fix: check when detectron threshold is none

nmn/evaluate/program_executor.py

@@ -9,7 +9,11 @@ import torch.nn as nn
 
 from nmn.modules.executors.filter import FilterExecutor
 from nmn.modules.executors.query import QueryExecutor
-from nmn.modules.executors.relate import ExtremeRelateExecutor, ImmediateRelateExecutor, RelateExecutor
+from nmn.modules.executors.relate import (
+    ExtremeRelateExecutor,
+    ImmediateRelateExecutor,
+    RelateExecutor,
+)
 from nmn.modules.executors.same import Same
 from nmn.modules.executors.segment import SegmentExecutor
 from nmn.modules.symbolic.compare import Equal, GreaterThan, LessThan

nmn/modules/executors/segment.py

@@ -30,28 +30,31 @@ class SegmentExecutor(nn.Module):
     def filter_doubles(
         self, detected_instances: list[torch.Tensor]
     ) -> list[torch.Tensor]:
-
-        jaccards = {}
-        for object_i in range(len(detected_instances)):
-            for object_j in range(object_i + 1, len(detected_instances)):
-                iou = self.jaccard(
-                    detected_instances[object_i], detected_instances[object_j]
-                )[1].item()
-                jaccards[(object_i, object_j)] = iou
-
-        largest_obj_pair = max(jaccards, key=jaccards.get)
-        largest_obj_iou = jaccards[largest_obj_pair]
-        largest_obj = largest_obj_pair[0]
-
-        if self.threshold and largest_obj_iou < self.threshold:
+        if not self.threshold:
             return detected_instances
         else:
-            logging.info(f" ---> filtered: {largest_obj_pair} from {jaccards}")
-            return [
-                detected_instances[i]
-                for i in range(len(detected_instances))
-                if i != largest_obj
-            ]
+            jaccards = {}
+            for object_i in range(len(detected_instances)):
+                for object_j in range(object_i + 1, len(detected_instances)):
+                    iou = self.jaccard(
+                        detected_instances[object_i], detected_instances[object_j]
+                    )[1].item()
+                    jaccards[(object_i, object_j)] = iou
+
+            largest_obj_pair = max(jaccards, key=jaccards.get)
+            largest_obj_iou = jaccards[largest_obj_pair]
+            largest_obj = largest_obj_pair[0]
+
+            if largest_obj_iou < self.threshold:
+                # jaccard metric does not exceed threshold, thus do not filter!
+                return detected_instances
+            else:
+                logging.info(f" ---> filtered: {largest_obj_pair} from {jaccards}")
+                return [
+                    detected_instances[i]
+                    for i in range(len(detected_instances))
+                    if i != largest_obj
+                ]
 
     def forward(self, image) -> AttentionSet: