Explicit cast to uint8 when bool inputs passed to argsort in MAP

CHANGELOG.md

@@ -43,6 +43,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
+- Fixed "Sort currently does not support bool dtype on CUDA" error in MAP for empty preds ([#983](https://github.com/PyTorchLightning/metrics/pull/983))
+
+
 - Fixed `BinnedPrecisionRecallCurve` when `thresholds` argument is not provided ([#968](https://github.com/PyTorchLightning/metrics/pull/968))
 
 

tests/detection/test_map.py

@@ -133,6 +133,25 @@
     ],
 )
 
+# Test empty preds case, to ensure bool inputs are properly casted to uint8
+# From https://github.com/PyTorchLightning/metrics/issues/981
+_inputs3 = Input(
+    preds=[
+        [
+            dict(boxes=torch.tensor([]), scores=torch.tensor([]), labels=torch.tensor([])),
+        ],
+    ],
+    target=[
+        [
+            dict(
+                boxes=torch.tensor([[1.0, 2.0, 3.0, 4.0]]),
+                scores=torch.tensor([0.8]),
+                labels=torch.tensor([1]),
+            ),
+        ],
+    ],
+)
+
 
 def _compare_fn(preds, target) -> dict:
     """Comparison function for map implementation.
@@ -283,7 +302,7 @@ def _move_to_gpu(input):
 
 @pytest.mark.skipif(_pytest_condition, reason="test requires that torchvision=>0.8.0 is installed")
 @pytest.mark.skipif(_gpu_test_condition, reason="test requires CUDA availability")
-@pytest.mark.parametrize("inputs", [_inputs, _inputs2])
+@pytest.mark.parametrize("inputs", [_inputs, _inputs2, _inputs3])
 def test_map_gpu(inputs):
     """Test predictions on single gpu."""
     metric = MeanAveragePrecision()

torchmetrics/detection/mean_ap.py

@@ -694,7 +694,10 @@ def __calculate_recall_precision_scores(
 
         # different sorting method generates slightly different results.
         # mergesort is used to be consistent as Matlab implementation.
-        inds = torch.argsort(det_scores, descending=True)
+        # Sort in PyTorch does not support bool types on CUDA (yet, 1.11.0)
+        dtype = torch.uint8 if det_scores.is_cuda and det_scores.dtype is torch.bool else det_scores.dtype
+        # Explicitly cast to uint8 to avoid error for bool inputs on CUDA to argsort
+        inds = torch.argsort(det_scores.to(dtype), descending=True)
         det_scores_sorted = det_scores[inds]
 
         det_matches = torch.cat([e["dtMatches"][:, :max_det] for e in img_eval_cls_bbox], axis=1)[:, inds]