Fix XAI algorithm for Detection (#2609)

* Impove saliency maps algorithm for Detection

* Remove extra changes

* Update unit tests

* Changes for 1 class

* Fix pre-commit

* Update CHANGELOG

CHANGELOG.md

@@ -16,6 +16,7 @@ All notable changes to this project will be documented in this file.
 - Fix mmcls bug not wrapping model in DataParallel on CPUs (<https://github.com/openvinotoolkit/training_extensions/pull/2601>)
 - Fix h-label loss normalization issue w/ exclusive label group of singe label (<https://github.com/openvinotoolkit/training_extensions/pull/2604>)
 - Fix division by zero in class incremental learning for classification (<https://github.com/openvinotoolkit/training_extensions/pull/2606>)
+- Fix saliency maps calculation issue for detection models (<https://github.com/openvinotoolkit/training_extensions/pull/2609>)
 
 ## \[v1.4.3\]
 

src/otx/algorithms/detection/adapters/mmdet/hooks/det_class_probability_map_hook.py

@@ -60,12 +60,9 @@ def func(
         else:
             cls_scores = self._get_cls_scores_from_feature_map(feature_map)
 
-        # Don't use softmax for tiles in tiling detection, if the tile doesn't contain objects,
-        # it would highlight one of the class maps as a background class
-        if self.use_cls_softmax and self._num_cls_out_channels > 1:
-            cls_scores = [torch.softmax(t, dim=1) for t in cls_scores]
-
-        batch_size, _, height, width = cls_scores[-1].size()
+        middle_idx = len(cls_scores) // 2
+        # resize to the middle feature map
+        batch_size, _, height, width = cls_scores[middle_idx].size()
         saliency_maps = torch.empty(batch_size, self._num_cls_out_channels, height, width)
         for batch_idx in range(batch_size):
             cls_scores_anchorless = []
@@ -82,6 +79,11 @@ def func(
                 )
             saliency_maps[batch_idx] = torch.cat(cls_scores_anchorless_resized, dim=0).mean(dim=0)
 
+        # Don't use softmax for tiles in tiling detection, if the tile doesn't contain objects,
+        # it would highlight one of the class maps as a background class
+        if self.use_cls_softmax:
+            saliency_maps[0] = torch.stack([torch.softmax(t, dim=1) for t in saliency_maps[0]])
+
         if self._norm_saliency_maps:
             saliency_maps = saliency_maps.reshape((batch_size, self._num_cls_out_channels, -1))
             saliency_maps = self._normalize_map(saliency_maps)

tests/unit/algorithms/detection/test_xai_detection_validity.py

@@ -24,19 +24,19 @@
 
 class TestExplainMethods:
     ref_saliency_shapes = {
-        "MobileNetV2-ATSS": (2, 4, 4),
+        "MobileNetV2-ATSS": (2, 13, 13),
         "SSD": (81, 13, 13),
-        "YOLOX": (80, 13, 13),
+        "YOLOX": (80, 26, 26),
     }
 
     ref_saliency_vals_det = {
-        "MobileNetV2-ATSS": np.array([67, 216, 255, 57], dtype=np.uint8),
-        "YOLOX": np.array([80, 28, 42, 53, 49, 68, 72, 75, 69, 57, 65, 6, 157], dtype=np.uint8),
-        "SSD": np.array([119, 72, 118, 35, 39, 30, 31, 31, 36, 28, 44, 23, 61], dtype=np.uint8),
+        "MobileNetV2-ATSS": np.array([34, 67, 148, 132, 172, 147, 146, 155, 167, 159], dtype=np.uint8),
+        "YOLOX": np.array([177, 94, 147, 147, 161, 162, 164, 164, 163, 166], dtype=np.uint8),
+        "SSD": np.array([255, 178, 212, 90, 93, 79, 79, 80, 87, 83], dtype=np.uint8),
     }
 
     ref_saliency_vals_det_wo_postprocess = {
-        "MobileNetV2-ATSS": -0.10465062,
+        "MobileNetV2-ATSS": -0.014513552,
         "YOLOX": 0.04948914,
         "SSD": 0.6629989,
     }
@@ -80,8 +80,8 @@ def test_saliency_map_det(self, template):
         assert len(saliency_maps) == 2
         assert saliency_maps[0].ndim == 3
         assert saliency_maps[0].shape == self.ref_saliency_shapes[template.name]
-        actual_sal_vals = saliency_maps[0][0][0].astype(np.int8)
-        ref_sal_vals = self.ref_saliency_vals_det[template.name].astype(np.int8)
+        actual_sal_vals = saliency_maps[0][0][0][:10].astype(np.int16)
+        ref_sal_vals = self.ref_saliency_vals_det[template.name].astype(np.uint8)
         assert np.all(np.abs(actual_sal_vals - ref_sal_vals) <= 1)
 
     @e2e_pytest_unit