fix(image): prevent NaN in spectral_angle_mapper when a pixel has zero norm

CHANGELOG.md

@@ -34,6 +34,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Fixed
 
 - Fixed malformed LaTeX in `CLIPScore` and `HausdorffDistance` docstring math so it renders correctly ([#3427](https://github.com/Lightning-AI/torchmetrics/pull/3427))
+- Fixed `NaN` in `SpectralAngleMapper` / `spectral_angle_mapper` when a pixel has zero norm ([#3424](https://github.com/Lightning-AI/torchmetrics/pull/3424))
 
 
 ---

src/torchmetrics/functional/image/sam.py

@@ -74,7 +74,9 @@ def _sam_compute(
     dot_product = (preds * target).sum(dim=1)
     preds_norm = preds.norm(dim=1)
     target_norm = target.norm(dim=1)
-    sam_score = torch.clamp(dot_product / (preds_norm * target_norm), -1, 1).acos()
+    # Clamp denominator to avoid NaN when a pixel has zero norm (all-zero channels)
+    denom = (preds_norm * target_norm).clamp(min=torch.finfo(preds.dtype).eps)
+    sam_score = torch.clamp(dot_product / denom, -1, 1).acos()
     return reduce(sam_score, reduction)
 
 

tests/unittests/image/test_sam.py

@@ -130,3 +130,27 @@ def test_error_on_grayscale_image(metric_class=SpectralAngleMapper):
     metric = metric_class()
     with pytest.raises(ValueError, match="Expected channel dimension of `preds` and `target` to be larger than 1.*"):
         metric(torch.randn([16, 1, 16, 16]), torch.randn([16, 1, 16, 16]))
+
+
+def test_no_nan_on_zero_pixel():
+    """Regression test for https://github.com/Lightning-AI/torchmetrics/issues/3322.
+
+    spectral_angle_mapper should not produce NaN when a pixel has zero norm
+    (all channels zero).  Previously, the zero-norm denominator caused 0/0 = NaN.
+    """
+    preds = torch.ones(1, 3, 8, 8)  # N, C, H, W
+    target = torch.ones(1, 3, 8, 8)
+    preds[:, :, 5, 3] = 0  # zero-norm pixel — exact reproducer from the issue
+
+    # functional interface
+    result = spectral_angle_mapper(preds, target)
+    assert torch.isfinite(result).all(), f"spectral_angle_mapper returned non-finite value: {result}"
+
+    # class interface
+    metric = SpectralAngleMapper()
+    result_cls = metric(preds, target)
+    assert torch.isfinite(result_cls).all(), f"SpectralAngleMapper returned non-finite value: {result_cls}"
+
+    # Result should be a valid angle in [0, pi/2]
+    assert (result >= 0).all(), f"result is negative: {result}"
+    assert (result <= torch.pi / 2).all(), f"result exceeds pi/2: {result}"