fix FID metric

Summary:
had to add a check in to see if num of real/fake images is bigger than 2 because otherwise we divided in zero in:
```real_cov = real_cov_num / (self.num_real_images - 1)```
and
```fake_cov = fake_cov_num / (self.num_fake_images - 1)```

Reviewed By: JKSenthil

Differential Revision: D49570353

fbshipit-source-id: b829d21fcc39c8a76d839349ab13cd2330fdf104

torcheval/metrics/image/fid.py

@@ -168,9 +168,9 @@ def compute(self: TFrechetInceptionDistance) -> Tensor:
 
         # If the user has not already updated with at lease one
         # image from each distribution, then we raise an Error.
-        if (self.num_real_images == 0) or (self.num_fake_images == 0):
+        if (self.num_real_images < 2) or (self.num_fake_images < 2):
             warnings.warn(
-                "Computing FID requires at least 1 real image and 1 fake image,"
+                "Computing FID requires at least 2 real images and 2 fake images,"
                 f"but currently running with {self.num_real_images} real images and {self.num_fake_images} fake images."
                 "Returning 0.0",
                 RuntimeWarning,