Delete main/src/model_unet_multiclass_deepsupervision_configured_v1.py with huggingface_hub
Browse files
main/src/model_unet_multiclass_deepsupervision_configured_v1.py
DELETED
|
@@ -1,560 +0,0 @@
|
|
| 1 |
-
from functools import partial
|
| 2 |
-
from timeit import repeat
|
| 3 |
-
from torch import nn
|
| 4 |
-
from basic_typing import Batch
|
| 5 |
-
from layers.unet_base import LatentConv, UNetBase
|
| 6 |
-
from layers.blocks import BlockConvNormActivation, BlockUpsampleNnConvNormActivation
|
| 7 |
-
from typing import Dict, Optional
|
| 8 |
-
import torch
|
| 9 |
-
|
| 10 |
-
import torch
|
| 11 |
-
import torch.nn as nn
|
| 12 |
-
import torch.nn.functional as F
|
| 13 |
-
|
| 14 |
-
from model_unet_multiclass import ModelUNetMulticlass, LossDiceCrossEntropyFocal
|
| 15 |
-
from model_unet_multiclass_deepsupervision import ModelUNetMulticlassDeepSupervision, LossDiceCrossEntropyFocal2
|
| 16 |
-
|
| 17 |
-
from layers.blocks import BlockConvNormActivation, BlockUpDeconvSkipConv
|
| 18 |
-
from layers.layer_config import LayerConfig, default_layer_config, NormType
|
| 19 |
-
from layers.unet_base import BlockConvType, BlockTypeConvSkip, Up
|
| 20 |
-
from layers.unet_attention import MergeBlockAttention_Gating_Input
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
class Down(nn.Module):
|
| 24 |
-
def __init__(
|
| 25 |
-
self,
|
| 26 |
-
layer_config: LayerConfig,
|
| 27 |
-
bloc_level: int,
|
| 28 |
-
input_channels: int,
|
| 29 |
-
output_channels: int,
|
| 30 |
-
block: BlockConvType = BlockConvNormActivation,
|
| 31 |
-
**block_kwargs):
|
| 32 |
-
super().__init__()
|
| 33 |
-
|
| 34 |
-
ops = [
|
| 35 |
-
block(
|
| 36 |
-
config=layer_config,
|
| 37 |
-
input_channels=input_channels,
|
| 38 |
-
output_channels=output_channels,
|
| 39 |
-
**block_kwargs
|
| 40 |
-
),
|
| 41 |
-
block(
|
| 42 |
-
config=layer_config,
|
| 43 |
-
input_channels=output_channels,
|
| 44 |
-
output_channels=output_channels,
|
| 45 |
-
)
|
| 46 |
-
]
|
| 47 |
-
|
| 48 |
-
self.ops = nn.Sequential(*ops)
|
| 49 |
-
|
| 50 |
-
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 51 |
-
return self.ops(x)
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
SimpleMulticlassUNet_dice_ce_fov_v5 = partial(
|
| 55 |
-
ModelUNetMulticlass,
|
| 56 |
-
model=UNetBase(
|
| 57 |
-
dim=3,
|
| 58 |
-
input_channels=2,
|
| 59 |
-
channels=[32, 64, 128, 256, 320],
|
| 60 |
-
output_channels=2,
|
| 61 |
-
config=default_layer_config(
|
| 62 |
-
norm_type=NormType.InstanceNorm,
|
| 63 |
-
conv_kwargs={'padding': 'same', 'bias': False},
|
| 64 |
-
deconv_kwargs={'padding': 'same'},
|
| 65 |
-
norm_kwargs={'affine': True},
|
| 66 |
-
activation=partial(nn.LeakyReLU, negative_slope=0.01)
|
| 67 |
-
),
|
| 68 |
-
activation=None,
|
| 69 |
-
down_block_fn=Down,
|
| 70 |
-
# , deconv_kernel_size=2
|
| 71 |
-
up_block_fn=partial(Up, block=partial(BlockUpDeconvSkipConv, nb_repeats=2)),
|
| 72 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=3))
|
| 73 |
-
),
|
| 74 |
-
loss_fn=partial(LossDiceCrossEntropyFocal, ce_factor=1.0)
|
| 75 |
-
)
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
SimpleMulticlassUNet_dice_ce_fov_v6 = partial(
|
| 79 |
-
ModelUNetMulticlassDeepSupervision,
|
| 80 |
-
model=UNetBase(
|
| 81 |
-
dim=3,
|
| 82 |
-
input_channels=2,
|
| 83 |
-
channels=[48, 64, 128, 256, 320],
|
| 84 |
-
output_channels=2,
|
| 85 |
-
config=default_layer_config(
|
| 86 |
-
norm_type=NormType.InstanceNorm,
|
| 87 |
-
#conv_kwargs={'padding': 'same', 'bias': False},
|
| 88 |
-
conv_kwargs={'padding': 'same'},
|
| 89 |
-
deconv_kwargs={'padding': 'same'},
|
| 90 |
-
#norm_kwargs={'affine': True},
|
| 91 |
-
activation=partial(nn.LeakyReLU, negative_slope=0.01)
|
| 92 |
-
),
|
| 93 |
-
activation=None,
|
| 94 |
-
down_block_fn=Down,
|
| 95 |
-
up_block_fn=partial(Up, block=partial(BlockUpDeconvSkipConv, nb_repeats=2)),
|
| 96 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=3))
|
| 97 |
-
),
|
| 98 |
-
with_deep_supervision=False,
|
| 99 |
-
#loss_fn=LossDiceCrossEntropyFocal2_nnUnet
|
| 100 |
-
loss_fn=LossDiceCrossEntropyFocal2
|
| 101 |
-
)
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
SimpleMulticlassUNet_dice_ce_fov_v6_ds = partial(
|
| 105 |
-
ModelUNetMulticlassDeepSupervision,
|
| 106 |
-
model=UNetBase(
|
| 107 |
-
dim=3,
|
| 108 |
-
input_channels=2,
|
| 109 |
-
channels=[48, 64, 128, 256, 320],
|
| 110 |
-
output_channels=2,
|
| 111 |
-
config=default_layer_config(
|
| 112 |
-
norm_type=NormType.InstanceNorm,
|
| 113 |
-
#conv_kwargs={'padding': 'same', 'bias': False},
|
| 114 |
-
conv_kwargs={'padding': 'same'},
|
| 115 |
-
deconv_kwargs={'padding': 'same'},
|
| 116 |
-
#norm_kwargs={'affine': True},
|
| 117 |
-
activation=partial(nn.LeakyReLU, negative_slope=0.01)
|
| 118 |
-
),
|
| 119 |
-
activation=None,
|
| 120 |
-
down_block_fn=Down,
|
| 121 |
-
up_block_fn=partial(Up, block=partial(BlockUpDeconvSkipConv, nb_repeats=2)),
|
| 122 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=3))
|
| 123 |
-
),
|
| 124 |
-
with_deep_supervision=True,
|
| 125 |
-
#loss_fn=LossDiceCrossEntropyFocal2_nnUnet
|
| 126 |
-
loss_fn=LossDiceCrossEntropyFocal2
|
| 127 |
-
)
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
SimpleMulticlassUNet_dice_ce_fov_v7_ds_lung = partial(
|
| 131 |
-
ModelUNetMulticlassDeepSupervision,
|
| 132 |
-
model=UNetBase(
|
| 133 |
-
dim=3,
|
| 134 |
-
input_channels=3,
|
| 135 |
-
channels=[64, 96, 128, 156],
|
| 136 |
-
output_channels=2,
|
| 137 |
-
config=default_layer_config(
|
| 138 |
-
norm_type=NormType.InstanceNorm,
|
| 139 |
-
activation=nn.LeakyReLU,
|
| 140 |
-
norm_kwargs={'affine': True},
|
| 141 |
-
),
|
| 142 |
-
activation=None,
|
| 143 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9))
|
| 144 |
-
),
|
| 145 |
-
with_deep_supervision=True,
|
| 146 |
-
loss_fn=LossDiceCrossEntropyFocal2,
|
| 147 |
-
with_ct_lung=True,
|
| 148 |
-
input_target_shape=(3, 96, 96, 96)
|
| 149 |
-
)
|
| 150 |
-
|
| 151 |
-
SimpleMulticlassUNet_dice_ce_fov_v8_ds_lung = partial(
|
| 152 |
-
ModelUNetMulticlassDeepSupervision,
|
| 153 |
-
model=UNetBase(
|
| 154 |
-
dim=3,
|
| 155 |
-
input_channels=3,
|
| 156 |
-
channels=[64, 96, 128, 156],
|
| 157 |
-
output_channels=2,
|
| 158 |
-
config=default_layer_config(
|
| 159 |
-
norm_type=NormType.InstanceNorm,
|
| 160 |
-
activation=nn.LeakyReLU,
|
| 161 |
-
norm_kwargs={'affine': True},
|
| 162 |
-
),
|
| 163 |
-
activation=None,
|
| 164 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9))
|
| 165 |
-
),
|
| 166 |
-
with_deep_supervision=True,
|
| 167 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=1.0, gamma=1.0),
|
| 168 |
-
with_ct_lung=True,
|
| 169 |
-
input_target_shape=(3, 96, 96, 96)
|
| 170 |
-
)
|
| 171 |
-
|
| 172 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot = partial(
|
| 173 |
-
ModelUNetMulticlassDeepSupervision,
|
| 174 |
-
model=UNetBase(
|
| 175 |
-
dim=3,
|
| 176 |
-
input_channels=5,
|
| 177 |
-
channels=[64, 96, 128, 156],
|
| 178 |
-
output_channels=2,
|
| 179 |
-
config=default_layer_config(
|
| 180 |
-
norm_type=NormType.InstanceNorm,
|
| 181 |
-
activation=nn.LeakyReLU,
|
| 182 |
-
norm_kwargs={'affine': True},
|
| 183 |
-
),
|
| 184 |
-
activation=None,
|
| 185 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9))
|
| 186 |
-
),
|
| 187 |
-
with_deep_supervision=True,
|
| 188 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=1.0, gamma=1.0),
|
| 189 |
-
with_ct_lung=True,
|
| 190 |
-
with_ct_soft=True,
|
| 191 |
-
with_pet_hot=True,
|
| 192 |
-
input_target_shape=(5, 96, 96, 96)
|
| 193 |
-
)
|
| 194 |
-
|
| 195 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_boundary = partial(
|
| 196 |
-
ModelUNetMulticlassDeepSupervision,
|
| 197 |
-
model=UNetBase(
|
| 198 |
-
dim=3,
|
| 199 |
-
input_channels=5,
|
| 200 |
-
channels=[64, 96, 128, 156],
|
| 201 |
-
output_channels=2,
|
| 202 |
-
config=default_layer_config(
|
| 203 |
-
norm_type=NormType.InstanceNorm,
|
| 204 |
-
activation=nn.LeakyReLU,
|
| 205 |
-
norm_kwargs={'affine': True},
|
| 206 |
-
),
|
| 207 |
-
activation=None,
|
| 208 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9))
|
| 209 |
-
),
|
| 210 |
-
with_deep_supervision=True,
|
| 211 |
-
boundary_loss_factor=1.0,
|
| 212 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=0.5, gamma=1.0),
|
| 213 |
-
with_ct_lung=True,
|
| 214 |
-
with_ct_soft=True,
|
| 215 |
-
with_pet_hot=True,
|
| 216 |
-
input_target_shape=(5, 96, 96, 96)
|
| 217 |
-
)
|
| 218 |
-
|
| 219 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_boundary_sensitivity = partial(
|
| 220 |
-
ModelUNetMulticlassDeepSupervision,
|
| 221 |
-
model=UNetBase(
|
| 222 |
-
dim=3,
|
| 223 |
-
input_channels=5,
|
| 224 |
-
channels=[64, 96, 128, 156],
|
| 225 |
-
output_channels=2,
|
| 226 |
-
config=default_layer_config(
|
| 227 |
-
norm_type=NormType.InstanceNorm,
|
| 228 |
-
activation=nn.LeakyReLU,
|
| 229 |
-
norm_kwargs={'affine': True},
|
| 230 |
-
),
|
| 231 |
-
activation=None,
|
| 232 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9))
|
| 233 |
-
),
|
| 234 |
-
with_deep_supervision=True,
|
| 235 |
-
boundary_loss_factor=1.0,
|
| 236 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=0.5, gamma=1.0, sensitivity=2.0),
|
| 237 |
-
with_ct_lung=True,
|
| 238 |
-
with_ct_soft=True,
|
| 239 |
-
with_pet_hot=True,
|
| 240 |
-
input_target_shape=(5, 96, 96, 96)
|
| 241 |
-
)
|
| 242 |
-
|
| 243 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_boundary_coord = partial(
|
| 244 |
-
ModelUNetMulticlassDeepSupervision,
|
| 245 |
-
model=UNetBase(
|
| 246 |
-
dim=3,
|
| 247 |
-
input_channels=8,
|
| 248 |
-
channels=[64, 96, 128, 156],
|
| 249 |
-
output_channels=2,
|
| 250 |
-
config=default_layer_config(
|
| 251 |
-
norm_type=NormType.InstanceNorm,
|
| 252 |
-
activation=nn.LeakyReLU,
|
| 253 |
-
norm_kwargs={'affine': True},
|
| 254 |
-
),
|
| 255 |
-
activation=None,
|
| 256 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9))
|
| 257 |
-
),
|
| 258 |
-
with_deep_supervision=True,
|
| 259 |
-
boundary_loss_factor=1.0,
|
| 260 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=0.5, gamma=1.0),
|
| 261 |
-
with_ct_lung=True,
|
| 262 |
-
with_ct_soft=True,
|
| 263 |
-
with_pet_hot=True,
|
| 264 |
-
with_additional_features=('z_coords', 'y_coords', 'x_coords'),
|
| 265 |
-
input_target_shape=(8, 96, 96, 96)
|
| 266 |
-
)
|
| 267 |
-
|
| 268 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_boundary_composite = partial(
|
| 269 |
-
ModelUNetMulticlassDeepSupervision,
|
| 270 |
-
model=UNetBase(
|
| 271 |
-
dim=3,
|
| 272 |
-
input_channels=5,
|
| 273 |
-
channels=[64, 96, 128, 156],
|
| 274 |
-
output_channels=2,
|
| 275 |
-
config=default_layer_config(
|
| 276 |
-
norm_type=NormType.InstanceNorm,
|
| 277 |
-
activation=nn.LeakyReLU,
|
| 278 |
-
norm_kwargs={'affine': True},
|
| 279 |
-
),
|
| 280 |
-
activation=None,
|
| 281 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9))
|
| 282 |
-
),
|
| 283 |
-
with_deep_supervision=False,
|
| 284 |
-
with_deep_supervision_composite=True,
|
| 285 |
-
boundary_loss_factor=1.0,
|
| 286 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=0.5, gamma=1.0),
|
| 287 |
-
with_ct_lung=True,
|
| 288 |
-
with_ct_soft=True,
|
| 289 |
-
with_pet_hot=True,
|
| 290 |
-
input_target_shape=(5, 96, 96, 96)
|
| 291 |
-
)
|
| 292 |
-
|
| 293 |
-
|
| 294 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_larger = partial(
|
| 295 |
-
ModelUNetMulticlassDeepSupervision,
|
| 296 |
-
model=UNetBase(
|
| 297 |
-
dim=3,
|
| 298 |
-
input_channels=5,
|
| 299 |
-
channels=[96, 128, 156, 192],
|
| 300 |
-
output_channels=2,
|
| 301 |
-
config=default_layer_config(
|
| 302 |
-
norm_type=NormType.InstanceNorm,
|
| 303 |
-
activation=nn.LeakyReLU,
|
| 304 |
-
norm_kwargs={'affine': True},
|
| 305 |
-
),
|
| 306 |
-
activation=None,
|
| 307 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9))
|
| 308 |
-
),
|
| 309 |
-
with_deep_supervision=True,
|
| 310 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=1.0, gamma=1.0),
|
| 311 |
-
with_ct_lung=True,
|
| 312 |
-
with_ct_soft=True,
|
| 313 |
-
with_pet_hot=True,
|
| 314 |
-
input_target_shape=(5, 96, 96, 96)
|
| 315 |
-
)
|
| 316 |
-
|
| 317 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_larger2 = partial(
|
| 318 |
-
ModelUNetMulticlassDeepSupervision,
|
| 319 |
-
model=UNetBase(
|
| 320 |
-
dim=3,
|
| 321 |
-
input_channels=5,
|
| 322 |
-
channels=[128, 156, 192, 224],
|
| 323 |
-
output_channels=2,
|
| 324 |
-
config=default_layer_config(
|
| 325 |
-
norm_type=NormType.InstanceNorm,
|
| 326 |
-
activation=nn.LeakyReLU,
|
| 327 |
-
norm_kwargs={'affine': True},
|
| 328 |
-
),
|
| 329 |
-
activation=None,
|
| 330 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=7))
|
| 331 |
-
),
|
| 332 |
-
with_deep_supervision=True,
|
| 333 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=1.0, gamma=1.0),
|
| 334 |
-
with_ct_lung=True,
|
| 335 |
-
with_ct_soft=True,
|
| 336 |
-
with_pet_hot=True,
|
| 337 |
-
input_target_shape=(5, 96, 96, 96)
|
| 338 |
-
)
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_larger2_sensitivity = partial(
|
| 342 |
-
ModelUNetMulticlassDeepSupervision,
|
| 343 |
-
model=UNetBase(
|
| 344 |
-
dim=3,
|
| 345 |
-
input_channels=5,
|
| 346 |
-
channels=[128, 156, 192, 224],
|
| 347 |
-
output_channels=2,
|
| 348 |
-
config=default_layer_config(
|
| 349 |
-
norm_type=NormType.InstanceNorm,
|
| 350 |
-
activation=nn.LeakyReLU,
|
| 351 |
-
norm_kwargs={'affine': True},
|
| 352 |
-
),
|
| 353 |
-
activation=None,
|
| 354 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=7))
|
| 355 |
-
),
|
| 356 |
-
with_deep_supervision=True,
|
| 357 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=1.0, gamma=1.0, sensitivity=1.0),
|
| 358 |
-
with_ct_lung=True,
|
| 359 |
-
with_ct_soft=True,
|
| 360 |
-
with_pet_hot=True,
|
| 361 |
-
input_target_shape=(5, 96, 96, 96)
|
| 362 |
-
)
|
| 363 |
-
|
| 364 |
-
|
| 365 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_focal_composite = partial(
|
| 366 |
-
ModelUNetMulticlassDeepSupervision,
|
| 367 |
-
model=UNetBase(
|
| 368 |
-
dim=3,
|
| 369 |
-
input_channels=5,
|
| 370 |
-
channels=[64, 96, 128, 156],
|
| 371 |
-
output_channels=2,
|
| 372 |
-
config=default_layer_config(
|
| 373 |
-
norm_type=NormType.InstanceNorm,
|
| 374 |
-
activation=nn.LeakyReLU,
|
| 375 |
-
norm_kwargs={'affine': True},
|
| 376 |
-
),
|
| 377 |
-
activation=None,
|
| 378 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9))
|
| 379 |
-
),
|
| 380 |
-
with_deep_supervision=False,
|
| 381 |
-
with_deep_supervision_composite=True,
|
| 382 |
-
boundary_loss_factor=None,
|
| 383 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=0.5, gamma=2.0),
|
| 384 |
-
with_ct_lung=True,
|
| 385 |
-
with_ct_soft=True,
|
| 386 |
-
with_pet_hot=True,
|
| 387 |
-
input_target_shape=(5, 96, 96, 96)
|
| 388 |
-
)
|
| 389 |
-
|
| 390 |
-
|
| 391 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_focal = partial(
|
| 392 |
-
ModelUNetMulticlassDeepSupervision,
|
| 393 |
-
model=UNetBase(
|
| 394 |
-
dim=3,
|
| 395 |
-
input_channels=5,
|
| 396 |
-
channels=[64, 96, 128, 156],
|
| 397 |
-
output_channels=2,
|
| 398 |
-
config=default_layer_config(
|
| 399 |
-
norm_type=NormType.InstanceNorm,
|
| 400 |
-
activation=nn.LeakyReLU,
|
| 401 |
-
norm_kwargs={'affine': True},
|
| 402 |
-
),
|
| 403 |
-
activation=None,
|
| 404 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9))
|
| 405 |
-
),
|
| 406 |
-
with_deep_supervision=True,
|
| 407 |
-
with_deep_supervision_composite=False,
|
| 408 |
-
boundary_loss_factor=None,
|
| 409 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=0.5, gamma=2.0),
|
| 410 |
-
with_ct_lung=True,
|
| 411 |
-
with_ct_soft=True,
|
| 412 |
-
with_pet_hot=True,
|
| 413 |
-
input_target_shape=(5, 96, 96, 96)
|
| 414 |
-
)
|
| 415 |
-
|
| 416 |
-
|
| 417 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_boundary_attention = partial(
|
| 418 |
-
ModelUNetMulticlassDeepSupervision,
|
| 419 |
-
model=UNetBase(
|
| 420 |
-
dim=3,
|
| 421 |
-
input_channels=5,
|
| 422 |
-
channels=[64, 96, 128, 156],
|
| 423 |
-
output_channels=2,
|
| 424 |
-
config=default_layer_config(
|
| 425 |
-
norm_type=NormType.InstanceNorm,
|
| 426 |
-
activation=nn.LeakyReLU,
|
| 427 |
-
norm_kwargs={'affine': True},
|
| 428 |
-
),
|
| 429 |
-
activation=None,
|
| 430 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9)),
|
| 431 |
-
up_block_fn=partial(Up, block=partial(
|
| 432 |
-
BlockUpDeconvSkipConv,
|
| 433 |
-
merge_layer_fn=MergeBlockAttention_Gating_Input,
|
| 434 |
-
deconv_block=BlockUpsampleNnConvNormActivation)
|
| 435 |
-
)
|
| 436 |
-
),
|
| 437 |
-
with_deep_supervision=True,
|
| 438 |
-
boundary_loss_factor=1.0,
|
| 439 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=0.5, gamma=1.0),
|
| 440 |
-
with_ct_lung=True,
|
| 441 |
-
with_ct_soft=True,
|
| 442 |
-
with_pet_hot=True,
|
| 443 |
-
input_target_shape=(5, 96, 96, 96)
|
| 444 |
-
)
|
| 445 |
-
|
| 446 |
-
|
| 447 |
-
from basic_typing import KernelSize, Padding, Stride
|
| 448 |
-
from typing import Union, Sequence
|
| 449 |
-
from layers.blocks import BlockMerge, BlockDeconvNormActivation
|
| 450 |
-
from layers.non_local import BlockNonLocal, linear_embedding
|
| 451 |
-
|
| 452 |
-
class BlockUpDeconvSkipConvNonLocal(nn.Module):
|
| 453 |
-
def __init__(
|
| 454 |
-
self,
|
| 455 |
-
config: LayerConfig,
|
| 456 |
-
skip_channels: int,
|
| 457 |
-
input_channels: int,
|
| 458 |
-
output_channels: int,
|
| 459 |
-
*,
|
| 460 |
-
nb_repeats: int = 1,
|
| 461 |
-
kernel_size: Optional[KernelSize] = None,
|
| 462 |
-
deconv_kernel_size: Optional[KernelSize] = None,
|
| 463 |
-
padding: Optional[Padding] = None,
|
| 464 |
-
output_padding: Optional[Union[int, Sequence[int]]] = None,
|
| 465 |
-
deconv_block=BlockDeconvNormActivation,
|
| 466 |
-
stride: Optional[Stride] = None,
|
| 467 |
-
merge_layer_fn=BlockMerge):
|
| 468 |
-
super().__init__()
|
| 469 |
-
|
| 470 |
-
self.merge_layer = merge_layer_fn(config=config, layer_channels=[skip_channels, output_channels])
|
| 471 |
-
|
| 472 |
-
if deconv_kernel_size is None:
|
| 473 |
-
deconv_kernel_size = kernel_size
|
| 474 |
-
|
| 475 |
-
self.ops_deconv = deconv_block(
|
| 476 |
-
config,
|
| 477 |
-
input_channels=input_channels,
|
| 478 |
-
output_channels=output_channels,
|
| 479 |
-
kernel_size=deconv_kernel_size,
|
| 480 |
-
padding=padding,
|
| 481 |
-
output_padding=output_padding,
|
| 482 |
-
stride=stride
|
| 483 |
-
)
|
| 484 |
-
|
| 485 |
-
convs = [
|
| 486 |
-
BlockConvNormActivation(
|
| 487 |
-
config,
|
| 488 |
-
input_channels=self.merge_layer.get_output_channels(),
|
| 489 |
-
output_channels=output_channels,
|
| 490 |
-
kernel_size=kernel_size,
|
| 491 |
-
stride=1
|
| 492 |
-
)
|
| 493 |
-
]
|
| 494 |
-
|
| 495 |
-
for _ in range(1, nb_repeats):
|
| 496 |
-
conv = BlockConvNormActivation(
|
| 497 |
-
config,
|
| 498 |
-
input_channels=output_channels,
|
| 499 |
-
output_channels=output_channels,
|
| 500 |
-
kernel_size=kernel_size,
|
| 501 |
-
stride=1
|
| 502 |
-
)
|
| 503 |
-
convs.append(conv)
|
| 504 |
-
|
| 505 |
-
self.ops_conv: Union[nn.Sequential, BlockConvNormActivation]
|
| 506 |
-
if len(convs) == 1:
|
| 507 |
-
self.ops_conv = convs[0]
|
| 508 |
-
else:
|
| 509 |
-
self.ops_conv = nn.Sequential(*convs)
|
| 510 |
-
|
| 511 |
-
self.non_local = BlockNonLocal(
|
| 512 |
-
config=config,
|
| 513 |
-
input_channels=output_channels,
|
| 514 |
-
f_mapping_fn=linear_embedding,
|
| 515 |
-
g_mapping_fn=linear_embedding,
|
| 516 |
-
intermediate_channels=max(16, output_channels // 2)
|
| 517 |
-
)
|
| 518 |
-
|
| 519 |
-
self.skip_channels = skip_channels
|
| 520 |
-
self.input_channels = input_channels
|
| 521 |
-
self.output_channels = output_channels
|
| 522 |
-
|
| 523 |
-
def forward(self, skip: torch.Tensor, previous: torch.Tensor) -> torch.Tensor:
|
| 524 |
-
assert skip.shape[1] == self.skip_channels
|
| 525 |
-
assert previous.shape[1] == self.input_channels
|
| 526 |
-
x = self.ops_deconv(previous)
|
| 527 |
-
assert x.shape[2:] == skip.shape[2:], f'got shape={x.shape[2:]}, expected={skip.shape[2:]}'
|
| 528 |
-
|
| 529 |
-
x = self.merge_layer([skip, x])
|
| 530 |
-
x = self.ops_conv(x)
|
| 531 |
-
if x.shape[2] <= 32:
|
| 532 |
-
# else it is just too slow...
|
| 533 |
-
x = self.non_local(x)
|
| 534 |
-
return x
|
| 535 |
-
|
| 536 |
-
|
| 537 |
-
SimpleMulticlassUNet_dice_ce_fov_v9_ds_lung_soft_hot_boundary_non_local = partial(
|
| 538 |
-
ModelUNetMulticlassDeepSupervision,
|
| 539 |
-
model=UNetBase(
|
| 540 |
-
dim=3,
|
| 541 |
-
input_channels=5,
|
| 542 |
-
channels=[64, 96, 128, 156],
|
| 543 |
-
output_channels=2,
|
| 544 |
-
config=default_layer_config(
|
| 545 |
-
norm_type=NormType.InstanceNorm,
|
| 546 |
-
activation=nn.LeakyReLU,
|
| 547 |
-
norm_kwargs={'affine': True},
|
| 548 |
-
),
|
| 549 |
-
activation=None,
|
| 550 |
-
middle_block_fn=partial(LatentConv, block=partial(BlockConvNormActivation, kernel_size=9)),
|
| 551 |
-
up_block_fn=partial(Up, block=BlockUpDeconvSkipConvNonLocal)
|
| 552 |
-
),
|
| 553 |
-
with_deep_supervision=True,
|
| 554 |
-
boundary_loss_factor=1.0,
|
| 555 |
-
loss_fn=partial(LossDiceCrossEntropyFocal2, ce_factor=0.5, gamma=1.0),
|
| 556 |
-
with_ct_lung=True,
|
| 557 |
-
with_ct_soft=True,
|
| 558 |
-
with_pet_hot=True,
|
| 559 |
-
input_target_shape=(5, 96, 96, 96)
|
| 560 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|