Outlier detection: keep instance ids, make thread safe

Outlier detection did not keep instance ids, so subsets did not work.
Also, it was not thread safe: multiple calls to _OutlierModel.__call__
could result in undefined behaviour, because some caching was done at
the object level.

Author: markotoplak

Orange/classification/outlier_detection.py

@@ -9,43 +9,52 @@
 from sklearn.svm import OneClassSVM
 
 from Orange.base import SklLearner, SklModel
-from Orange.data import Table, Domain, DiscreteVariable, ContinuousVariable, \
-    Variable
-from Orange.data.util import get_unique_names
+from Orange.data import Table, Domain, DiscreteVariable, ContinuousVariable
+from Orange.data.util import get_unique_names, SharedComputeValue
 from Orange.preprocess import AdaptiveNormalize
-from Orange.util import wrap_callback, dummy_callback
+from Orange.util import dummy_callback
 
 __all__ = ["LocalOutlierFactorLearner", "IsolationForestLearner",
            "EllipticEnvelopeLearner", "OneClassSVMLearner"]
 
 
+class _CachedTransform:
+    # to be used with SharedComputeValue
+    def __init__(self, model):
+        self.model = model
+
+    def __call__(self, data):
+        return self.model.data_to_model_domain(data)
+
+
 class _OutlierModel(SklModel):
     def __init__(self, skl_model):
         super().__init__(skl_model)
         self._cached_data = None
         self.outlier_var = None
+        self.cached_transform = _CachedTransform(self)
 
     def predict(self, X: np.ndarray) -> np.ndarray:
         pred = self.skl_model.predict(X)
         pred[pred == -1] = 0
         return pred[:, None]
 
+    def new_domain(self, data: Table) -> Domain:
+        assert self.outlier_var is not None
+        return Domain(data.domain.attributes, data.domain.class_vars,
+               data.domain.metas + (self.outlier_var,))
+
     def __call__(self, data: Table, progress_callback: Callable = None) \
             -> Table:
         assert isinstance(data, Table)
-        assert self.outlier_var is not None
 
-        domain = Domain(data.domain.attributes, data.domain.class_vars,
-                        data.domain.metas + (self.outlier_var,))
+        domain = self.new_domain(data)
         if progress_callback is None:
             progress_callback = dummy_callback
-        progress_callback(0, "Preprocessing...")
-        self._cached_data = self.data_to_model_domain(
-            data, wrap_callback(progress_callback, end=0.1))
-        progress_callback(0.1, "Predicting...")
-        metas = np.hstack((data.metas, self.predict(self._cached_data.X)))
+        progress_callback(0, "Predicting...")
+        new_table = data.transform(domain)
         progress_callback(1)
-        return Table.from_numpy(domain, data.X, data.Y, metas)
+        return new_table
 
 
 class _OutlierLearner(SklLearner):
@@ -64,27 +73,17 @@ def _fit_model(self, data: Table) -> _OutlierModel:
             compute_value=transformer
         )
 
-        transformer.variable = variable
         model.outlier_var = variable
         return model
 
 
-class _Transformer:
+class _Transformer(SharedComputeValue):
     def __init__(self, model: _OutlierModel):
+        super().__init__(model.cached_transform)
         self._model = model
-        self._variable = None
-
-    @property
-    def variable(self) -> Variable:
-        return self._variable
 
-    @variable.setter
-    def variable(self, var: Variable):
-        self._variable = var
-
-    def __call__(self, data: Table) -> np.ndarray:
-        assert isinstance(self._variable, Variable)
-        return self._model(data).get_column_view(self._variable)[0]
+    def compute(self, data: Table, shared_data: Table) -> np.ndarray:
+        return self._model.predict(shared_data.X)[:, 0]
 
 
 class OneClassSVMLearner(_OutlierLearner):
@@ -142,13 +141,16 @@ def mahalanobis(self, observations: np.ndarray) -> np.ndarray:
         """
         return self.skl_model.mahalanobis(observations)[:, None]
 
-    def __call__(self, data: Table, progress_callback: Callable = None) \
-            -> Table:
-        pred = super().__call__(data, progress_callback)
-        domain = Domain(pred.domain.attributes, pred.domain.class_vars,
-                        pred.domain.metas + (self.mahal_var,))
-        metas = np.hstack((pred.metas, self.mahalanobis(self._cached_data.X)))
-        return Table.from_numpy(domain, pred.X, pred.Y, metas)
+    def new_domain(self, data: Table) -> Domain:
+        assert self.mahal_var is not None
+        domain = super().new_domain(data)
+        return Domain(domain.attributes, domain.class_vars,
+                        domain.metas + (self.mahal_var,))
+
+
+class _TransformerMahalanobis(_Transformer):
+    def compute(self, data: Table, shared_data: Table) -> np.ndarray:
+        return self._model.mahalanobis(shared_data.X)[:, 0]
 
 
 class EllipticEnvelopeLearner(_OutlierLearner):
@@ -166,13 +168,12 @@ def _fit_model(self, data: Table) -> EllipticEnvelopeClassifier:
         domain = data.domain
         model = super()._fit_model(data.transform(Domain(domain.attributes)))
 
-        transformer = _Transformer(model)
+        transformer = _TransformerMahalanobis(model)
         names = [v.name for v in domain.variables + domain.metas]
         variable = ContinuousVariable(
             get_unique_names(names, "Mahalanobis"),
             compute_value=transformer
         )
 
-        transformer.variable = variable
         model.mahal_var = variable
         return model

Orange/classification/tests/test_outlier_detection.py

@@ -3,7 +3,7 @@
 import pickle
 import tempfile
 import unittest
-from unittest.mock import Mock
+from unittest.mock import Mock, patch
 
 import numpy as np
 
@@ -36,6 +36,7 @@ def assert_table_equal(self, table1, table2):
         np.testing.assert_array_equal(table1.metas, table2.metas)
 
     def assert_table_appended_outlier(self, table1, table2, offset=1):
+        np.testing.assert_array_equal(table1.ids, table2.ids)
         np.testing.assert_array_equal(table1.X, table2.X)
         np.testing.assert_array_equal(table1.Y, table2.Y)
         np.testing.assert_array_equal(table1.metas, table2.metas[:, :-offset])
@@ -47,7 +48,6 @@ def assert_table_appended_outlier(self, table1, table2, offset=1):
         self.assertEqual(table2.domain.metas[-offset].name, "Outlier")
         self.assertIsNotNone(table2.domain.metas[-offset].compute_value)
 
-
 class TestOneClassSVMLearner(_TestDetector):
     def test_OneClassSVM(self):
         np.random.seed(42)
@@ -128,12 +128,19 @@ def test_EllipticEnvelope(self):
     def test_mahalanobis(self):
         n = len(self.X_all)
         pred = self.model(self.X_all)
+
         y_pred = pred[:, self.model.outlier_var].metas
         y_mahal = pred[:, self.model.mahal_var].metas
         y_mahal, y_pred = zip(*sorted(zip(y_mahal, y_pred), reverse=True))
         self.assertTrue(all(i == 0 for i in y_pred[:int(self.cont * n)]))
         self.assertTrue(all(i == 1 for i in y_pred[int(self.cont * n):]))
 
+    def test_single_data_to_model_domain(self):
+        with patch.object(self.model, "data_to_model_domain",
+                          wraps=self.model.data_to_model_domain) as call:
+            self.model(self.X_all)
+            self.assertEqual(call.call_count, 1)
+
     def test_EllipticEnvelope_ignores_y(self):
         domain = Domain((ContinuousVariable("x1"), ContinuousVariable("x2")),
                         (ContinuousVariable("y1"), ContinuousVariable("y2")))
@@ -231,7 +238,6 @@ def test_transformer(self):
         detect = self.detector(self.iris)
         pred = detect(self.iris)
         var = pred.domain.metas[0]
-        self.assertIs(var, var.compute_value.variable)
         np.testing.assert_array_equal(pred[:, "Outlier"].metas.ravel(),
                                       var.compute_value(self.iris))