File size: 13,403 Bytes
4a582ec |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 |
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# The gca code was heavily based on https://github.com/Yaoyi-Li/GCA-Matting
# and https://github.com/open-mmlab/mmediting
import paddle
import paddle.nn as nn
import paddle.nn.functional as F
from paddleseg.cvlibs import manager, param_init
from paddleseg.utils import utils
from ppmatting.models.layers import GuidedCxtAtten
class ResNet_D(nn.Layer):
def __init__(self,
input_channels,
layers,
late_downsample=False,
pretrained=None):
super().__init__()
self.pretrained = pretrained
self._norm_layer = nn.BatchNorm
self.inplanes = 64
self.late_downsample = late_downsample
self.midplanes = 64 if late_downsample else 32
self.start_stride = [1, 2, 1, 2] if late_downsample else [2, 1, 2, 1]
self.conv1 = nn.utils.spectral_norm(
nn.Conv2D(
input_channels,
32,
kernel_size=3,
stride=self.start_stride[0],
padding=1,
bias_attr=False))
self.conv2 = nn.utils.spectral_norm(
nn.Conv2D(
32,
self.midplanes,
kernel_size=3,
stride=self.start_stride[1],
padding=1,
bias_attr=False))
self.conv3 = nn.utils.spectral_norm(
nn.Conv2D(
self.midplanes,
self.inplanes,
kernel_size=3,
stride=self.start_stride[2],
padding=1,
bias_attr=False))
self.bn1 = self._norm_layer(32)
self.bn2 = self._norm_layer(self.midplanes)
self.bn3 = self._norm_layer(self.inplanes)
self.activation = nn.ReLU()
self.layer1 = self._make_layer(
BasicBlock, 64, layers[0], stride=self.start_stride[3])
self.layer2 = self._make_layer(BasicBlock, 128, layers[1], stride=2)
self.layer3 = self._make_layer(BasicBlock, 256, layers[2], stride=2)
self.layer_bottleneck = self._make_layer(
BasicBlock, 512, layers[3], stride=2)
self.init_weight()
def _make_layer(self, block, planes, block_num, stride=1):
if block_num == 0:
return nn.Sequential(nn.Identity())
norm_layer = self._norm_layer
downsample = None
if stride != 1:
downsample = nn.Sequential(
nn.AvgPool2D(2, stride),
nn.utils.spectral_norm(
conv1x1(self.inplanes, planes * block.expansion)),
norm_layer(planes * block.expansion), )
elif self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
nn.utils.spectral_norm(
conv1x1(self.inplanes, planes * block.expansion, stride)),
norm_layer(planes * block.expansion), )
layers = [block(self.inplanes, planes, stride, downsample, norm_layer)]
self.inplanes = planes * block.expansion
for _ in range(1, block_num):
layers.append(block(self.inplanes, planes, norm_layer=norm_layer))
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.activation(x)
x = self.conv2(x)
x = self.bn2(x)
x1 = self.activation(x) # N x 32 x 256 x 256
x = self.conv3(x1)
x = self.bn3(x)
x2 = self.activation(x) # N x 64 x 128 x 128
x3 = self.layer1(x2) # N x 64 x 128 x 128
x4 = self.layer2(x3) # N x 128 x 64 x 64
x5 = self.layer3(x4) # N x 256 x 32 x 32
x = self.layer_bottleneck(x5) # N x 512 x 16 x 16
return x, (x1, x2, x3, x4, x5)
def init_weight(self):
for layer in self.sublayers():
if isinstance(layer, nn.Conv2D):
if hasattr(layer, "weight_orig"):
param = layer.weight_orig
else:
param = layer.weight
param_init.xavier_uniform(param)
elif isinstance(layer, (nn.BatchNorm, nn.SyncBatchNorm)):
param_init.constant_init(layer.weight, value=1.0)
param_init.constant_init(layer.bias, value=0.0)
elif isinstance(layer, BasicBlock):
param_init.constant_init(layer.bn2.weight, value=0.0)
if self.pretrained is not None:
utils.load_pretrained_model(self, self.pretrained)
@manager.MODELS.add_component
class ResShortCut_D(ResNet_D):
def __init__(self,
input_channels,
layers,
late_downsample=False,
pretrained=None):
super().__init__(
input_channels,
layers,
late_downsample=late_downsample,
pretrained=pretrained)
self.shortcut_inplane = [input_channels, self.midplanes, 64, 128, 256]
self.shortcut_plane = [32, self.midplanes, 64, 128, 256]
self.shortcut = nn.LayerList()
for stage, inplane in enumerate(self.shortcut_inplane):
self.shortcut.append(
self._make_shortcut(inplane, self.shortcut_plane[stage]))
def _make_shortcut(self, inplane, planes):
return nn.Sequential(
nn.utils.spectral_norm(
nn.Conv2D(
inplane, planes, kernel_size=3, padding=1,
bias_attr=False)),
nn.ReLU(),
self._norm_layer(planes),
nn.utils.spectral_norm(
nn.Conv2D(
planes, planes, kernel_size=3, padding=1, bias_attr=False)),
nn.ReLU(),
self._norm_layer(planes))
def forward(self, x):
out = self.conv1(x)
out = self.bn1(out)
out = self.activation(out)
out = self.conv2(out)
out = self.bn2(out)
x1 = self.activation(out) # N x 32 x 256 x 256
out = self.conv3(x1)
out = self.bn3(out)
out = self.activation(out)
x2 = self.layer1(out) # N x 64 x 128 x 128
x3 = self.layer2(x2) # N x 128 x 64 x 64
x4 = self.layer3(x3) # N x 256 x 32 x 32
out = self.layer_bottleneck(x4) # N x 512 x 16 x 16
fea1 = self.shortcut[0](x) # input image and trimap
fea2 = self.shortcut[1](x1)
fea3 = self.shortcut[2](x2)
fea4 = self.shortcut[3](x3)
fea5 = self.shortcut[4](x4)
return out, {
'shortcut': (fea1, fea2, fea3, fea4, fea5),
'image': x[:, :3, ...]
}
@manager.MODELS.add_component
class ResGuidedCxtAtten(ResNet_D):
def __init__(self,
input_channels,
layers,
late_downsample=False,
pretrained=None):
super().__init__(
input_channels,
layers,
late_downsample=late_downsample,
pretrained=pretrained)
self.input_channels = input_channels
self.shortcut_inplane = [input_channels, self.midplanes, 64, 128, 256]
self.shortcut_plane = [32, self.midplanes, 64, 128, 256]
self.shortcut = nn.LayerList()
for stage, inplane in enumerate(self.shortcut_inplane):
self.shortcut.append(
self._make_shortcut(inplane, self.shortcut_plane[stage]))
self.guidance_head = nn.Sequential(
nn.Pad2D(
1, mode="reflect"),
nn.utils.spectral_norm(
nn.Conv2D(
3, 16, kernel_size=3, padding=0, stride=2,
bias_attr=False)),
nn.ReLU(),
self._norm_layer(16),
nn.Pad2D(
1, mode="reflect"),
nn.utils.spectral_norm(
nn.Conv2D(
16, 32, kernel_size=3, padding=0, stride=2,
bias_attr=False)),
nn.ReLU(),
self._norm_layer(32),
nn.Pad2D(
1, mode="reflect"),
nn.utils.spectral_norm(
nn.Conv2D(
32,
128,
kernel_size=3,
padding=0,
stride=2,
bias_attr=False)),
nn.ReLU(),
self._norm_layer(128))
self.gca = GuidedCxtAtten(128, 128)
self.init_weight()
def init_weight(self):
for layer in self.sublayers():
if isinstance(layer, nn.Conv2D):
initializer = nn.initializer.XavierUniform()
if hasattr(layer, "weight_orig"):
param = layer.weight_orig
else:
param = layer.weight
initializer(param, param.block)
elif isinstance(layer, (nn.BatchNorm, nn.SyncBatchNorm)):
param_init.constant_init(layer.weight, value=1.0)
param_init.constant_init(layer.bias, value=0.0)
elif isinstance(layer, BasicBlock):
param_init.constant_init(layer.bn2.weight, value=0.0)
if self.pretrained is not None:
utils.load_pretrained_model(self, self.pretrained)
def _make_shortcut(self, inplane, planes):
return nn.Sequential(
nn.utils.spectral_norm(
nn.Conv2D(
inplane, planes, kernel_size=3, padding=1,
bias_attr=False)),
nn.ReLU(),
self._norm_layer(planes),
nn.utils.spectral_norm(
nn.Conv2D(
planes, planes, kernel_size=3, padding=1, bias_attr=False)),
nn.ReLU(),
self._norm_layer(planes))
def forward(self, x):
out = self.conv1(x)
out = self.bn1(out)
out = self.activation(out)
out = self.conv2(out)
out = self.bn2(out)
x1 = self.activation(out) # N x 32 x 256 x 256
out = self.conv3(x1)
out = self.bn3(out)
out = self.activation(out)
im_fea = self.guidance_head(
x[:, :3, ...]) # downsample origin image and extract features
if self.input_channels == 6:
unknown = F.interpolate(
x[:, 4:5, ...], scale_factor=1 / 8, mode='nearest')
else:
unknown = x[:, 3:, ...].equal(paddle.to_tensor([1.]))
unknown = paddle.cast(unknown, dtype='float32')
unknown = F.interpolate(unknown, scale_factor=1 / 8, mode='nearest')
x2 = self.layer1(out) # N x 64 x 128 x 128
x3 = self.layer2(x2) # N x 128 x 64 x 64
x3 = self.gca(im_fea, x3, unknown) # contextual attention
x4 = self.layer3(x3) # N x 256 x 32 x 32
out = self.layer_bottleneck(x4) # N x 512 x 16 x 16
fea1 = self.shortcut[0](x) # input image and trimap
fea2 = self.shortcut[1](x1)
fea3 = self.shortcut[2](x2)
fea4 = self.shortcut[3](x3)
fea5 = self.shortcut[4](x4)
return out, {
'shortcut': (fea1, fea2, fea3, fea4, fea5),
'image_fea': im_fea,
'unknown': unknown,
}
class BasicBlock(nn.Layer):
expansion = 1
def __init__(self,
inplanes,
planes,
stride=1,
downsample=None,
norm_layer=None):
super().__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm
# Both self.conv1 and self.downsample layers downsample the input when stride != 1
self.conv1 = nn.utils.spectral_norm(conv3x3(inplanes, planes, stride))
self.bn1 = norm_layer(planes)
self.activation = nn.ReLU()
self.conv2 = nn.utils.spectral_norm(conv3x3(planes, planes))
self.bn2 = norm_layer(planes)
self.downsample = downsample
self.stride = stride
def forward(self, x):
identity = x
out = self.conv1(x)
out = self.bn1(out)
out = self.activation(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.activation(out)
return out
def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
"""3x3 convolution with padding"""
return nn.Conv2D(
in_planes,
out_planes,
kernel_size=3,
stride=stride,
padding=dilation,
groups=groups,
bias_attr=False,
dilation=dilation)
def conv1x1(in_planes, out_planes, stride=1):
"""1x1 convolution"""
return nn.Conv2D(
in_planes, out_planes, kernel_size=1, stride=stride, bias_attr=False)
|