briarmbg.py
| 1 | import torch |
| 2 | import torch.nn as nn |
| 3 | import torch.nn.functional as F |
| 4 | from transformers import PreTrainedModel |
| 5 | from .MyConfig import RMBGConfig |
| 6 | |
| 7 | class REBNCONV(nn.Module): |
| 8 | def __init__(self,in_ch=3,out_ch=3,dirate=1,stride=1): |
| 9 | super(REBNCONV,self).__init__() |
| 10 | |
| 11 | self.conv_s1 = nn.Conv2d(in_ch,out_ch,3,padding=1*dirate,dilation=1*dirate,stride=stride) |
| 12 | self.bn_s1 = nn.BatchNorm2d(out_ch) |
| 13 | self.relu_s1 = nn.ReLU(inplace=True) |
| 14 | |
| 15 | def forward(self,x): |
| 16 | |
| 17 | hx = x |
| 18 | xout = self.relu_s1(self.bn_s1(self.conv_s1(hx))) |
| 19 | |
| 20 | return xout |
| 21 | |
| 22 | ## upsample tensor 'src' to have the same spatial size with tensor 'tar' |
| 23 | def _upsample_like(src,tar): |
| 24 | |
| 25 | src = F.interpolate(src,size=tar.shape[2:],mode='bilinear') |
| 26 | |
| 27 | return src |
| 28 | |
| 29 | |
| 30 | ### RSU-7 ### |
| 31 | class RSU7(nn.Module): |
| 32 | |
| 33 | def __init__(self, in_ch=3, mid_ch=12, out_ch=3, img_size=512): |
| 34 | super(RSU7,self).__init__() |
| 35 | |
| 36 | self.in_ch = in_ch |
| 37 | self.mid_ch = mid_ch |
| 38 | self.out_ch = out_ch |
| 39 | |
| 40 | self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) ## 1 -> 1/2 |
| 41 | |
| 42 | self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1) |
| 43 | self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 44 | |
| 45 | self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 46 | self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 47 | |
| 48 | self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 49 | self.pool3 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 50 | |
| 51 | self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 52 | self.pool4 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 53 | |
| 54 | self.rebnconv5 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 55 | self.pool5 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 56 | |
| 57 | self.rebnconv6 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 58 | |
| 59 | self.rebnconv7 = REBNCONV(mid_ch,mid_ch,dirate=2) |
| 60 | |
| 61 | self.rebnconv6d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 62 | self.rebnconv5d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 63 | self.rebnconv4d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 64 | self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 65 | self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 66 | self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1) |
| 67 | |
| 68 | def forward(self,x): |
| 69 | b, c, h, w = x.shape |
| 70 | |
| 71 | hx = x |
| 72 | hxin = self.rebnconvin(hx) |
| 73 | |
| 74 | hx1 = self.rebnconv1(hxin) |
| 75 | hx = self.pool1(hx1) |
| 76 | |
| 77 | hx2 = self.rebnconv2(hx) |
| 78 | hx = self.pool2(hx2) |
| 79 | |
| 80 | hx3 = self.rebnconv3(hx) |
| 81 | hx = self.pool3(hx3) |
| 82 | |
| 83 | hx4 = self.rebnconv4(hx) |
| 84 | hx = self.pool4(hx4) |
| 85 | |
| 86 | hx5 = self.rebnconv5(hx) |
| 87 | hx = self.pool5(hx5) |
| 88 | |
| 89 | hx6 = self.rebnconv6(hx) |
| 90 | |
| 91 | hx7 = self.rebnconv7(hx6) |
| 92 | |
| 93 | hx6d = self.rebnconv6d(torch.cat((hx7,hx6),1)) |
| 94 | hx6dup = _upsample_like(hx6d,hx5) |
| 95 | |
| 96 | hx5d = self.rebnconv5d(torch.cat((hx6dup,hx5),1)) |
| 97 | hx5dup = _upsample_like(hx5d,hx4) |
| 98 | |
| 99 | hx4d = self.rebnconv4d(torch.cat((hx5dup,hx4),1)) |
| 100 | hx4dup = _upsample_like(hx4d,hx3) |
| 101 | |
| 102 | hx3d = self.rebnconv3d(torch.cat((hx4dup,hx3),1)) |
| 103 | hx3dup = _upsample_like(hx3d,hx2) |
| 104 | |
| 105 | hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1)) |
| 106 | hx2dup = _upsample_like(hx2d,hx1) |
| 107 | |
| 108 | hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1)) |
| 109 | |
| 110 | return hx1d + hxin |
| 111 | |
| 112 | |
| 113 | ### RSU-6 ### |
| 114 | class RSU6(nn.Module): |
| 115 | |
| 116 | def __init__(self, in_ch=3, mid_ch=12, out_ch=3): |
| 117 | super(RSU6,self).__init__() |
| 118 | |
| 119 | self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) |
| 120 | |
| 121 | self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1) |
| 122 | self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 123 | |
| 124 | self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 125 | self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 126 | |
| 127 | self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 128 | self.pool3 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 129 | |
| 130 | self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 131 | self.pool4 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 132 | |
| 133 | self.rebnconv5 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 134 | |
| 135 | self.rebnconv6 = REBNCONV(mid_ch,mid_ch,dirate=2) |
| 136 | |
| 137 | self.rebnconv5d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 138 | self.rebnconv4d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 139 | self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 140 | self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 141 | self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1) |
| 142 | |
| 143 | def forward(self,x): |
| 144 | |
| 145 | hx = x |
| 146 | |
| 147 | hxin = self.rebnconvin(hx) |
| 148 | |
| 149 | hx1 = self.rebnconv1(hxin) |
| 150 | hx = self.pool1(hx1) |
| 151 | |
| 152 | hx2 = self.rebnconv2(hx) |
| 153 | hx = self.pool2(hx2) |
| 154 | |
| 155 | hx3 = self.rebnconv3(hx) |
| 156 | hx = self.pool3(hx3) |
| 157 | |
| 158 | hx4 = self.rebnconv4(hx) |
| 159 | hx = self.pool4(hx4) |
| 160 | |
| 161 | hx5 = self.rebnconv5(hx) |
| 162 | |
| 163 | hx6 = self.rebnconv6(hx5) |
| 164 | |
| 165 | |
| 166 | hx5d = self.rebnconv5d(torch.cat((hx6,hx5),1)) |
| 167 | hx5dup = _upsample_like(hx5d,hx4) |
| 168 | |
| 169 | hx4d = self.rebnconv4d(torch.cat((hx5dup,hx4),1)) |
| 170 | hx4dup = _upsample_like(hx4d,hx3) |
| 171 | |
| 172 | hx3d = self.rebnconv3d(torch.cat((hx4dup,hx3),1)) |
| 173 | hx3dup = _upsample_like(hx3d,hx2) |
| 174 | |
| 175 | hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1)) |
| 176 | hx2dup = _upsample_like(hx2d,hx1) |
| 177 | |
| 178 | hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1)) |
| 179 | |
| 180 | return hx1d + hxin |
| 181 | |
| 182 | ### RSU-5 ### |
| 183 | class RSU5(nn.Module): |
| 184 | |
| 185 | def __init__(self, in_ch=3, mid_ch=12, out_ch=3): |
| 186 | super(RSU5,self).__init__() |
| 187 | |
| 188 | self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) |
| 189 | |
| 190 | self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1) |
| 191 | self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 192 | |
| 193 | self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 194 | self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 195 | |
| 196 | self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 197 | self.pool3 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 198 | |
| 199 | self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 200 | |
| 201 | self.rebnconv5 = REBNCONV(mid_ch,mid_ch,dirate=2) |
| 202 | |
| 203 | self.rebnconv4d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 204 | self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 205 | self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 206 | self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1) |
| 207 | |
| 208 | def forward(self,x): |
| 209 | |
| 210 | hx = x |
| 211 | |
| 212 | hxin = self.rebnconvin(hx) |
| 213 | |
| 214 | hx1 = self.rebnconv1(hxin) |
| 215 | hx = self.pool1(hx1) |
| 216 | |
| 217 | hx2 = self.rebnconv2(hx) |
| 218 | hx = self.pool2(hx2) |
| 219 | |
| 220 | hx3 = self.rebnconv3(hx) |
| 221 | hx = self.pool3(hx3) |
| 222 | |
| 223 | hx4 = self.rebnconv4(hx) |
| 224 | |
| 225 | hx5 = self.rebnconv5(hx4) |
| 226 | |
| 227 | hx4d = self.rebnconv4d(torch.cat((hx5,hx4),1)) |
| 228 | hx4dup = _upsample_like(hx4d,hx3) |
| 229 | |
| 230 | hx3d = self.rebnconv3d(torch.cat((hx4dup,hx3),1)) |
| 231 | hx3dup = _upsample_like(hx3d,hx2) |
| 232 | |
| 233 | hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1)) |
| 234 | hx2dup = _upsample_like(hx2d,hx1) |
| 235 | |
| 236 | hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1)) |
| 237 | |
| 238 | return hx1d + hxin |
| 239 | |
| 240 | ### RSU-4 ### |
| 241 | class RSU4(nn.Module): |
| 242 | |
| 243 | def __init__(self, in_ch=3, mid_ch=12, out_ch=3): |
| 244 | super(RSU4,self).__init__() |
| 245 | |
| 246 | self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) |
| 247 | |
| 248 | self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1) |
| 249 | self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 250 | |
| 251 | self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 252 | self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 253 | |
| 254 | self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1) |
| 255 | |
| 256 | self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=2) |
| 257 | |
| 258 | self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 259 | self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1) |
| 260 | self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1) |
| 261 | |
| 262 | def forward(self,x): |
| 263 | |
| 264 | hx = x |
| 265 | |
| 266 | hxin = self.rebnconvin(hx) |
| 267 | |
| 268 | hx1 = self.rebnconv1(hxin) |
| 269 | hx = self.pool1(hx1) |
| 270 | |
| 271 | hx2 = self.rebnconv2(hx) |
| 272 | hx = self.pool2(hx2) |
| 273 | |
| 274 | hx3 = self.rebnconv3(hx) |
| 275 | |
| 276 | hx4 = self.rebnconv4(hx3) |
| 277 | |
| 278 | hx3d = self.rebnconv3d(torch.cat((hx4,hx3),1)) |
| 279 | hx3dup = _upsample_like(hx3d,hx2) |
| 280 | |
| 281 | hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1)) |
| 282 | hx2dup = _upsample_like(hx2d,hx1) |
| 283 | |
| 284 | hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1)) |
| 285 | |
| 286 | return hx1d + hxin |
| 287 | |
| 288 | ### RSU-4F ### |
| 289 | class RSU4F(nn.Module): |
| 290 | |
| 291 | def __init__(self, in_ch=3, mid_ch=12, out_ch=3): |
| 292 | super(RSU4F,self).__init__() |
| 293 | |
| 294 | self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) |
| 295 | |
| 296 | self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1) |
| 297 | self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=2) |
| 298 | self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=4) |
| 299 | |
| 300 | self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=8) |
| 301 | |
| 302 | self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=4) |
| 303 | self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=2) |
| 304 | self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1) |
| 305 | |
| 306 | def forward(self,x): |
| 307 | |
| 308 | hx = x |
| 309 | |
| 310 | hxin = self.rebnconvin(hx) |
| 311 | |
| 312 | hx1 = self.rebnconv1(hxin) |
| 313 | hx2 = self.rebnconv2(hx1) |
| 314 | hx3 = self.rebnconv3(hx2) |
| 315 | |
| 316 | hx4 = self.rebnconv4(hx3) |
| 317 | |
| 318 | hx3d = self.rebnconv3d(torch.cat((hx4,hx3),1)) |
| 319 | hx2d = self.rebnconv2d(torch.cat((hx3d,hx2),1)) |
| 320 | hx1d = self.rebnconv1d(torch.cat((hx2d,hx1),1)) |
| 321 | |
| 322 | return hx1d + hxin |
| 323 | |
| 324 | |
| 325 | class myrebnconv(nn.Module): |
| 326 | def __init__(self, in_ch=3, |
| 327 | out_ch=1, |
| 328 | kernel_size=3, |
| 329 | stride=1, |
| 330 | padding=1, |
| 331 | dilation=1, |
| 332 | groups=1): |
| 333 | super(myrebnconv,self).__init__() |
| 334 | |
| 335 | self.conv = nn.Conv2d(in_ch, |
| 336 | out_ch, |
| 337 | kernel_size=kernel_size, |
| 338 | stride=stride, |
| 339 | padding=padding, |
| 340 | dilation=dilation, |
| 341 | groups=groups) |
| 342 | self.bn = nn.BatchNorm2d(out_ch) |
| 343 | self.rl = nn.ReLU(inplace=True) |
| 344 | |
| 345 | def forward(self,x): |
| 346 | return self.rl(self.bn(self.conv(x))) |
| 347 | |
| 348 | |
| 349 | class BriaRMBG(PreTrainedModel): |
| 350 | config_class = RMBGConfig |
| 351 | def __init__(self,config:RMBGConfig = RMBGConfig()): |
| 352 | super().__init__(config) |
| 353 | in_ch = config.in_ch # 3 |
| 354 | out_ch = config.out_ch # 1 |
| 355 | self.conv_in = nn.Conv2d(in_ch,64,3,stride=2,padding=1) |
| 356 | self.pool_in = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 357 | |
| 358 | self.stage1 = RSU7(64,32,64) |
| 359 | self.pool12 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 360 | |
| 361 | self.stage2 = RSU6(64,32,128) |
| 362 | self.pool23 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 363 | |
| 364 | self.stage3 = RSU5(128,64,256) |
| 365 | self.pool34 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 366 | |
| 367 | self.stage4 = RSU4(256,128,512) |
| 368 | self.pool45 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 369 | |
| 370 | self.stage5 = RSU4F(512,256,512) |
| 371 | self.pool56 = nn.MaxPool2d(2,stride=2,ceil_mode=True) |
| 372 | |
| 373 | self.stage6 = RSU4F(512,256,512) |
| 374 | |
| 375 | # decoder |
| 376 | self.stage5d = RSU4F(1024,256,512) |
| 377 | self.stage4d = RSU4(1024,128,256) |
| 378 | self.stage3d = RSU5(512,64,128) |
| 379 | self.stage2d = RSU6(256,32,64) |
| 380 | self.stage1d = RSU7(128,16,64) |
| 381 | |
| 382 | self.side1 = nn.Conv2d(64,out_ch,3,padding=1) |
| 383 | self.side2 = nn.Conv2d(64,out_ch,3,padding=1) |
| 384 | self.side3 = nn.Conv2d(128,out_ch,3,padding=1) |
| 385 | self.side4 = nn.Conv2d(256,out_ch,3,padding=1) |
| 386 | self.side5 = nn.Conv2d(512,out_ch,3,padding=1) |
| 387 | self.side6 = nn.Conv2d(512,out_ch,3,padding=1) |
| 388 | |
| 389 | # self.outconv = nn.Conv2d(6*out_ch,out_ch,1) |
| 390 | |
| 391 | def forward(self,x): |
| 392 | |
| 393 | hx = x |
| 394 | |
| 395 | hxin = self.conv_in(hx) |
| 396 | #hx = self.pool_in(hxin) |
| 397 | |
| 398 | #stage 1 |
| 399 | hx1 = self.stage1(hxin) |
| 400 | hx = self.pool12(hx1) |
| 401 | |
| 402 | #stage 2 |
| 403 | hx2 = self.stage2(hx) |
| 404 | hx = self.pool23(hx2) |
| 405 | |
| 406 | #stage 3 |
| 407 | hx3 = self.stage3(hx) |
| 408 | hx = self.pool34(hx3) |
| 409 | |
| 410 | #stage 4 |
| 411 | hx4 = self.stage4(hx) |
| 412 | hx = self.pool45(hx4) |
| 413 | |
| 414 | #stage 5 |
| 415 | hx5 = self.stage5(hx) |
| 416 | hx = self.pool56(hx5) |
| 417 | |
| 418 | #stage 6 |
| 419 | hx6 = self.stage6(hx) |
| 420 | hx6up = _upsample_like(hx6,hx5) |
| 421 | |
| 422 | #-------------------- decoder -------------------- |
| 423 | hx5d = self.stage5d(torch.cat((hx6up,hx5),1)) |
| 424 | hx5dup = _upsample_like(hx5d,hx4) |
| 425 | |
| 426 | hx4d = self.stage4d(torch.cat((hx5dup,hx4),1)) |
| 427 | hx4dup = _upsample_like(hx4d,hx3) |
| 428 | |
| 429 | hx3d = self.stage3d(torch.cat((hx4dup,hx3),1)) |
| 430 | hx3dup = _upsample_like(hx3d,hx2) |
| 431 | |
| 432 | hx2d = self.stage2d(torch.cat((hx3dup,hx2),1)) |
| 433 | hx2dup = _upsample_like(hx2d,hx1) |
| 434 | |
| 435 | hx1d = self.stage1d(torch.cat((hx2dup,hx1),1)) |
| 436 | |
| 437 | |
| 438 | #side output |
| 439 | d1 = self.side1(hx1d) |
| 440 | d1 = _upsample_like(d1,x) |
| 441 | |
| 442 | d2 = self.side2(hx2d) |
| 443 | d2 = _upsample_like(d2,x) |
| 444 | |
| 445 | d3 = self.side3(hx3d) |
| 446 | d3 = _upsample_like(d3,x) |
| 447 | |
| 448 | d4 = self.side4(hx4d) |
| 449 | d4 = _upsample_like(d4,x) |
| 450 | |
| 451 | d5 = self.side5(hx5d) |
| 452 | d5 = _upsample_like(d5,x) |
| 453 | |
| 454 | d6 = self.side6(hx6) |
| 455 | d6 = _upsample_like(d6,x) |
| 456 | |
| 457 | return [F.sigmoid(d1), F.sigmoid(d2), F.sigmoid(d3), F.sigmoid(d4), F.sigmoid(d5), F.sigmoid(d6)],[hx1d,hx2d,hx3d,hx4d,hx5d,hx6] |
| 458 | |
| 459 | |