Fid: error on too few samples

CHANGELOG.md

@@ -96,6 +96,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Changed `__iter__` method from raising `NotImplementedError` to `TypeError` by setting to `None` ([#1538](https://github.com/Lightning-AI/metrics/pull/1538))
 
 
+- `FID` metric will now raise an error if too few samples are provided ([#1655](https://github.com/Lightning-AI/metrics/pull/1655))
+
+
 - Allowed FID with `torch.float64` ([#1628](https://github.com/Lightning-AI/metrics/pull/1628))
 
 

src/torchmetrics/image/fid.py

@@ -380,6 +380,8 @@ def update(self, imgs: Tensor, real: bool) -> None:
 
     def compute(self) -> Tensor:
         """Calculate FID score based on accumulated extracted features from the two distributions."""
+        if self.real_features_num_samples < 2 or self.fake_features_num_samples < 2:
+            raise RuntimeError("More than one sample is required for both the real and fake distributed to compute FID")
         mean_real = (self.real_features_sum / self.real_features_num_samples).unsqueeze(0)
         mean_fake = (self.fake_features_sum / self.fake_features_num_samples).unsqueeze(0)
 

tests/unittests/image/test_fid.py

@@ -201,6 +201,18 @@ def test_normalize_arg_false():
         metric.update(img, real=True)
 
 
+def test_not_enough_samples():
+    """Test that an error is raised if not enough samples were provided."""
+    img = torch.randint(0, 255, (1, 3, 299, 299), dtype=torch.uint8)
+    metric = FrechetInceptionDistance()
+    metric.update(img, real=True)
+    metric.update(img, real=False)
+    with pytest.raises(
+        RuntimeError, match="More than one sample is required for both the real and fake distributed to compute FID"
+    ):
+        metric.compute()
+
+
 def test_dtype_transfer_to_submodule():
     """Test that change in dtype also changes the default inception net."""
     imgs = torch.randn(1, 3, 256, 256)