Let Normalize() and RandomPhotometricDistort return datapoints instead of tensors (#7113)

Author: NicolasHug

test/prototype_transforms_dispatcher_infos.py

@@ -426,7 +426,6 @@ def fill_sequence_needs_broadcast(args_kwargs):
             datapoints.Video: F.normalize_video,
         },
         test_marks=[
-            skip_dispatch_feature,
             xfail_jit_python_scalar_arg("mean"),
             xfail_jit_python_scalar_arg("std"),
         ],

test/test_prototype_transforms_functional.py

@@ -13,7 +13,7 @@
 
 import torchvision.prototype.transforms.utils
 from common_utils import cache, cpu_and_gpu, needs_cuda, set_rng_seed
-from prototype_common_utils import assert_close, make_bounding_boxes, make_image, parametrized_error_message
+from prototype_common_utils import assert_close, make_bounding_boxes, parametrized_error_message
 from prototype_transforms_dispatcher_infos import DISPATCHER_INFOS
 from prototype_transforms_kernel_infos import KERNEL_INFOS
 from torch.utils._pytree import tree_map
@@ -1185,18 +1185,6 @@ def test_correctness_gaussian_blur_image_tensor(device, spatial_size, dt, ksize,
     torch.testing.assert_close(out, true_out, rtol=0.0, atol=1.0, msg=f"{ksize}, {sigma}")
 
 
-def test_normalize_output_type():
-    inpt = torch.rand(1, 3, 32, 32)
-    output = F.normalize(inpt, mean=[0.5, 0.5, 0.5], std=[1.0, 1.0, 1.0])
-    assert type(output) is torch.Tensor
-    torch.testing.assert_close(inpt - 0.5, output)
-
-    inpt = make_image(color_space=datapoints.ColorSpace.RGB)
-    output = F.normalize(inpt, mean=[0.5, 0.5, 0.5], std=[1.0, 1.0, 1.0])
-    assert type(output) is torch.Tensor
-    torch.testing.assert_close(inpt - 0.5, output)
-
-
 @pytest.mark.parametrize(
     "inpt",
     [

torchvision/prototype/datapoints/_image.py

@@ -289,6 +289,10 @@ def gaussian_blur(self, kernel_size: List[int], sigma: Optional[List[float]] = N
         )
         return Image.wrap_like(self, output)
 
+    def normalize(self, mean: List[float], std: List[float], inplace: bool = False) -> Image:
+        output = self._F.normalize_image_tensor(self.as_subclass(torch.Tensor), mean=mean, std=std, inplace=inplace)
+        return Image.wrap_like(self, output)
+
 
 ImageType = Union[torch.Tensor, PIL.Image.Image, Image]
 ImageTypeJIT = torch.Tensor

torchvision/prototype/datapoints/_video.py

@@ -241,6 +241,10 @@ def gaussian_blur(self, kernel_size: List[int], sigma: Optional[List[float]] = N
         output = self._F.gaussian_blur_video(self.as_subclass(torch.Tensor), kernel_size=kernel_size, sigma=sigma)
         return Video.wrap_like(self, output)
 
+    def normalize(self, mean: List[float], std: List[float], inplace: bool = False) -> Video:
+        output = self._F.normalize_video(self.as_subclass(torch.Tensor), mean=mean, std=std, inplace=inplace)
+        return Video.wrap_like(self, output)
+
 
 VideoType = Union[torch.Tensor, Video]
 VideoTypeJIT = torch.Tensor

torchvision/prototype/transforms/_color.py

@@ -82,6 +82,7 @@ def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
         return output
 
 
+# TODO: This class seems to be untested
 class RandomPhotometricDistort(Transform):
     _transformed_types = (
         datapoints.Image,
@@ -119,15 +120,14 @@ def _get_params(self, flat_inputs: List[Any]) -> Dict[str, Any]:
     def _permute_channels(
         self, inpt: Union[datapoints.ImageType, datapoints.VideoType], permutation: torch.Tensor
     ) -> Union[datapoints.ImageType, datapoints.VideoType]:
-        if isinstance(inpt, PIL.Image.Image):
+
+        orig_inpt = inpt
+        if isinstance(orig_inpt, PIL.Image.Image):
             inpt = F.pil_to_tensor(inpt)
 
         output = inpt[..., permutation, :, :]
 
-        if isinstance(inpt, (datapoints.Image, datapoints.Video)):
-            output = inpt.wrap_like(inpt, output, color_space=datapoints.ColorSpace.OTHER)  # type: ignore[arg-type]
-
-        elif isinstance(inpt, PIL.Image.Image):
+        if isinstance(orig_inpt, PIL.Image.Image):
             output = F.to_image_pil(output)
 
         return output

torchvision/prototype/transforms/functional/_misc.py

@@ -60,18 +60,14 @@ def normalize(
 ) -> torch.Tensor:
     if not torch.jit.is_scripting():
         _log_api_usage_once(normalize)
-
-        if isinstance(inpt, (datapoints.Image, datapoints.Video)):
-            inpt = inpt.as_subclass(torch.Tensor)
-        elif not is_simple_tensor(inpt):
-            raise TypeError(
-                f"Input can either be a plain tensor or an `Image` or `Video` datapoint, "
-                f"but got {type(inpt)} instead."
-            )
-
-    # Image or Video type should not be retained after normalization due to unknown data range
-    # Thus we return Tensor for input Image
-    return normalize_image_tensor(inpt, mean=mean, std=std, inplace=inplace)
+    if torch.jit.is_scripting() or is_simple_tensor(inpt):
+        return normalize_image_tensor(inpt, mean=mean, std=std, inplace=inplace)
+    elif isinstance(inpt, (datapoints.Image, datapoints.Video)):
+        return inpt.normalize(mean=mean, std=std, inplace=inplace)
+    else:
+        raise TypeError(
+            f"Input can either be a plain tensor or an `Image` or `Video` datapoint, " f"but got {type(inpt)} instead."
+        )
 
 
 def _get_gaussian_kernel1d(kernel_size: int, sigma: float, dtype: torch.dtype, device: torch.device) -> torch.Tensor: