cvcuda rotate verified correct visualizly and passing all tests

Author: justincdavis

test/test_transforms_v2.py

@@ -2152,7 +2152,8 @@ def test_functional_image_correctness(self, angle, center, interpolation, expand
 
         actual = F.rotate(image, angle=angle, center=center, interpolation=interpolation, expand=expand, fill=fill)
 
-        if make_input == make_image_cvcuda:
+        if make_input is make_image_cvcuda:
+            actual = cvcuda_to_pil_compatible_tensor(actual)
             image = cvcuda_to_pil_compatible_tensor(image)
 
         expected = F.to_image(
@@ -2162,7 +2163,7 @@ def test_functional_image_correctness(self, angle, center, interpolation, expand
         )
 
         mae = (actual.float() - expected.float()).abs().mean()
-        if make_input == make_image_cvcuda:
+        if make_input is make_image_cvcuda:
             # CV-CUDA nearest interpolation differs significantly from PIL, set much higher bound
             assert mae < (122.5) if interpolation is transforms.InterpolationMode.NEAREST else 6, f"MAE: {mae}"
         else:
@@ -2202,16 +2203,14 @@ def test_transform_image_correctness(self, center, interpolation, expand, fill,
 
         torch.manual_seed(seed)
 
-        if make_input == make_image_cvcuda:
-            actual = F.cvcuda_to_tensor(actual).to(device="cpu")
-            image = F.cvcuda_to_tensor(image)
-            # drop the batch dimensions
-            image = image.squeeze(0)
+        if make_input is make_image_cvcuda:
+            actual = cvcuda_to_pil_compatible_tensor(actual)
+            image = cvcuda_to_pil_compatible_tensor(image)
 
         expected = F.to_image(transform(F.to_pil_image(image)))
 
         mae = (actual.float() - expected.float()).abs().mean()
-        if make_input == make_image_cvcuda:
+        if make_input is make_image_cvcuda:
             # CV-CUDA nearest interpolation differs significantly from PIL, set much higher bound
             assert mae < (122.5) if interpolation is transforms.InterpolationMode.NEAREST else 6, f"MAE: {mae}"
         else:

torchvision/transforms/v2/functional/_geometry.py

@@ -1584,8 +1584,11 @@ def _rotate_cvcuda(
     # We need to calculate a shift to effectively rotate around the desired center
     if center is None:
         cx, cy = input_width / 2.0, input_height / 2.0
+        center_f = [0.0, 0.0]
     else:
         cx, cy = float(center[0]), float(center[1])
+        # Convert to image-center-relative coordinates (same as torchvision)
+        center_f = [cx - input_width * 0.5, cy - input_height * 0.5]
 
     angle_rad = math.radians(angle)
     cos_angle = math.cos(angle_rad)
@@ -1599,7 +1602,8 @@ def _rotate_cvcuda(
         return cvcuda.rotate(inpt, angle_deg=angle, shift=(shift_x, shift_y), interpolation=interp)
 
     # if we need to expand, use much of the same logic as torchvision, for output size/pad
-    matrix = _get_inverse_affine_matrix([0.0, 0.0], -angle, [0.0, 0.0], 1.0, [0.0, 0.0])
+    # Use center_f (image-center-relative coords) to match torchvision's output size calculation
+    matrix = _get_inverse_affine_matrix(center_f, -angle, [0.0, 0.0], 1.0, [0.0, 0.0])
     output_width, output_height = _compute_affine_output_size(matrix, input_width, input_height)
 
     pad_left = (output_width - input_width) // 2