compositional-explanations 0.0.3.3__tar.gz → 0.0.4__tar.gz

{compositional_explanations-0.0.3.3/src/compositional_explanations.egg-info → compositional_explanations-0.0.4}/PKG-INFO

@@ -1,11 +1,12 @@
 Metadata-Version: 2.1
 Name: compositional_explanations
-Version: 0.0.3.3
+Version: 0.0.4
 Summary: Package to compute Compositional Explanations both in their optimal and beam format.
 Author-email: Biagio La Rosa <bilarosa@ucsc.edu>
 License: MIT
 Project-URL: Homepage, https://github.com/aiea-lab/optimal-compositional-explanations
 Project-URL: Issues, https://github.com/aiea-lab/optimal-compositional-explanations/issues
+Platform: UNKNOWN
 Classifier: Programming Language :: Python :: 3
 Classifier: Operating System :: OS Independent
 Requires-Python: >=3.8

{compositional_explanations-0.0.3.3 → compositional_explanations-0.0.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "compositional_explanations"
-version = "0.0.3.3"
+version = "0.0.4"
 authors = [
   { name="Biagio La Rosa", email="bilarosa@ucsc.edu" },
 ]

{compositional_explanations-0.0.3.3 → compositional_explanations-0.0.4}/src/compositional_explanations/optimal.py

@@ -121,6 +121,14 @@ def apply_distributive_property(node):
         # considering the transformations since computing ((A AND NOT C) OR (B AND NOT C)) is more accurate than
         # computing ((A OR B) AND NOT C) since everythin is a pariwise estimation.
         return label
+    if external_op == F.Or:
+        # We do not rewrite (A AND B) OR C as (A OR C) AND (B OR C).
+        # Otherwise, it would put C on both sides of an AND.
+        # Our code assumes that AND removes always the unique elements,
+        # since it aassumes the concepts cannot be repeated.
+        # Therefore, it would underestimate the denominator of the IoU,
+        # leading to a lower upper bound.
+        return label
     label_left = label.left
     num_left_op = len(set(label_left.get_ops()))
     if num_left_op == 1:

{compositional_explanations-0.0.3.3 → compositional_explanations-0.0.4}/src/compositional_explanations/optimal_sample_heuristic.py

@@ -368,8 +368,8 @@ def estimate_label_info(label, left_quantities, right_quantities, neuron_quantit
             # OR can't reduce the extras
             return None
         
-        left_can_be_improved = np.where(min_left_common_intersection < neuron_common, True, False) & np.where(min_right_common_intersection > 0, True, False)
-        right_can_be_improved = np.where(min_right_common_intersection < neuron_common, True, False) & np.where(min_left_common_intersection > 0, True, False)
+        left_can_be_improved = np.where(min_left_common_intersection < neuron_common, True, False) & np.where(max_right_common_intersection > 0, True, False)
+        right_can_be_improved = np.where(min_right_common_intersection < neuron_common, True, False) & np.where(max_left_common_intersection > 0, True, False)
         if (right_unique_intersection_sum == 0 and not left_can_be_improved.any()) or (left_unique_intersection_sum == 0 and not right_can_be_improved.any()):
             # If for a sample the min left intersection is already maximum, the right component cannot improve the score
             # Similarly, if the minimum common intersection of the right element is 0, it cannot improve the score for that sample

{compositional_explanations-0.0.3.3 → compositional_explanations-0.0.4/src/compositional_explanations.egg-info}/PKG-INFO

@@ -1,11 +1,12 @@
 Metadata-Version: 2.1
 Name: compositional_explanations
-Version: 0.0.3.3
+Version: 0.0.4
 Summary: Package to compute Compositional Explanations both in their optimal and beam format.
 Author-email: Biagio La Rosa <bilarosa@ucsc.edu>
 License: MIT
 Project-URL: Homepage, https://github.com/aiea-lab/optimal-compositional-explanations
 Project-URL: Issues, https://github.com/aiea-lab/optimal-compositional-explanations/issues
+Platform: UNKNOWN
 Classifier: Programming Language :: Python :: 3
 Classifier: Operating System :: OS Independent
 Requires-Python: >=3.8

{compositional_explanations-0.0.3.3 → compositional_explanations-0.0.4}/src/package_utils/mask_utils.py

@@ -21,10 +21,9 @@ def get_dataset_info(masks, info_directory):
     file_path = f"{info_directory}/quantities.pkl"
     if info_directory is not None and os.path.exists(file_path):
         with open(file_path, "rb") as file:
-            print("Loading dataset info from cache...")
             common_elements, unique_elements, uncoverable_elements = pickle.load(file)
     else:
-        print("Computing dataset info...")
+        print("Dataset info not found in cache. Computing it. This process may take a while.")
         # Compute unique elements
         sum_elements = torch.zeros_like(
             general_utils.parse_mask_by_type(masks[1]), dtype=torch.int32