Add ToPureTensor transform (#7823)

Co-authored-by: Philip Meier <github.pmeier@posteo.de>

docs/source/transforms.rst

@@ -237,6 +237,7 @@ Conversion
     v2.ConvertImageDtype
     v2.ToDtype
     v2.ConvertBoundingBoxFormat
+    v2.ToPureTensor
 
 Auto-Augmentation
 -----------------

references/classification/presets.py

@@ -68,6 +68,9 @@ class ClassificationPresetTrain:
         if random_erase_prob > 0:
             transforms.append(T.RandomErasing(p=random_erase_prob))
 
+        if use_v2:
+            transforms.append(T.ToPureTensor())
+
         self.transforms = T.Compose(transforms)
 
     def __call__(self, img):
@@ -107,6 +110,9 @@ class ClassificationPresetEval:
             T.Normalize(mean=mean, std=std),
         ]
 
+        if use_v2:
+            transforms.append(T.ToPureTensor())
+
         self.transforms = T.Compose(transforms)
 
     def __call__(self, img):

references/detection/presets.py

@@ -79,6 +79,7 @@ class DetectionPresetTrain:
             transforms += [
                 T.ConvertBoundingBoxFormat(datapoints.BoundingBoxFormat.XYXY),
                 T.SanitizeBoundingBoxes(),
+                T.ToPureTensor(),
             ]
 
         self.transforms = T.Compose(transforms)
@@ -103,6 +104,10 @@ class DetectionPresetEval:
             raise ValueError(f"backend can be 'datapoint', 'tensor' or 'pil', but got {backend}")
 
         transforms += [T.ConvertImageDtype(torch.float)]
+
+        if use_v2:
+            transforms += [T.ToPureTensor()]
+
         self.transforms = T.Compose(transforms)
 
     def __call__(self, img, target):

references/segmentation/presets.py

@@ -63,6 +63,8 @@ class SegmentationPresetTrain:
             transforms += [T.ConvertImageDtype(torch.float)]
 
         transforms += [T.Normalize(mean=mean, std=std)]
+        if use_v2:
+            transforms += [T.ToPureTensor()]
 
         self.transforms = T.Compose(transforms)
 
@@ -98,6 +100,9 @@ class SegmentationPresetEval:
             T.ConvertImageDtype(torch.float),
             T.Normalize(mean=mean, std=std),
         ]
+        if use_v2:
+            transforms += [T.ToPureTensor()]
+
         self.transforms = T.Compose(transforms)
 
     def __call__(self, img, target):

test/test_transforms_v2_refactored.py

@@ -2353,3 +2353,24 @@ class TestElastic:
     @pytest.mark.parametrize("device", cpu_and_cuda())
     def test_transform(self, make_input, size, device):
         check_transform(transforms.ElasticTransform, make_input(size, device=device))
+
+
+class TestToPureTensor:
+    def test_correctness(self):
+        input = {
+            "img": make_image(),
+            "img_tensor": make_image_tensor(),
+            "img_pil": make_image_pil(),
+            "mask": make_detection_mask(),
+            "video": make_video(),
+            "bbox": make_bounding_box(),
+            "str": "str",
+        }
+
+        out = transforms.ToPureTensor()(input)
+
+        for input_value, out_value in zip(input.values(), out.values()):
+            if isinstance(input_value, datapoints.Datapoint):
+                assert isinstance(out_value, torch.Tensor) and not isinstance(out_value, datapoints.Datapoint)
+            else:
+                assert isinstance(out_value, type(input_value))

torchvision/transforms/v2/__init__.py

@@ -52,7 +52,7 @@ from ._misc import (
     ToDtype,
 )
 from ._temporal import UniformTemporalSubsample
-from ._type_conversion import PILToTensor, ToImage, ToPILImage
+from ._type_conversion import PILToTensor, ToImage, ToPILImage, ToPureTensor
 
 from ._deprecated import ToTensor  # usort: skip
 

torchvision/transforms/v2/_type_conversion.py

@@ -75,3 +75,17 @@ class ToPILImage(Transform):
         self, inpt: Union[torch.Tensor, PIL.Image.Image, np.ndarray], params: Dict[str, Any]
     ) -> PIL.Image.Image:
         return F.to_pil_image(inpt, mode=self.mode)
+
+
+class ToPureTensor(Transform):
+    """[BETA] Convert all datapoints to pure tensors, removing associated metadata (if any).
+
+    .. v2betastatus:: ToPureTensor transform
+
+    This doesn't scale or change the values, only the type.
+    """
+
+    _transformed_types = (datapoints.Datapoint,)
+
+    def _transform(self, inpt: Any, params: Dict[str, Any]) -> torch.Tensor:
+        return inpt.as_subclass(torch.Tensor)