Merge pull request #621 from dianna-ai/LIME_coefficients

Integrated LIME surrogate model scores segmentation filling as an option in LIME.

Author: WillemSpek

dianna/methods/lime.py

@@ -1,4 +1,5 @@
 import numpy as np
+from lime.lime_image import ImageExplanation
 from lime.lime_image import LimeImageExplainer
 from lime.lime_text import LimeTextExplainer
 from dianna import utils
@@ -164,6 +165,7 @@ def explain(self,
                 top_labels=None,
                 num_features=10,
                 num_samples=5000,
+                return_masks=True,
                 positive_only=False,
                 hide_rest=True,
                 **kwargs,
@@ -179,6 +181,7 @@ def explain(self,
             top_labels: Top labels
             num_features (int): Number of features
             num_samples (int): Number of samples
+            return_masks (bool): If true, return discretized masks. Otherwise, return LIME scores
             positive_only (bool): Positive only
             hide_rest (bool): Hide rest
             kwargs: These parameters are passed on
@@ -205,12 +208,14 @@ def explain(self,
                                                       num_samples=num_samples,
                                                       **explain_instance_kwargs,
                                                       )
-
-        get_image_and_mask_kwargs = utils.get_kwargs_applicable_to_function(explanation.get_image_and_mask, kwargs)
-        masks = [explanation.get_image_and_mask(label, positive_only=positive_only, hide_rest=hide_rest,
-                                                num_features=num_features, **get_image_and_mask_kwargs)[1]
-                 for label in labels]
-        return masks
+        if return_masks:
+            get_image_and_mask_kwargs = utils.get_kwargs_applicable_to_function(explanation.get_image_and_mask, kwargs)
+            maps = [explanation.get_image_and_mask(label, positive_only=positive_only, hide_rest=hide_rest,
+                                                   num_features=num_features, **get_image_and_mask_kwargs)[1]
+                    for label in labels]
+        else:
+            maps = [self._get_explanation_values(label, explanation) for label in labels]
+        return maps
 
     def _prepare_image_data(self, input_data):
         """Transforms the data to be of the shape and type LIME expects.
@@ -276,3 +281,27 @@ def moveaxis_function(data):
         if self.preprocess_function is None:
             return moveaxis_function
         return lambda data: self.preprocess_function(moveaxis_function(data))
+
+    def _get_explanation_values(self, label: int, explanation: ImageExplanation) -> np.array:
+        """Get the importance scores from LIME in a salience map.
+
+        Leverages the `ImageExplanation` class from LIME to generate salience maps.
+        These salience maps are constructed using the segmentation masks from
+        the explanation and fills these with the scores from the surrogate model
+        (default for LIME is Ridge regression) used for the explanation.
+
+        Args:
+            label: The class label for the given explanation
+            explanation: An Image Explanation generated by LIME
+
+        Returns:
+            A salience map containing the feature importances from LIME
+        """
+        class_explanation = explanation.local_exp[label]
+        salience_map = np.zeros(explanation.segments.shape,
+                                dtype=class_explanation[0][1].dtype) # Ensure same dataype for segment values
+
+        # Fill segments
+        for segment_id, segment_val in class_explanation:
+            salience_map[segment_id == explanation.segments] = segment_val
+        return salience_map

tests/test_data/heatmap_lime_values.npy

@@ -48,6 +48,23 @@ def test_lime_filename():
         assert heatmap[0].shape == input_data[0].shape
         assert np.allclose(heatmap, heatmap_expected, atol=1e-5)
 
+    @staticmethod
+    def test_lime_values():
+        """Test if get_explanation_values function works correctly."""
+        input_data = np.random.random((224, 224, 3))
+        heatmap_expected = np.load('tests/test_data/heatmap_lime_values.npy')
+        labels = [1]
+
+        explainer = LIMEImage(random_state=42)
+        heatmap = explainer.explain(run_model,
+                                    input_data,
+                                    labels,
+                                    return_masks=False,
+                                    num_samples=100)
+
+        assert heatmap[0].shape == input_data.shape[:2]
+        assert np.allclose(heatmap, heatmap_expected, atol=1e-5)
+
     def setUp(self) -> None:
         """Set seed."""
         np.random.seed(42)