File size: 6,418 Bytes
23e2777 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 |
import torch
Tensor = torch.Tensor
Device = torch.DeviceObjType
Dtype = torch.Type
pad = torch.nn.functional.pad
def _compute_zero_padding(kernel_size: tuple[int, int] | int) -> tuple[int, int]:
ky, kx = _unpack_2d_ks(kernel_size)
return (ky - 1) // 2, (kx - 1) // 2
def _unpack_2d_ks(kernel_size: tuple[int, int] | int) -> tuple[int, int]:
if isinstance(kernel_size, int):
ky = kx = kernel_size
else:
assert len(kernel_size) == 2, '2D Kernel size should have a length of 2.'
ky, kx = kernel_size
ky = int(ky)
kx = int(kx)
return ky, kx
def gaussian(
window_size: int, sigma: Tensor | float, *, device: Device | None = None, dtype: Dtype | None = None
) -> Tensor:
batch_size = sigma.shape[0]
x = (torch.arange(window_size, device=sigma.device, dtype=sigma.dtype) - window_size // 2).expand(batch_size, -1)
if window_size % 2 == 0:
x = x + 0.5
gauss = torch.exp(-x.pow(2.0) / (2 * sigma.pow(2.0)))
return gauss / gauss.sum(-1, keepdim=True)
def get_gaussian_kernel1d(
kernel_size: int,
sigma: float | Tensor,
force_even: bool = False,
*,
device: Device | None = None,
dtype: Dtype | None = None,
) -> Tensor:
return gaussian(kernel_size, sigma, device=device, dtype=dtype)
def get_gaussian_kernel2d(
kernel_size: tuple[int, int] | int,
sigma: tuple[float, float] | Tensor,
force_even: bool = False,
*,
device: Device | None = None,
dtype: Dtype | None = None,
) -> Tensor:
sigma = torch.Tensor([[sigma, sigma]]).to(device=device, dtype=dtype)
ksize_y, ksize_x = _unpack_2d_ks(kernel_size)
sigma_y, sigma_x = sigma[:, 0, None], sigma[:, 1, None]
kernel_y = get_gaussian_kernel1d(ksize_y, sigma_y, force_even, device=device, dtype=dtype)[..., None]
kernel_x = get_gaussian_kernel1d(ksize_x, sigma_x, force_even, device=device, dtype=dtype)[..., None]
return kernel_y * kernel_x.view(-1, 1, ksize_x)
def _bilateral_blur(
input: Tensor,
guidance: Tensor | None,
kernel_size: tuple[int, int] | int,
sigma_color: float | Tensor,
sigma_space: tuple[float, float] | Tensor,
border_type: str = 'reflect',
color_distance_type: str = 'l1',
) -> Tensor:
if isinstance(sigma_color, Tensor):
sigma_color = sigma_color.to(device=input.device, dtype=input.dtype).view(-1, 1, 1, 1, 1)
ky, kx = _unpack_2d_ks(kernel_size)
pad_y, pad_x = _compute_zero_padding(kernel_size)
padded_input = pad(input, (pad_x, pad_x, pad_y, pad_y), mode=border_type)
unfolded_input = padded_input.unfold(2, ky, 1).unfold(3, kx, 1).flatten(-2) # (B, C, H, W, Ky x Kx)
if guidance is None:
guidance = input
unfolded_guidance = unfolded_input
else:
padded_guidance = pad(guidance, (pad_x, pad_x, pad_y, pad_y), mode=border_type)
unfolded_guidance = padded_guidance.unfold(2, ky, 1).unfold(3, kx, 1).flatten(-2) # (B, C, H, W, Ky x Kx)
diff = unfolded_guidance - guidance.unsqueeze(-1)
if color_distance_type == "l1":
color_distance_sq = diff.abs().sum(1, keepdim=True).square()
elif color_distance_type == "l2":
color_distance_sq = diff.square().sum(1, keepdim=True)
else:
raise ValueError("color_distance_type only acceps l1 or l2")
color_kernel = (-0.5 / sigma_color**2 * color_distance_sq).exp() # (B, 1, H, W, Ky x Kx)
space_kernel = get_gaussian_kernel2d(kernel_size, sigma_space, device=input.device, dtype=input.dtype)
space_kernel = space_kernel.view(-1, 1, 1, 1, kx * ky)
kernel = space_kernel * color_kernel
out = (unfolded_input * kernel).sum(-1) / kernel.sum(-1)
return out
def bilateral_blur(
input: Tensor,
kernel_size: tuple[int, int] | int = (13, 13),
sigma_color: float | Tensor = 3.0,
sigma_space: tuple[float, float] | Tensor = 3.0,
border_type: str = 'reflect',
color_distance_type: str = 'l1',
) -> Tensor:
return _bilateral_blur(input, None, kernel_size, sigma_color, sigma_space, border_type, color_distance_type)
def adaptive_anisotropic_filter(x, g=None):
if g is None:
g = x
s, m = torch.std_mean(g, dim=(1, 2, 3), keepdim=True)
s = s + 1e-5
guidance = (g - m) / s
y = _bilateral_blur(x, guidance,
kernel_size=(13, 13),
sigma_color=3.0,
sigma_space=3.0,
border_type='reflect',
color_distance_type='l1')
return y
def joint_bilateral_blur(
input: Tensor,
guidance: Tensor,
kernel_size: tuple[int, int] | int,
sigma_color: float | Tensor,
sigma_space: tuple[float, float] | Tensor,
border_type: str = 'reflect',
color_distance_type: str = 'l1',
) -> Tensor:
return _bilateral_blur(input, guidance, kernel_size, sigma_color, sigma_space, border_type, color_distance_type)
class _BilateralBlur(torch.nn.Module):
def __init__(
self,
kernel_size: tuple[int, int] | int,
sigma_color: float | Tensor,
sigma_space: tuple[float, float] | Tensor,
border_type: str = 'reflect',
color_distance_type: str = "l1",
) -> None:
super().__init__()
self.kernel_size = kernel_size
self.sigma_color = sigma_color
self.sigma_space = sigma_space
self.border_type = border_type
self.color_distance_type = color_distance_type
def __repr__(self) -> str:
return (
f"{self.__class__.__name__}"
f"(kernel_size={self.kernel_size}, "
f"sigma_color={self.sigma_color}, "
f"sigma_space={self.sigma_space}, "
f"border_type={self.border_type}, "
f"color_distance_type={self.color_distance_type})"
)
class BilateralBlur(_BilateralBlur):
def forward(self, input: Tensor) -> Tensor:
return bilateral_blur(
input, self.kernel_size, self.sigma_color, self.sigma_space, self.border_type, self.color_distance_type
)
class JointBilateralBlur(_BilateralBlur):
def forward(self, input: Tensor, guidance: Tensor) -> Tensor:
return joint_bilateral_blur(
input,
guidance,
self.kernel_size,
self.sigma_color,
self.sigma_space,
self.border_type,
self.color_distance_type,
)
|