Spaces:
Running
on
Zero
Running
on
Zero
File size: 12,128 Bytes
1b34a12 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 |
"""
U-shaped DISCO Neural Operator
"""
from typing import List, Tuple
import torch
import torch.nn as nn
from torch.nn import functional as F
from torch_harmonics_local.convolution import (
EquidistantDiscreteContinuousConv2d as DISCO2d,
)
class UDNO(nn.Module):
"""
U-shaped DISCO Neural Operator in PyTorch
"""
def __init__(
self,
in_chans: int,
out_chans: int,
radius_cutoff: float,
chans: int = 32,
num_pool_layers: int = 4,
drop_prob: float = 0.0,
in_shape: Tuple[int, int] = (320, 320),
kernel_shape: Tuple[int, int] = (3, 4),
):
"""
Parameters
----------
in_chans : int
Number of channels in the input to the U-Net model.
out_chans : int
Number of channels in the output to the U-Net model.
radius_cutoff : float
Control the effective radius of the DISCO kernel. Values are
between 0.0 and 1.0. The radius_cutoff is represented as a proportion
of the normalized input space, to ensure that kernels are resolution
invaraint.
chans : int, optional
Number of output channels of the first DISCO layer. Default is 32.
num_pool_layers : int, optional
Number of down-sampling and up-sampling layers. Default is 4.
drop_prob : float, optional
Dropout probability. Default is 0.0.
in_shape : Tuple[int, int]
Shape of the input to the UDNO. This is required to dynamically
compile DISCO kernels for resolution invariance.
kernel_shape : Tuple[int, int], optional
Shape of the DISCO kernel. Default is (3, 4). This corresponds to 3
rings and 4 anisotropic basis functions. Under the hood, each DISCO
kernel has (3 - 1) * 4 + 1 = 9 parameters, equivalent to a standard
3x3 convolution kernel.
Note: This is NOT kernel_size, as under the DISCO framework,
kernels are dynamically compiled to support resolution invariance.
"""
super().__init__()
assert len(in_shape) == 2, "Input shape must be 2D"
self.in_chans = in_chans
self.out_chans = out_chans
self.chans = chans
self.num_pool_layers = num_pool_layers
self.drop_prob = drop_prob
self.in_shape = in_shape
self.kernel_shape = kernel_shape
self.down_sample_layers = nn.ModuleList(
[
DISCOBlock(
in_chans,
chans,
radius_cutoff,
drop_prob,
in_shape,
kernel_shape,
)
]
)
ch = chans
shape = (in_shape[0] // 2, in_shape[1] // 2)
radius_cutoff = radius_cutoff * 2
for _ in range(num_pool_layers - 1):
self.down_sample_layers.append(
DISCOBlock(
ch,
ch * 2,
radius_cutoff,
drop_prob,
in_shape=shape,
kernel_shape=kernel_shape,
)
)
ch *= 2
shape = (shape[0] // 2, shape[1] // 2)
radius_cutoff *= 2
# test commit
self.bottleneck = DISCOBlock(
ch,
ch * 2,
radius_cutoff,
drop_prob,
in_shape=shape,
kernel_shape=kernel_shape,
)
self.up = nn.ModuleList()
self.up_transpose = nn.ModuleList()
for _ in range(num_pool_layers - 1):
self.up_transpose.append(
TransposeDISCOBlock(
ch * 2,
ch,
radius_cutoff,
in_shape=shape,
kernel_shape=kernel_shape,
)
)
shape = (shape[0] * 2, shape[1] * 2)
radius_cutoff /= 2
self.up.append(
DISCOBlock(
ch * 2,
ch,
radius_cutoff,
drop_prob,
in_shape=shape,
kernel_shape=kernel_shape,
)
)
ch //= 2
self.up_transpose.append(
TransposeDISCOBlock(
ch * 2,
ch,
radius_cutoff,
in_shape=shape,
kernel_shape=kernel_shape,
)
)
shape = (shape[0] * 2, shape[1] * 2)
radius_cutoff /= 2
self.up.append(
nn.Sequential(
DISCOBlock(
ch * 2,
ch,
radius_cutoff,
drop_prob,
in_shape=shape,
kernel_shape=kernel_shape,
),
nn.Conv2d(
ch, self.out_chans, kernel_size=1, stride=1
), # 1x1 conv is always res-invariant (pixel wise channel transformation)
)
)
def forward(self, image: torch.Tensor) -> torch.Tensor:
"""
Parameters
----------
image : torch.Tensor
Input 4D tensor of shape `(N, in_chans, H, W)`.
Returns
-------
torch.Tensor
Output tensor of shape `(N, out_chans, H, W)`.
"""
stack = []
output = image
# apply down-sampling layers
for layer in self.down_sample_layers:
output = layer(output)
stack.append(output)
output = F.avg_pool2d(output, kernel_size=2, stride=2, padding=0)
output = self.bottleneck(output)
# apply up-sampling layers
for transpose, disco in zip(self.up_transpose, self.up):
downsample_layer = stack.pop()
output = transpose(output)
# reflect pad on the right/botton if needed to handle odd input dimensions
padding = [0, 0, 0, 0]
if output.shape[-1] != downsample_layer.shape[-1]:
padding[1] = 1 # padding right
if output.shape[-2] != downsample_layer.shape[-2]:
padding[3] = 1 # padding bottom
if torch.sum(torch.tensor(padding)) != 0:
output = F.pad(output, padding, "reflect")
output = torch.cat([output, downsample_layer], dim=1)
output = disco(output)
return output
class DISCOBlock(nn.Module):
"""
A DISCO Block that consists of two DISCO layers each followed by
instance normalization, LeakyReLU activation and dropout.
"""
def __init__(
self,
in_chans: int,
out_chans: int,
radius_cutoff: float,
drop_prob: float,
in_shape: Tuple[int, int],
kernel_shape: Tuple[int, int] = (3, 4),
):
"""
Parameters
----------
in_chans : int
Number of channels in the input.
out_chans : int
Number of channels in the output.
radius_cutoff : float
Control the effective radius of the DISCO kernel. Values are
between 0.0 and 1.0. The radius_cutoff is represented as a proportion
of the normalized input space, to ensure that kernels are resolution
invaraint.
in_shape : Tuple[int]
Unbatched spatial 2D shape of the input to this block.
Rrequired to dynamically compile DISCO kernels for resolution invariance.
kernel_shape : Tuple[int, int], optional
Shape of the DISCO kernel. Default is (3, 4). This corresponds to 3
rings and 4 anisotropic basis functions. Under the hood, each DISCO
kernel has (3 - 1) * 4 + 1 = 9 parameters, equivalent to a standard
3x3 convolution kernel.
Note: This is NOT kernel_size, as under the DISCO framework,
kernels are dynamically compiled to support resolution invariance.
drop_prob : float
Dropout probability.
"""
super().__init__()
self.in_chans = in_chans
self.out_chans = out_chans
self.drop_prob = drop_prob
self.layers = nn.Sequential(
DISCO2d(
in_chans,
out_chans,
kernel_shape=kernel_shape,
in_shape=in_shape,
bias=False,
radius_cutoff=radius_cutoff,
padding_mode="constant",
),
nn.InstanceNorm2d(out_chans),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Dropout2d(drop_prob),
DISCO2d(
out_chans,
out_chans,
kernel_shape=kernel_shape,
in_shape=in_shape,
bias=False,
radius_cutoff=radius_cutoff,
padding_mode="constant",
),
nn.InstanceNorm2d(out_chans),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Dropout2d(drop_prob),
)
def forward(self, image: torch.Tensor) -> torch.Tensor:
"""
Parameters
----------
image : ndarray
Input 4D tensor of shape `(N, in_chans, H, W)`.
Returns
-------
ndarray
Output tensor of shape `(N, out_chans, H, W)`.
"""
return self.layers(image)
class TransposeDISCOBlock(nn.Module):
"""
A transpose DISCO Block that consists of an up-sampling layer followed by a
DISCO layer, instance normalization, and LeakyReLU activation.
"""
def __init__(
self,
in_chans: int,
out_chans: int,
radius_cutoff: float,
in_shape: Tuple[int, int],
kernel_shape: Tuple[int, int] = (3, 4),
):
"""
Parameters
----------
in_chans : int
Number of channels in the input.
out_chans : int
Number of channels in the output.
radius_cutoff : float
Control the effective radius of the DISCO kernel. Values are
between 0.0 and 1.0. The radius_cutoff is represented as a proportion
of the normalized input space, to ensure that kernels are resolution
invaraint.
in_shape : Tuple[int]
Unbatched spatial 2D shape of the input to this block.
Rrequired to dynamically compile DISCO kernels for resolution invariance.
kernel_shape : Tuple[int, int], optional
Shape of the DISCO kernel. Default is (3, 4). This corresponds to 3
rings and 4 anisotropic basis functions. Under the hood, each DISCO
kernel has (3 - 1) * 4 + 1 = 9 parameters, equivalent to a standard
3x3 convolution kernel.
Note: This is NOT kernel_size, as under the DISCO framework,
kernels are dynamically compiled to support resolution invariance
"""
super().__init__()
self.in_chans = in_chans
self.out_chans = out_chans
self.layers = nn.Sequential(
nn.Upsample(scale_factor=2, mode="bilinear", align_corners=True),
DISCO2d(
in_chans,
out_chans,
kernel_shape=kernel_shape,
in_shape=(2 * in_shape[0], 2 * in_shape[1]),
bias=False,
radius_cutoff=(radius_cutoff / 2),
padding_mode="constant",
),
nn.InstanceNorm2d(out_chans),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
)
def forward(self, image: torch.Tensor) -> torch.Tensor:
"""
Parameters
----------
image : torch.Tensor
Input 4D tensor of shape `(N, in_chans, H, W)`.
Returns
-------
torch.Tensor
Output tensor of shape `(N, out_chans, H*2, W*2)`.
"""
return self.layers(image)
|