from typing import Dict, List, Optional, Tuple, Union, cast
import torch
from torch.nn.functional import pad
from kornia.augmentation import random_generator as rg
from kornia.augmentation._2d.geometric.base import GeometricAugmentationBase2D
from kornia.augmentation.base import TensorWithTransformMat
from kornia.augmentation.utils import _transform_input
from kornia.constants import Resample
from kornia.geometry.transform import crop_by_transform_mat, get_perspective_transform
[docs]class RandomCrop(GeometricAugmentationBase2D):
r"""Crop random patches of a tensor image on a given size.
.. image:: _static/img/RandomCrop.png
Args:
p: probability of applying the transformation for the whole batch.
size: Desired output size (out_h, out_w) of the crop.
Must be Tuple[int, int], then out_h = size[0], out_w = size[1].
padding: Optional padding on each border
of the image. Default is None, i.e no padding. If a sequence of length
4 is provided, it is used to pad left, top, right, bottom borders
respectively. If a sequence of length 2 is provided, it is used to
pad left/right, top/bottom borders, respectively.
pad_if_needed: It will pad the image if smaller than the
desired size to avoid raising an exception. Since cropping is done
after padding, the padding seems to be done at a random offset.
fill: Pixel fill value for constant fill. Default is 0. If a tuple of
length 3, it is used to fill R, G, B channels respectively.
This value is only used when the padding_mode is constant.
padding_mode: Type of padding. Should be: constant, edge, reflect or symmetric.
resample: the interpolation mode.
return_transform: if ``True`` return the matrix describing the transformation applied to each
input tensor. If ``False`` and the input is a tuple the applied transformation
won't be concatenated.
same_on_batch: apply the same transformation across the batch.
align_corners: interpolation flag.
keepdim: whether to keep the output shape the same as input (True) or broadcast it
to the batch form (False).
cropping_mode: The used algorithm to crop. ``slice`` will use advanced slicing to extract the tensor based
on the sampled indices. ``resample`` will use `warp_affine` using the affine transformation
to extract and resize at once. Use `slice` for efficiency, or `resample` for proper
differentiability.
Shape:
- Input: :math:`(C, H, W)` or :math:`(B, C, H, W)`, Optional: :math:`(B, 3, 3)`
- Output: :math:`(B, C, out_h, out_w)`
Note:
Input tensor must be float and normalized into [0, 1] for the best differentiability support.
Additionally, this function accepts another transformation tensor (:math:`(B, 3, 3)`), then the
applied transformation will be merged int to the input transformation tensor and returned.
Examples:
>>> _ = torch.manual_seed(0)
>>> inputs = torch.arange(1*1*3*3.).view(1, 1, 3, 3)
>>> aug = RandomCrop((2, 2), p=1., cropping_mode="resample")
>>> out = aug(inputs)
>>> out
tensor([[[[3., 4.],
[6., 7.]]]])
>>> aug.inverse(out, padding_mode="border")
tensor([[[[3., 4., 4.],
[3., 4., 4.],
[6., 7., 7.]]]])
To apply the exact augmenation again, you may take the advantage of the previous parameter state:
>>> input = torch.randn(1, 3, 32, 32)
>>> aug = RandomCrop((2, 2), p=1., cropping_mode="resample")
>>> (aug(input) == aug(input, params=aug._params)).all()
tensor(True)
"""
def __init__(
self,
size: Tuple[int, int],
padding: Optional[Union[int, Tuple[int, int], Tuple[int, int, int, int]]] = None,
pad_if_needed: Optional[bool] = False,
fill: int = 0,
padding_mode: str = "constant",
resample: Union[str, int, Resample] = Resample.BILINEAR.name,
return_transform: bool = False,
same_on_batch: bool = False,
align_corners: bool = True,
p: float = 1.0,
keepdim: bool = False,
cropping_mode: str = "slice",
) -> None:
# Since PyTorch does not support ragged tensor. So cropping function happens batch-wisely.
super().__init__(
p=1.0, return_transform=return_transform, same_on_batch=same_on_batch, p_batch=p, keepdim=keepdim
)
self._param_generator = cast(rg.CropGenerator, rg.CropGenerator(size))
self.flags = dict(
size=size,
padding=padding,
pad_if_needed=pad_if_needed,
fill=fill,
padding_mode=padding_mode,
resample=Resample.get(resample),
align_corners=align_corners,
cropping_mode=cropping_mode,
)
def compute_padding(self, shape: torch.Size) -> List[int]:
if len(shape) != 4:
raise AssertionError(f"Expected BCHW. Got {shape}.")
padding = [0, 0, 0, 0]
if self.flags["padding"] is not None:
if isinstance(self.flags["padding"], int):
padding = [self.flags["padding"]] * 4
elif isinstance(self.flags["padding"], tuple) and len(self.flags["padding"]) == 2:
padding = [
self.flags["padding"][1],
self.flags["padding"][1],
self.flags["padding"][0],
self.flags["padding"][0],
]
elif isinstance(self.flags["padding"], tuple) and len(self.flags["padding"]) == 4:
padding = [
self.flags["padding"][3],
self.flags["padding"][2],
self.flags["padding"][1],
self.flags["padding"][0],
]
else:
raise RuntimeError(f"Expect `padding` to be a scalar, or length 2/4 list. Got {self.flags['padding']}.")
if self.flags["pad_if_needed"] and shape[-2] < self.flags["size"][0]:
padding = [0, 0, (self.flags["size"][0] - shape[-2]), self.flags["size"][0] - shape[-2]]
if self.flags["pad_if_needed"] and shape[-1] < self.flags["size"][1]:
padding = [self.flags["size"][1] - shape[-1], self.flags["size"][1] - shape[-1], 0, 0]
return padding
def precrop_padding(self, input: torch.Tensor, padding: List[int] = None) -> torch.Tensor:
if padding is None:
padding = self.compute_padding(input.shape)
input = pad(input, padding, value=self.flags["fill"], mode=self.flags["padding_mode"])
return input
def compute_transformation(self, input: torch.Tensor, params: Dict[str, torch.Tensor]) -> torch.Tensor:
transform: torch.Tensor = get_perspective_transform(params["src"].to(input), params["dst"].to(input))
return transform
def apply_transform(
self, input: torch.Tensor, params: Dict[str, torch.Tensor], transform: Optional[torch.Tensor] = None
) -> torch.Tensor:
if self.flags["cropping_mode"] == "resample": # uses bilinear interpolation to crop
transform = cast(torch.Tensor, transform)
return crop_by_transform_mat(
input,
transform,
self.flags["size"],
mode=self.flags["resample"].name.lower(),
padding_mode="zeros",
align_corners=self.flags["align_corners"],
)
if self.flags["cropping_mode"] == "slice": # uses advanced slicing to crop
B, C, _, _ = input.shape
out = torch.empty(B, C, *self.flags["size"], device=input.device, dtype=input.dtype)
for i in range(B):
x1 = int(params["src"][i, 0, 0])
x2 = int(params["src"][i, 1, 0]) + 1
y1 = int(params["src"][i, 0, 1])
y2 = int(params["src"][i, 3, 1]) + 1
out[i] = input[i : i + 1, :, y1:y2, x1:x2]
return out
raise NotImplementedError(f"Not supported type: {self.flags['cropping_mode']}.")
def inverse_transform(
self,
input: torch.Tensor,
transform: Optional[torch.Tensor] = None,
size: Optional[Tuple[int, int]] = None,
**kwargs,
) -> torch.Tensor:
if self.flags["cropping_mode"] != "resample":
raise NotImplementedError(
f"`inverse` is only applicable for resample cropping mode. Got {self.flags['cropping_mode']}."
)
size = cast(Tuple[int, int], size)
mode = self.flags["resample"].name.lower() if "mode" not in kwargs else kwargs["mode"]
align_corners = self.flags["align_corners"] if "align_corners" not in kwargs else kwargs["align_corners"]
padding_mode = "zeros" if "padding_mode" not in kwargs else kwargs["padding_mode"]
transform = cast(torch.Tensor, transform)
return crop_by_transform_mat(input, transform[:, :2, :], size, mode, padding_mode, align_corners)
def inverse(
self,
input: TensorWithTransformMat,
params: Optional[Dict[str, torch.Tensor]] = None,
size: Optional[Tuple[int, int]] = None,
**kwargs,
) -> torch.Tensor:
out = super().inverse(input, params, size, **kwargs)
if params is None:
params = self._params
if "padding_size" in params:
padding_size = params["padding_size"].unique(dim=0).cpu().squeeze().numpy().tolist()
padding_size = [-padding_size[0], -padding_size[1], -padding_size[2], -padding_size[3]]
else:
padding_size = [0, 0, 0, 0]
return self.precrop_padding(out, padding_size)
def forward_parameters_precrop(self, batch_shape) -> Dict[str, torch.Tensor]:
input_pad = self.compute_padding(batch_shape)
batch_shape_new = (
*batch_shape[:2],
batch_shape[2] + input_pad[2] + input_pad[3], # original height + top + bottom padding
batch_shape[3] + input_pad[0] + input_pad[1], # original width + left + right padding
)
padding_size = torch.tensor(tuple(input_pad), dtype=torch.long).expand(batch_shape[0], -1)
_params = super().forward_parameters(batch_shape_new)
_params.update({"padding_size": padding_size})
return _params
def forward(
self,
input: TensorWithTransformMat,
params: Optional[Dict[str, torch.Tensor]] = None,
return_transform: Optional[bool] = None,
) -> TensorWithTransformMat:
padding_size = params.get("padding_size") if params else None
if padding_size is not None:
input_pad = padding_size.unique(dim=0).cpu().squeeze().numpy().tolist()
else:
input_pad = None
if isinstance(input, (tuple, list)):
input_temp = _transform_input(input[0])
input_pad = self.compute_padding(input[0].shape) if input_pad is None else input_pad
_input = (self.precrop_padding(input_temp, input_pad), input[1])
else:
input = cast(torch.Tensor, input) # TODO: weird that cast is not working under this context.
input_temp = _transform_input(input)
input_pad = self.compute_padding(input_temp.shape) if input_pad is None else input_pad
_input = self.precrop_padding(input_temp, input_pad) # type: ignore
out = super().forward(_input, params, return_transform)
# Update the actual input size for inverse
if "padding_size" not in self._params:
_padding_size = torch.tensor(tuple(input_pad), device=input_temp.device, dtype=torch.long).expand(
input_temp.size(0), -1
)
self._params.update({"padding_size": _padding_size})
if not self._params["batch_prob"].all():
# undo the pre-crop if nothing happened.
if isinstance(out, tuple) and isinstance(input, tuple):
return input[0], out[1]
if isinstance(out, tuple) and not isinstance(input, tuple):
return input, out[1]
return input
return out