Merge pull request #104 from automl/development

Release 1.0.6

Author: shukon

pimp/__version__.py

@@ -1,2 +1,2 @@
 """Version information."""
-__version__ = "1.0.5"
+__version__ = "1.0.6"

pimp/epm/base_epm.py

@@ -4,6 +4,12 @@
 from pyrfr import regression
 
 from smac.epm.base_epm import AbstractEPM
+from smac.configspace import (
+    CategoricalHyperparameter,
+    UniformFloatHyperparameter,
+    UniformIntegerHyperparameter,
+    Constant,
+)
 
 
 class RandomForestWithInstances(AbstractEPM):
@@ -28,8 +34,11 @@ class RandomForestWithInstances(AbstractEPM):
     logger : logging.logger
     """
 
-    def __init__(self, types: np.ndarray,
+    def __init__(self,
+                 configspace,
+                 types: np.ndarray,
                  bounds: np.ndarray,
+                 seed: int,
                  num_trees: int = 10,
                  do_bootstrapping: bool = True,
                  n_points_per_tree: int = -1,
@@ -39,13 +48,14 @@ def __init__(self, types: np.ndarray,
                  max_depth: int = 20,
                  eps_purity: int = 1e-8,
                  max_num_nodes: int = 2 ** 20,
-                 seed: int = 42,
                  logged_y: bool = True,
                  **kwargs):
         """Constructor
 
         Parameters
         ----------
+        configspace: ConfigurationSpace
+            configspace to be passed to random forest (used to impute inactive parameter-values)
         types : np.ndarray (D)
             Specifies the number of categorical values of an input dimension where
             the i-th entry corresponds to the i-th input dimension. Let's say we
@@ -54,6 +64,8 @@ def __init__(self, types: np.ndarray,
             have to pass np.array([2, 0]). Note that we count starting from 0.
         bounds : np.ndarray (D, 2)
             Specifies the bounds for continuous features.
+        seed : int
+            The seed that is passed to the random_forest_run library.
         num_trees : int
             The number of trees in the random forest.
         do_bootstrapping : bool
@@ -74,17 +86,20 @@ def __init__(self, types: np.ndarray,
             different
         max_num_nodes : int
             The maxmimum total number of nodes in a tree
-        seed : int
-            The seed that is passed to the random_forest_run library.
         logged_y: bool
             Indicates if the y data is transformed (i.e. put on logscale) or not
         """
         try:
-            super().__init__(types, bounds, **kwargs)
+            super().__init__(configspace=configspace, types=types, bounds=bounds, seed=seed, **kwargs)
         except TypeError:
-            # To ensure backwards-compatibility with smac<0.9.0
-            super().__init__(**kwargs)
-
+            try:
+                # To ensure backwards-compatibility with smac==0.10.0
+                super().__init__(types, bounds, **kwargs)
+            except TypeError:
+                # To ensure backwards-compatibility with smac<0.9.0
+                super().__init__(**kwargs)
+
+        self.configspace = configspace
         self.types = types
         self.bounds = bounds
         self.rng = regression.default_random_engine(seed)
@@ -112,9 +127,33 @@ def __init__(self, types: np.ndarray,
                        min_samples_leaf, max_depth, eps_purity, seed]
         self.seed = seed
 
+        self.impute_values = {}
+
         self.logger = logging.getLogger(self.__module__ + "." +
                                         self.__class__.__name__)
 
+    def _impute_inactive(self, X: np.ndarray) -> np.ndarray:
+        X = X.copy()
+        for idx, hp in enumerate(self.configspace.get_hyperparameters()):
+            if idx not in self.impute_values:
+                parents = self.configspace.get_parents_of(hp.name)
+                if len(parents) == 0:
+                    self.impute_values[idx] = None
+                else:
+                    if isinstance(hp, CategoricalHyperparameter):
+                        self.impute_values[idx] = len(hp.choices)
+                    elif isinstance(hp, (UniformFloatHyperparameter, UniformIntegerHyperparameter)):
+                        self.impute_values[idx] = -1
+                    elif isinstance(hp, Constant):
+                        self.impute_values[idx] = 1
+                    else:
+                        raise ValueError
+
+            nonfinite_mask = ~np.isfinite(X[:, idx])
+            X[nonfinite_mask, idx] = self.impute_values[idx]
+
+        return X
+
     def _train(self, X: np.ndarray, y: np.ndarray, **kwargs):
         """Trains the random forest on X and y.
 
@@ -130,9 +169,8 @@ def _train(self, X: np.ndarray, y: np.ndarray, **kwargs):
         self
         """
 
-        self.X = X
+        self.X = self._impute_inactive(X)
         self.y = y.flatten()
-
         if self.n_points_per_tree <= 0:
             self.rf_opts.num_data_points_per_tree = self.X.shape[0]
         else:
@@ -141,6 +179,7 @@ def _train(self, X: np.ndarray, y: np.ndarray, **kwargs):
         self.rf.options = self.rf_opts
         data = self.__init_data_container(self.X, self.y)
         self.rf.fit(data, rng=self.rng)
+
         return self
 
     def __init_data_container(self, X: np.ndarray, y: np.ndarray):
@@ -200,6 +239,9 @@ def _predict(self, X: np.ndarray):
                              (self.types.shape[0], X.shape[1]))
 
         means, vars_ = [], []
+
+        X = self._impute_inactive(X)
+
         for row_X in X:
             mean, var = self.rf.predict_mean_var(row_X)
             means.append(mean)

pimp/epm/epar_x_rfwi.py

@@ -12,7 +12,7 @@
 
 class EPARrfi(rfi):
 
-    def __init__(self, types, bounds,
+    def __init__(self, configspace, types, bounds, seed,
                  cutoff=0,
                  threshold=0, **kwargs):
         """
@@ -21,6 +21,8 @@ def __init__(self, types, bounds,
 
         Parameters
         ----------
+        configspace: ConfigurationSpace
+            configspace to be passed to random forest (used to impute inactive parameter-values)
         types: np.ndarray (D)
             Specifies the number of categorical values of an input dimension. Where
             the i-th entry corresponds to the i-th input dimension. Let say we have
@@ -29,6 +31,8 @@ def __init__(self, types, bounds,
             have to pass np.array([2, 0]). Note that we count starting from 0.
         bounds: np.ndarray (D)
             Specifies the bounds
+        seed: int
+            The seed that is passed to the random_forest_run library.
         instance_features: np.ndarray (I, K)
             Contains the K dimensional instance features
             of the I different instances
@@ -48,16 +52,13 @@ def __init__(self, types, bounds,
 
         max_num_nodes: int
 
-        seed: int
-            The seed that is passed to the random_forest_run library.
-
         cutoff: int
             The cutoff used in the specified scenario
 
         threshold:
             Maximal possible value
         """
-        super().__init__(types=types, bounds=bounds, **kwargs)
+        super().__init__(configspace=configspace, types=types, bounds=bounds, seed=seed, **kwargs)
         np.seterr(divide='ignore', invalid='ignore')
         self.cutoff = cutoff
         self.threshold = threshold

pimp/epm/unlogged_epar_x_rfwi.py

@@ -10,14 +10,14 @@
 
 class UnloggedEPARXrfi(EPARrfi, Urfi):
 
-    def __init__(self, types, bounds,
+    def __init__(self, configspace, types, bounds, seed,
                  cutoff=0,
                  threshold=0, **kwargs):
         """
         TODO
         """
-        Urfi.__init__(self, types=types, bounds=bounds, **kwargs)
-        EPARrfi.__init__(self, types=types, bounds=bounds, cutoff=cutoff, threshold=threshold, **kwargs)
+        Urfi.__init__(self, configspace=configspace, types=types, bounds=bounds, seed=seed, **kwargs)
+        EPARrfi.__init__(self, configspace=configspace, types=types, bounds=bounds, seed=seed, cutoff=cutoff, threshold=threshold, **kwargs)
 
     def predict(self, X):
         """

pimp/epm/unlogged_rfwi.py

@@ -12,13 +12,15 @@
 
 class Unloggedrfwi(rfi):
 
-    def __init__(self, types, bounds, **kwargs):
+    def __init__(self, configspace, types, bounds, seed, **kwargs):
         """
         Interface to the random forest that takes instance features
         into account.
 
         Parameters
         ----------
+        configspace: ConfigurationSpace
+            configspace to be passed to random forest (used to impute inactive parameter-values)
         types: np.ndarray (D)
             Specifies the number of categorical values of an input dimension. Where
             the i-th entry corresponds to the i-th input dimension. Let say we have
@@ -27,6 +29,9 @@ def __init__(self, types, bounds, **kwargs):
             have to pass np.array([2, 0]). Note that we count starting from 0.
         bounds: np.ndarray (D)
             Specifies the bounds
+        seed: int
+            The seed that is passed to the random_forest_run library.
+
         instance_features: np.ndarray (I, K)
             Contains the K dimensional instance features
             of the I different instances
@@ -46,16 +51,13 @@ def __init__(self, types, bounds, **kwargs):
 
         max_num_nodes: int
 
-        seed: int
-            The seed that is passed to the random_forest_run library.
-
         cutoff: int
             The cutoff used in the specified scenario
 
         threshold:
             Maximal possible value
         """
-        super().__init__(types=types, bounds=bounds, **kwargs)
+        super().__init__(configspace=configspace, types=types, bounds=bounds, seed=seed, **kwargs)
 
     # With the usage of pyrfr 0.8.0 this method is obsolete.
     # def _predict(self, X):

pimp/evaluator/base_evaluator.py

@@ -3,6 +3,7 @@
 from collections import OrderedDict
 
 from smac.epm.rf_with_instances import RandomForestWithInstances
+from ConfigSpace.hyperparameters import UniformFloatHyperparameter
 
 from pimp.configspace import ConfigurationSpace
 from pimp.utils import Scenario
@@ -102,6 +103,9 @@ def _refit_model(self, types, bounds, X, y):
         y:ndarray
             corresponding y vector
         """
-        self.model = RandomForestWithInstances(types, bounds, do_bootstrapping=True)
+        # We need to fake config-space bypass imputation of inactive values in random forest implementation
+        fake_cs = ConfigurationSpace(name="fake-cs-for-configurator-footprint")
+
+        self.model = RandomForestWithInstances(fake_cs, types, bounds, seed=12345, do_bootstrapping=True)
         self.model.rf_opts.compute_oob_error = True
         self.model.train(X, y)