MAP metric - fix empty predictions (#594)

* Apply suggestions from code review Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Jirka Borovec <[email protected]>
Lightning-AI · Oct 29, 2021 · 1c42f66 · 1c42f66
1 parent e44b51b
commit 1c42f66
Show file tree

Hide file tree

Showing 3 changed files with 23 additions and 4 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -23,16 +23,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
+- Fix empty predictions in MAP metric ([#594](https://github.com/PyTorchLightning/metrics/pull/594))
+
 
 ## [0.6.0] - 2021-10-28
 
 ### Added
 
 - Added audio metrics:
-  - Perceptual Evaluation of Speech Quality (PESQ) ([#353](https://github.com/PyTorchLightning/metrics/issues/353))
-  - Short Term Objective Intelligibility (STOI) ([#353](https://github.com/PyTorchLightning/metrics/issues/353))
+  - Perceptual Evaluation of Speech Quality (PESQ) ([#353](https://github.com/PyTorchLightning/metrics/pull/353))
+  - Short Term Objective Intelligibility (STOI) ([#353](https://github.com/PyTorchLightning/metrics/pull/353))
 - Added Information retrieval metrics:
-  - `RetrievalRPrecision` ([#577](https://github.com/PyTorchLightning/metrics/pull/577/))
+  - `RetrievalRPrecision` ([#577](https://github.com/PyTorchLightning/metrics/pull/577))
   - `RetrievalHitRate` ([#576](https://github.com/PyTorchLightning/metrics/pull/576))
 - Added NLP metrics:
   - `SacreBLEUScore` ([#546](https://github.com/PyTorchLightning/metrics/pull/546))

diff --git a/tests/detection/test_map.py b/tests/detection/test_map.py
@@ -206,6 +206,23 @@ def test_error_on_wrong_init():
         MAP(class_metrics=0)
 
 
+@pytest.mark.skipif(_pytest_condition, reason="test requires that pycocotools and torchvision=>0.8.0 is installed")
+def test_empty_preds():
+    """Test empty predictions."""
+
+    metric = MAP()
+
+    metric.update(
+        [
+            dict(boxes=torch.Tensor([[]]), scores=torch.Tensor([]), labels=torch.IntTensor([])),
+        ],
+        [
+            dict(boxes=torch.Tensor([[214.1500, 41.2900, 562.4100, 285.0700]]), labels=torch.IntTensor([4])),
+        ],
+    )
+    metric.compute()
+
+
 @pytest.mark.skipif(_pytest_condition, reason="test requires that pycocotools and torchvision=>0.8.0 is installed")
 def test_error_on_wrong_input():
     """Test class input validation."""

diff --git a/torchmetrics/detection/map.py b/torchmetrics/detection/map.py
@@ -365,7 +365,7 @@ def _get_coco_format(
         annotations = []
         annotation_id = 1  # has to start with 1, otherwise COCOEval results are wrong
 
-        boxes = [box_convert(box, in_fmt="xyxy", out_fmt="xywh") for box in boxes]
+        boxes = [box_convert(box, in_fmt="xyxy", out_fmt="xywh") if boxes[0].size(1) == 4 else box for box in boxes]
         for image_id, (image_boxes, image_labels) in enumerate(zip(boxes, labels)):
             image_boxes = image_boxes.cpu().tolist()
             image_labels = image_labels.cpu().tolist()