Move _q_tau from AADForest to AADEvaluator

q_tau is Hinge-loss specific value

Author: matwey

src/coniferest/aadforest.py

@@ -14,8 +14,20 @@
 class AADEvaluator(ConiferestEvaluator):
     def __init__(self, aad):
         super(AADEvaluator, self).__init__(aad, map_value=aad.map_value)
+        self.budget = aad.budget
         self.weights = np.full(shape=(self.n_leaves,), fill_value=np.reciprocal(np.sqrt(self.n_leaves)))
 
+    def _q_tau(self, scores):
+        if isinstance(self.budget, int):
+            if self.budget >= len(scores):
+                return np.max(scores)
+
+            return np.partition(scores, self.budget)[self.budget]
+        elif isinstance(self.budget, float):
+            return np.quantile(scores, self.budget)
+
+        raise ValueError("self.budget must be an int or float")
+
     def score_samples(self, x, weights=None):
         """
         Perform the computations.
@@ -250,17 +262,6 @@ def _build_trees(self, data):
             self.trees = self.build_trees(data, self.n_trees)
             self.evaluator = AADEvaluator(self)
 
-    def _q_tau(self, scores):
-        if isinstance(self.budget, int):
-            if self.budget >= len(scores):
-                return np.max(scores)
-
-            return np.partition(scores, self.budget)[self.budget]
-        elif isinstance(self.budget, float):
-            return np.quantile(scores, self.budget)
-
-        raise ValueError("self.budget must be an int or float")
-
     def fit(self, data, labels=None):
         """
         Build the trees with the data `data`.
@@ -324,7 +325,7 @@ def fit_known(self, data, known_data=None, known_labels=None):
             return self
 
         scores = self.score_samples(data)
-        q_tau = self._q_tau(scores)
+        q_tau = self.evaluator._q_tau(scores)
 
         anomaly_count = np.count_nonzero(known_labels == Label.ANOMALY)
         nominal_count = np.count_nonzero(known_labels == Label.REGULAR)

tests/test_aadforest.py

@@ -113,7 +113,7 @@ def test_benchmark_loss_gradient(n_samples, n_trees, n_jobs, benchmark):
     anomaly_count = np.count_nonzero(known_labels == Label.ANOMALY)
     nominal_count = np.count_nonzero(known_labels == Label.REGULAR)
     scores = forest.score_samples(data)
-    q_tau = forest._q_tau(scores)
+    q_tau = forest.evaluator._q_tau(scores)
 
     benchmark(
         forest.evaluator.loss_gradient,