Spaces:
Sleeping
Sleeping
File size: 8,869 Bytes
b2c5353 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 | # This code was originally written by Jose Javier Gonzalez Ortiz
# for use in UniverSeg (https://github.com/JJGO/UniverSeg).
# It is included here with their permission, without modifications.
import random
from typing import Any, Optional, Union
import kornia as K
import kornia.augmentation as KA
import numpy as np
import torch
from kornia.constants import BorderType
from pydantic import validate_arguments
from .common import AugmentationBase2D, _as2tuple, _as_single_val
class RandomBrightnessContrast(AugmentationBase2D):
def __init__(
self,
brightness: Union[float, tuple[float, float]] = 0.0,
contrast: Union[float, tuple[float, float]] = 1.0,
same_on_batch: bool = False,
p: float = 0.5,
keepdim: bool = False,
) -> None:
super().__init__(
p=p,
same_on_batch=same_on_batch,
p_batch=1.0,
keepdim=keepdim,
)
self.brightness = brightness
self.contrast = contrast
def generate_parameters(self, input_shape: torch.Size):
brightness = _as_single_val(self.brightness)
contrast = _as_single_val(self.contrast)
order = np.random.permutation(2)
return dict(brightness=brightness, contrast=contrast, order=order)
def apply_transform(
self,
input: torch.Tensor,
params: dict[str, torch.Tensor],
flags: dict[str, Any],
transform: Optional[torch.Tensor] = None,
) -> torch.Tensor:
transforms = [
lambda img: K.enhance.adjust_brightness(img, params["brightness"]),
lambda img: K.enhance.adjust_contrast(img, params["contrast"]),
]
jittered = input
for idx in params["order"].tolist():
t = transforms[idx]
jittered = t(jittered)
return jittered
class FilterBase(AugmentationBase2D):
@validate_arguments
def __init__(
self,
kernel_size: Union[int, tuple[int, int]],
sigma: Union[float, tuple[float, float]],
same_on_batch: bool = False,
p: float = 0.5,
keepdim: bool = False,
) -> None:
super().__init__(
p=p,
same_on_batch=same_on_batch,
p_batch=1.0,
keepdim=keepdim,
)
self.kernel_size = kernel_size
self.sigma = sigma
class VariableFilterBase(FilterBase):
"""Helper class for tasks that involve a random filter"""
def generate_parameters(self, input_shape: torch.Size):
kernel_size = _as_single_val(self.kernel_size)
sigma = _as_single_val(self.sigma)
return dict(kernel_size=kernel_size, sigma=sigma)
class RandomVariableGaussianBlur(VariableFilterBase):
def __init__(
self,
kernel_size: Union[int, tuple[int, int]],
sigma: Union[float, tuple[float, float]],
border_type: str = "reflect",
same_on_batch: bool = False,
p: float = 0.5,
keepdim: bool = False,
) -> None:
super().__init__(
kernel_size=kernel_size,
sigma=sigma,
p=p,
same_on_batch=same_on_batch,
keepdim=keepdim,
)
self.flags = dict(border_type=BorderType.get(border_type))
def apply_transform(
self,
input: torch.Tensor,
params: dict[str, torch.Tensor],
flags: dict[str, Any],
transform: Optional[torch.Tensor] = None,
) -> torch.Tensor:
kernel_size = _as2tuple(self.kernel_size)
sigma = _as2tuple(self.sigma)
return K.filters.gaussian_blur2d(
input, kernel_size, sigma, flags["border_type"].name.lower()
)
class RandomVariableBoxBlur(AugmentationBase2D):
def __init__(
self,
kernel_size: Union[int, tuple[int, int]] = 3,
border_type: str = "reflect",
normalized: bool = True,
same_on_batch: bool = False,
p: float = 0.5,
keepdim: bool = False,
) -> None:
super().__init__(
p=p,
same_on_batch=same_on_batch,
p_batch=1.0,
keepdim=keepdim,
)
self.flags = dict(border_type=border_type, normalized=normalized)
def generate_parameters(self, input_shape: torch.Size):
kernel_size = _as_single_val(self.kernel_size)
return dict(kernel_size=kernel_size)
def apply_transform(
self,
input: torch.Tensor,
params: dict[str, torch.Tensor],
flags: dict[str, Any],
transform: Optional[torch.Tensor] = None,
) -> torch.Tensor:
kernel_size = _as2tuple(params["kernel_size"])
return K.filters.box_blur(
input, kernel_size, flags["border_type"], flags["normalized"]
)
class RandomVariableGaussianNoise(AugmentationBase2D):
def __init__(
self,
mean: Union[float, tuple[float, float]] = 0.0,
std: Union[float, tuple[float, float]] = 1.0,
same_on_batch: bool = False,
p: float = 0.5,
keepdim: bool = False,
) -> None:
super().__init__(
p=p,
same_on_batch=same_on_batch,
p_batch=1.0,
keepdim=keepdim,
)
self.mean = mean
self.std = std
def generate_parameters(self, input_shape: torch.Size):
mean = _as_single_val(self.mean)
std = _as_single_val(self.std)
if torch.cuda.is_available():
noise = torch.empty(input_shape, dtype=torch.float32, device='cuda').normal_(mean, std)
else:
noise = torch.empty(input_shape, dtype=torch.float32, device='cpu').normal_(mean, std)
return dict(noise=noise)
def apply_transform(
self,
input: torch.Tensor,
params: dict[str, torch.Tensor],
flags: dict[str, Any],
transform: Optional[torch.Tensor] = None,
) -> torch.Tensor:
return input + params["noise"].to(input)
def validate_elastic_sigma_alpha(sigma, alpha):
if isinstance(alpha, (tuple, list)):
alpha = max(alpha)
if isinstance(sigma, (tuple, list)):
sigma = max(sigma)
if sigma / alpha < 1:
raise ValueError("Alpha and Sigma seem to be swapped")
class RandomVariableElasticTransform(AugmentationBase2D):
def __init__(
self,
kernel_size: Union[int, tuple[int, int]] = 63,
sigma: Union[float, tuple[float, float]] = 32,
alpha: Union[float, tuple[float, float]] = 1.0,
align_corners: bool = False,
mode: str = "bilinear",
padding_mode: str = "zeros",
same_on_batch: bool = False,
p: float = 0.5,
keepdim: bool = False,
) -> None:
super().__init__(
p=p,
same_on_batch=same_on_batch,
p_batch=1.0,
keepdim=keepdim,
)
validate_elastic_sigma_alpha(sigma, alpha)
self.flags = dict(
kernel_size=kernel_size,
sigma=sigma,
alpha=alpha,
align_corners=align_corners,
mode=mode,
padding_mode=padding_mode,
)
def generate_parameters(self, shape: torch.Size) -> dict[str, torch.Tensor]:
B, _, H, W = shape
# By default self.device (which is what kornia prefers, it default to cpu) so
# the conv2d's to lowpass filter the noise happen on the cpu regardless of
# input.device value. To bypass this, we force the noise device to 'cuda'
# whenever possible
device = "cuda" if torch.cuda.is_available() else "cpu"
if self.same_on_batch:
noise = torch.rand(1, 2, H, W, device=device, dtype=self.dtype).repeat(
B, 1, 1, 1
)
else:
noise = torch.rand(B, 2, H, W, device=device, dtype=self.dtype)
kernel_size = _as_single_val(self.flags["kernel_size"])
sigma = _as_single_val(self.flags["sigma"])
alpha = _as_single_val(self.flags["alpha"])
return dict(
noise=noise * 2 - 1, kernel_size=kernel_size, sigma=sigma, alpha=alpha
)
def apply_transform(
self,
input: torch.Tensor,
params: dict[str, torch.Tensor],
flags: dict[str, Any],
transform: Optional[torch.Tensor] = None,
) -> torch.Tensor:
assert (
input.device == params["noise"].device
), f"Input/Noise with different devices {input.device} & {params['noise'].device}"
return K.geometry.transform.elastic_transform2d(
input,
params["noise"], # .to(input),
_as2tuple(params["kernel_size"]),
_as2tuple(params["sigma"]),
_as2tuple(params["alpha"]),
flags["align_corners"],
flags["mode"],
flags["padding_mode"],
)
|