MeanAveragePrecision: Skip box conversion if no boxes are present (#1097

) * Skip box conversion if no boxes are present The `box_convert` function from torchvision expects the input to be a Tensor[N, 4], where N > 0. Should N == 0 and in_fmt != out_fmt, `unbind` will error out on the boxes tensor during the conversion process. The workaround is, therefore to skip the box conversion if boxes is an empty tensor. Co-authored-by: Jirka Borovec <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Jirka <[email protected]>
Lightning-AI · Jun 20, 2022 · d29919c · d29919c
1 parent 203ab6b
commit d29919c
Show file tree

Hide file tree

Showing 3 changed files with 62 additions and 39 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -42,9 +42,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - Fixed mAP calculation for areas with 0 predictions ([#1080](https://github.com/PyTorchLightning/metrics/pull/1080))
 
+
 - Fixed bug where avg precision state and auroc state was not merge when using MetricCollections ([#1086](https://github.com/PyTorchLightning/metrics/pull/1086))
 
 
+- Skip box conversion if no boxes are present in `MeanAveragePrecision` ([#1097](https://github.com/PyTorchLightning/metrics/pull/1097))
+
+
 ## [0.9.1] - 2022-06-08
 
 ### Added

diff --git a/src/torchmetrics/detection/mean_ap.py b/src/torchmetrics/detection/mean_ap.py
@@ -416,7 +416,8 @@ def _get_safe_item_values(self, item: Dict[str, Any]) -> Union[Tensor, Tuple]:
 
         if self.iou_type == "bbox":
             boxes = _fix_empty_tensors(item["boxes"])
-            boxes = box_convert(boxes, in_fmt=self.box_format, out_fmt="xyxy")
+            if boxes.numel() > 0:
+                boxes = box_convert(boxes, in_fmt=self.box_format, out_fmt="xyxy")
             return boxes
         elif self.iou_type == "segm":
             masks = []

diff --git a/tests/unittests/detection/test_map.py b/tests/unittests/detection/test_map.py
@@ -56,7 +56,6 @@
     ],
 )
 
-
 _inputs = Input(
     preds=[
         [
@@ -299,7 +298,6 @@ class TestMAP(MetricTester):
 
     @pytest.mark.parametrize("ddp", [False, True])
     def test_map_bbox(self, compute_on_cpu, ddp):
-
         """Test modular implementation for correctness."""
         self.run_class_metric_test(
             ddp=ddp,
@@ -344,12 +342,8 @@ def test_empty_preds():
     metric = MeanAveragePrecision()
 
     metric.update(
-        [
-            dict(boxes=Tensor([]), scores=Tensor([]), labels=IntTensor([])),
-        ],
-        [
-            dict(boxes=Tensor([[214.1500, 41.2900, 562.4100, 285.0700]]), labels=IntTensor([4])),
-        ],
+        [dict(boxes=Tensor([]), scores=Tensor([]), labels=IntTensor([]))],
+        [dict(boxes=Tensor([[214.1500, 41.2900, 562.4100, 285.0700]]), labels=IntTensor([4]))],
     )
     metric.compute()
 
@@ -360,16 +354,56 @@ def test_empty_ground_truths():
     metric = MeanAveragePrecision()
 
     metric.update(
-        [
-            dict(
-                boxes=Tensor([[214.1500, 41.2900, 562.4100, 285.0700]]),
-                scores=Tensor([0.5]),
-                labels=IntTensor([4]),
-            ),
-        ],
-        [
-            dict(boxes=Tensor([]), labels=IntTensor([])),
-        ],
+        [dict(boxes=Tensor([[214.1500, 41.2900, 562.4100, 285.0700]]), scores=Tensor([0.5]), labels=IntTensor([4]))],
+        [dict(boxes=Tensor([]), labels=IntTensor([]))],
+    )
+    metric.compute()
+
+
+@pytest.mark.skipif(_pytest_condition, reason="test requires that torchvision=>0.8.0 is installed")
+def test_empty_ground_truths_xywh():
+    """Test empty ground truths in xywh format."""
+    metric = MeanAveragePrecision(box_format="xywh")
+
+    metric.update(
+        [dict(boxes=Tensor([[214.1500, 41.2900, 348.2600, 243.7800]]), scores=Tensor([0.5]), labels=IntTensor([4]))],
+        [dict(boxes=Tensor([]), labels=IntTensor([]))],
+    )
+    metric.compute()
+
+
+@pytest.mark.skipif(_pytest_condition, reason="test requires that torchvision=>0.8.0 is installed")
+def test_empty_preds_xywh():
+    """Test empty predictions in xywh format."""
+    metric = MeanAveragePrecision(box_format="xywh")
+
+    metric.update(
+        [dict(boxes=Tensor([]), scores=Tensor([]), labels=IntTensor([]))],
+        [dict(boxes=Tensor([[214.1500, 41.2900, 348.2600, 243.7800]]), labels=IntTensor([4]))],
+    )
+    metric.compute()
+
+
+@pytest.mark.skipif(_pytest_condition, reason="test requires that torchvision=>0.8.0 is installed")
+def test_empty_ground_truths_cxcywh():
+    """Test empty ground truths in cxcywh format."""
+    metric = MeanAveragePrecision(box_format="cxcywh")
+
+    metric.update(
+        [dict(boxes=Tensor([[388.2800, 163.1800, 348.2600, 243.7800]]), scores=Tensor([0.5]), labels=IntTensor([4]))],
+        [dict(boxes=Tensor([]), labels=IntTensor([]))],
+    )
+    metric.compute()
+
+
+@pytest.mark.skipif(_pytest_condition, reason="test requires that torchvision=>0.8.0 is installed")
+def test_empty_preds_cxcywh():
+    """Test empty predictions in cxcywh format."""
+    metric = MeanAveragePrecision(box_format="cxcywh")
+
+    metric.update(
+        [dict(boxes=Tensor([]), scores=Tensor([]), labels=IntTensor([]))],
+        [dict(boxes=Tensor([[388.2800, 163.1800, 348.2600, 243.7800]]), labels=IntTensor([4]))],
     )
     metric.compute()
 
@@ -467,16 +501,8 @@ def test_segm_iou_empty_gt_mask():
     metric = MeanAveragePrecision(iou_type="segm")
 
     metric.update(
-        [
-            dict(
-                masks=torch.randint(0, 1, (1, 10, 10)).bool(),
-                scores=Tensor([0.5]),
-                labels=IntTensor([4]),
-            ),
-        ],
-        [
-            dict(masks=Tensor([]), labels=IntTensor([])),
-        ],
+        [dict(masks=torch.randint(0, 1, (1, 10, 10)).bool(), scores=Tensor([0.5]), labels=IntTensor([4]))],
+        [dict(masks=Tensor([]), labels=IntTensor([]))],
     )
 
     metric.compute()
@@ -488,16 +514,8 @@ def test_segm_iou_empty_pred_mask():
     metric = MeanAveragePrecision(iou_type="segm")
 
     metric.update(
-        [
-            dict(
-                masks=torch.BoolTensor([]),
-                scores=Tensor([]),
-                labels=IntTensor([]),
-            ),
-        ],
-        [
-            dict(masks=torch.randint(0, 1, (1, 10, 10)).bool(), labels=IntTensor([4])),
-        ],
+        [dict(masks=torch.BoolTensor([]), scores=Tensor([]), labels=IntTensor([]))],
+        [dict(masks=torch.randint(0, 1, (1, 10, 10)).bool(), labels=IntTensor([4]))],
     )
 
     metric.compute()