correct affine behavior, cvcuda center on top left

Author: justincdavis

test/test_transforms_v2.py

@@ -1575,11 +1575,8 @@ def test_functional_image_correctness(
         )
 
         if make_input is make_image_cvcuda:
-            actual = F.cvcuda_to_tensor(actual).to(device="cpu")
-            actual = actual.squeeze(0)
-            # drop the batch dimensions for image now
-            image = F.cvcuda_to_tensor(image)
-            image = image.squeeze(0)
+            actual = cvcuda_to_pil_compatible_tensor(actual)
+            image = cvcuda_to_pil_compatible_tensor(image)
 
         expected = F.to_image(
             F.affine(
@@ -1629,11 +1626,8 @@ def test_transform_image_correctness(self, center, interpolation, fill, seed, ma
         actual = transform(image)
 
         if make_input is make_image_cvcuda:
-            actual = F.cvcuda_to_tensor(actual).to(device="cpu")
-            actual = actual.squeeze(0)
-            # drop the batch dimensions for image now
-            image = F.cvcuda_to_tensor(image)
-            image = image.squeeze(0)
+            actual = cvcuda_to_pil_compatible_tensor(actual)
+            image = cvcuda_to_pil_compatible_tensor(image)
 
         torch.manual_seed(seed)
         expected = F.to_image(transform(F.to_pil_image(image)))

torchvision/transforms/v2/functional/_geometry.py

@@ -1361,12 +1361,17 @@ def _affine_cvcuda(
 
     height, width, num_channels = image.shape[1:]
 
-    center_f = [0.0, 0.0]
-    if center is not None:
-        center_f = [(c - s * 0.5) for c, s in zip(center, [width, height])]
+    # Determine the actual center point (cx, cy)
+    # torchvision uses image center by default, cvcuda transforms around upper-left (0,0)
+    # Unlike the tensor version which uses normalized coordinates centered at image center,
+    # CV-CUDA uses absolute pixel coordinates, so we pass actual center to _get_inverse_affine_matrix
+    if center is None:
+        cx, cy = width / 2.0, height / 2.0
+    else:
+        cx, cy = float(center[0]), float(center[1])
 
     translate_f = [float(t) for t in translate]
-    matrix = _get_inverse_affine_matrix(center_f, angle, translate_f, scale, shear)
+    matrix = _get_inverse_affine_matrix([cx, cy], angle, translate_f, scale, shear)
 
     interp = _cvcuda_interp.get(interpolation)
     if interp is None: