[FIX] Add stability to explain detection (#1901)

Add stability to explain detection

Author: sungmanc

otx/algorithms/detection/adapters/mmdet/models/heads/custom_atss_head.py

@@ -172,18 +172,16 @@ def loss_single(
             pos_centerness = centerness[pos_inds]
 
             centerness_targets = self.centerness_target(pos_anchors, pos_bbox_targets)
-            pos_decode_bbox_pred = self.bbox_coder.decode(pos_anchors, pos_bbox_pred)
-            pos_decode_bbox_targets = self.bbox_coder.decode(pos_anchors, pos_bbox_targets)
+            if self.reg_decoded_bbox:
+                pos_bbox_pred = self.bbox_coder.decode(pos_anchors, pos_bbox_pred)
 
             if self.use_qfl:
-                quality[pos_inds] = bbox_overlaps(
-                    pos_decode_bbox_pred.detach(), pos_decode_bbox_targets, is_aligned=True
-                ).clamp(min=1e-6)
+                quality[pos_inds] = bbox_overlaps(pos_bbox_pred.detach(), pos_bbox_targets, is_aligned=True).clamp(
+                    min=1e-6
+                )
 
             # regression loss
-            loss_bbox = self.loss_bbox(
-                pos_decode_bbox_pred, pos_decode_bbox_targets, weight=centerness_targets, avg_factor=1.0
-            )
+            loss_bbox = self.loss_bbox(pos_bbox_pred, pos_bbox_targets, weight=centerness_targets, avg_factor=1.0)
 
             # centerness loss
             loss_centerness = self.loss_centerness(pos_centerness, centerness_targets, avg_factor=num_total_samples)

otx/mpa/det/explainer.py

@@ -2,6 +2,7 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
+import torch
 from mmcv.utils import Config, ConfigDict
 from mmdet.datasets import build_dataloader as mmdet_build_dataloader
 from mmdet.datasets import build_dataset as mmdet_build_dataset
@@ -153,7 +154,8 @@ def explain(self, cfg, model_builder=None):
         eval_predictions = []
         with self.explainer_hook(feature_model) as saliency_hook:
             for data in test_dataloader:
-                result = model(return_loss=False, rescale=True, **data)
+                with torch.no_grad():
+                    result = model(return_loss=False, rescale=True, **data)
                 eval_predictions.extend(result)
             saliency_maps = saliency_hook.records