Fix a bug that training is stuck while detection model is trained on distrubited environment

src/otx/core/data/dataset/detection.py

@@ -49,8 +49,9 @@ def _get_item_impl(self, index: int) -> DetDataEntity | None:
                 bboxes,
                 format=tv_tensors.BoundingBoxFormat.XYXY,
                 canvas_size=img_shape,
+                dtype=torch.float32,
             ),
-            labels=torch.as_tensor([ann.label for ann in bbox_anns]),
+            labels=torch.as_tensor([ann.label for ann in bbox_anns], dtype=torch.long),
         )
 
         return self._apply_transforms(entity)

src/otx/core/data/dataset/instance_segmentation.py

@@ -75,6 +75,7 @@ def _get_item_impl(self, index: int) -> InstanceSegDataEntity | None:
                 bboxes,
                 format=tv_tensors.BoundingBoxFormat.XYXY,
                 canvas_size=img_shape,
+                dtype=torch.float32,
             ),
             masks=tv_tensors.Mask(masks, dtype=torch.uint8),
             labels=torch.as_tensor(labels),

src/otx/core/metrics/fmeasure.py

@@ -636,6 +636,8 @@ class FMeasure(Metric):
     IoU > threshold are reduced to one. This threshold can be determined automatically by setting `vary_nms_threshold`
     to True.
 
+    # TODO(someone): need to update for distriubted training. refer https://lightning.ai/docs/torchmetrics/stable/pages/implement.html
+
     Args:
         label_info (int): Dataclass including label information.
         vary_nms_threshold (bool): if True the maximal F-measure is determined by optimizing for different NMS threshold

src/otx/core/model/detection.py

@@ -297,7 +297,7 @@ def _convert_pred_entity_to_compute_metric(
             "preds": [
                 {
                     "boxes": bboxes.data,
-                    "scores": scores,
+                    "scores": scores.type(torch.float32),
                     "labels": labels,
                 }
                 for bboxes, scores, labels in zip(