From 4767fef93d69ed3c0de4773f171dee2f744360ef Mon Sep 17 00:00:00 2001
From: vfdev-5 <vfdev.5@gmail.com>
Date: Mon, 3 Oct 2022 10:11:43 +0000
Subject: [PATCH 01/11] [proto] Reduce number of calls of __torch_function__

---
 torchvision/prototype/features/_feature.py | 25 ++++++++++++++++++++++
 torchvision/prototype/features/_image.py   |  7 +++---
 2 files changed, 29 insertions(+), 3 deletions(-)

diff --git a/torchvision/prototype/features/_feature.py b/torchvision/prototype/features/_feature.py
index 9c0cece15be..9501ec7851d 100644
--- a/torchvision/prototype/features/_feature.py
+++ b/torchvision/prototype/features/_feature.py
@@ -50,6 +50,9 @@ def new_like(
         requires_grad: Optional[bool] = None,
         **kwargs: Any,
     ) -> F:
+        # Quick fix: Feature -> Tensor => won't go to __torch_function__
+        other = other.as_subclass(torch.Tensor)
+
         return cls(
             data,
             dtype=dtype if dtype is not None else other.dtype,
@@ -137,6 +140,28 @@ def _F(self) -> ModuleType:
             _Feature.__F = functional
         return _Feature.__F
 
+    # Add properties for common attributes like shape, dtype, device, ndim etc
+    # this way we return the result without passing into __torch_function__
+    @property
+    def shape(self):
+        return self.as_subclass(torch.Tensor).shape
+
+    @property
+    def ndim(self):
+        return self.as_subclass(torch.Tensor).ndim
+
+    @property
+    def device(self):
+        return self.as_subclass(torch.Tensor).device
+
+    @property
+    def dtype(self):
+        return self.as_subclass(torch.Tensor).dtype
+
+    @property
+    def requires_grad(self):
+        return self.as_subclass(torch.Tensor).requires_grad
+
     def horizontal_flip(self) -> _Feature:
         return self
 
diff --git a/torchvision/prototype/features/_image.py b/torchvision/prototype/features/_image.py
index 21126c7f254..0a1fcf7eefd 100644
--- a/torchvision/prototype/features/_image.py
+++ b/torchvision/prototype/features/_image.py
@@ -97,9 +97,10 @@ def __repr__(self, *, tensor_contents: Any = None) -> str:  # type: ignore[overr
     def new_like(
         cls, other: Image, data: Any, *, color_space: Optional[Union[ColorSpace, str]] = None, **kwargs: Any
     ) -> Image:
-        return super().new_like(
-            other, data, color_space=color_space if color_space is not None else other.color_space, **kwargs
-        )
+        # Question: Is it safe to assume data to be a tensor ?
+        out = data.as_subclass(Image)
+        out.color_space = color_space if color_space is not None else other.color_space
+        return out
 
     @property
     def image_size(self) -> Tuple[int, int]:

From 5ba3c0d58ff1d9acf498b60f47101651b602a655 Mon Sep 17 00:00:00 2001
From: vfdev-5 <vfdev.5@gmail.com>
Date: Wed, 5 Oct 2022 11:05:24 +0000
Subject: [PATCH 02/11] Use DisableTorchFunction and super

---
 torchvision/prototype/features/_feature.py | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/torchvision/prototype/features/_feature.py b/torchvision/prototype/features/_feature.py
index 9501ec7851d..926a8bd50a3 100644
--- a/torchvision/prototype/features/_feature.py
+++ b/torchvision/prototype/features/_feature.py
@@ -144,23 +144,28 @@ def _F(self) -> ModuleType:
     # this way we return the result without passing into __torch_function__
     @property
     def shape(self):
-        return self.as_subclass(torch.Tensor).shape
+        with DisableTorchFunction():
+            return super().shape
 
     @property
     def ndim(self):
-        return self.as_subclass(torch.Tensor).ndim
+        with DisableTorchFunction():
+            return super().ndim
 
     @property
     def device(self):
-        return self.as_subclass(torch.Tensor).device
+        with DisableTorchFunction():
+            return super().device
 
     @property
     def dtype(self):
-        return self.as_subclass(torch.Tensor).dtype
+        with DisableTorchFunction():
+            return super().dtype
 
     @property
     def requires_grad(self):
-        return self.as_subclass(torch.Tensor).requires_grad
+        with DisableTorchFunction():
+            return super().requires_grad
 
     def horizontal_flip(self) -> _Feature:
         return self

From 6435c095aaf48c8e157082f6d39be054eb0b32e1 Mon Sep 17 00:00:00 2001
From: vfdev-5 <vfdev.5@gmail.com>
Date: Wed, 5 Oct 2022 11:10:28 +0000
Subject: [PATCH 03/11] Use self._tensor

---
 torchvision/prototype/features/_feature.py | 30 +++++++++-------------
 torchvision/prototype/features/_image.py   |  4 +++
 2 files changed, 16 insertions(+), 18 deletions(-)

diff --git a/torchvision/prototype/features/_feature.py b/torchvision/prototype/features/_feature.py
index 926a8bd50a3..746061f20a2 100644
--- a/torchvision/prototype/features/_feature.py
+++ b/torchvision/prototype/features/_feature.py
@@ -29,15 +29,14 @@ def __new__(
         device: Optional[Union[torch.device, str, int]] = None,
         requires_grad: bool = False,
     ) -> F:
-        return (
-            torch.as_tensor(  # type: ignore[return-value]
-                data,
-                dtype=dtype,  # type: ignore[arg-type]
-                device=device,  # type: ignore[arg-type]
-            )
-            .as_subclass(cls)  # type: ignore[arg-type]
-            .requires_grad_(requires_grad)
+        tensor = torch.as_tensor(  # type: ignore[return-value]
+            data,
+            dtype=dtype,  # type: ignore[arg-type]
+            device=device,  # type: ignore[arg-type]
         )
+        output = tensor.as_subclass(cls).requires_grad_(requires_grad)  # type: ignore[arg-type]
+        output._tensor = tensor
+        return output
 
     @classmethod
     def new_like(
@@ -144,28 +143,23 @@ def _F(self) -> ModuleType:
     # this way we return the result without passing into __torch_function__
     @property
     def shape(self):
-        with DisableTorchFunction():
-            return super().shape
+        return self._tensor.shape
 
     @property
     def ndim(self):
-        with DisableTorchFunction():
-            return super().ndim
+        return self._tensor.ndim
 
     @property
     def device(self):
-        with DisableTorchFunction():
-            return super().device
+        return self._tensor.device
 
     @property
     def dtype(self):
-        with DisableTorchFunction():
-            return super().dtype
+        return self._tensor.dtype
 
     @property
     def requires_grad(self):
-        with DisableTorchFunction():
-            return super().requires_grad
+        return self._tensor.requires_grad
 
     def horizontal_flip(self) -> _Feature:
         return self
diff --git a/torchvision/prototype/features/_image.py b/torchvision/prototype/features/_image.py
index 0a1fcf7eefd..eaa7a167271 100644
--- a/torchvision/prototype/features/_image.py
+++ b/torchvision/prototype/features/_image.py
@@ -100,7 +100,11 @@ def new_like(
         # Question: Is it safe to assume data to be a tensor ?
         out = data.as_subclass(Image)
         out.color_space = color_space if color_space is not None else other.color_space
+        out._tensor = data
         return out
+        # return super().new_like(
+        #     other, data, color_space=color_space if color_space is not None else other.color_space, **kwargs
+        # )
 
     @property
     def image_size(self) -> Tuple[int, int]:

From b7694e6287ffe9406114b1a4d4c2b5c6ddacb12d Mon Sep 17 00:00:00 2001
From: vfdev-5 <vfdev.5@gmail.com>
Date: Wed, 5 Oct 2022 16:58:20 +0000
Subject: [PATCH 04/11] Fixes mypy and color space handling

---
 torchvision/prototype/features/_feature.py | 29 ++++++++++---------
 torchvision/prototype/features/_image.py   | 33 ++++++++++++----------
 2 files changed, 33 insertions(+), 29 deletions(-)

diff --git a/torchvision/prototype/features/_feature.py b/torchvision/prototype/features/_feature.py
index 746061f20a2..d0c73e46cf8 100644
--- a/torchvision/prototype/features/_feature.py
+++ b/torchvision/prototype/features/_feature.py
@@ -6,6 +6,7 @@
 import PIL.Image
 import torch
 from torch._C import DisableTorchFunction
+from torch.types import _device, _dtype, _size
 from torchvision.transforms import InterpolationMode
 
 
@@ -29,14 +30,14 @@ def __new__(
         device: Optional[Union[torch.device, str, int]] = None,
         requires_grad: bool = False,
     ) -> F:
-        tensor = torch.as_tensor(  # type: ignore[return-value]
+        tensor = torch.as_tensor(
             data,
             dtype=dtype,  # type: ignore[arg-type]
             device=device,  # type: ignore[arg-type]
         )
         output = tensor.as_subclass(cls).requires_grad_(requires_grad)  # type: ignore[arg-type]
-        output._tensor = tensor
-        return output
+        output._tensor = tensor  # type: ignore[attr-defined]
+        return output  # type: ignore[return-value]
 
     @classmethod
     def new_like(
@@ -50,7 +51,7 @@ def new_like(
         **kwargs: Any,
     ) -> F:
         # Quick fix: Feature -> Tensor => won't go to __torch_function__
-        other = other.as_subclass(torch.Tensor)
+        other = other.as_subclass(torch.Tensor)  # type: ignore[arg-type]
 
         return cls(
             data,
@@ -142,24 +143,24 @@ def _F(self) -> ModuleType:
     # Add properties for common attributes like shape, dtype, device, ndim etc
     # this way we return the result without passing into __torch_function__
     @property
-    def shape(self):
-        return self._tensor.shape
+    def shape(self) -> _size:  # type: ignore[override]
+        return self._tensor.shape  # type: ignore[attr-defined, no-any-return]
 
     @property
-    def ndim(self):
-        return self._tensor.ndim
+    def ndim(self) -> int:  # type: ignore[override]
+        return self._tensor.ndim  # type: ignore[attr-defined, no-any-return]
 
     @property
-    def device(self):
-        return self._tensor.device
+    def device(self, *args: Any, **kwargs: Any) -> _device:  # type: ignore[override]
+        return self._tensor.device  # type: ignore[attr-defined, no-any-return]
 
     @property
-    def dtype(self):
-        return self._tensor.dtype
+    def dtype(self) -> _dtype:  # type: ignore[override]
+        return self._tensor.dtype  # type: ignore[attr-defined, no-any-return]
 
     @property
-    def requires_grad(self):
-        return self._tensor.requires_grad
+    def requires_grad(self) -> bool:  # type: ignore[override]
+        return self._tensor.requires_grad  # type: ignore[attr-defined, no-any-return]
 
     def horizontal_flip(self) -> _Feature:
         return self
diff --git a/torchvision/prototype/features/_image.py b/torchvision/prototype/features/_image.py
index eaa7a167271..9357d01ab6e 100644
--- a/torchvision/prototype/features/_image.py
+++ b/torchvision/prototype/features/_image.py
@@ -59,6 +59,18 @@ def _from_tensor_shape(shape: List[int]) -> ColorSpace:
         return ColorSpace.OTHER
 
 
+def _setup_color_space(color_space: Union[None, ColorSpace, str], shape: List[int]) -> ColorSpace:
+    if color_space is None:
+        color_space = ColorSpace.from_tensor_shape(shape)
+        if color_space == ColorSpace.OTHER:
+            warnings.warn("Unable to guess a specific color space. Consider passing it explicitly.")
+        return color_space
+    elif isinstance(color_space, str):
+        return ColorSpace.from_str(color_space.upper())
+
+    raise ValueError
+
+
 class Image(_Feature):
     color_space: ColorSpace
 
@@ -78,15 +90,7 @@ def __new__(
             data = data.unsqueeze(0)
         image = super().__new__(cls, data, requires_grad=requires_grad)
 
-        if color_space is None:
-            color_space = ColorSpace.from_tensor_shape(image.shape)  # type: ignore[arg-type]
-            if color_space == ColorSpace.OTHER:
-                warnings.warn("Unable to guess a specific color space. Consider passing it explicitly.")
-        elif isinstance(color_space, str):
-            color_space = ColorSpace.from_str(color_space.upper())
-        elif not isinstance(color_space, ColorSpace):
-            raise ValueError
-        image.color_space = color_space
+        image.color_space = _setup_color_space(color_space, list(image.shape))
 
         return image
 
@@ -98,13 +102,12 @@ def new_like(
         cls, other: Image, data: Any, *, color_space: Optional[Union[ColorSpace, str]] = None, **kwargs: Any
     ) -> Image:
         # Question: Is it safe to assume data to be a tensor ?
-        out = data.as_subclass(Image)
-        out.color_space = color_space if color_space is not None else other.color_space
-        out._tensor = data
+        out: Image = data.as_subclass(Image)
+        out.color_space = _setup_color_space(
+            color_space if color_space is not None else other.color_space, list(data.shape)
+        )
+        out._tensor = data  # type: ignore[attr-defined]
         return out
-        # return super().new_like(
-        #     other, data, color_space=color_space if color_space is not None else other.color_space, **kwargs
-        # )
 
     @property
     def image_size(self) -> Tuple[int, int]:

From 628826c9901075755834904b44469fbad13848f7 Mon Sep 17 00:00:00 2001
From: vfdev-5 <vfdev.5@gmail.com>
Date: Thu, 6 Oct 2022 12:06:51 +0000
Subject: [PATCH 05/11] revert Image.new_like

---
 torchvision/prototype/features/_image.py | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)

diff --git a/torchvision/prototype/features/_image.py b/torchvision/prototype/features/_image.py
index 9357d01ab6e..ffdff2694a6 100644
--- a/torchvision/prototype/features/_image.py
+++ b/torchvision/prototype/features/_image.py
@@ -60,7 +60,9 @@ def _from_tensor_shape(shape: List[int]) -> ColorSpace:
 
 
 def _setup_color_space(color_space: Union[None, ColorSpace, str], shape: List[int]) -> ColorSpace:
-    if color_space is None:
+    if isinstance(color_space, ColorSpace):
+        return color_space
+    elif color_space is None:
         color_space = ColorSpace.from_tensor_shape(shape)
         if color_space == ColorSpace.OTHER:
             warnings.warn("Unable to guess a specific color space. Consider passing it explicitly.")
@@ -68,7 +70,7 @@ def _setup_color_space(color_space: Union[None, ColorSpace, str], shape: List[in
     elif isinstance(color_space, str):
         return ColorSpace.from_str(color_space.upper())
 
-    raise ValueError
+    raise ValueError(f"Unsupported color space '{color_space}'")
 
 
 class Image(_Feature):
@@ -101,13 +103,9 @@ def __repr__(self, *, tensor_contents: Any = None) -> str:  # type: ignore[overr
     def new_like(
         cls, other: Image, data: Any, *, color_space: Optional[Union[ColorSpace, str]] = None, **kwargs: Any
     ) -> Image:
-        # Question: Is it safe to assume data to be a tensor ?
-        out: Image = data.as_subclass(Image)
-        out.color_space = _setup_color_space(
-            color_space if color_space is not None else other.color_space, list(data.shape)
+        return super().new_like(
+            other, data, color_space=color_space if color_space is not None else other.color_space, **kwargs
         )
-        out._tensor = data  # type: ignore[attr-defined]
-        return out
 
     @property
     def image_size(self) -> Tuple[int, int]:

From 345790be347cf19fe4b15ffc6c4f7c2b23f3b621 Mon Sep 17 00:00:00 2001
From: vfdev-5 <vfdev.5@gmail.com>
Date: Fri, 7 Oct 2022 10:00:39 +0000
Subject: [PATCH 06/11] WIP

---
 torchvision/prototype/features/_feature.py |  1 -
 torchvision/prototype/features/_image.py   | 22 +++++++++++++++++++---
 2 files changed, 19 insertions(+), 4 deletions(-)

diff --git a/torchvision/prototype/features/_feature.py b/torchvision/prototype/features/_feature.py
index d0c73e46cf8..992831e855b 100644
--- a/torchvision/prototype/features/_feature.py
+++ b/torchvision/prototype/features/_feature.py
@@ -52,7 +52,6 @@ def new_like(
     ) -> F:
         # Quick fix: Feature -> Tensor => won't go to __torch_function__
         other = other.as_subclass(torch.Tensor)  # type: ignore[arg-type]
-
         return cls(
             data,
             dtype=dtype if dtype is not None else other.dtype,
diff --git a/torchvision/prototype/features/_image.py b/torchvision/prototype/features/_image.py
index ffdff2694a6..88b173fd302 100644
--- a/torchvision/prototype/features/_image.py
+++ b/torchvision/prototype/features/_image.py
@@ -103,9 +103,25 @@ def __repr__(self, *, tensor_contents: Any = None) -> str:  # type: ignore[overr
     def new_like(
         cls, other: Image, data: Any, *, color_space: Optional[Union[ColorSpace, str]] = None, **kwargs: Any
     ) -> Image:
-        return super().new_like(
-            other, data, color_space=color_space if color_space is not None else other.color_space, **kwargs
-        )
+        # Question: Is it safe to assume data to be a tensor ?
+        out = data.as_subclass(Image)
+        out.color_space = color_space if color_space is not None else other.color_space
+        out._tensor = data  # type: ignore[attr-defined]
+        return out
+
+    @classmethod
+    def _wrap(tensor, color_space=None):
+        image = tensor.as_subclass(Image)  # type: ignore[arg-type]
+        image.color_space = _setup_color_space(color_space, list(image.shape))
+        return image
+
+    # @classmethod
+    # def new_like(
+    #     cls, other: Image, data: Any, *, color_space: Optional[Union[ColorSpace, str]] = None, **kwargs: Any
+    # ) -> Image:
+    #     return super().new_like(
+    #         other, data, color_space=color_space if color_space is not None else other.color_space, **kwargs
+    #     )
 
     @property
     def image_size(self) -> Tuple[int, int]:

From db0eef5fc794be9bb8180f7edd0596d92728e0e0 Mon Sep 17 00:00:00 2001
From: vfdev-5 <vfdev.5@gmail.com>
Date: Mon, 10 Oct 2022 10:36:52 +0000
Subject: [PATCH 07/11] Perf opt with ref to tensor and properties

---
 torchvision/prototype/features/_bounding_box.py |  1 +
 torchvision/prototype/features/_encoded.py      |  4 +++-
 torchvision/prototype/features/_feature.py      |  8 ++++++--
 torchvision/prototype/features/_image.py        | 15 +--------------
 torchvision/prototype/features/_label.py        |  1 +
 torchvision/prototype/features/_mask.py         |  4 +++-
 torchvision/prototype/features/_video.py        |  7 ++++---
 7 files changed, 19 insertions(+), 21 deletions(-)

diff --git a/torchvision/prototype/features/_bounding_box.py b/torchvision/prototype/features/_bounding_box.py
index 7b69af5f9bb..4204f12aad2 100644
--- a/torchvision/prototype/features/_bounding_box.py
+++ b/torchvision/prototype/features/_bounding_box.py
@@ -24,6 +24,7 @@ def _wrap(cls, tensor: torch.Tensor, *, format: BoundingBoxFormat, image_size: T
         bounding_box = tensor.as_subclass(cls)
         bounding_box.format = format
         bounding_box.image_size = image_size
+        bounding_box._tensor = tensor
         return bounding_box
 
     def __new__(
diff --git a/torchvision/prototype/features/_encoded.py b/torchvision/prototype/features/_encoded.py
index 4b963986b4f..8344ea850ae 100644
--- a/torchvision/prototype/features/_encoded.py
+++ b/torchvision/prototype/features/_encoded.py
@@ -16,7 +16,9 @@
 class EncodedData(_Feature):
     @classmethod
     def _wrap(cls: Type[D], tensor: torch.Tensor) -> D:
-        return tensor.as_subclass(cls)
+        output = tensor.as_subclass(cls)
+        output._tensor = tensor
+        return output
 
     def __new__(
         cls,
diff --git a/torchvision/prototype/features/_feature.py b/torchvision/prototype/features/_feature.py
index e4c7b88ed5b..0f342113b59 100644
--- a/torchvision/prototype/features/_feature.py
+++ b/torchvision/prototype/features/_feature.py
@@ -43,14 +43,18 @@ def __new__(
         requires_grad: bool = False,
     ) -> _Feature:
         tensor = cls._to_tensor(data, dtype=dtype, device=device, requires_grad=requires_grad)
-        return tensor.as_subclass(_Feature)
+        output = tensor.as_subclass(_Feature)
+        output._tensor = tensor
+        return output
 
     @classmethod
     def wrap_like(cls: Type[F], other: F, tensor: torch.Tensor) -> F:
         # FIXME: this is just here for BC with the prototype datasets. See __new__ for details. If that is resolved,
         #  this method should be made abstract
         # raise NotImplementedError
-        return tensor.as_subclass(cls)
+        output = tensor.as_subclass(cls)
+        output._tensor = tensor
+        return output
 
     _NO_WRAPPING_EXCEPTIONS = {
         torch.Tensor.clone: lambda cls, input, output: cls.wrap_like(input, output),
diff --git a/torchvision/prototype/features/_image.py b/torchvision/prototype/features/_image.py
index c0c9326ec27..1936aae3f6c 100644
--- a/torchvision/prototype/features/_image.py
+++ b/torchvision/prototype/features/_image.py
@@ -59,20 +59,6 @@ def _from_tensor_shape(shape: List[int]) -> ColorSpace:
         return ColorSpace.OTHER
 
 
-def _setup_color_space(color_space: Union[None, ColorSpace, str], shape: List[int]) -> ColorSpace:
-    if isinstance(color_space, ColorSpace):
-        return color_space
-    elif color_space is None:
-        color_space = ColorSpace.from_tensor_shape(shape)
-        if color_space == ColorSpace.OTHER:
-            warnings.warn("Unable to guess a specific color space. Consider passing it explicitly.")
-        return color_space
-    elif isinstance(color_space, str):
-        return ColorSpace.from_str(color_space.upper())
-
-    raise ValueError(f"Unsupported color space '{color_space}'")
-
-
 class Image(_Feature):
     color_space: ColorSpace
 
@@ -80,6 +66,7 @@ class Image(_Feature):
     def _wrap(cls, tensor: torch.Tensor, *, color_space: ColorSpace) -> Image:
         image = tensor.as_subclass(cls)
         image.color_space = color_space
+        image._tensor = tensor
         return image
 
     def __new__(
diff --git a/torchvision/prototype/features/_label.py b/torchvision/prototype/features/_label.py
index 9c2bcfc0fb1..7743ef4970a 100644
--- a/torchvision/prototype/features/_label.py
+++ b/torchvision/prototype/features/_label.py
@@ -18,6 +18,7 @@ class _LabelBase(_Feature):
     def _wrap(cls: Type[L], tensor: torch.Tensor, *, categories: Optional[Sequence[str]]) -> L:
         label_base = tensor.as_subclass(cls)
         label_base.categories = categories
+        label_base._tensor = tensor
         return label_base
 
     def __new__(
diff --git a/torchvision/prototype/features/_mask.py b/torchvision/prototype/features/_mask.py
index 7b49ce8e85e..fb316f5b054 100644
--- a/torchvision/prototype/features/_mask.py
+++ b/torchvision/prototype/features/_mask.py
@@ -11,7 +11,9 @@
 class Mask(_Feature):
     @classmethod
     def _wrap(cls, tensor: torch.Tensor) -> Mask:
-        return tensor.as_subclass(cls)
+        output = tensor.as_subclass(cls)
+        output._tensor = tensor
+        return output
 
     def __new__(
         cls,
diff --git a/torchvision/prototype/features/_video.py b/torchvision/prototype/features/_video.py
index a58027243cf..dd663de8384 100644
--- a/torchvision/prototype/features/_video.py
+++ b/torchvision/prototype/features/_video.py
@@ -15,9 +15,10 @@ class Video(_Feature):
 
     @classmethod
     def _wrap(cls, tensor: torch.Tensor, *, color_space: ColorSpace) -> Video:
-        image = tensor.as_subclass(cls)
-        image.color_space = color_space
-        return image
+        video = tensor.as_subclass(cls)
+        video.color_space = color_space
+        video._tensor = tensor
+        return video
 
     def __new__(
         cls,

From c8b3ac85c2dc55092b86c10f4732ba11407a90fd Mon Sep 17 00:00:00 2001
From: vfdev-5 <vfdev.5@gmail.com>
Date: Mon, 10 Oct 2022 12:35:28 +0000
Subject: [PATCH 08/11] Removed requires_grad property

---
 torchvision/prototype/features/_feature.py | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/torchvision/prototype/features/_feature.py b/torchvision/prototype/features/_feature.py
index 0f342113b59..695a802154b 100644
--- a/torchvision/prototype/features/_feature.py
+++ b/torchvision/prototype/features/_feature.py
@@ -151,10 +151,6 @@ def device(self, *args: Any, **kwargs: Any) -> _device:  # type: ignore[override
     def dtype(self) -> _dtype:  # type: ignore[override]
         return self._tensor.dtype  # type: ignore[attr-defined, no-any-return]
 
-    @property
-    def requires_grad(self) -> bool:  # type: ignore[override]
-        return self._tensor.requires_grad  # type: ignore[attr-defined, no-any-return]
-
     def horizontal_flip(self) -> _Feature:
         return self
 

From 38f8e21242830fed46ddf31287edb67c1abd124a Mon Sep 17 00:00:00 2001
From: vfdev-5 <vfdev.5@gmail.com>
Date: Fri, 14 Oct 2022 15:40:24 +0000
Subject: [PATCH 09/11] Use _tensor ref

---
 torchvision/prototype/features/_bounding_box.py |  1 +
 torchvision/prototype/features/_encoded.py      |  4 +++-
 torchvision/prototype/features/_feature.py      | 17 ++++++++---------
 torchvision/prototype/features/_image.py        |  1 +
 torchvision/prototype/features/_label.py        |  1 +
 torchvision/prototype/features/_mask.py         |  4 +++-
 torchvision/prototype/features/_video.py        |  1 +
 7 files changed, 18 insertions(+), 11 deletions(-)

diff --git a/torchvision/prototype/features/_bounding_box.py b/torchvision/prototype/features/_bounding_box.py
index 18c607d4d16..db9af91c89f 100644
--- a/torchvision/prototype/features/_bounding_box.py
+++ b/torchvision/prototype/features/_bounding_box.py
@@ -24,6 +24,7 @@ def _wrap(cls, tensor: torch.Tensor, *, format: BoundingBoxFormat, spatial_size:
         bounding_box = tensor.as_subclass(cls)
         bounding_box.format = format
         bounding_box.spatial_size = spatial_size
+        bounding_box._tensor = tensor
         return bounding_box
 
     def __new__(
diff --git a/torchvision/prototype/features/_encoded.py b/torchvision/prototype/features/_encoded.py
index ffa347a3ef6..390263d3aea 100644
--- a/torchvision/prototype/features/_encoded.py
+++ b/torchvision/prototype/features/_encoded.py
@@ -16,7 +16,9 @@
 class EncodedData(_Feature):
     @classmethod
     def _wrap(cls: Type[D], tensor: torch.Tensor) -> D:
-        return tensor.as_subclass(cls)
+        output = tensor.as_subclass(cls)
+        output._tensor = tensor
+        return output
 
     def __new__(
         cls,
diff --git a/torchvision/prototype/features/_feature.py b/torchvision/prototype/features/_feature.py
index ec6db649025..834fea426d4 100644
--- a/torchvision/prototype/features/_feature.py
+++ b/torchvision/prototype/features/_feature.py
@@ -44,6 +44,7 @@ def __new__(
     ) -> _Feature:
         tensor = cls._to_tensor(data, dtype=dtype, device=device, requires_grad=requires_grad)
         output = tensor.as_subclass(_Feature)
+        output._tensor = tensor
         return output
 
     @classmethod
@@ -108,7 +109,9 @@ def __torch_function__(
             # Inplace `func`'s, canonically identified with a trailing underscore in their name like `.add_(...)`,
             # will retain the input type. Thus, we need to unwrap here.
             if isinstance(output, cls):
-                return output.as_subclass(torch.Tensor)
+                tensor = output.as_subclass(torch.Tensor)
+                output._tensor = tensor
+                return tensor
 
             return output
 
@@ -134,23 +137,19 @@ def _F(self) -> ModuleType:
     # this way we return the result without passing into __torch_function__
     @property
     def shape(self) -> _size:  # type: ignore[override]
-        with DisableTorchFunction():
-            return super().shape
+        return self._tensor.shape
 
     @property
     def ndim(self) -> int:  # type: ignore[override]
-        with DisableTorchFunction():
-            return super().ndim
+        return self._tensor.ndim
 
     @property
     def device(self, *args: Any, **kwargs: Any) -> _device:  # type: ignore[override]
-        with DisableTorchFunction():
-            return super().device
+        return self._tensor.device
 
     @property
     def dtype(self) -> _dtype:  # type: ignore[override]
-        with DisableTorchFunction():
-            return super().dtype
+        return self._tensor.dtype
 
     def horizontal_flip(self) -> _Feature:
         return self
diff --git a/torchvision/prototype/features/_image.py b/torchvision/prototype/features/_image.py
index e9128b94be0..224f1978611 100644
--- a/torchvision/prototype/features/_image.py
+++ b/torchvision/prototype/features/_image.py
@@ -64,6 +64,7 @@ class Image(_Feature):
     def _wrap(cls, tensor: torch.Tensor, *, color_space: ColorSpace) -> Image:
         image = tensor.as_subclass(cls)
         image.color_space = color_space
+        image._tensor = tensor
         return image
 
     def __new__(
diff --git a/torchvision/prototype/features/_label.py b/torchvision/prototype/features/_label.py
index 9c2bcfc0fb1..7743ef4970a 100644
--- a/torchvision/prototype/features/_label.py
+++ b/torchvision/prototype/features/_label.py
@@ -18,6 +18,7 @@ class _LabelBase(_Feature):
     def _wrap(cls: Type[L], tensor: torch.Tensor, *, categories: Optional[Sequence[str]]) -> L:
         label_base = tensor.as_subclass(cls)
         label_base.categories = categories
+        label_base._tensor = tensor
         return label_base
 
     def __new__(
diff --git a/torchvision/prototype/features/_mask.py b/torchvision/prototype/features/_mask.py
index 2da10195e80..35f9f67a5bd 100644
--- a/torchvision/prototype/features/_mask.py
+++ b/torchvision/prototype/features/_mask.py
@@ -11,7 +11,9 @@
 class Mask(_Feature):
     @classmethod
     def _wrap(cls, tensor: torch.Tensor) -> Mask:
-        return tensor.as_subclass(cls)
+        output = tensor.as_subclass(cls)
+        output._tensor = tensor
+        return output
 
     def __new__(
         cls,
diff --git a/torchvision/prototype/features/_video.py b/torchvision/prototype/features/_video.py
index 26f97549ac5..54c20b2d8b4 100644
--- a/torchvision/prototype/features/_video.py
+++ b/torchvision/prototype/features/_video.py
@@ -17,6 +17,7 @@ class Video(_Feature):
     def _wrap(cls, tensor: torch.Tensor, *, color_space: ColorSpace) -> Video:
         video = tensor.as_subclass(cls)
         video.color_space = color_space
+        video._tensor = tensor
         return video
 
     def __new__(

From 8441dbc247418aedc27bb5f015d0bacd89b479da Mon Sep 17 00:00:00 2001
From: vfdev-5 <vfdev.5@gmail.com>
Date: Fri, 14 Oct 2022 15:59:46 +0000
Subject: [PATCH 10/11] Revert "Use _tensor ref"

This reverts commit 38f8e21242830fed46ddf31287edb67c1abd124a.
---
 torchvision/prototype/features/_bounding_box.py |  1 -
 torchvision/prototype/features/_encoded.py      |  4 +---
 torchvision/prototype/features/_feature.py      | 17 +++++++++--------
 torchvision/prototype/features/_image.py        |  1 -
 torchvision/prototype/features/_label.py        |  1 -
 torchvision/prototype/features/_mask.py         |  4 +---
 torchvision/prototype/features/_video.py        |  1 -
 7 files changed, 11 insertions(+), 18 deletions(-)

diff --git a/torchvision/prototype/features/_bounding_box.py b/torchvision/prototype/features/_bounding_box.py
index db9af91c89f..18c607d4d16 100644
--- a/torchvision/prototype/features/_bounding_box.py
+++ b/torchvision/prototype/features/_bounding_box.py
@@ -24,7 +24,6 @@ def _wrap(cls, tensor: torch.Tensor, *, format: BoundingBoxFormat, spatial_size:
         bounding_box = tensor.as_subclass(cls)
         bounding_box.format = format
         bounding_box.spatial_size = spatial_size
-        bounding_box._tensor = tensor
         return bounding_box
 
     def __new__(
diff --git a/torchvision/prototype/features/_encoded.py b/torchvision/prototype/features/_encoded.py
index 390263d3aea..ffa347a3ef6 100644
--- a/torchvision/prototype/features/_encoded.py
+++ b/torchvision/prototype/features/_encoded.py
@@ -16,9 +16,7 @@
 class EncodedData(_Feature):
     @classmethod
     def _wrap(cls: Type[D], tensor: torch.Tensor) -> D:
-        output = tensor.as_subclass(cls)
-        output._tensor = tensor
-        return output
+        return tensor.as_subclass(cls)
 
     def __new__(
         cls,
diff --git a/torchvision/prototype/features/_feature.py b/torchvision/prototype/features/_feature.py
index 834fea426d4..ec6db649025 100644
--- a/torchvision/prototype/features/_feature.py
+++ b/torchvision/prototype/features/_feature.py
@@ -44,7 +44,6 @@ def __new__(
     ) -> _Feature:
         tensor = cls._to_tensor(data, dtype=dtype, device=device, requires_grad=requires_grad)
         output = tensor.as_subclass(_Feature)
-        output._tensor = tensor
         return output
 
     @classmethod
@@ -109,9 +108,7 @@ def __torch_function__(
             # Inplace `func`'s, canonically identified with a trailing underscore in their name like `.add_(...)`,
             # will retain the input type. Thus, we need to unwrap here.
             if isinstance(output, cls):
-                tensor = output.as_subclass(torch.Tensor)
-                output._tensor = tensor
-                return tensor
+                return output.as_subclass(torch.Tensor)
 
             return output
 
@@ -137,19 +134,23 @@ def _F(self) -> ModuleType:
     # this way we return the result without passing into __torch_function__
     @property
     def shape(self) -> _size:  # type: ignore[override]
-        return self._tensor.shape
+        with DisableTorchFunction():
+            return super().shape
 
     @property
     def ndim(self) -> int:  # type: ignore[override]
-        return self._tensor.ndim
+        with DisableTorchFunction():
+            return super().ndim
 
     @property
     def device(self, *args: Any, **kwargs: Any) -> _device:  # type: ignore[override]
-        return self._tensor.device
+        with DisableTorchFunction():
+            return super().device
 
     @property
     def dtype(self) -> _dtype:  # type: ignore[override]
-        return self._tensor.dtype
+        with DisableTorchFunction():
+            return super().dtype
 
     def horizontal_flip(self) -> _Feature:
         return self
diff --git a/torchvision/prototype/features/_image.py b/torchvision/prototype/features/_image.py
index 224f1978611..e9128b94be0 100644
--- a/torchvision/prototype/features/_image.py
+++ b/torchvision/prototype/features/_image.py
@@ -64,7 +64,6 @@ class Image(_Feature):
     def _wrap(cls, tensor: torch.Tensor, *, color_space: ColorSpace) -> Image:
         image = tensor.as_subclass(cls)
         image.color_space = color_space
-        image._tensor = tensor
         return image
 
     def __new__(
diff --git a/torchvision/prototype/features/_label.py b/torchvision/prototype/features/_label.py
index 7743ef4970a..9c2bcfc0fb1 100644
--- a/torchvision/prototype/features/_label.py
+++ b/torchvision/prototype/features/_label.py
@@ -18,7 +18,6 @@ class _LabelBase(_Feature):
     def _wrap(cls: Type[L], tensor: torch.Tensor, *, categories: Optional[Sequence[str]]) -> L:
         label_base = tensor.as_subclass(cls)
         label_base.categories = categories
-        label_base._tensor = tensor
         return label_base
 
     def __new__(
diff --git a/torchvision/prototype/features/_mask.py b/torchvision/prototype/features/_mask.py
index 35f9f67a5bd..2da10195e80 100644
--- a/torchvision/prototype/features/_mask.py
+++ b/torchvision/prototype/features/_mask.py
@@ -11,9 +11,7 @@
 class Mask(_Feature):
     @classmethod
     def _wrap(cls, tensor: torch.Tensor) -> Mask:
-        output = tensor.as_subclass(cls)
-        output._tensor = tensor
-        return output
+        return tensor.as_subclass(cls)
 
     def __new__(
         cls,
diff --git a/torchvision/prototype/features/_video.py b/torchvision/prototype/features/_video.py
index 54c20b2d8b4..26f97549ac5 100644
--- a/torchvision/prototype/features/_video.py
+++ b/torchvision/prototype/features/_video.py
@@ -17,7 +17,6 @@ class Video(_Feature):
     def _wrap(cls, tensor: torch.Tensor, *, color_space: ColorSpace) -> Video:
         video = tensor.as_subclass(cls)
         video.color_space = color_space
-        video._tensor = tensor
         return video
 
     def __new__(

From 7e8eb46c75a32c35450dc6705c78045526cb6402 Mon Sep 17 00:00:00 2001
From: vfdev <vfdev.5@gmail.com>
Date: Mon, 17 Oct 2022 09:31:02 +0200
Subject: [PATCH 11/11] Update torchvision/prototype/features/_feature.py

Co-authored-by: Philip Meier <github.pmeier@posteo.de>
---
 torchvision/prototype/features/_feature.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/torchvision/prototype/features/_feature.py b/torchvision/prototype/features/_feature.py
index ec6db649025..1cc2d8d4bb7 100644
--- a/torchvision/prototype/features/_feature.py
+++ b/torchvision/prototype/features/_feature.py
@@ -43,8 +43,7 @@ def __new__(
         requires_grad: bool = False,
     ) -> _Feature:
         tensor = cls._to_tensor(data, dtype=dtype, device=device, requires_grad=requires_grad)
-        output = tensor.as_subclass(_Feature)
-        return output
+        return tensor.as_subclass(_Feature)
 
     @classmethod
     def wrap_like(cls: Type[F], other: F, tensor: torch.Tensor) -> F: