Add tests and proper support for videos in ConvertImageDtype (#6783)

* add KernelInfo

* split dtype and device consistency tests

* add proper support for video

* fix tests and add DispatcherInfo

* add aliases

* cleanup

* fix typo

test/prototype_common_utils.py

@@ -22,7 +22,7 @@
     UnsupportedInputs,
 )
 from torchvision.prototype import features
-from torchvision.prototype.transforms.functional import convert_image_dtype, to_image_tensor
+from torchvision.prototype.transforms.functional import convert_dtype_image_tensor, to_image_tensor
 from torchvision.transforms.functional_tensor import _max_value as get_max_value
 
 __all__ = [
@@ -97,8 +97,8 @@ def _process_inputs(self, actual, expected, *, id, allow_subclasses):
     def _equalize_attributes(self, actual, expected):
         if actual.dtype != expected.dtype:
             dtype = torch.promote_types(actual.dtype, expected.dtype)
-            actual = convert_image_dtype(actual, dtype)
-            expected = convert_image_dtype(expected, dtype)
+            actual = convert_dtype_image_tensor(actual, dtype)
+            expected = convert_dtype_image_tensor(expected, dtype)
 
         return super()._equalize_attributes(actual, expected)
 

test/prototype_transforms_dispatcher_infos.py

@@ -416,4 +416,14 @@ def xfail_all_tests(*, reason, condition):
             skip_dispatch_feature,
         ],
     ),
+    DispatcherInfo(
+        F.convert_dtype,
+        kernels={
+            features.Image: F.convert_dtype_image_tensor,
+            features.Video: F.convert_dtype_video,
+        },
+        test_marks=[
+            skip_dispatch_feature,
+        ],
+    ),
 ]

test/prototype_transforms_kernel_infos.py

@@ -1979,7 +1979,7 @@ def sample_inputs_normalize_video():
 )
 
 
-def sample_inputs_convert_image_dtype():
+def sample_inputs_convert_dtype_image_tensor():
     for input_dtype, output_dtype in itertools.product(
         [torch.uint8, torch.int64, torch.float32, torch.float64], repeat=2
     ):
@@ -1992,10 +1992,8 @@ def sample_inputs_convert_image_dtype():
         ):
             yield ArgsKwargs(image_loader, dtype=output_dtype)
 
-    yield ArgsKwargs(make_image_loader(color_space=features.ColorSpace.RGB), dtype=torch.uint8)
 
-
-def reference_convert_image_dtype(image, dtype=torch.float):
+def reference_convert_dtype_image_tensor(image, dtype=torch.float):
     input_dtype = image.dtype
     output_dtype = dtype
 
@@ -2026,7 +2024,7 @@ def fn(value):
     return torch.tensor(tree_map(fn, image.tolist()), dtype=dtype)
 
 
-def reference_inputs_convert_image_dtype():
+def reference_inputs_convert_dtype_image_tensor():
     for input_dtype, output_dtype in itertools.product(
         [
             torch.uint8,
@@ -2055,41 +2053,50 @@ def reference_inputs_convert_image_dtype():
         yield ArgsKwargs(image, dtype=output_dtype)
 
 
+def sample_inputs_convert_dtype_video():
+    for video_loader in make_video_loaders(sizes=["random"], num_frames=["random"]):
+        yield ArgsKwargs(video_loader)
+
+
+_common_convert_dtype_marks = [
+    TestMark(
+        ("TestKernels", "test_dtype_and_device_consistency"),
+        pytest.mark.skip(reason="`convert_dtype_*` kernels convert the dtype by design"),
+        condition=lambda args_kwargs: args_kwargs.args[0].dtype != args_kwargs.kwargs.get("dtype", torch.float32),
+    ),
+    TestMark(
+        ("TestKernels", "test_scripted_vs_eager"),
+        pytest.mark.filterwarnings(f"ignore:{re.escape('operator() profile_node %')}:UserWarning"),
+    ),
+]
+
 KERNEL_INFOS.extend(
     [
         KernelInfo(
-            F.convert_image_dtype,
-            sample_inputs_fn=sample_inputs_convert_image_dtype,
-            reference_fn=reference_convert_image_dtype,
-            reference_inputs_fn=reference_inputs_convert_image_dtype,
+            F.convert_dtype_image_tensor,
+            sample_inputs_fn=sample_inputs_convert_dtype_image_tensor,
+            reference_fn=reference_convert_dtype_image_tensor,
+            reference_inputs_fn=reference_inputs_convert_dtype_image_tensor,
             test_marks=[
-                TestMark(
-                    ("TestKernels", "test_scripted_vs_eager"),
-                    pytest.mark.filterwarnings(f"ignore:{re.escape('operator() profile_node %41')}:UserWarning"),
-                ),
-                TestMark(
-                    ("TestKernels", "test_dtype_and_device_consistency"),
-                    pytest.mark.skip(reason="`convert_dtype_*` kernels convert the dtype by design"),
-                    condition=lambda args_kwargs: args_kwargs.args[0].dtype
-                    != args_kwargs.kwargs.get("dtype", torch.float32),
-                ),
+                *_common_convert_dtype_marks,
                 TestMark(
                     ("TestKernels", "test_against_reference"),
                     pytest.mark.xfail(reason="Conversion overflows"),
                     condition=lambda args_kwargs: (
                         args_kwargs.args[0].dtype in {torch.float16, torch.bfloat16}
                         and not args_kwargs.kwargs["dtype"].is_floating_point
                     )
-                    or (
-                        args_kwargs.args[0].dtype in {torch.float16, torch.bfloat16}
-                        and args_kwargs.kwargs["dtype"] == torch.int64
-                    )
                     or (
                         args_kwargs.args[0].dtype in {torch.int32, torch.int64}
                         and args_kwargs.kwargs["dtype"] == torch.float16
                     ),
                 ),
             ],
         ),
+        KernelInfo(
+            F.convert_dtype_video,
+            sample_inputs_fn=sample_inputs_convert_dtype_video,
+            test_marks=_common_convert_dtype_marks,
+        ),
     ]
 )

test/test_prototype_transforms.py

@@ -92,7 +92,7 @@ class TestSmoke:
         transforms.RandomErasing(p=1.0),
         transforms.Resize([16, 16]),
         transforms.CenterCrop([16, 16]),
-        transforms.ConvertImageDtype(),
+        transforms.ConvertDtype(),
         transforms.RandomHorizontalFlip(),
         transforms.Pad(5),
         transforms.RandomZoomOut(),

test/test_prototype_transforms_consistency.py

@@ -153,7 +153,7 @@ def __init__(
         ),
     ),
     ConsistencyConfig(
-        prototype_transforms.ConvertImageDtype,
+        prototype_transforms.ConvertDtype,
         legacy_transforms.ConvertImageDtype,
         [
             ArgsKwargs(torch.float16),

test/test_prototype_transforms_functional.py

@@ -307,6 +307,7 @@ def test_dispatch_feature(self, info, args_kwargs, spy_on):
             (F.get_image_num_channels, F.get_num_channels),
             (F.to_pil_image, F.to_image_pil),
             (F.elastic_transform, F.elastic),
+            (F.convert_image_dtype, F.convert_dtype_image_tensor),
         ]
     ],
 )

torchvision/prototype/transforms/__init__.py

@@ -39,7 +39,7 @@
     ScaleJitter,
     TenCrop,
 )
-from ._meta import ClampBoundingBoxes, ConvertBoundingBoxFormat, ConvertColorSpace, ConvertImageDtype
+from ._meta import ClampBoundingBoxes, ConvertBoundingBoxFormat, ConvertColorSpace, ConvertDtype, ConvertImageDtype
 from ._misc import (
     GaussianBlur,
     Identity,

torchvision/prototype/transforms/_meta.py

@@ -25,7 +25,7 @@ def _transform(self, inpt: features.BoundingBox, params: Dict[str, Any]) -> feat
         return features.BoundingBox.wrap_like(inpt, output, format=params["format"])
 
 
-class ConvertImageDtype(Transform):
+class ConvertDtype(Transform):
     _transformed_types = (features.is_simple_tensor, features.Image, features.Video)
 
     def __init__(self, dtype: torch.dtype = torch.float32) -> None:
@@ -35,12 +35,12 @@ def __init__(self, dtype: torch.dtype = torch.float32) -> None:
     def _transform(
         self, inpt: Union[features.TensorImageType, features.TensorVideoType], params: Dict[str, Any]
     ) -> Union[features.TensorImageType, features.TensorVideoType]:
-        # TODO: the `inpt.as_subclass(torch.Tensor)` call can be removed as soon as we have a proper dispatcher that
-        #  handles this. See https://github.com/pytorch/vision/pull/6783 for details.
-        output = F.convert_image_dtype(inpt.as_subclass(torch.Tensor), dtype=self.dtype)
-        return (
-            output if features.is_simple_tensor(inpt) else type(inpt).wrap_like(inpt, output)  # type: ignore[attr-defined]
-        )
+        return F.convert_dtype(inpt, self.dtype)
+
+
+# We changed the name to align it with the new naming scheme. Still, `ConvertImageDtype` is
+# prevalent and well understood. Thus, we just alias it without deprecating the old name.
+ConvertImageDtype = ConvertDtype
 
 
 class ConvertColorSpace(Transform):

torchvision/prototype/transforms/functional/__init__.py

@@ -8,6 +8,10 @@
     convert_color_space_image_pil,
     convert_color_space_video,
     convert_color_space,
+    convert_dtype_image_tensor,
+    convert_dtype,
+    convert_dtype_video,
+    convert_image_dtype,
     get_dimensions_image_tensor,
     get_dimensions_image_pil,
     get_dimensions,
@@ -162,7 +166,6 @@
     normalize_video,
 )
 from ._type_conversion import (
-    convert_image_dtype,
     decode_image_with_pil,
     decode_video_with_av,
     pil_to_tensor,

torchvision/prototype/transforms/functional/_meta.py

@@ -285,3 +285,99 @@ def convert_color_space(
         return features.Video.wrap_like(inpt, output, color_space=color_space)
     else:
         return convert_color_space_image_pil(inpt, color_space)
+
+
+def _num_value_bits(dtype: torch.dtype) -> int:
+    if dtype == torch.uint8:
+        return 8
+    elif dtype == torch.int8:
+        return 7
+    elif dtype == torch.int16:
+        return 15
+    elif dtype == torch.int32:
+        return 31
+    elif dtype == torch.int64:
+        return 63
+    else:
+        raise TypeError(f"Number of value bits is only defined for integer dtypes, but got {dtype}.")
+
+
+def convert_dtype_image_tensor(image: torch.Tensor, dtype: torch.dtype = torch.float) -> torch.Tensor:
+    if image.dtype == dtype:
+        return image
+
+    float_input = image.is_floating_point()
+    if torch.jit.is_scripting():
+        # TODO: remove this branch as soon as `dtype.is_floating_point` is supported by JIT
+        float_output = torch.tensor(0, dtype=dtype).is_floating_point()
+    else:
+        float_output = dtype.is_floating_point
+
+    if float_input:
+        # float to float
+        if float_output:
+            return image.to(dtype)
+
+        # float to int
+        if (image.dtype == torch.float32 and dtype in (torch.int32, torch.int64)) or (
+            image.dtype == torch.float64 and dtype == torch.int64
+        ):
+            raise RuntimeError(f"The conversion from {image.dtype} to {dtype} cannot be performed safely.")
+
+        # For data in the range `[0.0, 1.0]`, just multiplying by the maximum value of the integer range and converting
+        # to the integer dtype  is not sufficient. For example, `torch.rand(...).mul(255).to(torch.uint8)` will only
+        # be `255` if the input is exactly `1.0`. See https://github.com/pytorch/vision/pull/2078#issuecomment-612045321
+        # for a detailed analysis.
+        # To mitigate this, we could round before we convert to the integer dtype, but this is an extra operation.
+        # Instead, we can also multiply by the maximum value plus something close to `1`. See
+        # https://github.com/pytorch/vision/pull/2078#issuecomment-613524965 for details.
+        eps = 1e-3
+        max_value = float(_FT._max_value(dtype))
+        # We need to scale first since the conversion would otherwise turn the input range `[0.0, 1.0]` into the
+        # discrete set `{0, 1}`.
+        return image.mul(max_value + 1.0 - eps).to(dtype)
+    else:
+        # int to float
+        if float_output:
+            return image.to(dtype).div_(_FT._max_value(image.dtype))
+
+        # int to int
+        num_value_bits_input = _num_value_bits(image.dtype)
+        num_value_bits_output = _num_value_bits(dtype)
+
+        if num_value_bits_input > num_value_bits_output:
+            return image.bitwise_right_shift(num_value_bits_input - num_value_bits_output).to(dtype)
+        else:
+            # The bitshift kernel is not vectorized
+            #  https://github.com/pytorch/pytorch/blob/703c19008df4700b6a522b0ae5c4b6d5ffc0906f/aten/src/ATen/native/cpu/BinaryOpsKernel.cpp#L315-L322
+            #  This results in the multiplication actually being faster.
+            # TODO: If the bitshift kernel is optimized in core, replace the computation below with
+            #  `image.to(dtype).bitwise_left_shift_(num_value_bits_output - num_value_bits_input)`
+            max_value_input = float(_FT._max_value(dtype))
+            max_value_output = float(_FT._max_value(image.dtype))
+            factor = int((max_value_input + 1) // (max_value_output + 1))
+            return image.to(dtype).mul_(factor)
+
+
+# We changed the name to align it with the new naming scheme. Still, `convert_image_dtype` is
+# prevalent and well understood. Thus, we just alias it without deprecating the old name.
+convert_image_dtype = convert_dtype_image_tensor
+
+
+def convert_dtype_video(video: torch.Tensor, dtype: torch.dtype = torch.float) -> torch.Tensor:
+    return convert_dtype_image_tensor(video, dtype)
+
+
+def convert_dtype(
+    inpt: Union[features.ImageTypeJIT, features.VideoTypeJIT], dtype: torch.dtype = torch.float
+) -> torch.Tensor:
+    if isinstance(inpt, torch.Tensor) and (
+        torch.jit.is_scripting() or not isinstance(inpt, (features.Image, features.Video))
+    ):
+        return convert_dtype_image_tensor(inpt, dtype)
+    elif isinstance(inpt, features.Image):
+        output = convert_dtype_image_tensor(inpt.as_subclass(torch.Tensor), dtype)
+        return features.Image.wrap_like(inpt, output)
+    else:  # isinstance(inpt, features.Video):
+        output = convert_dtype_video(inpt.as_subclass(torch.Tensor), dtype)
+        return features.Video.wrap_like(inpt, output)