port RandomZoomOut from detection references to prototype transforms (#5551)

* port RandomZoomOut from detection references to prototype transforms

* copy by default

* move padding parsing to stable API

* merge zoom out kernels into transform

* lint

* fix filling if no padding was done

* resolve todo

Author: pmeier

torchvision/prototype/features/_bounding_box.py

@@ -64,7 +64,7 @@ def to_format(self, format: Union[str, BoundingBoxFormat]) -> BoundingBox:
         from torchvision.prototype.transforms.functional import convert_bounding_box_format
 
         if isinstance(format, str):
-            format = BoundingBoxFormat[format]
+            format = BoundingBoxFormat.from_str(format.upper())
 
         return BoundingBox.new_like(
             self, convert_bounding_box_format(self, old_format=self.format, new_format=format), format=format

torchvision/prototype/transforms/__init__.py

@@ -7,7 +7,16 @@
 from ._augment import RandomErasing, RandomMixup, RandomCutmix
 from ._auto_augment import RandAugment, TrivialAugmentWide, AutoAugment, AugMix
 from ._container import Compose, RandomApply, RandomChoice, RandomOrder
-from ._geometry import HorizontalFlip, Resize, CenterCrop, RandomResizedCrop, FiveCrop, TenCrop, BatchMultiCrop
+from ._geometry import (
+    HorizontalFlip,
+    Resize,
+    CenterCrop,
+    RandomResizedCrop,
+    FiveCrop,
+    TenCrop,
+    BatchMultiCrop,
+    RandomZoomOut,
+)
 from ._meta import ConvertBoundingBoxFormat, ConvertImageDtype, ConvertImageColorSpace
 from ._misc import Identity, Normalize, ToDtype, Lambda
 from ._presets import (

torchvision/prototype/transforms/_geometry.py

@@ -256,3 +256,88 @@ def apply_recursively(obj: Any) -> Any:
                 return obj
 
         return apply_recursively(inputs if len(inputs) > 1 else inputs[0])
+
+
+class RandomZoomOut(Transform):
+    def __init__(
+        self, fill: Union[float, Sequence[float]] = 0.0, side_range: Tuple[float, float] = (1.0, 4.0), p: float = 0.5
+    ) -> None:
+        super().__init__()
+
+        if fill is None:
+            fill = 0.0
+        self.fill = fill
+
+        self.side_range = side_range
+        if side_range[0] < 1.0 or side_range[0] > side_range[1]:
+            raise ValueError(f"Invalid canvas side range provided {side_range}.")
+
+        self.p = p
+
+    def _get_params(self, sample: Any) -> Dict[str, Any]:
+        image = query_image(sample)
+        orig_c, orig_h, orig_w = get_image_dimensions(image)
+
+        r = self.side_range[0] + torch.rand(1) * (self.side_range[1] - self.side_range[0])
+        canvas_width = int(orig_w * r)
+        canvas_height = int(orig_h * r)
+
+        r = torch.rand(2)
+        left = int((canvas_width - orig_w) * r[0])
+        top = int((canvas_height - orig_h) * r[1])
+        right = canvas_width - (left + orig_w)
+        bottom = canvas_height - (top + orig_h)
+        padding = [left, top, right, bottom]
+
+        fill = self.fill
+        if not isinstance(fill, collections.abc.Sequence):
+            fill = [fill] * orig_c
+
+        return dict(padding=padding, fill=fill)
+
+    def _transform(self, input: Any, params: Dict[str, Any]) -> Any:
+        if isinstance(input, features.Image) or is_simple_tensor(input):
+            # PyTorch's pad supports only integers on fill. So we need to overwrite the colour
+            output = F.pad_image_tensor(input, params["padding"], fill=0, padding_mode="constant")
+
+            left, top, right, bottom = params["padding"]
+            fill = torch.tensor(params["fill"], dtype=input.dtype, device=input.device).to().view(-1, 1, 1)
+
+            if top > 0:
+                output[..., :top, :] = fill
+            if left > 0:
+                output[..., :, :left] = fill
+            if bottom > 0:
+                output[..., -bottom:, :] = fill
+            if right > 0:
+                output[..., :, -right:] = fill
+
+            if isinstance(input, features.Image):
+                output = features.Image.new_like(input, output)
+
+            return output
+        elif isinstance(input, PIL.Image.Image):
+            return F.pad_image_pil(
+                input,
+                params["padding"],
+                fill=tuple(int(v) if input.mode != "F" else v for v in params["fill"]),
+                padding_mode="constant",
+            )
+        elif isinstance(input, features.BoundingBox):
+            output = F.pad_bounding_box(input, params["padding"], format=input.format)
+
+            left, top, right, bottom = params["padding"]
+            height, width = input.image_size
+            height += top + bottom
+            width += left + right
+
+            return features.BoundingBox.new_like(input, output, image_size=(height, width))
+        else:
+            return input
+
+    def forward(self, *inputs: Any) -> Any:
+        sample = inputs if len(inputs) > 1 else inputs[0]
+        if torch.rand(1) >= self.p:
+            return sample
+
+        return super().forward(sample)

torchvision/prototype/transforms/functional/__init__.py

@@ -54,6 +54,7 @@
     rotate_image_pil,
     pad_image_tensor,
     pad_image_pil,
+    pad_bounding_box,
     crop_image_tensor,
     crop_image_pil,
     perspective_image_tensor,

torchvision/prototype/transforms/functional/_geometry.py

@@ -27,7 +27,7 @@ def horizontal_flip_bounding_box(
     bounding_box[:, [0, 2]] = image_size[1] - bounding_box[:, [2, 0]]
 
     return convert_bounding_box_format(
-        bounding_box, old_format=features.BoundingBoxFormat.XYXY, new_format=format
+        bounding_box, old_format=features.BoundingBoxFormat.XYXY, new_format=format, copy=False
     ).view(shape)
 
 
@@ -210,6 +210,26 @@ def rotate_image_pil(
 pad_image_tensor = _FT.pad
 pad_image_pil = _FP.pad
 
+
+def pad_bounding_box(
+    bounding_box: torch.Tensor, padding: List[int], format: features.BoundingBoxFormat
+) -> torch.Tensor:
+    left, _, top, _ = _FT._parse_pad_padding(padding)
+
+    shape = bounding_box.shape
+
+    bounding_box = convert_bounding_box_format(
+        bounding_box, old_format=format, new_format=features.BoundingBoxFormat.XYXY
+    ).view(-1, 4)
+
+    bounding_box[:, 0::2] += left
+    bounding_box[:, 1::2] += top
+
+    return convert_bounding_box_format(
+        bounding_box, old_format=features.BoundingBoxFormat.XYXY, new_format=format, copy=False
+    ).view(shape)
+
+
 crop_image_tensor = _FT.crop
 crop_image_pil = _FP.crop
 

torchvision/prototype/transforms/functional/_meta.py

@@ -40,10 +40,13 @@ def _xyxy_to_cxcywh(xyxy: torch.Tensor) -> torch.Tensor:
 
 
 def convert_bounding_box_format(
-    bounding_box: torch.Tensor, *, old_format: BoundingBoxFormat, new_format: BoundingBoxFormat
+    bounding_box: torch.Tensor, old_format: BoundingBoxFormat, new_format: BoundingBoxFormat, copy: bool = True
 ) -> torch.Tensor:
     if new_format == old_format:
-        return bounding_box.clone()
+        if copy:
+            return bounding_box.clone()
+        else:
+            return bounding_box
 
     if old_format == BoundingBoxFormat.XYWH:
         bounding_box = _xywh_to_xyxy(bounding_box)
@@ -89,10 +92,13 @@ def _gray_to_rgb(grayscale: torch.Tensor) -> torch.Tensor:
 
 
 def convert_image_color_space_tensor(
-    image: torch.Tensor, old_color_space: ColorSpace, new_color_space: ColorSpace
+    image: torch.Tensor, old_color_space: ColorSpace, new_color_space: ColorSpace, copy: bool = True
 ) -> torch.Tensor:
     if new_color_space == old_color_space:
-        return image.clone()
+        if copy:
+            return image.clone()
+        else:
+            return image
 
     if old_color_space == ColorSpace.OTHER or new_color_space == ColorSpace.OTHER:
         raise RuntimeError(f"Conversion to or from {ColorSpace.OTHER} is not supported.")
@@ -135,11 +141,16 @@ def convert_image_color_space_tensor(
 }
 
 
-def convert_image_color_space_pil(image: PIL.Image.Image, color_space: ColorSpace) -> PIL.Image.Image:
+def convert_image_color_space_pil(
+    image: PIL.Image.Image, color_space: ColorSpace, copy: bool = True
+) -> PIL.Image.Image:
     old_mode = image.mode
     try:
         new_mode = _COLOR_SPACE_TO_PIL_MODE[color_space]
     except KeyError:
         raise ValueError(f"Conversion from {ColorSpace.from_pil_mode(old_mode)} to {color_space} is not supported.")
 
+    if not copy and image.mode == new_mode:
+        return image
+
     return image.convert(new_mode)

torchvision/transforms/functional_tensor.py

@@ -350,6 +350,26 @@ def _pad_symmetric(img: Tensor, padding: List[int]) -> Tensor:
         raise RuntimeError("Symmetric padding of N-D tensors are not supported yet")
 
 
+def _parse_pad_padding(padding: List[int]) -> List[int]:
+    if isinstance(padding, int):
+        if torch.jit.is_scripting():
+            # This maybe unreachable
+            raise ValueError("padding can't be an int while torchscripting, set it as a list [value, ]")
+        pad_left = pad_right = pad_top = pad_bottom = padding
+    elif len(padding) == 1:
+        pad_left = pad_right = pad_top = pad_bottom = padding[0]
+    elif len(padding) == 2:
+        pad_left = pad_right = padding[0]
+        pad_top = pad_bottom = padding[1]
+    else:
+        pad_left = padding[0]
+        pad_top = padding[1]
+        pad_right = padding[2]
+        pad_bottom = padding[3]
+
+    return [pad_left, pad_right, pad_top, pad_bottom]
+
+
 def pad(img: Tensor, padding: List[int], fill: int = 0, padding_mode: str = "constant") -> Tensor:
     _assert_image_tensor(img)
 
@@ -369,23 +389,7 @@ def pad(img: Tensor, padding: List[int], fill: int = 0, padding_mode: str = "con
     if padding_mode not in ["constant", "edge", "reflect", "symmetric"]:
         raise ValueError("Padding mode should be either constant, edge, reflect or symmetric")
 
-    if isinstance(padding, int):
-        if torch.jit.is_scripting():
-            # This maybe unreachable
-            raise ValueError("padding can't be an int while torchscripting, set it as a list [value, ]")
-        pad_left = pad_right = pad_top = pad_bottom = padding
-    elif len(padding) == 1:
-        pad_left = pad_right = pad_top = pad_bottom = padding[0]
-    elif len(padding) == 2:
-        pad_left = pad_right = padding[0]
-        pad_top = pad_bottom = padding[1]
-    else:
-        pad_left = padding[0]
-        pad_top = padding[1]
-        pad_right = padding[2]
-        pad_bottom = padding[3]
-
-    p = [pad_left, pad_right, pad_top, pad_bottom]
+    p = _parse_pad_padding(padding)
 
     if padding_mode == "edge":
         # remap padding_mode str