SMAC3 Bayesian Optimisation Integration [🆕 Search Method] 🥉

.gitignore

@@ -10,3 +10,4 @@ docs/source/generated/*
 *.log
 *.ipynb_checkpoints*
 venv/
+smac3_output/*

gama/GamaClassifier.py

@@ -1,41 +1,135 @@
 import inspect
 from typing import Union, Optional
+import logging
 
 import numpy as np
 import pandas as pd
+from ConfigSpace import ForbiddenEqualsClause
 from sklearn.base import ClassifierMixin
 from sklearn.preprocessing import LabelEncoder
+import ConfigSpace as cs
 
-from .gama import Gama
+from gama.configuration.classification import config_space as clf_config
 from gama.data_loading import X_y_from_file
-from gama.configuration.classification import clf_config
-from gama.utilities.metrics import scoring_to_metric
+from gama.utilities.metrics import scoring_to_metric, Metric
+from .gama import Gama
+from .utilities.config_space import get_estimator_by_name
+
+# Avoid stopit from logging warnings every time a pipeline evaluation times out
+logging.getLogger("stopit").setLevel(logging.ERROR)
+log = logging.getLogger(__name__)
 
 
 class GamaClassifier(Gama):
     """Gama with adaptations for (multi-class) classification."""
 
-    def __init__(self, search_space=None, scoring="neg_log_loss", *args, **kwargs):
+    def __init__(
+        self,
+        search_space: Optional[cs.ConfigurationSpace] = None,
+        scoring: Metric = "neg_log_loss",  # type: ignore
+        *args,
+        **kwargs,
+    ):
         if not search_space:
             # Do this to avoid the whole dictionary being included in the documentation.
             search_space = clf_config
 
         self._metrics = scoring_to_metric(scoring)
+
+        search_space = self._search_space_check(search_space)
+
+        self._label_encoder = None
+        super().__init__(
+            *args, search_space=search_space, scoring=scoring, **kwargs
+        )  # type: ignore
+
+    def _search_space_check(
+        self,
+        search_space: cs.ConfigurationSpace,
+    ) -> cs.ConfigurationSpace:
+        """Check if the search space is valid for classification."""
+
+        # Check if the search space contains a classifier hyperparameter.
+        if (
+            "estimators" not in search_space.meta
+            or (
+                search_space.meta["estimators"]
+                not in search_space.get_hyperparameters_dict()
+            )
+            or not isinstance(
+                search_space.get_hyperparameter(search_space.meta["estimators"]),
+                cs.CategoricalHyperparameter,
+            )
+        ):
+            raise ValueError(
+                "The search space must include a hyperparameter for the classifiers "
+                "that is a CategoricalHyperparameter with choices for all desired "
+                "classifiers. Please double-check the spelling of the name, and review "
+                "the `meta` object in the search space configuration located at "
+                "`configurations/classification.py`. The `meta` object should contain "
+                "a key `estimators` with a value that is the name of the hyperparameter"
+                " that contains the classifier choices."
+            )
+
+        # Check if the search space contains a preprocessor hyperparameter
+        # if it is specified in the meta.
+        if (
+            "preprocessors" in search_space.meta
+            and (
+                search_space.meta["preprocessors"]
+                not in search_space.get_hyperparameters_dict()
+            )
+            or "preprocessors" in search_space.meta
+            and not isinstance(
+                search_space.get_hyperparameter(search_space.meta["preprocessors"]),
+                cs.CategoricalHyperparameter,
+            )
+        ):
+            raise ValueError(
+                "The search space must include a hyperparameter for the preprocessors "
+                "that is a CategoricalHyperparameter with choices for all desired "
+                "preprocessors. Please double-check the spelling of the name, and "
+                "review the `meta` object in the search space configuration located at "
+                "`configurations/classification.py`. The `meta` object should contain "
+                "a key `preprocessors` with a value that is the name of the "
+                "hyperparameter that contains the preprocessor choices. "
+            )
+
+        # Check if the search space contains only classifiers that have predict_proba
+        # if the scoring requires probabilities.
         if any(metric.requires_probabilities for metric in self._metrics):
             # we don't want classifiers that do not have `predict_proba`,
             # because then we have to start doing one hot encodings of predictions etc.
-            search_space = {
-                alg: hp
-                for (alg, hp) in search_space.items()
-                if not (
-                    inspect.isclass(alg)
-                    and issubclass(alg, ClassifierMixin)
-                    and not hasattr(alg(), "predict_proba")
-                )
-            }
-
-        self._label_encoder = None
-        super().__init__(*args, search_space=search_space, scoring=scoring, **kwargs)
+            no_proba_clfs = []
+            for classifier in search_space.get_hyperparameter(
+                search_space.meta["estimators"]
+            ).choices:
+                estimator = get_estimator_by_name(classifier)
+                if (
+                    estimator is not None
+                    and issubclass(estimator, ClassifierMixin)
+                    and not hasattr(estimator(), "predict_proba")
+                ):
+                    no_proba_clfs.append(classifier)
+
+            log.info(
+                f"The following classifiers do not have a predict_proba method "
+                f"and will be excluded from the search space: {no_proba_clfs}"
+            )
+            search_space.add_forbidden_clauses(
+                [
+                    ForbiddenEqualsClause(
+                        search_space.get_hyperparameter(
+                            search_space.meta["estimators"]
+                        ),
+                        classifier,
+                    )
+                    for classifier in no_proba_clfs
+                    if classifier
+                ]
+            )
+
+        return search_space
 
     def _predict(self, x: pd.DataFrame):
         """Predict the target for input X.
@@ -52,8 +146,8 @@ def _predict(self, x: pd.DataFrame):
         """
         y = self.model.predict(x)  # type: ignore
         # Decode the predicted labels - necessary only if ensemble is not used.
-        if y[0] not in list(self._label_encoder.classes_):
-            y = self._label_encoder.inverse_transform(y)
+        if y[0] not in list(self._label_encoder.classes_):  # type: ignore
+            y = self._label_encoder.inverse_transform(y)  # type: ignore
         return y
 
     def _predict_proba(self, x: pd.DataFrame):

gama/GamaRegressor.py

@@ -1,7 +1,8 @@
 import pandas as pd
 
 from .gama import Gama
-from gama.configuration.regression import reg_config
+from gama.configuration.regression import config_space as reg_config
+import ConfigSpace as cs
 
 
 class GamaRegressor(Gama):
@@ -16,8 +17,64 @@ def __init__(
 
         if not search_space:
             search_space = reg_config
+
+        search_space = self._search_space_check(search_space)
+
         super().__init__(*args, search_space=search_space, scoring=scoring, **kwargs)
 
+    def _search_space_check(
+        self, search_space: cs.ConfigurationSpace
+    ) -> cs.ConfigurationSpace:
+        """Check if the search space is valid for regression."""
+
+        # Check if the search space contains a regressor hyperparameter.
+        if (
+            "estimators" not in search_space.meta
+            or (
+                search_space.meta["estimators"]
+                not in search_space.get_hyperparameters_dict()
+            )
+            or not isinstance(
+                search_space.get_hyperparameter(search_space.meta["estimators"]),
+                cs.CategoricalHyperparameter,
+            )
+        ):
+            raise ValueError(
+                "The search space must include a hyperparameter for the regressors "
+                "that is a CategoricalHyperparameter with choices for all desired "
+                "regressors. Please double-check the spelling of the name, and review "
+                "the `meta` object in the search space configuration located at "
+                "`configurations/regression.py`. The `meta` object should contain "
+                "a key `estimators` with a value that is the name of the hyperparameter"
+                " that contains the regressor choices."
+            )
+
+        # Check if the search space contains a preprocessor hyperparameter
+        # if it is specified in the meta.
+        if (
+            "preprocessors" in search_space.meta
+            and (
+                search_space.meta["preprocessors"]
+                not in search_space.get_hyperparameters_dict()
+            )
+            or "preprocessors" in search_space.meta
+            and not isinstance(
+                search_space.get_hyperparameter(search_space.meta["preprocessors"]),
+                cs.CategoricalHyperparameter,
+            )
+        ):
+            raise ValueError(
+                "The search space must include a hyperparameter for the preprocessors "
+                "that is a CategoricalHyperparameter with choices for all desired "
+                "preprocessors. Please double-check the spelling of the name, and "
+                "review the `meta` object in the search space configuration located at "
+                "`configurations/regression.py`. The `meta` object should contain "
+                "a key `preprocessors` with a value that is the name of the "
+                "hyperparameter that contains the preprocessor choices. "
+            )
+
+        return search_space
+
     def _predict(self, x: pd.DataFrame):
         """Predict the target for input X.