├── .gitignore
    ├── MobileNetV2_deploy.prototxt
    ├── benchmark_mobilenetv2.sh
    ├── conv_dw_layer.cpp
    ├── conv_dw_layer.cu
    ├── conv_dw_layer.hpp
    └── mobilenet_v2_new.caffemodel
├── LICENSE
├── MobileUnet1by2.prototxt
├── README.md
├── SmallRefineDet_mobilev2_ssdlite.prototxt
└── forMobileUNet
    ├── caffe.proto
    ├── dice_coef_loss_layer.cpp
    ├── dice_coef_loss_layer.cu
    ├── dice_coef_loss_layer.hpp
    ├── hello.txt
    ├── resize_layer.cpp
    ├── resize_layer.cu
    └── resize_layer.hpp


/.gitignore/MobileNetV2_deploy.prototxt:
--------------------------------------------------------------------------------
   1 | name: "MOBILENETV2"
   2 | 
   3 | #input: "data"
   4 | #input_dim: 1
   5 | #input_dim: 3
   6 | #input_dim: 224
   7 | #input_dim: 224
   8 | 
   9 | layer {
  10 |   name: "data"
  11 |   type: "Data"
  12 |   top: "data"
  13 |   top: "label"
  14 |   include {
  15 |     phase: TEST
  16 |   }
  17 |   transform_param {
  18 |     crop_size: 224
  19 |     scale: 0.017
  20 |     mean_value: 103.94
  21 |     mean_value: 116.78
  22 |     mean_value: 123.68
  23 |   }
  24 |   data_param {
  25 |     source: "****/ImageNet2012/lmdb/ilsvrc12_val_lmdb"
  26 |     batch_size: 25
  27 |     backend: LMDB
  28 |   }
  29 | }
  30 | 
  31 | layer {
  32 |   name: "conv1_new"
  33 |   type: "Convolution"
  34 |   bottom: "data"
  35 |   top: "conv1_new"
  36 |   param {
  37 |     lr_mult: 1
  38 |     decay_mult: 1
  39 |   }
  40 |   convolution_param {
  41 |     num_output: 32
  42 |     bias_term: false
  43 |     pad: 1
  44 |     kernel_size: 3
  45 |     stride: 2
  46 |     weight_filler {
  47 |       type: "msra"
  48 |     }
  49 |   }
  50 | }
  51 | layer {
  52 |   name: "conv1/bn_new"
  53 |   type: "BatchNorm"
  54 |   bottom: "conv1_new"
  55 |   top: "conv1_new"
  56 |   param {
  57 |     lr_mult: 0
  58 |     decay_mult: 0
  59 |   }
  60 |   param {
  61 |     lr_mult: 0
  62 |     decay_mult: 0
  63 |   }
  64 |   param {
  65 |     lr_mult: 0
  66 |     decay_mult: 0
  67 |   }
  68 | }
  69 | layer {
  70 |   name: "conv1/scale_new"
  71 |   type: "Scale"
  72 |   bottom: "conv1_new"
  73 |   top: "conv1_new"
  74 |   param {
  75 |     lr_mult: 1
  76 |     decay_mult: 0
  77 |   }
  78 |   param {
  79 |     lr_mult: 1
  80 |     decay_mult: 0
  81 |   }
  82 |   scale_param {
  83 |     filler {
  84 |       value: 1
  85 |     }
  86 |     bias_term: true
  87 |     bias_filler {
  88 |       value: 0
  89 |     }
  90 |   }
  91 | }
  92 | layer {
  93 |   name: "relu1_new"
  94 |   type: "ReLU"
  95 |   bottom: "conv1_new"
  96 |   top: "conv1_new"
  97 | }
  98 | 
  99 | ##################################################################
 100 | 
 101 | layer {
 102 |   name: "conv1_1/in/pw_new"
 103 |   type: "Convolution"
 104 |   bottom: "conv1_new"
 105 |   top: "conv1_1/in/pw_new"
 106 |   param {
 107 |     lr_mult: 1
 108 |     decay_mult: 1
 109 |   }
 110 |   convolution_param {
 111 |     num_output: 32
 112 |     bias_term: false
 113 |     pad: 0
 114 |     kernel_size: 1
 115 |     engine: CAFFE
 116 |     stride: 1
 117 |     weight_filler {
 118 |       type: "msra"
 119 |     }
 120 |   }
 121 | }
 122 | layer {
 123 |   name: "conv1_1/in/pw/bn_new"
 124 |   type: "BatchNorm"
 125 |   bottom: "conv1_1/in/pw_new"
 126 |   top: "conv1_1/in/pw_new"
 127 |   param {
 128 |     lr_mult: 0
 129 |     decay_mult: 0
 130 |   }
 131 |   param {
 132 |     lr_mult: 0
 133 |     decay_mult: 0
 134 |   }
 135 |   param {
 136 |     lr_mult: 0
 137 |     decay_mult: 0
 138 |   }
 139 | }
 140 | layer {
 141 |   name: "conv1_1/in/pw/scale_new"
 142 |   type: "Scale"
 143 |   bottom: "conv1_1/in/pw_new"
 144 |   top: "conv1_1/in/pw_new"
 145 |   param {
 146 |     lr_mult: 1
 147 |     decay_mult: 0
 148 |   }
 149 |   param {
 150 |     lr_mult: 1
 151 |     decay_mult: 0
 152 |   }
 153 |   scale_param {
 154 |     filler {
 155 |       value: 1
 156 |     }
 157 |     bias_term: true
 158 |     bias_filler {
 159 |       value: 0
 160 |     }
 161 |   }
 162 | }
 163 | layer {
 164 |   name: "relu1_1/in/pw_new"
 165 |   type: "ReLU"
 166 |   bottom: "conv1_1/in/pw_new"
 167 |   top: "conv1_1/in/pw_new"
 168 | }
 169 | layer {
 170 |   name: "conv1_1/dw_new"
 171 |   type: "ConvolutionDepthwise"
 172 |   bottom: "conv1_1/in/pw_new"
 173 |   top: "conv1_1/dw_new"
 174 |   param {
 175 |     lr_mult: 1
 176 |     decay_mult: 1
 177 |   }
 178 |   convolution_param {
 179 |     num_output: 32
 180 |     bias_term: false
 181 |     pad: 1
 182 |     kernel_size: 3
 183 |     engine: CAFFE
 184 |     stride: 1
 185 |     weight_filler {
 186 |       type: "msra"
 187 |     }
 188 |   }
 189 | }
 190 | layer {
 191 |   name: "conv1_1/dw/bn_new"
 192 |   type: "BatchNorm"
 193 |   bottom: "conv1_1/dw_new"
 194 |   top: "conv1_1/dw_new"
 195 |   param {
 196 |     lr_mult: 0
 197 |     decay_mult: 0
 198 |   }
 199 |   param {
 200 |     lr_mult: 0
 201 |     decay_mult: 0
 202 |   }
 203 |   param {
 204 |     lr_mult: 0
 205 |     decay_mult: 0
 206 |   }
 207 | }
 208 | layer {
 209 |   name: "conv1_1/dw/scale_new"
 210 |   type: "Scale"
 211 |   bottom: "conv1_1/dw_new"
 212 |   top: "conv1_1/dw_new"
 213 |   param {
 214 |     lr_mult: 1
 215 |     decay_mult: 0
 216 |   }
 217 |   param {
 218 |     lr_mult: 1
 219 |     decay_mult: 0
 220 |   }
 221 |   scale_param {
 222 |     filler {
 223 |       value: 1
 224 |     }
 225 |     bias_term: true
 226 |     bias_filler {
 227 |       value: 0
 228 |     }
 229 |   }
 230 | }
 231 | layer {
 232 |   name: "relu1_1/dw_new"
 233 |   type: "ReLU"
 234 |   bottom: "conv1_1/dw_new"
 235 |   top: "conv1_1/dw_new"
 236 | }
 237 | layer {
 238 |   name: "conv1_1/out/pw_new"
 239 |   type: "Convolution"
 240 |   bottom: "conv1_1/dw_new"
 241 |   top: "conv1_1/out/pw_new"
 242 |   param {
 243 |     lr_mult: 1
 244 |     decay_mult: 1
 245 |   }
 246 |   convolution_param {
 247 |     num_output: 16
 248 |     bias_term: false
 249 |     pad: 0
 250 |     kernel_size: 1
 251 |     engine: CAFFE
 252 |     stride: 1
 253 |     weight_filler {
 254 |       type: "msra"
 255 |     }
 256 |   }
 257 | }
 258 | layer {
 259 |   name: "conv1_1/out/pw/bn_new"
 260 |   type: "BatchNorm"
 261 |   bottom: "conv1_1/out/pw_new"
 262 |   top: "conv1_1/out/pw_new"
 263 |   param {
 264 |     lr_mult: 0
 265 |     decay_mult: 0
 266 |   }
 267 |   param {
 268 |     lr_mult: 0
 269 |     decay_mult: 0
 270 |   }
 271 |   param {
 272 |     lr_mult: 0
 273 |     decay_mult: 0
 274 |   }
 275 | }
 276 | layer {
 277 |   name: "conv1_1/out/pw/scale_new"
 278 |   type: "Scale"
 279 |   bottom: "conv1_1/out/pw_new"
 280 |   top: "conv1_1/out/pw_new"
 281 |   param {
 282 |     lr_mult: 1
 283 |     decay_mult: 0
 284 |   }
 285 |   param {
 286 |     lr_mult: 1
 287 |     decay_mult: 0
 288 |   }
 289 |   scale_param {
 290 |     filler {
 291 |       value: 1
 292 |     }
 293 |     bias_term: true
 294 |     bias_filler {
 295 |       value: 0
 296 |     }
 297 |   }
 298 | }
 299 | ##################################################################
 300 | layer {
 301 |   name: "conv2_1/in/pw_new"
 302 |   type: "Convolution"
 303 |   bottom: "conv1_1/out/pw_new"
 304 |   top: "conv2_1/in/pw_new"
 305 |   param {
 306 |     lr_mult: 1
 307 |     decay_mult: 1
 308 |   }
 309 |   convolution_param {
 310 |     num_output: 96
 311 |     bias_term: false
 312 |     pad: 0
 313 |     kernel_size: 1
 314 |     engine: CAFFE
 315 |     stride: 1
 316 |     weight_filler {
 317 |       type: "msra"
 318 |     }
 319 |   }
 320 | }
 321 | layer {
 322 |   name: "conv2_1/in/pw/bn_new"
 323 |   type: "BatchNorm"
 324 |   bottom: "conv2_1/in/pw_new"
 325 |   top: "conv2_1/in/pw_new"
 326 |   param {
 327 |     lr_mult: 0
 328 |     decay_mult: 0
 329 |   }
 330 |   param {
 331 |     lr_mult: 0
 332 |     decay_mult: 0
 333 |   }
 334 |   param {
 335 |     lr_mult: 0
 336 |     decay_mult: 0
 337 |   }
 338 | }
 339 | layer {
 340 |   name: "conv2_1/in/pw/scale_new"
 341 |   type: "Scale"
 342 |   bottom: "conv2_1/in/pw_new"
 343 |   top: "conv2_1/in/pw_new"
 344 |   param {
 345 |     lr_mult: 1
 346 |     decay_mult: 0
 347 |   }
 348 |   param {
 349 |     lr_mult: 1
 350 |     decay_mult: 0
 351 |   }
 352 |   scale_param {
 353 |     filler {
 354 |       value: 1
 355 |     }
 356 |     bias_term: true
 357 |     bias_filler {
 358 |       value: 0
 359 |     }
 360 |   }
 361 | }
 362 | layer {
 363 |   name: "relu2_1/in/pw_new"
 364 |   type: "ReLU"
 365 |   bottom: "conv2_1/in/pw_new"
 366 |   top: "conv2_1/in/pw_new"
 367 | }
 368 | layer {
 369 |   name: "conv2_1/dw_new"
 370 |   type: "ConvolutionDepthwise"
 371 |   bottom: "conv2_1/in/pw_new"
 372 |   top: "conv2_1/dw_new"
 373 |   param {
 374 |     lr_mult: 1
 375 |     decay_mult: 1
 376 |   }
 377 |   convolution_param {
 378 |     num_output: 96
 379 |     bias_term: false
 380 |     pad: 1
 381 |     kernel_size: 3
 382 |     engine: CAFFE
 383 |     stride: 2
 384 |     weight_filler {
 385 |       type: "msra"
 386 |     }
 387 |   }
 388 | }
 389 | layer {
 390 |   name: "conv2_1/dw/bn_new"
 391 |   type: "BatchNorm"
 392 |   bottom: "conv2_1/dw_new"
 393 |   top: "conv2_1/dw_new"
 394 |   param {
 395 |     lr_mult: 0
 396 |     decay_mult: 0
 397 |   }
 398 |   param {
 399 |     lr_mult: 0
 400 |     decay_mult: 0
 401 |   }
 402 |   param {
 403 |     lr_mult: 0
 404 |     decay_mult: 0
 405 |   }
 406 | }
 407 | layer {
 408 |   name: "conv2_1/dw/scale_new"
 409 |   type: "Scale"
 410 |   bottom: "conv2_1/dw_new"
 411 |   top: "conv2_1/dw_new"
 412 |   param {
 413 |     lr_mult: 1
 414 |     decay_mult: 0
 415 |   }
 416 |   param {
 417 |     lr_mult: 1
 418 |     decay_mult: 0
 419 |   }
 420 |   scale_param {
 421 |     filler {
 422 |       value: 1
 423 |     }
 424 |     bias_term: true
 425 |     bias_filler {
 426 |       value: 0
 427 |     }
 428 |   }
 429 | }
 430 | layer {
 431 |   name: "relu2_1/dw_new"
 432 |   type: "ReLU"
 433 |   bottom: "conv2_1/dw_new"
 434 |   top: "conv2_1/dw_new"
 435 | }
 436 | layer {
 437 |   name: "conv2_1/out/pw_new"
 438 |   type: "Convolution"
 439 |   bottom: "conv2_1/dw_new"
 440 |   top: "conv2_1/out/pw_new"
 441 |   param {
 442 |     lr_mult: 1
 443 |     decay_mult: 1
 444 |   }
 445 |   convolution_param {
 446 |     num_output: 24
 447 |     bias_term: false
 448 |     pad: 0
 449 |     kernel_size: 1
 450 |     engine: CAFFE
 451 |     stride: 1
 452 |     weight_filler {
 453 |       type: "msra"
 454 |     }
 455 |   }
 456 | }
 457 | layer {
 458 |   name: "conv2_1/out/pw/bn_new"
 459 |   type: "BatchNorm"
 460 |   bottom: "conv2_1/out/pw_new"
 461 |   top: "conv2_1/out/pw_new"
 462 |   param {
 463 |     lr_mult: 0
 464 |     decay_mult: 0
 465 |   }
 466 |   param {
 467 |     lr_mult: 0
 468 |     decay_mult: 0
 469 |   }
 470 |   param {
 471 |     lr_mult: 0
 472 |     decay_mult: 0
 473 |   }
 474 | }
 475 | layer {
 476 |   name: "conv2_1/out/pw/scale_new"
 477 |   type: "Scale"
 478 |   bottom: "conv2_1/out/pw_new"
 479 |   top: "conv2_1/out/pw_new"
 480 |   param {
 481 |     lr_mult: 1
 482 |     decay_mult: 0
 483 |   }
 484 |   param {
 485 |     lr_mult: 1
 486 |     decay_mult: 0
 487 |   }
 488 |   scale_param {
 489 |     filler {
 490 |       value: 1
 491 |     }
 492 |     bias_term: true
 493 |     bias_filler {
 494 |       value: 0
 495 |     }
 496 |   }
 497 | }
 498 | layer {
 499 |   name: "conv2_2/in/pw_new"
 500 |   type: "Convolution"
 501 |   bottom: "conv2_1/out/pw_new"
 502 |   top: "conv2_2/in/pw_new"
 503 |   param {
 504 |     lr_mult: 1
 505 |     decay_mult: 1
 506 |   }
 507 |   convolution_param {
 508 |     num_output: 144
 509 |     bias_term: false
 510 |     pad: 0
 511 |     kernel_size: 1
 512 |     engine: CAFFE
 513 |     stride: 1
 514 |     weight_filler {
 515 |       type: "msra"
 516 |     }
 517 |   }
 518 | }
 519 | layer {
 520 |   name: "conv2_2/in/pw/bn_new"
 521 |   type: "BatchNorm"
 522 |   bottom: "conv2_2/in/pw_new"
 523 |   top: "conv2_2/in/pw_new"
 524 |   param {
 525 |     lr_mult: 0
 526 |     decay_mult: 0
 527 |   }
 528 |   param {
 529 |     lr_mult: 0
 530 |     decay_mult: 0
 531 |   }
 532 |   param {
 533 |     lr_mult: 0
 534 |     decay_mult: 0
 535 |   }
 536 | }
 537 | layer {
 538 |   name: "conv2_2/in/pw/scale_new"
 539 |   type: "Scale"
 540 |   bottom: "conv2_2/in/pw_new"
 541 |   top: "conv2_2/in/pw_new"
 542 |   param {
 543 |     lr_mult: 1
 544 |     decay_mult: 0
 545 |   }
 546 |   param {
 547 |     lr_mult: 1
 548 |     decay_mult: 0
 549 |   }
 550 |   scale_param {
 551 |     filler {
 552 |       value: 1
 553 |     }
 554 |     bias_term: true
 555 |     bias_filler {
 556 |       value: 0
 557 |     }
 558 |   }
 559 | }
 560 | layer {
 561 |   name: "relu2_2/in/pw_new"
 562 |   type: "ReLU"
 563 |   bottom: "conv2_2/in/pw_new"
 564 |   top: "conv2_2/in/pw_new"
 565 | }
 566 | layer {
 567 |   name: "conv2_2/dw_new"
 568 |   type: "ConvolutionDepthwise"
 569 |   bottom: "conv2_2/in/pw_new"
 570 |   top: "conv2_2/dw_new"
 571 |   param {
 572 |     lr_mult: 1
 573 |     decay_mult: 1
 574 |   }
 575 |   convolution_param {
 576 |     num_output: 144
 577 |     bias_term: false
 578 |     pad: 1
 579 |     kernel_size: 3
 580 |     engine: CAFFE
 581 |     stride: 1
 582 |     weight_filler {
 583 |       type: "msra"
 584 |     }
 585 |   }
 586 | }
 587 | layer {
 588 |   name: "conv2_2/dw/bn_new"
 589 |   type: "BatchNorm"
 590 |   bottom: "conv2_2/dw_new"
 591 |   top: "conv2_2/dw_new"
 592 |   param {
 593 |     lr_mult: 0
 594 |     decay_mult: 0
 595 |   }
 596 |   param {
 597 |     lr_mult: 0
 598 |     decay_mult: 0
 599 |   }
 600 |   param {
 601 |     lr_mult: 0
 602 |     decay_mult: 0
 603 |   }
 604 | }
 605 | layer {
 606 |   name: "conv2_2/dw/scale_new"
 607 |   type: "Scale"
 608 |   bottom: "conv2_2/dw_new"
 609 |   top: "conv2_2/dw_new"
 610 |   param {
 611 |     lr_mult: 1
 612 |     decay_mult: 0
 613 |   }
 614 |   param {
 615 |     lr_mult: 1
 616 |     decay_mult: 0
 617 |   }
 618 |   scale_param {
 619 |     filler {
 620 |       value: 1
 621 |     }
 622 |     bias_term: true
 623 |     bias_filler {
 624 |       value: 0
 625 |     }
 626 |   }
 627 | }
 628 | layer {
 629 |   name: "relu2_2/dw_new"
 630 |   type: "ReLU"
 631 |   bottom: "conv2_2/dw_new"
 632 |   top: "conv2_2/dw_new"
 633 | }
 634 | layer {
 635 |   name: "conv2_2/out/pw_new"
 636 |   type: "Convolution"
 637 |   bottom: "conv2_2/dw_new"
 638 |   top: "conv2_2/out/pw_new"
 639 |   param {
 640 |     lr_mult: 1
 641 |     decay_mult: 1
 642 |   }
 643 |   convolution_param {
 644 |     num_output: 24
 645 |     bias_term: false
 646 |     pad: 0
 647 |     kernel_size: 1
 648 |     engine: CAFFE
 649 |     stride: 1
 650 |     weight_filler {
 651 |       type: "msra"
 652 |     }
 653 |   }
 654 | }
 655 | layer {
 656 |   name: "conv2_2/out/pw/bn_new"
 657 |   type: "BatchNorm"
 658 |   bottom: "conv2_2/out/pw_new"
 659 |   top: "conv2_2/out/pw_new"
 660 |   param {
 661 |     lr_mult: 0
 662 |     decay_mult: 0
 663 |   }
 664 |   param {
 665 |     lr_mult: 0
 666 |     decay_mult: 0
 667 |   }
 668 |   param {
 669 |     lr_mult: 0
 670 |     decay_mult: 0
 671 |   }
 672 | }
 673 | layer {
 674 |   name: "conv2_2/out/pw/scale_new"
 675 |   type: "Scale"
 676 |   bottom: "conv2_2/out/pw_new"
 677 |   top: "conv2_2/out/pw_new"
 678 |   param {
 679 |     lr_mult: 1
 680 |     decay_mult: 0
 681 |   }
 682 |   param {
 683 |     lr_mult: 1
 684 |     decay_mult: 0
 685 |   }
 686 |   scale_param {
 687 |     filler {
 688 |       value: 1
 689 |     }
 690 |     bias_term: true
 691 |     bias_filler {
 692 |       value: 0
 693 |     }
 694 |   }
 695 | }
 696 | layer {
 697 |   name: "fuse_conv2_2"
 698 |   type: "Eltwise"
 699 |   bottom: "conv2_1/out/pw_new"
 700 |   bottom: "conv2_2/out/pw_new"
 701 |   top: "fuse_conv2_2"
 702 |   eltwise_param {
 703 |     operation: SUM
 704 |   }
 705 | }
 706 | 
 707 | ###############################################
 708 | 
 709 | layer {
 710 |   name: "conv3_1/in/pw_new"
 711 |   type: "Convolution"
 712 |   bottom: "fuse_conv2_2"
 713 |   top: "conv3_1/in/pw_new"
 714 |   param {
 715 |     lr_mult: 1
 716 |     decay_mult: 1
 717 |   }
 718 |   convolution_param {
 719 |     num_output: 144
 720 |     bias_term: false
 721 |     pad: 0
 722 |     kernel_size: 1
 723 |     engine: CAFFE
 724 |     stride: 1
 725 |     weight_filler {
 726 |       type: "msra"
 727 |     }
 728 |   }
 729 | }
 730 | layer {
 731 |   name: "conv3_1/in/pw/bn_new"
 732 |   type: "BatchNorm"
 733 |   bottom: "conv3_1/in/pw_new"
 734 |   top: "conv3_1/in/pw_new"
 735 |   param {
 736 |     lr_mult: 0
 737 |     decay_mult: 0
 738 |   }
 739 |   param {
 740 |     lr_mult: 0
 741 |     decay_mult: 0
 742 |   }
 743 |   param {
 744 |     lr_mult: 0
 745 |     decay_mult: 0
 746 |   }
 747 | }
 748 | layer {
 749 |   name: "conv3_1/in/pw/scale_new"
 750 |   type: "Scale"
 751 |   bottom: "conv3_1/in/pw_new"
 752 |   top: "conv3_1/in/pw_new"
 753 |   param {
 754 |     lr_mult: 1
 755 |     decay_mult: 0
 756 |   }
 757 |   param {
 758 |     lr_mult: 1
 759 |     decay_mult: 0
 760 |   }
 761 |   scale_param {
 762 |     filler {
 763 |       value: 1
 764 |     }
 765 |     bias_term: true
 766 |     bias_filler {
 767 |       value: 0
 768 |     }
 769 |   }
 770 | }
 771 | layer {
 772 |   name: "relu3_1/in/pw_new"
 773 |   type: "ReLU"
 774 |   bottom: "conv3_1/in/pw_new"
 775 |   top: "conv3_1/in/pw_new"
 776 | }
 777 | layer {
 778 |   name: "conv3_1/dw_new"
 779 |   type: "ConvolutionDepthwise"
 780 |   bottom: "conv3_1/in/pw_new"
 781 |   top: "conv3_1/dw_new"
 782 |   param {
 783 |     lr_mult: 1
 784 |     decay_mult: 1
 785 |   }
 786 |   convolution_param {
 787 |     num_output: 144
 788 |     bias_term: false
 789 |     pad: 1
 790 |     kernel_size: 3
 791 |     engine: CAFFE
 792 |     stride: 2
 793 |     weight_filler {
 794 |       type: "msra"
 795 |     }
 796 |   }
 797 | }
 798 | layer {
 799 |   name: "conv3_1/dw/bn_new"
 800 |   type: "BatchNorm"
 801 |   bottom: "conv3_1/dw_new"
 802 |   top: "conv3_1/dw_new"
 803 |   param {
 804 |     lr_mult: 0
 805 |     decay_mult: 0
 806 |   }
 807 |   param {
 808 |     lr_mult: 0
 809 |     decay_mult: 0
 810 |   }
 811 |   param {
 812 |     lr_mult: 0
 813 |     decay_mult: 0
 814 |   }
 815 | }
 816 | layer {
 817 |   name: "conv3_1/dw/scale_new"
 818 |   type: "Scale"
 819 |   bottom: "conv3_1/dw_new"
 820 |   top: "conv3_1/dw_new"
 821 |   param {
 822 |     lr_mult: 1
 823 |     decay_mult: 0
 824 |   }
 825 |   param {
 826 |     lr_mult: 1
 827 |     decay_mult: 0
 828 |   }
 829 |   scale_param {
 830 |     filler {
 831 |       value: 1
 832 |     }
 833 |     bias_term: true
 834 |     bias_filler {
 835 |       value: 0
 836 |     }
 837 |   }
 838 | }
 839 | layer {
 840 |   name: "relu3_1/dw_new"
 841 |   type: "ReLU"
 842 |   bottom: "conv3_1/dw_new"
 843 |   top: "conv3_1/dw_new"
 844 | }
 845 | layer {
 846 |   name: "conv3_1/out/pw_new"
 847 |   type: "Convolution"
 848 |   bottom: "conv3_1/dw_new"
 849 |   top: "conv3_1/out/pw_new"
 850 |   param {
 851 |     lr_mult: 1
 852 |     decay_mult: 1
 853 |   }
 854 |   convolution_param {
 855 |     num_output: 32
 856 |     bias_term: false
 857 |     pad: 0
 858 |     kernel_size: 1
 859 |     engine: CAFFE
 860 |     stride: 1
 861 |     weight_filler {
 862 |       type: "msra"
 863 |     }
 864 |   }
 865 | }
 866 | layer {
 867 |   name: "conv3_1/out/pw/bn_new"
 868 |   type: "BatchNorm"
 869 |   bottom: "conv3_1/out/pw_new"
 870 |   top: "conv3_1/out/pw_new"
 871 |   param {
 872 |     lr_mult: 0
 873 |     decay_mult: 0
 874 |   }
 875 |   param {
 876 |     lr_mult: 0
 877 |     decay_mult: 0
 878 |   }
 879 |   param {
 880 |     lr_mult: 0
 881 |     decay_mult: 0
 882 |   }
 883 | }
 884 | layer {
 885 |   name: "conv3_1/out/pw/scale_new"
 886 |   type: "Scale"
 887 |   bottom: "conv3_1/out/pw_new"
 888 |   top: "conv3_1/out/pw_new"
 889 |   param {
 890 |     lr_mult: 1
 891 |     decay_mult: 0
 892 |   }
 893 |   param {
 894 |     lr_mult: 1
 895 |     decay_mult: 0
 896 |   }
 897 |   scale_param {
 898 |     filler {
 899 |       value: 1
 900 |     }
 901 |     bias_term: true
 902 |     bias_filler {
 903 |       value: 0
 904 |     }
 905 |   }
 906 | }
 907 | layer {
 908 |   name: "conv3_2/in/pw_new"
 909 |   type: "Convolution"
 910 |   bottom: "conv3_1/out/pw_new"
 911 |   top: "conv3_2/in/pw_new"
 912 |   param {
 913 |     lr_mult: 1
 914 |     decay_mult: 1
 915 |   }
 916 |   convolution_param {
 917 |     num_output: 192
 918 |     bias_term: false
 919 |     pad: 0
 920 |     kernel_size: 1
 921 |     engine: CAFFE
 922 |     stride: 1
 923 |     weight_filler {
 924 |       type: "msra"
 925 |     }
 926 |   }
 927 | }
 928 | layer {
 929 |   name: "conv3_2/in/pw/bn_new"
 930 |   type: "BatchNorm"
 931 |   bottom: "conv3_2/in/pw_new"
 932 |   top: "conv3_2/in/pw_new"
 933 |   param {
 934 |     lr_mult: 0
 935 |     decay_mult: 0
 936 |   }
 937 |   param {
 938 |     lr_mult: 0
 939 |     decay_mult: 0
 940 |   }
 941 |   param {
 942 |     lr_mult: 0
 943 |     decay_mult: 0
 944 |   }
 945 | }
 946 | layer {
 947 |   name: "conv3_2/in/pw/scale_new"
 948 |   type: "Scale"
 949 |   bottom: "conv3_2/in/pw_new"
 950 |   top: "conv3_2/in/pw_new"
 951 |   param {
 952 |     lr_mult: 1
 953 |     decay_mult: 0
 954 |   }
 955 |   param {
 956 |     lr_mult: 1
 957 |     decay_mult: 0
 958 |   }
 959 |   scale_param {
 960 |     filler {
 961 |       value: 1
 962 |     }
 963 |     bias_term: true
 964 |     bias_filler {
 965 |       value: 0
 966 |     }
 967 |   }
 968 | }
 969 | layer {
 970 |   name: "relu3_2/in/pw_new"
 971 |   type: "ReLU"
 972 |   bottom: "conv3_2/in/pw_new"
 973 |   top: "conv3_2/in/pw_new"
 974 | }
 975 | layer {
 976 |   name: "conv3_2/dw_new"
 977 |   type: "ConvolutionDepthwise"
 978 |   bottom: "conv3_2/in/pw_new"
 979 |   top: "conv3_2/dw_new"
 980 |   param {
 981 |     lr_mult: 1
 982 |     decay_mult: 1
 983 |   }
 984 |   convolution_param {
 985 |     num_output: 192
 986 |     bias_term: false
 987 |     pad: 1
 988 |     kernel_size: 3
 989 |     engine: CAFFE
 990 |     stride: 1
 991 |     weight_filler {
 992 |       type: "msra"
 993 |     }
 994 |   }
 995 | }
 996 | layer {
 997 |   name: "conv3_2/dw/bn_new"
 998 |   type: "BatchNorm"
 999 |   bottom: "conv3_2/dw_new"
1000 |   top: "conv3_2/dw_new"
1001 |   param {
1002 |     lr_mult: 0
1003 |     decay_mult: 0
1004 |   }
1005 |   param {
1006 |     lr_mult: 0
1007 |     decay_mult: 0
1008 |   }
1009 |   param {
1010 |     lr_mult: 0
1011 |     decay_mult: 0
1012 |   }
1013 | }
1014 | layer {
1015 |   name: "conv3_2/dw/scale_new"
1016 |   type: "Scale"
1017 |   bottom: "conv3_2/dw_new"
1018 |   top: "conv3_2/dw_new"
1019 |   param {
1020 |     lr_mult: 1
1021 |     decay_mult: 0
1022 |   }
1023 |   param {
1024 |     lr_mult: 1
1025 |     decay_mult: 0
1026 |   }
1027 |   scale_param {
1028 |     filler {
1029 |       value: 1
1030 |     }
1031 |     bias_term: true
1032 |     bias_filler {
1033 |       value: 0
1034 |     }
1035 |   }
1036 | }
1037 | layer {
1038 |   name: "relu3_2/dw_new"
1039 |   type: "ReLU"
1040 |   bottom: "conv3_2/dw_new"
1041 |   top: "conv3_2/dw_new"
1042 | }
1043 | layer {
1044 |   name: "conv3_2/out/pw_new"
1045 |   type: "Convolution"
1046 |   bottom: "conv3_2/dw_new"
1047 |   top: "conv3_2/out/pw_new"
1048 |   param {
1049 |     lr_mult: 1
1050 |     decay_mult: 1
1051 |   }
1052 |   convolution_param {
1053 |     num_output: 32
1054 |     bias_term: false
1055 |     pad: 0
1056 |     kernel_size: 1
1057 |     engine: CAFFE
1058 |     stride: 1
1059 |     weight_filler {
1060 |       type: "msra"
1061 |     }
1062 |   }
1063 | }
1064 | layer {
1065 |   name: "conv3_2/out/pw/bn_new"
1066 |   type: "BatchNorm"
1067 |   bottom: "conv3_2/out/pw_new"
1068 |   top: "conv3_2/out/pw_new"
1069 |   param {
1070 |     lr_mult: 0
1071 |     decay_mult: 0
1072 |   }
1073 |   param {
1074 |     lr_mult: 0
1075 |     decay_mult: 0
1076 |   }
1077 |   param {
1078 |     lr_mult: 0
1079 |     decay_mult: 0
1080 |   }
1081 | }
1082 | layer {
1083 |   name: "conv3_2/out/pw/scale_new"
1084 |   type: "Scale"
1085 |   bottom: "conv3_2/out/pw_new"
1086 |   top: "conv3_2/out/pw_new"
1087 |   param {
1088 |     lr_mult: 1
1089 |     decay_mult: 0
1090 |   }
1091 |   param {
1092 |     lr_mult: 1
1093 |     decay_mult: 0
1094 |   }
1095 |   scale_param {
1096 |     filler {
1097 |       value: 1
1098 |     }
1099 |     bias_term: true
1100 |     bias_filler {
1101 |       value: 0
1102 |     }
1103 |   }
1104 | }
1105 | layer {
1106 |   name: "fuse_conv3_2"
1107 |   type: "Eltwise"
1108 |   bottom: "conv3_1/out/pw_new"
1109 |   bottom: "conv3_2/out/pw_new"
1110 |   top: "fuse_conv3_2"
1111 |   eltwise_param {
1112 |     operation: SUM
1113 |   }
1114 | }
1115 | layer {
1116 |   name: "conv3_3/in/pw_new"
1117 |   type: "Convolution"
1118 |   bottom: "fuse_conv3_2"
1119 |   top: "conv3_3/in/pw_new"
1120 |   param {
1121 |     lr_mult: 1
1122 |     decay_mult: 1
1123 |   }
1124 |   convolution_param {
1125 |     num_output: 192
1126 |     bias_term: false
1127 |     pad: 0
1128 |     kernel_size: 1
1129 |     engine: CAFFE
1130 |     stride: 1
1131 |     weight_filler {
1132 |       type: "msra"
1133 |     }
1134 |   }
1135 | }
1136 | layer {
1137 |   name: "conv3_3/in/pw/bn_new"
1138 |   type: "BatchNorm"
1139 |   bottom: "conv3_3/in/pw_new"
1140 |   top: "conv3_3/in/pw_new"
1141 |   param {
1142 |     lr_mult: 0
1143 |     decay_mult: 0
1144 |   }
1145 |   param {
1146 |     lr_mult: 0
1147 |     decay_mult: 0
1148 |   }
1149 |   param {
1150 |     lr_mult: 0
1151 |     decay_mult: 0
1152 |   }
1153 | }
1154 | layer {
1155 |   name: "conv3_3/in/pw/scale_new"
1156 |   type: "Scale"
1157 |   bottom: "conv3_3/in/pw_new"
1158 |   top: "conv3_3/in/pw_new"
1159 |   param {
1160 |     lr_mult: 1
1161 |     decay_mult: 0
1162 |   }
1163 |   param {
1164 |     lr_mult: 1
1165 |     decay_mult: 0
1166 |   }
1167 |   scale_param {
1168 |     filler {
1169 |       value: 1
1170 |     }
1171 |     bias_term: true
1172 |     bias_filler {
1173 |       value: 0
1174 |     }
1175 |   }
1176 | }
1177 | layer {
1178 |   name: "relu3_3/in/pw_new"
1179 |   type: "ReLU"
1180 |   bottom: "conv3_3/in/pw_new"
1181 |   top: "conv3_3/in/pw_new"
1182 | }
1183 | layer {
1184 |   name: "conv3_3/dw_new"
1185 |   type: "ConvolutionDepthwise"
1186 |   bottom: "conv3_3/in/pw_new"
1187 |   top: "conv3_3/dw_new"
1188 |   param {
1189 |     lr_mult: 1
1190 |     decay_mult: 1
1191 |   }
1192 |   convolution_param {
1193 |     num_output: 192
1194 |     bias_term: false
1195 |     pad: 1
1196 |     kernel_size: 3
1197 |     engine: CAFFE
1198 |     stride: 1
1199 |     weight_filler {
1200 |       type: "msra"
1201 |     }
1202 |   }
1203 | }
1204 | layer {
1205 |   name: "conv3_3/dw/bn_new"
1206 |   type: "BatchNorm"
1207 |   bottom: "conv3_3/dw_new"
1208 |   top: "conv3_3/dw_new"
1209 |   param {
1210 |     lr_mult: 0
1211 |     decay_mult: 0
1212 |   }
1213 |   param {
1214 |     lr_mult: 0
1215 |     decay_mult: 0
1216 |   }
1217 |   param {
1218 |     lr_mult: 0
1219 |     decay_mult: 0
1220 |   }
1221 | }
1222 | layer {
1223 |   name: "conv3_3/dw/scale_new"
1224 |   type: "Scale"
1225 |   bottom: "conv3_3/dw_new"
1226 |   top: "conv3_3/dw_new"
1227 |   param {
1228 |     lr_mult: 1
1229 |     decay_mult: 0
1230 |   }
1231 |   param {
1232 |     lr_mult: 1
1233 |     decay_mult: 0
1234 |   }
1235 |   scale_param {
1236 |     filler {
1237 |       value: 1
1238 |     }
1239 |     bias_term: true
1240 |     bias_filler {
1241 |       value: 0
1242 |     }
1243 |   }
1244 | }
1245 | layer {
1246 |   name: "relu3_3/dw_new"
1247 |   type: "ReLU"
1248 |   bottom: "conv3_3/dw_new"
1249 |   top: "conv3_3/dw_new"
1250 | }
1251 | layer {
1252 |   name: "conv3_3/out/pw_new"
1253 |   type: "Convolution"
1254 |   bottom: "conv3_3/dw_new"
1255 |   top: "conv3_3/out/pw_new"
1256 |   param {
1257 |     lr_mult: 1
1258 |     decay_mult: 1
1259 |   }
1260 |   convolution_param {
1261 |     num_output: 32
1262 |     bias_term: false
1263 |     pad: 0
1264 |     kernel_size: 1
1265 |     engine: CAFFE
1266 |     stride: 1
1267 |     weight_filler {
1268 |       type: "msra"
1269 |     }
1270 |   }
1271 | }
1272 | layer {
1273 |   name: "conv3_3/out/pw/bn_new"
1274 |   type: "BatchNorm"
1275 |   bottom: "conv3_3/out/pw_new"
1276 |   top: "conv3_3/out/pw_new"
1277 |   param {
1278 |     lr_mult: 0
1279 |     decay_mult: 0
1280 |   }
1281 |   param {
1282 |     lr_mult: 0
1283 |     decay_mult: 0
1284 |   }
1285 |   param {
1286 |     lr_mult: 0
1287 |     decay_mult: 0
1288 |   }
1289 | }
1290 | layer {
1291 |   name: "conv3_3/out/pw/scale_new"
1292 |   type: "Scale"
1293 |   bottom: "conv3_3/out/pw_new"
1294 |   top: "conv3_3/out/pw_new"
1295 |   param {
1296 |     lr_mult: 1
1297 |     decay_mult: 0
1298 |   }
1299 |   param {
1300 |     lr_mult: 1
1301 |     decay_mult: 0
1302 |   }
1303 |   scale_param {
1304 |     filler {
1305 |       value: 1
1306 |     }
1307 |     bias_term: true
1308 |     bias_filler {
1309 |       value: 0
1310 |     }
1311 |   }
1312 | }
1313 | layer {
1314 |   name: "fuse_conv3_3"
1315 |   type: "Eltwise"
1316 |   bottom: "fuse_conv3_2"
1317 |   bottom: "conv3_3/out/pw_new"
1318 |   top: "fuse_conv3_3"
1319 |   eltwise_param {
1320 |     operation: SUM
1321 |   }
1322 | }
1323 | 
1324 | ###############################################
1325 | 
1326 | layer {
1327 |   name: "conv4_1/in/pw_new"
1328 |   type: "Convolution"
1329 |   bottom: "fuse_conv3_3"
1330 |   top: "conv4_1/in/pw_new"
1331 |   param {
1332 |     lr_mult: 1
1333 |     decay_mult: 1
1334 |   }
1335 |   convolution_param {
1336 |     num_output: 192
1337 |     bias_term: false
1338 |     pad: 0
1339 |     kernel_size: 1
1340 |     engine: CAFFE
1341 |     stride: 1
1342 |     weight_filler {
1343 |       type: "msra"
1344 |     }
1345 |   }
1346 | }
1347 | layer {
1348 |   name: "conv4_1/in/pw/bn_new"
1349 |   type: "BatchNorm"
1350 |   bottom: "conv4_1/in/pw_new"
1351 |   top: "conv4_1/in/pw_new"
1352 |   param {
1353 |     lr_mult: 0
1354 |     decay_mult: 0
1355 |   }
1356 |   param {
1357 |     lr_mult: 0
1358 |     decay_mult: 0
1359 |   }
1360 |   param {
1361 |     lr_mult: 0
1362 |     decay_mult: 0
1363 |   }
1364 | }
1365 | layer {
1366 |   name: "conv4_1/in/pw/scale_new"
1367 |   type: "Scale"
1368 |   bottom: "conv4_1/in/pw_new"
1369 |   top: "conv4_1/in/pw_new"
1370 |   param {
1371 |     lr_mult: 1
1372 |     decay_mult: 0
1373 |   }
1374 |   param {
1375 |     lr_mult: 1
1376 |     decay_mult: 0
1377 |   }
1378 |   scale_param {
1379 |     filler {
1380 |       value: 1
1381 |     }
1382 |     bias_term: true
1383 |     bias_filler {
1384 |       value: 0
1385 |     }
1386 |   }
1387 | }
1388 | layer {
1389 |   name: "relu4_1/in/pw_new"
1390 |   type: "ReLU"
1391 |   bottom: "conv4_1/in/pw_new"
1392 |   top: "conv4_1/in/pw_new"
1393 | }
1394 | layer {
1395 |   name: "conv4_1/dw_new"
1396 |   type: "ConvolutionDepthwise"
1397 |   bottom: "conv4_1/in/pw_new"
1398 |   top: "conv4_1/dw_new"
1399 |   param {
1400 |     lr_mult: 1
1401 |     decay_mult: 1
1402 |   }
1403 |   convolution_param {
1404 |     num_output: 192
1405 |     bias_term: false
1406 |     pad: 1
1407 |     kernel_size: 3
1408 |     engine: CAFFE
1409 |     stride: 1
1410 |     weight_filler {
1411 |       type: "msra"
1412 |     }
1413 |   }
1414 | }
1415 | layer {
1416 |   name: "conv4_1/dw/bn_new"
1417 |   type: "BatchNorm"
1418 |   bottom: "conv4_1/dw_new"
1419 |   top: "conv4_1/dw_new"
1420 |   param {
1421 |     lr_mult: 0
1422 |     decay_mult: 0
1423 |   }
1424 |   param {
1425 |     lr_mult: 0
1426 |     decay_mult: 0
1427 |   }
1428 |   param {
1429 |     lr_mult: 0
1430 |     decay_mult: 0
1431 |   }
1432 | }
1433 | layer {
1434 |   name: "conv4_1/dw/scale_new"
1435 |   type: "Scale"
1436 |   bottom: "conv4_1/dw_new"
1437 |   top: "conv4_1/dw_new"
1438 |   param {
1439 |     lr_mult: 1
1440 |     decay_mult: 0
1441 |   }
1442 |   param {
1443 |     lr_mult: 1
1444 |     decay_mult: 0
1445 |   }
1446 |   scale_param {
1447 |     filler {
1448 |       value: 1
1449 |     }
1450 |     bias_term: true
1451 |     bias_filler {
1452 |       value: 0
1453 |     }
1454 |   }
1455 | }
1456 | layer {
1457 |   name: "relu4_1/dw_new"
1458 |   type: "ReLU"
1459 |   bottom: "conv4_1/dw_new"
1460 |   top: "conv4_1/dw_new"
1461 | }
1462 | layer {
1463 |   name: "conv4_1/out/pw_new"
1464 |   type: "Convolution"
1465 |   bottom: "conv4_1/dw_new"
1466 |   top: "conv4_1/out/pw_new"
1467 |   param {
1468 |     lr_mult: 1
1469 |     decay_mult: 1
1470 |   }
1471 |   convolution_param {
1472 |     num_output: 64
1473 |     bias_term: false
1474 |     pad: 0
1475 |     kernel_size: 1
1476 |     engine: CAFFE
1477 |     stride: 1
1478 |     weight_filler {
1479 |       type: "msra"
1480 |     }
1481 |   }
1482 | }
1483 | layer {
1484 |   name: "conv4_1/out/pw/bn_new"
1485 |   type: "BatchNorm"
1486 |   bottom: "conv4_1/out/pw_new"
1487 |   top: "conv4_1/out/pw_new"
1488 |   param {
1489 |     lr_mult: 0
1490 |     decay_mult: 0
1491 |   }
1492 |   param {
1493 |     lr_mult: 0
1494 |     decay_mult: 0
1495 |   }
1496 |   param {
1497 |     lr_mult: 0
1498 |     decay_mult: 0
1499 |   }
1500 | }
1501 | layer {
1502 |   name: "conv4_1/out/pw/scale_new"
1503 |   type: "Scale"
1504 |   bottom: "conv4_1/out/pw_new"
1505 |   top: "conv4_1/out/pw_new"
1506 |   param {
1507 |     lr_mult: 1
1508 |     decay_mult: 0
1509 |   }
1510 |   param {
1511 |     lr_mult: 1
1512 |     decay_mult: 0
1513 |   }
1514 |   scale_param {
1515 |     filler {
1516 |       value: 1
1517 |     }
1518 |     bias_term: true
1519 |     bias_filler {
1520 |       value: 0
1521 |     }
1522 |   }
1523 | }
1524 | layer {
1525 |   name: "conv4_2/in/pw_new"
1526 |   type: "Convolution"
1527 |   bottom: "conv4_1/out/pw_new"
1528 |   top: "conv4_2/in/pw_new"
1529 |   param {
1530 |     lr_mult: 1
1531 |     decay_mult: 1
1532 |   }
1533 |   convolution_param {
1534 |     num_output: 384
1535 |     bias_term: false
1536 |     pad: 0
1537 |     kernel_size: 1
1538 |     engine: CAFFE
1539 |     stride: 1
1540 |     weight_filler {
1541 |       type: "msra"
1542 |     }
1543 |   }
1544 | }
1545 | layer {
1546 |   name: "conv4_2/in/pw/bn_new"
1547 |   type: "BatchNorm"
1548 |   bottom: "conv4_2/in/pw_new"
1549 |   top: "conv4_2/in/pw_new"
1550 |   param {
1551 |     lr_mult: 0
1552 |     decay_mult: 0
1553 |   }
1554 |   param {
1555 |     lr_mult: 0
1556 |     decay_mult: 0
1557 |   }
1558 |   param {
1559 |     lr_mult: 0
1560 |     decay_mult: 0
1561 |   }
1562 | }
1563 | layer {
1564 |   name: "conv4_2/in/pw/scale_new"
1565 |   type: "Scale"
1566 |   bottom: "conv4_2/in/pw_new"
1567 |   top: "conv4_2/in/pw_new"
1568 |   param {
1569 |     lr_mult: 1
1570 |     decay_mult: 0
1571 |   }
1572 |   param {
1573 |     lr_mult: 1
1574 |     decay_mult: 0
1575 |   }
1576 |   scale_param {
1577 |     filler {
1578 |       value: 1
1579 |     }
1580 |     bias_term: true
1581 |     bias_filler {
1582 |       value: 0
1583 |     }
1584 |   }
1585 | }
1586 | layer {
1587 |   name: "relu4_2/in/pw_new"
1588 |   type: "ReLU"
1589 |   bottom: "conv4_2/in/pw_new"
1590 |   top: "conv4_2/in/pw_new"
1591 | }
1592 | layer {
1593 |   name: "conv4_2/dw_new"
1594 |   type: "ConvolutionDepthwise"
1595 |   bottom: "conv4_2/in/pw_new"
1596 |   top: "conv4_2/dw_new"
1597 |   param {
1598 |     lr_mult: 1
1599 |     decay_mult: 1
1600 |   }
1601 |   convolution_param {
1602 |     num_output: 384
1603 |     bias_term: false
1604 |     pad: 1
1605 |     kernel_size: 3
1606 |     engine: CAFFE
1607 |     stride: 1
1608 |     weight_filler {
1609 |       type: "msra"
1610 |     }
1611 |   }
1612 | }
1613 | layer {
1614 |   name: "conv4_2/dw/bn_new"
1615 |   type: "BatchNorm"
1616 |   bottom: "conv4_2/dw_new"
1617 |   top: "conv4_2/dw_new"
1618 |   param {
1619 |     lr_mult: 0
1620 |     decay_mult: 0
1621 |   }
1622 |   param {
1623 |     lr_mult: 0
1624 |     decay_mult: 0
1625 |   }
1626 |   param {
1627 |     lr_mult: 0
1628 |     decay_mult: 0
1629 |   }
1630 | }
1631 | layer {
1632 |   name: "conv4_2/dw/scale_new"
1633 |   type: "Scale"
1634 |   bottom: "conv4_2/dw_new"
1635 |   top: "conv4_2/dw_new"
1636 |   param {
1637 |     lr_mult: 1
1638 |     decay_mult: 0
1639 |   }
1640 |   param {
1641 |     lr_mult: 1
1642 |     decay_mult: 0
1643 |   }
1644 |   scale_param {
1645 |     filler {
1646 |       value: 1
1647 |     }
1648 |     bias_term: true
1649 |     bias_filler {
1650 |       value: 0
1651 |     }
1652 |   }
1653 | }
1654 | layer {
1655 |   name: "relu4_2/dw_new"
1656 |   type: "ReLU"
1657 |   bottom: "conv4_2/dw_new"
1658 |   top: "conv4_2/dw_new"
1659 | }
1660 | layer {
1661 |   name: "conv4_2/out/pw_new"
1662 |   type: "Convolution"
1663 |   bottom: "conv4_2/dw_new"
1664 |   top: "conv4_2/out/pw_new"
1665 |   param {
1666 |     lr_mult: 1
1667 |     decay_mult: 1
1668 |   }
1669 |   convolution_param {
1670 |     num_output: 64
1671 |     bias_term: false
1672 |     pad: 0
1673 |     kernel_size: 1
1674 |     engine: CAFFE
1675 |     stride: 1
1676 |     weight_filler {
1677 |       type: "msra"
1678 |     }
1679 |   }
1680 | }
1681 | layer {
1682 |   name: "conv4_2/out/pw/bn_new"
1683 |   type: "BatchNorm"
1684 |   bottom: "conv4_2/out/pw_new"
1685 |   top: "conv4_2/out/pw_new"
1686 |   param {
1687 |     lr_mult: 0
1688 |     decay_mult: 0
1689 |   }
1690 |   param {
1691 |     lr_mult: 0
1692 |     decay_mult: 0
1693 |   }
1694 |   param {
1695 |     lr_mult: 0
1696 |     decay_mult: 0
1697 |   }
1698 | }
1699 | layer {
1700 |   name: "conv4_2/out/pw/scale_new"
1701 |   type: "Scale"
1702 |   bottom: "conv4_2/out/pw_new"
1703 |   top: "conv4_2/out/pw_new"
1704 |   param {
1705 |     lr_mult: 1
1706 |     decay_mult: 0
1707 |   }
1708 |   param {
1709 |     lr_mult: 1
1710 |     decay_mult: 0
1711 |   }
1712 |   scale_param {
1713 |     filler {
1714 |       value: 1
1715 |     }
1716 |     bias_term: true
1717 |     bias_filler {
1718 |       value: 0
1719 |     }
1720 |   }
1721 | }
1722 | layer {
1723 |   name: "fuse_conv4_2"
1724 |   type: "Eltwise"
1725 |   bottom: "conv4_1/out/pw_new"
1726 |   bottom: "conv4_2/out/pw_new"
1727 |   top: "fuse_conv4_2"
1728 |   eltwise_param {
1729 |     operation: SUM
1730 |   }
1731 | }
1732 | layer {
1733 |   name: "conv4_3/in/pw_new"
1734 |   type: "Convolution"
1735 |   bottom: "fuse_conv4_2"
1736 |   top: "conv4_3/in/pw_new"
1737 |   param {
1738 |     lr_mult: 1
1739 |     decay_mult: 1
1740 |   }
1741 |   convolution_param {
1742 |     num_output: 384
1743 |     bias_term: false
1744 |     pad: 0
1745 |     kernel_size: 1
1746 |     engine: CAFFE
1747 |     stride: 1
1748 |     weight_filler {
1749 |       type: "msra"
1750 |     }
1751 |   }
1752 | }
1753 | layer {
1754 |   name: "conv4_3/in/pw/bn_new"
1755 |   type: "BatchNorm"
1756 |   bottom: "conv4_3/in/pw_new"
1757 |   top: "conv4_3/in/pw_new"
1758 |   param {
1759 |     lr_mult: 0
1760 |     decay_mult: 0
1761 |   }
1762 |   param {
1763 |     lr_mult: 0
1764 |     decay_mult: 0
1765 |   }
1766 |   param {
1767 |     lr_mult: 0
1768 |     decay_mult: 0
1769 |   }
1770 | }
1771 | layer {
1772 |   name: "conv4_3/in/pw/scale_new"
1773 |   type: "Scale"
1774 |   bottom: "conv4_3/in/pw_new"
1775 |   top: "conv4_3/in/pw_new"
1776 |   param {
1777 |     lr_mult: 1
1778 |     decay_mult: 0
1779 |   }
1780 |   param {
1781 |     lr_mult: 1
1782 |     decay_mult: 0
1783 |   }
1784 |   scale_param {
1785 |     filler {
1786 |       value: 1
1787 |     }
1788 |     bias_term: true
1789 |     bias_filler {
1790 |       value: 0
1791 |     }
1792 |   }
1793 | }
1794 | layer {
1795 |   name: "relu4_3/in/pw_new"
1796 |   type: "ReLU"
1797 |   bottom: "conv4_3/in/pw_new"
1798 |   top: "conv4_3/in/pw_new"
1799 | }
1800 | layer {
1801 |   name: "conv4_3/dw_new"
1802 |   type: "ConvolutionDepthwise"
1803 |   bottom: "conv4_3/in/pw_new"
1804 |   top: "conv4_3/dw_new"
1805 |   param {
1806 |     lr_mult: 1
1807 |     decay_mult: 1
1808 |   }
1809 |   convolution_param {
1810 |     num_output: 384
1811 |     bias_term: false
1812 |     pad: 1
1813 |     kernel_size: 3
1814 |     engine: CAFFE
1815 |     stride: 1
1816 |     weight_filler {
1817 |       type: "msra"
1818 |     }
1819 |   }
1820 | }
1821 | layer {
1822 |   name: "conv4_3/dw/bn_new"
1823 |   type: "BatchNorm"
1824 |   bottom: "conv4_3/dw_new"
1825 |   top: "conv4_3/dw_new"
1826 |   param {
1827 |     lr_mult: 0
1828 |     decay_mult: 0
1829 |   }
1830 |   param {
1831 |     lr_mult: 0
1832 |     decay_mult: 0
1833 |   }
1834 |   param {
1835 |     lr_mult: 0
1836 |     decay_mult: 0
1837 |   }
1838 | }
1839 | layer {
1840 |   name: "conv4_3/dw/scale_new"
1841 |   type: "Scale"
1842 |   bottom: "conv4_3/dw_new"
1843 |   top: "conv4_3/dw_new"
1844 |   param {
1845 |     lr_mult: 1
1846 |     decay_mult: 0
1847 |   }
1848 |   param {
1849 |     lr_mult: 1
1850 |     decay_mult: 0
1851 |   }
1852 |   scale_param {
1853 |     filler {
1854 |       value: 1
1855 |     }
1856 |     bias_term: true
1857 |     bias_filler {
1858 |       value: 0
1859 |     }
1860 |   }
1861 | }
1862 | layer {
1863 |   name: "relu4_3/dw_new"
1864 |   type: "ReLU"
1865 |   bottom: "conv4_3/dw_new"
1866 |   top: "conv4_3/dw_new"
1867 | }
1868 | layer {
1869 |   name: "conv4_3/out/pw_new"
1870 |   type: "Convolution"
1871 |   bottom: "conv4_3/dw_new"
1872 |   top: "conv4_3/out/pw_new"
1873 |   param {
1874 |     lr_mult: 1
1875 |     decay_mult: 1
1876 |   }
1877 |   convolution_param {
1878 |     num_output: 64
1879 |     bias_term: false
1880 |     pad: 0
1881 |     kernel_size: 1
1882 |     engine: CAFFE
1883 |     stride: 1
1884 |     weight_filler {
1885 |       type: "msra"
1886 |     }
1887 |   }
1888 | }
1889 | layer {
1890 |   name: "conv4_3/out/pw/bn_new"
1891 |   type: "BatchNorm"
1892 |   bottom: "conv4_3/out/pw_new"
1893 |   top: "conv4_3/out/pw_new"
1894 |   param {
1895 |     lr_mult: 0
1896 |     decay_mult: 0
1897 |   }
1898 |   param {
1899 |     lr_mult: 0
1900 |     decay_mult: 0
1901 |   }
1902 |   param {
1903 |     lr_mult: 0
1904 |     decay_mult: 0
1905 |   }
1906 | }
1907 | layer {
1908 |   name: "conv4_3/out/pw/scale_new"
1909 |   type: "Scale"
1910 |   bottom: "conv4_3/out/pw_new"
1911 |   top: "conv4_3/out/pw_new"
1912 |   param {
1913 |     lr_mult: 1
1914 |     decay_mult: 0
1915 |   }
1916 |   param {
1917 |     lr_mult: 1
1918 |     decay_mult: 0
1919 |   }
1920 |   scale_param {
1921 |     filler {
1922 |       value: 1
1923 |     }
1924 |     bias_term: true
1925 |     bias_filler {
1926 |       value: 0
1927 |     }
1928 |   }
1929 | }
1930 | layer {
1931 |   name: "fuse_conv4_3"
1932 |   type: "Eltwise"
1933 |   bottom: "fuse_conv4_2"
1934 |   bottom: "conv4_3/out/pw_new"
1935 |   top: "fuse_conv4_3"
1936 |   eltwise_param {
1937 |     operation: SUM
1938 |   }
1939 | }
1940 | layer {
1941 |   name: "conv4_4/in/pw_new"
1942 |   type: "Convolution"
1943 |   bottom: "fuse_conv4_3"
1944 |   top: "conv4_4/in/pw_new"
1945 |   param {
1946 |     lr_mult: 1
1947 |     decay_mult: 1
1948 |   }
1949 |   convolution_param {
1950 |     num_output: 384
1951 |     bias_term: false
1952 |     pad: 0
1953 |     kernel_size: 1
1954 |     engine: CAFFE
1955 |     stride: 1
1956 |     weight_filler {
1957 |       type: "msra"
1958 |     }
1959 |   }
1960 | }
1961 | layer {
1962 |   name: "conv4_4/in/pw/bn_new"
1963 |   type: "BatchNorm"
1964 |   bottom: "conv4_4/in/pw_new"
1965 |   top: "conv4_4/in/pw_new"
1966 |   param {
1967 |     lr_mult: 0
1968 |     decay_mult: 0
1969 |   }
1970 |   param {
1971 |     lr_mult: 0
1972 |     decay_mult: 0
1973 |   }
1974 |   param {
1975 |     lr_mult: 0
1976 |     decay_mult: 0
1977 |   }
1978 | }
1979 | layer {
1980 |   name: "conv4_4/in/pw/scale_new"
1981 |   type: "Scale"
1982 |   bottom: "conv4_4/in/pw_new"
1983 |   top: "conv4_4/in/pw_new"
1984 |   param {
1985 |     lr_mult: 1
1986 |     decay_mult: 0
1987 |   }
1988 |   param {
1989 |     lr_mult: 1
1990 |     decay_mult: 0
1991 |   }
1992 |   scale_param {
1993 |     filler {
1994 |       value: 1
1995 |     }
1996 |     bias_term: true
1997 |     bias_filler {
1998 |       value: 0
1999 |     }
2000 |   }
2001 | }
2002 | layer {
2003 |   name: "relu4_4/in/pw_new"
2004 |   type: "ReLU"
2005 |   bottom: "conv4_4/in/pw_new"
2006 |   top: "conv4_4/in/pw_new"
2007 | }
2008 | layer {
2009 |   name: "conv4_4/dw_new"
2010 |   type: "ConvolutionDepthwise"
2011 |   bottom: "conv4_4/in/pw_new"
2012 |   top: "conv4_4/dw_new"
2013 |   param {
2014 |     lr_mult: 1
2015 |     decay_mult: 1
2016 |   }
2017 |   convolution_param {
2018 |     num_output: 384
2019 |     bias_term: false
2020 |     pad: 1
2021 |     kernel_size: 3
2022 |     engine: CAFFE
2023 |     stride: 1
2024 |     weight_filler {
2025 |       type: "msra"
2026 |     }
2027 |   }
2028 | }
2029 | layer {
2030 |   name: "conv4_4/dw/bn_new"
2031 |   type: "BatchNorm"
2032 |   bottom: "conv4_4/dw_new"
2033 |   top: "conv4_4/dw_new"
2034 |   param {
2035 |     lr_mult: 0
2036 |     decay_mult: 0
2037 |   }
2038 |   param {
2039 |     lr_mult: 0
2040 |     decay_mult: 0
2041 |   }
2042 |   param {
2043 |     lr_mult: 0
2044 |     decay_mult: 0
2045 |   }
2046 | }
2047 | layer {
2048 |   name: "conv4_4/dw/scale_new"
2049 |   type: "Scale"
2050 |   bottom: "conv4_4/dw_new"
2051 |   top: "conv4_4/dw_new"
2052 |   param {
2053 |     lr_mult: 1
2054 |     decay_mult: 0
2055 |   }
2056 |   param {
2057 |     lr_mult: 1
2058 |     decay_mult: 0
2059 |   }
2060 |   scale_param {
2061 |     filler {
2062 |       value: 1
2063 |     }
2064 |     bias_term: true
2065 |     bias_filler {
2066 |       value: 0
2067 |     }
2068 |   }
2069 | }
2070 | layer {
2071 |   name: "relu4_4/dw_new"
2072 |   type: "ReLU"
2073 |   bottom: "conv4_4/dw_new"
2074 |   top: "conv4_4/dw_new"
2075 | }
2076 | layer {
2077 |   name: "conv4_4/out/pw_new"
2078 |   type: "Convolution"
2079 |   bottom: "conv4_4/dw_new"
2080 |   top: "conv4_4/out/pw_new"
2081 |   param {
2082 |     lr_mult: 1
2083 |     decay_mult: 1
2084 |   }
2085 |   convolution_param {
2086 |     num_output: 64
2087 |     bias_term: false
2088 |     pad: 0
2089 |     kernel_size: 1
2090 |     engine: CAFFE
2091 |     stride: 1
2092 |     weight_filler {
2093 |       type: "msra"
2094 |     }
2095 |   }
2096 | }
2097 | layer {
2098 |   name: "conv4_4/out/pw/bn_new"
2099 |   type: "BatchNorm"
2100 |   bottom: "conv4_4/out/pw_new"
2101 |   top: "conv4_4/out/pw_new"
2102 |   param {
2103 |     lr_mult: 0
2104 |     decay_mult: 0
2105 |   }
2106 |   param {
2107 |     lr_mult: 0
2108 |     decay_mult: 0
2109 |   }
2110 |   param {
2111 |     lr_mult: 0
2112 |     decay_mult: 0
2113 |   }
2114 | }
2115 | layer {
2116 |   name: "conv4_4/out/pw/scale_new"
2117 |   type: "Scale"
2118 |   bottom: "conv4_4/out/pw_new"
2119 |   top: "conv4_4/out/pw_new"
2120 |   param {
2121 |     lr_mult: 1
2122 |     decay_mult: 0
2123 |   }
2124 |   param {
2125 |     lr_mult: 1
2126 |     decay_mult: 0
2127 |   }
2128 |   scale_param {
2129 |     filler {
2130 |       value: 1
2131 |     }
2132 |     bias_term: true
2133 |     bias_filler {
2134 |       value: 0
2135 |     }
2136 |   }
2137 | }
2138 | layer {
2139 |   name: "fuse_conv4_4"
2140 |   type: "Eltwise"
2141 |   bottom: "fuse_conv4_3"
2142 |   bottom: "conv4_4/out/pw_new"
2143 |   top: "fuse_conv4_4"
2144 |   eltwise_param {
2145 |     operation: SUM
2146 |   }
2147 | }
2148 | ###############################################
2149 | 
2150 | layer {
2151 |   name: "conv5_1/in/pw_new"
2152 |   type: "Convolution"
2153 |   bottom: "fuse_conv4_4"
2154 |   top: "conv5_1/in/pw_new"
2155 |   param {
2156 |     lr_mult: 1
2157 |     decay_mult: 1
2158 |   }
2159 |   convolution_param {
2160 |     num_output: 384
2161 |     bias_term: false
2162 |     pad: 0
2163 |     kernel_size: 1
2164 |     engine: CAFFE
2165 |     stride: 1
2166 |     weight_filler {
2167 |       type: "msra"
2168 |     }
2169 |   }
2170 | }
2171 | layer {
2172 |   name: "conv5_1/in/pw/bn_new"
2173 |   type: "BatchNorm"
2174 |   bottom: "conv5_1/in/pw_new"
2175 |   top: "conv5_1/in/pw_new"
2176 |   param {
2177 |     lr_mult: 0
2178 |     decay_mult: 0
2179 |   }
2180 |   param {
2181 |     lr_mult: 0
2182 |     decay_mult: 0
2183 |   }
2184 |   param {
2185 |     lr_mult: 0
2186 |     decay_mult: 0
2187 |   }
2188 | }
2189 | layer {
2190 |   name: "conv5_1/in/pw/scale_new"
2191 |   type: "Scale"
2192 |   bottom: "conv5_1/in/pw_new"
2193 |   top: "conv5_1/in/pw_new"
2194 |   param {
2195 |     lr_mult: 1
2196 |     decay_mult: 0
2197 |   }
2198 |   param {
2199 |     lr_mult: 1
2200 |     decay_mult: 0
2201 |   }
2202 |   scale_param {
2203 |     filler {
2204 |       value: 1
2205 |     }
2206 |     bias_term: true
2207 |     bias_filler {
2208 |       value: 0
2209 |     }
2210 |   }
2211 | }
2212 | layer {
2213 |   name: "relu5_1/in/pw_new"
2214 |   type: "ReLU"
2215 |   bottom: "conv5_1/in/pw_new"
2216 |   top: "conv5_1/in/pw_new"
2217 | }
2218 | layer {
2219 |   name: "conv5_1/dw_new"
2220 |   type: "ConvolutionDepthwise"
2221 |   bottom: "conv5_1/in/pw_new"
2222 |   top: "conv5_1/dw_new"
2223 |   param {
2224 |     lr_mult: 1
2225 |     decay_mult: 1
2226 |   }
2227 |   convolution_param {
2228 |     num_output: 384
2229 |     bias_term: false
2230 |     pad: 1
2231 |     kernel_size: 3
2232 |     engine: CAFFE
2233 |     stride: 2
2234 |     weight_filler {
2235 |       type: "msra"
2236 |     }
2237 |   }
2238 | }
2239 | layer {
2240 |   name: "conv5_1/dw/bn_new"
2241 |   type: "BatchNorm"
2242 |   bottom: "conv5_1/dw_new"
2243 |   top: "conv5_1/dw_new"
2244 |   param {
2245 |     lr_mult: 0
2246 |     decay_mult: 0
2247 |   }
2248 |   param {
2249 |     lr_mult: 0
2250 |     decay_mult: 0
2251 |   }
2252 |   param {
2253 |     lr_mult: 0
2254 |     decay_mult: 0
2255 |   }
2256 | }
2257 | layer {
2258 |   name: "conv5_1/dw/scale_new"
2259 |   type: "Scale"
2260 |   bottom: "conv5_1/dw_new"
2261 |   top: "conv5_1/dw_new"
2262 |   param {
2263 |     lr_mult: 1
2264 |     decay_mult: 0
2265 |   }
2266 |   param {
2267 |     lr_mult: 1
2268 |     decay_mult: 0
2269 |   }
2270 |   scale_param {
2271 |     filler {
2272 |       value: 1
2273 |     }
2274 |     bias_term: true
2275 |     bias_filler {
2276 |       value: 0
2277 |     }
2278 |   }
2279 | }
2280 | layer {
2281 |   name: "relu5_1/dw_new"
2282 |   type: "ReLU"
2283 |   bottom: "conv5_1/dw_new"
2284 |   top: "conv5_1/dw_new"
2285 | }
2286 | layer {
2287 |   name: "conv5_1/out/pw_new"
2288 |   type: "Convolution"
2289 |   bottom: "conv5_1/dw_new"
2290 |   top: "conv5_1/out/pw_new"
2291 |   param {
2292 |     lr_mult: 1
2293 |     decay_mult: 1
2294 |   }
2295 |   convolution_param {
2296 |     num_output: 96
2297 |     bias_term: false
2298 |     pad: 0
2299 |     kernel_size: 1
2300 |     engine: CAFFE
2301 |     stride: 1
2302 |     weight_filler {
2303 |       type: "msra"
2304 |     }
2305 |   }
2306 | }
2307 | layer {
2308 |   name: "conv5_1/out/pw/bn_new"
2309 |   type: "BatchNorm"
2310 |   bottom: "conv5_1/out/pw_new"
2311 |   top: "conv5_1/out/pw_new"
2312 |   param {
2313 |     lr_mult: 0
2314 |     decay_mult: 0
2315 |   }
2316 |   param {
2317 |     lr_mult: 0
2318 |     decay_mult: 0
2319 |   }
2320 |   param {
2321 |     lr_mult: 0
2322 |     decay_mult: 0
2323 |   }
2324 | }
2325 | layer {
2326 |   name: "conv5_1/out/pw/scale_new"
2327 |   type: "Scale"
2328 |   bottom: "conv5_1/out/pw_new"
2329 |   top: "conv5_1/out/pw_new"
2330 |   param {
2331 |     lr_mult: 1
2332 |     decay_mult: 0
2333 |   }
2334 |   param {
2335 |     lr_mult: 1
2336 |     decay_mult: 0
2337 |   }
2338 |   scale_param {
2339 |     filler {
2340 |       value: 1
2341 |     }
2342 |     bias_term: true
2343 |     bias_filler {
2344 |       value: 0
2345 |     }
2346 |   }
2347 | }
2348 | layer {
2349 |   name: "conv5_2/in/pw_new"
2350 |   type: "Convolution"
2351 |   bottom: "conv5_1/out/pw_new"
2352 |   top: "conv5_2/in/pw_new"
2353 |   param {
2354 |     lr_mult: 1
2355 |     decay_mult: 1
2356 |   }
2357 |   convolution_param {
2358 |     num_output: 576
2359 |     bias_term: false
2360 |     pad: 0
2361 |     kernel_size: 1
2362 |     engine: CAFFE
2363 |     stride: 1
2364 |     weight_filler {
2365 |       type: "msra"
2366 |     }
2367 |   }
2368 | }
2369 | layer {
2370 |   name: "conv5_2/in/pw/bn_new"
2371 |   type: "BatchNorm"
2372 |   bottom: "conv5_2/in/pw_new"
2373 |   top: "conv5_2/in/pw_new"
2374 |   param {
2375 |     lr_mult: 0
2376 |     decay_mult: 0
2377 |   }
2378 |   param {
2379 |     lr_mult: 0
2380 |     decay_mult: 0
2381 |   }
2382 |   param {
2383 |     lr_mult: 0
2384 |     decay_mult: 0
2385 |   }
2386 | }
2387 | layer {
2388 |   name: "conv5_2/in/pw/scale_new"
2389 |   type: "Scale"
2390 |   bottom: "conv5_2/in/pw_new"
2391 |   top: "conv5_2/in/pw_new"
2392 |   param {
2393 |     lr_mult: 1
2394 |     decay_mult: 0
2395 |   }
2396 |   param {
2397 |     lr_mult: 1
2398 |     decay_mult: 0
2399 |   }
2400 |   scale_param {
2401 |     filler {
2402 |       value: 1
2403 |     }
2404 |     bias_term: true
2405 |     bias_filler {
2406 |       value: 0
2407 |     }
2408 |   }
2409 | }
2410 | layer {
2411 |   name: "relu5_2/in/pw_new"
2412 |   type: "ReLU"
2413 |   bottom: "conv5_2/in/pw_new"
2414 |   top: "conv5_2/in/pw_new"
2415 | }
2416 | layer {
2417 |   name: "conv5_2/dw_new"
2418 |   type: "ConvolutionDepthwise"
2419 |   bottom: "conv5_2/in/pw_new"
2420 |   top: "conv5_2/dw_new"
2421 |   param {
2422 |     lr_mult: 1
2423 |     decay_mult: 1
2424 |   }
2425 |   convolution_param {
2426 |     num_output: 576
2427 |     bias_term: false
2428 |     pad: 1
2429 |     kernel_size: 3
2430 |     engine: CAFFE
2431 |     stride: 1
2432 |     weight_filler {
2433 |       type: "msra"
2434 |     }
2435 |   }
2436 | }
2437 | layer {
2438 |   name: "conv5_2/dw/bn_new"
2439 |   type: "BatchNorm"
2440 |   bottom: "conv5_2/dw_new"
2441 |   top: "conv5_2/dw_new"
2442 |   param {
2443 |     lr_mult: 0
2444 |     decay_mult: 0
2445 |   }
2446 |   param {
2447 |     lr_mult: 0
2448 |     decay_mult: 0
2449 |   }
2450 |   param {
2451 |     lr_mult: 0
2452 |     decay_mult: 0
2453 |   }
2454 | }
2455 | layer {
2456 |   name: "conv5_2/dw/scale_new"
2457 |   type: "Scale"
2458 |   bottom: "conv5_2/dw_new"
2459 |   top: "conv5_2/dw_new"
2460 |   param {
2461 |     lr_mult: 1
2462 |     decay_mult: 0
2463 |   }
2464 |   param {
2465 |     lr_mult: 1
2466 |     decay_mult: 0
2467 |   }
2468 |   scale_param {
2469 |     filler {
2470 |       value: 1
2471 |     }
2472 |     bias_term: true
2473 |     bias_filler {
2474 |       value: 0
2475 |     }
2476 |   }
2477 | }
2478 | layer {
2479 |   name: "relu5_2/dw_new"
2480 |   type: "ReLU"
2481 |   bottom: "conv5_2/dw_new"
2482 |   top: "conv5_2/dw_new"
2483 | }
2484 | layer {
2485 |   name: "conv5_2/out/pw_new"
2486 |   type: "Convolution"
2487 |   bottom: "conv5_2/dw_new"
2488 |   top: "conv5_2/out/pw_new"
2489 |   param {
2490 |     lr_mult: 1
2491 |     decay_mult: 1
2492 |   }
2493 |   convolution_param {
2494 |     num_output: 96
2495 |     bias_term: false
2496 |     pad: 0
2497 |     kernel_size: 1
2498 |     engine: CAFFE
2499 |     stride: 1
2500 |     weight_filler {
2501 |       type: "msra"
2502 |     }
2503 |   }
2504 | }
2505 | layer {
2506 |   name: "conv5_2/out/pw/bn_new"
2507 |   type: "BatchNorm"
2508 |   bottom: "conv5_2/out/pw_new"
2509 |   top: "conv5_2/out/pw_new"
2510 |   param {
2511 |     lr_mult: 0
2512 |     decay_mult: 0
2513 |   }
2514 |   param {
2515 |     lr_mult: 0
2516 |     decay_mult: 0
2517 |   }
2518 |   param {
2519 |     lr_mult: 0
2520 |     decay_mult: 0
2521 |   }
2522 | }
2523 | layer {
2524 |   name: "conv5_2/out/pw/scale_new"
2525 |   type: "Scale"
2526 |   bottom: "conv5_2/out/pw_new"
2527 |   top: "conv5_2/out/pw_new"
2528 |   param {
2529 |     lr_mult: 1
2530 |     decay_mult: 0
2531 |   }
2532 |   param {
2533 |     lr_mult: 1
2534 |     decay_mult: 0
2535 |   }
2536 |   scale_param {
2537 |     filler {
2538 |       value: 1
2539 |     }
2540 |     bias_term: true
2541 |     bias_filler {
2542 |       value: 0
2543 |     }
2544 |   }
2545 | }
2546 | layer {
2547 |   name: "fuse_conv5_2"
2548 |   type: "Eltwise"
2549 |   bottom: "conv5_1/out/pw_new"
2550 |   bottom: "conv5_2/out/pw_new"
2551 |   top: "fuse_conv5_2"
2552 |   eltwise_param {
2553 |     operation: SUM
2554 |   }
2555 | }
2556 | layer {
2557 |   name: "conv5_3/in/pw_new"
2558 |   type: "Convolution"
2559 |   bottom: "fuse_conv5_2"
2560 |   top: "conv5_3/in/pw_new"
2561 |   param {
2562 |     lr_mult: 1
2563 |     decay_mult: 1
2564 |   }
2565 |   convolution_param {
2566 |     num_output: 576
2567 |     bias_term: false
2568 |     pad: 0
2569 |     kernel_size: 1
2570 |     engine: CAFFE
2571 |     stride: 1
2572 |     weight_filler {
2573 |       type: "msra"
2574 |     }
2575 |   }
2576 | }
2577 | layer {
2578 |   name: "conv5_3/in/pw/bn_new"
2579 |   type: "BatchNorm"
2580 |   bottom: "conv5_3/in/pw_new"
2581 |   top: "conv5_3/in/pw_new"
2582 |   param {
2583 |     lr_mult: 0
2584 |     decay_mult: 0
2585 |   }
2586 |   param {
2587 |     lr_mult: 0
2588 |     decay_mult: 0
2589 |   }
2590 |   param {
2591 |     lr_mult: 0
2592 |     decay_mult: 0
2593 |   }
2594 | }
2595 | layer {
2596 |   name: "conv5_3/in/pw/scale_new"
2597 |   type: "Scale"
2598 |   bottom: "conv5_3/in/pw_new"
2599 |   top: "conv5_3/in/pw_new"
2600 |   param {
2601 |     lr_mult: 1
2602 |     decay_mult: 0
2603 |   }
2604 |   param {
2605 |     lr_mult: 1
2606 |     decay_mult: 0
2607 |   }
2608 |   scale_param {
2609 |     filler {
2610 |       value: 1
2611 |     }
2612 |     bias_term: true
2613 |     bias_filler {
2614 |       value: 0
2615 |     }
2616 |   }
2617 | }
2618 | layer {
2619 |   name: "relu5_3/in/pw_new"
2620 |   type: "ReLU"
2621 |   bottom: "conv5_3/in/pw_new"
2622 |   top: "conv5_3/in/pw_new"
2623 | }
2624 | layer {
2625 |   name: "conv5_3/dw_new"
2626 |   type: "ConvolutionDepthwise"
2627 |   bottom: "conv5_3/in/pw_new"
2628 |   top: "conv5_3/dw_new"
2629 |   param {
2630 |     lr_mult: 1
2631 |     decay_mult: 1
2632 |   }
2633 |   convolution_param {
2634 |     num_output: 576
2635 |     bias_term: false
2636 |     pad: 1
2637 |     kernel_size: 3
2638 |     engine: CAFFE
2639 |     stride: 1
2640 |     weight_filler {
2641 |       type: "msra"
2642 |     }
2643 |   }
2644 | }
2645 | layer {
2646 |   name: "conv5_3/dw/bn_new"
2647 |   type: "BatchNorm"
2648 |   bottom: "conv5_3/dw_new"
2649 |   top: "conv5_3/dw_new"
2650 |   param {
2651 |     lr_mult: 0
2652 |     decay_mult: 0
2653 |   }
2654 |   param {
2655 |     lr_mult: 0
2656 |     decay_mult: 0
2657 |   }
2658 |   param {
2659 |     lr_mult: 0
2660 |     decay_mult: 0
2661 |   }
2662 | }
2663 | layer {
2664 |   name: "conv5_3/dw/scale_new"
2665 |   type: "Scale"
2666 |   bottom: "conv5_3/dw_new"
2667 |   top: "conv5_3/dw_new"
2668 |   param {
2669 |     lr_mult: 1
2670 |     decay_mult: 0
2671 |   }
2672 |   param {
2673 |     lr_mult: 1
2674 |     decay_mult: 0
2675 |   }
2676 |   scale_param {
2677 |     filler {
2678 |       value: 1
2679 |     }
2680 |     bias_term: true
2681 |     bias_filler {
2682 |       value: 0
2683 |     }
2684 |   }
2685 | }
2686 | layer {
2687 |   name: "relu5_3/dw_new"
2688 |   type: "ReLU"
2689 |   bottom: "conv5_3/dw_new"
2690 |   top: "conv5_3/dw_new"
2691 | }
2692 | layer {
2693 |   name: "conv5_3/out/pw_new"
2694 |   type: "Convolution"
2695 |   bottom: "conv5_3/dw_new"
2696 |   top: "conv5_3/out/pw_new"
2697 |   param {
2698 |     lr_mult: 1
2699 |     decay_mult: 1
2700 |   }
2701 |   convolution_param {
2702 |     num_output: 96
2703 |     bias_term: false
2704 |     pad: 0
2705 |     kernel_size: 1
2706 |     engine: CAFFE
2707 |     stride: 1
2708 |     weight_filler {
2709 |       type: "msra"
2710 |     }
2711 |   }
2712 | }
2713 | layer {
2714 |   name: "conv5_3/out/pw/bn_new"
2715 |   type: "BatchNorm"
2716 |   bottom: "conv5_3/out/pw_new"
2717 |   top: "conv5_3/out/pw_new"
2718 |   param {
2719 |     lr_mult: 0
2720 |     decay_mult: 0
2721 |   }
2722 |   param {
2723 |     lr_mult: 0
2724 |     decay_mult: 0
2725 |   }
2726 |   param {
2727 |     lr_mult: 0
2728 |     decay_mult: 0
2729 |   }
2730 | }
2731 | layer {
2732 |   name: "conv5_3/out/pw/scale_new"
2733 |   type: "Scale"
2734 |   bottom: "conv5_3/out/pw_new"
2735 |   top: "conv5_3/out/pw_new"
2736 |   param {
2737 |     lr_mult: 1
2738 |     decay_mult: 0
2739 |   }
2740 |   param {
2741 |     lr_mult: 1
2742 |     decay_mult: 0
2743 |   }
2744 |   scale_param {
2745 |     filler {
2746 |       value: 1
2747 |     }
2748 |     bias_term: true
2749 |     bias_filler {
2750 |       value: 0
2751 |     }
2752 |   }
2753 | }
2754 | layer {
2755 |   name: "fuse_conv5_3"
2756 |   type: "Eltwise"
2757 |   bottom: "fuse_conv5_2"
2758 |   bottom: "conv5_3/out/pw_new"
2759 |   top: "fuse_conv5_3"
2760 |   eltwise_param {
2761 |     operation: SUM
2762 |   }
2763 | }
2764 | ###############################################
2765 | 
2766 | layer {
2767 |   name: "conv6_1/in/pw_new"
2768 |   type: "Convolution"
2769 |   bottom: "fuse_conv5_3"
2770 |   top: "conv6_1/in/pw_new"
2771 |   param {
2772 |     lr_mult: 1
2773 |     decay_mult: 1
2774 |   }
2775 |   convolution_param {
2776 |     num_output: 576
2777 |     bias_term: false
2778 |     pad: 0
2779 |     kernel_size: 1
2780 |     engine: CAFFE
2781 |     stride: 1
2782 |     weight_filler {
2783 |       type: "msra"
2784 |     }
2785 |   }
2786 | }
2787 | layer {
2788 |   name: "conv6_1/in/pw/bn_new"
2789 |   type: "BatchNorm"
2790 |   bottom: "conv6_1/in/pw_new"
2791 |   top: "conv6_1/in/pw_new"
2792 |   param {
2793 |     lr_mult: 0
2794 |     decay_mult: 0
2795 |   }
2796 |   param {
2797 |     lr_mult: 0
2798 |     decay_mult: 0
2799 |   }
2800 |   param {
2801 |     lr_mult: 0
2802 |     decay_mult: 0
2803 |   }
2804 | }
2805 | layer {
2806 |   name: "conv6_1/in/pw/scale_new"
2807 |   type: "Scale"
2808 |   bottom: "conv6_1/in/pw_new"
2809 |   top: "conv6_1/in/pw_new"
2810 |   param {
2811 |     lr_mult: 1
2812 |     decay_mult: 0
2813 |   }
2814 |   param {
2815 |     lr_mult: 1
2816 |     decay_mult: 0
2817 |   }
2818 |   scale_param {
2819 |     filler {
2820 |       value: 1
2821 |     }
2822 |     bias_term: true
2823 |     bias_filler {
2824 |       value: 0
2825 |     }
2826 |   }
2827 | }
2828 | layer {
2829 |   name: "relu6_1/in/pw_new"
2830 |   type: "ReLU"
2831 |   bottom: "conv6_1/in/pw_new"
2832 |   top: "conv6_1/in/pw_new"
2833 | }
2834 | layer {
2835 |   name: "conv6_1/dw_new"
2836 |   type: "ConvolutionDepthwise"
2837 |   bottom: "conv6_1/in/pw_new"
2838 |   top: "conv6_1/dw_new"
2839 |   param {
2840 |     lr_mult: 1
2841 |     decay_mult: 1
2842 |   }
2843 |   convolution_param {
2844 |     num_output: 576
2845 |     bias_term: false
2846 |     pad: 1
2847 |     kernel_size: 3
2848 |     engine: CAFFE
2849 |     stride: 2
2850 |     weight_filler {
2851 |       type: "msra"
2852 |     }
2853 |   }
2854 | }
2855 | layer {
2856 |   name: "conv6_1/dw/bn_new"
2857 |   type: "BatchNorm"
2858 |   bottom: "conv6_1/dw_new"
2859 |   top: "conv6_1/dw_new"
2860 |   param {
2861 |     lr_mult: 0
2862 |     decay_mult: 0
2863 |   }
2864 |   param {
2865 |     lr_mult: 0
2866 |     decay_mult: 0
2867 |   }
2868 |   param {
2869 |     lr_mult: 0
2870 |     decay_mult: 0
2871 |   }
2872 | }
2873 | layer {
2874 |   name: "conv6_1/dw/scale_new"
2875 |   type: "Scale"
2876 |   bottom: "conv6_1/dw_new"
2877 |   top: "conv6_1/dw_new"
2878 |   param {
2879 |     lr_mult: 1
2880 |     decay_mult: 0
2881 |   }
2882 |   param {
2883 |     lr_mult: 1
2884 |     decay_mult: 0
2885 |   }
2886 |   scale_param {
2887 |     filler {
2888 |       value: 1
2889 |     }
2890 |     bias_term: true
2891 |     bias_filler {
2892 |       value: 0
2893 |     }
2894 |   }
2895 | }
2896 | layer {
2897 |   name: "relu6_1/dw_new"
2898 |   type: "ReLU"
2899 |   bottom: "conv6_1/dw_new"
2900 |   top: "conv6_1/dw_new"
2901 | }
2902 | layer {
2903 |   name: "conv6_1/out/pw_new"
2904 |   type: "Convolution"
2905 |   bottom: "conv6_1/dw_new"
2906 |   top: "conv6_1/out/pw_new"
2907 |   param {
2908 |     lr_mult: 1
2909 |     decay_mult: 1
2910 |   }
2911 |   convolution_param {
2912 |     num_output: 160
2913 |     bias_term: false
2914 |     pad: 0
2915 |     kernel_size: 1
2916 |     engine: CAFFE
2917 |     stride: 1
2918 |     weight_filler {
2919 |       type: "msra"
2920 |     }
2921 |   }
2922 | }
2923 | layer {
2924 |   name: "conv6_1/out/pw/bn_new"
2925 |   type: "BatchNorm"
2926 |   bottom: "conv6_1/out/pw_new"
2927 |   top: "conv6_1/out/pw_new"
2928 |   param {
2929 |     lr_mult: 0
2930 |     decay_mult: 0
2931 |   }
2932 |   param {
2933 |     lr_mult: 0
2934 |     decay_mult: 0
2935 |   }
2936 |   param {
2937 |     lr_mult: 0
2938 |     decay_mult: 0
2939 |   }
2940 | }
2941 | layer {
2942 |   name: "conv6_1/out/pw/scale_new"
2943 |   type: "Scale"
2944 |   bottom: "conv6_1/out/pw_new"
2945 |   top: "conv6_1/out/pw_new"
2946 |   param {
2947 |     lr_mult: 1
2948 |     decay_mult: 0
2949 |   }
2950 |   param {
2951 |     lr_mult: 1
2952 |     decay_mult: 0
2953 |   }
2954 |   scale_param {
2955 |     filler {
2956 |       value: 1
2957 |     }
2958 |     bias_term: true
2959 |     bias_filler {
2960 |       value: 0
2961 |     }
2962 |   }
2963 | }
2964 | layer {
2965 |   name: "conv6_2/in/pw_new"
2966 |   type: "Convolution"
2967 |   bottom: "conv6_1/out/pw_new"
2968 |   top: "conv6_2/in/pw_new"
2969 |   param {
2970 |     lr_mult: 1
2971 |     decay_mult: 1
2972 |   }
2973 |   convolution_param {
2974 |     num_output: 960
2975 |     bias_term: false
2976 |     pad: 0
2977 |     kernel_size: 1
2978 |     engine: CAFFE
2979 |     stride: 1
2980 |     weight_filler {
2981 |       type: "msra"
2982 |     }
2983 |   }
2984 | }
2985 | layer {
2986 |   name: "conv6_2/in/pw/bn_new"
2987 |   type: "BatchNorm"
2988 |   bottom: "conv6_2/in/pw_new"
2989 |   top: "conv6_2/in/pw_new"
2990 |   param {
2991 |     lr_mult: 0
2992 |     decay_mult: 0
2993 |   }
2994 |   param {
2995 |     lr_mult: 0
2996 |     decay_mult: 0
2997 |   }
2998 |   param {
2999 |     lr_mult: 0
3000 |     decay_mult: 0
3001 |   }
3002 | }
3003 | layer {
3004 |   name: "conv6_2/in/pw/scale_new"
3005 |   type: "Scale"
3006 |   bottom: "conv6_2/in/pw_new"
3007 |   top: "conv6_2/in/pw_new"
3008 |   param {
3009 |     lr_mult: 1
3010 |     decay_mult: 0
3011 |   }
3012 |   param {
3013 |     lr_mult: 1
3014 |     decay_mult: 0
3015 |   }
3016 |   scale_param {
3017 |     filler {
3018 |       value: 1
3019 |     }
3020 |     bias_term: true
3021 |     bias_filler {
3022 |       value: 0
3023 |     }
3024 |   }
3025 | }
3026 | layer {
3027 |   name: "relu6_2/in/pw_new"
3028 |   type: "ReLU"
3029 |   bottom: "conv6_2/in/pw_new"
3030 |   top: "conv6_2/in/pw_new"
3031 | }
3032 | layer {
3033 |   name: "conv6_2/dw_new"
3034 |   type: "ConvolutionDepthwise"
3035 |   bottom: "conv6_2/in/pw_new"
3036 |   top: "conv6_2/dw_new"
3037 |   param {
3038 |     lr_mult: 1
3039 |     decay_mult: 1
3040 |   }
3041 |   convolution_param {
3042 |     num_output: 960
3043 |     bias_term: false
3044 |     pad: 1
3045 |     kernel_size: 3
3046 |     engine: CAFFE
3047 |     stride: 1
3048 |     weight_filler {
3049 |       type: "msra"
3050 |     }
3051 |   }
3052 | }
3053 | layer {
3054 |   name: "conv6_2/dw/bn_new"
3055 |   type: "BatchNorm"
3056 |   bottom: "conv6_2/dw_new"
3057 |   top: "conv6_2/dw_new"
3058 |   param {
3059 |     lr_mult: 0
3060 |     decay_mult: 0
3061 |   }
3062 |   param {
3063 |     lr_mult: 0
3064 |     decay_mult: 0
3065 |   }
3066 |   param {
3067 |     lr_mult: 0
3068 |     decay_mult: 0
3069 |   }
3070 | }
3071 | layer {
3072 |   name: "conv6_2/dw/scale_new"
3073 |   type: "Scale"
3074 |   bottom: "conv6_2/dw_new"
3075 |   top: "conv6_2/dw_new"
3076 |   param {
3077 |     lr_mult: 1
3078 |     decay_mult: 0
3079 |   }
3080 |   param {
3081 |     lr_mult: 1
3082 |     decay_mult: 0
3083 |   }
3084 |   scale_param {
3085 |     filler {
3086 |       value: 1
3087 |     }
3088 |     bias_term: true
3089 |     bias_filler {
3090 |       value: 0
3091 |     }
3092 |   }
3093 | }
3094 | layer {
3095 |   name: "relu6_2/dw_new"
3096 |   type: "ReLU"
3097 |   bottom: "conv6_2/dw_new"
3098 |   top: "conv6_2/dw_new"
3099 | }
3100 | layer {
3101 |   name: "conv6_2/out/pw_new"
3102 |   type: "Convolution"
3103 |   bottom: "conv6_2/dw_new"
3104 |   top: "conv6_2/out/pw_new"
3105 |   param {
3106 |     lr_mult: 1
3107 |     decay_mult: 1
3108 |   }
3109 |   convolution_param {
3110 |     num_output: 160
3111 |     bias_term: false
3112 |     pad: 0
3113 |     kernel_size: 1
3114 |     engine: CAFFE
3115 |     stride: 1
3116 |     weight_filler {
3117 |       type: "msra"
3118 |     }
3119 |   }
3120 | }
3121 | layer {
3122 |   name: "conv6_2/out/pw/bn_new"
3123 |   type: "BatchNorm"
3124 |   bottom: "conv6_2/out/pw_new"
3125 |   top: "conv6_2/out/pw_new"
3126 |   param {
3127 |     lr_mult: 0
3128 |     decay_mult: 0
3129 |   }
3130 |   param {
3131 |     lr_mult: 0
3132 |     decay_mult: 0
3133 |   }
3134 |   param {
3135 |     lr_mult: 0
3136 |     decay_mult: 0
3137 |   }
3138 | }
3139 | layer {
3140 |   name: "conv6_2/out/pw/scale_new"
3141 |   type: "Scale"
3142 |   bottom: "conv6_2/out/pw_new"
3143 |   top: "conv6_2/out/pw_new"
3144 |   param {
3145 |     lr_mult: 1
3146 |     decay_mult: 0
3147 |   }
3148 |   param {
3149 |     lr_mult: 1
3150 |     decay_mult: 0
3151 |   }
3152 |   scale_param {
3153 |     filler {
3154 |       value: 1
3155 |     }
3156 |     bias_term: true
3157 |     bias_filler {
3158 |       value: 0
3159 |     }
3160 |   }
3161 | }
3162 | layer {
3163 |   name: "fuse_conv6_2"
3164 |   type: "Eltwise"
3165 |   bottom: "conv6_1/out/pw_new"
3166 |   bottom: "conv6_2/out/pw_new"
3167 |   top: "fuse_conv6_2"
3168 |   eltwise_param {
3169 |     operation: SUM
3170 |   }
3171 | }
3172 | layer {
3173 |   name: "conv6_3/in/pw_new"
3174 |   type: "Convolution"
3175 |   bottom: "fuse_conv6_2"
3176 |   top: "conv6_3/in/pw_new"
3177 |   param {
3178 |     lr_mult: 1
3179 |     decay_mult: 1
3180 |   }
3181 |   convolution_param {
3182 |     num_output: 960
3183 |     bias_term: false
3184 |     pad: 0
3185 |     kernel_size: 1
3186 |     engine: CAFFE
3187 |     stride: 1
3188 |     weight_filler {
3189 |       type: "msra"
3190 |     }
3191 |   }
3192 | }
3193 | layer {
3194 |   name: "conv6_3/in/pw/bn_new"
3195 |   type: "BatchNorm"
3196 |   bottom: "conv6_3/in/pw_new"
3197 |   top: "conv6_3/in/pw_new"
3198 |   param {
3199 |     lr_mult: 0
3200 |     decay_mult: 0
3201 |   }
3202 |   param {
3203 |     lr_mult: 0
3204 |     decay_mult: 0
3205 |   }
3206 |   param {
3207 |     lr_mult: 0
3208 |     decay_mult: 0
3209 |   }
3210 | }
3211 | layer {
3212 |   name: "conv6_3/in/pw/scale_new"
3213 |   type: "Scale"
3214 |   bottom: "conv6_3/in/pw_new"
3215 |   top: "conv6_3/in/pw_new"
3216 |   param {
3217 |     lr_mult: 1
3218 |     decay_mult: 0
3219 |   }
3220 |   param {
3221 |     lr_mult: 1
3222 |     decay_mult: 0
3223 |   }
3224 |   scale_param {
3225 |     filler {
3226 |       value: 1
3227 |     }
3228 |     bias_term: true
3229 |     bias_filler {
3230 |       value: 0
3231 |     }
3232 |   }
3233 | }
3234 | layer {
3235 |   name: "relu6_3/in/pw_new"
3236 |   type: "ReLU"
3237 |   bottom: "conv6_3/in/pw_new"
3238 |   top: "conv6_3/in/pw_new"
3239 | }
3240 | layer {
3241 |   name: "conv6_3/dw_new"
3242 |   type: "ConvolutionDepthwise"
3243 |   bottom: "conv6_3/in/pw_new"
3244 |   top: "conv6_3/dw_new"
3245 |   param {
3246 |     lr_mult: 1
3247 |     decay_mult: 1
3248 |   }
3249 |   convolution_param {
3250 |     num_output: 960
3251 |     bias_term: false
3252 |     pad: 1
3253 |     kernel_size: 3
3254 |     engine: CAFFE
3255 |     stride: 1
3256 |     weight_filler {
3257 |       type: "msra"
3258 |     }
3259 |   }
3260 | }
3261 | layer {
3262 |   name: "conv6_3/dw/bn_new"
3263 |   type: "BatchNorm"
3264 |   bottom: "conv6_3/dw_new"
3265 |   top: "conv6_3/dw_new"
3266 |   param {
3267 |     lr_mult: 0
3268 |     decay_mult: 0
3269 |   }
3270 |   param {
3271 |     lr_mult: 0
3272 |     decay_mult: 0
3273 |   }
3274 |   param {
3275 |     lr_mult: 0
3276 |     decay_mult: 0
3277 |   }
3278 | }
3279 | layer {
3280 |   name: "conv6_3/dw/scale_new"
3281 |   type: "Scale"
3282 |   bottom: "conv6_3/dw_new"
3283 |   top: "conv6_3/dw_new"
3284 |   param {
3285 |     lr_mult: 1
3286 |     decay_mult: 0
3287 |   }
3288 |   param {
3289 |     lr_mult: 1
3290 |     decay_mult: 0
3291 |   }
3292 |   scale_param {
3293 |     filler {
3294 |       value: 1
3295 |     }
3296 |     bias_term: true
3297 |     bias_filler {
3298 |       value: 0
3299 |     }
3300 |   }
3301 | }
3302 | layer {
3303 |   name: "relu6_3/dw_new"
3304 |   type: "ReLU"
3305 |   bottom: "conv6_3/dw_new"
3306 |   top: "conv6_3/dw_new"
3307 | }
3308 | layer {
3309 |   name: "conv6_3/out/pw_new"
3310 |   type: "Convolution"
3311 |   bottom: "conv6_3/dw_new"
3312 |   top: "conv6_3/out/pw_new"
3313 |   param {
3314 |     lr_mult: 1
3315 |     decay_mult: 1
3316 |   }
3317 |   convolution_param {
3318 |     num_output: 160
3319 |     bias_term: false
3320 |     pad: 0
3321 |     kernel_size: 1
3322 |     engine: CAFFE
3323 |     stride: 1
3324 |     weight_filler {
3325 |       type: "msra"
3326 |     }
3327 |   }
3328 | }
3329 | layer {
3330 |   name: "conv6_3/out/pw/bn_new"
3331 |   type: "BatchNorm"
3332 |   bottom: "conv6_3/out/pw_new"
3333 |   top: "conv6_3/out/pw_new"
3334 |   param {
3335 |     lr_mult: 0
3336 |     decay_mult: 0
3337 |   }
3338 |   param {
3339 |     lr_mult: 0
3340 |     decay_mult: 0
3341 |   }
3342 |   param {
3343 |     lr_mult: 0
3344 |     decay_mult: 0
3345 |   }
3346 | }
3347 | layer {
3348 |   name: "conv6_3/out/pw/scale_new"
3349 |   type: "Scale"
3350 |   bottom: "conv6_3/out/pw_new"
3351 |   top: "conv6_3/out/pw_new"
3352 |   param {
3353 |     lr_mult: 1
3354 |     decay_mult: 0
3355 |   }
3356 |   param {
3357 |     lr_mult: 1
3358 |     decay_mult: 0
3359 |   }
3360 |   scale_param {
3361 |     filler {
3362 |       value: 1
3363 |     }
3364 |     bias_term: true
3365 |     bias_filler {
3366 |       value: 0
3367 |     }
3368 |   }
3369 | }
3370 | layer {
3371 |   name: "fuse_conv6_3"
3372 |   type: "Eltwise"
3373 |   bottom: "fuse_conv6_2"
3374 |   bottom: "conv6_3/out/pw_new"
3375 |   top: "fuse_conv6_3"
3376 |   eltwise_param {
3377 |     operation: SUM
3378 |   }
3379 | }
3380 | ###############################################
3381 | 
3382 | layer {
3383 |   name: "conv7_1/in/pw_new"
3384 |   type: "Convolution"
3385 |   bottom: "fuse_conv6_3"
3386 |   top: "conv7_1/in/pw_new"
3387 |   param {
3388 |     lr_mult: 1
3389 |     decay_mult: 1
3390 |   }
3391 |   convolution_param {
3392 |     num_output: 960
3393 |     bias_term: false
3394 |     pad: 0
3395 |     kernel_size: 1
3396 |     engine: CAFFE
3397 |     stride: 1
3398 |     weight_filler {
3399 |       type: "msra"
3400 |     }
3401 |   }
3402 | }
3403 | layer {
3404 |   name: "conv7_1/in/pw/bn_new"
3405 |   type: "BatchNorm"
3406 |   bottom: "conv7_1/in/pw_new"
3407 |   top: "conv7_1/in/pw_new"
3408 |   param {
3409 |     lr_mult: 0
3410 |     decay_mult: 0
3411 |   }
3412 |   param {
3413 |     lr_mult: 0
3414 |     decay_mult: 0
3415 |   }
3416 |   param {
3417 |     lr_mult: 0
3418 |     decay_mult: 0
3419 |   }
3420 | }
3421 | layer {
3422 |   name: "conv7_1/in/pw/scale_new"
3423 |   type: "Scale"
3424 |   bottom: "conv7_1/in/pw_new"
3425 |   top: "conv7_1/in/pw_new"
3426 |   param {
3427 |     lr_mult: 1
3428 |     decay_mult: 0
3429 |   }
3430 |   param {
3431 |     lr_mult: 1
3432 |     decay_mult: 0
3433 |   }
3434 |   scale_param {
3435 |     filler {
3436 |       value: 1
3437 |     }
3438 |     bias_term: true
3439 |     bias_filler {
3440 |       value: 0
3441 |     }
3442 |   }
3443 | }
3444 | layer {
3445 |   name: "relu7_1/in/pw_new"
3446 |   type: "ReLU"
3447 |   bottom: "conv7_1/in/pw_new"
3448 |   top: "conv7_1/in/pw_new"
3449 | }
3450 | layer {
3451 |   name: "conv7_1/dw_new"
3452 |   type: "ConvolutionDepthwise"
3453 |   bottom: "conv7_1/in/pw_new"
3454 |   top: "conv7_1/dw_new"
3455 |   param {
3456 |     lr_mult: 1
3457 |     decay_mult: 1
3458 |   }
3459 |   convolution_param {
3460 |     num_output: 960
3461 |     bias_term: false
3462 |     pad: 1
3463 |     kernel_size: 3
3464 |     engine: CAFFE
3465 |     stride: 1
3466 |     weight_filler {
3467 |       type: "msra"
3468 |     }
3469 |   }
3470 | }
3471 | layer {
3472 |   name: "conv7_1/dw/bn_new"
3473 |   type: "BatchNorm"
3474 |   bottom: "conv7_1/dw_new"
3475 |   top: "conv7_1/dw_new"
3476 |   param {
3477 |     lr_mult: 0
3478 |     decay_mult: 0
3479 |   }
3480 |   param {
3481 |     lr_mult: 0
3482 |     decay_mult: 0
3483 |   }
3484 |   param {
3485 |     lr_mult: 0
3486 |     decay_mult: 0
3487 |   }
3488 | }
3489 | layer {
3490 |   name: "conv7_1/dw/scale_new"
3491 |   type: "Scale"
3492 |   bottom: "conv7_1/dw_new"
3493 |   top: "conv7_1/dw_new"
3494 |   param {
3495 |     lr_mult: 1
3496 |     decay_mult: 0
3497 |   }
3498 |   param {
3499 |     lr_mult: 1
3500 |     decay_mult: 0
3501 |   }
3502 |   scale_param {
3503 |     filler {
3504 |       value: 1
3505 |     }
3506 |     bias_term: true
3507 |     bias_filler {
3508 |       value: 0
3509 |     }
3510 |   }
3511 | }
3512 | layer {
3513 |   name: "relu7_1/dw_new"
3514 |   type: "ReLU"
3515 |   bottom: "conv7_1/dw_new"
3516 |   top: "conv7_1/dw_new"
3517 | }
3518 | layer {
3519 |   name: "conv7_1/out/pw_new"
3520 |   type: "Convolution"
3521 |   bottom: "conv7_1/dw_new"
3522 |   top: "conv7_1/out/pw_new"
3523 |   param {
3524 |     lr_mult: 1
3525 |     decay_mult: 1
3526 |   }
3527 |   convolution_param {
3528 |     num_output: 320
3529 |     bias_term: false
3530 |     pad: 0
3531 |     kernel_size: 1
3532 |     engine: CAFFE
3533 |     stride: 1
3534 |     weight_filler {
3535 |       type: "msra"
3536 |     }
3537 |   }
3538 | }
3539 | layer {
3540 |   name: "conv7_1/out/pw/bn_new"
3541 |   type: "BatchNorm"
3542 |   bottom: "conv7_1/out/pw_new"
3543 |   top: "conv7_1/out/pw_new"
3544 |   param {
3545 |     lr_mult: 0
3546 |     decay_mult: 0
3547 |   }
3548 |   param {
3549 |     lr_mult: 0
3550 |     decay_mult: 0
3551 |   }
3552 |   param {
3553 |     lr_mult: 0
3554 |     decay_mult: 0
3555 |   }
3556 | }
3557 | layer {
3558 |   name: "conv7_1/out/pw/scale_new"
3559 |   type: "Scale"
3560 |   bottom: "conv7_1/out/pw_new"
3561 |   top: "conv7_1/out/pw_new"
3562 |   param {
3563 |     lr_mult: 1
3564 |     decay_mult: 0
3565 |   }
3566 |   param {
3567 |     lr_mult: 1
3568 |     decay_mult: 0
3569 |   }
3570 |   scale_param {
3571 |     filler {
3572 |       value: 1
3573 |     }
3574 |     bias_term: true
3575 |     bias_filler {
3576 |       value: 0
3577 |     }
3578 |   }
3579 | }
3580 | layer {
3581 |   name: "conv8_new"
3582 |   type: "Convolution"
3583 |   bottom: "conv7_1/out/pw_new"
3584 |   top: "conv8_new"
3585 |   param {
3586 |     lr_mult: 1
3587 |     decay_mult: 1
3588 |   }
3589 |   convolution_param {
3590 |     num_output: 1280
3591 |     bias_term: false
3592 |     pad: 0
3593 |     kernel_size: 1
3594 |     stride: 1
3595 |     weight_filler {
3596 |       type: "msra"
3597 |     }
3598 |   }
3599 | }
3600 | layer {
3601 |   name: "conv8/bn_new"
3602 |   type: "BatchNorm"
3603 |   bottom: "conv8_new"
3604 |   top: "conv8_new"
3605 |   param {
3606 |     lr_mult: 0
3607 |     decay_mult: 0
3608 |   }
3609 |   param {
3610 |     lr_mult: 0
3611 |     decay_mult: 0
3612 |   }
3613 |   param {
3614 |     lr_mult: 0
3615 |     decay_mult: 0
3616 |   }
3617 | }
3618 | layer {
3619 |   name: "conv8/scale_new"
3620 |   type: "Scale"
3621 |   bottom: "conv8_new"
3622 |   top: "conv8_new"
3623 |   param {
3624 |     lr_mult: 1
3625 |     decay_mult: 0
3626 |   }
3627 |   param {
3628 |     lr_mult: 1
3629 |     decay_mult: 0
3630 |   }
3631 |   scale_param {
3632 |     filler {
3633 |       value: 1
3634 |     }
3635 |     bias_term: true
3636 |     bias_filler {
3637 |       value: 0
3638 |     }
3639 |   }
3640 | }
3641 | layer {
3642 |   name: "relu8_new"
3643 |   type: "ReLU"
3644 |   bottom: "conv8_new"
3645 |   top: "conv8_new"
3646 | }
3647 | layer {
3648 |   name: "pool8_new"
3649 |   type: "Pooling"
3650 |   bottom: "conv8_new"
3651 |   top: "pool8_new"
3652 |   pooling_param {
3653 |     pool: AVE
3654 |     global_pooling: true
3655 |   }
3656 | }
3657 | layer {
3658 |   name: "fc8_new"
3659 |   type: "Convolution"
3660 |   bottom: "pool8_new"
3661 |   top: "fc8_new"
3662 |   param {
3663 |     lr_mult: 1
3664 |     decay_mult: 1
3665 |   }
3666 |   param {
3667 |     lr_mult: 2
3668 |     decay_mult: 0
3669 |   }
3670 |   convolution_param {
3671 |     num_output: 1000
3672 |     kernel_size: 1
3673 |     weight_filler {
3674 |       type: "msra"
3675 |     }
3676 |     bias_filler {
3677 |       type: "constant"
3678 |       value: 0
3679 |     }
3680 |   }
3681 | }
3682 | 
3683 | layer {
3684 |   name: "accuracy/top1"
3685 |   type: "Accuracy"
3686 |   bottom: "fc8_new"
3687 |   bottom: "label"
3688 |   top: "accuracy@1"
3689 |   include: { phase: TEST }
3690 |   accuracy_param {
3691 |     top_k: 1
3692 |   }
3693 | }
3694 | layer {
3695 |   name: "accuracy/top5"
3696 |   type: "Accuracy"
3697 |   bottom: "fc8_new"
3698 |   bottom: "label"
3699 |   top: "accuracy@5"
3700 |   include: { phase: TEST }
3701 |   accuracy_param {
3702 |     top_k: 5
3703 |   }
3704 | }
3705 | 
3706 | 


--------------------------------------------------------------------------------
/.gitignore/benchmark_mobilenetv2.sh:
--------------------------------------------------------------------------------
1 | #!/usr/bin/env sh
2 | 
3 | ./build/tools/caffe test \
4 | 	--model=****/MobileNetV2_deploy.prototxt \
5 | 	--weights=****/mobilenet_v2_new.caffemodel \
6 | 	--gpu=0 --iterations=2000
7 | 


--------------------------------------------------------------------------------
/.gitignore/conv_dw_layer.cpp:
--------------------------------------------------------------------------------
  1 | #include <algorithm>
  2 | #include <vector>
  3 | #include "caffe/filler.hpp"
  4 | #include "caffe/layers/conv_dw_layer.hpp"
  5 | 
  6 | namespace caffe {
  7 | 
  8 | template <typename Dtype>
  9 | void ConvolutionDepthwiseLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
 10 |       const vector<Blob<Dtype>*>& top) {
 11 |   ConvolutionParameter conv_param = this->layer_param_.convolution_param();
 12 |   if (conv_param.has_kernel_h() && conv_param.has_kernel_w()) {
 13 |     kernel_h_ = conv_param.kernel_h();
 14 |     kernel_w_ = conv_param.kernel_w();
 15 |   } else {
 16 |     if (conv_param.kernel_size_size() == 1)
 17 |     {
 18 |       kernel_h_ = conv_param.kernel_size(0);
 19 |       kernel_w_ = conv_param.kernel_size(0);
 20 |     }
 21 |     else
 22 |     {
 23 |       kernel_h_ = conv_param.kernel_size(0);
 24 |       kernel_w_ = conv_param.kernel_size(1);
 25 |     }
 26 |   }
 27 |   if (conv_param.has_stride_h() && conv_param.has_stride_w()) {
 28 |     stride_h_ = conv_param.stride_h();
 29 |     stride_w_ = conv_param.stride_w();
 30 |   } else {
 31 |     if (conv_param.stride_size() == 1)
 32 |     {
 33 |       stride_h_ = conv_param.stride(0);
 34 |       stride_w_ = conv_param.stride(0);
 35 |     }
 36 |     else
 37 |     {
 38 |       stride_h_ = conv_param.stride(0);
 39 |       stride_w_ = conv_param.stride(1);
 40 |     }
 41 |   }
 42 |   if (conv_param.has_pad_h() && conv_param.has_pad_w()) {
 43 |     pad_h_ = conv_param.pad_h();
 44 |     pad_w_ = conv_param.pad_w();
 45 |   } else {
 46 |     if (conv_param.pad_size() == 1)
 47 |     {
 48 |       pad_h_ = conv_param.pad(0);
 49 |       pad_w_ = conv_param.pad(0);
 50 |     }
 51 |     else
 52 |     {
 53 |       pad_h_ = conv_param.pad(0);
 54 |       pad_w_ = conv_param.pad(1);
 55 |     }
 56 |   }
 57 |   if (conv_param.dilation_size() > 0)
 58 |   {
 59 |     if (conv_param.dilation_size() == 1)
 60 |     {
 61 |       dilation_h_ = conv_param.dilation(0);
 62 |       dilation_w_ = conv_param.dilation(0);
 63 |     }
 64 |     else
 65 |     {
 66 |       dilation_h_ = conv_param.dilation(0);
 67 |       dilation_w_ = conv_param.dilation(1);
 68 |     }
 69 |   }
 70 |   else
 71 |   {
 72 |     dilation_h_ = 1;
 73 |     dilation_w_ = 1;
 74 |   }
 75 |   vector<int> weight_shape(4);
 76 |   weight_shape[0] = bottom[0]->channels();
 77 |   weight_shape[1] = 1;
 78 |   weight_shape[2] = kernel_h_;
 79 |   weight_shape[3] = kernel_w_;
 80 |   vector<int> bias_shape;
 81 |   if (conv_param.bias_term())
 82 |   {
 83 |     bias_shape.push_back(bottom[0]->channels());
 84 |   }
 85 |   if (this->blobs_.size() == 0) {
 86 |     if (conv_param.bias_term()) {
 87 |       this->blobs_.resize(2);
 88 |     } else {
 89 |       this->blobs_.resize(1);
 90 |     }
 91 |     this->blobs_[0].reset(new Blob<Dtype>(weight_shape));
 92 |     shared_ptr<Filler<Dtype> > weight_filler(GetFiller<Dtype>(conv_param.weight_filler()));
 93 |     weight_filler->Fill(this->blobs_[0].get());
 94 |     if (conv_param.bias_term()) {
 95 |       this->blobs_[1].reset(new Blob<Dtype>(bias_shape));
 96 |       shared_ptr<Filler<Dtype> > bias_filler(GetFiller<Dtype>(conv_param.bias_filler()));
 97 |       bias_filler->Fill(this->blobs_[1].get());
 98 |     }
 99 |   }
100 |   this->param_propagate_down_.resize(this->blobs_.size(), true);
101 | }
102 | 
103 | template <typename Dtype>
104 | void ConvolutionDepthwiseLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
105 |       const vector<Blob<Dtype>*>& top) {
106 |   vector<int> top_shape;
107 |   top_shape.push_back(bottom[0]->num());
108 |   top_shape.push_back(bottom[0]->channels());
109 |   top_shape.push_back((bottom[0]->height() + 2 * pad_h_ - (dilation_h_ * (kernel_h_ - 1) + 1)) / stride_h_ + 1);
110 |   top_shape.push_back((bottom[0]->width() + 2 * pad_w_ - (dilation_w_ * (kernel_w_ - 1) + 1)) / stride_w_ + 1);
111 |   top[0]->Reshape(top_shape);
112 |   vector<int> weight_buffer_shape;
113 |   weight_buffer_shape.push_back(bottom[0]->channels());
114 |   weight_buffer_shape.push_back(kernel_h_);
115 |   weight_buffer_shape.push_back(kernel_w_);
116 |   weight_buffer_shape.push_back(bottom[0]->num());
117 |   weight_buffer_shape.push_back(top[0]->height());
118 |   weight_buffer_shape.push_back(top[0]->width());
119 |   weight_buffer_.Reshape(weight_buffer_shape);
120 |   vector<int> weight_multiplier_shape;
121 |   weight_multiplier_shape.push_back(bottom[0]->num());
122 |   weight_multiplier_shape.push_back(top[0]->height());
123 |   weight_multiplier_shape.push_back(top[0]->width());
124 |   weight_multiplier_.Reshape(weight_multiplier_shape);
125 |   caffe_gpu_set(weight_multiplier_.count(), Dtype(1), weight_multiplier_.mutable_gpu_data());
126 |   if (this->layer_param_.convolution_param().bias_term())
127 |   {
128 |     vector<int> bias_buffer_shape;
129 |     bias_buffer_shape.push_back(bottom[0]->channels());
130 |     bias_buffer_shape.push_back(bottom[0]->num());
131 |     bias_buffer_shape.push_back(top[0]->height());
132 |     bias_buffer_shape.push_back(top[0]->width());
133 |     bias_buffer_.Reshape(bias_buffer_shape);
134 |     vector<int> bias_multiplier_shape;
135 |     bias_multiplier_shape.push_back(bottom[0]->num());
136 |     bias_multiplier_shape.push_back(top[0]->height());
137 |     bias_multiplier_shape.push_back(top[0]->width());
138 |     bias_multiplier_.Reshape(bias_multiplier_shape);
139 |     caffe_gpu_set(bias_multiplier_.count(), Dtype(1), bias_multiplier_.mutable_gpu_data());
140 |   }
141 | }
142 | 
143 | template <typename Dtype>
144 | void ConvolutionDepthwiseLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
145 |       const vector<Blob<Dtype>*>& top)
146 | {
147 |   const int num = top[0]->num();
148 |   const int channels = top[0]->channels();
149 |   const int top_height = top[0]->height();
150 |   const int top_width = top[0]->width();
151 |   const int bottom_height = bottom[0]->height();
152 |   const int bottom_width = bottom[0]->width();
153 |   const Dtype* bottom_data = bottom[0]->cpu_data();
154 |   const Dtype* weight_data_base = this->blobs_[0]->cpu_data();
155 |   Dtype* top_data = top[0]->mutable_cpu_data();
156 |   for (int n = 0; n < num; ++n)
157 |   {
158 |     for (int c = 0; c < channels; ++c)
159 |     {
160 |       for (int h = 0; h < top_height; ++h)
161 |       {
162 |         for (int w = 0; w < top_width; ++w)
163 |         {
164 |           const Dtype* weight_data = weight_data_base + c * kernel_h_ * kernel_w_;
165 |           Dtype value = 0;
166 |           for (int kh = 0; kh < kernel_h_; ++kh)
167 |           {
168 |             for (int kw = 0; kw < kernel_w_; ++kw)
169 |             {
170 |               int h_in = -pad_h_ + h * stride_h_ + kh * dilation_h_;
171 |               int w_in = -pad_w_ + w * stride_w_ + kw * dilation_w_;
172 |               if ((h_in >= 0) && (h_in < bottom_height) && (w_in >= 0) && (w_in < bottom_width))
173 |               {
174 |                 int offset = ((n * channels + c) * bottom_height + h_in) * bottom_width + w_in;
175 |                 value += (*weight_data) * bottom_data[offset];
176 |               }
177 |               ++weight_data;
178 |             }
179 |           }
180 |           *top_data++ = value;
181 |         }
182 |       }
183 |     }
184 |   }
185 |   if (this->layer_param_.convolution_param().bias_term())
186 |   {
187 |     top_data = top[0]->mutable_cpu_data();
188 |     for (int n = 0; n < num; ++n)
189 |     {
190 |       const Dtype* bias_data = this->blobs_[1]->cpu_data();
191 |       for (int c = 0; c < channels; ++c)
192 |       {
193 |         for (int h = 0; h < top_height; ++h)
194 |         {
195 |           for (int w = 0; w < top_width; ++w)
196 |           {
197 |             *top_data += *bias_data;
198 |             ++top_data;
199 |           }
200 |         }
201 |         ++bias_data;
202 |       }
203 |     }
204 |   }
205 | }
206 | 
207 | template <typename Dtype>
208 | void ConvolutionDepthwiseLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
209 |       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom)
210 | {
211 |   const int num = top[0]->num();
212 |   const int channels = top[0]->channels();
213 |   const int top_height = top[0]->height();
214 |   const int top_width = top[0]->width();
215 |   const int bottom_height = bottom[0]->height();
216 |   const int bottom_width = bottom[0]->width();
217 |   caffe_set(bottom[0]->count(), Dtype(0), bottom[0]->mutable_cpu_diff());
218 |   if (this->layer_param_.convolution_param().bias_term() && this->param_propagate_down_[1])
219 |   {
220 |     const Dtype* top_diff = top[0]->cpu_diff();
221 |     for (int n = 0; n < num; ++n)
222 |     {
223 |       Dtype* bias_diff = this->blobs_[1]->mutable_cpu_diff();
224 |       for (int c = 0; c < channels; ++c)
225 |       {
226 |         for (int h = 0; h < top_height; ++h)
227 |         {
228 |           for (int w = 0; w < top_width; ++w)
229 |           {
230 |             *bias_diff += *top_diff;
231 |             ++top_diff;
232 |           }
233 |         }
234 |         ++bias_diff;
235 |       }
236 |     }
237 |   }
238 |   if (this->param_propagate_down_[0])
239 |   {
240 |     const Dtype* top_diff = top[0]->cpu_diff();
241 |     const Dtype* bottom_data = bottom[0]->cpu_data();
242 |     Dtype* weight_diff_base = this->blobs_[0]->mutable_cpu_diff();
243 |     for (int n = 0; n < num; ++n)
244 |     {
245 |       for (int c = 0; c < channels; ++c)
246 |       {
247 |         for (int h = 0; h < top_height; ++h)
248 |         {
249 |           for (int w = 0; w < top_width; ++w)
250 |           {
251 |             Dtype* weight_diff = weight_diff_base + c * kernel_h_ * kernel_w_;
252 |             for (int kh = 0; kh < kernel_h_; ++kh)
253 |             {
254 |               for (int kw = 0; kw < kernel_w_; ++kw)
255 |               {
256 |                 int h_in = -pad_h_ + h * stride_h_ + kh * dilation_h_;
257 |                 int w_in = -pad_w_ + w * stride_w_ + kw * dilation_w_;
258 |                 if ((h_in >= 0) && (h_in < bottom_height) && (w_in >= 0) && (w_in < bottom_width))
259 |                 {
260 |                   int offset = ((n * channels + c) * bottom_height + h_in) * bottom_width + w_in;
261 |                   *weight_diff += bottom_data[offset] * (*top_diff);
262 |                 }
263 |                 ++weight_diff;
264 |               }
265 |             }
266 |             ++top_diff;
267 |           }
268 |         }
269 |       }
270 |     }
271 |   }
272 |   if (propagate_down[0])
273 |   {
274 |     const Dtype* top_diff = top[0]->cpu_diff();
275 |     const Dtype* weight_data_base = this->blobs_[0]->cpu_data();
276 |     Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
277 |     for (int n = 0; n < num; ++n)
278 |     {
279 |       for (int c = 0; c < channels; ++c)
280 |       {
281 |         for (int h = 0; h < top_height; ++h)
282 |         {
283 |           for (int w = 0; w < top_width; ++w)
284 |           {
285 |             const Dtype* weight_data = weight_data_base + c * kernel_h_ * kernel_w_;
286 |             for (int kh = 0; kh < kernel_h_; ++kh)
287 |             {
288 |               for (int kw = 0; kw < kernel_w_; ++kw)
289 |               {
290 |                 int h_in = -pad_h_ + h * stride_h_ + kh * dilation_h_;
291 |                 int w_in = -pad_w_ + w * stride_w_ + kw * dilation_w_;
292 |                 if ((h_in >= 0) && (h_in < bottom_height) && (w_in >= 0) && (w_in < bottom_width))
293 |                 {
294 |                   int offset = ((n * channels + c) * bottom_height + h_in) * bottom_width + w_in;
295 |                   bottom_diff[offset] += (*weight_data) * (*top_diff);
296 |                 }
297 |                 ++weight_data;
298 |               }
299 |             }
300 |             ++top_diff;
301 |           }
302 |         }
303 |       }
304 |     }
305 |   }
306 | }
307 | 
308 | #ifdef CPU_ONLY
309 | STUB_GPU(ConvolutionDepthwiseLayer);
310 | #endif
311 | 
312 | INSTANTIATE_CLASS(ConvolutionDepthwiseLayer);
313 | REGISTER_LAYER_CLASS(ConvolutionDepthwise);
314 | 
315 | }  // namespace caffe
316 | 


--------------------------------------------------------------------------------
/.gitignore/conv_dw_layer.cu:
--------------------------------------------------------------------------------
  1 | #include <vector>
  2 | #include "caffe/layers/conv_dw_layer.hpp"
  3 | #include "caffe/util/gpu_util.cuh"
  4 | 
  5 | namespace caffe {
  6 | 
  7 | template <typename Dtype>
  8 | __global__ void ConvolutionDepthwiseWeightForward(const int nthreads,
  9 |     const Dtype* const bottom_data, const Dtype* const weight_data, const int num, const int channels,
 10 |     const int top_height, const int top_width, const int bottom_height, const int bottom_width,
 11 |     const int kernel_h, const int kernel_w, const int stride_h, const int stride_w,
 12 |     const int pad_h, const int pad_w, const int dilation_h, const int dilation_w,
 13 |     Dtype* const top_data) {
 14 |   CUDA_KERNEL_LOOP(index, nthreads) {
 15 |     const int n = index / channels / top_height / top_width;
 16 |     const int c = (index / top_height / top_width) % channels;
 17 |     const int h = (index / top_width) % top_height;
 18 |     const int w = index % top_width;
 19 |     const Dtype* weight = weight_data + c * kernel_h * kernel_w;
 20 |     Dtype value = 0;
 21 |     for (int kh = 0; kh < kernel_h; ++kh)
 22 |     {
 23 |       for (int kw = 0; kw < kernel_w; ++kw)
 24 |       {
 25 |         const int h_in = -pad_h + h * stride_h + kh * dilation_h;
 26 |         const int w_in = -pad_w + w * stride_w + kw * dilation_w;
 27 |         if ((h_in >= 0) && (h_in < bottom_height) && (w_in >= 0) && (w_in < bottom_width))
 28 |         {
 29 |           const int offset = ((n * channels + c) * bottom_height + h_in) * bottom_width + w_in;
 30 |           value += (*weight) * bottom_data[offset];
 31 |         }
 32 |         ++weight;
 33 |       }
 34 |     }
 35 |     top_data[index] = value;
 36 |   }
 37 | }
 38 | 
 39 | template <typename Dtype>
 40 | __global__ void ConvolutionDepthwiseBiasForward(const int nthreads,
 41 |     const Dtype* const bias_data, const int num, const int channels,
 42 |     const int top_height, const int top_width, Dtype* const top_data) {
 43 |   CUDA_KERNEL_LOOP(index, nthreads) {
 44 |     const int c = (index / top_height / top_width) % channels;
 45 |     top_data[index] += bias_data[c];
 46 |   }
 47 | }
 48 | 
 49 | template <typename Dtype>
 50 | void ConvolutionDepthwiseLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
 51 |       const vector<Blob<Dtype>*>& top) {
 52 |   const Dtype* bottom_data = bottom[0]->gpu_data();
 53 |   Dtype* top_data = top[0]->mutable_gpu_data();
 54 |   const Dtype* weight_data = this->blobs_[0]->gpu_data();
 55 |   const int count = top[0]->count();
 56 |   const int num = top[0]->num();
 57 |   const int channels = top[0]->channels();
 58 |   const int top_height = top[0]->height();
 59 |   const int top_width = top[0]->width();
 60 |   const int bottom_height = bottom[0]->height();
 61 |   const int bottom_width = bottom[0]->width();
 62 |   ConvolutionDepthwiseWeightForward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(
 63 |       count, bottom_data, weight_data, num, channels,
 64 |       top_height, top_width, bottom_height, bottom_width,
 65 |       kernel_h_, kernel_w_, stride_h_, stride_w_,
 66 |       pad_h_, pad_w_, dilation_h_, dilation_w_, top_data);
 67 |   if (this->layer_param_.convolution_param().bias_term())
 68 |   {
 69 |     const Dtype* bias_data = this->blobs_[1]->gpu_data();
 70 |     ConvolutionDepthwiseBiasForward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(
 71 |         count, bias_data, num, channels,
 72 |         top_height, top_width, top_data);
 73 |   }
 74 | }
 75 | 
 76 | template <typename Dtype>
 77 | __global__ void ConvolutionDepthwiseWeightBackward(const int nthreads,
 78 |     const Dtype* const top_diff, const Dtype* const bottom_data, const int num, const int channels,
 79 |     const int top_height, const int top_width, const int bottom_height, const int bottom_width,
 80 |     const int kernel_h, const int kernel_w, const int stride_h, const int stride_w,
 81 |     const int pad_h, const int pad_w, const int dilation_h, const int dilation_w,
 82 |     Dtype* const buffer_data) {
 83 |   CUDA_KERNEL_LOOP(index, nthreads) {
 84 |     const int h = (index / top_width) % top_height;
 85 |     const int w = index % top_width;
 86 |     const int kh = (index / kernel_w / num / top_height / top_width) % kernel_h;
 87 |     const int kw = (index / num / top_height / top_width) % kernel_w;
 88 |     const int h_in = -pad_h + h * stride_h + kh * dilation_h;
 89 |     const int w_in = -pad_w + w * stride_w + kw * dilation_w;
 90 |     if ((h_in >= 0) && (h_in < bottom_height) && (w_in >= 0) && (w_in < bottom_width))
 91 |     {
 92 |       const int c = index / kernel_h / kernel_w / num / top_height / top_width;
 93 |       const int n = (index / top_height / top_width) % num;
 94 |       const int top_offset = ((n * channels + c) * top_height + h) * top_width + w;
 95 |       const int bottom_offset = ((n * channels + c) * bottom_height + h_in) * bottom_width + w_in;
 96 |       buffer_data[index] = top_diff[top_offset] * bottom_data[bottom_offset];
 97 |     }
 98 |     else
 99 |     {
100 |       buffer_data[index] = 0;
101 |     }
102 |   }
103 | }
104 | 
105 | template <typename Dtype>
106 | __global__ void ConvolutionDepthwiseBottomBackward(const int nthreads,
107 |     const Dtype* const top_diff, const Dtype* const weight_data, const int num, const int channels,
108 |     const int top_height, const int top_width, const int bottom_height, const int bottom_width,
109 |     const int kernel_h, const int kernel_w, const int stride_h, const int stride_w,
110 |     const int pad_h, const int pad_w, const int dilation_h, const int dilation_w,
111 |     Dtype* const bottom_diff) {
112 |   CUDA_KERNEL_LOOP(index, nthreads) {
113 |     const int n = index / channels / bottom_height / bottom_width;
114 |     const int c = (index / bottom_height / bottom_width) % channels;
115 |     const int h = (index / bottom_width) % bottom_height;
116 |     const int w = index % bottom_width;
117 |     const Dtype* weight = weight_data + c * kernel_h * kernel_w;
118 |     Dtype value = 0;
119 |     for (int kh = 0; kh < kernel_h; ++kh)
120 |     {
121 |       for (int kw = 0; kw < kernel_w; ++kw)
122 |       {
123 |         const int h_out_s = h + pad_h - kh * dilation_h;
124 |         const int w_out_s = w + pad_w - kw * dilation_w;
125 |         if (((h_out_s % stride_h) == 0) && ((w_out_s % stride_w) == 0))
126 |         {
127 |           const int h_out = h_out_s / stride_h;
128 |           const int w_out = w_out_s / stride_w;
129 |           if ((h_out >= 0) && (h_out < top_height) && (w_out >= 0) && (w_out < top_width))
130 |           {
131 |             const int offset = ((n * channels + c) * top_height + h_out) * top_width + w_out;
132 |             value += (*weight) * top_diff[offset];
133 |           }
134 |         }
135 |         ++weight;
136 |       }
137 |     }
138 |     bottom_diff[index] += value;
139 |   }
140 | }
141 | 
142 | template <typename Dtype>
143 | __global__ void ConvolutionDepthwiseBiasBackward(const int nthreads,
144 |     const Dtype* const top_diff, const int num, const int channels,
145 |     const int top_height, const int top_width, Dtype* const buffer_data) {
146 |   CUDA_KERNEL_LOOP(index, nthreads) {
147 |     const int c = index / num / top_height / top_width;
148 |     const int n = (index / top_height / top_width) % num;
149 |     const int h = (index / top_width) % top_height;
150 |     const int w = index % top_width;
151 |     const int offset = ((n * channels + c) * top_height + h) * top_width + w;
152 |     buffer_data[index] = top_diff[offset];
153 |   }
154 | }
155 | 
156 | template <typename Dtype>
157 | void ConvolutionDepthwiseLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
158 |       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
159 |   const Dtype* top_diff = top[0]->gpu_diff();
160 |   const int bottom_count = bottom[0]->count();
161 |   const int num = top[0]->num();
162 |   const int channels = top[0]->channels();
163 |   const int top_height = top[0]->height();
164 |   const int top_width = top[0]->width();
165 |   const int bottom_height = bottom[0]->height();
166 |   const int bottom_width = bottom[0]->width();
167 |   const int length = num * top_height * top_width;
168 |   caffe_gpu_set(bottom_count, Dtype(0), bottom[0]->mutable_gpu_diff());
169 |   if (this->layer_param_.convolution_param().bias_term() && this->param_propagate_down_[1])
170 |   {
171 |     const int bias_buffer_count = bias_buffer_.count();
172 |     Dtype* bias_buffer_mutable_data = bias_buffer_.mutable_gpu_data();
173 |     ConvolutionDepthwiseBiasBackward<Dtype><<<CAFFE_GET_BLOCKS(bias_buffer_count), CAFFE_CUDA_NUM_THREADS>>>(
174 |         bias_buffer_count, top_diff, num, channels,
175 |         top_height, top_width, bias_buffer_mutable_data);
176 |     const int bias_count = this->blobs_[1]->count();
177 |     const Dtype* bias_buffer_data = bias_buffer_.gpu_data();
178 |     Dtype* bias_diff = this->blobs_[1]->mutable_gpu_diff();
179 |     const Dtype* bias_multiplier_data = bias_multiplier_.gpu_data();
180 |     caffe_gpu_gemv(CblasNoTrans, bias_count, length, Dtype(1), bias_buffer_data, bias_multiplier_data, Dtype(1), bias_diff);
181 |   }
182 |   if (this->param_propagate_down_[0])
183 |   {
184 |     const int weight_buffer_count = weight_buffer_.count();
185 |     const Dtype* bottom_data = bottom[0]->gpu_data();
186 |     Dtype* weight_buffer_mutable_data = weight_buffer_.mutable_gpu_data();
187 |     ConvolutionDepthwiseWeightBackward<Dtype><<<CAFFE_GET_BLOCKS(weight_buffer_count), CAFFE_CUDA_NUM_THREADS>>>(
188 |         weight_buffer_count, top_diff, bottom_data, num, channels,
189 |         top_height, top_width, bottom_height, bottom_width,
190 |         kernel_h_, kernel_w_, stride_h_, stride_w_,
191 |         pad_h_, pad_w_, dilation_h_, dilation_w_, weight_buffer_mutable_data);
192 |     const int weight_count = this->blobs_[0]->count();
193 |     const Dtype* weight_buffer_data = weight_buffer_.gpu_data();
194 |     Dtype* weight_diff = this->blobs_[0]->mutable_gpu_diff();
195 |     const Dtype* weight_multiplier_data = weight_multiplier_.gpu_data();
196 |     caffe_gpu_gemv(CblasNoTrans, weight_count, length, Dtype(1), weight_buffer_data, weight_multiplier_data, Dtype(1), weight_diff);
197 |   }
198 |   if (propagate_down[0])
199 |   {
200 |     const Dtype* weight_data = this->blobs_[0]->gpu_data();
201 |     Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
202 |     ConvolutionDepthwiseBottomBackward<Dtype><<<CAFFE_GET_BLOCKS(bottom_count), CAFFE_CUDA_NUM_THREADS>>>(
203 |         bottom_count, top_diff, weight_data, num, channels,
204 |         top_height, top_width, bottom_height, bottom_width,
205 |         kernel_h_, kernel_w_, stride_h_, stride_w_,
206 |         pad_h_, pad_w_, dilation_h_, dilation_w_, bottom_diff);
207 |   }
208 | }
209 | 
210 | INSTANTIATE_LAYER_GPU_FUNCS(ConvolutionDepthwiseLayer);
211 | 
212 | }  // namespace caffe
213 | 


--------------------------------------------------------------------------------
/.gitignore/conv_dw_layer.hpp:
--------------------------------------------------------------------------------
 1 | #ifndef CAFFE_CONV_DW_LAYER_HPP_
 2 | #define CAFFE_CONV_DW_LAYER_HPP_
 3 | 
 4 | #include <vector>
 5 | #include "caffe/blob.hpp"
 6 | #include "caffe/layer.hpp"
 7 | #include "caffe/proto/caffe.pb.h"
 8 | 
 9 | namespace caffe {
10 | 
11 | template <typename Dtype>
12 | class ConvolutionDepthwiseLayer : public Layer<Dtype> {
13 |  public:
14 |   explicit ConvolutionDepthwiseLayer(const LayerParameter& param)
15 |       : Layer<Dtype>(param) {}
16 |   virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
17 |       const vector<Blob<Dtype>*>& top);
18 |   virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
19 |       const vector<Blob<Dtype>*>& top);
20 |   virtual inline int ExactNumBottomBlobs() const { return 1; }
21 |   virtual inline int ExactNumTopBlobs() const { return 1; }
22 |   virtual inline const char* type() const { return "ConvolutionDepthwise"; }
23 |  protected:
24 |   virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
25 |       const vector<Blob<Dtype>*>& top);
26 |   virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
27 |       const vector<Blob<Dtype>*>& top);
28 |   virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
29 |       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
30 |   virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
31 |       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
32 |   unsigned int kernel_h_;
33 |   unsigned int kernel_w_;
34 |   unsigned int stride_h_;
35 |   unsigned int stride_w_;
36 |   unsigned int pad_h_;
37 |   unsigned int pad_w_;
38 |   unsigned int dilation_h_;
39 |   unsigned int dilation_w_;
40 |   Blob<Dtype> weight_buffer_;
41 |   Blob<Dtype> weight_multiplier_;
42 |   Blob<Dtype> bias_buffer_;
43 |   Blob<Dtype> bias_multiplier_;
44 | };
45 | 
46 | }  // namespace caffe
47 | 
48 | #endif  // CAFFE_CONV_DW_LAYER_HPP_
49 | 


--------------------------------------------------------------------------------
/.gitignore/mobilenet_v2_new.caffemodel:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/suzhenghang/MobileNetv2/abe0e3ddfd8bed0900eb7238c49d3adcc180f7f2/.gitignore/mobilenet_v2_new.caffemodel


--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
 1 | BSD 3-Clause License
 2 | 
 3 | Copyright (c) 2018, suzhenghang
 4 | All rights reserved.
 5 | 
 6 | Redistribution and use in source and binary forms, with or without
 7 | modification, are permitted provided that the following conditions are met:
 8 | 
 9 | * Redistributions of source code must retain the above copyright notice, this
10 |   list of conditions and the following disclaimer.
11 | 
12 | * Redistributions in binary form must reproduce the above copyright notice,
13 |   this list of conditions and the following disclaimer in the documentation
14 |   and/or other materials provided with the distribution.
15 | 
16 | * Neither the name of the copyright holder nor the names of its
17 |   contributors may be used to endorse or promote products derived from
18 |   this software without specific prior written permission.
19 | 
20 | THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
21 | AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 | IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
23 | DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
24 | FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 | DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
26 | SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
27 | CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
28 | OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
29 | OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30 | 


--------------------------------------------------------------------------------
/MobileUnet1by2.prototxt:
--------------------------------------------------------------------------------
   1 | name: "MOBILENET1by2_hp_Set24_AUX"
   2 | #  transform_param {
   3 | #    scale: 0.017
   4 | #    mirror: false
   5 | #    crop_size: 224
   6 | #    mean_value: [103.94,116.78,123.68]
   7 | #  }
   8 | #input: "data"
   9 | #input_shape { 
  10 | #  dim: 1
  11 | #  dim: 3
  12 | #  dim: 224
  13 | #  dim: 224
  14 | #}
  15 | 
  16 | layer {
  17 |   name: "train_data"
  18 |   type: "ImageSegData"
  19 |   top: "data"
  20 |   top: "label"
  21 |   image_data_param {
  22 |     root_folder: ""
  23 |     source: "trainlist.txt"
  24 |     label_type: PIXEL
  25 |     batch_size: 16
  26 |     shuffle: true
  27 |     ignore_label: 255
  28 |     new_width: 400
  29 |     new_height: 400
  30 |   }
  31 |   transform_param {
  32 |     mirror: true
  33 |     crop_size: 320
  34 |     mean_value: 103.52
  35 |     mean_value: 116.28
  36 |     mean_value: 123.675
  37 |     scale: 0.017
  38 |     #### multi-scale ####
  39 | #    scale_factors: 0.75
  40 | #    scale_factors: 1
  41 | #    scale_factors: 1.25
  42 | #    scale_factors: 1.5
  43 | #    scale_factors: 2.0
  44 |     #### rolation ####
  45 |     max_rotation_angle: 10
  46 |     #### max_color_shift ####
  47 |     max_color_shift: 20
  48 |     #### brightness ####
  49 |     contrast_brightness_adjustment: true
  50 |     min_contrast: 0.5
  51 |     max_contrast: 2.5
  52 |     #### smooth ####
  53 |     smooth_filtering: true
  54 |     max_smooth: 6
  55 | 
  56 |     apply_probability: 0.5
  57 |     debug_params: false
  58 | 
  59 |   }
  60 |   include: { phase: TRAIN }
  61 | }
  62 | layer {
  63 |   name: "test_data"
  64 |   type: "ImageSegData"
  65 |   top: "data"
  66 |   top: "label"
  67 |   image_data_param {
  68 |     root_folder: ""
  69 |     source: "vallist.txt"
  70 |     label_type: PIXEL
  71 |     batch_size: 1
  72 |     shuffle: true
  73 |     ignore_label: 255
  74 |     new_width: 400
  75 |     new_height: 400
  76 |   }
  77 |   transform_param {
  78 |     mean_value: 103.94
  79 |     mean_value: 116.78
  80 |     mean_value: 123.68
  81 |     crop_size: 320
  82 |     scale: 0.017
  83 |     mirror: false
  84 |   }
  85 |   include: { phase: TEST }
  86 | }
  87 | 
  88 | layer {
  89 |   name: "conv1_1by2"
  90 |   type: "Convolution"
  91 |   bottom: "data"
  92 |   top: "conv1_1by2"
  93 |   param {
  94 |     lr_mult: 1
  95 |     decay_mult: 1
  96 |   }
  97 |   convolution_param {
  98 |     num_output: 16
  99 |     bias_term: false
 100 |     pad: 1
 101 |     kernel_size: 3
 102 |     stride: 2
 103 |     engine: CAFFE
 104 |     weight_filler {
 105 |       type: "msra"
 106 |     }
 107 |   }
 108 | }
 109 | layer {
 110 |   name: "conv1/bn_1by2"
 111 |   type: "BatchNorm"
 112 |   bottom: "conv1_1by2"
 113 |   top: "conv1_1by2"
 114 |   param {
 115 |     lr_mult: 0
 116 |     decay_mult: 0
 117 |   }
 118 |   param {
 119 |     lr_mult: 0
 120 |     decay_mult: 0
 121 |   }
 122 |   param {
 123 |     lr_mult: 0
 124 |     decay_mult: 0
 125 |   }
 126 | }
 127 | layer {
 128 |   name: "conv1/scale_1by2"
 129 |   type: "Scale"
 130 |   bottom: "conv1_1by2"
 131 |   top: "conv1_1by2"
 132 |   scale_param {
 133 |     filler {
 134 |       value: 1
 135 |     }
 136 |     bias_term: true
 137 |     bias_filler {
 138 |       value: 0
 139 |     }
 140 |   }
 141 | }
 142 | layer {
 143 |   name: "relu1_1by2"
 144 |   type: "ReLU"
 145 |   bottom: "conv1_1by2"
 146 |   top: "conv1_1by2"
 147 | }
 148 | layer {
 149 |   name: "conv2_1/dw_1by2"
 150 |   type: "ConvolutionDepthwise"
 151 |   bottom: "conv1_1by2"
 152 |   top: "conv2_1/dw_1by2"
 153 |   param {
 154 |     lr_mult: 1
 155 |     decay_mult: 1
 156 |   }
 157 |   convolution_param {
 158 |     num_output: 16
 159 |     bias_term: false
 160 |     pad: 1
 161 |     kernel_size: 3
 162 |     stride: 1
 163 |     engine: CAFFE
 164 |     weight_filler {
 165 |       type: "msra"
 166 |     }
 167 |   }
 168 | }
 169 | layer {
 170 |   name: "conv2_1/dw/bn_1by2"
 171 |   type: "BatchNorm"
 172 |   bottom: "conv2_1/dw_1by2"
 173 |   top: "conv2_1/dw_1by2"
 174 |   param {
 175 |     lr_mult: 0
 176 |     decay_mult: 0
 177 |   }
 178 |   param {
 179 |     lr_mult: 0
 180 |     decay_mult: 0
 181 |   }
 182 |   param {
 183 |     lr_mult: 0
 184 |     decay_mult: 0
 185 |   }
 186 | }
 187 | layer {
 188 |   name: "conv2_1/dw/scale_1by2"
 189 |   type: "Scale"
 190 |   bottom: "conv2_1/dw_1by2"
 191 |   top: "conv2_1/dw_1by2"
 192 |   scale_param {
 193 |     filler {
 194 |       value: 1
 195 |     }
 196 |     bias_term: true
 197 |     bias_filler {
 198 |       value: 0
 199 |     }
 200 |   }
 201 | }
 202 | layer {
 203 |   name: "relu2_1/dw_1by2"
 204 |   type: "ReLU"
 205 |   bottom: "conv2_1/dw_1by2"
 206 |   top: "conv2_1/dw_1by2"
 207 | }
 208 | layer {
 209 |   name: "conv2_1/sep_1by2"
 210 |   type: "Convolution"
 211 |   bottom: "conv2_1/dw_1by2"
 212 |   top: "conv2_1/sep_1by2"
 213 |   param {
 214 |     lr_mult: 1
 215 |     decay_mult: 1
 216 |   }
 217 |   convolution_param {
 218 |     num_output: 32
 219 |     bias_term: false
 220 |     pad: 0
 221 |     kernel_size: 1
 222 |     stride: 1
 223 |     engine: CAFFE
 224 |     weight_filler {
 225 |       type: "msra"
 226 |     }
 227 |   }
 228 | }
 229 | layer {
 230 |   name: "conv2_1/sep/bn_1by2"
 231 |   type: "BatchNorm"
 232 |   bottom: "conv2_1/sep_1by2"
 233 |   top: "conv2_1/sep_1by2"
 234 |   param {
 235 |     lr_mult: 0
 236 |     decay_mult: 0
 237 |   }
 238 |   param {
 239 |     lr_mult: 0
 240 |     decay_mult: 0
 241 |   }
 242 |   param {
 243 |     lr_mult: 0
 244 |     decay_mult: 0
 245 |   }
 246 | }
 247 | layer {
 248 |   name: "conv2_1/sep/scale_1by2"
 249 |   type: "Scale"
 250 |   bottom: "conv2_1/sep_1by2"
 251 |   top: "conv2_1/sep_1by2"
 252 |   scale_param {
 253 |     filler {
 254 |       value: 1
 255 |     }
 256 |     bias_term: true
 257 |     bias_filler {
 258 |       value: 0
 259 |     }
 260 |   }
 261 | }
 262 | layer {
 263 |   name: "relu2_1/sep_1by2"
 264 |   type: "ReLU"
 265 |   bottom: "conv2_1/sep_1by2"
 266 |   top: "conv2_1/sep_1by2"
 267 | }
 268 | layer {
 269 |   name: "conv2_2/dw_1by2"
 270 |   type: "ConvolutionDepthwise"
 271 |   bottom: "conv2_1/sep_1by2"
 272 |   top: "conv2_2/dw_1by2"
 273 |   param {
 274 |     lr_mult: 1
 275 |     decay_mult: 1
 276 |   }
 277 |   convolution_param {
 278 |     num_output: 32
 279 |     bias_term: false
 280 |     pad: 1
 281 |     kernel_size: 3
 282 |     stride: 2
 283 |     engine: CAFFE
 284 |     weight_filler {
 285 |       type: "msra"
 286 |     }
 287 |   }
 288 | }
 289 | layer {
 290 |   name: "conv2_2/dw/bn_1by2"
 291 |   type: "BatchNorm"
 292 |   bottom: "conv2_2/dw_1by2"
 293 |   top: "conv2_2/dw_1by2"
 294 |   param {
 295 |     lr_mult: 0
 296 |     decay_mult: 0
 297 |   }
 298 |   param {
 299 |     lr_mult: 0
 300 |     decay_mult: 0
 301 |   }
 302 |   param {
 303 |     lr_mult: 0
 304 |     decay_mult: 0
 305 |   }
 306 | }
 307 | layer {
 308 |   name: "conv2_2/dw/scale_1by2"
 309 |   type: "Scale"
 310 |   bottom: "conv2_2/dw_1by2"
 311 |   top: "conv2_2/dw_1by2"
 312 |   scale_param {
 313 |     filler {
 314 |       value: 1
 315 |     }
 316 |     bias_term: true
 317 |     bias_filler {
 318 |       value: 0
 319 |     }
 320 |   }
 321 | }
 322 | layer {
 323 |   name: "relu2_2/dw_1by2"
 324 |   type: "ReLU"
 325 |   bottom: "conv2_2/dw_1by2"
 326 |   top: "conv2_2/dw_1by2"
 327 | }
 328 | layer {
 329 |   name: "conv2_2/sep_1by2"
 330 |   type: "Convolution"
 331 |   bottom: "conv2_2/dw_1by2"
 332 |   top: "conv2_2/sep_1by2"
 333 |   param {
 334 |     lr_mult: 1
 335 |     decay_mult: 1
 336 |   }
 337 |   convolution_param {
 338 |     num_output: 64
 339 |     bias_term: false
 340 |     pad: 0
 341 |     kernel_size: 1
 342 |     stride: 1
 343 |     engine: CAFFE
 344 |     weight_filler {
 345 |       type: "msra"
 346 |     }
 347 |   }
 348 | }
 349 | layer {
 350 |   name: "conv2_2/sep/bn_1by2"
 351 |   type: "BatchNorm"
 352 |   bottom: "conv2_2/sep_1by2"
 353 |   top: "conv2_2/sep_1by2"
 354 |   param {
 355 |     lr_mult: 0
 356 |     decay_mult: 0
 357 |   }
 358 |   param {
 359 |     lr_mult: 0
 360 |     decay_mult: 0
 361 |   }
 362 |   param {
 363 |     lr_mult: 0
 364 |     decay_mult: 0
 365 |   }
 366 | }
 367 | layer {
 368 |   name: "conv2_2/sep/scale_1by2"
 369 |   type: "Scale"
 370 |   bottom: "conv2_2/sep_1by2"
 371 |   top: "conv2_2/sep_1by2"
 372 |   scale_param {
 373 |     filler {
 374 |       value: 1
 375 |     }
 376 |     bias_term: true
 377 |     bias_filler {
 378 |       value: 0
 379 |     }
 380 |   }
 381 | }
 382 | layer {
 383 |   name: "relu2_2/sep_1by2"
 384 |   type: "ReLU"
 385 |   bottom: "conv2_2/sep_1by2"
 386 |   top: "conv2_2/sep_1by2"
 387 | }
 388 | layer {
 389 |   name: "conv3_1/dw_1by2"
 390 |   type: "ConvolutionDepthwise"
 391 |   bottom: "conv2_2/sep_1by2"
 392 |   top: "conv3_1/dw_1by2"
 393 |   param {
 394 |     lr_mult: 1
 395 |     decay_mult: 1
 396 |   }
 397 |   convolution_param {
 398 |     num_output: 64
 399 |     bias_term: false
 400 |     pad: 1
 401 |     kernel_size: 3
 402 |     stride: 1
 403 |     engine: CAFFE
 404 |     weight_filler {
 405 |       type: "msra"
 406 |     }
 407 |   }
 408 | }
 409 | layer {
 410 |   name: "conv3_1/dw/bn_1by2"
 411 |   type: "BatchNorm"
 412 |   bottom: "conv3_1/dw_1by2"
 413 |   top: "conv3_1/dw_1by2"
 414 |   param {
 415 |     lr_mult: 0
 416 |     decay_mult: 0
 417 |   }
 418 |   param {
 419 |     lr_mult: 0
 420 |     decay_mult: 0
 421 |   }
 422 |   param {
 423 |     lr_mult: 0
 424 |     decay_mult: 0
 425 |   }
 426 | }
 427 | layer {
 428 |   name: "conv3_1/dw/scale_1by2"
 429 |   type: "Scale"
 430 |   bottom: "conv3_1/dw_1by2"
 431 |   top: "conv3_1/dw_1by2"
 432 |   scale_param {
 433 |     filler {
 434 |       value: 1
 435 |     }
 436 |     bias_term: true
 437 |     bias_filler {
 438 |       value: 0
 439 |     }
 440 |   }
 441 | }
 442 | layer {
 443 |   name: "relu3_1/dw_1by2"
 444 |   type: "ReLU"
 445 |   bottom: "conv3_1/dw_1by2"
 446 |   top: "conv3_1/dw_1by2"
 447 | }
 448 | layer {
 449 |   name: "conv3_1/sep_1by2"
 450 |   type: "Convolution"
 451 |   bottom: "conv3_1/dw_1by2"
 452 |   top: "conv3_1/sep_1by2"
 453 |   param {
 454 |     lr_mult: 1
 455 |     decay_mult: 1
 456 |   }
 457 |   convolution_param {
 458 |     num_output: 64
 459 |     bias_term: false
 460 |     pad: 0
 461 |     kernel_size: 1
 462 |     stride: 1
 463 |     engine: CAFFE
 464 |     weight_filler {
 465 |       type: "msra"
 466 |     }
 467 |   }
 468 | }
 469 | layer {
 470 |   name: "conv3_1/sep/bn_1by2"
 471 |   type: "BatchNorm"
 472 |   bottom: "conv3_1/sep_1by2"
 473 |   top: "conv3_1/sep_1by2"
 474 |   param {
 475 |     lr_mult: 0
 476 |     decay_mult: 0
 477 |   }
 478 |   param {
 479 |     lr_mult: 0
 480 |     decay_mult: 0
 481 |   }
 482 |   param {
 483 |     lr_mult: 0
 484 |     decay_mult: 0
 485 |   }
 486 | }
 487 | layer {
 488 |   name: "conv3_1/sep/scale_1by2"
 489 |   type: "Scale"
 490 |   bottom: "conv3_1/sep_1by2"
 491 |   top: "conv3_1/sep_1by2"
 492 |   scale_param {
 493 |     filler {
 494 |       value: 1
 495 |     }
 496 |     bias_term: true
 497 |     bias_filler {
 498 |       value: 0
 499 |     }
 500 |   }
 501 | }
 502 | layer {
 503 |   name: "relu3_1/sep_1by2"
 504 |   type: "ReLU"
 505 |   bottom: "conv3_1/sep_1by2"
 506 |   top: "conv3_1/sep_1by2"
 507 | }
 508 | layer {
 509 |   name: "conv3_2/dw_1by2"
 510 |   type: "ConvolutionDepthwise"
 511 |   bottom: "conv3_1/sep_1by2"
 512 |   top: "conv3_2/dw_1by2"
 513 |   param {
 514 |     lr_mult: 1
 515 |     decay_mult: 1
 516 |   }
 517 |   convolution_param {
 518 |     num_output: 64
 519 |     bias_term: false
 520 |     pad: 1
 521 |     kernel_size: 3
 522 |     stride: 2
 523 |     engine: CAFFE
 524 |     weight_filler {
 525 |       type: "msra"
 526 |     }
 527 |   }
 528 | }
 529 | layer {
 530 |   name: "conv3_2/dw/bn_1by2"
 531 |   type: "BatchNorm"
 532 |   bottom: "conv3_2/dw_1by2"
 533 |   top: "conv3_2/dw_1by2"
 534 |   param {
 535 |     lr_mult: 0
 536 |     decay_mult: 0
 537 |   }
 538 |   param {
 539 |     lr_mult: 0
 540 |     decay_mult: 0
 541 |   }
 542 |   param {
 543 |     lr_mult: 0
 544 |     decay_mult: 0
 545 |   }
 546 | }
 547 | layer {
 548 |   name: "conv3_2/dw/scale_1by2"
 549 |   type: "Scale"
 550 |   bottom: "conv3_2/dw_1by2"
 551 |   top: "conv3_2/dw_1by2"
 552 |   scale_param {
 553 |     filler {
 554 |       value: 1
 555 |     }
 556 |     bias_term: true
 557 |     bias_filler {
 558 |       value: 0
 559 |     }
 560 |   }
 561 | }
 562 | layer {
 563 |   name: "relu3_2/dw_1by2"
 564 |   type: "ReLU"
 565 |   bottom: "conv3_2/dw_1by2"
 566 |   top: "conv3_2/dw_1by2"
 567 | }
 568 | layer {
 569 |   name: "conv3_2/sep_1by2"
 570 |   type: "Convolution"
 571 |   bottom: "conv3_2/dw_1by2"
 572 |   top: "conv3_2/sep_1by2"
 573 |   param {
 574 |     lr_mult: 1
 575 |     decay_mult: 1
 576 |   }
 577 |   convolution_param {
 578 |     num_output: 128
 579 |     bias_term: false
 580 |     pad: 0
 581 |     kernel_size: 1
 582 |     stride: 1
 583 |     engine: CAFFE
 584 |     weight_filler {
 585 |       type: "msra"
 586 |     }
 587 |   }
 588 | }
 589 | layer {
 590 |   name: "conv3_2/sep/bn_1by2"
 591 |   type: "BatchNorm"
 592 |   bottom: "conv3_2/sep_1by2"
 593 |   top: "conv3_2/sep_1by2"
 594 |   param {
 595 |     lr_mult: 0
 596 |     decay_mult: 0
 597 |   }
 598 |   param {
 599 |     lr_mult: 0
 600 |     decay_mult: 0
 601 |   }
 602 |   param {
 603 |     lr_mult: 0
 604 |     decay_mult: 0
 605 |   }
 606 | }
 607 | layer {
 608 |   name: "conv3_2/sep/scale_1by2"
 609 |   type: "Scale"
 610 |   bottom: "conv3_2/sep_1by2"
 611 |   top: "conv3_2/sep_1by2"
 612 |   scale_param {
 613 |     filler {
 614 |       value: 1
 615 |     }
 616 |     bias_term: true
 617 |     bias_filler {
 618 |       value: 0
 619 |     }
 620 |   }
 621 | }
 622 | layer {
 623 |   name: "relu3_2/sep_1by2"
 624 |   type: "ReLU"
 625 |   bottom: "conv3_2/sep_1by2"
 626 |   top: "conv3_2/sep_1by2"
 627 | }
 628 | layer {
 629 |   name: "conv4_1/dw_1by2"
 630 |   type: "ConvolutionDepthwise"
 631 |   bottom: "conv3_2/sep_1by2"
 632 |   top: "conv4_1/dw_1by2"
 633 |   param {
 634 |     lr_mult: 1
 635 |     decay_mult: 1
 636 |   }
 637 |   convolution_param {
 638 |     num_output: 128
 639 |     bias_term: false
 640 |     pad: 1
 641 |     kernel_size: 3
 642 |     stride: 1
 643 |     engine: CAFFE
 644 |     weight_filler {
 645 |       type: "msra"
 646 |     }
 647 |   }
 648 | }
 649 | layer {
 650 |   name: "conv4_1/dw/bn_1by2"
 651 |   type: "BatchNorm"
 652 |   bottom: "conv4_1/dw_1by2"
 653 |   top: "conv4_1/dw_1by2"
 654 |   param {
 655 |     lr_mult: 0
 656 |     decay_mult: 0
 657 |   }
 658 |   param {
 659 |     lr_mult: 0
 660 |     decay_mult: 0
 661 |   }
 662 |   param {
 663 |     lr_mult: 0
 664 |     decay_mult: 0
 665 |   }
 666 | }
 667 | layer {
 668 |   name: "conv4_1/dw/scale_1by2"
 669 |   type: "Scale"
 670 |   bottom: "conv4_1/dw_1by2"
 671 |   top: "conv4_1/dw_1by2"
 672 |   scale_param {
 673 |     filler {
 674 |       value: 1
 675 |     }
 676 |     bias_term: true
 677 |     bias_filler {
 678 |       value: 0
 679 |     }
 680 |   }
 681 | }
 682 | layer {
 683 |   name: "relu4_1/dw_1by2"
 684 |   type: "ReLU"
 685 |   bottom: "conv4_1/dw_1by2"
 686 |   top: "conv4_1/dw_1by2"
 687 | }
 688 | layer {
 689 |   name: "conv4_1/sep_1by2"
 690 |   type: "Convolution"
 691 |   bottom: "conv4_1/dw_1by2"
 692 |   top: "conv4_1/sep_1by2"
 693 |   param {
 694 |     lr_mult: 1
 695 |     decay_mult: 1
 696 |   }
 697 |   convolution_param {
 698 |     num_output: 128
 699 |     bias_term: false
 700 |     pad: 0
 701 |     kernel_size: 1
 702 |     engine: CAFFE
 703 |     stride: 1
 704 |     weight_filler {
 705 |       type: "msra"
 706 |     }
 707 |   }
 708 | }
 709 | layer {
 710 |   name: "conv4_1/sep/bn_1by2"
 711 |   type: "BatchNorm"
 712 |   bottom: "conv4_1/sep_1by2"
 713 |   top: "conv4_1/sep_1by2"
 714 |   param {
 715 |     lr_mult: 0
 716 |     decay_mult: 0
 717 |   }
 718 |   param {
 719 |     lr_mult: 0
 720 |     decay_mult: 0
 721 |   }
 722 |   param {
 723 |     lr_mult: 0
 724 |     decay_mult: 0
 725 |   }
 726 | }
 727 | layer {
 728 |   name: "conv4_1/sep/scale_1by2"
 729 |   type: "Scale"
 730 |   bottom: "conv4_1/sep_1by2"
 731 |   top: "conv4_1/sep_1by2"
 732 |   scale_param {
 733 |     filler {
 734 |       value: 1
 735 |     }
 736 |     bias_term: true
 737 |     bias_filler {
 738 |       value: 0
 739 |     }
 740 |   }
 741 | }
 742 | layer {
 743 |   name: "relu4_1/sep_1by2"
 744 |   type: "ReLU"
 745 |   bottom: "conv4_1/sep_1by2"
 746 |   top: "conv4_1/sep_1by2"
 747 | }
 748 | layer {
 749 |   name: "conv4_2/dw_1by2"
 750 |   type: "ConvolutionDepthwise"
 751 |   bottom: "conv4_1/sep_1by2"
 752 |   top: "conv4_2/dw_1by2"
 753 |   param {
 754 |     lr_mult: 1
 755 |     decay_mult: 1
 756 |   }
 757 |   convolution_param {
 758 |     num_output: 128
 759 |     bias_term: false
 760 |     pad: 1
 761 |     kernel_size: 3
 762 |     stride: 2
 763 |     engine: CAFFE
 764 |     weight_filler {
 765 |       type: "msra"
 766 |     }
 767 |   }
 768 | }
 769 | layer {
 770 |   name: "conv4_2/dw/bn_1by2"
 771 |   type: "BatchNorm"
 772 |   bottom: "conv4_2/dw_1by2"
 773 |   top: "conv4_2/dw_1by2"
 774 |   param {
 775 |     lr_mult: 0
 776 |     decay_mult: 0
 777 |   }
 778 |   param {
 779 |     lr_mult: 0
 780 |     decay_mult: 0
 781 |   }
 782 |   param {
 783 |     lr_mult: 0
 784 |     decay_mult: 0
 785 |   }
 786 | }
 787 | layer {
 788 |   name: "conv4_2/dw/scale_1by2"
 789 |   type: "Scale"
 790 |   bottom: "conv4_2/dw_1by2"
 791 |   top: "conv4_2/dw_1by2"
 792 |   scale_param {
 793 |     filler {
 794 |       value: 1
 795 |     }
 796 |     bias_term: true
 797 |     bias_filler {
 798 |       value: 0
 799 |     }
 800 |   }
 801 | }
 802 | layer {
 803 |   name: "relu4_2/dw_1by2"
 804 |   type: "ReLU"
 805 |   bottom: "conv4_2/dw_1by2"
 806 |   top: "conv4_2/dw_1by2"
 807 | }
 808 | layer {
 809 |   name: "conv4_2/sep_1by2"
 810 |   type: "Convolution"
 811 |   bottom: "conv4_2/dw_1by2"
 812 |   top: "conv4_2/sep_1by2"
 813 |   param {
 814 |     lr_mult: 1
 815 |     decay_mult: 1
 816 |   }
 817 |   convolution_param {
 818 |     num_output: 128
 819 |     bias_term: false
 820 |     pad: 0
 821 |     kernel_size: 1
 822 |     stride: 1
 823 |     engine: CAFFE
 824 |     weight_filler {
 825 |       type: "msra"
 826 |     }
 827 |   }
 828 | }
 829 | layer {
 830 |   name: "conv4_2/sep/bn_1by2"
 831 |   type: "BatchNorm"
 832 |   bottom: "conv4_2/sep_1by2"
 833 |   top: "conv4_2/sep_1by2"
 834 |   param {
 835 |     lr_mult: 0
 836 |     decay_mult: 0
 837 |   }
 838 |   param {
 839 |     lr_mult: 0
 840 |     decay_mult: 0
 841 |   }
 842 |   param {
 843 |     lr_mult: 0
 844 |     decay_mult: 0
 845 |   }
 846 | }
 847 | layer {
 848 |   name: "conv4_2/sep/scale_1by2"
 849 |   type: "Scale"
 850 |   bottom: "conv4_2/sep_1by2"
 851 |   top: "conv4_2/sep_1by2"
 852 |   scale_param {
 853 |     filler {
 854 |       value: 1
 855 |     }
 856 |     bias_term: true
 857 |     bias_filler {
 858 |       value: 0
 859 |     }
 860 |   }
 861 | }
 862 | layer {
 863 |   name: "relu4_2/sep_1by2"
 864 |   type: "ReLU"
 865 |   bottom: "conv4_2/sep_1by2"
 866 |   top: "conv4_2/sep_1by2"
 867 | }
 868 | layer {
 869 |   name: "conv5_1/dw_1by2"
 870 |   type: "ConvolutionDepthwise"
 871 |   bottom: "conv4_2/sep_1by2"
 872 |   top: "conv5_1/dw_1by2"
 873 |   param {
 874 |     lr_mult: 1
 875 |     decay_mult: 1
 876 |   }
 877 |   convolution_param {
 878 |     num_output: 128
 879 |     bias_term: false
 880 |     pad: 1
 881 |     kernel_size: 3
 882 |     stride: 1
 883 |     engine: CAFFE
 884 |     dilation: 1
 885 |     weight_filler {
 886 |       type: "msra"
 887 |     }
 888 |   }
 889 | }
 890 | layer {
 891 |   name: "conv5_1/dw/bn_1by2"
 892 |   type: "BatchNorm"
 893 |   bottom: "conv5_1/dw_1by2"
 894 |   top: "conv5_1/dw_1by2"
 895 |   param {
 896 |     lr_mult: 0
 897 |     decay_mult: 0
 898 |   }
 899 |   param {
 900 |     lr_mult: 0
 901 |     decay_mult: 0
 902 |   }
 903 |   param {
 904 |     lr_mult: 0
 905 |     decay_mult: 0
 906 |   }
 907 | }
 908 | layer {
 909 |   name: "conv5_1/dw/scale_1by2"
 910 |   type: "Scale"
 911 |   bottom: "conv5_1/dw_1by2"
 912 |   top: "conv5_1/dw_1by2"
 913 |   scale_param {
 914 |     filler {
 915 |       value: 1
 916 |     }
 917 |     bias_term: true
 918 |     bias_filler {
 919 |       value: 0
 920 |     }
 921 |   }
 922 | }
 923 | layer {
 924 |   name: "relu5_1/dw_1by2"
 925 |   type: "ReLU"
 926 |   bottom: "conv5_1/dw_1by2"
 927 |   top: "conv5_1/dw_1by2"
 928 | }
 929 | layer {
 930 |   name: "conv5_1/sep_1by2"
 931 |   type: "Convolution"
 932 |   bottom: "conv5_1/dw_1by2"
 933 |   top: "conv5_1/sep_1by2"
 934 |   param {
 935 |     lr_mult: 1
 936 |     decay_mult: 1
 937 |   }
 938 |   convolution_param {
 939 |     num_output: 128
 940 |     bias_term: false
 941 |     pad: 0
 942 |     kernel_size: 1
 943 |     engine: CAFFE
 944 |     stride: 1
 945 |     weight_filler {
 946 |       type: "msra"
 947 |     }
 948 |   }
 949 | }
 950 | layer {
 951 |   name: "conv5_1/sep/bn_1by2"
 952 |   type: "BatchNorm"
 953 |   bottom: "conv5_1/sep_1by2"
 954 |   top: "conv5_1/sep_1by2"
 955 |   param {
 956 |     lr_mult: 0
 957 |     decay_mult: 0
 958 |   }
 959 |   param {
 960 |     lr_mult: 0
 961 |     decay_mult: 0
 962 |   }
 963 |   param {
 964 |     lr_mult: 0
 965 |     decay_mult: 0
 966 |   }
 967 | }
 968 | layer {
 969 |   name: "conv5_1/sep/scale_1by2"
 970 |   type: "Scale"
 971 |   bottom: "conv5_1/sep_1by2"
 972 |   top: "conv5_1/sep_1by2"
 973 |   scale_param {
 974 |     filler {
 975 |       value: 1
 976 |     }
 977 |     bias_term: true
 978 |     bias_filler {
 979 |       value: 0
 980 |     }
 981 |   }
 982 | }
 983 | layer {
 984 |   name: "relu5_1/sep_1by2"
 985 |   type: "ReLU"
 986 |   bottom: "conv5_1/sep_1by2"
 987 |   top: "conv5_1/sep_1by2"
 988 | }
 989 | layer {
 990 |   name: "conv5_2/dw_1by2"
 991 |   type: "ConvolutionDepthwise"
 992 |   bottom: "conv5_1/sep_1by2"
 993 |   top: "conv5_2/dw_1by2"
 994 |   param {
 995 |     lr_mult: 1
 996 |     decay_mult: 1
 997 |   }
 998 |   convolution_param {
 999 |     num_output: 128
1000 |     bias_term: false
1001 |     pad: 2
1002 |     kernel_size: 3
1003 |     stride: 1
1004 |     engine: CAFFE
1005 |     dilation: 2
1006 |     weight_filler {
1007 |       type: "msra"
1008 |     }
1009 |   }
1010 | }
1011 | layer {
1012 |   name: "conv5_2/dw/bn_1by2"
1013 |   type: "BatchNorm"
1014 |   bottom: "conv5_2/dw_1by2"
1015 |   top: "conv5_2/dw_1by2"
1016 |   param {
1017 |     lr_mult: 0
1018 |     decay_mult: 0
1019 |   }
1020 |   param {
1021 |     lr_mult: 0
1022 |     decay_mult: 0
1023 |   }
1024 |   param {
1025 |     lr_mult: 0
1026 |     decay_mult: 0
1027 |   }
1028 | }
1029 | layer {
1030 |   name: "conv5_2/dw/scale_1by2"
1031 |   type: "Scale"
1032 |   bottom: "conv5_2/dw_1by2"
1033 |   top: "conv5_2/dw_1by2"
1034 |   scale_param {
1035 |     filler {
1036 |       value: 1
1037 |     }
1038 |     bias_term: true
1039 |     bias_filler {
1040 |       value: 0
1041 |     }
1042 |   }
1043 | }
1044 | layer {
1045 |   name: "relu5_2/dw_1by2"
1046 |   type: "ReLU"
1047 |   bottom: "conv5_2/dw_1by2"
1048 |   top: "conv5_2/dw_1by2"
1049 | }
1050 | layer {
1051 |   name: "conv5_2/sep_1by2"
1052 |   type: "Convolution"
1053 |   bottom: "conv5_2/dw_1by2"
1054 |   top: "conv5_2/sep_1by2"
1055 |   param {
1056 |     lr_mult: 1
1057 |     decay_mult: 1
1058 |   }
1059 |   convolution_param {
1060 |     num_output: 128
1061 |     bias_term: false
1062 |     pad: 0
1063 |     kernel_size: 1
1064 |     stride: 1
1065 |     engine: CAFFE
1066 |     weight_filler {
1067 |       type: "msra"
1068 |     }
1069 |   }
1070 | }
1071 | layer {
1072 |   name: "conv5_2/sep/bn_1by2"
1073 |   type: "BatchNorm"
1074 |   bottom: "conv5_2/sep_1by2"
1075 |   top: "conv5_2/sep_1by2"
1076 |   param {
1077 |     lr_mult: 0
1078 |     decay_mult: 0
1079 |   }
1080 |   param {
1081 |     lr_mult: 0
1082 |     decay_mult: 0
1083 |   }
1084 |   param {
1085 |     lr_mult: 0
1086 |     decay_mult: 0
1087 |   }
1088 | }
1089 | layer {
1090 |   name: "conv5_2/sep/scale_1by2"
1091 |   type: "Scale"
1092 |   bottom: "conv5_2/sep_1by2"
1093 |   top: "conv5_2/sep_1by2"
1094 |   scale_param {
1095 |     filler {
1096 |       value: 1
1097 |     }
1098 |     bias_term: true
1099 |     bias_filler {
1100 |       value: 0
1101 |     }
1102 |   }
1103 | }
1104 | layer {
1105 |   name: "relu5_2/sep_1by2"
1106 |   type: "ReLU"
1107 |   bottom: "conv5_2/sep_1by2"
1108 |   top: "conv5_2/sep_1by2"
1109 | }
1110 | layer {
1111 |   name: "conv5_3/dw_1by2"
1112 |   type: "ConvolutionDepthwise"
1113 |   bottom: "conv5_2/sep_1by2"
1114 |   top: "conv5_3/dw_1by2"
1115 |   param {
1116 |     lr_mult: 1
1117 |     decay_mult: 1
1118 |   }
1119 |   convolution_param {
1120 |     num_output: 128
1121 |     bias_term: false
1122 |     pad: 5
1123 |     kernel_size: 3
1124 |     stride: 1
1125 |     engine: CAFFE
1126 |     dilation: 5
1127 |     weight_filler {
1128 |       type: "msra"
1129 |     }
1130 |   }
1131 | }
1132 | layer {
1133 |   name: "conv5_3/dw/bn_1by2"
1134 |   type: "BatchNorm"
1135 |   bottom: "conv5_3/dw_1by2"
1136 |   top: "conv5_3/dw_1by2"
1137 |   param {
1138 |     lr_mult: 0
1139 |     decay_mult: 0
1140 |   }
1141 |   param {
1142 |     lr_mult: 0
1143 |     decay_mult: 0
1144 |   }
1145 |   param {
1146 |     lr_mult: 0
1147 |     decay_mult: 0
1148 |   }
1149 | }
1150 | layer {
1151 |   name: "conv5_3/dw/scale_1by2"
1152 |   type: "Scale"
1153 |   bottom: "conv5_3/dw_1by2"
1154 |   top: "conv5_3/dw_1by2"
1155 |   scale_param {
1156 |     filler {
1157 |       value: 1
1158 |     }
1159 |     bias_term: true
1160 |     bias_filler {
1161 |       value: 0
1162 |     }
1163 |   }
1164 | }
1165 | layer {
1166 |   name: "relu5_3/dw_1by2"
1167 |   type: "ReLU"
1168 |   bottom: "conv5_3/dw_1by2"
1169 |   top: "conv5_3/dw_1by2"
1170 | }
1171 | layer {
1172 |   name: "conv5_3/sep_1by2"
1173 |   type: "Convolution"
1174 |   bottom: "conv5_3/dw_1by2"
1175 |   top: "conv5_3/sep_1by2"
1176 |   param {
1177 |     lr_mult: 1
1178 |     decay_mult: 1
1179 |   }
1180 |   convolution_param {
1181 |     num_output: 128
1182 |     bias_term: false
1183 |     pad: 0
1184 |     kernel_size: 1
1185 |     stride: 1
1186 |     engine: CAFFE
1187 |     weight_filler {
1188 |       type: "msra"
1189 |     }
1190 |   }
1191 | }
1192 | layer {
1193 |   name: "conv5_3/sep/bn_1by2"
1194 |   type: "BatchNorm"
1195 |   bottom: "conv5_3/sep_1by2"
1196 |   top: "conv5_3/sep_1by2"
1197 |   param {
1198 |     lr_mult: 0
1199 |     decay_mult: 0
1200 |   }
1201 |   param {
1202 |     lr_mult: 0
1203 |     decay_mult: 0
1204 |   }
1205 |   param {
1206 |     lr_mult: 0
1207 |     decay_mult: 0
1208 |   }
1209 | }
1210 | layer {
1211 |   name: "conv5_3/sep/scale_1by2"
1212 |   type: "Scale"
1213 |   bottom: "conv5_3/sep_1by2"
1214 |   top: "conv5_3/sep_1by2"
1215 |   scale_param {
1216 |     filler {
1217 |       value: 1
1218 |     }
1219 |     bias_term: true
1220 |     bias_filler {
1221 |       value: 0
1222 |     }
1223 |   }
1224 | }
1225 | layer {
1226 |   name: "relu5_3/sep_1by2"
1227 |   type: "ReLU"
1228 |   bottom: "conv5_3/sep_1by2"
1229 |   top: "conv5_3/sep_1by2"
1230 | }
1231 | layer {
1232 |   name: "conv5_4/dw_1by2"
1233 |   type: "ConvolutionDepthwise"
1234 |   bottom: "conv5_3/sep_1by2"
1235 |   top: "conv5_4/dw_1by2"
1236 |   param {
1237 |     lr_mult: 1
1238 |     decay_mult: 1
1239 |   }
1240 |   convolution_param {
1241 |     num_output: 128
1242 |     bias_term: false
1243 |     pad: 7
1244 |     kernel_size: 3
1245 |     stride: 1
1246 |     engine: CAFFE
1247 |     dilation: 7
1248 |     weight_filler {
1249 |       type: "msra"
1250 |     }
1251 |   }
1252 | }
1253 | layer {
1254 |   name: "conv5_4/dw/bn_1by2"
1255 |   type: "BatchNorm"
1256 |   bottom: "conv5_4/dw_1by2"
1257 |   top: "conv5_4/dw_1by2"
1258 |   param {
1259 |     lr_mult: 0
1260 |     decay_mult: 0
1261 |   }
1262 |   param {
1263 |     lr_mult: 0
1264 |     decay_mult: 0
1265 |   }
1266 |   param {
1267 |     lr_mult: 0
1268 |     decay_mult: 0
1269 |   }
1270 | }
1271 | layer {
1272 |   name: "conv5_4/dw/scale_1by2"
1273 |   type: "Scale"
1274 |   bottom: "conv5_4/dw_1by2"
1275 |   top: "conv5_4/dw_1by2"
1276 |   scale_param {
1277 |     filler {
1278 |       value: 1
1279 |     }
1280 |     bias_term: true
1281 |     bias_filler {
1282 |       value: 0
1283 |     }
1284 |   }
1285 | }
1286 | layer {
1287 |   name: "relu5_4/dw_1by2"
1288 |   type: "ReLU"
1289 |   bottom: "conv5_4/dw_1by2"
1290 |   top: "conv5_4/dw_1by2"
1291 | }
1292 | layer {
1293 |   name: "conv5_4/sep_1by2"
1294 |   type: "Convolution"
1295 |   bottom: "conv5_4/dw_1by2"
1296 |   top: "conv5_4/sep_1by2"
1297 |   param {
1298 |     lr_mult: 1
1299 |     decay_mult: 1
1300 |   }
1301 |   convolution_param {
1302 |     num_output: 128
1303 |     bias_term: false
1304 |     pad: 0
1305 |     kernel_size: 1
1306 |     stride: 1
1307 |     engine: CAFFE
1308 |     weight_filler {
1309 |       type: "msra"
1310 |     }
1311 |   }
1312 | }
1313 | layer {
1314 |   name: "conv5_4/sep/bn_1by2"
1315 |   type: "BatchNorm"
1316 |   bottom: "conv5_4/sep_1by2"
1317 |   top: "conv5_4/sep_1by2"
1318 |   param {
1319 |     lr_mult: 0
1320 |     decay_mult: 0
1321 |   }
1322 |   param {
1323 |     lr_mult: 0
1324 |     decay_mult: 0
1325 |   }
1326 |   param {
1327 |     lr_mult: 0
1328 |     decay_mult: 0
1329 |   }
1330 | }
1331 | layer {
1332 |   name: "conv5_4/sep/scale_1by2"
1333 |   type: "Scale"
1334 |   bottom: "conv5_4/sep_1by2"
1335 |   top: "conv5_4/sep_1by2"
1336 |   scale_param {
1337 |     filler {
1338 |       value: 1
1339 |     }
1340 |     bias_term: true
1341 |     bias_filler {
1342 |       value: 0
1343 |     }
1344 |   }
1345 | }
1346 | layer {
1347 |   name: "relu5_4/sep_1by2"
1348 |   type: "ReLU"
1349 |   bottom: "conv5_4/sep_1by2"
1350 |   top: "conv5_4/sep_1by2"
1351 | }
1352 | layer {
1353 |   name: "conv5_5/dw_1by2"
1354 |   type: "ConvolutionDepthwise"
1355 |   bottom: "conv5_4/sep_1by2"
1356 |   top: "conv5_5/dw_1by2"
1357 |   param {
1358 |     lr_mult: 1
1359 |     decay_mult: 1
1360 |   }
1361 |   convolution_param {
1362 |     num_output: 128
1363 |     bias_term: false
1364 |     pad: 1
1365 |     kernel_size: 3
1366 |     engine: CAFFE
1367 |     stride: 1
1368 |     dilation: 1
1369 |     weight_filler {
1370 |       type: "msra"
1371 |     }
1372 |   }
1373 | }
1374 | layer {
1375 |   name: "conv5_5/dw/bn_1by2"
1376 |   type: "BatchNorm"
1377 |   bottom: "conv5_5/dw_1by2"
1378 |   top: "conv5_5/dw_1by2"
1379 |   param {
1380 |     lr_mult: 0
1381 |     decay_mult: 0
1382 |   }
1383 |   param {
1384 |     lr_mult: 0
1385 |     decay_mult: 0
1386 |   }
1387 |   param {
1388 |     lr_mult: 0
1389 |     decay_mult: 0
1390 |   }
1391 | }
1392 | layer {
1393 |   name: "conv5_5/dw/scale_1by2"
1394 |   type: "Scale"
1395 |   bottom: "conv5_5/dw_1by2"
1396 |   top: "conv5_5/dw_1by2"
1397 |   scale_param {
1398 |     filler {
1399 |       value: 1
1400 |     }
1401 |     bias_term: true
1402 |     bias_filler {
1403 |       value: 0
1404 |     }
1405 |   }
1406 | }
1407 | layer {
1408 |   name: "relu5_5/dw_1by2"
1409 |   type: "ReLU"
1410 |   bottom: "conv5_5/dw_1by2"
1411 |   top: "conv5_5/dw_1by2"
1412 | }
1413 | layer {
1414 |   name: "conv5_5/sep_1by2"
1415 |   type: "Convolution"
1416 |   bottom: "conv5_5/dw_1by2"
1417 |   top: "conv5_5/sep_1by2"
1418 |   param {
1419 |     lr_mult: 1
1420 |     decay_mult: 1
1421 |   }
1422 |   convolution_param {
1423 |     num_output: 128
1424 |     bias_term: false
1425 |     pad: 0
1426 |     engine: CAFFE
1427 |     kernel_size: 1
1428 |     stride: 1
1429 |     weight_filler {
1430 |       type: "msra"
1431 |     }
1432 |   }
1433 | }
1434 | layer {
1435 |   name: "conv5_5/sep/bn_1by2"
1436 |   type: "BatchNorm"
1437 |   bottom: "conv5_5/sep_1by2"
1438 |   top: "conv5_5/sep_1by2"
1439 |   param {
1440 |     lr_mult: 0
1441 |     decay_mult: 0
1442 |   }
1443 |   param {
1444 |     lr_mult: 0
1445 |     decay_mult: 0
1446 |   }
1447 |   param {
1448 |     lr_mult: 0
1449 |     decay_mult: 0
1450 |   }
1451 | }
1452 | layer {
1453 |   name: "conv5_5/sep/scale_1by2"
1454 |   type: "Scale"
1455 |   bottom: "conv5_5/sep_1by2"
1456 |   top: "conv5_5/sep_1by2"
1457 |   scale_param {
1458 |     filler {
1459 |       value: 1
1460 |     }
1461 |     bias_term: true
1462 |     bias_filler {
1463 |       value: 0
1464 |     }
1465 |   }
1466 | }
1467 | layer {
1468 |   name: "relu5_5/sep_1by2"
1469 |   type: "ReLU"
1470 |   bottom: "conv5_5/sep_1by2"
1471 |   top: "conv5_5/sep_1by2"
1472 | }
1473 | layer {
1474 |   name: "conv5_6/dw_1by2"
1475 |   type: "ConvolutionDepthwise"
1476 |   bottom: "conv5_5/sep_1by2"
1477 |   top: "conv5_6/dw_1by2"
1478 |   param {
1479 |     lr_mult: 1
1480 |     decay_mult: 1
1481 |   }
1482 |   convolution_param {
1483 |     num_output: 128
1484 |     bias_term: false
1485 |     pad: 2
1486 |     engine: CAFFE
1487 |     kernel_size: 3
1488 |     stride: 1
1489 |     dilation: 2
1490 |     weight_filler {
1491 |       type: "msra"
1492 |     }
1493 |   }
1494 | }
1495 | layer {
1496 |   name: "conv5_6/dw/bn_1by2"
1497 |   type: "BatchNorm"
1498 |   bottom: "conv5_6/dw_1by2"
1499 |   top: "conv5_6/dw_1by2"
1500 |   param {
1501 |     lr_mult: 0
1502 |     decay_mult: 0
1503 |   }
1504 |   param {
1505 |     lr_mult: 0
1506 |     decay_mult: 0
1507 |   }
1508 |   param {
1509 |     lr_mult: 0
1510 |     decay_mult: 0
1511 |   }
1512 | }
1513 | layer {
1514 |   name: "conv5_6/dw/scale_1by2"
1515 |   type: "Scale"
1516 |   bottom: "conv5_6/dw_1by2"
1517 |   top: "conv5_6/dw_1by2"
1518 |   scale_param {
1519 |     filler {
1520 |       value: 1
1521 |     }
1522 |     bias_term: true
1523 |     bias_filler {
1524 |       value: 0
1525 |     }
1526 |   }
1527 | }
1528 | layer {
1529 |   name: "relu5_6/dw_1by2"
1530 |   type: "ReLU"
1531 |   bottom: "conv5_6/dw_1by2"
1532 |   top: "conv5_6/dw_1by2"
1533 | }
1534 | layer {
1535 |   name: "conv5_6/sep_1by2"
1536 |   type: "Convolution"
1537 |   bottom: "conv5_6/dw_1by2"
1538 |   top: "conv5_6/sep_1by2"
1539 |   param {
1540 |     lr_mult: 1
1541 |     decay_mult: 1
1542 |   }
1543 |   convolution_param {
1544 |     num_output: 128
1545 |     bias_term: false
1546 |     pad: 0
1547 |     engine: CAFFE
1548 |     kernel_size: 1
1549 |     stride: 1
1550 |     weight_filler {
1551 |       type: "msra"
1552 |     }
1553 |   }
1554 | }
1555 | layer {
1556 |   name: "conv5_6/sep/bn_1by2"
1557 |   type: "BatchNorm"
1558 |   bottom: "conv5_6/sep_1by2"
1559 |   top: "conv5_6/sep_1by2"
1560 |   param {
1561 |     lr_mult: 0
1562 |     decay_mult: 0
1563 |   }
1564 |   param {
1565 |     lr_mult: 0
1566 |     decay_mult: 0
1567 |   }
1568 |   param {
1569 |     lr_mult: 0
1570 |     decay_mult: 0
1571 |   }
1572 | }
1573 | layer {
1574 |   name: "conv5_6/sep/scale_1by2"
1575 |   type: "Scale"
1576 |   bottom: "conv5_6/sep_1by2"
1577 |   top: "conv5_6/sep_1by2"
1578 |   scale_param {
1579 |     filler {
1580 |       value: 1
1581 |     }
1582 |     bias_term: true
1583 |     bias_filler {
1584 |       value: 0
1585 |     }
1586 |   }
1587 | }
1588 | layer {
1589 |   name: "relu5_6/sep_1by2"
1590 |   type: "ReLU"
1591 |   bottom: "conv5_6/sep_1by2"
1592 |   top: "conv5_6/sep_1by2"
1593 | }
1594 | layer {
1595 |   name: "conv6/dw_1by2"
1596 |   type: "ConvolutionDepthwise"
1597 |   bottom: "conv5_6/sep_1by2"
1598 |   top: "conv6/dw_1by2"
1599 |   param {
1600 |     lr_mult: 1
1601 |     decay_mult: 1
1602 |   }
1603 |   convolution_param {
1604 |     num_output: 128
1605 |     bias_term: false
1606 |     pad: 5
1607 |     kernel_size: 3
1608 |     stride: 1
1609 |     engine: CAFFE
1610 |     dilation: 5
1611 |     weight_filler {
1612 |       type: "msra"
1613 |     }
1614 |   }
1615 | }
1616 | layer {
1617 |   name: "conv6/dw/bn_1by2"
1618 |   type: "BatchNorm"
1619 |   bottom: "conv6/dw_1by2"
1620 |   top: "conv6/dw_1by2"
1621 |   param {
1622 |     lr_mult: 0
1623 |     decay_mult: 0
1624 |   }
1625 |   param {
1626 |     lr_mult: 0
1627 |     decay_mult: 0
1628 |   }
1629 |   param {
1630 |     lr_mult: 0
1631 |     decay_mult: 0
1632 |   }
1633 | }
1634 | layer {
1635 |   name: "conv6/dw/scale_1by2"
1636 |   type: "Scale"
1637 |   bottom: "conv6/dw_1by2"
1638 |   top: "conv6/dw_1by2"
1639 |   scale_param {
1640 |     filler {
1641 |       value: 1
1642 |     }
1643 |     bias_term: true
1644 |     bias_filler {
1645 |       value: 0
1646 |     }
1647 |   }
1648 | }
1649 | layer {
1650 |   name: "relu6/dw_1by2"
1651 |   type: "ReLU"
1652 |   bottom: "conv6/dw_1by2"
1653 |   top: "conv6/dw_1by2"
1654 | }
1655 | layer {
1656 |   name: "conv6/sep_1by2"
1657 |   type: "Convolution"
1658 |   bottom: "conv6/dw_1by2"
1659 |   top: "conv6/sep_1by2"
1660 |   param {
1661 |     lr_mult: 1
1662 |     decay_mult: 1
1663 |   }
1664 |   convolution_param {
1665 |     num_output: 128
1666 |     bias_term: false
1667 |     pad: 0
1668 |     kernel_size: 1
1669 |     stride: 1
1670 |     engine: CAFFE
1671 |     weight_filler {
1672 |       type: "msra"
1673 |     }
1674 |   }
1675 | }
1676 | layer {
1677 |   name: "conv6/sep/bn_1by2"
1678 |   type: "BatchNorm"
1679 |   bottom: "conv6/sep_1by2"
1680 |   top: "conv6/sep_1by2"
1681 |   param {
1682 |     lr_mult: 0
1683 |     decay_mult: 0
1684 |   }
1685 |   param {
1686 |     lr_mult: 0
1687 |     decay_mult: 0
1688 |   }
1689 |   param {
1690 |     lr_mult: 0
1691 |     decay_mult: 0
1692 |   }
1693 | }
1694 | layer {
1695 |   name: "conv6/sep/scale_1by2"
1696 |   type: "Scale"
1697 |   bottom: "conv6/sep_1by2"
1698 |   top: "conv6/sep_1by2"
1699 |   scale_param {
1700 |     filler {
1701 |       value: 1
1702 |     }
1703 |     bias_term: true
1704 |     bias_filler {
1705 |       value: 0
1706 |     }
1707 |   }
1708 | }
1709 | layer {
1710 |   name: "relu6/sep_1by2"
1711 |   type: "ReLU"
1712 |   bottom: "conv6/sep_1by2"
1713 |   top: "conv6/sep_1by2"
1714 | }
1715 | layer {
1716 |   name: "concat1_1by2"
1717 |   type: "Concat"
1718 |   bottom: "conv6/sep_1by2"
1719 |   bottom: "conv5_5/sep_1by2"
1720 |   top: "concat1_1by2"
1721 |   concat_param {
1722 |     concat_dim:1
1723 |   }
1724 | }
1725 | 
1726 | 
1727 | layer {
1728 |   name: "concat1_1by2/dw"
1729 |   type: "ConvolutionDepthwise"
1730 |   bottom: "concat1_1by2"
1731 |   top: "concat1_1by2/dw"
1732 |   param {
1733 |     lr_mult: 1
1734 |     decay_mult: 0
1735 |   }
1736 |   convolution_param {
1737 |     num_output: 256
1738 |     bias_term: false
1739 |     pad: 7
1740 |     kernel_size: 3
1741 |     stride: 1
1742 |     dilation: 7
1743 |     engine: CAFFE
1744 |     weight_filler {
1745 |       type: "msra"
1746 |     }
1747 |   }
1748 | }
1749 | layer {
1750 |   name: "concat1_1by2/dw/bn"
1751 |   type: "BatchNorm"
1752 |   bottom: "concat1_1by2/dw"
1753 |   top: "concat1_1by2/dw"
1754 |   param {
1755 |     lr_mult: 0
1756 |     decay_mult: 0
1757 |   }
1758 |   param {
1759 |     lr_mult: 0
1760 |     decay_mult: 0
1761 |   }
1762 |   param {
1763 |     lr_mult: 0
1764 |     decay_mult: 0
1765 |   }
1766 | }
1767 | layer {
1768 |   name: "concat1_1by2/dw/scale"
1769 |   type: "Scale"
1770 |   bottom: "concat1_1by2/dw"
1771 |   top: "concat1_1by2/dw"
1772 |   scale_param {
1773 |     filler {
1774 |       value: 1
1775 |     }
1776 |     bias_term: true
1777 |     bias_filler {
1778 |       value: 0
1779 |     }
1780 |   }
1781 | }
1782 | layer {
1783 |   name: "reluconcat1_1by2/dw"
1784 |   type: "ReLU"
1785 |   bottom: "concat1_1by2/dw"
1786 |   top: "concat1_1by2/dw"
1787 | }
1788 | layer {
1789 |   name: "concat1_1by2/sep"
1790 |   type: "Convolution"
1791 |   bottom: "concat1_1by2/dw"
1792 |   top: "concat1_1by2/sep"
1793 |   param {
1794 |     lr_mult: 1
1795 |     decay_mult: 1
1796 |   }
1797 |   convolution_param {
1798 |     num_output: 128
1799 |     bias_term: false
1800 |     pad: 0
1801 |     kernel_size: 1
1802 |     stride: 1
1803 |     engine: CAFFE
1804 |     weight_filler {
1805 |       type: "msra"
1806 |     }
1807 |   }
1808 | }
1809 | layer {
1810 |   name: "concat1_1by2/sep/bn"
1811 |   type: "BatchNorm"
1812 |   bottom: "concat1_1by2/sep"
1813 |   top: "concat1_1by2/sep"
1814 |   param {
1815 |     lr_mult: 0
1816 |     decay_mult: 0
1817 |   }
1818 |   param {
1819 |     lr_mult: 0
1820 |     decay_mult: 0
1821 |   }
1822 |   param {
1823 |     lr_mult: 0
1824 |     decay_mult: 0
1825 |   }
1826 | }
1827 | layer {
1828 |   name: "concat1_1by2/sep/scale"
1829 |   type: "Scale"
1830 |   bottom: "concat1_1by2/sep"
1831 |   top: "concat1_1by2/sep"
1832 |   scale_param {
1833 |     filler {
1834 |       value: 1
1835 |     }
1836 |     bias_term: true
1837 |     bias_filler {
1838 |       value: 0
1839 |     }
1840 |   }
1841 | }
1842 | layer {
1843 |   name: "reluconcat1_1by2/sep"
1844 |   type: "ReLU"
1845 |   bottom: "concat1_1by2/sep"
1846 |   top: "concat1_1by2/sep"
1847 | }
1848 | layer {
1849 |   name: "up2_1by2"
1850 |   type: "Deconvolution"
1851 |   bottom: "concat1_1by2/sep"
1852 |   top: "up2_1by2"
1853 |   param {
1854 |     lr_mult: 0.1
1855 |     decay_mult: 1
1856 |   }
1857 |   convolution_param {
1858 |     weight_filler: { type: "bilinear" }
1859 |     num_output: 128
1860 |     group: 128
1861 |     bias_term: false
1862 |     kernel_size: 4
1863 |     stride: 2
1864 |     pad: 1
1865 |   }
1866 | }
1867 | layer {
1868 |   name: "concat2_1by2"
1869 |   type: "Concat"
1870 |   bottom: "up2_1by2"
1871 |   bottom: "conv4_1/sep_1by2"
1872 |   top: "concat2_1by2"
1873 |   concat_param {
1874 |     concat_dim:1
1875 |   }
1876 | }
1877 | 
1878 | 
1879 | layer {
1880 |   name: "concat2_1by2/dw"
1881 |   type: "ConvolutionDepthwise"
1882 |   bottom: "concat2_1by2"
1883 |   top: "concat2_1by2/dw"
1884 |   param {
1885 |     lr_mult: 1
1886 |     decay_mult: 0
1887 |   }
1888 |   convolution_param {
1889 |     num_output: 256
1890 |     bias_term: false
1891 |     pad: 1
1892 |     kernel_size: 3
1893 |     stride: 1
1894 |     dilation: 1
1895 |     engine: CAFFE
1896 |     weight_filler {
1897 |       type: "msra"
1898 |     }
1899 |   }
1900 | }
1901 | layer {
1902 |   name: "concat2_1by2/dw/bn"
1903 |   type: "BatchNorm"
1904 |   bottom: "concat2_1by2/dw"
1905 |   top: "concat2_1by2/dw"
1906 |   param {
1907 |     lr_mult: 0
1908 |     decay_mult: 0
1909 |   }
1910 |   param {
1911 |     lr_mult: 0
1912 |     decay_mult: 0
1913 |   }
1914 |   param {
1915 |     lr_mult: 0
1916 |     decay_mult: 0
1917 |   }
1918 | }
1919 | layer {
1920 |   name: "concat2_1by2/dw/scale"
1921 |   type: "Scale"
1922 |   bottom: "concat2_1by2/dw"
1923 |   top: "concat2_1by2/dw"
1924 |   scale_param {
1925 |     filler {
1926 |       value: 1
1927 |     }
1928 |     bias_term: true
1929 |     bias_filler {
1930 |       value: 0
1931 |     }
1932 |   }
1933 | }
1934 | layer {
1935 |   name: "reluconcat2_1by2/dw"
1936 |   type: "ReLU"
1937 |   bottom: "concat2_1by2/dw"
1938 |   top: "concat2_1by2/dw"
1939 | }
1940 | layer {
1941 |   name: "concat2_1by2/sep"
1942 |   type: "Convolution"
1943 |   bottom: "concat2_1by2/dw"
1944 |   top: "concat2_1by2/sep"
1945 |   param {
1946 |     lr_mult: 1
1947 |     decay_mult: 1
1948 |   }
1949 |   convolution_param {
1950 |     num_output: 128
1951 |     bias_term: false
1952 |     pad: 0
1953 |     kernel_size: 1
1954 |     stride: 1
1955 |     engine: CAFFE
1956 |     weight_filler {
1957 |       type: "msra"
1958 |     }
1959 |   }
1960 | }
1961 | layer {
1962 |   name: "concat2_1by2/sep/bn"
1963 |   type: "BatchNorm"
1964 |   bottom: "concat2_1by2/sep"
1965 |   top: "concat2_1by2/sep"
1966 |   param {
1967 |     lr_mult: 0
1968 |     decay_mult: 0
1969 |   }
1970 |   param {
1971 |     lr_mult: 0
1972 |     decay_mult: 0
1973 |   }
1974 |   param {
1975 |     lr_mult: 0
1976 |     decay_mult: 0
1977 |   }
1978 | }
1979 | layer {
1980 |   name: "concat2_1by2/sep/scale"
1981 |   type: "Scale"
1982 |   bottom: "concat2_1by2/sep"
1983 |   top: "concat2_1by2/sep"
1984 |   scale_param {
1985 |     filler {
1986 |       value: 1
1987 |     }
1988 |     bias_term: true
1989 |     bias_filler {
1990 |       value: 0
1991 |     }
1992 |   }
1993 | }
1994 | layer {
1995 |   name: "reluconcat2_1by2/sep"
1996 |   type: "ReLU"
1997 |   bottom: "concat2_1by2/sep"
1998 |   top: "concat2_1by2/sep"
1999 | }
2000 | layer {
2001 |   name: "up3_1by2"
2002 |   type: "Deconvolution"
2003 |   bottom: "concat2_1by2/sep"
2004 |   top: "up3_1by2"
2005 |   param {
2006 |     lr_mult: 0.1
2007 |     decay_mult: 1
2008 |   }
2009 |   convolution_param {
2010 |     weight_filler: { type: "bilinear" }
2011 |     num_output: 64
2012 |     group: 64
2013 |     bias_term: false
2014 |     kernel_size: 4
2015 |     stride: 2
2016 |     pad: 1
2017 |   }
2018 | }
2019 | layer {
2020 |   name: "concat3_1by2"
2021 |   type: "Concat"
2022 |   bottom: "up3_1by2"
2023 |   bottom: "conv3_1/sep_1by2"
2024 |   top: "concat3_1by2"
2025 |   concat_param {
2026 |     concat_dim:1
2027 |   }
2028 | }
2029 | layer {
2030 |   name: "concat3_1by2/dw"
2031 |   type: "ConvolutionDepthwise"
2032 |   bottom: "concat3_1by2"
2033 |   top: "concat3_1by2/dw"
2034 |   param {
2035 |     lr_mult: 1
2036 |     decay_mult: 0
2037 |   }
2038 |   convolution_param {
2039 |     num_output: 128
2040 |     bias_term: false
2041 |     pad: 2
2042 |     kernel_size: 3
2043 |     stride: 1
2044 |     dilation: 2
2045 |     engine: CAFFE
2046 |     weight_filler {
2047 |       type: "msra"
2048 |     }
2049 |   }
2050 | }
2051 | layer {
2052 |   name: "concat3_1by2/dw/bn"
2053 |   type: "BatchNorm"
2054 |   bottom: "concat3_1by2/dw"
2055 |   top: "concat3_1by2/dw"
2056 |   param {
2057 |     lr_mult: 0
2058 |     decay_mult: 0
2059 |   }
2060 |   param {
2061 |     lr_mult: 0
2062 |     decay_mult: 0
2063 |   }
2064 |   param {
2065 |     lr_mult: 0
2066 |     decay_mult: 0
2067 |   }
2068 | }
2069 | layer {
2070 |   name: "concat3_1by2/dw/scale"
2071 |   type: "Scale"
2072 |   bottom: "concat3_1by2/dw"
2073 |   top: "concat3_1by2/dw"
2074 |   scale_param {
2075 |     filler {
2076 |       value: 1
2077 |     }
2078 |     bias_term: true
2079 |     bias_filler {
2080 |       value: 0
2081 |     }
2082 |   }
2083 | }
2084 | layer {
2085 |   name: "reluconcat3_1by2/dw"
2086 |   type: "ReLU"
2087 |   bottom: "concat3_1by2/dw"
2088 |   top: "concat3_1by2/dw"
2089 | }
2090 | layer {
2091 |   name: "concat3_1by2/sep"
2092 |   type: "Convolution"
2093 |   bottom: "concat3_1by2/dw"
2094 |   top: "concat3_1by2/sep"
2095 |   param {
2096 |     lr_mult: 1
2097 |     decay_mult: 1
2098 |   }
2099 |   convolution_param {
2100 |     num_output: 64
2101 |     bias_term: false
2102 |     pad: 0
2103 |     kernel_size: 1
2104 |     stride: 1
2105 |     engine: CAFFE
2106 |     weight_filler {
2107 |       type: "msra"
2108 |     }
2109 |   }
2110 | }
2111 | layer {
2112 |   name: "concat3_1by2/sep/bn"
2113 |   type: "BatchNorm"
2114 |   bottom: "concat3_1by2/sep"
2115 |   top: "concat3_1by2/sep"
2116 |   param {
2117 |     lr_mult: 0
2118 |     decay_mult: 0
2119 |   }
2120 |   param {
2121 |     lr_mult: 0
2122 |     decay_mult: 0
2123 |   }
2124 |   param {
2125 |     lr_mult: 0
2126 |     decay_mult: 0
2127 |   }
2128 | }
2129 | layer {
2130 |   name: "concat3_1by2/sep/scale"
2131 |   type: "Scale"
2132 |   bottom: "concat3_1by2/sep"
2133 |   top: "concat3_1by2/sep"
2134 |   scale_param {
2135 |     filler {
2136 |       value: 1
2137 |     }
2138 |     bias_term: true
2139 |     bias_filler {
2140 |       value: 0
2141 |     }
2142 |   }
2143 | }
2144 | layer {
2145 |   name: "reluconcat3_1by2/sep"
2146 |   type: "ReLU"
2147 |   bottom: "concat3_1by2/sep"
2148 |   top: "concat3_1by2/sep"
2149 | }
2150 | layer {
2151 |   name: "up4_1by2"
2152 |   type: "Deconvolution"
2153 |   bottom: "concat3_1by2/sep"
2154 |   top: "up4_1by2"
2155 |   param {
2156 |     lr_mult: 0.1
2157 |     decay_mult: 1
2158 |   }
2159 |   convolution_param {
2160 |     weight_filler: { type: "bilinear" }
2161 |     num_output: 32
2162 |     group: 32
2163 |     bias_term: false
2164 |     kernel_size: 4
2165 |     stride: 2
2166 |     pad: 1
2167 |   }
2168 | }
2169 | layer {
2170 |   name: "concat4_1by2"
2171 |   type: "Concat"
2172 |   bottom: "up4_1by2"
2173 |   bottom: "conv2_1/sep_1by2"
2174 |   top: "concat4_1by2"
2175 |   concat_param {
2176 |     concat_dim:1
2177 |   }
2178 | }
2179 | layer {
2180 |   name: "concat4_1by2/dw"
2181 |   type: "ConvolutionDepthwise"
2182 |   bottom: "concat4_1by2"
2183 |   top: "concat4_1by2/dw"
2184 |   param {
2185 |     lr_mult: 1
2186 |     decay_mult: 0
2187 |   }
2188 |   convolution_param {
2189 |     num_output: 64
2190 |     bias_term: false
2191 |     pad: 5
2192 |     kernel_size: 3
2193 |     stride: 1
2194 |     dilation: 5
2195 |     engine: CAFFE
2196 |     weight_filler {
2197 |       type: "msra"
2198 |     }
2199 |   }
2200 | }
2201 | layer {
2202 |   name: "concat4_1by2/dw/bn"
2203 |   type: "BatchNorm"
2204 |   bottom: "concat4_1by2/dw"
2205 |   top: "concat4_1by2/dw"
2206 |   param {
2207 |     lr_mult: 0
2208 |     decay_mult: 0
2209 |   }
2210 |   param {
2211 |     lr_mult: 0
2212 |     decay_mult: 0
2213 |   }
2214 |   param {
2215 |     lr_mult: 0
2216 |     decay_mult: 0
2217 |   }
2218 | }
2219 | layer {
2220 |   name: "concat4_1by2/dw/scale"
2221 |   type: "Scale"
2222 |   bottom: "concat4_1by2/dw"
2223 |   top: "concat4_1by2/dw"
2224 |   scale_param {
2225 |     filler {
2226 |       value: 1
2227 |     }
2228 |     bias_term: true
2229 |     bias_filler {
2230 |       value: 0
2231 |     }
2232 |   }
2233 | }
2234 | layer {
2235 |   name: "reluconcat4_1by2/dw"
2236 |   type: "ReLU"
2237 |   bottom: "concat4_1by2/dw"
2238 |   top: "concat4_1by2/dw"
2239 | }
2240 | layer {
2241 |   name: "concat4_1by2/sep"
2242 |   type: "Convolution"
2243 |   bottom: "concat4_1by2/dw"
2244 |   top: "concat4_1by2/sep"
2245 |   param {
2246 |     lr_mult: 1
2247 |     decay_mult: 1
2248 |   }
2249 |   convolution_param {
2250 |     num_output: 32
2251 |     bias_term: false
2252 |     pad: 0
2253 |     kernel_size: 1
2254 |     stride: 1
2255 |     engine: CAFFE
2256 |     weight_filler {
2257 |       type: "msra"
2258 |     }
2259 |   }
2260 | }
2261 | layer {
2262 |   name: "concat4_1by2/sep/bn"
2263 |   type: "BatchNorm"
2264 |   bottom: "concat4_1by2/sep"
2265 |   top: "concat4_1by2/sep"
2266 |   param {
2267 |     lr_mult: 0
2268 |     decay_mult: 0
2269 |   }
2270 |   param {
2271 |     lr_mult: 0
2272 |     decay_mult: 0
2273 |   }
2274 |   param {
2275 |     lr_mult: 0
2276 |     decay_mult: 0
2277 |   }
2278 | }
2279 | layer {
2280 |   name: "concat4_1by2/sep/scale"
2281 |   type: "Scale"
2282 |   bottom: "concat4_1by2/sep"
2283 |   top: "concat4_1by2/sep"
2284 |   scale_param {
2285 |     filler {
2286 |       value: 1
2287 |     }
2288 |     bias_term: true
2289 |     bias_filler {
2290 |       value: 0
2291 |     }
2292 |   }
2293 | }
2294 | layer {
2295 |   name: "reluconcat4_1by2/sep"
2296 |   type: "ReLU"
2297 |   bottom: "concat4_1by2/sep"
2298 |   top: "concat4_1by2/sep"
2299 | }
2300 | layer {
2301 |   name: "concat5_1by2"
2302 |   type: "Concat"
2303 |   bottom: "concat4_1by2/sep"
2304 |   bottom: "conv1_1by2"
2305 |   top: "concat5_1by2"
2306 |   concat_param {
2307 |     concat_dim:1
2308 |   }
2309 | }
2310 | layer {
2311 |   name: "conv_concat5_1by2"
2312 |   type: "Convolution"
2313 |   bottom: "concat5_1by2"
2314 |   top: "conv_concat5_1by2"
2315 |   param {
2316 |     lr_mult: 1
2317 |     decay_mult: 1
2318 |   }
2319 |   convolution_param {
2320 |     num_output: 16
2321 |     bias_term: false
2322 |     pad: 7
2323 |     kernel_size: 3
2324 |     stride: 1
2325 |     dilation: 7
2326 |     engine: CAFFE
2327 |     weight_filler {
2328 |       type: "msra"
2329 |     }
2330 |   }
2331 | }
2332 | layer {
2333 |   name: "conv_concat5_1by2/bn"
2334 |   type: "BatchNorm"
2335 |   bottom: "conv_concat5_1by2"
2336 |   top: "conv_concat5_1by2"
2337 |   param {
2338 |     lr_mult: 0
2339 |     decay_mult: 0
2340 |   }
2341 |   param {
2342 |     lr_mult: 0
2343 |     decay_mult: 0
2344 |   }
2345 |   param {
2346 |     lr_mult: 0
2347 |     decay_mult: 0
2348 |   }
2349 | }
2350 | layer {
2351 |   name: "conv_concat5_1by2/scale"
2352 |   type: "Scale"
2353 |   bottom: "conv_concat5_1by2"
2354 |   top: "conv_concat5_1by2"
2355 |   scale_param {
2356 |     filler {
2357 |       value: 1
2358 |     }
2359 |     bias_term: true
2360 |     bias_filler {
2361 |       value: 0
2362 |     }
2363 |   }
2364 | }
2365 | layer {
2366 |   name: "reluconv_concat5_1by2"
2367 |   type: "ReLU"
2368 |   bottom: "conv_concat5_1by2"
2369 |   top: "conv_concat5_1by2"
2370 | }
2371 | layer {
2372 |   name: "score_fr_1by2"
2373 |   type: "Convolution"
2374 |   bottom: "conv_concat5_1by2"
2375 |   top: "score_fr_1by2"
2376 |   param {
2377 |     lr_mult: 1
2378 |     decay_mult: 1
2379 |   }
2380 |   param {
2381 |     lr_mult: 2
2382 |     decay_mult: 0
2383 |   }
2384 |   convolution_param {
2385 |     num_output: 1
2386 |     engine: CAFFE
2387 |     pad: 0
2388 |     kernel_size: 1
2389 |     weight_filler {
2390 |       type: "msra"
2391 |     }
2392 |     bias_filler {
2393 |       type: "constant"
2394 |       value: 0
2395 |     }
2396 |   }
2397 | }
2398 | layer {
2399 |   name: "up5_1by2"
2400 |   type: "Deconvolution"
2401 |   bottom: "score_fr_1by2"
2402 |   top: "up5_1by2"
2403 |   param {
2404 |     lr_mult: 0.1
2405 |     decay_mult: 1
2406 |   }
2407 |   convolution_param {
2408 |     weight_filler: { type: "bilinear" }
2409 |     num_output: 1
2410 |     group: 1
2411 |     bias_term: false
2412 |     kernel_size: 4
2413 |     stride: 2
2414 |     pad: 1
2415 |   }
2416 | }
2417 | layer {
2418 |   name: "sigmoid_1by2"
2419 |   type: "Sigmoid"
2420 |   bottom: "up5_1by2"
2421 |   top: "up5_1by2"
2422 | }
2423 | layer {
2424 |   name: "loss_1by2"
2425 |   type: "DiceCoefLoss"
2426 |   bottom: "up5_1by2"
2427 |   bottom: "label"
2428 |   top: "loss_1by2"
2429 |   loss_weight: 1
2430 |   include: { phase: TRAIN }
2431 | }
2432 | layer {
2433 |   name: "accuracy_1by2"
2434 |   type: "Accuracy"
2435 |   bottom: "up5_1by2"
2436 |   bottom: "label"
2437 |   top: "IOU_test_1by2"
2438 |   include {
2439 |     phase: TEST
2440 |   }
2441 | }
2442 | 
2443 | #### auxi loss 16 ####
2444 | 
2445 | layer {
2446 |   name: "seg_scale1by2"
2447 |   type: "Resize"
2448 |   bottom: "label"
2449 |   top: "seg_scale1by2"
2450 |   resize_param {
2451 |     function_type: FACTOR_GIVEN
2452 |     intepolation_type: NEAREST
2453 |     resize_factor: 0.5
2454 |   }
2455 |   include: { phase: TRAIN }
2456 | }
2457 | 
2458 | layer {
2459 |   name: "concat4_1by2_scale1by2/1x1"
2460 |   type: "Convolution"
2461 |   bottom: "concat4_1by2"
2462 |   top: "concat4_1by2_scale1by2/1x1"
2463 |   param { lr_mult: 1.0 decay_mult: 1.0}
2464 |   param { lr_mult: 2.0 decay_mult: 0 }
2465 |   convolution_param {
2466 |     engine: CAFFE
2467 |     num_output: 64
2468 |     kernel_size: 1 pad: 0 stride: 1
2469 |     weight_filler { type: "msra"  } 
2470 |     bias_filler { type: "constant" value: 0 }
2471 |   }
2472 |   include: { phase: TRAIN }
2473 | }
2474 | layer {
2475 |   name: "concat4_1by2_scale1by2/3x3"
2476 |   type: "Convolution"
2477 |   bottom: "concat4_1by2_scale1by2/1x1"
2478 |   top: "concat4_1by2_scale1by2/3x3"
2479 |   param {
2480 |     lr_mult: 1
2481 |     decay_mult: 1
2482 |   }
2483 |   param {
2484 |     lr_mult: 2
2485 |     decay_mult: 0
2486 |   }
2487 |   convolution_param {
2488 |     engine: CAFFE
2489 |     num_output: 1
2490 |     pad: 1
2491 |     kernel_size: 3
2492 |     weight_filler {
2493 |       type: "msra"
2494 |     }
2495 |     bias_filler {
2496 |       type: "constant"
2497 |       value: 0
2498 |     }
2499 |   }
2500 |   include: { phase: TRAIN }
2501 | }
2502 | layer {
2503 |   name: "loss_scale1by2"
2504 |   type: "DiceCoefLoss"
2505 |   bottom: "concat4_1by2_scale1by2/3x3"
2506 |   bottom: "seg_scale1by2"
2507 |   top: "loss_scale1by2"
2508 |   loss_weight: 0.6
2509 |   include: { phase: TRAIN }
2510 | }
2511 | 
2512 | 
2513 | #### auxi loss 8 ####
2514 | 
2515 | layer {
2516 |   name: "seg_scale1by4"
2517 |   type: "Resize"
2518 |   bottom: "label"
2519 |   top: "seg_scale1by4"
2520 |   resize_param {
2521 |     function_type: FACTOR_GIVEN
2522 |     intepolation_type: NEAREST
2523 |     resize_factor: 0.25
2524 |   }
2525 |   include: { phase: TRAIN }
2526 | }
2527 | 
2528 | layer {
2529 |   name: "concat4_1by2_scale1by4/1x1"
2530 |   type: "Convolution"
2531 |   bottom: "concat3_1by2"
2532 |   top: "concat4_1by2_scale1by4/1x1"
2533 |   param { lr_mult: 1.0 decay_mult: 1.0}
2534 |   param { lr_mult: 2.0 decay_mult: 0 }
2535 |   convolution_param {
2536 |     engine: CAFFE
2537 |     num_output: 64
2538 |     kernel_size: 1 pad: 0 stride: 1
2539 |     weight_filler { type: "msra"  } 
2540 |     bias_filler { type: "constant" value: 0 }
2541 |   }
2542 |   include: { phase: TRAIN }
2543 | }
2544 | layer {
2545 |   name: "concat4_1by2_scale1by4/3x3"
2546 |   type: "Convolution"
2547 |   bottom: "concat4_1by2_scale1by4/1x1"
2548 |   top: "concat4_1by2_scale1by4/3x3"
2549 |   param {
2550 |     lr_mult: 1
2551 |     decay_mult: 1
2552 |   }
2553 |   param {
2554 |     lr_mult: 2
2555 |     decay_mult: 0
2556 |   }
2557 |   convolution_param {
2558 |     engine: CAFFE
2559 |     num_output: 1
2560 |     pad: 1
2561 |     kernel_size: 3
2562 |     weight_filler {
2563 |       type: "msra"
2564 |     }
2565 |     bias_filler {
2566 |       type: "constant"
2567 |       value: 0
2568 |     }
2569 |   }
2570 |   include: { phase: TRAIN }
2571 | }
2572 | layer {
2573 |   name: "loss_scale1by4"
2574 |   type: "DiceCoefLoss"
2575 |   bottom: "concat4_1by2_scale1by4/3x3"
2576 |   bottom: "seg_scale1by4"
2577 |   top: "loss_scale1by4"
2578 |   loss_weight: 0.3
2579 |   include: { phase: TRAIN }
2580 | }
2581 | 
2582 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # MobileNetv2
 2 | 
 3 | Run the script benchmark_mobilenetv2.sh and you can get the result: top1/top5: 0.7123/0.9018
 4 | 
 5 | You can use this model to do a lot of things such as training a smaller mobilenetv2 (By moving params or knowledge distillation)
 6 | 
 7 | Small MobileNetV1/V2 is friendly used on a mobile device (U-net for semanic segmentation and refinedet for object detection)
 8 | 
 9 | Training details for ImageNet2012 :
10 |                                    type: "SGD"
11 |                                    lr_policy: "poly"
12 |                                    base_lr: 0.045
13 |                                    power: 1
14 |                                    momentum: 0.9
15 |                                    weight_decay: 0.00004
16 |  
17 |                                   
18 |  The pytorch version: https://github.com/suzhenghang/MobileNetV2_Pytorch
19 |  
20 |  
21 |  The prototxt of SmallMobileUnet: https://github.com/suzhenghang/MobileNetv2/blob/master/MobileUnet1by2.prototxt
22 |  
23 |  The prototxt of SmallRefineDet_mobilev2_ssdlite: https://github.com/suzhenghang/MobileNetv2/blob/master/SmallRefineDet_mobilev2_ssdlite.prototxt
24 | 


--------------------------------------------------------------------------------
/forMobileUNet/dice_coef_loss_layer.cpp:
--------------------------------------------------------------------------------
 1 | #include <vector>
 2 | 
 3 | #include "caffe/layers/dice_coef_loss_layer.hpp"
 4 | #include "caffe/util/math_functions.hpp"
 5 | 
 6 | namespace caffe {
 7 | 
 8 | template <typename Dtype>
 9 | void DiceCoefLossLayer<Dtype>::Reshape(
10 |   const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
11 |   LossLayer<Dtype>::Reshape(bottom, top);
12 |   CHECK_EQ(bottom[0]->count(1), bottom[1]->count(1))
13 |       << "Inputs must have the same dimension.";
14 |   const int batchsize = bottom[0]->num();
15 |   const int dim = bottom[0]->count(1);
16 |   
17 |   vector<int> multiplier_shape(1, dim);
18 |   vector<int> result_shape(1, batchsize);
19 |   result_.Reshape(result_shape);
20 |   result_tmp_.Reshape(result_shape);
21 |   multiplier_.Reshape(multiplier_shape);
22 |   tmp_.ReshapeLike(*bottom[0]);
23 |   smooth = Dtype(1.);
24 |   caffe_set(dim, Dtype(1), multiplier_.mutable_cpu_data());
25 |   caffe_set(batchsize, smooth, result_tmp_.mutable_cpu_data());
26 |   caffe_set(batchsize, smooth, result_.mutable_cpu_data());
27 | }
28 | 
29 | template <typename Dtype>
30 | void DiceCoefLossLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
31 |     const vector<Blob<Dtype>*>& top) {
32 |   caffe_mul(bottom[0]->count(), bottom[0]->cpu_data(), bottom[0]->cpu_data(), 
33 |                 tmp_.mutable_cpu_data());
34 |   caffe_cpu_gemv(CblasNoTrans, bottom[0]->num(), bottom[0]->count(1), Dtype(1.), tmp_.cpu_data(), 
35 |                     multiplier_.cpu_data(), Dtype(1.), result_tmp_.mutable_cpu_data());
36 |   caffe_mul(bottom[1]->count(), bottom[1]->cpu_data(), bottom[1]->cpu_data(), 
37 |                 tmp_.mutable_cpu_data());
38 |   caffe_cpu_gemv(CblasNoTrans, bottom[1]->num(), bottom[1]->count(1), Dtype(1.), tmp_.cpu_data(), 
39 |                     multiplier_.cpu_data(), Dtype(1.), result_tmp_.mutable_cpu_data());
40 |   caffe_mul(bottom[0]->count(), bottom[0]->cpu_data(), bottom[1]->cpu_data(), 
41 |                 tmp_.mutable_cpu_data());
42 |   caffe_cpu_gemv(CblasNoTrans, bottom[1]->num(), bottom[1]->count(1), Dtype(2.), tmp_.cpu_data(), 
43 |                     multiplier_.cpu_data(), Dtype(1.), result_.mutable_cpu_data());
44 |   caffe_div(bottom[0]->num(), result_.cpu_data(), result_tmp_.cpu_data(), result_.mutable_cpu_data());
45 |   
46 |   Dtype loss = Dtype(1) - caffe_cpu_asum(bottom[0]->num(), result_.cpu_data()) / bottom[0]->num();
47 |   top[0]->mutable_cpu_data()[0] = loss;
48 | }
49 | 
50 | template <typename Dtype>
51 | void DiceCoefLossLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
52 |     const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
53 |   for (int i = 0; i < 2; ++i) {
54 |     if (propagate_down[i]) {
55 |       const Dtype sign = Dtype(1.0);
56 |       const int index = (i == 0) ? 1 : 0;
57 |       caffe_copy(bottom[i]->count(), bottom[i]->cpu_data(), bottom[i]->mutable_cpu_diff());
58 |       for (int j = 0; j < bottom[i]->num(); j++) {
59 |         const Dtype alpha = sign * top[0]->cpu_diff()[0] / result_tmp_.cpu_data()[j];
60 |         // LOG(INFO) << top[0]->cpu_diff()[0];
61 |         caffe_cpu_axpby(
62 |           bottom[i]->count(1),              // count
63 |           alpha*Dtype(-2),                  // alpha
64 |           bottom[index]->cpu_data()+j*bottom[i]->count(1),        // a
65 |           alpha*result_.cpu_data()[j]*Dtype(2),                      // beta
66 |           bottom[i]->mutable_cpu_diff()+j*bottom[i]->count(1)
67 |         );  // b
68 |       } 
69 |     }
70 |   }
71 | }
72 | 
73 | #ifdef CPU_ONLY
74 | STUB_GPU(DiceCoefLossLayer);
75 | #endif
76 | 
77 | INSTANTIATE_CLASS(DiceCoefLossLayer);
78 | REGISTER_LAYER_CLASS(DiceCoefLoss);
79 | 
80 | }  // namespace caffe
81 | 


--------------------------------------------------------------------------------
/forMobileUNet/dice_coef_loss_layer.cu:
--------------------------------------------------------------------------------
 1 | #include <vector>
 2 | 
 3 | #include "caffe/layers/dice_coef_loss_layer.hpp"
 4 | #include "caffe/util/math_functions.hpp"
 5 | 
 6 | namespace caffe {
 7 | 
 8 | template <typename Dtype>
 9 | void DiceCoefLossLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
10 |     const vector<Blob<Dtype>*>& top) {
11 |   int count = bottom[0]->count();
12 |   
13 |   caffe_gpu_mul(count, bottom[0]->gpu_data(), bottom[0]->gpu_data(), 
14 |                 tmp_.mutable_gpu_data());
15 |   
16 |   caffe_gpu_gemv(CblasNoTrans, bottom[0]->num(), bottom[0]->count(1), Dtype(1.), tmp_.gpu_data(), 
17 |                     multiplier_.gpu_data(), Dtype(1.), result_tmp_.mutable_gpu_data());
18 |   caffe_gpu_mul(count, bottom[1]->gpu_data(), bottom[1]->gpu_data(), 
19 |                 tmp_.mutable_gpu_data());
20 |   caffe_gpu_gemv(CblasNoTrans, bottom[1]->num(), bottom[1]->count(1), Dtype(1.), tmp_.gpu_data(), 
21 |                     multiplier_.gpu_data(), Dtype(1.), result_tmp_.mutable_gpu_data());
22 |   caffe_gpu_mul(count, bottom[0]->gpu_data(), bottom[1]->gpu_data(), 
23 |                 tmp_.mutable_gpu_data());
24 |   caffe_gpu_gemv(CblasNoTrans, bottom[1]->num(), bottom[1]->count(1), Dtype(2.), tmp_.gpu_data(), 
25 |                     multiplier_.gpu_data(), Dtype(1.), result_.mutable_gpu_data());
26 |   caffe_gpu_div(bottom[0]->num(), result_.gpu_data(), result_tmp_.gpu_data(), 
27 |                 result_.mutable_gpu_data());
28 |   Dtype loss;
29 |   caffe_gpu_asum(bottom[0]->num(), result_.gpu_data(), &loss);
30 |   loss /= bottom[0]->num();
31 |   loss = Dtype(1.) - loss;
32 |   top[0]->mutable_cpu_data()[0] = loss;
33 | }
34 | 
35 | template <typename Dtype>
36 | __global__ void DiceCoefLossBackward(const int n, const Dtype* x,
37 |     const Dtype* y, Dtype* out_diff, const Dtype sign,
38 |     const Dtype* loss, const Dtype* denominator, const int dim) {
39 |   CUDA_KERNEL_LOOP(index, n) {
40 |     const int i = index / dim;
41 |     const Dtype alpha = Dtype(-2.) / denominator[i] * sign;
42 |     const Dtype beta = Dtype(2.) * loss[i] / denominator[i] * sign;
43 |     out_diff[index] = alpha * x[index] + beta * y[index];
44 |   }
45 | }
46 | 
47 | template <typename Dtype>
48 | void DiceCoefLossLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
49 |     const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
50 |   for (int i = 0; i < 2; ++i) {
51 |     if (propagate_down[i]) {
52 |       int count = bottom[i]->count();
53 |       const Dtype sign = Dtype(1.0)*top[0]->cpu_diff()[0];
54 |       const int index = (i == 0) ? 1 : 0;
55 |       DiceCoefLossBackward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(
56 |         count, bottom[index]->gpu_data(), bottom[i]->gpu_data(), bottom[i]->mutable_gpu_diff(), 
57 |         sign, result_.gpu_data(), result_tmp_.gpu_data(), bottom[i]->count(1));
58 |       CUDA_POST_KERNEL_CHECK;
59 |     }
60 |   }
61 | }
62 | 
63 | INSTANTIATE_LAYER_GPU_FUNCS(DiceCoefLossLayer);
64 | 
65 | }  // namespace caffe
66 | 


--------------------------------------------------------------------------------
/forMobileUNet/dice_coef_loss_layer.hpp:
--------------------------------------------------------------------------------
  1 | #ifndef CAFFE_DICE_COEF_LOSS_LAYER_HPP_
  2 | #define CAFFE_DICE_COEF_LOSS_LAYER_HPP_
  3 | 
  4 | #include <vector>
  5 | 
  6 | #include "caffe/blob.hpp"
  7 | #include "caffe/layer.hpp"
  8 | #include "caffe/proto/caffe.pb.h"
  9 | 
 10 | #include "caffe/layers/loss_layer.hpp"
 11 | 
 12 | namespace caffe {
 13 | 
 14 | /**
 15 |  * @brief Computes the Euclidean (L2) loss @f$
 16 |  *          E = \frac{1}{2N} \sum\limits_{n=1}^N \left| \left| \hat{y}_n - y_n
 17 |  *        \right| \right|_2^2 @f$ for real-valued regression tasks.
 18 |  *
 19 |  * @param bottom input Blob vector (length 2)
 20 |  *   -# @f$ (N \times C \times H \times W) @f$
 21 |  *      the predictions @f$ \hat{y} \in [-\infty, +\infty]@f$
 22 |  *   -# @f$ (N \times C \times H \times W) @f$
 23 |  *      the targets @f$ y \in [-\infty, +\infty]@f$
 24 |  * @param top output Blob vector (length 1)
 25 |  *   -# @f$ (1 \times 1 \times 1 \times 1) @f$
 26 |  *      the computed Euclidean loss: @f$ E =
 27 |  *          \frac{1}{2n} \sum\limits_{n=1}^N \left| \left| \hat{y}_n - y_n
 28 |  *        \right| \right|_2^2 @f$
 29 |  *
 30 |  * This can be used for least-squares regression tasks.  An InnerProductLayer
 31 |  * input to a DiceCoefLossLayer exactly formulates a linear least squares
 32 |  * regression problem. With non-zero weight decay the problem becomes one of
 33 |  * ridge regression -- see src/caffe/test/test_sgd_solver.cpp for a concrete
 34 |  * example wherein we check that the gradients computed for a Net with exactly
 35 |  * this structure match hand-computed gradient formulas for ridge regression.
 36 |  *
 37 |  * (Note: Caffe, and SGD in general, is certainly \b not the best way to solve
 38 |  * linear least squares problems! We use it only as an instructive example.)
 39 |  */
 40 | template <typename Dtype>
 41 | class DiceCoefLossLayer : public LossLayer<Dtype> {
 42 |  public:
 43 |   explicit DiceCoefLossLayer(const LayerParameter& param)
 44 |       : LossLayer<Dtype>(param), multiplier_() ,result_() ,result_tmp_() ,tmp_(){}
 45 |   virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
 46 |       const vector<Blob<Dtype>*>& top);
 47 | 
 48 |   virtual inline const char* type() const { return "DiceCoefLoss"; }
 49 |   /**
 50 |    * Unlike most loss layers, in the DiceCoefLossLayer we can backpropagate
 51 |    * to both inputs -- override to return true and always allow force_backward.
 52 |    */
 53 |   virtual inline bool AllowForceBackward(const int bottom_index) const {
 54 |     return true;
 55 |   }
 56 | 
 57 |  protected:
 58 |   /// @copydoc DiceCoefLossLayer
 59 |   virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
 60 |       const vector<Blob<Dtype>*>& top);
 61 |   virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
 62 |       const vector<Blob<Dtype>*>& top);
 63 | 
 64 |   /**
 65 |    * @brief Computes the Euclidean error gradient w.r.t. the inputs.
 66 |    *
 67 |    * Unlike other children of LossLayer, DiceCoefLossLayer \b can compute
 68 |    * gradients with respect to the label inputs bottom[1] (but still only will
 69 |    * if propagate_down[1] is set, due to being produced by learnable parameters
 70 |    * or if force_backward is set). In fact, this layer is "commutative" -- the
 71 |    * result is the same regardless of the order of the two bottoms.
 72 |    *
 73 |    * @param top output Blob vector (length 1), providing the error gradient with
 74 |    *      respect to the outputs
 75 |    *   -# @f$ (1 \times 1 \times 1 \times 1) @f$
 76 |    *      This Blob's diff will simply contain the loss_weight* @f$ \lambda @f$,
 77 |    *      as @f$ \lambda @f$ is the coefficient of this layer's output
 78 |    *      @f$\ell_i@f$ in the overall Net loss
 79 |    *      @f$ E = \lambda_i \ell_i + \mbox{other loss terms}@f$; hence
 80 |    *      @f$ \frac{\partial E}{\partial \ell_i} = \lambda_i @f$.
 81 |    *      (*Assuming that this top Blob is not used as a bottom (input) by any
 82 |    *      other layer of the Net.)
 83 |    * @param propagate_down see Layer::Backward.
 84 |    * @param bottom input Blob vector (length 2)
 85 |    *   -# @f$ (N \times C \times H \times W) @f$
 86 |    *      the predictions @f$\hat{y}@f$; Backward fills their diff with
 87 |    *      gradients @f$
 88 |    *        \frac{\partial E}{\partial \hat{y}} =
 89 |    *            \frac{1}{n} \sum\limits_{n=1}^N (\hat{y}_n - y_n)
 90 |    *      @f$ if propagate_down[0]
 91 |    *   -# @f$ (N \times C \times H \times W) @f$
 92 |    *      the targets @f$y@f$; Backward fills their diff with gradients
 93 |    *      @f$ \frac{\partial E}{\partial y} =
 94 |    *          \frac{1}{n} \sum\limits_{n=1}^N (y_n - \hat{y}_n)
 95 |    *      @f$ if propagate_down[1]
 96 |    */
 97 |   virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
 98 |       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
 99 |   virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
100 |       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
101 | 
102 |   Blob<Dtype> multiplier_;
103 |   Blob<Dtype> result_;
104 |   Blob<Dtype> result_tmp_;
105 |   Blob<Dtype> tmp_;
106 |   
107 |   Dtype smooth;
108 | };
109 | 
110 | }  // namespace caffe
111 | 
112 | #endif  // CAFFE_EUCLIDEAN_LOSS_LAYER_HPP_
113 | 


--------------------------------------------------------------------------------
/forMobileUNet/hello.txt:
--------------------------------------------------------------------------------
1 | 
2 | 


--------------------------------------------------------------------------------
/forMobileUNet/resize_layer.cpp:
--------------------------------------------------------------------------------
  1 | /*
  2 | * Integrate imgResize and blobAlign python layers together.
  3 | *
  4 | * Author: Wei Zhen @ IIE, CAS
  5 | * Last modified: 2017-06-11
  6 | */
  7 | 
  8 | #include <algorithm>
  9 | #include <vector>
 10 | 
 11 | #include "caffe/layer.hpp"
 12 | #include "caffe/util/math_functions.hpp"
 13 | #include "caffe/layers/resize_layer.hpp"
 14 | 
 15 | #define MAX(a, b) ((a) > (b) ? (a) : (b))
 16 | #define MIN(a, b) ((a) < (b) ? (a) : (b))
 17 | 
 18 | namespace caffe {
 19 | 
 20 | template <typename Dtype>
 21 | void ResizeLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
 22 |       const vector<Blob<Dtype>*>& top) {
 23 |   // do nothing
 24 | }
 25 | 
 26 | template <typename Dtype>
 27 | void ResizeLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
 28 |       const vector<Blob<Dtype>*>& top) {
 29 |   // parse params and get resize factor
 30 |   this->input_size_ = bottom[0]->height();
 31 | 
 32 |   SegResizeParameter resize_param = this->layer_param_.resize_param();
 33 |   switch (resize_param.function_type()) {
 34 | 	case SegResizeParameter_FunctionType_SIZE_GIVEN:
 35 | 		if (resize_param.output_size() <= 0) {
 36 | 		    LOG(INFO)<< "Illegal output size (" << resize_param.output_size() << "), use default resize factor 1.";
 37 | 		}
 38 | 		else {
 39 | 		    this->output_size_ = resize_param.output_size();
 40 | 		    this->resize_factor_ = (float)this->output_size_ / this->input_size_;
 41 | 		}
 42 | 		break;
 43 | 	case SegResizeParameter_FunctionType_BLOB_ALIGN:
 44 | 		CHECK_EQ(bottom.size(), 2) << "When using BLOB_ALIGN option, bottom[1] must be given.";
 45 | 		this->output_size_ = bottom[1]->height();
 46 | 		this->resize_factor_ = (float)this->output_size_ / this->input_size_;
 47 | 		break;
 48 | 	case SegResizeParameter_FunctionType_FACTOR_GIVEN:
 49 | 		CHECK_GT(resize_param.resize_factor(), 0) << "Illegal resize factor (" << resize_param.resize_factor() << ").";
 50 | 		this->resize_factor_ = resize_param.resize_factor();
 51 | 		this->output_size_ = static_cast<int>(this->input_size_ * this->resize_factor_);
 52 | 		break;
 53 | 	default:
 54 | 		this->resize_factor_ = 1;
 55 | 		this->output_size_ = this->input_size_;
 56 |   }
 57 | 
 58 |   // reshape top (assume that all features are square)
 59 |   top[0]->Reshape(bottom[0]->num(), bottom[0]->channels(), this->output_size_, this->output_size_);
 60 | }
 61 | 
 62 | template <typename Dtype>
 63 | void ResizeLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
 64 |     const vector<Blob<Dtype>*>& top) {
 65 | 
 66 |   const Dtype* bottom_data = bottom[0]->cpu_data();
 67 |   Dtype* top_data = top[0]->mutable_cpu_data();
 68 | 
 69 |   // perform interpolation based on different interpolation types
 70 |   switch (this->layer_param_.resize_param().intepolation_type()) {
 71 | 	case SegResizeParameter_InterpolationType_NEAREST:
 72 | 		for (int n = 0; n < bottom[0]->num(); ++n) {
 73 | 		    for (int c = 0; c < bottom[0]->channels(); ++c) {
 74 | 			for (int rh = 0; rh < this->output_size_; ++rh) {
 75 | 			    for (int rw = 0; rw < this->output_size_; ++rw) {
 76 | 				int h = int(rh / this->resize_factor_);
 77 | 				int w = int(rw / this->resize_factor_);
 78 | 				h = std::min(h, this->input_size_);
 79 | 				w = std::min(w, this->input_size_);
 80 | 				top_data[rh * this->output_size_ + rw] = bottom_data[h * this->input_size_ + w];
 81 | 			    }
 82 | 			}
 83 | 			// compute offset
 84 | 			bottom_data += bottom[0]->offset(0, 1);
 85 | 			top_data += top[0]->offset(0, 1);
 86 | 		    }
 87 | 		}
 88 | 		break;
 89 | 	case SegResizeParameter_InterpolationType_BILINEAR:
 90 | 		for (int n = 0; n < bottom[0]->num(); ++n) {
 91 | 		    for (int c = 0; c < bottom[0]->channels(); ++c) {
 92 | 			for (int rh = 0; rh < this->output_size_; ++rh) {
 93 | 			    for (int rw = 0; rw < this->output_size_; ++rw) {
 94 | 				float h = rh / this->resize_factor_;
 95 | 				float w = rw / this->resize_factor_;
 96 | 				//h = std::min(h, this->input_size_);
 97 | 				//w = std::min(w, this->input_size_);
 98 | 				int top_idx = rh * this->output_size_ + rw;
 99 | 				top_data[top_idx] = 0;		// DO NOT forget to reset before accumulation
100 | 				for (int n = MAX(static_cast<int>(h-1) + 1, 0); n < MIN(h + 1, this->input_size_); n++) {
101 | 				    for (int m = MAX(static_cast<int>(w-1) + 1, 0); m < MIN(w + 1, this->input_size_); m++) {
102 | 					top_data[top_idx] += bottom_data[n * this->input_size_ + m] * (1 - std::abs(w-m)) * (1 - std::abs(h-n));
103 | 				    }
104 | 				}
105 | 			    }
106 | 			}
107 | 			// compute offset
108 | 			bottom_data += bottom[0]->offset(0, 1);
109 | 			top_data += top[0]->offset(0, 1);
110 | 		    }
111 | 		}
112 | 		break;
113 | 	default:
114 | 		LOG(FATAL) << "Unknown interpolation type.";
115 |   }
116 | }
117 | 
118 | template <typename Dtype>
119 | void ResizeLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
120 |     const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
121 |   if (!propagate_down[0]) {    return;  }
122 | 
123 |   const Dtype* top_diff = top[0]->cpu_diff();
124 |   Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
125 |   caffe_set(bottom[0]->count(), Dtype(0), bottom_diff);
126 | 
127 |   // data gradient
128 |   switch (this->layer_param_.resize_param().intepolation_type()) {
129 | 	case SegResizeParameter_InterpolationType_NEAREST:
130 | 		for (int n = 0; n < top[0]->num(); ++n) {
131 | 		    for (int c = 0; c < top[0]->channels(); ++c) {
132 | 			for (int h = 0; h < this->input_size_; ++h) {
133 | 			    for (int w = 0; w < this->input_size_; ++w) {
134 | 				int rh = int(h * this->resize_factor_);
135 | 				int rw = int(w * this->resize_factor_);
136 | 				rh = std::min(rh, this->output_size_ - 1);
137 | 				rw = std::min(rw, this->output_size_ - 1);
138 | 				bottom_diff[h * this->input_size_ + w] = top_diff[rh * this->output_size_ + rw];
139 | 			    }
140 | 			}
141 | 			// offset
142 | 			bottom_diff += bottom[0]->offset(0, 1);
143 | 			top_diff += top[0]->offset(0, 1);
144 | 		    }
145 | 		}
146 | 		break;
147 | 	case SegResizeParameter_InterpolationType_BILINEAR:
148 | 		for (int n = 0; n < top[0]->num(); ++n) {
149 | 		    for (int c = 0; c < top[0]->channels(); ++c) {
150 | 			for (int rh = 0; rh < this->output_size_; ++rh) {
151 | 			    for (int rw = 0; rw < this->output_size_; ++rw) {
152 | 				float h = rh / this->resize_factor_;
153 | 				float w = rw / this->resize_factor_;
154 | 
155 | 				int top_idx = rh * this->output_size_ + rw;
156 | 				int h0 = std::max(int(h), 0);
157 | 				int w0 = std::max(int(w), 0);
158 | 				int h1 = 0;
159 | 				float weight_h_h1 = 0;
160 | 				float weight_w_w0 = std::max(1 - std::abs(w - w0), float (0));
161 | 				float weight_h_h0 = std::max(1 - std::abs(h - h0), float (0));
162 | 				if (h < this->input_size_-1) {				// h1 does not exceed input_size
163 | 					h1 = std::min(int(h)+1, this->input_size_ - 1);
164 | 					weight_h_h1 = std::max(1 - std::abs(h - h1), float (0));
165 | 					if (w0 == 0) {					// the first column
166 | 						bottom_diff[h1 * this->input_size_ + w0] += top_diff[top_idx] * weight_w_w0 * weight_h_h1 * 2;
167 | 					} else {
168 | 						bottom_diff[h1 * this->input_size_ + w0] += top_diff[top_idx] * weight_w_w0 * weight_h_h1;
169 | 					}
170 | 				}
171 | 				if (w < this->input_size_-1) {				// w1 does not exceed input_size
172 | 					int w1 = std::min(int(w)+1, this->input_size_ - 1);
173 | 					float weight_w_w1 = std::max(1 - std::abs(w - w1), float (0));
174 | 					if (h0 == 0) {					// the first row
175 | 						bottom_diff[h0 * this->input_size_ + w1] += top_diff[top_idx] * weight_w_w1 * weight_h_h0*2;
176 | 					} else {
177 | 						bottom_diff[h0 * this->input_size_ + w1] += top_diff[top_idx] * weight_w_w1 * weight_h_h0;
178 | 					}
179 | 					if (h < this->input_size_-1) {			// for the very left-bottom one
180 | 						bottom_diff[h1 * this->input_size_ + w1] += top_diff[top_idx] * weight_w_w1 * weight_h_h1;
181 | 					}
182 | 				}
183 | 
184 | 				bottom_diff[h0 * this->input_size_ + w0] += top_diff[top_idx] * weight_w_w0 * weight_h_h0;
185 | 			    }
186 | 			}
187 | 			// normalize
188 | 			caffe_scal(bottom[0]->count(), this->resize_factor_>1 ? 1/this->resize_factor_/this->resize_factor_ : 1, bottom_diff);
189 | 			// offset
190 | 			bottom_diff += bottom[0]->offset(0, 1);
191 | 			top_diff += top[0]->offset(0, 1);
192 | 		    }
193 | 		}
194 | 		/*for (int n = 0; n < top[0]->num(); ++n) {
195 | 		    for (int c = 0; c < top[0]->channels(); ++c) {
196 | 			for (int h = 0; h < this->input_size_; ++h) {
197 | 			    for (int w = 0; w < this->input_size_; ++w) {
198 | 				int hstart = int((h - 1) * this->resize_factor_) + 1;
199 | 				int wstart = int((w - 1) * this->resize_factor_) + 1;
200 | 				int hend = int((h + 1) * this->resize_factor_);
201 | 				int wend = int((w + 1) * this->resize_factor_);
202 | 				hstart = std::max(hstart, 0);
203 | 				wstart = std::max(wstart, 0);
204 | 				hend = std::min(hend, this->output_size_ - 1);
205 | 				wend = std::min(wend, this->output_size_ - 1);
206 | 				const int bottom_idx = h * this->input_size_ + w;
207 | 				for (int rh = hstart; rh < hend; ++rh) {
208 | 				    for (int rw = wstart; rw < wend; ++rw) {
209 | 					bottom_diff[bottom_idx] += top_diff[rh * this->output_size_ + rw] 
210 |                                           * (1 - std::abs((rw / this->resize_factor_) - w)) * (1 - std::abs((rh / this->resize_factor_) - h));
211 | 				    }
212 | 				}
213 | 				bottom_diff[bottom_idx] /= (hend-hstart) * (wend-wstart);
214 | 			    }
215 | 			}
216 | 			// offset
217 | 			bottom_diff += bottom[0]->offset(0, 1);
218 | 			top_diff += top[0]->offset(0, 1);
219 | 		    }
220 | 		}*/
221 | 						
222 | 		break;
223 | 	default:
224 | 		LOG(FATAL) << "Unknown interpolation type.";
225 |   }
226 | }
227 | 
228 | #ifdef CPU_ONLY
229 | STUB_GPU(ResizeLayer);
230 | #endif
231 | 
232 | INSTANTIATE_CLASS(ResizeLayer);
233 | REGISTER_LAYER_CLASS(Resize);
234 | }  // namespace caffe
235 | 


--------------------------------------------------------------------------------
/forMobileUNet/resize_layer.cu:
--------------------------------------------------------------------------------
  1 | /*
  2 | * Integrate imgResize and blobAlign python layers together.
  3 | *
  4 | * Author: Wei Zhen @ IIE, CAS
  5 | * Last modified: 2017-06-11
  6 | */
  7 | 
  8 | #include <algorithm>
  9 | #include <vector>
 10 | 
 11 | #include "caffe/layer.hpp"
 12 | #include "caffe/util/math_functions.hpp"
 13 | #include "caffe/layers/resize_layer.hpp"
 14 | 
 15 | #define MAX(a, b) ((a) > (b) ? (a) : (b))
 16 | #define MIN(a, b) ((a) < (b) ? (a) : (b))
 17 | 
 18 | namespace caffe {
 19 | 
 20 | template <typename Dtype>
 21 | __global__ void nearestForwardGPU(const int nthreads,
 22 | 	const int input_channels, const int input_size, const int output_size, const float resize_factor,
 23 | 	const Dtype* bottom_data, Dtype* top_data) {
 24 | 
 25 | 	CUDA_KERNEL_LOOP(index, nthreads) {
 26 | 		// resume channel idx and pixel idx
 27 | 		int rw = index % output_size;
 28 | 		int rh = (index / output_size) % output_size;
 29 | 		int c = (index / output_size / output_size) % input_channels;
 30 | 		int n = index / output_size / output_size / input_channels;
 31 | 		// indexing and sampling
 32 | 		int h = int(rh / resize_factor);
 33 | 		int w = int(rw / resize_factor);
 34 | 		h = min(h, input_size);
 35 | 		w = min(w, input_size);
 36 | 		const Dtype* bottom_data_channel = bottom_data + (n * input_channels + c) * input_size * input_size;
 37 | 		top_data[index] = bottom_data_channel[h * input_size + w];
 38 | 	}
 39 | }
 40 | 
 41 | template <typename Dtype>
 42 | __global__  void bilinearForwardGPU(const int nthreads,
 43 | 	const int input_channels, const int input_size, const int output_size, const float resize_factor,
 44 | 	const Dtype *bottom_data, Dtype *top_data) {
 45 | 
 46 | 	CUDA_KERNEL_LOOP(index, nthreads) {
 47 | 		// resume channel idx and pixel idx
 48 | 		int rw = index % output_size;
 49 | 		int rh = (index / output_size) % output_size;
 50 | 		int c = (index / output_size / output_size) % input_channels;
 51 | 		int num = index / output_size / output_size / input_channels;
 52 | 		// indexing and sampling
 53 | 		float h = rh / resize_factor;
 54 | 		float w = rw / resize_factor;
 55 | 		//h = min(h, float(input_size));
 56 | 		//w = min(w, float(input_size));
 57 | 		const Dtype* bottom_data_channel = bottom_data + (num * input_channels + c) * input_size * input_size;
 58 | 		top_data[index] = 0;		// DO NOT forget to reset before accumulation
 59 | 		for (int n = MAX(floor(h-1) + 1, 0); n < MIN(h + 1, input_size); n++) {
 60 | 		    for (int m = MAX(floor(w-1) + 1, 0); m < MIN(w + 1, input_size); m++) {
 61 | 			top_data[index] += bottom_data_channel[n * input_size + m] * (1 - abs(w - m)) * (1 - abs(h - n));
 62 | 		    }
 63 | 		}
 64 | 	}
 65 | }
 66 | 
 67 | template <typename Dtype>
 68 | void ResizeLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
 69 |     const vector<Blob<Dtype>*>& top) {
 70 | 
 71 |   const Dtype* bottom_data = bottom[0]->gpu_data();
 72 |   Dtype* top_data = top[0]->mutable_gpu_data();
 73 |   int count = top[0]->count();
 74 | 
 75 |   // perform interpolation based on different interpolation types
 76 |   switch (this->layer_param_.resize_param().intepolation_type()) {
 77 | 	case SegResizeParameter_InterpolationType_NEAREST:
 78 | 		// parallel at pixel level
 79 | 		nearestForwardGPU<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>
 80 | 			(count, bottom[0]->channels(), this->input_size_, this->output_size_,
 81 | 			 this->resize_factor_, bottom_data, top_data);
 82 | 		break;
 83 | 	case SegResizeParameter_InterpolationType_BILINEAR:
 84 | 		// parallel at pixel level
 85 | 		bilinearForwardGPU<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>
 86 | 			(count, bottom[0]->channels(), this->input_size_, this->output_size_,
 87 | 			 this->resize_factor_, bottom_data, top_data);
 88 | 		break;
 89 | 	default:
 90 | 		LOG(FATAL) << "Unknown interpolation type.";
 91 |   }
 92 | }
 93 | 
 94 | template <typename Dtype>
 95 | __global__ void nearestBackwardGPU(const int nthreads,
 96 | 	const int input_channels, const int input_size, const int output_size, const float resize_factor,
 97 | 	Dtype* bottom_diff, const Dtype* top_diff) {
 98 | 
 99 | 	CUDA_KERNEL_LOOP(index, nthreads) {
100 | 		// resume channel idx and pixel idx
101 | 		int w = index % input_size;
102 | 		int h = (index / input_size) % input_size;
103 | 		int c = (index / input_size / input_size) % input_channels;
104 | 		int n = index / input_size / input_size / input_channels;
105 | 		// indexing and sampling
106 | 		int rh = int(h * resize_factor);
107 | 		int rw = int(w * resize_factor);
108 | 		rh = min(rh, output_size - 1);
109 | 		rw = min(rw, output_size - 1);
110 | 		top_diff += (n*input_channels + c) * output_size * output_size;
111 | 		bottom_diff[index] = top_diff[rh * output_size + rw];
112 | 	}
113 | }
114 | 
115 | template <typename Dtype>
116 | __global__ void bilinearBackwardGPU(const int nthreads,
117 | 	const int input_channels, const int input_size, const int output_size, const float resize_factor,
118 | 	Dtype* bottom_diff, const Dtype* top_diff) {
119 | 
120 | 	/*CUDA_KERNEL_LOOP(index, nthreads) {
121 | 		// resume channel idx
122 | 		int c = index % input_channels;
123 | 		int n = index / input_channels;
124 | 
125 | 		// indexing and sampling
126 | 		for (int rh = 0; rh < output_size; ++rh){
127 | 		    for (int rw = 0; rw < output_size; ++rw){
128 | 			float h = rh / resize_factor;
129 | 			float w = rw / resize_factor;
130 | 			int top_idx = ((n*input_channels + c) * output_size + rh ) * output_size + rw;
131 | 			Dtype* bottom_diff_channel = bottom_diff + (n*input_channels + c) * input_size * input_size;
132 | 
133 | 			int h0 = max(int(h), 0);
134 | 			int w0 = max(int(w), 0);
135 | 			int h1 = 0;
136 | 			float weight_h_h1 = 0;
137 | 
138 | 			float weight_w_w0 = max(1 - abs(w - w0), 0.);
139 | 			float weight_h_h0 = max(1 - abs(h - h0), 0.);
140 | 			if (h < input_size-1) {				// h1 does not exceed input_size
141 | 				h1 = min(int(h)+1, input_size - 1);
142 | 				weight_h_h1 = max(1 - std::abs(h - h1), float (0));
143 | 				if (w0 == 0) {				// the first column
144 | 					bottom_diff_channel[h1 * input_size + w0] += top_diff[top_idx] * weight_w_w0 * weight_h_h1 * 2;
145 | 				} else {
146 | 					bottom_diff_channel[h1 * input_size + w0] += top_diff[top_idx] * weight_w_w0 * weight_h_h1;
147 | 				}
148 | 			}
149 | 			if (w < input_size-1) {				// w1 does not exceed input_size
150 | 				int w1 = min(int(w)+1, input_size - 1);
151 | 				float weight_w_w1 = max(1 - std::abs(w - w1), float (0));
152 | 				if (h0 == 0) {				// the first row
153 | 					bottom_diff_channel[h0 * input_size + w1] += top_diff[top_idx] * weight_w_w1 * weight_h_h0 * 2;
154 | 				} else {
155 | 					bottom_diff_channel[h0 * input_size + w1] += top_diff[top_idx] * weight_w_w1 * weight_h_h0;
156 | 				}
157 | 				if (h < input_size-1) {			// for the very left-bottom one
158 | 					bottom_diff_channel[h1 * input_size + w1] += top_diff[top_idx] * weight_w_w1 * weight_h_h1;
159 | 				}
160 | 			}
161 | 		    }
162 | 		}
163 | 		// the normalization is outside this function
164 | 	}*/
165 | 	CUDA_KERNEL_LOOP(index, nthreads) {
166 | 		// resume channel idx and pixel idx
167 | 		int w = index % input_size;
168 | 		int h = (index / input_size) % input_size;
169 | 		int c = (index / input_size / input_size) % input_channels;
170 | 		int n = index / input_size / input_size / input_channels;
171 | 		// indexing and sampling
172 | 		float hstart = (h - 1) * resize_factor + 1;
173 | 		float wstart = (w - 1) * resize_factor + 1;
174 | 		float hend = (h + 1) * resize_factor;
175 | 		float wend = (w + 1) * resize_factor;
176 | 		hstart = max(hstart, 0.);
177 | 		wstart = max(wstart, 0.);
178 | 		hend = min(hend, float(output_size));
179 | 		wend = min(wend, float(output_size));
180 | 		const Dtype* top_diff_channel = top_diff + (n*input_channels + c) * output_size * output_size;
181 | 		bottom_diff[index] = 0;
182 | 		for (int rh = hstart; rh < hend; ++rh) {
183 | 		    for (int rw = wstart; rw < wend; ++rw) {
184 | 			bottom_diff[index] += top_diff_channel[rh * output_size + rw]
185 |                                           * (1 - abs((rw / resize_factor) - w)) * (1 - abs((rh / resize_factor) - h));
186 | 		    }
187 | 		}
188 | 		bottom_diff[index] /= resize_factor*resize_factor;
189 | 	}
190 | }
191 | 
192 | template <typename Dtype>
193 | void ResizeLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
194 |     const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
195 |   if (!propagate_down[0]) {    return;  }
196 | 
197 |   const Dtype* top_diff = top[0]->gpu_diff();
198 |   Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
199 |   const int bottom_count = bottom[0]->count();
200 |   //caffe_gpu_set(bottom[0]->count(), Dtype(0.), bottom_diff);
201 |   //const int top_count = top[0]->count();
202 |   int nthreads = top[0]->num() * top[0]->channels();
203 | 
204 |   // data gradient
205 |   switch (this->layer_param_.resize_param().intepolation_type()) {
206 | 	case SegResizeParameter_InterpolationType_NEAREST:
207 | 		// parallel at pixel level
208 | 		nearestBackwardGPU<Dtype><<<CAFFE_GET_BLOCKS(bottom_count), CAFFE_CUDA_NUM_THREADS>>>
209 | 			(bottom_count, bottom[0]->channels(), this->input_size_, this->output_size_,
210 | 			 this->resize_factor_, bottom_diff, top_diff);
211 | 		break;
212 | 	case SegResizeParameter_InterpolationType_BILINEAR:
213 | 		// parallel at channel level
214 | 		bilinearBackwardGPU<Dtype><<<CAFFE_GET_BLOCKS(nthreads), CAFFE_CUDA_NUM_THREADS>>>
215 | 			(bottom_count, bottom[0]->channels(), this->input_size_, this->output_size_,
216 | 			 this->resize_factor_, bottom_diff, top_diff);
217 | 		// normalize
218 | 		//caffe_gpu_scal(bottom_count, this->resize_factor_>1 ? 1/this->resize_factor_/this->resize_factor_ : 1, bottom_diff);
219 | 		break;
220 | 	default:
221 | 		LOG(FATAL) << "Unknown interpolation type.";
222 |   }
223 | 
224 | }
225 | 
226 | INSTANTIATE_LAYER_GPU_FUNCS(ResizeLayer);
227 | 
228 | }// namespace caffe
229 | 


--------------------------------------------------------------------------------
/forMobileUNet/resize_layer.hpp:
--------------------------------------------------------------------------------
 1 | /*
 2 | * Integrate imgResize and blobAlign python layers together.
 3 | *
 4 | * Author: Wei Zhen @ IIE, CAS
 5 | * Last modified: 2017-06-11
 6 | */
 7 | 
 8 | #ifndef CAFFE_RESIZE_LAYER_HPP_
 9 | #define CAFFE_RESIZE_LAYER_HPP_
10 | 
11 | #include <vector>
12 | 
13 | #include "caffe/blob.hpp"
14 | #include "caffe/layer.hpp"
15 | #include "caffe/proto/caffe.pb.h"
16 | 
17 | namespace caffe {
18 | 
19 | template <typename Dtype>
20 | class ResizeLayer : public Layer<Dtype> {
21 |   public:
22 |   explicit ResizeLayer(const LayerParameter& param)
23 |       : Layer<Dtype>(param) {}
24 |   virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
25 |       const vector<Blob<Dtype>*>& top);
26 |   virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
27 |       const vector<Blob<Dtype>*>& top);
28 | 
29 |   virtual inline const char* type() const { return "Resize"; }
30 |   virtual inline int MinBottomBlobs() const { return 1; }
31 |   virtual inline int MaxBottomBlobs() const { return 2; }
32 |   virtual inline int ExactNumTopBlobs() const { return 1; }
33 | 
34 |  protected:
35 |   virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
36 |       const vector<Blob<Dtype>*>& top);
37 |   virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
38 |       const vector<Blob<Dtype>*>& top);
39 |   virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
40 |       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
41 |   virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
42 |       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
43 | 
44 |   Dtype resize_factor_;
45 |   int input_size_;
46 |   int output_size_;
47 | 
48 | };
49 | 
50 | 
51 | }  // namespace caffe
52 | 
53 | #endif  // CAFFE_RESIZE_LAYER_HPP_
54 | 


--------------------------------------------------------------------------------