Fix FID computation for non equal size (#1028)

* fix fid computation
* changelog

CHANGELOG.md

@@ -62,6 +62,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 
 
+- Fixed `FID` calculation for non-equal size real and fake input ([#1028](https://github.com/PyTorchLightning/metrics/pull/1028))
+
+
 ## [0.8.2] - 2022-05-06
 
 

tests/image/test_fid.py

@@ -125,21 +125,25 @@ def __len__(self):
 
 @pytest.mark.skipif(not torch.cuda.is_available(), reason="test is too slow without gpu")
 @pytest.mark.skipif(not _TORCH_FIDELITY_AVAILABLE, reason="test requires torch-fidelity")
-def test_compare_fid(tmpdir, feature=2048):
+@pytest.mark.parametrize("equal_size", [False, True])
+def test_compare_fid(tmpdir, equal_size, feature=2048):
     """check that the hole pipeline give the same result as torch-fidelity."""
     from torch_fidelity import calculate_metrics
 
     metric = FrechetInceptionDistance(feature=feature).cuda()
 
+    n = 100
+    m = 100 if equal_size else 90
+
     # Generate some synthetic data
-    img1 = torch.randint(0, 180, (100, 3, 299, 299), dtype=torch.uint8)
-    img2 = torch.randint(100, 255, (100, 3, 299, 299), dtype=torch.uint8)
+    img1 = torch.randint(0, 180, (n, 3, 299, 299), dtype=torch.uint8)
+    img2 = torch.randint(100, 255, (m, 3, 299, 299), dtype=torch.uint8)
 
     batch_size = 10
-    for i in range(img1.shape[0] // batch_size):
+    for i in range(n // batch_size):
         metric.update(img1[batch_size * i : batch_size * (i + 1)].cuda(), real=True)
 
-    for i in range(img2.shape[0] // batch_size):
+    for i in range(m // batch_size):
         metric.update(img2[batch_size * i : batch_size * (i + 1)].cuda(), real=False)
 
     torch_fid = calculate_metrics(

torchmetrics/image/fid.py

@@ -268,12 +268,13 @@ def compute(self) -> Tensor:
 
         # calculate mean and covariance
         n = real_features.shape[0]
+        m = fake_features.shape[0]
         mean1 = real_features.mean(dim=0)
         mean2 = fake_features.mean(dim=0)
         diff1 = real_features - mean1
         diff2 = fake_features - mean2
         cov1 = 1.0 / (n - 1) * diff1.t().mm(diff1)
-        cov2 = 1.0 / (n - 1) * diff2.t().mm(diff2)
+        cov2 = 1.0 / (m - 1) * diff2.t().mm(diff2)
 
         # compute fid
         return _compute_fid(mean1, cov1, mean2, cov2).to(orig_dtype)