Skip to content

Updated fill arg typehint for affine, perspective and elastic ops #6595

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Sep 19, 2022
Merged

Updated fill arg typehint for affine, perspective and elastic ops #6595

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
e24d71a
Updated fill arg typehint for affine, perspective and elastic ops
vfdev-5 Sep 16, 2022
8bf31ba
Updated pad op on prototype side
vfdev-5 Sep 16, 2022
a36da7d
Code updates
vfdev-5 Sep 16, 2022
6d62612
Few other minor updates
vfdev-5 Sep 16, 2022
fc42d08
Merge branch 'main' of github.com:pytorch/vision into enh-fill-typehi…
vfdev-5 Sep 16, 2022
430911e
Merge branch 'main' into enh-fill-typehint-2
vfdev-5 Sep 19, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion test/prototype_transforms_kernel_infos.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -226,7 +226,7 @@ def sample_inputs_affine_image_tensor():
],
[None, (0, 0)],
):
for fill in [None, [0.5] * image_loader.num_channels]:
for fill in [None, 128.0, 128, [12.0], [0.5] * image_loader.num_channels]:
yield ArgsKwargs(
image_loader,
interpolation=interpolation_mode,
Expand Down
12 changes: 10 additions & 2 deletions test/test_prototype_transforms_functional.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -228,8 +228,12 @@ def perspective_image_tensor():
[1.2405, 0.1772, -6.9113, 0.0463, 1.251, -5.235, 0.00013, 0.0018],
[0.7366, -0.11724, 1.45775, -0.15012, 0.73406, 2.6019, -0.0072, -0.0063],
],
[None, [128], [12.0]], # fill
[None, 128.0, 128, [12.0], [1.0, 2.0, 3.0]], # fill
):
if isinstance(fill, list) and len(fill) == 3 and image.shape[1] != 3:
# skip the test with non-broadcastable fill value
continue

yield ArgsKwargs(image, perspective_coeffs=perspective_coeffs, fill=fill)


Expand Down Expand Up @@ -268,8 +272,12 @@ def perspective_mask():
def elastic_image_tensor():
for image, fill in itertools.product(
make_images(extra_dims=((), (4,))),
[None, [128], [12.0]], # fill
[None, 128.0, 128, [12.0], [1.0, 2.0, 3.0]], # fill
):
if isinstance(fill, list) and len(fill) == 3 and image.shape[1] != 3:
# skip the test with non-broadcastable fill value
continue

h, w = image.shape[-2:]
displacement = torch.rand(1, h, w, 2)
yield ArgsKwargs(image, displacement=displacement, fill=fill)
Expand Down
7 changes: 2 additions & 5 deletions torchvision/prototype/features/_image.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -177,12 +177,9 @@ def pad(
if not isinstance(padding, int):
padding = list(padding)

# PyTorch's pad supports only scalars on fill. So we need to overwrite the colour
if isinstance(fill, (int, float)) or fill is None:
output = self._F.pad_image_tensor(self, padding, fill=fill, padding_mode=padding_mode)
else:
output = self._F._geometry._pad_with_vector_fill(self, padding, fill=fill, padding_mode=padding_mode)
fill = self._F._geometry._convert_fill_arg(fill)

output = self._F.pad_image_tensor(self, padding, fill=fill, padding_mode=padding_mode)
return Image.new_like(self, output)

def rotate(
Expand Down
9 changes: 2 additions & 7 deletions torchvision/prototype/features/_mask.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,14 +58,9 @@ def pad(
if not isinstance(padding, int):
padding = list(padding)

if isinstance(fill, (int, float)) or fill is None:
if fill is None:
fill = 0
output = self._F.pad_mask(self, padding, padding_mode=padding_mode, fill=fill)
else:
# Let's raise an error for vector fill on masks
raise ValueError("Non-scalar fill value is not supported")
fill = self._F._geometry._convert_fill_arg(fill)

output = self._F.pad_mask(self, padding, padding_mode=padding_mode, fill=fill)
return Mask.new_like(self, output)

def rotate(
Expand Down
50 changes: 35 additions & 15 deletions torchvision/prototype/transforms/functional/_geometry.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -232,7 +232,7 @@ def affine_image_tensor(
scale: float,
shear: List[float],
interpolation: InterpolationMode = InterpolationMode.NEAREST,
fill: Optional[List[float]] = None,
fill: Optional[Union[int, float, List[float]]] = None,
center: Optional[List[float]] = None,
) -> torch.Tensor:
if img.numel() == 0:
Expand Down Expand Up @@ -405,7 +405,9 @@ def affine_mask(
return output


def _convert_fill_arg(fill: Optional[Union[int, float, Sequence[int], Sequence[float]]]) -> Optional[List[float]]:
def _convert_fill_arg(
fill: Optional[Union[int, float, Sequence[int], Sequence[float]]]
) -> Optional[Union[int, float, List[float]]]:
# Fill = 0 is not equivalent to None, https://github.com/pytorch/vision/issues/6517
# So, we can't reassign fill to 0
# if fill is None:
Expand All @@ -416,9 +418,6 @@ def _convert_fill_arg(fill: Optional[Union[int, float, Sequence[int], Sequence[f
# This cast does Sequence -> List[float] to please mypy and torch.jit.script
if not isinstance(fill, (int, float)):
fill = [float(v) for v in list(fill)]
else:
# It is OK to cast int to float as later we use inpt.dtype
fill = [float(fill)]
return fill


Expand Down Expand Up @@ -591,7 +590,23 @@ def rotate(
def pad_image_tensor(
img: torch.Tensor,
padding: Union[int, List[int]],
fill: Optional[Union[int, float]] = 0,
fill: Optional[Union[int, float, List[float]]] = None,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I know this kernel is private on stable and I agree with the change to ensure consistency. Any BC breakages that can creep upwards due to the change on the default value?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

pad_image_tensor is low-level new API so I think there is no BC to keep. In any case fill=None will be transformed to fill=0 inside

vision/torchvision/transforms/functional_tensor.py

Lines 381 to 382 in c0911e3

if fill is None:
fill = 0

padding_mode: str = "constant",
) -> torch.Tensor:
if fill is None:
# This is a JIT workaround
return _pad_with_scalar_fill(img, padding, fill=None, padding_mode=padding_mode)
elif isinstance(fill, (int, float)) or len(fill) == 1:
fill_number = fill[0] if isinstance(fill, list) else fill
return _pad_with_scalar_fill(img, padding, fill=fill_number, padding_mode=padding_mode)
Comment on lines +596 to +601
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I just wonder if we could do something like:

Suggested change
if fill is None:
# This is a JIT workaround
return _pad_with_scalar_fill(img, padding, fill=None, padding_mode=padding_mode)
elif isinstance(fill, (int, float)) or len(fill) == 1:
fill_number = fill[0] if isinstance(fill, list) else fill
return _pad_with_scalar_fill(img, padding, fill=fill_number, padding_mode=padding_mode)
if fill is None or isinstance(fill, (int, float)) or len(fill) == 1:
if fill is None:
fill = 0
fill_number = fill[0] if isinstance(fill, list) else fill
return _pad_with_scalar_fill(img, padding, fill=fill_number, padding_mode=padding_mode)

To avoid the duplicate dispatch to scalar. I assume that fill-None ends up becoming 0?

Copy link
Collaborator Author

@vfdev-5 vfdev-5 Sep 16, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Functional tensor is already doing

vision/torchvision/transforms/functional_tensor.py

Lines 381 to 382 in c0911e3

if fill is None:
fill = 0

Copy link
Collaborator Author

@vfdev-5 vfdev-5 Sep 16, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I was thinking that below code is valid, but JIT does not agree

    if fill is None or isinstance(fill, (int, float)) or len(fill) == 1:
        fill_number = fill[0] if isinstance(fill, list) else fill

else:
return _pad_with_vector_fill(img, padding, fill=fill, padding_mode=padding_mode)


def _pad_with_scalar_fill(
img: torch.Tensor,
padding: Union[int, List[int]],
fill: Union[int, float, None],
padding_mode: str = "constant",
) -> torch.Tensor:
num_channels, height, width = img.shape[-3:]
Expand All @@ -614,13 +629,13 @@ def pad_image_tensor(
def _pad_with_vector_fill(
img: torch.Tensor,
padding: Union[int, List[int]],
fill: Sequence[float] = [0.0],
fill: List[float],
padding_mode: str = "constant",
) -> torch.Tensor:
if padding_mode != "constant":
raise ValueError(f"Padding mode '{padding_mode}' is not supported if fill is not scalar")

output = pad_image_tensor(img, padding, fill=0, padding_mode="constant")
output = _pad_with_scalar_fill(img, padding, fill=0, padding_mode="constant")
left, right, top, bottom = _parse_pad_padding(padding)
fill = torch.tensor(fill, dtype=img.dtype, device=img.device).view(-1, 1, 1)

Expand All @@ -639,8 +654,14 @@ def pad_mask(
mask: torch.Tensor,
padding: Union[int, List[int]],
padding_mode: str = "constant",
fill: Optional[Union[int, float]] = 0,
fill: Optional[Union[int, float, List[float]]] = None,
) -> torch.Tensor:
if fill is None:
fill = 0

if isinstance(fill, list):
raise ValueError("Non-scalar fill value is not supported")

if mask.ndim < 3:
mask = mask.unsqueeze(0)
needs_squeeze = True
Expand Down Expand Up @@ -693,10 +714,9 @@ def pad(
if not isinstance(padding, int):
padding = list(padding)

# TODO: PyTorch's pad supports only scalars on fill. So we need to overwrite the colour
if isinstance(fill, (int, float)) or fill is None:
return pad_image_tensor(inpt, padding, fill=fill, padding_mode=padding_mode)
return _pad_with_vector_fill(inpt, padding, fill=fill, padding_mode=padding_mode)
fill = _convert_fill_arg(fill)

return pad_image_tensor(inpt, padding, fill=fill, padding_mode=padding_mode)


crop_image_tensor = _FT.crop
Expand Down Expand Up @@ -739,7 +759,7 @@ def perspective_image_tensor(
img: torch.Tensor,
perspective_coeffs: List[float],
interpolation: InterpolationMode = InterpolationMode.BILINEAR,
fill: Optional[List[float]] = None,
fill: Optional[Union[int, float, List[float]]] = None,
) -> torch.Tensor:
return _FT.perspective(img, perspective_coeffs, interpolation=interpolation.value, fill=fill)

Expand Down Expand Up @@ -878,7 +898,7 @@ def elastic_image_tensor(
img: torch.Tensor,
displacement: torch.Tensor,
interpolation: InterpolationMode = InterpolationMode.BILINEAR,
fill: Optional[List[float]] = None,
fill: Optional[Union[int, float, List[float]]] = None,
) -> torch.Tensor:
return _FT.elastic_transform(img, displacement, interpolation=interpolation.value, fill=fill)

Expand Down
12 changes: 9 additions & 3 deletions torchvision/transforms/functional_tensor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -600,7 +600,10 @@ def _gen_affine_grid(


def affine(
img: Tensor, matrix: List[float], interpolation: str = "nearest", fill: Optional[List[float]] = None
img: Tensor,
matrix: List[float],
interpolation: str = "nearest",
fill: Optional[Union[int, float, List[float]]] = None,
) -> Tensor:
_assert_grid_transform_inputs(img, matrix, interpolation, fill, ["nearest", "bilinear"])

Expand Down Expand Up @@ -693,7 +696,10 @@ def _perspective_grid(coeffs: List[float], ow: int, oh: int, dtype: torch.dtype,


def perspective(
img: Tensor, perspective_coeffs: List[float], interpolation: str = "bilinear", fill: Optional[List[float]] = None
img: Tensor,
perspective_coeffs: List[float],
interpolation: str = "bilinear",
fill: Optional[Union[int, float, List[float]]] = None,
) -> Tensor:
if not (isinstance(img, torch.Tensor)):
raise TypeError("Input img should be Tensor.")
Expand Down Expand Up @@ -950,7 +956,7 @@ def elastic_transform(
img: Tensor,
displacement: Tensor,
interpolation: str = "bilinear",
fill: Optional[List[float]] = None,
fill: Optional[Union[int, float, List[float]]] = None,
) -> Tensor:

if not (isinstance(img, torch.Tensor)):
Expand Down