "modified_huber", "squared_hinge", "perceptron",   # noqa
                                                   "squared_error", "huber", "epsilon_insensitive",
                                                   "squared_epsilon_insensitive"],
-          msg="Invalid parameter '%s' with value '%s'. Value must be in ['hinge', 'log', 'modified_huber', "
+          msg="Invalid parameter '%s' with value '%s'. Value must be in ['hinge', 'log_loss', 'modified_huber', "
               "'squared_hinge', 'perceptron', 'squared_error', 'huber', 'epsilon_insensitive', "
               "'squared_epsilon_insensitive'].")
 @validate(params=["class_weight"],
diff --git a/hpsklearn/components/neighbors/_nearest_centroid.py b/hpsklearn/components/neighbors/_nearest_centroid.py
@@ -11,7 +11,7 @@ def sklearn_NearestCentroid(*args, **kwargs):
 
 
 def nearest_centroid(name: str,
-                     metric: typing.Union[str, callable, Apply] = None,
+                     metric: typing.Union[str, Apply] = None,
                      shrink_threshold: float = None):
     """
     Return a pyll graph with hyperparameters that will construct
@@ -27,7 +27,7 @@ def _name(msg):
         return f"{name}.nearest_centroid_{msg}"
 
     hp_space = dict(
-        metric=hp.choice(_name("metric"), ["cityblock", "cosine", "l1", "l2", "minkowski", "euclidean", "manhattan"])
+        metric=hp.choice(_name("metric"), ["euclidean", "manhattan"])
         if metric is None else metric,
         shrink_threshold=shrink_threshold,
     )
diff --git a/hpsklearn/components/preprocessing/_discretization.py b/hpsklearn/components/preprocessing/_discretization.py
@@ -26,6 +26,7 @@ def k_bins_discretizer(name: str,
                        n_bins: typing.Union[int, npt.ArrayLike, Apply] = None,
                        encode: typing.Union[str, Apply] = None,
                        strategy: typing.Union[str, Apply] = None,
+                       subsample: typing.Union[int, None, Apply] = None,
                        dtype=None):
     """
     Return a pyll graph with hyperparameters that will construct
@@ -36,12 +37,14 @@ def k_bins_discretizer(name: str,
         n_bins: number of bins | int, npt.ArrayLike
         encode: encoding method | str
         strategy: strategy used to define width of bins | str
+        subsample: subsample size of training data | int, None
         dtype: dtype of output | type
     """
     rval = scope.sklearn_KBinsDiscretizer(
         n_bins=scope.int(hp.uniform(name + ".n_bins", 2, 20)) if n_bins is None else n_bins,
         encode=hp.choice(name + ".encode", ["onehot-dense", "ordinal"]) if encode is None else encode,
         strategy=hp.choice(name + ".strategy", ["uniform", "quantile", "kmeans"]) if strategy is None else strategy,
+        subsample=hp.choice(name + ".subsample", [200000, None] if subsample is None else subsample),
         dtype=dtype
     )
 
diff --git a/hpsklearn/components/preprocessing/_encoders.py b/hpsklearn/components/preprocessing/_encoders.py
@@ -27,7 +27,7 @@ def sklearn_OrdinalEncoder(*args, **kwargs):
 def one_hot_encoder(name: str,
                     categories: typing.Union[str, list] = "auto",
                     drop: typing.Union[str, np.ndarray, Apply] = None,
-                    sparse: bool = True,
+                    sparse_output: bool = True,
                     dtype: type = np.float64):
     """
     Return a pyll graph with hyperparameters that will construct
@@ -37,13 +37,13 @@ def one_hot_encoder(name: str,
         name: name | str
         categories: categories per feature | str or list
         drop: choose 'first' or 'if_binary' | str or np.ndarray
-        sparse: return sparse matrix or array | bool
+        sparse_output: return sparse_output matrix or array | bool
         dtype: desired dtype of output | type
     """
     rval = scope.sklearn_OneHotEncoder(
         categories=categories,
         drop=hp.choice(name + ".drop", ["first", "if_binary"]) if drop is None else drop,
-        sparse=sparse,
+        sparse_output=sparse_output,
         dtype=dtype
     )
 
diff --git a/hpsklearn/components/svm/_classes.py b/hpsklearn/components/svm/_classes.py
diff --git a/pyproject.toml b/pyproject.toml
diff --git a/requirements.txt b/requirements.txt
diff --git a/requirements_dev.txt b/requirements_dev.txt
diff --git a/setup.cfg b/setup.cfg
diff --git a/tests/test_components/test_lagselectors.py b/tests/test_components/test_lagselectors.py
diff --git a/tests/test_estimator/test_estimator.py b/tests/test_estimator/test_estimator.py
diff --git a/tox.ini b/tox.ini
