├── IRG.prototxt
├── IRG_Transformation.prototxt
├── LICENSE
├── README.md
├── ResNet-20.prototxt
├── batch_euclidean_map_layer.cpp
├── batch_euclidean_map_layer.cu
├── batch_euclidean_map_layer.hpp
├── batch_euclidean_vector_layer.cpp
├── batch_euclidean_vector_layer.cu
├── batch_euclidean_vector_layer.hpp
└── img
    ├── Exp_results.png
    └── framework.png


/IRG.prototxt:
--------------------------------------------------------------------------------
   1 | name: "resnet_cifar10_IRG"
   2 | layer {
   3 |   name: "Data1"
   4 |   type: "Data"
   5 |   top: "Data1"
   6 |   top: "Data2"
   7 |   include {
   8 |     phase: TRAIN
   9 |   }
  10 |   transform_param {
  11 |     mean_file: "yourpath/mean_pad.binaryproto"
  12 |     crop_size: 32
  13 |     mirror:true
  14 |   }
  15 |   data_param {
  16 |     source: "yourpath/cifar-10-train_lmdb"
  17 |     batch_size: 64 
  18 |     backend: LMDB
  19 |   }
  20 | }
  21 | layer {
  22 |   name: "Data1"
  23 |   type: "Data"
  24 |   top: "Data1"
  25 |   top: "Data2"
  26 |   include {
  27 |     phase: TEST
  28 |   }
  29 |   transform_param {
  30 |     mean_file: "yourpath/mean.binaryproto"
  31 |   }
  32 |   data_param {
  33 |     source: "yourpath/cifar10_test_lmdb"
  34 |     batch_size: 100
  35 |     backend: LMDB
  36 |   }
  37 | }
  38 | layer {
  39 |   name: "Convolution1"
  40 |   type: "Convolution"
  41 |   bottom: "Data1"
  42 |   top: "Convolution1"
  43 |   param {
  44 |     lr_mult: 0
  45 |     decay_mult: 0
  46 |   }
  47 |   param {
  48 |     lr_mult: 0
  49 |     decay_mult: 0
  50 |   }
  51 |   convolution_param {
  52 |     num_output: 16
  53 |     pad: 1
  54 |     kernel_size: 3
  55 |     stride: 1
  56 |     weight_filler {
  57 |       type: "msra"
  58 |     }
  59 |     bias_filler {
  60 |       type: "constant"
  61 |       value: 0
  62 |     }
  63 |   }
  64 | }
  65 | layer {
  66 |   name: "BatchNorm1"
  67 |   type: "BatchNorm"
  68 |   bottom: "Convolution1"
  69 |   top: "Convolution1"
  70 |   param {
  71 |     lr_mult: 0
  72 |     decay_mult: 0
  73 |   }
  74 |   param {
  75 |     lr_mult: 0
  76 |     decay_mult: 0
  77 |   }
  78 |   param {
  79 |     lr_mult: 0
  80 |     decay_mult: 0
  81 |   }
  82 |   batch_norm_param {
  83 |     use_global_stats: true
  84 |   }
  85 | }
  86 | layer {
  87 |   name: "Scale1"
  88 |   type: "Scale"
  89 |   bottom: "Convolution1"
  90 |   top: "Convolution1"
  91 |   param {
  92 |     lr_mult: 0
  93 |     decay_mult: 0
  94 |   }
  95 |   param {
  96 |     lr_mult: 0
  97 |     decay_mult: 0
  98 |   }
  99 |   scale_param {
 100 |     bias_term: true
 101 |   }
 102 | }
 103 | layer {
 104 |   name: "ReLU1"
 105 |   type: "ReLU"
 106 |   bottom: "Convolution1"
 107 |   top: "Convolution1"
 108 | }
 109 | layer {
 110 |   name: "Convolution2"
 111 |   type: "Convolution"
 112 |   bottom: "Convolution1"
 113 |   top: "Convolution2"
 114 |   param {
 115 |     lr_mult: 0
 116 |     decay_mult: 0
 117 |   }
 118 |   param {
 119 |     lr_mult: 0
 120 |     decay_mult: 0
 121 |   }
 122 |   convolution_param {
 123 |     num_output: 16
 124 |     pad: 1
 125 |     kernel_size: 3
 126 |     stride: 1
 127 |     weight_filler {
 128 |       type: "msra"
 129 |     }
 130 |     bias_filler {
 131 |       type: "constant"
 132 |       value: 0
 133 |     }
 134 |   }
 135 | }
 136 | layer {
 137 |   name: "BatchNorm2"
 138 |   type: "BatchNorm"
 139 |   bottom: "Convolution2"
 140 |   top: "Convolution2"
 141 |   param {
 142 |     lr_mult: 0
 143 |     decay_mult: 0
 144 |   }
 145 |   param {
 146 |     lr_mult: 0
 147 |     decay_mult: 0
 148 |   }
 149 |   param {
 150 |     lr_mult: 0
 151 |     decay_mult: 0
 152 |   }
 153 |   batch_norm_param {
 154 |     use_global_stats: true
 155 |   }
 156 | }
 157 | layer {
 158 |   name: "Scale2"
 159 |   type: "Scale"
 160 |   bottom: "Convolution2"
 161 |   top: "Convolution2"
 162 |   param {
 163 |     lr_mult: 0
 164 |     decay_mult: 0
 165 |   }
 166 |   param {
 167 |     lr_mult: 0
 168 |     decay_mult: 0
 169 |   }
 170 |   scale_param {
 171 |     bias_term: true
 172 |   }
 173 | }
 174 | layer {
 175 |   name: "ReLU2"
 176 |   type: "ReLU"
 177 |   bottom: "Convolution2"
 178 |   top: "Convolution2"
 179 | }
 180 | layer {
 181 |   name: "Convolution3"
 182 |   type: "Convolution"
 183 |   bottom: "Convolution2"
 184 |   top: "Convolution3"
 185 |   param {
 186 |     lr_mult: 0
 187 |     decay_mult: 0
 188 |   }
 189 |   param {
 190 |     lr_mult: 0
 191 |     decay_mult: 0
 192 |   }
 193 |   convolution_param {
 194 |     num_output: 16
 195 |     pad: 1
 196 |     kernel_size: 3
 197 |     stride: 1
 198 |     weight_filler {
 199 |       type: "msra"
 200 |     }
 201 |     bias_filler {
 202 |       type: "constant"
 203 |       value: 0
 204 |     }
 205 |   }
 206 | }
 207 | layer {
 208 |   name: "BatchNorm3"
 209 |   type: "BatchNorm"
 210 |   bottom: "Convolution3"
 211 |   top: "Convolution3"
 212 |   param {
 213 |     lr_mult: 0
 214 |     decay_mult: 0
 215 |   }
 216 |   param {
 217 |     lr_mult: 0
 218 |     decay_mult: 0
 219 |   }
 220 |   param {
 221 |     lr_mult: 0
 222 |     decay_mult: 0
 223 |   }
 224 |   batch_norm_param {
 225 |     use_global_stats: true
 226 |   }
 227 | }
 228 | layer {
 229 |   name: "Scale3"
 230 |   type: "Scale"
 231 |   bottom: "Convolution3"
 232 |   top: "Convolution3"
 233 |   param {
 234 |     lr_mult: 0
 235 |     decay_mult: 0
 236 |   }
 237 |   param {
 238 |     lr_mult: 0
 239 |     decay_mult: 0
 240 |   }
 241 |   scale_param {
 242 |     bias_term: true
 243 |   }
 244 | }
 245 | layer {
 246 |   name: "Eltwise1"
 247 |   type: "Eltwise"
 248 |   bottom: "Convolution1"
 249 |   bottom: "Convolution3"
 250 |   top: "Eltwise1"
 251 |   eltwise_param {
 252 |     operation: SUM
 253 |   }
 254 | }
 255 | layer {
 256 |   name: "ReLU3"
 257 |   type: "ReLU"
 258 |   bottom: "Eltwise1"
 259 |   top: "Eltwise1"
 260 | }
 261 | layer {
 262 |   name: "Convolution4"
 263 |   type: "Convolution"
 264 |   bottom: "Eltwise1"
 265 |   top: "Convolution4"
 266 |   param {
 267 |     lr_mult: 0
 268 |     decay_mult: 0
 269 |   }
 270 |   param {
 271 |     lr_mult: 0
 272 |     decay_mult: 0
 273 |   }
 274 |   convolution_param {
 275 |     num_output: 16
 276 |     pad: 1
 277 |     kernel_size: 3
 278 |     stride: 1
 279 |     weight_filler {
 280 |       type: "msra"
 281 |     }
 282 |     bias_filler {
 283 |       type: "constant"
 284 |       value: 0
 285 |     }
 286 |   }
 287 | }
 288 | layer {
 289 |   name: "BatchNorm4"
 290 |   type: "BatchNorm"
 291 |   bottom: "Convolution4"
 292 |   top: "Convolution4"
 293 |   param {
 294 |     lr_mult: 0
 295 |     decay_mult: 0
 296 |   }
 297 |   param {
 298 |     lr_mult: 0
 299 |     decay_mult: 0
 300 |   }
 301 |   param {
 302 |     lr_mult: 0
 303 |     decay_mult: 0
 304 |   }
 305 |   batch_norm_param {
 306 |     use_global_stats: true
 307 |   }
 308 | }
 309 | layer {
 310 |   name: "Scale4"
 311 |   type: "Scale"
 312 |   bottom: "Convolution4"
 313 |   top: "Convolution4"
 314 |   param {
 315 |     lr_mult: 0
 316 |     decay_mult: 0
 317 |   }
 318 |   param {
 319 |     lr_mult: 0
 320 |     decay_mult: 0
 321 |   }
 322 |   scale_param {
 323 |     bias_term: true
 324 |   }
 325 | }
 326 | layer {
 327 |   name: "ReLU4"
 328 |   type: "ReLU"
 329 |   bottom: "Convolution4"
 330 |   top: "Convolution4"
 331 | }
 332 | layer {
 333 |   name: "Convolution5"
 334 |   type: "Convolution"
 335 |   bottom: "Convolution4"
 336 |   top: "Convolution5"
 337 |   param {
 338 |     lr_mult: 0
 339 |     decay_mult: 0
 340 |   }
 341 |   param {
 342 |     lr_mult: 0
 343 |     decay_mult: 0
 344 |   }
 345 |   convolution_param {
 346 |     num_output: 16
 347 |     pad: 1
 348 |     kernel_size: 3
 349 |     stride: 1
 350 |     weight_filler {
 351 |       type: "msra"
 352 |     }
 353 |     bias_filler {
 354 |       type: "constant"
 355 |       value: 0
 356 |     }
 357 |   }
 358 | }
 359 | layer {
 360 |   name: "BatchNorm5"
 361 |   type: "BatchNorm"
 362 |   bottom: "Convolution5"
 363 |   top: "Convolution5"
 364 |   param {
 365 |     lr_mult: 0
 366 |     decay_mult: 0
 367 |   }
 368 |   param {
 369 |     lr_mult: 0
 370 |     decay_mult: 0
 371 |   }
 372 |   param {
 373 |     lr_mult: 0
 374 |     decay_mult: 0
 375 |   }
 376 |   batch_norm_param {
 377 |     use_global_stats: true
 378 |   }
 379 | }
 380 | layer {
 381 |   name: "Scale5"
 382 |   type: "Scale"
 383 |   bottom: "Convolution5"
 384 |   top: "Convolution5"
 385 |   param {
 386 |     lr_mult: 0
 387 |     decay_mult: 0
 388 |   }
 389 |   param {
 390 |     lr_mult: 0
 391 |     decay_mult: 0
 392 |   }
 393 |   scale_param {
 394 |     bias_term: true
 395 |   }
 396 | }
 397 | layer {
 398 |   name: "Eltwise2"
 399 |   type: "Eltwise"
 400 |   bottom: "Eltwise1"
 401 |   bottom: "Convolution5"
 402 |   top: "Eltwise2"
 403 |   eltwise_param {
 404 |     operation: SUM
 405 |   }
 406 | }
 407 | layer {
 408 |   name: "ReLU5"
 409 |   type: "ReLU"
 410 |   bottom: "Eltwise2"
 411 |   top: "Eltwise2"
 412 | }
 413 | layer {
 414 |   name: "Convolution6"
 415 |   type: "Convolution"
 416 |   bottom: "Eltwise2"
 417 |   top: "Convolution6"
 418 |   param {
 419 |     lr_mult: 0
 420 |     decay_mult: 0
 421 |   }
 422 |   param {
 423 |     lr_mult: 0
 424 |     decay_mult: 0
 425 |   }
 426 |   convolution_param {
 427 |     num_output: 16
 428 |     pad: 1
 429 |     kernel_size: 3
 430 |     stride: 1
 431 |     weight_filler {
 432 |       type: "msra"
 433 |     }
 434 |     bias_filler {
 435 |       type: "constant"
 436 |       value: 0
 437 |     }
 438 |   }
 439 | }
 440 | layer {
 441 |   name: "BatchNorm6"
 442 |   type: "BatchNorm"
 443 |   bottom: "Convolution6"
 444 |   top: "Convolution6"
 445 |   param {
 446 |     lr_mult: 0
 447 |     decay_mult: 0
 448 |   }
 449 |   param {
 450 |     lr_mult: 0
 451 |     decay_mult: 0
 452 |   }
 453 |   param {
 454 |     lr_mult: 0
 455 |     decay_mult: 0
 456 |   }
 457 |   batch_norm_param {
 458 |     use_global_stats: true
 459 |   }
 460 | }
 461 | layer {
 462 |   name: "Scale6"
 463 |   type: "Scale"
 464 |   bottom: "Convolution6"
 465 |   top: "Convolution6"
 466 |   param {
 467 |     lr_mult: 0
 468 |     decay_mult: 0
 469 |   }
 470 |   param {
 471 |     lr_mult: 0
 472 |     decay_mult: 0
 473 |   }
 474 |   scale_param {
 475 |     bias_term: true
 476 |   }
 477 | }
 478 | layer {
 479 |   name: "ReLU6"
 480 |   type: "ReLU"
 481 |   bottom: "Convolution6"
 482 |   top: "Convolution6"
 483 | }
 484 | layer {
 485 |   name: "Convolution7"
 486 |   type: "Convolution"
 487 |   bottom: "Convolution6"
 488 |   top: "Convolution7"
 489 |   param {
 490 |     lr_mult: 0
 491 |     decay_mult: 0
 492 |   }
 493 |   param {
 494 |     lr_mult: 0
 495 |     decay_mult: 0
 496 |   }
 497 |   convolution_param {
 498 |     num_output: 16
 499 |     pad: 1
 500 |     kernel_size: 3
 501 |     stride: 1
 502 |     weight_filler {
 503 |       type: "msra"
 504 |     }
 505 |     bias_filler {
 506 |       type: "constant"
 507 |       value: 0
 508 |     }
 509 |   }
 510 | }
 511 | layer {
 512 |   name: "BatchNorm7"
 513 |   type: "BatchNorm"
 514 |   bottom: "Convolution7"
 515 |   top: "Convolution7"
 516 |   param {
 517 |     lr_mult: 0
 518 |     decay_mult: 0
 519 |   }
 520 |   param {
 521 |     lr_mult: 0
 522 |     decay_mult: 0
 523 |   }
 524 |   param {
 525 |     lr_mult: 0
 526 |     decay_mult: 0
 527 |   }
 528 |   batch_norm_param {
 529 |     use_global_stats: true
 530 |   }
 531 | }
 532 | layer {
 533 |   name: "Scale7"
 534 |   type: "Scale"
 535 |   bottom: "Convolution7"
 536 |   top: "Convolution7"
 537 |   param {
 538 |     lr_mult: 0
 539 |     decay_mult: 0
 540 |   }
 541 |   param {
 542 |     lr_mult: 0
 543 |     decay_mult: 0
 544 |   }
 545 |   scale_param {
 546 |     bias_term: true
 547 |   }
 548 | }
 549 | layer {
 550 |   name: "Eltwise3"
 551 |   type: "Eltwise"
 552 |   bottom: "Eltwise2"
 553 |   bottom: "Convolution7"
 554 |   top: "Eltwise3"
 555 |   eltwise_param {
 556 |     operation: SUM
 557 |   }
 558 | }
 559 | layer {
 560 |   name: "ReLU7"
 561 |   type: "ReLU"
 562 |   bottom: "Eltwise3"
 563 |   top: "Eltwise3"
 564 | }
 565 | layer {
 566 |   name: "Convolution8"
 567 |   type: "Convolution"
 568 |   bottom: "Eltwise3"
 569 |   top: "Convolution8"
 570 |   param {
 571 |     lr_mult: 0
 572 |     decay_mult: 0
 573 |   }
 574 |   param {
 575 |     lr_mult: 0
 576 |     decay_mult: 0
 577 |   }
 578 |   convolution_param {
 579 |     num_output: 32
 580 |     pad: 0
 581 |     kernel_size: 1
 582 |     stride: 2
 583 |     weight_filler {
 584 |       type: "msra"
 585 |     }
 586 |     bias_filler {
 587 |       type: "constant"
 588 |       value: 0
 589 |     }
 590 |   }
 591 | }
 592 | layer {
 593 |   name: "BatchNorm8"
 594 |   type: "BatchNorm"
 595 |   bottom: "Convolution8"
 596 |   top: "Convolution8"
 597 |   param {
 598 |     lr_mult: 0
 599 |     decay_mult: 0
 600 |   }
 601 |   param {
 602 |     lr_mult: 0
 603 |     decay_mult: 0
 604 |   }
 605 |   param {
 606 |     lr_mult: 0
 607 |     decay_mult: 0
 608 |   }
 609 |   batch_norm_param {
 610 |     use_global_stats: true
 611 |   }
 612 | }
 613 | layer {
 614 |   name: "Scale8"
 615 |   type: "Scale"
 616 |   bottom: "Convolution8"
 617 |   top: "Convolution8"
 618 |   param {
 619 |     lr_mult: 0
 620 |     decay_mult: 0
 621 |   }
 622 |   param {
 623 |     lr_mult: 0
 624 |     decay_mult: 0
 625 |   }
 626 |   scale_param {
 627 |     bias_term: true
 628 |   }
 629 | }
 630 | layer {
 631 |   name: "Convolution9"
 632 |   type: "Convolution"
 633 |   bottom: "Eltwise3"
 634 |   top: "Convolution9"
 635 |   param {
 636 |     lr_mult: 0
 637 |     decay_mult: 0
 638 |   }
 639 |   param {
 640 |     lr_mult: 0
 641 |     decay_mult: 0
 642 |   }
 643 |   convolution_param {
 644 |     num_output: 32
 645 |     pad: 1
 646 |     kernel_size: 3
 647 |     stride: 2
 648 |     weight_filler {
 649 |       type: "msra"
 650 |     }
 651 |     bias_filler {
 652 |       type: "constant"
 653 |       value: 0
 654 |     }
 655 |   }
 656 | }
 657 | layer {
 658 |   name: "BatchNorm9"
 659 |   type: "BatchNorm"
 660 |   bottom: "Convolution9"
 661 |   top: "Convolution9"
 662 |   param {
 663 |     lr_mult: 0
 664 |     decay_mult: 0
 665 |   }
 666 |   param {
 667 |     lr_mult: 0
 668 |     decay_mult: 0
 669 |   }
 670 |   param {
 671 |     lr_mult: 0
 672 |     decay_mult: 0
 673 |   }
 674 |   batch_norm_param {
 675 |     use_global_stats: true
 676 |   }
 677 | }
 678 | layer {
 679 |   name: "Scale9"
 680 |   type: "Scale"
 681 |   bottom: "Convolution9"
 682 |   top: "Convolution9"
 683 |   param {
 684 |     lr_mult: 0
 685 |     decay_mult: 0
 686 |   }
 687 |   param {
 688 |     lr_mult: 0
 689 |     decay_mult: 0
 690 |   }
 691 |   scale_param {
 692 |     bias_term: true
 693 |   }
 694 | }
 695 | layer {
 696 |   name: "ReLU8"
 697 |   type: "ReLU"
 698 |   bottom: "Convolution9"
 699 |   top: "Convolution9"
 700 | }
 701 | layer {
 702 |   name: "Convolution10"
 703 |   type: "Convolution"
 704 |   bottom: "Convolution9"
 705 |   top: "Convolution10"
 706 |   param {
 707 |     lr_mult: 0
 708 |     decay_mult: 0
 709 |   }
 710 |   param {
 711 |     lr_mult: 0
 712 |     decay_mult: 0
 713 |   }
 714 |   convolution_param {
 715 |     num_output: 32
 716 |     pad: 1
 717 |     kernel_size: 3
 718 |     stride: 1
 719 |     weight_filler {
 720 |       type: "msra"
 721 |     }
 722 |     bias_filler {
 723 |       type: "constant"
 724 |       value: 0
 725 |     }
 726 |   }
 727 | }
 728 | layer {
 729 |   name: "BatchNorm10"
 730 |   type: "BatchNorm"
 731 |   bottom: "Convolution10"
 732 |   top: "Convolution10"
 733 |   param {
 734 |     lr_mult: 0
 735 |     decay_mult: 0
 736 |   }
 737 |   param {
 738 |     lr_mult: 0
 739 |     decay_mult: 0
 740 |   }
 741 |   param {
 742 |     lr_mult: 0
 743 |     decay_mult: 0
 744 |   }
 745 |   batch_norm_param {
 746 |     use_global_stats: true
 747 |   }
 748 | }
 749 | layer {
 750 |   name: "Scale10"
 751 |   type: "Scale"
 752 |   bottom: "Convolution10"
 753 |   top: "Convolution10"
 754 |   param {
 755 |     lr_mult: 0
 756 |     decay_mult: 0
 757 |   }
 758 |   param {
 759 |     lr_mult: 0
 760 |     decay_mult: 0
 761 |   }
 762 |   scale_param {
 763 |     bias_term: true
 764 |   }
 765 | }
 766 | layer {
 767 |   name: "Eltwise4"
 768 |   type: "Eltwise"
 769 |   bottom: "Convolution8"
 770 |   bottom: "Convolution10"
 771 |   top: "Eltwise4"
 772 |   eltwise_param {
 773 |     operation: SUM
 774 |   }
 775 | }
 776 | layer {
 777 |   name: "ReLU9"
 778 |   type: "ReLU"
 779 |   bottom: "Eltwise4"
 780 |   top: "Eltwise4"
 781 | }
 782 | layer {
 783 |   name: "Convolution11"
 784 |   type: "Convolution"
 785 |   bottom: "Eltwise4"
 786 |   top: "Convolution11"
 787 |   param {
 788 |     lr_mult: 0
 789 |     decay_mult: 0
 790 |   }
 791 |   param {
 792 |     lr_mult: 0
 793 |     decay_mult: 0
 794 |   }
 795 |   convolution_param {
 796 |     num_output: 32
 797 |     pad: 1
 798 |     kernel_size: 3
 799 |     stride: 1
 800 |     weight_filler {
 801 |       type: "msra"
 802 |     }
 803 |     bias_filler {
 804 |       type: "constant"
 805 |       value: 0
 806 |     }
 807 |   }
 808 | }
 809 | layer {
 810 |   name: "BatchNorm11"
 811 |   type: "BatchNorm"
 812 |   bottom: "Convolution11"
 813 |   top: "Convolution11"
 814 |   param {
 815 |     lr_mult: 0
 816 |     decay_mult: 0
 817 |   }
 818 |   param {
 819 |     lr_mult: 0
 820 |     decay_mult: 0
 821 |   }
 822 |   param {
 823 |     lr_mult: 0
 824 |     decay_mult: 0
 825 |   }
 826 |   batch_norm_param {
 827 |     use_global_stats: true
 828 |   }
 829 | }
 830 | layer {
 831 |   name: "Scale11"
 832 |   type: "Scale"
 833 |   bottom: "Convolution11"
 834 |   top: "Convolution11"
 835 |   param {
 836 |     lr_mult: 0
 837 |     decay_mult: 0
 838 |   }
 839 |   param {
 840 |     lr_mult: 0
 841 |     decay_mult: 0
 842 |   }
 843 |   scale_param {
 844 |     bias_term: true
 845 |   }
 846 | }
 847 | layer {
 848 |   name: "ReLU10"
 849 |   type: "ReLU"
 850 |   bottom: "Convolution11"
 851 |   top: "Convolution11"
 852 | }
 853 | layer {
 854 |   name: "Convolution12"
 855 |   type: "Convolution"
 856 |   bottom: "Convolution11"
 857 |   top: "Convolution12"
 858 |   param {
 859 |     lr_mult: 0
 860 |     decay_mult: 0
 861 |   }
 862 |   param {
 863 |     lr_mult: 0
 864 |     decay_mult: 0
 865 |   }
 866 |   convolution_param {
 867 |     num_output: 32
 868 |     pad: 1
 869 |     kernel_size: 3
 870 |     stride: 1
 871 |     weight_filler {
 872 |       type: "msra"
 873 |     }
 874 |     bias_filler {
 875 |       type: "constant"
 876 |       value: 0
 877 |     }
 878 |   }
 879 | }
 880 | layer {
 881 |   name: "BatchNorm12"
 882 |   type: "BatchNorm"
 883 |   bottom: "Convolution12"
 884 |   top: "Convolution12"
 885 |   param {
 886 |     lr_mult: 0
 887 |     decay_mult: 0
 888 |   }
 889 |   param {
 890 |     lr_mult: 0
 891 |     decay_mult: 0
 892 |   }
 893 |   param {
 894 |     lr_mult: 0
 895 |     decay_mult: 0
 896 |   }
 897 |   batch_norm_param {
 898 |     use_global_stats: true
 899 |   }
 900 | }
 901 | layer {
 902 |   name: "Scale12"
 903 |   type: "Scale"
 904 |   bottom: "Convolution12"
 905 |   top: "Convolution12"
 906 |   param {
 907 |     lr_mult: 0
 908 |     decay_mult: 0
 909 |   }
 910 |   param {
 911 |     lr_mult: 0
 912 |     decay_mult: 0
 913 |   }
 914 |   scale_param {
 915 |     bias_term: true
 916 |   }
 917 | }
 918 | layer {
 919 |   name: "Eltwise5"
 920 |   type: "Eltwise"
 921 |   bottom: "Eltwise4"
 922 |   bottom: "Convolution12"
 923 |   top: "Eltwise5"
 924 |   eltwise_param {
 925 |     operation: SUM
 926 |   }
 927 | }
 928 | layer {
 929 |   name: "ReLU11"
 930 |   type: "ReLU"
 931 |   bottom: "Eltwise5"
 932 |   top: "Eltwise5"
 933 | }
 934 | layer {
 935 |   name: "Convolution13"
 936 |   type: "Convolution"
 937 |   bottom: "Eltwise5"
 938 |   top: "Convolution13"
 939 |   param {
 940 |     lr_mult: 0
 941 |     decay_mult: 0
 942 |   }
 943 |   param {
 944 |     lr_mult: 0
 945 |     decay_mult: 0
 946 |   }
 947 |   convolution_param {
 948 |     num_output: 32
 949 |     pad: 1
 950 |     kernel_size: 3
 951 |     stride: 1
 952 |     weight_filler {
 953 |       type: "msra"
 954 |     }
 955 |     bias_filler {
 956 |       type: "constant"
 957 |       value: 0
 958 |     }
 959 |   }
 960 | }
 961 | layer {
 962 |   name: "BatchNorm13"
 963 |   type: "BatchNorm"
 964 |   bottom: "Convolution13"
 965 |   top: "Convolution13"
 966 |   param {
 967 |     lr_mult: 0
 968 |     decay_mult: 0
 969 |   }
 970 |   param {
 971 |     lr_mult: 0
 972 |     decay_mult: 0
 973 |   }
 974 |   param {
 975 |     lr_mult: 0
 976 |     decay_mult: 0
 977 |   }
 978 |   batch_norm_param {
 979 |     use_global_stats: true
 980 |   }
 981 | }
 982 | layer {
 983 |   name: "Scale13"
 984 |   type: "Scale"
 985 |   bottom: "Convolution13"
 986 |   top: "Convolution13"
 987 |   param {
 988 |     lr_mult: 0
 989 |     decay_mult: 0
 990 |   }
 991 |   param {
 992 |     lr_mult: 0
 993 |     decay_mult: 0
 994 |   }
 995 |   scale_param {
 996 |     bias_term: true
 997 |   }
 998 | }
 999 | layer {
1000 |   name: "ReLU12"
1001 |   type: "ReLU"
1002 |   bottom: "Convolution13"
1003 |   top: "Convolution13"
1004 | }
1005 | layer {
1006 |   name: "Convolution14"
1007 |   type: "Convolution"
1008 |   bottom: "Convolution13"
1009 |   top: "Convolution14"
1010 |   param {
1011 |     lr_mult: 0
1012 |     decay_mult: 0
1013 |   }
1014 |   param {
1015 |     lr_mult: 0
1016 |     decay_mult: 0
1017 |   }
1018 |   convolution_param {
1019 |     num_output: 32
1020 |     pad: 1
1021 |     kernel_size: 3
1022 |     stride: 1
1023 |     weight_filler {
1024 |       type: "msra"
1025 |     }
1026 |     bias_filler {
1027 |       type: "constant"
1028 |       value: 0
1029 |     }
1030 |   }
1031 | }
1032 | layer {
1033 |   name: "BatchNorm14"
1034 |   type: "BatchNorm"
1035 |   bottom: "Convolution14"
1036 |   top: "Convolution14"
1037 |   param {
1038 |     lr_mult: 0
1039 |     decay_mult: 0
1040 |   }
1041 |   param {
1042 |     lr_mult: 0
1043 |     decay_mult: 0
1044 |   }
1045 |   param {
1046 |     lr_mult: 0
1047 |     decay_mult: 0
1048 |   }
1049 |   batch_norm_param {
1050 |     use_global_stats: true
1051 |   }
1052 | }
1053 | layer {
1054 |   name: "Scale14"
1055 |   type: "Scale"
1056 |   bottom: "Convolution14"
1057 |   top: "Convolution14"
1058 |   param {
1059 |     lr_mult: 0
1060 |     decay_mult: 0
1061 |   }
1062 |   param {
1063 |     lr_mult: 0
1064 |     decay_mult: 0
1065 |   }
1066 |   scale_param {
1067 |     bias_term: true
1068 |   }
1069 | }
1070 | layer {
1071 |   name: "Eltwise6"
1072 |   type: "Eltwise"
1073 |   bottom: "Eltwise5"
1074 |   bottom: "Convolution14"
1075 |   top: "Eltwise6"
1076 |   eltwise_param {
1077 |     operation: SUM
1078 |   }
1079 | }
1080 | layer {
1081 |   name: "ReLU13"
1082 |   type: "ReLU"
1083 |   bottom: "Eltwise6"
1084 |   top: "Eltwise6"
1085 | }
1086 | layer {
1087 |   name: "Convolution15"
1088 |   type: "Convolution"
1089 |   bottom: "Eltwise6"
1090 |   top: "Convolution15"
1091 |   param {
1092 |     lr_mult: 0
1093 |     decay_mult: 0
1094 |   }
1095 |   param {
1096 |     lr_mult: 0
1097 |     decay_mult: 0
1098 |   }
1099 |   convolution_param {
1100 |     num_output: 64
1101 |     pad: 0
1102 |     kernel_size: 1
1103 |     stride: 2
1104 |     weight_filler {
1105 |       type: "msra"
1106 |     }
1107 |     bias_filler {
1108 |       type: "constant"
1109 |       value: 0
1110 |     }
1111 |   }
1112 | }
1113 | layer {
1114 |   name: "BatchNorm15"
1115 |   type: "BatchNorm"
1116 |   bottom: "Convolution15"
1117 |   top: "Convolution15"
1118 |   param {
1119 |     lr_mult: 0
1120 |     decay_mult: 0
1121 |   }
1122 |   param {
1123 |     lr_mult: 0
1124 |     decay_mult: 0
1125 |   }
1126 |   param {
1127 |     lr_mult: 0
1128 |     decay_mult: 0
1129 |   }
1130 |   batch_norm_param {
1131 |     use_global_stats: true
1132 |   }
1133 | }
1134 | layer {
1135 |   name: "Scale15"
1136 |   type: "Scale"
1137 |   bottom: "Convolution15"
1138 |   top: "Convolution15"
1139 |   param {
1140 |     lr_mult: 0
1141 |     decay_mult: 0
1142 |   }
1143 |   param {
1144 |     lr_mult: 0
1145 |     decay_mult: 0
1146 |   }
1147 |   scale_param {
1148 |     bias_term: true
1149 |   }
1150 | }
1151 | layer {
1152 |   name: "Convolution16"
1153 |   type: "Convolution"
1154 |   bottom: "Eltwise6"
1155 |   top: "Convolution16"
1156 |   param {
1157 |     lr_mult: 0
1158 |     decay_mult: 0
1159 |   }
1160 |   param {
1161 |     lr_mult: 0
1162 |     decay_mult: 0
1163 |   }
1164 |   convolution_param {
1165 |     num_output: 64
1166 |     pad: 1
1167 |     kernel_size: 3
1168 |     stride: 2
1169 |     weight_filler {
1170 |       type: "msra"
1171 |     }
1172 |     bias_filler {
1173 |       type: "constant"
1174 |       value: 0
1175 |     }
1176 |   }
1177 | }
1178 | layer {
1179 |   name: "BatchNorm16"
1180 |   type: "BatchNorm"
1181 |   bottom: "Convolution16"
1182 |   top: "Convolution16"
1183 |   param {
1184 |     lr_mult: 0
1185 |     decay_mult: 0
1186 |   }
1187 |   param {
1188 |     lr_mult: 0
1189 |     decay_mult: 0
1190 |   }
1191 |   param {
1192 |     lr_mult: 0
1193 |     decay_mult: 0
1194 |   }
1195 |   batch_norm_param {
1196 |     use_global_stats: true
1197 |   }
1198 | }
1199 | layer {
1200 |   name: "Scale16"
1201 |   type: "Scale"
1202 |   bottom: "Convolution16"
1203 |   top: "Convolution16"
1204 |   param {
1205 |     lr_mult: 0
1206 |     decay_mult: 0
1207 |   }
1208 |   param {
1209 |     lr_mult: 0
1210 |     decay_mult: 0
1211 |   }
1212 |   scale_param {
1213 |     bias_term: true
1214 |   }
1215 | }
1216 | layer {
1217 |   name: "ReLU14"
1218 |   type: "ReLU"
1219 |   bottom: "Convolution16"
1220 |   top: "Convolution16"
1221 | }
1222 | layer {
1223 |   name: "Convolution17"
1224 |   type: "Convolution"
1225 |   bottom: "Convolution16"
1226 |   top: "Convolution17"
1227 |   param {
1228 |     lr_mult: 0
1229 |     decay_mult: 0
1230 |   }
1231 |   param {
1232 |     lr_mult: 0
1233 |     decay_mult: 0
1234 |   }
1235 |   convolution_param {
1236 |     num_output: 64
1237 |     pad: 1
1238 |     kernel_size: 3
1239 |     stride: 1
1240 |     weight_filler {
1241 |       type: "msra"
1242 |     }
1243 |     bias_filler {
1244 |       type: "constant"
1245 |       value: 0
1246 |     }
1247 |   }
1248 | }
1249 | layer {
1250 |   name: "BatchNorm17"
1251 |   type: "BatchNorm"
1252 |   bottom: "Convolution17"
1253 |   top: "Convolution17"
1254 |   param {
1255 |     lr_mult: 0
1256 |     decay_mult: 0
1257 |   }
1258 |   param {
1259 |     lr_mult: 0
1260 |     decay_mult: 0
1261 |   }
1262 |   param {
1263 |     lr_mult: 0
1264 |     decay_mult: 0
1265 |   }
1266 |   batch_norm_param {
1267 |     use_global_stats: true
1268 |   }
1269 | }
1270 | layer {
1271 |   name: "Scale17"
1272 |   type: "Scale"
1273 |   bottom: "Convolution17"
1274 |   top: "Convolution17"
1275 |   param {
1276 |     lr_mult: 0
1277 |     decay_mult: 0
1278 |   }
1279 |   param {
1280 |     lr_mult: 0
1281 |     decay_mult: 0
1282 |   }
1283 |   scale_param {
1284 |     bias_term: true
1285 |   }
1286 | }
1287 | layer {
1288 |   name: "Eltwise7"
1289 |   type: "Eltwise"
1290 |   bottom: "Convolution15"
1291 |   bottom: "Convolution17"
1292 |   top: "Eltwise7"
1293 |   eltwise_param {
1294 |     operation: SUM
1295 |   }
1296 | }
1297 | layer {
1298 |   name: "ReLU15"
1299 |   type: "ReLU"
1300 |   bottom: "Eltwise7"
1301 |   top: "Eltwise7"
1302 | }
1303 | layer {
1304 |   name: "Convolution18"
1305 |   type: "Convolution"
1306 |   bottom: "Eltwise7"
1307 |   top: "Convolution18"
1308 |   param {
1309 |     lr_mult: 0
1310 |     decay_mult: 0
1311 |   }
1312 |   param {
1313 |     lr_mult: 0
1314 |     decay_mult: 0
1315 |   }
1316 |   convolution_param {
1317 |     num_output: 64
1318 |     pad: 1
1319 |     kernel_size: 3
1320 |     stride: 1
1321 |     weight_filler {
1322 |       type: "msra"
1323 |     }
1324 |     bias_filler {
1325 |       type: "constant"
1326 |       value: 0
1327 |     }
1328 |   }
1329 | }
1330 | layer {
1331 |   name: "BatchNorm18"
1332 |   type: "BatchNorm"
1333 |   bottom: "Convolution18"
1334 |   top: "Convolution18"
1335 |   param {
1336 |     lr_mult: 0
1337 |     decay_mult: 0
1338 |   }
1339 |   param {
1340 |     lr_mult: 0
1341 |     decay_mult: 0
1342 |   }
1343 |   param {
1344 |     lr_mult: 0
1345 |     decay_mult: 0
1346 |   }
1347 |   batch_norm_param {
1348 |     use_global_stats: true
1349 |   }
1350 | }
1351 | layer {
1352 |   name: "Scale18"
1353 |   type: "Scale"
1354 |   bottom: "Convolution18"
1355 |   top: "Convolution18"
1356 |   param {
1357 |     lr_mult: 0
1358 |     decay_mult: 0
1359 |   }
1360 |   param {
1361 |     lr_mult: 0
1362 |     decay_mult: 0
1363 |   }
1364 |   scale_param {
1365 |     bias_term: true
1366 |   }
1367 | }
1368 | layer {
1369 |   name: "ReLU16"
1370 |   type: "ReLU"
1371 |   bottom: "Convolution18"
1372 |   top: "Convolution18"
1373 | }
1374 | layer {
1375 |   name: "Convolution19"
1376 |   type: "Convolution"
1377 |   bottom: "Convolution18"
1378 |   top: "Convolution19"
1379 |   param {
1380 |     lr_mult: 0
1381 |     decay_mult: 0
1382 |   }
1383 |   param {
1384 |     lr_mult: 0
1385 |     decay_mult: 0
1386 |   }
1387 |   convolution_param {
1388 |     num_output: 64
1389 |     pad: 1
1390 |     kernel_size: 3
1391 |     stride: 1
1392 |     weight_filler {
1393 |       type: "msra"
1394 |     }
1395 |     bias_filler {
1396 |       type: "constant"
1397 |       value: 0
1398 |     }
1399 |   }
1400 | }
1401 | layer {
1402 |   name: "BatchNorm19"
1403 |   type: "BatchNorm"
1404 |   bottom: "Convolution19"
1405 |   top: "Convolution19"
1406 |   param {
1407 |     lr_mult: 0
1408 |     decay_mult: 0
1409 |   }
1410 |   param {
1411 |     lr_mult: 0
1412 |     decay_mult: 0
1413 |   }
1414 |   param {
1415 |     lr_mult: 0
1416 |     decay_mult: 0
1417 |   }
1418 |   batch_norm_param {
1419 |     use_global_stats: true
1420 |   }
1421 | }
1422 | layer {
1423 |   name: "Scale19"
1424 |   type: "Scale"
1425 |   bottom: "Convolution19"
1426 |   top: "Convolution19"
1427 |   param {
1428 |     lr_mult: 0
1429 |     decay_mult: 0
1430 |   }
1431 |   param {
1432 |     lr_mult: 0
1433 |     decay_mult: 0
1434 |   }
1435 |   scale_param {
1436 |     bias_term: true
1437 |   }
1438 | }
1439 | layer {
1440 |   name: "Eltwise8"
1441 |   type: "Eltwise"
1442 |   bottom: "Eltwise7"
1443 |   bottom: "Convolution19"
1444 |   top: "Eltwise8"
1445 |   eltwise_param {
1446 |     operation: SUM
1447 |   }
1448 | }
1449 | layer {
1450 |   name: "ReLU17"
1451 |   type: "ReLU"
1452 |   bottom: "Eltwise8"
1453 |   top: "Eltwise8"
1454 | }
1455 | layer {
1456 |   name: "Convolution20"
1457 |   type: "Convolution"
1458 |   bottom: "Eltwise8"
1459 |   top: "Convolution20"
1460 |   param {
1461 |     lr_mult: 0
1462 |     decay_mult: 0
1463 |   }
1464 |   param {
1465 |     lr_mult: 0
1466 |     decay_mult: 0
1467 |   }
1468 |   convolution_param {
1469 |     num_output: 64
1470 |     pad: 1
1471 |     kernel_size: 3
1472 |     stride: 1
1473 |     weight_filler {
1474 |       type: "msra"
1475 |     }
1476 |     bias_filler {
1477 |       type: "constant"
1478 |       value: 0
1479 |     }
1480 |   }
1481 | }
1482 | layer {
1483 |   name: "BatchNorm20"
1484 |   type: "BatchNorm"
1485 |   bottom: "Convolution20"
1486 |   top: "Convolution20"
1487 |   param {
1488 |     lr_mult: 0
1489 |     decay_mult: 0
1490 |   }
1491 |   param {
1492 |     lr_mult: 0
1493 |     decay_mult: 0
1494 |   }
1495 |   param {
1496 |     lr_mult: 0
1497 |     decay_mult: 0
1498 |   }
1499 |   batch_norm_param {
1500 |     use_global_stats: true
1501 |   }
1502 | }
1503 | layer {
1504 |   name: "Scale20"
1505 |   type: "Scale"
1506 |   bottom: "Convolution20"
1507 |   top: "Convolution20"
1508 |   param {
1509 |     lr_mult: 0
1510 |     decay_mult: 0
1511 |   }
1512 |   param {
1513 |     lr_mult: 0
1514 |     decay_mult: 0
1515 |   }
1516 |   scale_param {
1517 |     bias_term: true
1518 |   }
1519 | }
1520 | layer {
1521 |   name: "ReLU18"
1522 |   type: "ReLU"
1523 |   bottom: "Convolution20"
1524 |   top: "Convolution20"
1525 | }
1526 | layer {
1527 |   name: "Convolution21"
1528 |   type: "Convolution"
1529 |   bottom: "Convolution20"
1530 |   top: "Convolution21"
1531 |   param {
1532 |     lr_mult: 0
1533 |     decay_mult: 0
1534 |   }
1535 |   param {
1536 |     lr_mult: 0
1537 |     decay_mult: 0
1538 |   }
1539 |   convolution_param {
1540 |     num_output: 64
1541 |     pad: 1
1542 |     kernel_size: 3
1543 |     stride: 1
1544 |     weight_filler {
1545 |       type: "msra"
1546 |     }
1547 |     bias_filler {
1548 |       type: "constant"
1549 |       value: 0
1550 |     }
1551 |   }
1552 | }
1553 | layer {
1554 |   name: "BatchNorm21"
1555 |   type: "BatchNorm"
1556 |   bottom: "Convolution21"
1557 |   top: "Convolution21"
1558 |   param {
1559 |     lr_mult: 0
1560 |     decay_mult: 0
1561 |   }
1562 |   param {
1563 |     lr_mult: 0
1564 |     decay_mult: 0
1565 |   }
1566 |   param {
1567 |     lr_mult: 0
1568 |     decay_mult: 0
1569 |   }
1570 |   batch_norm_param {
1571 |     use_global_stats: true
1572 |   }
1573 | }
1574 | layer {
1575 |   name: "Scale21"
1576 |   type: "Scale"
1577 |   bottom: "Convolution21"
1578 |   top: "Convolution21"
1579 |   param {
1580 |     lr_mult: 0
1581 |     decay_mult: 0
1582 |   }
1583 |   param {
1584 |     lr_mult: 0
1585 |     decay_mult: 0
1586 |   }
1587 |   scale_param {
1588 |     bias_term: true
1589 |   }
1590 | }
1591 | layer {
1592 |   name: "Eltwise9"
1593 |   type: "Eltwise"
1594 |   bottom: "Eltwise8"
1595 |   bottom: "Convolution21"
1596 |   top: "Eltwise9"
1597 |   eltwise_param {
1598 |     operation: SUM
1599 |   }
1600 | }
1601 | layer {
1602 |   name: "ReLU19"
1603 |   type: "ReLU"
1604 |   bottom: "Eltwise9"
1605 |   top: "Eltwise9"
1606 | }
1607 | layer {
1608 |   name: "Pooling1"
1609 |   type: "Pooling"
1610 |   bottom: "Eltwise9"
1611 |   top: "Pooling1"
1612 |   pooling_param {
1613 |     pool: AVE
1614 |     global_pooling: true
1615 |   }
1616 | }
1617 | layer {
1618 |   name: "bem"
1619 |   type: "BatchEuclideanMap"
1620 |   bottom: "Pooling1"
1621 |   top: "bem"
1622 | }
1623 | layer {
1624 |   name: "InnerProduct1"
1625 |   type: "InnerProduct"
1626 |   bottom: "Pooling1"
1627 |   top: "InnerProduct1"
1628 |   param {
1629 |     lr_mult: 0
1630 |     decay_mult: 0
1631 |   }
1632 |   param {
1633 |     lr_mult: 0
1634 |     decay_mult: 0
1635 |   }
1636 |   inner_product_param {
1637 |     num_output: 10
1638 |     weight_filler {
1639 |       type: "msra"
1640 |     }
1641 |     bias_filler {
1642 |       type: "constant"
1643 |       value: 0
1644 |     }
1645 |   }
1646 | }
1647 | layer {
1648 |   name: "SoftmaxWithLoss1"
1649 |   type: "SoftmaxWithLoss"
1650 |   bottom: "InnerProduct1"
1651 |   bottom: "Data2"
1652 |   top: "SoftmaxWithLoss1"
1653 |   loss_weight: 0
1654 | }
1655 | layer {
1656 |   name: "Accuracy1"
1657 |   type: "Accuracy"
1658 |   bottom: "InnerProduct1"
1659 |   bottom: "Data2"
1660 |   top: "Accuracy1"
1661 |   include {
1662 |     phase: TEST
1663 |   }
1664 | }
1665 | 
1666 | layer {
1667 |   name: "stu_Convolution1"
1668 |   type: "Convolution"
1669 |   bottom: "Data1"
1670 |   top: "stu_Convolution1"
1671 |   param {
1672 |     lr_mult: 1
1673 |     decay_mult: 1
1674 |   }
1675 |   param {
1676 |     lr_mult: 2
1677 |     decay_mult: 0
1678 |   }
1679 |   convolution_param {
1680 |     num_output: 8
1681 |     pad: 1
1682 |     kernel_size: 3
1683 |     stride: 1
1684 |     weight_filler {
1685 |       type: "msra"
1686 |     }
1687 |     bias_filler {
1688 |       type: "constant"
1689 |       value: 0
1690 |     }
1691 |   }
1692 | }
1693 | layer {
1694 |   name: "stu_BatchNorm1"
1695 |   type: "BatchNorm"
1696 |   bottom: "stu_Convolution1"
1697 |   top: "stu_Convolution1"
1698 |   param {
1699 |     lr_mult: 0
1700 |     decay_mult: 0
1701 |   }
1702 |   param {
1703 |     lr_mult: 0
1704 |     decay_mult: 0
1705 |   }
1706 |   param {
1707 |     lr_mult: 0
1708 |     decay_mult: 0
1709 |   }
1710 | }
1711 | layer {
1712 |   name: "stu_Scale1"
1713 |   type: "Scale"
1714 |   bottom: "stu_Convolution1"
1715 |   top: "stu_Convolution1"
1716 |   scale_param {
1717 |     bias_term: true
1718 |   }
1719 | }
1720 | layer {
1721 |   name: "stu_ReLU1"
1722 |   type: "ReLU"
1723 |   bottom: "stu_Convolution1"
1724 |   top: "stu_Convolution1"
1725 | }
1726 | layer {
1727 |   name: "stu_Convolution2"
1728 |   type: "Convolution"
1729 |   bottom: "stu_Convolution1"
1730 |   top: "stu_Convolution2"
1731 |   param {
1732 |     lr_mult: 1
1733 |     decay_mult: 1
1734 |   }
1735 |   param {
1736 |     lr_mult: 2
1737 |     decay_mult: 0
1738 |   }
1739 |   convolution_param {
1740 |     num_output: 8
1741 |     pad: 1
1742 |     kernel_size: 3
1743 |     stride: 1
1744 |     weight_filler {
1745 |       type: "msra"
1746 |     }
1747 |     bias_filler {
1748 |       type: "constant"
1749 |       value: 0
1750 |     }
1751 |   }
1752 | }
1753 | layer {
1754 |   name: "stu_BatchNorm2"
1755 |   type: "BatchNorm"
1756 |   bottom: "stu_Convolution2"
1757 |   top: "stu_Convolution2"
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: "stu_Scale2"
1773 |   type: "Scale"
1774 |   bottom: "stu_Convolution2"
1775 |   top: "stu_Convolution2"
1776 |   scale_param {
1777 |     bias_term: true
1778 |   }
1779 | }
1780 | layer {
1781 |   name: "stu_ReLU2"
1782 |   type: "ReLU"
1783 |   bottom: "stu_Convolution2"
1784 |   top: "stu_Convolution2"
1785 | }
1786 | layer {
1787 |   name: "stu_Convolution3"
1788 |   type: "Convolution"
1789 |   bottom: "stu_Convolution2"
1790 |   top: "stu_Convolution3"
1791 |   param {
1792 |     lr_mult: 1
1793 |     decay_mult: 1
1794 |   }
1795 |   param {
1796 |     lr_mult: 2
1797 |     decay_mult: 0
1798 |   }
1799 |   convolution_param {
1800 |     num_output: 8
1801 |     pad: 1
1802 |     kernel_size: 3
1803 |     stride: 1
1804 |     weight_filler {
1805 |       type: "msra"
1806 |     }
1807 |     bias_filler {
1808 |       type: "constant"
1809 |       value: 0
1810 |     }
1811 |   }
1812 | }
1813 | layer {
1814 |   name: "stu_BatchNorm3"
1815 |   type: "BatchNorm"
1816 |   bottom: "stu_Convolution3"
1817 |   top: "stu_Convolution3"
1818 |   param {
1819 |     lr_mult: 0
1820 |     decay_mult: 0
1821 |   }
1822 |   param {
1823 |     lr_mult: 0
1824 |     decay_mult: 0
1825 |   }
1826 |   param {
1827 |     lr_mult: 0
1828 |     decay_mult: 0
1829 |   }
1830 | }
1831 | layer {
1832 |   name: "stu_Scale3"
1833 |   type: "Scale"
1834 |   bottom: "stu_Convolution3"
1835 |   top: "stu_Convolution3"
1836 |   scale_param {
1837 |     bias_term: true
1838 |   }
1839 | }
1840 | layer {
1841 |   name: "stu_Eltwise1"
1842 |   type: "Eltwise"
1843 |   bottom: "stu_Convolution1"
1844 |   bottom: "stu_Convolution3"
1845 |   top: "stu_Eltwise1"
1846 |   eltwise_param {
1847 |     operation: SUM
1848 |   }
1849 | }
1850 | layer {
1851 |   name: "stu_ReLU3"
1852 |   type: "ReLU"
1853 |   bottom: "stu_Eltwise1"
1854 |   top: "stu_Eltwise1"
1855 | }
1856 | layer {
1857 |   name: "stu_Convolution4"
1858 |   type: "Convolution"
1859 |   bottom: "stu_Eltwise1"
1860 |   top: "stu_Convolution4"
1861 |   param {
1862 |     lr_mult: 1
1863 |     decay_mult: 1
1864 |   }
1865 |   param {
1866 |     lr_mult: 2
1867 |     decay_mult: 0
1868 |   }
1869 |   convolution_param {
1870 |     num_output: 8
1871 |     pad: 1
1872 |     kernel_size: 3
1873 |     stride: 1
1874 |     weight_filler {
1875 |       type: "msra"
1876 |     }
1877 |     bias_filler {
1878 |       type: "constant"
1879 |       value: 0
1880 |     }
1881 |   }
1882 | }
1883 | layer {
1884 |   name: "stu_BatchNorm4"
1885 |   type: "BatchNorm"
1886 |   bottom: "stu_Convolution4"
1887 |   top: "stu_Convolution4"
1888 |   param {
1889 |     lr_mult: 0
1890 |     decay_mult: 0
1891 |   }
1892 |   param {
1893 |     lr_mult: 0
1894 |     decay_mult: 0
1895 |   }
1896 |   param {
1897 |     lr_mult: 0
1898 |     decay_mult: 0
1899 |   }
1900 | }
1901 | layer {
1902 |   name: "stu_Scale4"
1903 |   type: "Scale"
1904 |   bottom: "stu_Convolution4"
1905 |   top: "stu_Convolution4"
1906 |   scale_param {
1907 |     bias_term: true
1908 |   }
1909 | }
1910 | layer {
1911 |   name: "stu_ReLU4"
1912 |   type: "ReLU"
1913 |   bottom: "stu_Convolution4"
1914 |   top: "stu_Convolution4"
1915 | }
1916 | layer {
1917 |   name: "stu_Convolution5"
1918 |   type: "Convolution"
1919 |   bottom: "stu_Convolution4"
1920 |   top: "stu_Convolution5"
1921 |   param {
1922 |     lr_mult: 1
1923 |     decay_mult: 1
1924 |   }
1925 |   param {
1926 |     lr_mult: 2
1927 |     decay_mult: 0
1928 |   }
1929 |   convolution_param {
1930 |     num_output: 8
1931 |     pad: 1
1932 |     kernel_size: 3
1933 |     stride: 1
1934 |     weight_filler {
1935 |       type: "msra"
1936 |     }
1937 |     bias_filler {
1938 |       type: "constant"
1939 |       value: 0
1940 |     }
1941 |   }
1942 | }
1943 | layer {
1944 |   name: "stu_BatchNorm5"
1945 |   type: "BatchNorm"
1946 |   bottom: "stu_Convolution5"
1947 |   top: "stu_Convolution5"
1948 |   param {
1949 |     lr_mult: 0
1950 |     decay_mult: 0
1951 |   }
1952 |   param {
1953 |     lr_mult: 0
1954 |     decay_mult: 0
1955 |   }
1956 |   param {
1957 |     lr_mult: 0
1958 |     decay_mult: 0
1959 |   }
1960 | }
1961 | layer {
1962 |   name: "stu_Scale5"
1963 |   type: "Scale"
1964 |   bottom: "stu_Convolution5"
1965 |   top: "stu_Convolution5"
1966 |   scale_param {
1967 |     bias_term: true
1968 |   }
1969 | }
1970 | layer {
1971 |   name: "stu_Eltwise2"
1972 |   type: "Eltwise"
1973 |   bottom: "stu_Eltwise1"
1974 |   bottom: "stu_Convolution5"
1975 |   top: "stu_Eltwise2"
1976 |   eltwise_param {
1977 |     operation: SUM
1978 |   }
1979 | }
1980 | layer {
1981 |   name: "stu_ReLU5"
1982 |   type: "ReLU"
1983 |   bottom: "stu_Eltwise2"
1984 |   top: "stu_Eltwise2"
1985 | }
1986 | layer {
1987 |   name: "stu_Convolution6"
1988 |   type: "Convolution"
1989 |   bottom: "stu_Eltwise2"
1990 |   top: "stu_Convolution6"
1991 |   param {
1992 |     lr_mult: 1
1993 |     decay_mult: 1
1994 |   }
1995 |   param {
1996 |     lr_mult: 2
1997 |     decay_mult: 0
1998 |   }
1999 |   convolution_param {
2000 |     num_output: 8
2001 |     pad: 1
2002 |     kernel_size: 3
2003 |     stride: 1
2004 |     weight_filler {
2005 |       type: "msra"
2006 |     }
2007 |     bias_filler {
2008 |       type: "constant"
2009 |       value: 0
2010 |     }
2011 |   }
2012 | }
2013 | layer {
2014 |   name: "stu_BatchNorm6"
2015 |   type: "BatchNorm"
2016 |   bottom: "stu_Convolution6"
2017 |   top: "stu_Convolution6"
2018 |   param {
2019 |     lr_mult: 0
2020 |     decay_mult: 0
2021 |   }
2022 |   param {
2023 |     lr_mult: 0
2024 |     decay_mult: 0
2025 |   }
2026 |   param {
2027 |     lr_mult: 0
2028 |     decay_mult: 0
2029 |   }
2030 | }
2031 | layer {
2032 |   name: "stu_Scale6"
2033 |   type: "Scale"
2034 |   bottom: "stu_Convolution6"
2035 |   top: "stu_Convolution6"
2036 |   scale_param {
2037 |     bias_term: true
2038 |   }
2039 | }
2040 | layer {
2041 |   name: "stu_ReLU6"
2042 |   type: "ReLU"
2043 |   bottom: "stu_Convolution6"
2044 |   top: "stu_Convolution6"
2045 | }
2046 | layer {
2047 |   name: "stu_Convolution7"
2048 |   type: "Convolution"
2049 |   bottom: "stu_Convolution6"
2050 |   top: "stu_Convolution7"
2051 |   param {
2052 |     lr_mult: 1
2053 |     decay_mult: 1
2054 |   }
2055 |   param {
2056 |     lr_mult: 2
2057 |     decay_mult: 0
2058 |   }
2059 |   convolution_param {
2060 |     num_output: 8
2061 |     pad: 1
2062 |     kernel_size: 3
2063 |     stride: 1
2064 |     weight_filler {
2065 |       type: "msra"
2066 |     }
2067 |     bias_filler {
2068 |       type: "constant"
2069 |       value: 0
2070 |     }
2071 |   }
2072 | }
2073 | layer {
2074 |   name: "stu_BatchNorm7"
2075 |   type: "BatchNorm"
2076 |   bottom: "stu_Convolution7"
2077 |   top: "stu_Convolution7"
2078 |   param {
2079 |     lr_mult: 0
2080 |     decay_mult: 0
2081 |   }
2082 |   param {
2083 |     lr_mult: 0
2084 |     decay_mult: 0
2085 |   }
2086 |   param {
2087 |     lr_mult: 0
2088 |     decay_mult: 0
2089 |   }
2090 | }
2091 | layer {
2092 |   name: "stu_Scale7"
2093 |   type: "Scale"
2094 |   bottom: "stu_Convolution7"
2095 |   top: "stu_Convolution7"
2096 |   scale_param {
2097 |     bias_term: true
2098 |   }
2099 | }
2100 | layer {
2101 |   name: "stu_Eltwise3"
2102 |   type: "Eltwise"
2103 |   bottom: "stu_Eltwise2"
2104 |   bottom: "stu_Convolution7"
2105 |   top: "stu_Eltwise3"
2106 |   eltwise_param {
2107 |     operation: SUM
2108 |   }
2109 | }
2110 | layer {
2111 |   name: "stu_ReLU7"
2112 |   type: "ReLU"
2113 |   bottom: "stu_Eltwise3"
2114 |   top: "stu_Eltwise3"
2115 | }
2116 | layer {
2117 |   name: "stu_Convolution8"
2118 |   type: "Convolution"
2119 |   bottom: "stu_Eltwise3"
2120 |   top: "stu_Convolution8"
2121 |   param {
2122 |     lr_mult: 1
2123 |     decay_mult: 1
2124 |   }
2125 |   param {
2126 |     lr_mult: 2
2127 |     decay_mult: 0
2128 |   }
2129 |   convolution_param {
2130 |     num_output: 16
2131 |     pad: 0
2132 |     kernel_size: 1
2133 |     stride: 2
2134 |     weight_filler {
2135 |       type: "msra"
2136 |     }
2137 |     bias_filler {
2138 |       type: "constant"
2139 |       value: 0
2140 |     }
2141 |   }
2142 | }
2143 | layer {
2144 |   name: "stu_BatchNorm8"
2145 |   type: "BatchNorm"
2146 |   bottom: "stu_Convolution8"
2147 |   top: "stu_Convolution8"
2148 |   param {
2149 |     lr_mult: 0
2150 |     decay_mult: 0
2151 |   }
2152 |   param {
2153 |     lr_mult: 0
2154 |     decay_mult: 0
2155 |   }
2156 |   param {
2157 |     lr_mult: 0
2158 |     decay_mult: 0
2159 |   }
2160 | }
2161 | layer {
2162 |   name: "stu_Scale8"
2163 |   type: "Scale"
2164 |   bottom: "stu_Convolution8"
2165 |   top: "stu_Convolution8"
2166 |   scale_param {
2167 |     bias_term: true
2168 |   }
2169 | }
2170 | layer {
2171 |   name: "stu_Convolution9"
2172 |   type: "Convolution"
2173 |   bottom: "stu_Eltwise3"
2174 |   top: "stu_Convolution9"
2175 |   param {
2176 |     lr_mult: 1
2177 |     decay_mult: 1
2178 |   }
2179 |   param {
2180 |     lr_mult: 2
2181 |     decay_mult: 0
2182 |   }
2183 |   convolution_param {
2184 |     num_output: 16
2185 |     pad: 1
2186 |     kernel_size: 3
2187 |     stride: 2
2188 |     weight_filler {
2189 |       type: "msra"
2190 |     }
2191 |     bias_filler {
2192 |       type: "constant"
2193 |       value: 0
2194 |     }
2195 |   }
2196 | }
2197 | layer {
2198 |   name: "stu_BatchNorm9"
2199 |   type: "BatchNorm"
2200 |   bottom: "stu_Convolution9"
2201 |   top: "stu_Convolution9"
2202 |   param {
2203 |     lr_mult: 0
2204 |     decay_mult: 0
2205 |   }
2206 |   param {
2207 |     lr_mult: 0
2208 |     decay_mult: 0
2209 |   }
2210 |   param {
2211 |     lr_mult: 0
2212 |     decay_mult: 0
2213 |   }
2214 | }
2215 | layer {
2216 |   name: "stu_Scale9"
2217 |   type: "Scale"
2218 |   bottom: "stu_Convolution9"
2219 |   top: "stu_Convolution9"
2220 |   scale_param {
2221 |     bias_term: true
2222 |   }
2223 | }
2224 | layer {
2225 |   name: "stu_ReLU8"
2226 |   type: "ReLU"
2227 |   bottom: "stu_Convolution9"
2228 |   top: "stu_Convolution9"
2229 | }
2230 | layer {
2231 |   name: "stu_Convolution10"
2232 |   type: "Convolution"
2233 |   bottom: "stu_Convolution9"
2234 |   top: "stu_Convolution10"
2235 |   param {
2236 |     lr_mult: 1
2237 |     decay_mult: 1
2238 |   }
2239 |   param {
2240 |     lr_mult: 2
2241 |     decay_mult: 0
2242 |   }
2243 |   convolution_param {
2244 |     num_output: 16
2245 |     pad: 1
2246 |     kernel_size: 3
2247 |     stride: 1
2248 |     weight_filler {
2249 |       type: "msra"
2250 |     }
2251 |     bias_filler {
2252 |       type: "constant"
2253 |       value: 0
2254 |     }
2255 |   }
2256 | }
2257 | layer {
2258 |   name: "stu_BatchNorm10"
2259 |   type: "BatchNorm"
2260 |   bottom: "stu_Convolution10"
2261 |   top: "stu_Convolution10"
2262 |   param {
2263 |     lr_mult: 0
2264 |     decay_mult: 0
2265 |   }
2266 |   param {
2267 |     lr_mult: 0
2268 |     decay_mult: 0
2269 |   }
2270 |   param {
2271 |     lr_mult: 0
2272 |     decay_mult: 0
2273 |   }
2274 | }
2275 | layer {
2276 |   name: "stu_Scale10"
2277 |   type: "Scale"
2278 |   bottom: "stu_Convolution10"
2279 |   top: "stu_Convolution10"
2280 |   scale_param {
2281 |     bias_term: true
2282 |   }
2283 | }
2284 | layer {
2285 |   name: "stu_Eltwise4"
2286 |   type: "Eltwise"
2287 |   bottom: "stu_Convolution8"
2288 |   bottom: "stu_Convolution10"
2289 |   top: "stu_Eltwise4"
2290 |   eltwise_param {
2291 |     operation: SUM
2292 |   }
2293 | }
2294 | layer {
2295 |   name: "stu_ReLU9"
2296 |   type: "ReLU"
2297 |   bottom: "stu_Eltwise4"
2298 |   top: "stu_Eltwise4"
2299 | }
2300 | layer {
2301 |   name: "stu_Convolution11"
2302 |   type: "Convolution"
2303 |   bottom: "stu_Eltwise4"
2304 |   top: "stu_Convolution11"
2305 |   param {
2306 |     lr_mult: 1
2307 |     decay_mult: 1
2308 |   }
2309 |   param {
2310 |     lr_mult: 2
2311 |     decay_mult: 0
2312 |   }
2313 |   convolution_param {
2314 |     num_output: 16
2315 |     pad: 1
2316 |     kernel_size: 3
2317 |     stride: 1
2318 |     weight_filler {
2319 |       type: "msra"
2320 |     }
2321 |     bias_filler {
2322 |       type: "constant"
2323 |       value: 0
2324 |     }
2325 |   }
2326 | }
2327 | layer {
2328 |   name: "stu_BatchNorm11"
2329 |   type: "BatchNorm"
2330 |   bottom: "stu_Convolution11"
2331 |   top: "stu_Convolution11"
2332 |   param {
2333 |     lr_mult: 0
2334 |     decay_mult: 0
2335 |   }
2336 |   param {
2337 |     lr_mult: 0
2338 |     decay_mult: 0
2339 |   }
2340 |   param {
2341 |     lr_mult: 0
2342 |     decay_mult: 0
2343 |   }
2344 | }
2345 | layer {
2346 |   name: "stu_Scale11"
2347 |   type: "Scale"
2348 |   bottom: "stu_Convolution11"
2349 |   top: "stu_Convolution11"
2350 |   scale_param {
2351 |     bias_term: true
2352 |   }
2353 | }
2354 | layer {
2355 |   name: "stu_ReLU10"
2356 |   type: "ReLU"
2357 |   bottom: "stu_Convolution11"
2358 |   top: "stu_Convolution11"
2359 | }
2360 | layer {
2361 |   name: "stu_Convolution12"
2362 |   type: "Convolution"
2363 |   bottom: "stu_Convolution11"
2364 |   top: "stu_Convolution12"
2365 |   param {
2366 |     lr_mult: 1
2367 |     decay_mult: 1
2368 |   }
2369 |   param {
2370 |     lr_mult: 2
2371 |     decay_mult: 0
2372 |   }
2373 |   convolution_param {
2374 |     num_output: 16
2375 |     pad: 1
2376 |     kernel_size: 3
2377 |     stride: 1
2378 |     weight_filler {
2379 |       type: "msra"
2380 |     }
2381 |     bias_filler {
2382 |       type: "constant"
2383 |       value: 0
2384 |     }
2385 |   }
2386 | }
2387 | layer {
2388 |   name: "stu_BatchNorm12"
2389 |   type: "BatchNorm"
2390 |   bottom: "stu_Convolution12"
2391 |   top: "stu_Convolution12"
2392 |   param {
2393 |     lr_mult: 0
2394 |     decay_mult: 0
2395 |   }
2396 |   param {
2397 |     lr_mult: 0
2398 |     decay_mult: 0
2399 |   }
2400 |   param {
2401 |     lr_mult: 0
2402 |     decay_mult: 0
2403 |   }
2404 | }
2405 | layer {
2406 |   name: "stu_Scale12"
2407 |   type: "Scale"
2408 |   bottom: "stu_Convolution12"
2409 |   top: "stu_Convolution12"
2410 |   scale_param {
2411 |     bias_term: true
2412 |   }
2413 | }
2414 | layer {
2415 |   name: "stu_Eltwise5"
2416 |   type: "Eltwise"
2417 |   bottom: "stu_Eltwise4"
2418 |   bottom: "stu_Convolution12"
2419 |   top: "stu_Eltwise5"
2420 |   eltwise_param {
2421 |     operation: SUM
2422 |   }
2423 | }
2424 | layer {
2425 |   name: "stu_ReLU11"
2426 |   type: "ReLU"
2427 |   bottom: "stu_Eltwise5"
2428 |   top: "stu_Eltwise5"
2429 | }
2430 | layer {
2431 |   name: "stu_Convolution13"
2432 |   type: "Convolution"
2433 |   bottom: "stu_Eltwise5"
2434 |   top: "stu_Convolution13"
2435 |   param {
2436 |     lr_mult: 1
2437 |     decay_mult: 1
2438 |   }
2439 |   param {
2440 |     lr_mult: 2
2441 |     decay_mult: 0
2442 |   }
2443 |   convolution_param {
2444 |     num_output: 16
2445 |     pad: 1
2446 |     kernel_size: 3
2447 |     stride: 1
2448 |     weight_filler {
2449 |       type: "msra"
2450 |     }
2451 |     bias_filler {
2452 |       type: "constant"
2453 |       value: 0
2454 |     }
2455 |   }
2456 | }
2457 | layer {
2458 |   name: "stu_BatchNorm13"
2459 |   type: "BatchNorm"
2460 |   bottom: "stu_Convolution13"
2461 |   top: "stu_Convolution13"
2462 |   param {
2463 |     lr_mult: 0
2464 |     decay_mult: 0
2465 |   }
2466 |   param {
2467 |     lr_mult: 0
2468 |     decay_mult: 0
2469 |   }
2470 |   param {
2471 |     lr_mult: 0
2472 |     decay_mult: 0
2473 |   }
2474 | }
2475 | layer {
2476 |   name: "stu_Scale13"
2477 |   type: "Scale"
2478 |   bottom: "stu_Convolution13"
2479 |   top: "stu_Convolution13"
2480 |   scale_param {
2481 |     bias_term: true
2482 |   }
2483 | }
2484 | layer {
2485 |   name: "stu_ReLU12"
2486 |   type: "ReLU"
2487 |   bottom: "stu_Convolution13"
2488 |   top: "stu_Convolution13"
2489 | }
2490 | layer {
2491 |   name: "stu_Convolution14"
2492 |   type: "Convolution"
2493 |   bottom: "stu_Convolution13"
2494 |   top: "stu_Convolution14"
2495 |   param {
2496 |     lr_mult: 1
2497 |     decay_mult: 1
2498 |   }
2499 |   param {
2500 |     lr_mult: 2
2501 |     decay_mult: 0
2502 |   }
2503 |   convolution_param {
2504 |     num_output: 16
2505 |     pad: 1
2506 |     kernel_size: 3
2507 |     stride: 1
2508 |     weight_filler {
2509 |       type: "msra"
2510 |     }
2511 |     bias_filler {
2512 |       type: "constant"
2513 |       value: 0
2514 |     }
2515 |   }
2516 | }
2517 | layer {
2518 |   name: "stu_BatchNorm14"
2519 |   type: "BatchNorm"
2520 |   bottom: "stu_Convolution14"
2521 |   top: "stu_Convolution14"
2522 |   param {
2523 |     lr_mult: 0
2524 |     decay_mult: 0
2525 |   }
2526 |   param {
2527 |     lr_mult: 0
2528 |     decay_mult: 0
2529 |   }
2530 |   param {
2531 |     lr_mult: 0
2532 |     decay_mult: 0
2533 |   }
2534 | }
2535 | layer {
2536 |   name: "stu_Scale14"
2537 |   type: "Scale"
2538 |   bottom: "stu_Convolution14"
2539 |   top: "stu_Convolution14"
2540 |   scale_param {
2541 |     bias_term: true
2542 |   }
2543 | }
2544 | layer {
2545 |   name: "stu_Eltwise6"
2546 |   type: "Eltwise"
2547 |   bottom: "stu_Eltwise5"
2548 |   bottom: "stu_Convolution14"
2549 |   top: "stu_Eltwise6"
2550 |   eltwise_param {
2551 |     operation: SUM
2552 |   }
2553 | }
2554 | layer {
2555 |   name: "stu_ReLU13"
2556 |   type: "ReLU"
2557 |   bottom: "stu_Eltwise6"
2558 |   top: "stu_Eltwise6"
2559 | }
2560 | layer {
2561 |   name: "stu_Convolution15"
2562 |   type: "Convolution"
2563 |   bottom: "stu_Eltwise6"
2564 |   top: "stu_Convolution15"
2565 |   param {
2566 |     lr_mult: 1
2567 |     decay_mult: 1
2568 |   }
2569 |   param {
2570 |     lr_mult: 2
2571 |     decay_mult: 0
2572 |   }
2573 |   convolution_param {
2574 |     num_output: 32
2575 |     pad: 0
2576 |     kernel_size: 1
2577 |     stride: 2
2578 |     weight_filler {
2579 |       type: "msra"
2580 |     }
2581 |     bias_filler {
2582 |       type: "constant"
2583 |       value: 0
2584 |     }
2585 |   }
2586 | }
2587 | layer {
2588 |   name: "stu_BatchNorm15"
2589 |   type: "BatchNorm"
2590 |   bottom: "stu_Convolution15"
2591 |   top: "stu_Convolution15"
2592 |   param {
2593 |     lr_mult: 0
2594 |     decay_mult: 0
2595 |   }
2596 |   param {
2597 |     lr_mult: 0
2598 |     decay_mult: 0
2599 |   }
2600 |   param {
2601 |     lr_mult: 0
2602 |     decay_mult: 0
2603 |   }
2604 | }
2605 | layer {
2606 |   name: "stu_Scale15"
2607 |   type: "Scale"
2608 |   bottom: "stu_Convolution15"
2609 |   top: "stu_Convolution15"
2610 |   scale_param {
2611 |     bias_term: true
2612 |   }
2613 | }
2614 | layer {
2615 |   name: "stu_Convolution16"
2616 |   type: "Convolution"
2617 |   bottom: "stu_Eltwise6"
2618 |   top: "stu_Convolution16"
2619 |   param {
2620 |     lr_mult: 1
2621 |     decay_mult: 1
2622 |   }
2623 |   param {
2624 |     lr_mult: 2
2625 |     decay_mult: 0
2626 |   }
2627 |   convolution_param {
2628 |     num_output: 32
2629 |     pad: 1
2630 |     kernel_size: 3
2631 |     stride: 2
2632 |     weight_filler {
2633 |       type: "msra"
2634 |     }
2635 |     bias_filler {
2636 |       type: "constant"
2637 |       value: 0
2638 |     }
2639 |   }
2640 | }
2641 | layer {
2642 |   name: "stu_BatchNorm16"
2643 |   type: "BatchNorm"
2644 |   bottom: "stu_Convolution16"
2645 |   top: "stu_Convolution16"
2646 |   param {
2647 |     lr_mult: 0
2648 |     decay_mult: 0
2649 |   }
2650 |   param {
2651 |     lr_mult: 0
2652 |     decay_mult: 0
2653 |   }
2654 |   param {
2655 |     lr_mult: 0
2656 |     decay_mult: 0
2657 |   }
2658 | }
2659 | layer {
2660 |   name: "stu_Scale16"
2661 |   type: "Scale"
2662 |   bottom: "stu_Convolution16"
2663 |   top: "stu_Convolution16"
2664 |   scale_param {
2665 |     bias_term: true
2666 |   }
2667 | }
2668 | layer {
2669 |   name: "stu_ReLU14"
2670 |   type: "ReLU"
2671 |   bottom: "stu_Convolution16"
2672 |   top: "stu_Convolution16"
2673 | }
2674 | layer {
2675 |   name: "stu_Convolution17"
2676 |   type: "Convolution"
2677 |   bottom: "stu_Convolution16"
2678 |   top: "stu_Convolution17"
2679 |   param {
2680 |     lr_mult: 1
2681 |     decay_mult: 1
2682 |   }
2683 |   param {
2684 |     lr_mult: 2
2685 |     decay_mult: 0
2686 |   }
2687 |   convolution_param {
2688 |     num_output: 32
2689 |     pad: 1
2690 |     kernel_size: 3
2691 |     stride: 1
2692 |     weight_filler {
2693 |       type: "msra"
2694 |     }
2695 |     bias_filler {
2696 |       type: "constant"
2697 |       value: 0
2698 |     }
2699 |   }
2700 | }
2701 | layer {
2702 |   name: "stu_BatchNorm17"
2703 |   type: "BatchNorm"
2704 |   bottom: "stu_Convolution17"
2705 |   top: "stu_Convolution17"
2706 |   param {
2707 |     lr_mult: 0
2708 |     decay_mult: 0
2709 |   }
2710 |   param {
2711 |     lr_mult: 0
2712 |     decay_mult: 0
2713 |   }
2714 |   param {
2715 |     lr_mult: 0
2716 |     decay_mult: 0
2717 |   }
2718 | }
2719 | layer {
2720 |   name: "stu_Scale17"
2721 |   type: "Scale"
2722 |   bottom: "stu_Convolution17"
2723 |   top: "stu_Convolution17"
2724 |   scale_param {
2725 |     bias_term: true
2726 |   }
2727 | }
2728 | layer {
2729 |   name: "stu_Eltwise7"
2730 |   type: "Eltwise"
2731 |   bottom: "stu_Convolution15"
2732 |   bottom: "stu_Convolution17"
2733 |   top: "stu_Eltwise7"
2734 |   eltwise_param {
2735 |     operation: SUM
2736 |   }
2737 | }
2738 | layer {
2739 |   name: "stu_ReLU15"
2740 |   type: "ReLU"
2741 |   bottom: "stu_Eltwise7"
2742 |   top: "stu_Eltwise7"
2743 | }
2744 | layer {
2745 |   name: "stu_Convolution18"
2746 |   type: "Convolution"
2747 |   bottom: "stu_Eltwise7"
2748 |   top: "stu_Convolution18"
2749 |   param {
2750 |     lr_mult: 1
2751 |     decay_mult: 1
2752 |   }
2753 |   param {
2754 |     lr_mult: 2
2755 |     decay_mult: 0
2756 |   }
2757 |   convolution_param {
2758 |     num_output: 32
2759 |     pad: 1
2760 |     kernel_size: 3
2761 |     stride: 1
2762 |     weight_filler {
2763 |       type: "msra"
2764 |     }
2765 |     bias_filler {
2766 |       type: "constant"
2767 |       value: 0
2768 |     }
2769 |   }
2770 | }
2771 | layer {
2772 |   name: "stu_BatchNorm18"
2773 |   type: "BatchNorm"
2774 |   bottom: "stu_Convolution18"
2775 |   top: "stu_Convolution18"
2776 |   param {
2777 |     lr_mult: 0
2778 |     decay_mult: 0
2779 |   }
2780 |   param {
2781 |     lr_mult: 0
2782 |     decay_mult: 0
2783 |   }
2784 |   param {
2785 |     lr_mult: 0
2786 |     decay_mult: 0
2787 |   }
2788 | }
2789 | layer {
2790 |   name: "stu_Scale18"
2791 |   type: "Scale"
2792 |   bottom: "stu_Convolution18"
2793 |   top: "stu_Convolution18"
2794 |   scale_param {
2795 |     bias_term: true
2796 |   }
2797 | }
2798 | layer {
2799 |   name: "stu_ReLU16"
2800 |   type: "ReLU"
2801 |   bottom: "stu_Convolution18"
2802 |   top: "stu_Convolution18"
2803 | }
2804 | layer {
2805 |   name: "stu_Convolution19"
2806 |   type: "Convolution"
2807 |   bottom: "stu_Convolution18"
2808 |   top: "stu_Convolution19"
2809 |   param {
2810 |     lr_mult: 1
2811 |     decay_mult: 1
2812 |   }
2813 |   param {
2814 |     lr_mult: 2
2815 |     decay_mult: 0
2816 |   }
2817 |   convolution_param {
2818 |     num_output: 32
2819 |     pad: 1
2820 |     kernel_size: 3
2821 |     stride: 1
2822 |     weight_filler {
2823 |       type: "msra"
2824 |     }
2825 |     bias_filler {
2826 |       type: "constant"
2827 |       value: 0
2828 |     }
2829 |   }
2830 | }
2831 | layer {
2832 |   name: "stu_BatchNorm19"
2833 |   type: "BatchNorm"
2834 |   bottom: "stu_Convolution19"
2835 |   top: "stu_Convolution19"
2836 |   param {
2837 |     lr_mult: 0
2838 |     decay_mult: 0
2839 |   }
2840 |   param {
2841 |     lr_mult: 0
2842 |     decay_mult: 0
2843 |   }
2844 |   param {
2845 |     lr_mult: 0
2846 |     decay_mult: 0
2847 |   }
2848 | }
2849 | layer {
2850 |   name: "stu_Scale19"
2851 |   type: "Scale"
2852 |   bottom: "stu_Convolution19"
2853 |   top: "stu_Convolution19"
2854 |   scale_param {
2855 |     bias_term: true
2856 |   }
2857 | }
2858 | layer {
2859 |   name: "stu_Eltwise8"
2860 |   type: "Eltwise"
2861 |   bottom: "stu_Eltwise7"
2862 |   bottom: "stu_Convolution19"
2863 |   top: "stu_Eltwise8"
2864 |   eltwise_param {
2865 |     operation: SUM
2866 |   }
2867 | }
2868 | layer {
2869 |   name: "stu_ReLU17"
2870 |   type: "ReLU"
2871 |   bottom: "stu_Eltwise8"
2872 |   top: "stu_Eltwise8"
2873 | }
2874 | layer {
2875 |   name: "stu_Convolution20"
2876 |   type: "Convolution"
2877 |   bottom: "stu_Eltwise8"
2878 |   top: "stu_Convolution20"
2879 |   param {
2880 |     lr_mult: 1
2881 |     decay_mult: 1
2882 |   }
2883 |   param {
2884 |     lr_mult: 2
2885 |     decay_mult: 0
2886 |   }
2887 |   convolution_param {
2888 |     num_output: 32
2889 |     pad: 1
2890 |     kernel_size: 3
2891 |     stride: 1
2892 |     weight_filler {
2893 |       type: "msra"
2894 |     }
2895 |     bias_filler {
2896 |       type: "constant"
2897 |       value: 0
2898 |     }
2899 |   }
2900 | }
2901 | layer {
2902 |   name: "stu_BatchNorm20"
2903 |   type: "BatchNorm"
2904 |   bottom: "stu_Convolution20"
2905 |   top: "stu_Convolution20"
2906 |   param {
2907 |     lr_mult: 0
2908 |     decay_mult: 0
2909 |   }
2910 |   param {
2911 |     lr_mult: 0
2912 |     decay_mult: 0
2913 |   }
2914 |   param {
2915 |     lr_mult: 0
2916 |     decay_mult: 0
2917 |   }
2918 | }
2919 | layer {
2920 |   name: "stu_Scale20"
2921 |   type: "Scale"
2922 |   bottom: "stu_Convolution20"
2923 |   top: "stu_Convolution20"
2924 |   scale_param {
2925 |     bias_term: true
2926 |   }
2927 | }
2928 | layer {
2929 |   name: "stu_ReLU18"
2930 |   type: "ReLU"
2931 |   bottom: "stu_Convolution20"
2932 |   top: "stu_Convolution20"
2933 | }
2934 | layer {
2935 |   name: "stu_Convolution21"
2936 |   type: "Convolution"
2937 |   bottom: "stu_Convolution20"
2938 |   top: "stu_Convolution21"
2939 |   param {
2940 |     lr_mult: 1
2941 |     decay_mult: 1
2942 |   }
2943 |   param {
2944 |     lr_mult: 2
2945 |     decay_mult: 0
2946 |   }
2947 |   convolution_param {
2948 |     num_output: 32
2949 |     pad: 1
2950 |     kernel_size: 3
2951 |     stride: 1
2952 |     weight_filler {
2953 |       type: "msra"
2954 |     }
2955 |     bias_filler {
2956 |       type: "constant"
2957 |       value: 0
2958 |     }
2959 |   }
2960 | }
2961 | layer {
2962 |   name: "stu_BatchNorm21"
2963 |   type: "BatchNorm"
2964 |   bottom: "stu_Convolution21"
2965 |   top: "stu_Convolution21"
2966 |   param {
2967 |     lr_mult: 0
2968 |     decay_mult: 0
2969 |   }
2970 |   param {
2971 |     lr_mult: 0
2972 |     decay_mult: 0
2973 |   }
2974 |   param {
2975 |     lr_mult: 0
2976 |     decay_mult: 0
2977 |   }
2978 | }
2979 | layer {
2980 |   name: "stu_Scale21"
2981 |   type: "Scale"
2982 |   bottom: "stu_Convolution21"
2983 |   top: "stu_Convolution21"
2984 |   scale_param {
2985 |     bias_term: true
2986 |   }
2987 | }
2988 | layer {
2989 |   name: "stu_Eltwise9"
2990 |   type: "Eltwise"
2991 |   bottom: "stu_Eltwise8"
2992 |   bottom: "stu_Convolution21"
2993 |   top: "stu_Eltwise9"
2994 |   eltwise_param {
2995 |     operation: SUM
2996 |   }
2997 | }
2998 | layer {
2999 |   name: "stu_ReLU19"
3000 |   type: "ReLU"
3001 |   bottom: "stu_Eltwise9"
3002 |   top: "stu_Eltwise9"
3003 | }
3004 | layer {
3005 |   name: "stu_Pooling1"
3006 |   type: "Pooling"
3007 |   bottom: "stu_Eltwise9"
3008 |   top: "stu_Pooling1"
3009 |   pooling_param {
3010 |     pool: AVE
3011 |     global_pooling: true
3012 |   }
3013 | }
3014 | layer {
3015 |   name: "stu_bem"
3016 |   type: "BatchEuclideanMap"
3017 |   bottom: "stu_Pooling1"
3018 |   top: "stu_bem"
3019 | }
3020 | layer {
3021 |   name: "bem_loss"
3022 |   type: "EuclideanLoss"
3023 |   bottom: "bem"
3024 |   bottom: "stu_bem"
3025 |   top: "bem_loss"
3026 |   loss_weight: 0.005
3027 | }
3028 | layer {
3029 |   name: "stu_InnerProduct1"
3030 |   type: "InnerProduct"
3031 |   bottom: "stu_Pooling1"
3032 |   top: "stu_InnerProduct1"
3033 |   param {
3034 |     lr_mult: 1
3035 |     decay_mult: 1
3036 |   }
3037 |   param {
3038 |     lr_mult: 2
3039 |     decay_mult: 1
3040 |   }
3041 |   inner_product_param {
3042 |     num_output: 10
3043 |     weight_filler {
3044 |       type: "msra"
3045 |     }
3046 |     bias_filler {
3047 |       type: "constant"
3048 |       value: 0
3049 |     }
3050 |   }
3051 | }
3052 | layer {
3053 |   name: "stu_SoftmaxWithLoss1"
3054 |   type: "SoftmaxWithLoss"
3055 |   bottom: "stu_InnerProduct1"
3056 |   bottom: "Data2"
3057 |   top: "stu_SoftmaxWithLoss1"
3058 | }
3059 | layer {
3060 |   name: "stu_Accuracy1"
3061 |   type: "Accuracy"
3062 |   bottom: "stu_InnerProduct1"
3063 |   bottom: "Data2"
3064 |   top: "stu_Accuracy1"
3065 |   include {
3066 |     phase: TEST
3067 |   }
3068 | }
3069 | 


--------------------------------------------------------------------------------
/IRG_Transformation.prototxt:
--------------------------------------------------------------------------------
   1 | name: "resnet_cifar10_IRGTransformation"
   2 | layer {
   3 |   name: "Data1"
   4 |   type: "Data"
   5 |   top: "Data1"
   6 |   top: "Data2"
   7 |   include {
   8 |     phase: TRAIN
   9 |   }
  10 |   transform_param {
  11 |     mean_file: "yourpath/mean_pad.binaryproto"
  12 |     crop_size: 32
  13 |     mirror:true
  14 |   }
  15 |   data_param {
  16 |     source: "yourpath/cifar-10-train_lmdb"
  17 |     batch_size: 64 
  18 |     backend: LMDB
  19 |   }
  20 | }
  21 | layer {
  22 |   name: "Data1"
  23 |   type: "Data"
  24 |   top: "Data1"
  25 |   top: "Data2"
  26 |   include {
  27 |     phase: TEST
  28 |   }
  29 |   transform_param {
  30 |     mean_file: "yourpath/mean.binaryproto"
  31 |   }
  32 |   data_param {
  33 |     source: "yourpath/cifar10_test_lmdb"
  34 |     batch_size: 100
  35 |     backend: LMDB
  36 |   }
  37 | }
  38 | layer {
  39 |   name: "Convolution1"
  40 |   type: "Convolution"
  41 |   bottom: "Data1"
  42 |   top: "Convolution1"
  43 |   param {
  44 |     lr_mult: 0
  45 |     decay_mult: 0
  46 |   }
  47 |   param {
  48 |     lr_mult: 0
  49 |     decay_mult: 0
  50 |   }
  51 |   convolution_param {
  52 |     num_output: 16
  53 |     pad: 1
  54 |     kernel_size: 3
  55 |     stride: 1
  56 |     weight_filler {
  57 |       type: "msra"
  58 |     }
  59 |     bias_filler {
  60 |       type: "constant"
  61 |       value: 0
  62 |     }
  63 |   }
  64 | }
  65 | layer {
  66 |   name: "BatchNorm1"
  67 |   type: "BatchNorm"
  68 |   bottom: "Convolution1"
  69 |   top: "Convolution1"
  70 |   param {
  71 |     lr_mult: 0
  72 |     decay_mult: 0
  73 |   }
  74 |   param {
  75 |     lr_mult: 0
  76 |     decay_mult: 0
  77 |   }
  78 |   param {
  79 |     lr_mult: 0
  80 |     decay_mult: 0
  81 |   }
  82 |   batch_norm_param {
  83 |     use_global_stats: true
  84 |   }
  85 | }
  86 | layer {
  87 |   name: "Scale1"
  88 |   type: "Scale"
  89 |   bottom: "Convolution1"
  90 |   top: "Convolution1"
  91 |   param {
  92 |     lr_mult: 0
  93 |     decay_mult: 0
  94 |   }
  95 |   param {
  96 |     lr_mult: 0
  97 |     decay_mult: 0
  98 |   }
  99 |   scale_param {
 100 |     bias_term: true
 101 |   }
 102 | }
 103 | layer {
 104 |   name: "ReLU1"
 105 |   type: "ReLU"
 106 |   bottom: "Convolution1"
 107 |   top: "Convolution1"
 108 | }
 109 | layer {
 110 |   name: "Convolution2"
 111 |   type: "Convolution"
 112 |   bottom: "Convolution1"
 113 |   top: "Convolution2"
 114 |   param {
 115 |     lr_mult: 0
 116 |     decay_mult: 0
 117 |   }
 118 |   param {
 119 |     lr_mult: 0
 120 |     decay_mult: 0
 121 |   }
 122 |   convolution_param {
 123 |     num_output: 16
 124 |     pad: 1
 125 |     kernel_size: 3
 126 |     stride: 1
 127 |     weight_filler {
 128 |       type: "msra"
 129 |     }
 130 |     bias_filler {
 131 |       type: "constant"
 132 |       value: 0
 133 |     }
 134 |   }
 135 | }
 136 | layer {
 137 |   name: "BatchNorm2"
 138 |   type: "BatchNorm"
 139 |   bottom: "Convolution2"
 140 |   top: "Convolution2"
 141 |   param {
 142 |     lr_mult: 0
 143 |     decay_mult: 0
 144 |   }
 145 |   param {
 146 |     lr_mult: 0
 147 |     decay_mult: 0
 148 |   }
 149 |   param {
 150 |     lr_mult: 0
 151 |     decay_mult: 0
 152 |   }
 153 |   batch_norm_param {
 154 |     use_global_stats: true
 155 |   }
 156 | }
 157 | layer {
 158 |   name: "Scale2"
 159 |   type: "Scale"
 160 |   bottom: "Convolution2"
 161 |   top: "Convolution2"
 162 |   param {
 163 |     lr_mult: 0
 164 |     decay_mult: 0
 165 |   }
 166 |   param {
 167 |     lr_mult: 0
 168 |     decay_mult: 0
 169 |   }
 170 |   scale_param {
 171 |     bias_term: true
 172 |   }
 173 | }
 174 | layer {
 175 |   name: "ReLU2"
 176 |   type: "ReLU"
 177 |   bottom: "Convolution2"
 178 |   top: "Convolution2"
 179 | }
 180 | layer {
 181 |   name: "Convolution3"
 182 |   type: "Convolution"
 183 |   bottom: "Convolution2"
 184 |   top: "Convolution3"
 185 |   param {
 186 |     lr_mult: 0
 187 |     decay_mult: 0
 188 |   }
 189 |   param {
 190 |     lr_mult: 0
 191 |     decay_mult: 0
 192 |   }
 193 |   convolution_param {
 194 |     num_output: 16
 195 |     pad: 1
 196 |     kernel_size: 3
 197 |     stride: 1
 198 |     weight_filler {
 199 |       type: "msra"
 200 |     }
 201 |     bias_filler {
 202 |       type: "constant"
 203 |       value: 0
 204 |     }
 205 |   }
 206 | }
 207 | layer {
 208 |   name: "BatchNorm3"
 209 |   type: "BatchNorm"
 210 |   bottom: "Convolution3"
 211 |   top: "Convolution3"
 212 |   param {
 213 |     lr_mult: 0
 214 |     decay_mult: 0
 215 |   }
 216 |   param {
 217 |     lr_mult: 0
 218 |     decay_mult: 0
 219 |   }
 220 |   param {
 221 |     lr_mult: 0
 222 |     decay_mult: 0
 223 |   }
 224 |   batch_norm_param {
 225 |     use_global_stats: true
 226 |   }
 227 | }
 228 | layer {
 229 |   name: "Scale3"
 230 |   type: "Scale"
 231 |   bottom: "Convolution3"
 232 |   top: "Convolution3"
 233 |   param {
 234 |     lr_mult: 0
 235 |     decay_mult: 0
 236 |   }
 237 |   param {
 238 |     lr_mult: 0
 239 |     decay_mult: 0
 240 |   }
 241 |   scale_param {
 242 |     bias_term: true
 243 |   }
 244 | }
 245 | layer {
 246 |   name: "Eltwise1"
 247 |   type: "Eltwise"
 248 |   bottom: "Convolution1"
 249 |   bottom: "Convolution3"
 250 |   top: "Eltwise1"
 251 |   eltwise_param {
 252 |     operation: SUM
 253 |   }
 254 | }
 255 | layer {
 256 |   name: "ReLU3"
 257 |   type: "ReLU"
 258 |   bottom: "Eltwise1"
 259 |   top: "Eltwise1"
 260 | }
 261 | layer {
 262 |   name: "Convolution4"
 263 |   type: "Convolution"
 264 |   bottom: "Eltwise1"
 265 |   top: "Convolution4"
 266 |   param {
 267 |     lr_mult: 0
 268 |     decay_mult: 0
 269 |   }
 270 |   param {
 271 |     lr_mult: 0
 272 |     decay_mult: 0
 273 |   }
 274 |   convolution_param {
 275 |     num_output: 16
 276 |     pad: 1
 277 |     kernel_size: 3
 278 |     stride: 1
 279 |     weight_filler {
 280 |       type: "msra"
 281 |     }
 282 |     bias_filler {
 283 |       type: "constant"
 284 |       value: 0
 285 |     }
 286 |   }
 287 | }
 288 | layer {
 289 |   name: "BatchNorm4"
 290 |   type: "BatchNorm"
 291 |   bottom: "Convolution4"
 292 |   top: "Convolution4"
 293 |   param {
 294 |     lr_mult: 0
 295 |     decay_mult: 0
 296 |   }
 297 |   param {
 298 |     lr_mult: 0
 299 |     decay_mult: 0
 300 |   }
 301 |   param {
 302 |     lr_mult: 0
 303 |     decay_mult: 0
 304 |   }
 305 |   batch_norm_param {
 306 |     use_global_stats: true
 307 |   }
 308 | }
 309 | layer {
 310 |   name: "Scale4"
 311 |   type: "Scale"
 312 |   bottom: "Convolution4"
 313 |   top: "Convolution4"
 314 |   param {
 315 |     lr_mult: 0
 316 |     decay_mult: 0
 317 |   }
 318 |   param {
 319 |     lr_mult: 0
 320 |     decay_mult: 0
 321 |   }
 322 |   scale_param {
 323 |     bias_term: true
 324 |   }
 325 | }
 326 | layer {
 327 |   name: "ReLU4"
 328 |   type: "ReLU"
 329 |   bottom: "Convolution4"
 330 |   top: "Convolution4"
 331 | }
 332 | layer {
 333 |   name: "Convolution5"
 334 |   type: "Convolution"
 335 |   bottom: "Convolution4"
 336 |   top: "Convolution5"
 337 |   param {
 338 |     lr_mult: 0
 339 |     decay_mult: 0
 340 |   }
 341 |   param {
 342 |     lr_mult: 0
 343 |     decay_mult: 0
 344 |   }
 345 |   convolution_param {
 346 |     num_output: 16
 347 |     pad: 1
 348 |     kernel_size: 3
 349 |     stride: 1
 350 |     weight_filler {
 351 |       type: "msra"
 352 |     }
 353 |     bias_filler {
 354 |       type: "constant"
 355 |       value: 0
 356 |     }
 357 |   }
 358 | }
 359 | layer {
 360 |   name: "BatchNorm5"
 361 |   type: "BatchNorm"
 362 |   bottom: "Convolution5"
 363 |   top: "Convolution5"
 364 |   param {
 365 |     lr_mult: 0
 366 |     decay_mult: 0
 367 |   }
 368 |   param {
 369 |     lr_mult: 0
 370 |     decay_mult: 0
 371 |   }
 372 |   param {
 373 |     lr_mult: 0
 374 |     decay_mult: 0
 375 |   }
 376 |   batch_norm_param {
 377 |     use_global_stats: true
 378 |   }
 379 | }
 380 | layer {
 381 |   name: "Scale5"
 382 |   type: "Scale"
 383 |   bottom: "Convolution5"
 384 |   top: "Convolution5"
 385 |   param {
 386 |     lr_mult: 0
 387 |     decay_mult: 0
 388 |   }
 389 |   param {
 390 |     lr_mult: 0
 391 |     decay_mult: 0
 392 |   }
 393 |   scale_param {
 394 |     bias_term: true
 395 |   }
 396 | }
 397 | layer {
 398 |   name: "Eltwise2"
 399 |   type: "Eltwise"
 400 |   bottom: "Eltwise1"
 401 |   bottom: "Convolution5"
 402 |   top: "Eltwise2"
 403 |   eltwise_param {
 404 |     operation: SUM
 405 |   }
 406 | }
 407 | layer {
 408 |   name: "ReLU5"
 409 |   type: "ReLU"
 410 |   bottom: "Eltwise2"
 411 |   top: "Eltwise2"
 412 | }
 413 | layer {
 414 |   name: "Convolution6"
 415 |   type: "Convolution"
 416 |   bottom: "Eltwise2"
 417 |   top: "Convolution6"
 418 |   param {
 419 |     lr_mult: 0
 420 |     decay_mult: 0
 421 |   }
 422 |   param {
 423 |     lr_mult: 0
 424 |     decay_mult: 0
 425 |   }
 426 |   convolution_param {
 427 |     num_output: 16
 428 |     pad: 1
 429 |     kernel_size: 3
 430 |     stride: 1
 431 |     weight_filler {
 432 |       type: "msra"
 433 |     }
 434 |     bias_filler {
 435 |       type: "constant"
 436 |       value: 0
 437 |     }
 438 |   }
 439 | }
 440 | layer {
 441 |   name: "BatchNorm6"
 442 |   type: "BatchNorm"
 443 |   bottom: "Convolution6"
 444 |   top: "Convolution6"
 445 |   param {
 446 |     lr_mult: 0
 447 |     decay_mult: 0
 448 |   }
 449 |   param {
 450 |     lr_mult: 0
 451 |     decay_mult: 0
 452 |   }
 453 |   param {
 454 |     lr_mult: 0
 455 |     decay_mult: 0
 456 |   }
 457 |   batch_norm_param {
 458 |     use_global_stats: true
 459 |   }
 460 | }
 461 | layer {
 462 |   name: "Scale6"
 463 |   type: "Scale"
 464 |   bottom: "Convolution6"
 465 |   top: "Convolution6"
 466 |   param {
 467 |     lr_mult: 0
 468 |     decay_mult: 0
 469 |   }
 470 |   param {
 471 |     lr_mult: 0
 472 |     decay_mult: 0
 473 |   }
 474 |   scale_param {
 475 |     bias_term: true
 476 |   }
 477 | }
 478 | layer {
 479 |   name: "ReLU6"
 480 |   type: "ReLU"
 481 |   bottom: "Convolution6"
 482 |   top: "Convolution6"
 483 | }
 484 | layer {
 485 |   name: "Convolution7"
 486 |   type: "Convolution"
 487 |   bottom: "Convolution6"
 488 |   top: "Convolution7"
 489 |   param {
 490 |     lr_mult: 0
 491 |     decay_mult: 0
 492 |   }
 493 |   param {
 494 |     lr_mult: 0
 495 |     decay_mult: 0
 496 |   }
 497 |   convolution_param {
 498 |     num_output: 16
 499 |     pad: 1
 500 |     kernel_size: 3
 501 |     stride: 1
 502 |     weight_filler {
 503 |       type: "msra"
 504 |     }
 505 |     bias_filler {
 506 |       type: "constant"
 507 |       value: 0
 508 |     }
 509 |   }
 510 | }
 511 | layer {
 512 |   name: "BatchNorm7"
 513 |   type: "BatchNorm"
 514 |   bottom: "Convolution7"
 515 |   top: "Convolution7"
 516 |   param {
 517 |     lr_mult: 0
 518 |     decay_mult: 0
 519 |   }
 520 |   param {
 521 |     lr_mult: 0
 522 |     decay_mult: 0
 523 |   }
 524 |   param {
 525 |     lr_mult: 0
 526 |     decay_mult: 0
 527 |   }
 528 |   batch_norm_param {
 529 |     use_global_stats: true
 530 |   }
 531 | }
 532 | layer {
 533 |   name: "Scale7"
 534 |   type: "Scale"
 535 |   bottom: "Convolution7"
 536 |   top: "Convolution7"
 537 |   param {
 538 |     lr_mult: 0
 539 |     decay_mult: 0
 540 |   }
 541 |   param {
 542 |     lr_mult: 0
 543 |     decay_mult: 0
 544 |   }
 545 |   scale_param {
 546 |     bias_term: true
 547 |   }
 548 | }
 549 | layer {
 550 |   name: "Eltwise3"
 551 |   type: "Eltwise"
 552 |   bottom: "Eltwise2"
 553 |   bottom: "Convolution7"
 554 |   top: "Eltwise3"
 555 |   eltwise_param {
 556 |     operation: SUM
 557 |   }
 558 | }
 559 | layer {
 560 |   name: "ReLU7"
 561 |   type: "ReLU"
 562 |   bottom: "Eltwise3"
 563 |   top: "Eltwise3"
 564 | }
 565 | layer {
 566 |   name: "Convolution8"
 567 |   type: "Convolution"
 568 |   bottom: "Eltwise3"
 569 |   top: "Convolution8"
 570 |   param {
 571 |     lr_mult: 0
 572 |     decay_mult: 0
 573 |   }
 574 |   param {
 575 |     lr_mult: 0
 576 |     decay_mult: 0
 577 |   }
 578 |   convolution_param {
 579 |     num_output: 32
 580 |     pad: 0
 581 |     kernel_size: 1
 582 |     stride: 2
 583 |     weight_filler {
 584 |       type: "msra"
 585 |     }
 586 |     bias_filler {
 587 |       type: "constant"
 588 |       value: 0
 589 |     }
 590 |   }
 591 | }
 592 | layer {
 593 |   name: "BatchNorm8"
 594 |   type: "BatchNorm"
 595 |   bottom: "Convolution8"
 596 |   top: "Convolution8"
 597 |   param {
 598 |     lr_mult: 0
 599 |     decay_mult: 0
 600 |   }
 601 |   param {
 602 |     lr_mult: 0
 603 |     decay_mult: 0
 604 |   }
 605 |   param {
 606 |     lr_mult: 0
 607 |     decay_mult: 0
 608 |   }
 609 |   batch_norm_param {
 610 |     use_global_stats: true
 611 |   }
 612 | }
 613 | layer {
 614 |   name: "Scale8"
 615 |   type: "Scale"
 616 |   bottom: "Convolution8"
 617 |   top: "Convolution8"
 618 |   param {
 619 |     lr_mult: 0
 620 |     decay_mult: 0
 621 |   }
 622 |   param {
 623 |     lr_mult: 0
 624 |     decay_mult: 0
 625 |   }
 626 |   scale_param {
 627 |     bias_term: true
 628 |   }
 629 | }
 630 | layer {
 631 |   name: "Convolution9"
 632 |   type: "Convolution"
 633 |   bottom: "Eltwise3"
 634 |   top: "Convolution9"
 635 |   param {
 636 |     lr_mult: 0
 637 |     decay_mult: 0
 638 |   }
 639 |   param {
 640 |     lr_mult: 0
 641 |     decay_mult: 0
 642 |   }
 643 |   convolution_param {
 644 |     num_output: 32
 645 |     pad: 1
 646 |     kernel_size: 3
 647 |     stride: 2
 648 |     weight_filler {
 649 |       type: "msra"
 650 |     }
 651 |     bias_filler {
 652 |       type: "constant"
 653 |       value: 0
 654 |     }
 655 |   }
 656 | }
 657 | layer {
 658 |   name: "BatchNorm9"
 659 |   type: "BatchNorm"
 660 |   bottom: "Convolution9"
 661 |   top: "Convolution9"
 662 |   param {
 663 |     lr_mult: 0
 664 |     decay_mult: 0
 665 |   }
 666 |   param {
 667 |     lr_mult: 0
 668 |     decay_mult: 0
 669 |   }
 670 |   param {
 671 |     lr_mult: 0
 672 |     decay_mult: 0
 673 |   }
 674 |   batch_norm_param {
 675 |     use_global_stats: true
 676 |   }
 677 | }
 678 | layer {
 679 |   name: "Scale9"
 680 |   type: "Scale"
 681 |   bottom: "Convolution9"
 682 |   top: "Convolution9"
 683 |   param {
 684 |     lr_mult: 0
 685 |     decay_mult: 0
 686 |   }
 687 |   param {
 688 |     lr_mult: 0
 689 |     decay_mult: 0
 690 |   }
 691 |   scale_param {
 692 |     bias_term: true
 693 |   }
 694 | }
 695 | layer {
 696 |   name: "ReLU8"
 697 |   type: "ReLU"
 698 |   bottom: "Convolution9"
 699 |   top: "Convolution9"
 700 | }
 701 | layer {
 702 |   name: "Convolution10"
 703 |   type: "Convolution"
 704 |   bottom: "Convolution9"
 705 |   top: "Convolution10"
 706 |   param {
 707 |     lr_mult: 0
 708 |     decay_mult: 0
 709 |   }
 710 |   param {
 711 |     lr_mult: 0
 712 |     decay_mult: 0
 713 |   }
 714 |   convolution_param {
 715 |     num_output: 32
 716 |     pad: 1
 717 |     kernel_size: 3
 718 |     stride: 1
 719 |     weight_filler {
 720 |       type: "msra"
 721 |     }
 722 |     bias_filler {
 723 |       type: "constant"
 724 |       value: 0
 725 |     }
 726 |   }
 727 | }
 728 | layer {
 729 |   name: "BatchNorm10"
 730 |   type: "BatchNorm"
 731 |   bottom: "Convolution10"
 732 |   top: "Convolution10"
 733 |   param {
 734 |     lr_mult: 0
 735 |     decay_mult: 0
 736 |   }
 737 |   param {
 738 |     lr_mult: 0
 739 |     decay_mult: 0
 740 |   }
 741 |   param {
 742 |     lr_mult: 0
 743 |     decay_mult: 0
 744 |   }
 745 |   batch_norm_param {
 746 |     use_global_stats: true
 747 |   }
 748 | }
 749 | layer {
 750 |   name: "Scale10"
 751 |   type: "Scale"
 752 |   bottom: "Convolution10"
 753 |   top: "Convolution10"
 754 |   param {
 755 |     lr_mult: 0
 756 |     decay_mult: 0
 757 |   }
 758 |   param {
 759 |     lr_mult: 0
 760 |     decay_mult: 0
 761 |   }
 762 |   scale_param {
 763 |     bias_term: true
 764 |   }
 765 | }
 766 | layer {
 767 |   name: "Eltwise4"
 768 |   type: "Eltwise"
 769 |   bottom: "Convolution8"
 770 |   bottom: "Convolution10"
 771 |   top: "Eltwise4"
 772 |   eltwise_param {
 773 |     operation: SUM
 774 |   }
 775 | }
 776 | layer {
 777 |   name: "ReLU9"
 778 |   type: "ReLU"
 779 |   bottom: "Eltwise4"
 780 |   top: "Eltwise4"
 781 | }
 782 | layer {
 783 |   name: "Convolution11"
 784 |   type: "Convolution"
 785 |   bottom: "Eltwise4"
 786 |   top: "Convolution11"
 787 |   param {
 788 |     lr_mult: 0
 789 |     decay_mult: 0
 790 |   }
 791 |   param {
 792 |     lr_mult: 0
 793 |     decay_mult: 0
 794 |   }
 795 |   convolution_param {
 796 |     num_output: 32
 797 |     pad: 1
 798 |     kernel_size: 3
 799 |     stride: 1
 800 |     weight_filler {
 801 |       type: "msra"
 802 |     }
 803 |     bias_filler {
 804 |       type: "constant"
 805 |       value: 0
 806 |     }
 807 |   }
 808 | }
 809 | layer {
 810 |   name: "BatchNorm11"
 811 |   type: "BatchNorm"
 812 |   bottom: "Convolution11"
 813 |   top: "Convolution11"
 814 |   param {
 815 |     lr_mult: 0
 816 |     decay_mult: 0
 817 |   }
 818 |   param {
 819 |     lr_mult: 0
 820 |     decay_mult: 0
 821 |   }
 822 |   param {
 823 |     lr_mult: 0
 824 |     decay_mult: 0
 825 |   }
 826 |   batch_norm_param {
 827 |     use_global_stats: true
 828 |   }
 829 | }
 830 | layer {
 831 |   name: "Scale11"
 832 |   type: "Scale"
 833 |   bottom: "Convolution11"
 834 |   top: "Convolution11"
 835 |   param {
 836 |     lr_mult: 0
 837 |     decay_mult: 0
 838 |   }
 839 |   param {
 840 |     lr_mult: 0
 841 |     decay_mult: 0
 842 |   }
 843 |   scale_param {
 844 |     bias_term: true
 845 |   }
 846 | }
 847 | layer {
 848 |   name: "ReLU10"
 849 |   type: "ReLU"
 850 |   bottom: "Convolution11"
 851 |   top: "Convolution11"
 852 | }
 853 | layer {
 854 |   name: "Convolution12"
 855 |   type: "Convolution"
 856 |   bottom: "Convolution11"
 857 |   top: "Convolution12"
 858 |   param {
 859 |     lr_mult: 0
 860 |     decay_mult: 0
 861 |   }
 862 |   param {
 863 |     lr_mult: 0
 864 |     decay_mult: 0
 865 |   }
 866 |   convolution_param {
 867 |     num_output: 32
 868 |     pad: 1
 869 |     kernel_size: 3
 870 |     stride: 1
 871 |     weight_filler {
 872 |       type: "msra"
 873 |     }
 874 |     bias_filler {
 875 |       type: "constant"
 876 |       value: 0
 877 |     }
 878 |   }
 879 | }
 880 | layer {
 881 |   name: "BatchNorm12"
 882 |   type: "BatchNorm"
 883 |   bottom: "Convolution12"
 884 |   top: "Convolution12"
 885 |   param {
 886 |     lr_mult: 0
 887 |     decay_mult: 0
 888 |   }
 889 |   param {
 890 |     lr_mult: 0
 891 |     decay_mult: 0
 892 |   }
 893 |   param {
 894 |     lr_mult: 0
 895 |     decay_mult: 0
 896 |   }
 897 |   batch_norm_param {
 898 |     use_global_stats: true
 899 |   }
 900 | }
 901 | layer {
 902 |   name: "Scale12"
 903 |   type: "Scale"
 904 |   bottom: "Convolution12"
 905 |   top: "Convolution12"
 906 |   param {
 907 |     lr_mult: 0
 908 |     decay_mult: 0
 909 |   }
 910 |   param {
 911 |     lr_mult: 0
 912 |     decay_mult: 0
 913 |   }
 914 |   scale_param {
 915 |     bias_term: true
 916 |   }
 917 | }
 918 | layer {
 919 |   name: "Eltwise5"
 920 |   type: "Eltwise"
 921 |   bottom: "Eltwise4"
 922 |   bottom: "Convolution12"
 923 |   top: "Eltwise5"
 924 |   eltwise_param {
 925 |     operation: SUM
 926 |   }
 927 | }
 928 | layer {
 929 |   name: "ReLU11"
 930 |   type: "ReLU"
 931 |   bottom: "Eltwise5"
 932 |   top: "Eltwise5"
 933 | }
 934 | layer {
 935 |   name: "Convolution13"
 936 |   type: "Convolution"
 937 |   bottom: "Eltwise5"
 938 |   top: "Convolution13"
 939 |   param {
 940 |     lr_mult: 0
 941 |     decay_mult: 0
 942 |   }
 943 |   param {
 944 |     lr_mult: 0
 945 |     decay_mult: 0
 946 |   }
 947 |   convolution_param {
 948 |     num_output: 32
 949 |     pad: 1
 950 |     kernel_size: 3
 951 |     stride: 1
 952 |     weight_filler {
 953 |       type: "msra"
 954 |     }
 955 |     bias_filler {
 956 |       type: "constant"
 957 |       value: 0
 958 |     }
 959 |   }
 960 | }
 961 | layer {
 962 |   name: "BatchNorm13"
 963 |   type: "BatchNorm"
 964 |   bottom: "Convolution13"
 965 |   top: "Convolution13"
 966 |   param {
 967 |     lr_mult: 0
 968 |     decay_mult: 0
 969 |   }
 970 |   param {
 971 |     lr_mult: 0
 972 |     decay_mult: 0
 973 |   }
 974 |   param {
 975 |     lr_mult: 0
 976 |     decay_mult: 0
 977 |   }
 978 |   batch_norm_param {
 979 |     use_global_stats: true
 980 |   }
 981 | }
 982 | layer {
 983 |   name: "Scale13"
 984 |   type: "Scale"
 985 |   bottom: "Convolution13"
 986 |   top: "Convolution13"
 987 |   param {
 988 |     lr_mult: 0
 989 |     decay_mult: 0
 990 |   }
 991 |   param {
 992 |     lr_mult: 0
 993 |     decay_mult: 0
 994 |   }
 995 |   scale_param {
 996 |     bias_term: true
 997 |   }
 998 | }
 999 | layer {
1000 |   name: "ReLU12"
1001 |   type: "ReLU"
1002 |   bottom: "Convolution13"
1003 |   top: "Convolution13"
1004 | }
1005 | layer {
1006 |   name: "Convolution14"
1007 |   type: "Convolution"
1008 |   bottom: "Convolution13"
1009 |   top: "Convolution14"
1010 |   param {
1011 |     lr_mult: 0
1012 |     decay_mult: 0
1013 |   }
1014 |   param {
1015 |     lr_mult: 0
1016 |     decay_mult: 0
1017 |   }
1018 |   convolution_param {
1019 |     num_output: 32
1020 |     pad: 1
1021 |     kernel_size: 3
1022 |     stride: 1
1023 |     weight_filler {
1024 |       type: "msra"
1025 |     }
1026 |     bias_filler {
1027 |       type: "constant"
1028 |       value: 0
1029 |     }
1030 |   }
1031 | }
1032 | layer {
1033 |   name: "BatchNorm14"
1034 |   type: "BatchNorm"
1035 |   bottom: "Convolution14"
1036 |   top: "Convolution14"
1037 |   param {
1038 |     lr_mult: 0
1039 |     decay_mult: 0
1040 |   }
1041 |   param {
1042 |     lr_mult: 0
1043 |     decay_mult: 0
1044 |   }
1045 |   param {
1046 |     lr_mult: 0
1047 |     decay_mult: 0
1048 |   }
1049 |   batch_norm_param {
1050 |     use_global_stats: true
1051 |   }
1052 | }
1053 | layer {
1054 |   name: "Scale14"
1055 |   type: "Scale"
1056 |   bottom: "Convolution14"
1057 |   top: "Convolution14"
1058 |   param {
1059 |     lr_mult: 0
1060 |     decay_mult: 0
1061 |   }
1062 |   param {
1063 |     lr_mult: 0
1064 |     decay_mult: 0
1065 |   }
1066 |   scale_param {
1067 |     bias_term: true
1068 |   }
1069 | }
1070 | layer {
1071 |   name: "Eltwise6"
1072 |   type: "Eltwise"
1073 |   bottom: "Eltwise5"
1074 |   bottom: "Convolution14"
1075 |   top: "Eltwise6"
1076 |   eltwise_param {
1077 |     operation: SUM
1078 |   }
1079 | }
1080 | layer {
1081 |   name: "ReLU13"
1082 |   type: "ReLU"
1083 |   bottom: "Eltwise6"
1084 |   top: "Eltwise6"
1085 | }
1086 | layer {
1087 |   name: "Convolution15"
1088 |   type: "Convolution"
1089 |   bottom: "Eltwise6"
1090 |   top: "Convolution15"
1091 |   param {
1092 |     lr_mult: 0
1093 |     decay_mult: 0
1094 |   }
1095 |   param {
1096 |     lr_mult: 0
1097 |     decay_mult: 0
1098 |   }
1099 |   convolution_param {
1100 |     num_output: 64
1101 |     pad: 0
1102 |     kernel_size: 1
1103 |     stride: 2
1104 |     weight_filler {
1105 |       type: "msra"
1106 |     }
1107 |     bias_filler {
1108 |       type: "constant"
1109 |       value: 0
1110 |     }
1111 |   }
1112 | }
1113 | layer {
1114 |   name: "BatchNorm15"
1115 |   type: "BatchNorm"
1116 |   bottom: "Convolution15"
1117 |   top: "Convolution15"
1118 |   param {
1119 |     lr_mult: 0
1120 |     decay_mult: 0
1121 |   }
1122 |   param {
1123 |     lr_mult: 0
1124 |     decay_mult: 0
1125 |   }
1126 |   param {
1127 |     lr_mult: 0
1128 |     decay_mult: 0
1129 |   }
1130 |   batch_norm_param {
1131 |     use_global_stats: true
1132 |   }
1133 | }
1134 | layer {
1135 |   name: "Scale15"
1136 |   type: "Scale"
1137 |   bottom: "Convolution15"
1138 |   top: "Convolution15"
1139 |   param {
1140 |     lr_mult: 0
1141 |     decay_mult: 0
1142 |   }
1143 |   param {
1144 |     lr_mult: 0
1145 |     decay_mult: 0
1146 |   }
1147 |   scale_param {
1148 |     bias_term: true
1149 |   }
1150 | }
1151 | layer {
1152 |   name: "Convolution16"
1153 |   type: "Convolution"
1154 |   bottom: "Eltwise6"
1155 |   top: "Convolution16"
1156 |   param {
1157 |     lr_mult: 0
1158 |     decay_mult: 0
1159 |   }
1160 |   param {
1161 |     lr_mult: 0
1162 |     decay_mult: 0
1163 |   }
1164 |   convolution_param {
1165 |     num_output: 64
1166 |     pad: 1
1167 |     kernel_size: 3
1168 |     stride: 2
1169 |     weight_filler {
1170 |       type: "msra"
1171 |     }
1172 |     bias_filler {
1173 |       type: "constant"
1174 |       value: 0
1175 |     }
1176 |   }
1177 | }
1178 | layer {
1179 |   name: "BatchNorm16"
1180 |   type: "BatchNorm"
1181 |   bottom: "Convolution16"
1182 |   top: "Convolution16"
1183 |   param {
1184 |     lr_mult: 0
1185 |     decay_mult: 0
1186 |   }
1187 |   param {
1188 |     lr_mult: 0
1189 |     decay_mult: 0
1190 |   }
1191 |   param {
1192 |     lr_mult: 0
1193 |     decay_mult: 0
1194 |   }
1195 |   batch_norm_param {
1196 |     use_global_stats: true
1197 |   }
1198 | }
1199 | layer {
1200 |   name: "Scale16"
1201 |   type: "Scale"
1202 |   bottom: "Convolution16"
1203 |   top: "Convolution16"
1204 |   param {
1205 |     lr_mult: 0
1206 |     decay_mult: 0
1207 |   }
1208 |   param {
1209 |     lr_mult: 0
1210 |     decay_mult: 0
1211 |   }
1212 |   scale_param {
1213 |     bias_term: true
1214 |   }
1215 | }
1216 | layer {
1217 |   name: "ReLU14"
1218 |   type: "ReLU"
1219 |   bottom: "Convolution16"
1220 |   top: "Convolution16"
1221 | }
1222 | layer {
1223 |   name: "Convolution17"
1224 |   type: "Convolution"
1225 |   bottom: "Convolution16"
1226 |   top: "Convolution17"
1227 |   param {
1228 |     lr_mult: 0
1229 |     decay_mult: 0
1230 |   }
1231 |   param {
1232 |     lr_mult: 0
1233 |     decay_mult: 0
1234 |   }
1235 |   convolution_param {
1236 |     num_output: 64
1237 |     pad: 1
1238 |     kernel_size: 3
1239 |     stride: 1
1240 |     weight_filler {
1241 |       type: "msra"
1242 |     }
1243 |     bias_filler {
1244 |       type: "constant"
1245 |       value: 0
1246 |     }
1247 |   }
1248 | }
1249 | layer {
1250 |   name: "BatchNorm17"
1251 |   type: "BatchNorm"
1252 |   bottom: "Convolution17"
1253 |   top: "Convolution17"
1254 |   param {
1255 |     lr_mult: 0
1256 |     decay_mult: 0
1257 |   }
1258 |   param {
1259 |     lr_mult: 0
1260 |     decay_mult: 0
1261 |   }
1262 |   param {
1263 |     lr_mult: 0
1264 |     decay_mult: 0
1265 |   }
1266 |   batch_norm_param {
1267 |     use_global_stats: true
1268 |   }
1269 | }
1270 | layer {
1271 |   name: "Scale17"
1272 |   type: "Scale"
1273 |   bottom: "Convolution17"
1274 |   top: "Convolution17"
1275 |   param {
1276 |     lr_mult: 0
1277 |     decay_mult: 0
1278 |   }
1279 |   param {
1280 |     lr_mult: 0
1281 |     decay_mult: 0
1282 |   }
1283 |   scale_param {
1284 |     bias_term: true
1285 |   }
1286 | }
1287 | layer {
1288 |   name: "Eltwise7"
1289 |   type: "Eltwise"
1290 |   bottom: "Convolution15"
1291 |   bottom: "Convolution17"
1292 |   top: "Eltwise7"
1293 |   eltwise_param {
1294 |     operation: SUM
1295 |   }
1296 | }
1297 | layer {
1298 |   name: "ReLU15"
1299 |   type: "ReLU"
1300 |   bottom: "Eltwise7"
1301 |   top: "Eltwise7"
1302 | }
1303 | layer {
1304 |   name: "Convolution18"
1305 |   type: "Convolution"
1306 |   bottom: "Eltwise7"
1307 |   top: "Convolution18"
1308 |   param {
1309 |     lr_mult: 0
1310 |     decay_mult: 0
1311 |   }
1312 |   param {
1313 |     lr_mult: 0
1314 |     decay_mult: 0
1315 |   }
1316 |   convolution_param {
1317 |     num_output: 64
1318 |     pad: 1
1319 |     kernel_size: 3
1320 |     stride: 1
1321 |     weight_filler {
1322 |       type: "msra"
1323 |     }
1324 |     bias_filler {
1325 |       type: "constant"
1326 |       value: 0
1327 |     }
1328 |   }
1329 | }
1330 | layer {
1331 |   name: "BatchNorm18"
1332 |   type: "BatchNorm"
1333 |   bottom: "Convolution18"
1334 |   top: "Convolution18"
1335 |   param {
1336 |     lr_mult: 0
1337 |     decay_mult: 0
1338 |   }
1339 |   param {
1340 |     lr_mult: 0
1341 |     decay_mult: 0
1342 |   }
1343 |   param {
1344 |     lr_mult: 0
1345 |     decay_mult: 0
1346 |   }
1347 |   batch_norm_param {
1348 |     use_global_stats: true
1349 |   }
1350 | }
1351 | layer {
1352 |   name: "Scale18"
1353 |   type: "Scale"
1354 |   bottom: "Convolution18"
1355 |   top: "Convolution18"
1356 |   param {
1357 |     lr_mult: 0
1358 |     decay_mult: 0
1359 |   }
1360 |   param {
1361 |     lr_mult: 0
1362 |     decay_mult: 0
1363 |   }
1364 |   scale_param {
1365 |     bias_term: true
1366 |   }
1367 | }
1368 | layer {
1369 |   name: "ReLU16"
1370 |   type: "ReLU"
1371 |   bottom: "Convolution18"
1372 |   top: "Convolution18"
1373 | }
1374 | layer {
1375 |   name: "Convolution19"
1376 |   type: "Convolution"
1377 |   bottom: "Convolution18"
1378 |   top: "Convolution19"
1379 |   param {
1380 |     lr_mult: 0
1381 |     decay_mult: 0
1382 |   }
1383 |   param {
1384 |     lr_mult: 0
1385 |     decay_mult: 0
1386 |   }
1387 |   convolution_param {
1388 |     num_output: 64
1389 |     pad: 1
1390 |     kernel_size: 3
1391 |     stride: 1
1392 |     weight_filler {
1393 |       type: "msra"
1394 |     }
1395 |     bias_filler {
1396 |       type: "constant"
1397 |       value: 0
1398 |     }
1399 |   }
1400 | }
1401 | layer {
1402 |   name: "BatchNorm19"
1403 |   type: "BatchNorm"
1404 |   bottom: "Convolution19"
1405 |   top: "Convolution19"
1406 |   param {
1407 |     lr_mult: 0
1408 |     decay_mult: 0
1409 |   }
1410 |   param {
1411 |     lr_mult: 0
1412 |     decay_mult: 0
1413 |   }
1414 |   param {
1415 |     lr_mult: 0
1416 |     decay_mult: 0
1417 |   }
1418 |   batch_norm_param {
1419 |     use_global_stats: true
1420 |   }
1421 | }
1422 | layer {
1423 |   name: "Scale19"
1424 |   type: "Scale"
1425 |   bottom: "Convolution19"
1426 |   top: "Convolution19"
1427 |   param {
1428 |     lr_mult: 0
1429 |     decay_mult: 0
1430 |   }
1431 |   param {
1432 |     lr_mult: 0
1433 |     decay_mult: 0
1434 |   }
1435 |   scale_param {
1436 |     bias_term: true
1437 |   }
1438 | }
1439 | layer {
1440 |   name: "Eltwise8"
1441 |   type: "Eltwise"
1442 |   bottom: "Eltwise7"
1443 |   bottom: "Convolution19"
1444 |   top: "Eltwise8"
1445 |   eltwise_param {
1446 |     operation: SUM
1447 |   }
1448 | }
1449 | layer {
1450 |   name: "ReLU17"
1451 |   type: "ReLU"
1452 |   bottom: "Eltwise8"
1453 |   top: "Eltwise8"
1454 | }
1455 | layer {
1456 |   name: "Convolution20"
1457 |   type: "Convolution"
1458 |   bottom: "Eltwise8"
1459 |   top: "Convolution20"
1460 |   param {
1461 |     lr_mult: 0
1462 |     decay_mult: 0
1463 |   }
1464 |   param {
1465 |     lr_mult: 0
1466 |     decay_mult: 0
1467 |   }
1468 |   convolution_param {
1469 |     num_output: 64
1470 |     pad: 1
1471 |     kernel_size: 3
1472 |     stride: 1
1473 |     weight_filler {
1474 |       type: "msra"
1475 |     }
1476 |     bias_filler {
1477 |       type: "constant"
1478 |       value: 0
1479 |     }
1480 |   }
1481 | }
1482 | layer {
1483 |   name: "BatchNorm20"
1484 |   type: "BatchNorm"
1485 |   bottom: "Convolution20"
1486 |   top: "Convolution20"
1487 |   param {
1488 |     lr_mult: 0
1489 |     decay_mult: 0
1490 |   }
1491 |   param {
1492 |     lr_mult: 0
1493 |     decay_mult: 0
1494 |   }
1495 |   param {
1496 |     lr_mult: 0
1497 |     decay_mult: 0
1498 |   }
1499 |   batch_norm_param {
1500 |     use_global_stats: true
1501 |   }
1502 | }
1503 | layer {
1504 |   name: "Scale20"
1505 |   type: "Scale"
1506 |   bottom: "Convolution20"
1507 |   top: "Convolution20"
1508 |   param {
1509 |     lr_mult: 0
1510 |     decay_mult: 0
1511 |   }
1512 |   param {
1513 |     lr_mult: 0
1514 |     decay_mult: 0
1515 |   }
1516 |   scale_param {
1517 |     bias_term: true
1518 |   }
1519 | }
1520 | layer {
1521 |   name: "ReLU18"
1522 |   type: "ReLU"
1523 |   bottom: "Convolution20"
1524 |   top: "Convolution20"
1525 | }
1526 | layer {
1527 |   name: "Convolution21"
1528 |   type: "Convolution"
1529 |   bottom: "Convolution20"
1530 |   top: "Convolution21"
1531 |   param {
1532 |     lr_mult: 0
1533 |     decay_mult: 0
1534 |   }
1535 |   param {
1536 |     lr_mult: 0
1537 |     decay_mult: 0
1538 |   }
1539 |   convolution_param {
1540 |     num_output: 64
1541 |     pad: 1
1542 |     kernel_size: 3
1543 |     stride: 1
1544 |     weight_filler {
1545 |       type: "msra"
1546 |     }
1547 |     bias_filler {
1548 |       type: "constant"
1549 |       value: 0
1550 |     }
1551 |   }
1552 | }
1553 | layer {
1554 |   name: "BatchNorm21"
1555 |   type: "BatchNorm"
1556 |   bottom: "Convolution21"
1557 |   top: "Convolution21"
1558 |   param {
1559 |     lr_mult: 0
1560 |     decay_mult: 0
1561 |   }
1562 |   param {
1563 |     lr_mult: 0
1564 |     decay_mult: 0
1565 |   }
1566 |   param {
1567 |     lr_mult: 0
1568 |     decay_mult: 0
1569 |   }
1570 |   batch_norm_param {
1571 |     use_global_stats: true
1572 |   }
1573 | }
1574 | layer {
1575 |   name: "Scale21"
1576 |   type: "Scale"
1577 |   bottom: "Convolution21"
1578 |   top: "Convolution21"
1579 |   param {
1580 |     lr_mult: 0
1581 |     decay_mult: 0
1582 |   }
1583 |   param {
1584 |     lr_mult: 0
1585 |     decay_mult: 0
1586 |   }
1587 |   scale_param {
1588 |     bias_term: true
1589 |   }
1590 | }
1591 | layer {
1592 |   name: "Eltwise9"
1593 |   type: "Eltwise"
1594 |   bottom: "Eltwise8"
1595 |   bottom: "Convolution21"
1596 |   top: "Eltwise9"
1597 |   eltwise_param {
1598 |     operation: SUM
1599 |   }
1600 | }
1601 | layer {
1602 |   name: "ReLU19"
1603 |   type: "ReLU"
1604 |   bottom: "Eltwise9"
1605 |   top: "Eltwise9"
1606 | }
1607 | layer {
1608 |   name: "Pooling1"
1609 |   type: "Pooling"
1610 |   bottom: "Eltwise9"
1611 |   top: "Pooling1"
1612 |   pooling_param {
1613 |     pool: AVE
1614 |     global_pooling: true
1615 |   }
1616 | }
1617 | layer {
1618 |   name: "InnerProduct1"
1619 |   type: "InnerProduct"
1620 |   bottom: "Pooling1"
1621 |   top: "InnerProduct1"
1622 |   param {
1623 |     lr_mult: 0
1624 |     decay_mult: 0
1625 |   }
1626 |   param {
1627 |     lr_mult: 0
1628 |     decay_mult: 0
1629 |   }
1630 |   inner_product_param {
1631 |     num_output: 10
1632 |     weight_filler {
1633 |       type: "msra"
1634 |     }
1635 |     bias_filler {
1636 |       type: "constant"
1637 |       value: 0
1638 |     }
1639 |   }
1640 | }
1641 | layer {
1642 |   name: "SoftmaxWithLoss1"
1643 |   type: "SoftmaxWithLoss"
1644 |   bottom: "InnerProduct1"
1645 |   bottom: "Data2"
1646 |   top: "SoftmaxWithLoss1"
1647 |   loss_weight: 0
1648 | }
1649 | layer {
1650 |   name: "Accuracy1"
1651 |   type: "Accuracy"
1652 |   bottom: "InnerProduct1"
1653 |   bottom: "Data2"
1654 |   top: "Accuracy1"
1655 |   include {
1656 |     phase: TEST
1657 |   }
1658 | }
1659 | 
1660 | layer {
1661 |   name: "stu_Convolution1"
1662 |   type: "Convolution"
1663 |   bottom: "Data1"
1664 |   top: "stu_Convolution1"
1665 |   param {
1666 |     lr_mult: 1
1667 |     decay_mult: 1
1668 |   }
1669 |   param {
1670 |     lr_mult: 2
1671 |     decay_mult: 0
1672 |   }
1673 |   convolution_param {
1674 |     num_output: 8
1675 |     pad: 1
1676 |     kernel_size: 3
1677 |     stride: 1
1678 |     weight_filler {
1679 |       type: "msra"
1680 |     }
1681 |     bias_filler {
1682 |       type: "constant"
1683 |       value: 0
1684 |     }
1685 |   }
1686 | }
1687 | layer {
1688 |   name: "stu_BatchNorm1"
1689 |   type: "BatchNorm"
1690 |   bottom: "stu_Convolution1"
1691 |   top: "stu_Convolution1"
1692 |   param {
1693 |     lr_mult: 0
1694 |     decay_mult: 0
1695 |   }
1696 |   param {
1697 |     lr_mult: 0
1698 |     decay_mult: 0
1699 |   }
1700 |   param {
1701 |     lr_mult: 0
1702 |     decay_mult: 0
1703 |   }
1704 | }
1705 | layer {
1706 |   name: "stu_Scale1"
1707 |   type: "Scale"
1708 |   bottom: "stu_Convolution1"
1709 |   top: "stu_Convolution1"
1710 |   scale_param {
1711 |     bias_term: true
1712 |   }
1713 | }
1714 | layer {
1715 |   name: "stu_ReLU1"
1716 |   type: "ReLU"
1717 |   bottom: "stu_Convolution1"
1718 |   top: "stu_Convolution1"
1719 | }
1720 | layer {
1721 |   name: "stu_Convolution2"
1722 |   type: "Convolution"
1723 |   bottom: "stu_Convolution1"
1724 |   top: "stu_Convolution2"
1725 |   param {
1726 |     lr_mult: 1
1727 |     decay_mult: 1
1728 |   }
1729 |   param {
1730 |     lr_mult: 2
1731 |     decay_mult: 0
1732 |   }
1733 |   convolution_param {
1734 |     num_output: 8
1735 |     pad: 1
1736 |     kernel_size: 3
1737 |     stride: 1
1738 |     weight_filler {
1739 |       type: "msra"
1740 |     }
1741 |     bias_filler {
1742 |       type: "constant"
1743 |       value: 0
1744 |     }
1745 |   }
1746 | }
1747 | layer {
1748 |   name: "stu_BatchNorm2"
1749 |   type: "BatchNorm"
1750 |   bottom: "stu_Convolution2"
1751 |   top: "stu_Convolution2"
1752 |   param {
1753 |     lr_mult: 0
1754 |     decay_mult: 0
1755 |   }
1756 |   param {
1757 |     lr_mult: 0
1758 |     decay_mult: 0
1759 |   }
1760 |   param {
1761 |     lr_mult: 0
1762 |     decay_mult: 0
1763 |   }
1764 | }
1765 | layer {
1766 |   name: "stu_Scale2"
1767 |   type: "Scale"
1768 |   bottom: "stu_Convolution2"
1769 |   top: "stu_Convolution2"
1770 |   scale_param {
1771 |     bias_term: true
1772 |   }
1773 | }
1774 | layer {
1775 |   name: "stu_ReLU2"
1776 |   type: "ReLU"
1777 |   bottom: "stu_Convolution2"
1778 |   top: "stu_Convolution2"
1779 | }
1780 | layer {
1781 |   name: "stu_Convolution3"
1782 |   type: "Convolution"
1783 |   bottom: "stu_Convolution2"
1784 |   top: "stu_Convolution3"
1785 |   param {
1786 |     lr_mult: 1
1787 |     decay_mult: 1
1788 |   }
1789 |   param {
1790 |     lr_mult: 2
1791 |     decay_mult: 0
1792 |   }
1793 |   convolution_param {
1794 |     num_output: 8
1795 |     pad: 1
1796 |     kernel_size: 3
1797 |     stride: 1
1798 |     weight_filler {
1799 |       type: "msra"
1800 |     }
1801 |     bias_filler {
1802 |       type: "constant"
1803 |       value: 0
1804 |     }
1805 |   }
1806 | }
1807 | layer {
1808 |   name: "stu_BatchNorm3"
1809 |   type: "BatchNorm"
1810 |   bottom: "stu_Convolution3"
1811 |   top: "stu_Convolution3"
1812 |   param {
1813 |     lr_mult: 0
1814 |     decay_mult: 0
1815 |   }
1816 |   param {
1817 |     lr_mult: 0
1818 |     decay_mult: 0
1819 |   }
1820 |   param {
1821 |     lr_mult: 0
1822 |     decay_mult: 0
1823 |   }
1824 | }
1825 | layer {
1826 |   name: "stu_Scale3"
1827 |   type: "Scale"
1828 |   bottom: "stu_Convolution3"
1829 |   top: "stu_Convolution3"
1830 |   scale_param {
1831 |     bias_term: true
1832 |   }
1833 | }
1834 | layer {
1835 |   name: "stu_Eltwise1"
1836 |   type: "Eltwise"
1837 |   bottom: "stu_Convolution1"
1838 |   bottom: "stu_Convolution3"
1839 |   top: "stu_Eltwise1"
1840 |   eltwise_param {
1841 |     operation: SUM
1842 |   }
1843 | }
1844 | layer {
1845 |   name: "stu_ReLU3"
1846 |   type: "ReLU"
1847 |   bottom: "stu_Eltwise1"
1848 |   top: "stu_Eltwise1"
1849 | }
1850 | layer {
1851 |   name: "stu_Convolution6"
1852 |   type: "Convolution"
1853 |   bottom: "stu_Eltwise1"
1854 |   top: "stu_Convolution6"
1855 |   param {
1856 |     lr_mult: 1
1857 |     decay_mult: 1
1858 |   }
1859 |   param {
1860 |     lr_mult: 2
1861 |     decay_mult: 0
1862 |   }
1863 |   convolution_param {
1864 |     num_output: 8
1865 |     pad: 1
1866 |     kernel_size: 3
1867 |     stride: 1
1868 |     weight_filler {
1869 |       type: "msra"
1870 |     }
1871 |     bias_filler {
1872 |       type: "constant"
1873 |       value: 0
1874 |     }
1875 |   }
1876 | }
1877 | layer {
1878 |   name: "stu_BatchNorm6"
1879 |   type: "BatchNorm"
1880 |   bottom: "stu_Convolution6"
1881 |   top: "stu_Convolution6"
1882 |   param {
1883 |     lr_mult: 0
1884 |     decay_mult: 0
1885 |   }
1886 |   param {
1887 |     lr_mult: 0
1888 |     decay_mult: 0
1889 |   }
1890 |   param {
1891 |     lr_mult: 0
1892 |     decay_mult: 0
1893 |   }
1894 | }
1895 | layer {
1896 |   name: "stu_Scale6"
1897 |   type: "Scale"
1898 |   bottom: "stu_Convolution6"
1899 |   top: "stu_Convolution6"
1900 |   scale_param {
1901 |     bias_term: true
1902 |   }
1903 | }
1904 | layer {
1905 |   name: "stu_ReLU6"
1906 |   type: "ReLU"
1907 |   bottom: "stu_Convolution6"
1908 |   top: "stu_Convolution6"
1909 | }
1910 | layer {
1911 |   name: "stu_Convolution7"
1912 |   type: "Convolution"
1913 |   bottom: "stu_Convolution6"
1914 |   top: "stu_Convolution7"
1915 |   param {
1916 |     lr_mult: 1
1917 |     decay_mult: 1
1918 |   }
1919 |   param {
1920 |     lr_mult: 2
1921 |     decay_mult: 0
1922 |   }
1923 |   convolution_param {
1924 |     num_output: 8
1925 |     pad: 1
1926 |     kernel_size: 3
1927 |     stride: 1
1928 |     weight_filler {
1929 |       type: "msra"
1930 |     }
1931 |     bias_filler {
1932 |       type: "constant"
1933 |       value: 0
1934 |     }
1935 |   }
1936 | }
1937 | layer {
1938 |   name: "stu_BatchNorm7"
1939 |   type: "BatchNorm"
1940 |   bottom: "stu_Convolution7"
1941 |   top: "stu_Convolution7"
1942 |   param {
1943 |     lr_mult: 0
1944 |     decay_mult: 0
1945 |   }
1946 |   param {
1947 |     lr_mult: 0
1948 |     decay_mult: 0
1949 |   }
1950 |   param {
1951 |     lr_mult: 0
1952 |     decay_mult: 0
1953 |   }
1954 | }
1955 | layer {
1956 |   name: "stu_Scale7"
1957 |   type: "Scale"
1958 |   bottom: "stu_Convolution7"
1959 |   top: "stu_Convolution7"
1960 |   scale_param {
1961 |     bias_term: true
1962 |   }
1963 | }
1964 | layer {
1965 |   name: "stu_Eltwise3"
1966 |   type: "Eltwise"
1967 |   bottom: "stu_Eltwise1"
1968 |   bottom: "stu_Convolution7"
1969 |   top: "stu_Eltwise3"
1970 |   eltwise_param {
1971 |     operation: SUM
1972 |   }
1973 | }
1974 | layer {
1975 |   name: "stu_ReLU7"
1976 |   type: "ReLU"
1977 |   bottom: "stu_Eltwise3"
1978 |   top: "stu_Eltwise3"
1979 | }
1980 | layer {
1981 |   name: "stu_Convolution8"
1982 |   type: "Convolution"
1983 |   bottom: "stu_Eltwise3"
1984 |   top: "stu_Convolution8"
1985 |   param {
1986 |     lr_mult: 1
1987 |     decay_mult: 1
1988 |   }
1989 |   param {
1990 |     lr_mult: 2
1991 |     decay_mult: 0
1992 |   }
1993 |   convolution_param {
1994 |     num_output: 16
1995 |     pad: 0
1996 |     kernel_size: 1
1997 |     stride: 2
1998 |     weight_filler {
1999 |       type: "msra"
2000 |     }
2001 |     bias_filler {
2002 |       type: "constant"
2003 |       value: 0
2004 |     }
2005 |   }
2006 | }
2007 | layer {
2008 |   name: "stu_BatchNorm8"
2009 |   type: "BatchNorm"
2010 |   bottom: "stu_Convolution8"
2011 |   top: "stu_Convolution8"
2012 |   param {
2013 |     lr_mult: 0
2014 |     decay_mult: 0
2015 |   }
2016 |   param {
2017 |     lr_mult: 0
2018 |     decay_mult: 0
2019 |   }
2020 |   param {
2021 |     lr_mult: 0
2022 |     decay_mult: 0
2023 |   }
2024 | }
2025 | layer {
2026 |   name: "stu_Scale8"
2027 |   type: "Scale"
2028 |   bottom: "stu_Convolution8"
2029 |   top: "stu_Convolution8"
2030 |   scale_param {
2031 |     bias_term: true
2032 |   }
2033 | }
2034 | layer {
2035 |   name: "stu_Convolution9"
2036 |   type: "Convolution"
2037 |   bottom: "stu_Eltwise3"
2038 |   top: "stu_Convolution9"
2039 |   param {
2040 |     lr_mult: 1
2041 |     decay_mult: 1
2042 |   }
2043 |   param {
2044 |     lr_mult: 2
2045 |     decay_mult: 0
2046 |   }
2047 |   convolution_param {
2048 |     num_output: 16
2049 |     pad: 1
2050 |     kernel_size: 3
2051 |     stride: 2
2052 |     weight_filler {
2053 |       type: "msra"
2054 |     }
2055 |     bias_filler {
2056 |       type: "constant"
2057 |       value: 0
2058 |     }
2059 |   }
2060 | }
2061 | layer {
2062 |   name: "stu_BatchNorm9"
2063 |   type: "BatchNorm"
2064 |   bottom: "stu_Convolution9"
2065 |   top: "stu_Convolution9"
2066 |   param {
2067 |     lr_mult: 0
2068 |     decay_mult: 0
2069 |   }
2070 |   param {
2071 |     lr_mult: 0
2072 |     decay_mult: 0
2073 |   }
2074 |   param {
2075 |     lr_mult: 0
2076 |     decay_mult: 0
2077 |   }
2078 | }
2079 | layer {
2080 |   name: "stu_Scale9"
2081 |   type: "Scale"
2082 |   bottom: "stu_Convolution9"
2083 |   top: "stu_Convolution9"
2084 |   scale_param {
2085 |     bias_term: true
2086 |   }
2087 | }
2088 | layer {
2089 |   name: "stu_ReLU8"
2090 |   type: "ReLU"
2091 |   bottom: "stu_Convolution9"
2092 |   top: "stu_Convolution9"
2093 | }
2094 | layer {
2095 |   name: "stu_Convolution10"
2096 |   type: "Convolution"
2097 |   bottom: "stu_Convolution9"
2098 |   top: "stu_Convolution10"
2099 |   param {
2100 |     lr_mult: 1
2101 |     decay_mult: 1
2102 |   }
2103 |   param {
2104 |     lr_mult: 2
2105 |     decay_mult: 0
2106 |   }
2107 |   convolution_param {
2108 |     num_output: 16
2109 |     pad: 1
2110 |     kernel_size: 3
2111 |     stride: 1
2112 |     weight_filler {
2113 |       type: "msra"
2114 |     }
2115 |     bias_filler {
2116 |       type: "constant"
2117 |       value: 0
2118 |     }
2119 |   }
2120 | }
2121 | layer {
2122 |   name: "stu_BatchNorm10"
2123 |   type: "BatchNorm"
2124 |   bottom: "stu_Convolution10"
2125 |   top: "stu_Convolution10"
2126 |   param {
2127 |     lr_mult: 0
2128 |     decay_mult: 0
2129 |   }
2130 |   param {
2131 |     lr_mult: 0
2132 |     decay_mult: 0
2133 |   }
2134 |   param {
2135 |     lr_mult: 0
2136 |     decay_mult: 0
2137 |   }
2138 | }
2139 | layer {
2140 |   name: "stu_Scale10"
2141 |   type: "Scale"
2142 |   bottom: "stu_Convolution10"
2143 |   top: "stu_Convolution10"
2144 |   scale_param {
2145 |     bias_term: true
2146 |   }
2147 | }
2148 | layer {
2149 |   name: "stu_Eltwise4"
2150 |   type: "Eltwise"
2151 |   bottom: "stu_Convolution8"
2152 |   bottom: "stu_Convolution10"
2153 |   top: "stu_Eltwise4"
2154 |   eltwise_param {
2155 |     operation: SUM
2156 |   }
2157 | }
2158 | layer {
2159 |   name: "stu_ReLU9"
2160 |   type: "ReLU"
2161 |   bottom: "stu_Eltwise4"
2162 |   top: "stu_Eltwise4"
2163 | }
2164 | layer {
2165 |   name: "stu_Convolution11"
2166 |   type: "Convolution"
2167 |   bottom: "stu_Eltwise4"
2168 |   top: "stu_Convolution11"
2169 |   param {
2170 |     lr_mult: 1
2171 |     decay_mult: 1
2172 |   }
2173 |   param {
2174 |     lr_mult: 2
2175 |     decay_mult: 0
2176 |   }
2177 |   convolution_param {
2178 |     num_output: 16
2179 |     pad: 1
2180 |     kernel_size: 3
2181 |     stride: 1
2182 |     weight_filler {
2183 |       type: "msra"
2184 |     }
2185 |     bias_filler {
2186 |       type: "constant"
2187 |       value: 0
2188 |     }
2189 |   }
2190 | }
2191 | layer {
2192 |   name: "stu_BatchNorm11"
2193 |   type: "BatchNorm"
2194 |   bottom: "stu_Convolution11"
2195 |   top: "stu_Convolution11"
2196 |   param {
2197 |     lr_mult: 0
2198 |     decay_mult: 0
2199 |   }
2200 |   param {
2201 |     lr_mult: 0
2202 |     decay_mult: 0
2203 |   }
2204 |   param {
2205 |     lr_mult: 0
2206 |     decay_mult: 0
2207 |   }
2208 | }
2209 | layer {
2210 |   name: "stu_Scale11"
2211 |   type: "Scale"
2212 |   bottom: "stu_Convolution11"
2213 |   top: "stu_Convolution11"
2214 |   scale_param {
2215 |     bias_term: true
2216 |   }
2217 | }
2218 | layer {
2219 |   name: "stu_ReLU10"
2220 |   type: "ReLU"
2221 |   bottom: "stu_Convolution11"
2222 |   top: "stu_Convolution11"
2223 | }
2224 | layer {
2225 |   name: "stu_Convolution12"
2226 |   type: "Convolution"
2227 |   bottom: "stu_Convolution11"
2228 |   top: "stu_Convolution12"
2229 |   param {
2230 |     lr_mult: 1
2231 |     decay_mult: 1
2232 |   }
2233 |   param {
2234 |     lr_mult: 2
2235 |     decay_mult: 0
2236 |   }
2237 |   convolution_param {
2238 |     num_output: 16
2239 |     pad: 1
2240 |     kernel_size: 3
2241 |     stride: 1
2242 |     weight_filler {
2243 |       type: "msra"
2244 |     }
2245 |     bias_filler {
2246 |       type: "constant"
2247 |       value: 0
2248 |     }
2249 |   }
2250 | }
2251 | layer {
2252 |   name: "stu_BatchNorm12"
2253 |   type: "BatchNorm"
2254 |   bottom: "stu_Convolution12"
2255 |   top: "stu_Convolution12"
2256 |   param {
2257 |     lr_mult: 0
2258 |     decay_mult: 0
2259 |   }
2260 |   param {
2261 |     lr_mult: 0
2262 |     decay_mult: 0
2263 |   }
2264 |   param {
2265 |     lr_mult: 0
2266 |     decay_mult: 0
2267 |   }
2268 | }
2269 | layer {
2270 |   name: "stu_Scale12"
2271 |   type: "Scale"
2272 |   bottom: "stu_Convolution12"
2273 |   top: "stu_Convolution12"
2274 |   scale_param {
2275 |     bias_term: true
2276 |   }
2277 | }
2278 | layer {
2279 |   name: "stu_Eltwise5"
2280 |   type: "Eltwise"
2281 |   bottom: "stu_Eltwise4"
2282 |   bottom: "stu_Convolution12"
2283 |   top: "stu_Eltwise5"
2284 |   eltwise_param {
2285 |     operation: SUM
2286 |   }
2287 | }
2288 | layer {
2289 |   name: "stu_ReLU11"
2290 |   type: "ReLU"
2291 |   bottom: "stu_Eltwise5"
2292 |   top: "stu_Eltwise5"
2293 | }
2294 | layer {
2295 |   name: "stu_Convolution15"
2296 |   type: "Convolution"
2297 |   bottom: "stu_Eltwise5"
2298 |   top: "stu_Convolution15"
2299 |   param {
2300 |     lr_mult: 1
2301 |     decay_mult: 1
2302 |   }
2303 |   param {
2304 |     lr_mult: 2
2305 |     decay_mult: 0
2306 |   }
2307 |   convolution_param {
2308 |     num_output: 32
2309 |     pad: 0
2310 |     kernel_size: 1
2311 |     stride: 2
2312 |     weight_filler {
2313 |       type: "msra"
2314 |     }
2315 |     bias_filler {
2316 |       type: "constant"
2317 |       value: 0
2318 |     }
2319 |   }
2320 | }
2321 | layer {
2322 |   name: "stu_BatchNorm15"
2323 |   type: "BatchNorm"
2324 |   bottom: "stu_Convolution15"
2325 |   top: "stu_Convolution15"
2326 |   param {
2327 |     lr_mult: 0
2328 |     decay_mult: 0
2329 |   }
2330 |   param {
2331 |     lr_mult: 0
2332 |     decay_mult: 0
2333 |   }
2334 |   param {
2335 |     lr_mult: 0
2336 |     decay_mult: 0
2337 |   }
2338 | }
2339 | layer {
2340 |   name: "stu_Scale15"
2341 |   type: "Scale"
2342 |   bottom: "stu_Convolution15"
2343 |   top: "stu_Convolution15"
2344 |   scale_param {
2345 |     bias_term: true
2346 |   }
2347 | }
2348 | layer {
2349 |   name: "stu_Convolution16"
2350 |   type: "Convolution"
2351 |   bottom: "stu_Eltwise5"
2352 |   top: "stu_Convolution16"
2353 |   param {
2354 |     lr_mult: 1
2355 |     decay_mult: 1
2356 |   }
2357 |   param {
2358 |     lr_mult: 2
2359 |     decay_mult: 0
2360 |   }
2361 |   convolution_param {
2362 |     num_output: 32
2363 |     pad: 1
2364 |     kernel_size: 3
2365 |     stride: 2
2366 |     weight_filler {
2367 |       type: "msra"
2368 |     }
2369 |     bias_filler {
2370 |       type: "constant"
2371 |       value: 0
2372 |     }
2373 |   }
2374 | }
2375 | layer {
2376 |   name: "stu_BatchNorm16"
2377 |   type: "BatchNorm"
2378 |   bottom: "stu_Convolution16"
2379 |   top: "stu_Convolution16"
2380 |   param {
2381 |     lr_mult: 0
2382 |     decay_mult: 0
2383 |   }
2384 |   param {
2385 |     lr_mult: 0
2386 |     decay_mult: 0
2387 |   }
2388 |   param {
2389 |     lr_mult: 0
2390 |     decay_mult: 0
2391 |   }
2392 | }
2393 | layer {
2394 |   name: "stu_Scale16"
2395 |   type: "Scale"
2396 |   bottom: "stu_Convolution16"
2397 |   top: "stu_Convolution16"
2398 |   scale_param {
2399 |     bias_term: true
2400 |   }
2401 | }
2402 | layer {
2403 |   name: "stu_ReLU14"
2404 |   type: "ReLU"
2405 |   bottom: "stu_Convolution16"
2406 |   top: "stu_Convolution16"
2407 | }
2408 | layer {
2409 |   name: "stu_Convolution17"
2410 |   type: "Convolution"
2411 |   bottom: "stu_Convolution16"
2412 |   top: "stu_Convolution17"
2413 |   param {
2414 |     lr_mult: 1
2415 |     decay_mult: 1
2416 |   }
2417 |   param {
2418 |     lr_mult: 2
2419 |     decay_mult: 0
2420 |   }
2421 |   convolution_param {
2422 |     num_output: 32
2423 |     pad: 1
2424 |     kernel_size: 3
2425 |     stride: 1
2426 |     weight_filler {
2427 |       type: "msra"
2428 |     }
2429 |     bias_filler {
2430 |       type: "constant"
2431 |       value: 0
2432 |     }
2433 |   }
2434 | }
2435 | layer {
2436 |   name: "stu_BatchNorm17"
2437 |   type: "BatchNorm"
2438 |   bottom: "stu_Convolution17"
2439 |   top: "stu_Convolution17"
2440 |   param {
2441 |     lr_mult: 0
2442 |     decay_mult: 0
2443 |   }
2444 |   param {
2445 |     lr_mult: 0
2446 |     decay_mult: 0
2447 |   }
2448 |   param {
2449 |     lr_mult: 0
2450 |     decay_mult: 0
2451 |   }
2452 | }
2453 | layer {
2454 |   name: "stu_Scale17"
2455 |   type: "Scale"
2456 |   bottom: "stu_Convolution17"
2457 |   top: "stu_Convolution17"
2458 |   scale_param {
2459 |     bias_term: true
2460 |   }
2461 | }
2462 | layer {
2463 |   name: "stu_Eltwise7"
2464 |   type: "Eltwise"
2465 |   bottom: "stu_Convolution15"
2466 |   bottom: "stu_Convolution17"
2467 |   top: "stu_Eltwise7"
2468 |   eltwise_param {
2469 |     operation: SUM
2470 |   }
2471 | }
2472 | layer {
2473 |   name: "stu_ReLU15"
2474 |   type: "ReLU"
2475 |   bottom: "stu_Eltwise7"
2476 |   top: "stu_Eltwise7"
2477 | }
2478 | 
2479 | layer {
2480 |   name: "stu_Convolution20"
2481 |   type: "Convolution"
2482 |   bottom: "stu_Eltwise7"
2483 |   top: "stu_Convolution20"
2484 |   param {
2485 |     lr_mult: 1
2486 |     decay_mult: 1
2487 |   }
2488 |   param {
2489 |     lr_mult: 2
2490 |     decay_mult: 0
2491 |   }
2492 |   convolution_param {
2493 |     num_output: 32
2494 |     pad: 1
2495 |     kernel_size: 3
2496 |     stride: 1
2497 |     weight_filler {
2498 |       type: "msra"
2499 |     }
2500 |     bias_filler {
2501 |       type: "constant"
2502 |       value: 0
2503 |     }
2504 |   }
2505 | }
2506 | layer {
2507 |   name: "stu_BatchNorm20"
2508 |   type: "BatchNorm"
2509 |   bottom: "stu_Convolution20"
2510 |   top: "stu_Convolution20"
2511 |   param {
2512 |     lr_mult: 0
2513 |     decay_mult: 0
2514 |   }
2515 |   param {
2516 |     lr_mult: 0
2517 |     decay_mult: 0
2518 |   }
2519 |   param {
2520 |     lr_mult: 0
2521 |     decay_mult: 0
2522 |   }
2523 | }
2524 | layer {
2525 |   name: "stu_Scale20"
2526 |   type: "Scale"
2527 |   bottom: "stu_Convolution20"
2528 |   top: "stu_Convolution20"
2529 |   scale_param {
2530 |     bias_term: true
2531 |   }
2532 | }
2533 | layer {
2534 |   name: "stu_ReLU18"
2535 |   type: "ReLU"
2536 |   bottom: "stu_Convolution20"
2537 |   top: "stu_Convolution20"
2538 | }
2539 | layer {
2540 |   name: "stu_Convolution21"
2541 |   type: "Convolution"
2542 |   bottom: "stu_Convolution20"
2543 |   top: "stu_Convolution21"
2544 |   param {
2545 |     lr_mult: 1
2546 |     decay_mult: 1
2547 |   }
2548 |   param {
2549 |     lr_mult: 2
2550 |     decay_mult: 0
2551 |   }
2552 |   convolution_param {
2553 |     num_output: 32
2554 |     pad: 1
2555 |     kernel_size: 3
2556 |     stride: 1
2557 |     weight_filler {
2558 |       type: "msra"
2559 |     }
2560 |     bias_filler {
2561 |       type: "constant"
2562 |       value: 0
2563 |     }
2564 |   }
2565 | }
2566 | layer {
2567 |   name: "stu_BatchNorm21"
2568 |   type: "BatchNorm"
2569 |   bottom: "stu_Convolution21"
2570 |   top: "stu_Convolution21"
2571 |   param {
2572 |     lr_mult: 0
2573 |     decay_mult: 0
2574 |   }
2575 |   param {
2576 |     lr_mult: 0
2577 |     decay_mult: 0
2578 |   }
2579 |   param {
2580 |     lr_mult: 0
2581 |     decay_mult: 0
2582 |   }
2583 | }
2584 | layer {
2585 |   name: "stu_Scale21"
2586 |   type: "Scale"
2587 |   bottom: "stu_Convolution21"
2588 |   top: "stu_Convolution21"
2589 |   scale_param {
2590 |     bias_term: true
2591 |   }
2592 | }
2593 | layer {
2594 |   name: "stu_Eltwise9"
2595 |   type: "Eltwise"
2596 |   bottom: "stu_Eltwise7"
2597 |   bottom: "stu_Convolution21"
2598 |   top: "stu_Eltwise9"
2599 |   eltwise_param {
2600 |     operation: SUM
2601 |   }
2602 | }
2603 | layer {
2604 |   name: "stu_ReLU19"
2605 |   type: "ReLU"
2606 |   bottom: "stu_Eltwise9"
2607 |   top: "stu_Eltwise9"
2608 | }
2609 | layer {
2610 |   name: "stu_Pooling1"
2611 |   type: "Pooling"
2612 |   bottom: "stu_Eltwise9"
2613 |   top: "stu_Pooling1"
2614 |   pooling_param {
2615 |     pool: AVE
2616 |     global_pooling: true
2617 |   }
2618 | }
2619 | layer {
2620 |   name: "stu_InnerProduct1"
2621 |   type: "InnerProduct"
2622 |   bottom: "stu_Pooling1"
2623 |   top: "stu_InnerProduct1"
2624 |   param {
2625 |     lr_mult: 1
2626 |     decay_mult: 1
2627 |   }
2628 |   param {
2629 |     lr_mult: 2
2630 |     decay_mult: 1
2631 |   }
2632 |   inner_product_param {
2633 |     num_output: 10
2634 |     weight_filler {
2635 |       type: "msra"
2636 |     }
2637 |     bias_filler {
2638 |       type: "constant"
2639 |       value: 0
2640 |     }
2641 |   }
2642 | }
2643 | layer {
2644 |   name: "stu_SoftmaxWithLoss1"
2645 |   type: "SoftmaxWithLoss"
2646 |   bottom: "stu_InnerProduct1"
2647 |   bottom: "Data2"
2648 |   top: "stu_SoftmaxWithLoss1"
2649 | }
2650 | layer {
2651 |   name: "stu_Accuracy1"
2652 |   type: "Accuracy"
2653 |   bottom: "stu_InnerProduct1"
2654 |   bottom: "Data2"
2655 |   top: "stu_Accuracy1"
2656 |   include {
2657 |     phase: TEST
2658 |   }
2659 | }
2660 | 
2661 | layer {
2662 |   name: "pool_eltwise1"
2663 |   type: "Pooling"
2664 |   bottom: "Eltwise1"
2665 |   top: "pool_eltwise1"
2666 |   pooling_param {
2667 |     pool: AVE
2668 |     global_pooling: true
2669 |   }
2670 |   include {
2671 |     phase: TRAIN
2672 |   }
2673 | }
2674 | layer {
2675 |   name: "pool_eltwise3"
2676 |   type: "Pooling"
2677 |   bottom: "Eltwise3"
2678 |   top: "pool_eltwise3"
2679 |   pooling_param {
2680 |     pool: AVE
2681 |     global_pooling: true
2682 |   }
2683 |   include {
2684 |     phase: TRAIN
2685 |   }
2686 | }
2687 | layer {
2688 |   name: "pool_eltwise4"
2689 |   type: "Pooling"
2690 |   bottom: "Eltwise4"
2691 |   top: "pool_eltwise4"
2692 |   pooling_param {
2693 |     pool: AVE
2694 |     global_pooling: true
2695 |   }
2696 |   include {
2697 |     phase: TRAIN
2698 |   }
2699 | }
2700 | layer {
2701 |   name: "pool_eltwise6"
2702 |   type: "Pooling"
2703 |   bottom: "Eltwise6"
2704 |   top: "pool_eltwise6"
2705 |   pooling_param {
2706 |     pool: AVE
2707 |     global_pooling: true
2708 |   }
2709 |   include {
2710 |     phase: TRAIN
2711 |   }
2712 | }
2713 | layer {
2714 |   name: "pool_eltwise7"
2715 |   type: "Pooling"
2716 |   bottom: "Eltwise7"
2717 |   top: "pool_eltwise7"
2718 |   pooling_param {
2719 |     pool: AVE
2720 |     global_pooling: true
2721 |   }
2722 |   include {
2723 |     phase: TRAIN
2724 |   }
2725 | }
2726 | layer {
2727 |   name: "bem13"
2728 |   type: "BatchEuclideanVector"
2729 |   bottom: "pool_eltwise3"
2730 |   bottom: "pool_eltwise1"
2731 |   top: "bem13"
2732 |   include {
2733 |     phase: TRAIN
2734 |   }
2735 | }
2736 | layer {
2737 |   name: "bem46"
2738 |   type: "BatchEuclideanVector"
2739 |   bottom: "pool_eltwise6"
2740 |   bottom: "pool_eltwise4"
2741 |   top: "bem46"
2742 |   include {
2743 |     phase: TRAIN
2744 |   }
2745 | }
2746 | layer {
2747 |   name: "bem79"
2748 |   type: "BatchEuclideanVector"
2749 |   bottom: "Pooling1"
2750 |   bottom: "pool_eltwise7"
2751 |   top: "bem79"
2752 |   include {
2753 |     phase: TRAIN
2754 |   }
2755 | }
2756 | layer {
2757 |   name: "stu_pool_eltwise1"
2758 |   type: "Pooling"
2759 |   bottom: "stu_Eltwise1"
2760 |   top: "stu_pool_eltwise1"
2761 |   pooling_param {
2762 |     pool: AVE
2763 |     global_pooling: true
2764 |   }
2765 |   include {
2766 |     phase: TRAIN
2767 |   }
2768 | }
2769 | layer {
2770 |   name: "stu_pool_eltwise3"
2771 |   type: "Pooling"
2772 |   bottom: "stu_Eltwise3"
2773 |   top: "stu_pool_eltwise3"
2774 |   pooling_param {
2775 |     pool: AVE
2776 |     global_pooling: true
2777 |   }
2778 |   include {
2779 |     phase: TRAIN
2780 |   }
2781 | }
2782 | layer {
2783 |   name: "stu_pool_eltwise4"
2784 |   type: "Pooling"
2785 |   bottom: "stu_Eltwise4"
2786 |   top: "stu_pool_eltwise4"
2787 |   pooling_param {
2788 |     pool: AVE
2789 |     global_pooling: true
2790 |   }
2791 |   include {
2792 |     phase: TRAIN
2793 |   }
2794 | }
2795 | layer {
2796 |   name: "stu_pool_eltwise5"
2797 |   type: "Pooling"
2798 |   bottom: "stu_Eltwise5"
2799 |   top: "stu_pool_eltwise5"
2800 |   pooling_param {
2801 |     pool: AVE
2802 |     global_pooling: true
2803 |   }
2804 |   include {
2805 |     phase: TRAIN
2806 |   }
2807 | }
2808 | layer {
2809 |   name: "stu_pool_eltwise7"
2810 |   type: "Pooling"
2811 |   bottom: "stu_Eltwise7"
2812 |   top: "stu_pool_eltwise7"
2813 |   pooling_param {
2814 |     pool: AVE
2815 |     global_pooling: true
2816 |   }
2817 |   include {
2818 |     phase: TRAIN
2819 |   }
2820 | }
2821 | layer {
2822 |   name: "stu_bem13"
2823 |   type: "BatchEuclideanVector"
2824 |   bottom: "stu_pool_eltwise3"
2825 |   bottom: "stu_pool_eltwise1"
2826 |   top: "stu_bem13"
2827 |   include {
2828 |     phase: TRAIN
2829 |   }
2830 | }
2831 | layer {
2832 |   name: "stu_bem46"
2833 |   type: "BatchEuclideanVector"
2834 |   bottom: "stu_pool_eltwise5"
2835 |   bottom: "stu_pool_eltwise4"
2836 |   top: "stu_bem46"
2837 |   include {
2838 |     phase: TRAIN
2839 |   }
2840 | }
2841 | layer {
2842 |   name: "stu_bem79"
2843 |   type: "BatchEuclideanVector"
2844 |   bottom: "stu_Pooling1"
2845 |   bottom: "stu_pool_eltwise7"
2846 |   top: "stu_bem79"
2847 |   include {
2848 |     phase: TRAIN
2849 |   }
2850 | }
2851 | layer {
2852 |   name: "bem13_loss"
2853 |   type: "EuclideanLoss"
2854 |   bottom: "bem13"
2855 |   bottom: "stu_bem13"
2856 |   top: "bem13_loss"
2857 |   loss_weight: 0.005
2858 |   include {
2859 |     phase: TRAIN
2860 |   }
2861 | }
2862 | layer {
2863 |   name: "bem46_loss"
2864 |   type: "EuclideanLoss"
2865 |   bottom: "bem46"
2866 |   bottom: "stu_bem46"
2867 |   top: "bem46_loss"
2868 |   loss_weight: 0.005
2869 |   include {
2870 |     phase: TRAIN
2871 |   }
2872 | }
2873 | layer {
2874 |   name: "bem79_loss"
2875 |   type: "EuclideanLoss"
2876 |   bottom: "bem79"
2877 |   bottom: "stu_bem79"
2878 |   top: "bem79_loss"
2879 |   loss_weight: 0.005
2880 |   include {
2881 |     phase: TRAIN
2882 |   }
2883 | }
2884 | 


--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
 1 | 
 2 | COPYRIGHT
 3 | 
 4 | All contributions by Institute of Automation, Chinese Academy of Sciences:
 5 | Copyright (c) 2018-2019 The Regents of Institute of Automation, Chinese Academy of Sciences (Regents)
 6 | All rights reserved.
 7 | 
 8 | All other contributions:
 9 | Copyright (c) 2018-2019, the respective contributors
10 | All rights reserved.
11 | 
12 | LICENSE
13 | 
14 | Redistribution and use in source and binary forms, with or without
15 | modification, are permitted provided that the following conditions are met:
16 | 
17 | 1. Redistributions of source code must retain the above copyright notice, this
18 |    list of conditions and the following disclaimer.
19 | 2. Redistributions in binary form must reproduce the above copyright notice,
20 |    this list of conditions and the following disclaimer in the documentation
21 |    and/or other materials provided with the distribution.
22 | 
23 | THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
24 | ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
25 | WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
26 | DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
27 | ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
28 | (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 | LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
30 | ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31 | (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
32 | SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33 | 
34 | CONTRIBUTION AGREEMENT
35 | 
36 | By contributing to the repository through pull-request, comment,
37 | or otherwise, the contributor releases their content to the
38 | license and copyright terms herein.
39 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # IRG for Knowledge Distillation
 2 | [![License](https://img.shields.io/badge/license-BSD-blue.svg)](LICENSE)
 3 | 
 4 | IRG is an open source implementation of the paper called "Knowledge Distillation via Instance Relationship Graph":
 5 | > Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu *et al*. Knowledge Distillation via Instance Relationship Graph. IEEE Conference on Computer Vision and Pattern Recognition (CVPR): Long Beach, USA. 2019.06.16-2019.06.20.
 6 | 
 7 | ## Abstract
 8 | *The key challenge of knowledge distillation is to extract
 9 | general, moderate and sufficient knowledge from a teacher
10 | network to guide a student network. In this paper, a novel
11 | Instance Relationship Graph (IRG) is proposed for knowledge
12 | distillation. It models three kinds of knowledge, including
13 | instance features, instance relationships and feature
14 | space transformation, while the latter two kinds of
15 | knowledge are neglected by previous methods. Firstly, the
16 | IRG is constructed to model the distilled knowledge of
17 | one network layer, by considering instance features and
18 | instance relationships as vertexes and edges respectively.
19 | Secondly, an IRG transformation is proposed to models the
20 | feature space transformation across layers. It is more moderate
21 | than directly mimicking the features at intermediate
22 | layers. Finally, hint loss functions are designed to force a
23 | student’s IRGs to mimic the structures of a teacher’s IRGs.
24 | The proposed method effectively captures the knowledge
25 | along the whole network via IRGs, and thus shows stable
26 | convergence and strong robustness to different network architectures.
27 | In addition, the proposed method shows superior
28 | performance over existing methods on datasets of various
29 | scales.*
30 | 
31 | ## Citation
32 | If you find IRG useful in your research, please cite our paper: 
33 | ````
34 | @inproceedings{liu2019knowledge,
35 |   title={Knowledge Distillation via Instance Relationship Graph},
36 |   author={Liu, Yufan and Cao, Jiajiong and Li, Bing and Yuan, Chunfeng and Hu, Weiming and Li, Yangxi and Duan, Yunqiang},
37 |   booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
38 |   pages={7096--7104},
39 |   year={2019}
40 | }
41 | ````
42 | 
43 | ## Framework
44 | ![Framework](./img/framework.png "Framework")
45 | 
46 | ## Requirements
47 | 
48 | - Caffe-windows
49 | 
50 | ## Usage
51 | 1. Put \*.cpp/\*.hpp/\*.cu files into corresponding directories of caffe.
52 | 
53 | 2. Compile caffe.
54 | ```make
55 |   make all -j 8
56 |   ```
57 | 3. Use prototxt files (e.g., IRG.prototxt) to train.
58 |    IRG.prototxt is the train prototxt of single IRG (Teacher: ResNet20; Student: ResNet20-x0.5).
59 |    IRG_Transformation.prototxt is the train prototxt of single IRG Transformation (Teacher: ResNet20; Student: ResNet20-x0.5).
60 |    ResNet-20.prototxt is the original train prototxt of ResNet20.
61 | 
62 | ## Experiments
63 | ![Experimental results](./img/Exp_results.png "Experimental results")
64 | 
65 | ## Contact
66 | If any question, please contact yufan.liu@ia.ac.cn, or use public issues section of this repository.
67 | 


--------------------------------------------------------------------------------
/ResNet-20.prototxt:
--------------------------------------------------------------------------------
   1 | name: "resnet_cifar10"
   2 | layer {
   3 |   name: "Data1"
   4 |   type: "ImageData"
   5 |   top: "Data1"
   6 |   top: "Data2"
   7 |   include {
   8 |     phase: TRAIN
   9 |   }
  10 |   transform_param {
  11 |     mean_value: 127.5
  12 |     mean_value: 127.5
  13 |     mean_value: 127.5
  14 |     scale: 0.00390625
  15 |     crop_size: 32
  16 |     mirror:true
  17 |   }
  18 |   image_data_param {
  19 |     source: "yourpath/train.txt"
  20 |     new_height: 36
  21 |     new_width: 36
  22 |     batch_size: 128
  23 |     is_color: true
  24 |     shuffle: true
  25 |   }
  26 | }
  27 | layer {
  28 |   name: "Data1"
  29 |   type: "ImageData"
  30 |   top: "Data1"
  31 |   top: "Data2"
  32 |   include {
  33 |     phase: TEST
  34 |   }
  35 |   transform_param {
  36 |     mean_value: 127.5
  37 |     mean_value: 127.5
  38 |     mean_value: 127.5
  39 |     scale: 0.00390625
  40 |   }
  41 |   image_data_param {
  42 |     source: "yourpath/val.txt"
  43 |     new_height: 36
  44 |     new_width: 36
  45 |     batch_size: 100
  46 |     is_color: true
  47 |     shuffle: false
  48 |   }
  49 | }
  50 | layer {
  51 |   name: "Convolution1"
  52 |   type: "Convolution"
  53 |   bottom: "Data1"
  54 |   top: "Convolution1"
  55 |   param {
  56 |     lr_mult: 0
  57 |     decay_mult: 0
  58 |   }
  59 |   param {
  60 |     lr_mult: 0
  61 |     decay_mult: 0
  62 |   }
  63 |   convolution_param {
  64 |     num_output: 16
  65 |     pad: 1
  66 |     kernel_size: 3
  67 |     stride: 1
  68 |     weight_filler {
  69 |       type: "msra"
  70 |     }
  71 |     bias_filler {
  72 |       type: "constant"
  73 |       value: 0
  74 |     }
  75 |   }
  76 | }
  77 | layer {
  78 |   name: "BatchNorm1"
  79 |   type: "BatchNorm"
  80 |   bottom: "Convolution1"
  81 |   top: "Convolution1"
  82 |   param {
  83 |     lr_mult: 0
  84 |     decay_mult: 0
  85 |   }
  86 |   param {
  87 |     lr_mult: 0
  88 |     decay_mult: 0
  89 |   }
  90 |   param {
  91 |     lr_mult: 0
  92 |     decay_mult: 0
  93 |   }
  94 |   batch_norm_param {
  95 |     use_global_stats: true
  96 |   }
  97 | }
  98 | layer {
  99 |   name: "Scale1"
 100 |   type: "Scale"
 101 |   bottom: "Convolution1"
 102 |   top: "Convolution1"
 103 |   param {
 104 |     lr_mult: 0
 105 |     decay_mult: 0
 106 |   }
 107 |   param {
 108 |     lr_mult: 0
 109 |     decay_mult: 0
 110 |   }
 111 |   scale_param {
 112 |     bias_term: true
 113 |   }
 114 | }
 115 | layer {
 116 |   name: "ReLU1"
 117 |   type: "ReLU"
 118 |   bottom: "Convolution1"
 119 |   top: "Convolution1"
 120 | }
 121 | layer {
 122 |   name: "Convolution2"
 123 |   type: "Convolution"
 124 |   bottom: "Convolution1"
 125 |   top: "Convolution2"
 126 |   param {
 127 |     lr_mult: 0
 128 |     decay_mult: 0
 129 |   }
 130 |   param {
 131 |     lr_mult: 0
 132 |     decay_mult: 0
 133 |   }
 134 |   convolution_param {
 135 |     num_output: 16
 136 |     pad: 1
 137 |     kernel_size: 3
 138 |     stride: 1
 139 |     weight_filler {
 140 |       type: "msra"
 141 |     }
 142 |     bias_filler {
 143 |       type: "constant"
 144 |       value: 0
 145 |     }
 146 |   }
 147 | }
 148 | layer {
 149 |   name: "BatchNorm2"
 150 |   type: "BatchNorm"
 151 |   bottom: "Convolution2"
 152 |   top: "Convolution2"
 153 |   param {
 154 |     lr_mult: 0
 155 |     decay_mult: 0
 156 |   }
 157 |   param {
 158 |     lr_mult: 0
 159 |     decay_mult: 0
 160 |   }
 161 |   param {
 162 |     lr_mult: 0
 163 |     decay_mult: 0
 164 |   }
 165 |   batch_norm_param {
 166 |     use_global_stats: true
 167 |   }
 168 | }
 169 | layer {
 170 |   name: "Scale2"
 171 |   type: "Scale"
 172 |   bottom: "Convolution2"
 173 |   top: "Convolution2"
 174 |   param {
 175 |     lr_mult: 0
 176 |     decay_mult: 0
 177 |   }
 178 |   param {
 179 |     lr_mult: 0
 180 |     decay_mult: 0
 181 |   }
 182 |   scale_param {
 183 |     bias_term: true
 184 |   }
 185 | }
 186 | layer {
 187 |   name: "ReLU2"
 188 |   type: "ReLU"
 189 |   bottom: "Convolution2"
 190 |   top: "Convolution2"
 191 | }
 192 | layer {
 193 |   name: "Convolution3"
 194 |   type: "Convolution"
 195 |   bottom: "Convolution2"
 196 |   top: "Convolution3"
 197 |   param {
 198 |     lr_mult: 0
 199 |     decay_mult: 0
 200 |   }
 201 |   param {
 202 |     lr_mult: 0
 203 |     decay_mult: 0
 204 |   }
 205 |   convolution_param {
 206 |     num_output: 16
 207 |     pad: 1
 208 |     kernel_size: 3
 209 |     stride: 1
 210 |     weight_filler {
 211 |       type: "msra"
 212 |     }
 213 |     bias_filler {
 214 |       type: "constant"
 215 |       value: 0
 216 |     }
 217 |   }
 218 | }
 219 | layer {
 220 |   name: "BatchNorm3"
 221 |   type: "BatchNorm"
 222 |   bottom: "Convolution3"
 223 |   top: "Convolution3"
 224 |   param {
 225 |     lr_mult: 0
 226 |     decay_mult: 0
 227 |   }
 228 |   param {
 229 |     lr_mult: 0
 230 |     decay_mult: 0
 231 |   }
 232 |   param {
 233 |     lr_mult: 0
 234 |     decay_mult: 0
 235 |   }
 236 |   batch_norm_param {
 237 |     use_global_stats: true
 238 |   }
 239 | }
 240 | layer {
 241 |   name: "Scale3"
 242 |   type: "Scale"
 243 |   bottom: "Convolution3"
 244 |   top: "Convolution3"
 245 |   param {
 246 |     lr_mult: 0
 247 |     decay_mult: 0
 248 |   }
 249 |   param {
 250 |     lr_mult: 0
 251 |     decay_mult: 0
 252 |   }
 253 |   scale_param {
 254 |     bias_term: true
 255 |   }
 256 | }
 257 | layer {
 258 |   name: "Eltwise1"
 259 |   type: "Eltwise"
 260 |   bottom: "Convolution1"
 261 |   bottom: "Convolution3"
 262 |   top: "Eltwise1"
 263 |   eltwise_param {
 264 |     operation: SUM
 265 |   }
 266 | }
 267 | layer {
 268 |   name: "ReLU3"
 269 |   type: "ReLU"
 270 |   bottom: "Eltwise1"
 271 |   top: "Eltwise1"
 272 | }
 273 | layer {
 274 |   name: "Convolution4"
 275 |   type: "Convolution"
 276 |   bottom: "Eltwise1"
 277 |   top: "Convolution4"
 278 |   param {
 279 |     lr_mult: 0
 280 |     decay_mult: 0
 281 |   }
 282 |   param {
 283 |     lr_mult: 0
 284 |     decay_mult: 0
 285 |   }
 286 |   convolution_param {
 287 |     num_output: 16
 288 |     pad: 1
 289 |     kernel_size: 3
 290 |     stride: 1
 291 |     weight_filler {
 292 |       type: "msra"
 293 |     }
 294 |     bias_filler {
 295 |       type: "constant"
 296 |       value: 0
 297 |     }
 298 |   }
 299 | }
 300 | layer {
 301 |   name: "BatchNorm4"
 302 |   type: "BatchNorm"
 303 |   bottom: "Convolution4"
 304 |   top: "Convolution4"
 305 |   param {
 306 |     lr_mult: 0
 307 |     decay_mult: 0
 308 |   }
 309 |   param {
 310 |     lr_mult: 0
 311 |     decay_mult: 0
 312 |   }
 313 |   param {
 314 |     lr_mult: 0
 315 |     decay_mult: 0
 316 |   }
 317 |   batch_norm_param {
 318 |     use_global_stats: true
 319 |   }
 320 | }
 321 | layer {
 322 |   name: "Scale4"
 323 |   type: "Scale"
 324 |   bottom: "Convolution4"
 325 |   top: "Convolution4"
 326 |   param {
 327 |     lr_mult: 0
 328 |     decay_mult: 0
 329 |   }
 330 |   param {
 331 |     lr_mult: 0
 332 |     decay_mult: 0
 333 |   }
 334 |   scale_param {
 335 |     bias_term: true
 336 |   }
 337 | }
 338 | layer {
 339 |   name: "ReLU4"
 340 |   type: "ReLU"
 341 |   bottom: "Convolution4"
 342 |   top: "Convolution4"
 343 | }
 344 | layer {
 345 |   name: "Convolution5"
 346 |   type: "Convolution"
 347 |   bottom: "Convolution4"
 348 |   top: "Convolution5"
 349 |   param {
 350 |     lr_mult: 0
 351 |     decay_mult: 0
 352 |   }
 353 |   param {
 354 |     lr_mult: 0
 355 |     decay_mult: 0
 356 |   }
 357 |   convolution_param {
 358 |     num_output: 16
 359 |     pad: 1
 360 |     kernel_size: 3
 361 |     stride: 1
 362 |     weight_filler {
 363 |       type: "msra"
 364 |     }
 365 |     bias_filler {
 366 |       type: "constant"
 367 |       value: 0
 368 |     }
 369 |   }
 370 | }
 371 | layer {
 372 |   name: "BatchNorm5"
 373 |   type: "BatchNorm"
 374 |   bottom: "Convolution5"
 375 |   top: "Convolution5"
 376 |   param {
 377 |     lr_mult: 0
 378 |     decay_mult: 0
 379 |   }
 380 |   param {
 381 |     lr_mult: 0
 382 |     decay_mult: 0
 383 |   }
 384 |   param {
 385 |     lr_mult: 0
 386 |     decay_mult: 0
 387 |   }
 388 |   batch_norm_param {
 389 |     use_global_stats: true
 390 |   }
 391 | }
 392 | layer {
 393 |   name: "Scale5"
 394 |   type: "Scale"
 395 |   bottom: "Convolution5"
 396 |   top: "Convolution5"
 397 |   param {
 398 |     lr_mult: 0
 399 |     decay_mult: 0
 400 |   }
 401 |   param {
 402 |     lr_mult: 0
 403 |     decay_mult: 0
 404 |   }
 405 |   scale_param {
 406 |     bias_term: true
 407 |   }
 408 | }
 409 | layer {
 410 |   name: "Eltwise2"
 411 |   type: "Eltwise"
 412 |   bottom: "Eltwise1"
 413 |   bottom: "Convolution5"
 414 |   top: "Eltwise2"
 415 |   eltwise_param {
 416 |     operation: SUM
 417 |   }
 418 | }
 419 | layer {
 420 |   name: "ReLU5"
 421 |   type: "ReLU"
 422 |   bottom: "Eltwise2"
 423 |   top: "Eltwise2"
 424 | }
 425 | layer {
 426 |   name: "Convolution6"
 427 |   type: "Convolution"
 428 |   bottom: "Eltwise2"
 429 |   top: "Convolution6"
 430 |   param {
 431 |     lr_mult: 0
 432 |     decay_mult: 0
 433 |   }
 434 |   param {
 435 |     lr_mult: 0
 436 |     decay_mult: 0
 437 |   }
 438 |   convolution_param {
 439 |     num_output: 16
 440 |     pad: 1
 441 |     kernel_size: 3
 442 |     stride: 1
 443 |     weight_filler {
 444 |       type: "msra"
 445 |     }
 446 |     bias_filler {
 447 |       type: "constant"
 448 |       value: 0
 449 |     }
 450 |   }
 451 | }
 452 | layer {
 453 |   name: "BatchNorm6"
 454 |   type: "BatchNorm"
 455 |   bottom: "Convolution6"
 456 |   top: "Convolution6"
 457 |   param {
 458 |     lr_mult: 0
 459 |     decay_mult: 0
 460 |   }
 461 |   param {
 462 |     lr_mult: 0
 463 |     decay_mult: 0
 464 |   }
 465 |   param {
 466 |     lr_mult: 0
 467 |     decay_mult: 0
 468 |   }
 469 |   batch_norm_param {
 470 |     use_global_stats: true
 471 |   }
 472 | }
 473 | layer {
 474 |   name: "Scale6"
 475 |   type: "Scale"
 476 |   bottom: "Convolution6"
 477 |   top: "Convolution6"
 478 |   param {
 479 |     lr_mult: 0
 480 |     decay_mult: 0
 481 |   }
 482 |   param {
 483 |     lr_mult: 0
 484 |     decay_mult: 0
 485 |   }
 486 |   scale_param {
 487 |     bias_term: true
 488 |   }
 489 | }
 490 | layer {
 491 |   name: "ReLU6"
 492 |   type: "ReLU"
 493 |   bottom: "Convolution6"
 494 |   top: "Convolution6"
 495 | }
 496 | layer {
 497 |   name: "Convolution7"
 498 |   type: "Convolution"
 499 |   bottom: "Convolution6"
 500 |   top: "Convolution7"
 501 |   param {
 502 |     lr_mult: 0
 503 |     decay_mult: 0
 504 |   }
 505 |   param {
 506 |     lr_mult: 0
 507 |     decay_mult: 0
 508 |   }
 509 |   convolution_param {
 510 |     num_output: 16
 511 |     pad: 1
 512 |     kernel_size: 3
 513 |     stride: 1
 514 |     weight_filler {
 515 |       type: "msra"
 516 |     }
 517 |     bias_filler {
 518 |       type: "constant"
 519 |       value: 0
 520 |     }
 521 |   }
 522 | }
 523 | layer {
 524 |   name: "BatchNorm7"
 525 |   type: "BatchNorm"
 526 |   bottom: "Convolution7"
 527 |   top: "Convolution7"
 528 |   param {
 529 |     lr_mult: 0
 530 |     decay_mult: 0
 531 |   }
 532 |   param {
 533 |     lr_mult: 0
 534 |     decay_mult: 0
 535 |   }
 536 |   param {
 537 |     lr_mult: 0
 538 |     decay_mult: 0
 539 |   }
 540 |   batch_norm_param {
 541 |     use_global_stats: true
 542 |   }
 543 | }
 544 | layer {
 545 |   name: "Scale7"
 546 |   type: "Scale"
 547 |   bottom: "Convolution7"
 548 |   top: "Convolution7"
 549 |   param {
 550 |     lr_mult: 0
 551 |     decay_mult: 0
 552 |   }
 553 |   param {
 554 |     lr_mult: 0
 555 |     decay_mult: 0
 556 |   }
 557 |   scale_param {
 558 |     bias_term: true
 559 |   }
 560 | }
 561 | layer {
 562 |   name: "Eltwise3"
 563 |   type: "Eltwise"
 564 |   bottom: "Eltwise2"
 565 |   bottom: "Convolution7"
 566 |   top: "Eltwise3"
 567 |   eltwise_param {
 568 |     operation: SUM
 569 |   }
 570 | }
 571 | layer {
 572 |   name: "ReLU7"
 573 |   type: "ReLU"
 574 |   bottom: "Eltwise3"
 575 |   top: "Eltwise3"
 576 | }
 577 | layer {
 578 |   name: "Convolution8"
 579 |   type: "Convolution"
 580 |   bottom: "Eltwise3"
 581 |   top: "Convolution8"
 582 |   param {
 583 |     lr_mult: 0
 584 |     decay_mult: 0
 585 |   }
 586 |   param {
 587 |     lr_mult: 0
 588 |     decay_mult: 0
 589 |   }
 590 |   convolution_param {
 591 |     num_output: 32
 592 |     pad: 0
 593 |     kernel_size: 1
 594 |     stride: 2
 595 |     weight_filler {
 596 |       type: "msra"
 597 |     }
 598 |     bias_filler {
 599 |       type: "constant"
 600 |       value: 0
 601 |     }
 602 |   }
 603 | }
 604 | layer {
 605 |   name: "BatchNorm8"
 606 |   type: "BatchNorm"
 607 |   bottom: "Convolution8"
 608 |   top: "Convolution8"
 609 |   param {
 610 |     lr_mult: 0
 611 |     decay_mult: 0
 612 |   }
 613 |   param {
 614 |     lr_mult: 0
 615 |     decay_mult: 0
 616 |   }
 617 |   param {
 618 |     lr_mult: 0
 619 |     decay_mult: 0
 620 |   }
 621 |   batch_norm_param {
 622 |     use_global_stats: true
 623 |   }
 624 | }
 625 | layer {
 626 |   name: "Scale8"
 627 |   type: "Scale"
 628 |   bottom: "Convolution8"
 629 |   top: "Convolution8"
 630 |   param {
 631 |     lr_mult: 0
 632 |     decay_mult: 0
 633 |   }
 634 |   param {
 635 |     lr_mult: 0
 636 |     decay_mult: 0
 637 |   }
 638 |   scale_param {
 639 |     bias_term: true
 640 |   }
 641 | }
 642 | layer {
 643 |   name: "Convolution9"
 644 |   type: "Convolution"
 645 |   bottom: "Eltwise3"
 646 |   top: "Convolution9"
 647 |   param {
 648 |     lr_mult: 0
 649 |     decay_mult: 0
 650 |   }
 651 |   param {
 652 |     lr_mult: 0
 653 |     decay_mult: 0
 654 |   }
 655 |   convolution_param {
 656 |     num_output: 32
 657 |     pad: 1
 658 |     kernel_size: 3
 659 |     stride: 2
 660 |     weight_filler {
 661 |       type: "msra"
 662 |     }
 663 |     bias_filler {
 664 |       type: "constant"
 665 |       value: 0
 666 |     }
 667 |   }
 668 | }
 669 | layer {
 670 |   name: "BatchNorm9"
 671 |   type: "BatchNorm"
 672 |   bottom: "Convolution9"
 673 |   top: "Convolution9"
 674 |   param {
 675 |     lr_mult: 0
 676 |     decay_mult: 0
 677 |   }
 678 |   param {
 679 |     lr_mult: 0
 680 |     decay_mult: 0
 681 |   }
 682 |   param {
 683 |     lr_mult: 0
 684 |     decay_mult: 0
 685 |   }
 686 |   batch_norm_param {
 687 |     use_global_stats: true
 688 |   }
 689 | }
 690 | layer {
 691 |   name: "Scale9"
 692 |   type: "Scale"
 693 |   bottom: "Convolution9"
 694 |   top: "Convolution9"
 695 |   param {
 696 |     lr_mult: 0
 697 |     decay_mult: 0
 698 |   }
 699 |   param {
 700 |     lr_mult: 0
 701 |     decay_mult: 0
 702 |   }
 703 |   scale_param {
 704 |     bias_term: true
 705 |   }
 706 | }
 707 | layer {
 708 |   name: "ReLU8"
 709 |   type: "ReLU"
 710 |   bottom: "Convolution9"
 711 |   top: "Convolution9"
 712 | }
 713 | layer {
 714 |   name: "Convolution10"
 715 |   type: "Convolution"
 716 |   bottom: "Convolution9"
 717 |   top: "Convolution10"
 718 |   param {
 719 |     lr_mult: 0
 720 |     decay_mult: 0
 721 |   }
 722 |   param {
 723 |     lr_mult: 0
 724 |     decay_mult: 0
 725 |   }
 726 |   convolution_param {
 727 |     num_output: 32
 728 |     pad: 1
 729 |     kernel_size: 3
 730 |     stride: 1
 731 |     weight_filler {
 732 |       type: "msra"
 733 |     }
 734 |     bias_filler {
 735 |       type: "constant"
 736 |       value: 0
 737 |     }
 738 |   }
 739 | }
 740 | layer {
 741 |   name: "BatchNorm10"
 742 |   type: "BatchNorm"
 743 |   bottom: "Convolution10"
 744 |   top: "Convolution10"
 745 |   param {
 746 |     lr_mult: 0
 747 |     decay_mult: 0
 748 |   }
 749 |   param {
 750 |     lr_mult: 0
 751 |     decay_mult: 0
 752 |   }
 753 |   param {
 754 |     lr_mult: 0
 755 |     decay_mult: 0
 756 |   }
 757 |   batch_norm_param {
 758 |     use_global_stats: true
 759 |   }
 760 | }
 761 | layer {
 762 |   name: "Scale10"
 763 |   type: "Scale"
 764 |   bottom: "Convolution10"
 765 |   top: "Convolution10"
 766 |   param {
 767 |     lr_mult: 0
 768 |     decay_mult: 0
 769 |   }
 770 |   param {
 771 |     lr_mult: 0
 772 |     decay_mult: 0
 773 |   }
 774 |   scale_param {
 775 |     bias_term: true
 776 |   }
 777 | }
 778 | layer {
 779 |   name: "Eltwise4"
 780 |   type: "Eltwise"
 781 |   bottom: "Convolution8"
 782 |   bottom: "Convolution10"
 783 |   top: "Eltwise4"
 784 |   eltwise_param {
 785 |     operation: SUM
 786 |   }
 787 | }
 788 | layer {
 789 |   name: "ReLU9"
 790 |   type: "ReLU"
 791 |   bottom: "Eltwise4"
 792 |   top: "Eltwise4"
 793 | }
 794 | layer {
 795 |   name: "Convolution11"
 796 |   type: "Convolution"
 797 |   bottom: "Eltwise4"
 798 |   top: "Convolution11"
 799 |   param {
 800 |     lr_mult: 0
 801 |     decay_mult: 0
 802 |   }
 803 |   param {
 804 |     lr_mult: 0
 805 |     decay_mult: 0
 806 |   }
 807 |   convolution_param {
 808 |     num_output: 32
 809 |     pad: 1
 810 |     kernel_size: 3
 811 |     stride: 1
 812 |     weight_filler {
 813 |       type: "msra"
 814 |     }
 815 |     bias_filler {
 816 |       type: "constant"
 817 |       value: 0
 818 |     }
 819 |   }
 820 | }
 821 | layer {
 822 |   name: "BatchNorm11"
 823 |   type: "BatchNorm"
 824 |   bottom: "Convolution11"
 825 |   top: "Convolution11"
 826 |   param {
 827 |     lr_mult: 0
 828 |     decay_mult: 0
 829 |   }
 830 |   param {
 831 |     lr_mult: 0
 832 |     decay_mult: 0
 833 |   }
 834 |   param {
 835 |     lr_mult: 0
 836 |     decay_mult: 0
 837 |   }
 838 |   batch_norm_param {
 839 |     use_global_stats: true
 840 |   }
 841 | }
 842 | layer {
 843 |   name: "Scale11"
 844 |   type: "Scale"
 845 |   bottom: "Convolution11"
 846 |   top: "Convolution11"
 847 |   param {
 848 |     lr_mult: 0
 849 |     decay_mult: 0
 850 |   }
 851 |   param {
 852 |     lr_mult: 0
 853 |     decay_mult: 0
 854 |   }
 855 |   scale_param {
 856 |     bias_term: true
 857 |   }
 858 | }
 859 | layer {
 860 |   name: "ReLU10"
 861 |   type: "ReLU"
 862 |   bottom: "Convolution11"
 863 |   top: "Convolution11"
 864 | }
 865 | layer {
 866 |   name: "Convolution12"
 867 |   type: "Convolution"
 868 |   bottom: "Convolution11"
 869 |   top: "Convolution12"
 870 |   param {
 871 |     lr_mult: 0
 872 |     decay_mult: 0
 873 |   }
 874 |   param {
 875 |     lr_mult: 0
 876 |     decay_mult: 0
 877 |   }
 878 |   convolution_param {
 879 |     num_output: 32
 880 |     pad: 1
 881 |     kernel_size: 3
 882 |     stride: 1
 883 |     weight_filler {
 884 |       type: "msra"
 885 |     }
 886 |     bias_filler {
 887 |       type: "constant"
 888 |       value: 0
 889 |     }
 890 |   }
 891 | }
 892 | layer {
 893 |   name: "BatchNorm12"
 894 |   type: "BatchNorm"
 895 |   bottom: "Convolution12"
 896 |   top: "Convolution12"
 897 |   param {
 898 |     lr_mult: 0
 899 |     decay_mult: 0
 900 |   }
 901 |   param {
 902 |     lr_mult: 0
 903 |     decay_mult: 0
 904 |   }
 905 |   param {
 906 |     lr_mult: 0
 907 |     decay_mult: 0
 908 |   }
 909 |   batch_norm_param {
 910 |     use_global_stats: true
 911 |   }
 912 | }
 913 | layer {
 914 |   name: "Scale12"
 915 |   type: "Scale"
 916 |   bottom: "Convolution12"
 917 |   top: "Convolution12"
 918 |   param {
 919 |     lr_mult: 0
 920 |     decay_mult: 0
 921 |   }
 922 |   param {
 923 |     lr_mult: 0
 924 |     decay_mult: 0
 925 |   }
 926 |   scale_param {
 927 |     bias_term: true
 928 |   }
 929 | }
 930 | layer {
 931 |   name: "Eltwise5"
 932 |   type: "Eltwise"
 933 |   bottom: "Eltwise4"
 934 |   bottom: "Convolution12"
 935 |   top: "Eltwise5"
 936 |   eltwise_param {
 937 |     operation: SUM
 938 |   }
 939 | }
 940 | layer {
 941 |   name: "ReLU11"
 942 |   type: "ReLU"
 943 |   bottom: "Eltwise5"
 944 |   top: "Eltwise5"
 945 | }
 946 | layer {
 947 |   name: "Convolution13"
 948 |   type: "Convolution"
 949 |   bottom: "Eltwise5"
 950 |   top: "Convolution13"
 951 |   param {
 952 |     lr_mult: 0
 953 |     decay_mult: 0
 954 |   }
 955 |   param {
 956 |     lr_mult: 0
 957 |     decay_mult: 0
 958 |   }
 959 |   convolution_param {
 960 |     num_output: 32
 961 |     pad: 1
 962 |     kernel_size: 3
 963 |     stride: 1
 964 |     weight_filler {
 965 |       type: "msra"
 966 |     }
 967 |     bias_filler {
 968 |       type: "constant"
 969 |       value: 0
 970 |     }
 971 |   }
 972 | }
 973 | layer {
 974 |   name: "BatchNorm13"
 975 |   type: "BatchNorm"
 976 |   bottom: "Convolution13"
 977 |   top: "Convolution13"
 978 |   param {
 979 |     lr_mult: 0
 980 |     decay_mult: 0
 981 |   }
 982 |   param {
 983 |     lr_mult: 0
 984 |     decay_mult: 0
 985 |   }
 986 |   param {
 987 |     lr_mult: 0
 988 |     decay_mult: 0
 989 |   }
 990 |   batch_norm_param {
 991 |     use_global_stats: true
 992 |   }
 993 | }
 994 | layer {
 995 |   name: "Scale13"
 996 |   type: "Scale"
 997 |   bottom: "Convolution13"
 998 |   top: "Convolution13"
 999 |   param {
1000 |     lr_mult: 0
1001 |     decay_mult: 0
1002 |   }
1003 |   param {
1004 |     lr_mult: 0
1005 |     decay_mult: 0
1006 |   }
1007 |   scale_param {
1008 |     bias_term: true
1009 |   }
1010 | }
1011 | layer {
1012 |   name: "ReLU12"
1013 |   type: "ReLU"
1014 |   bottom: "Convolution13"
1015 |   top: "Convolution13"
1016 | }
1017 | layer {
1018 |   name: "Convolution14"
1019 |   type: "Convolution"
1020 |   bottom: "Convolution13"
1021 |   top: "Convolution14"
1022 |   param {
1023 |     lr_mult: 0
1024 |     decay_mult: 0
1025 |   }
1026 |   param {
1027 |     lr_mult: 0
1028 |     decay_mult: 0
1029 |   }
1030 |   convolution_param {
1031 |     num_output: 32
1032 |     pad: 1
1033 |     kernel_size: 3
1034 |     stride: 1
1035 |     weight_filler {
1036 |       type: "msra"
1037 |     }
1038 |     bias_filler {
1039 |       type: "constant"
1040 |       value: 0
1041 |     }
1042 |   }
1043 | }
1044 | layer {
1045 |   name: "BatchNorm14"
1046 |   type: "BatchNorm"
1047 |   bottom: "Convolution14"
1048 |   top: "Convolution14"
1049 |   param {
1050 |     lr_mult: 0
1051 |     decay_mult: 0
1052 |   }
1053 |   param {
1054 |     lr_mult: 0
1055 |     decay_mult: 0
1056 |   }
1057 |   param {
1058 |     lr_mult: 0
1059 |     decay_mult: 0
1060 |   }
1061 |   batch_norm_param {
1062 |     use_global_stats: true
1063 |   }
1064 | }
1065 | layer {
1066 |   name: "Scale14"
1067 |   type: "Scale"
1068 |   bottom: "Convolution14"
1069 |   top: "Convolution14"
1070 |   param {
1071 |     lr_mult: 0
1072 |     decay_mult: 0
1073 |   }
1074 |   param {
1075 |     lr_mult: 0
1076 |     decay_mult: 0
1077 |   }
1078 |   scale_param {
1079 |     bias_term: true
1080 |   }
1081 | }
1082 | layer {
1083 |   name: "Eltwise6"
1084 |   type: "Eltwise"
1085 |   bottom: "Eltwise5"
1086 |   bottom: "Convolution14"
1087 |   top: "Eltwise6"
1088 |   eltwise_param {
1089 |     operation: SUM
1090 |   }
1091 | }
1092 | layer {
1093 |   name: "ReLU13"
1094 |   type: "ReLU"
1095 |   bottom: "Eltwise6"
1096 |   top: "Eltwise6"
1097 | }
1098 | layer {
1099 |   name: "Convolution15"
1100 |   type: "Convolution"
1101 |   bottom: "Eltwise6"
1102 |   top: "Convolution15"
1103 |   param {
1104 |     lr_mult: 0
1105 |     decay_mult: 0
1106 |   }
1107 |   param {
1108 |     lr_mult: 0
1109 |     decay_mult: 0
1110 |   }
1111 |   convolution_param {
1112 |     num_output: 64
1113 |     pad: 0
1114 |     kernel_size: 1
1115 |     stride: 2
1116 |     weight_filler {
1117 |       type: "msra"
1118 |     }
1119 |     bias_filler {
1120 |       type: "constant"
1121 |       value: 0
1122 |     }
1123 |   }
1124 | }
1125 | layer {
1126 |   name: "BatchNorm15"
1127 |   type: "BatchNorm"
1128 |   bottom: "Convolution15"
1129 |   top: "Convolution15"
1130 |   param {
1131 |     lr_mult: 0
1132 |     decay_mult: 0
1133 |   }
1134 |   param {
1135 |     lr_mult: 0
1136 |     decay_mult: 0
1137 |   }
1138 |   param {
1139 |     lr_mult: 0
1140 |     decay_mult: 0
1141 |   }
1142 |   batch_norm_param {
1143 |     use_global_stats: true
1144 |   }
1145 | }
1146 | layer {
1147 |   name: "Scale15"
1148 |   type: "Scale"
1149 |   bottom: "Convolution15"
1150 |   top: "Convolution15"
1151 |   param {
1152 |     lr_mult: 0
1153 |     decay_mult: 0
1154 |   }
1155 |   param {
1156 |     lr_mult: 0
1157 |     decay_mult: 0
1158 |   }
1159 |   scale_param {
1160 |     bias_term: true
1161 |   }
1162 | }
1163 | layer {
1164 |   name: "Convolution16"
1165 |   type: "Convolution"
1166 |   bottom: "Eltwise6"
1167 |   top: "Convolution16"
1168 |   param {
1169 |     lr_mult: 0
1170 |     decay_mult: 0
1171 |   }
1172 |   param {
1173 |     lr_mult: 0
1174 |     decay_mult: 0
1175 |   }
1176 |   convolution_param {
1177 |     num_output: 64
1178 |     pad: 1
1179 |     kernel_size: 3
1180 |     stride: 2
1181 |     weight_filler {
1182 |       type: "msra"
1183 |     }
1184 |     bias_filler {
1185 |       type: "constant"
1186 |       value: 0
1187 |     }
1188 |   }
1189 | }
1190 | layer {
1191 |   name: "BatchNorm16"
1192 |   type: "BatchNorm"
1193 |   bottom: "Convolution16"
1194 |   top: "Convolution16"
1195 |   param {
1196 |     lr_mult: 0
1197 |     decay_mult: 0
1198 |   }
1199 |   param {
1200 |     lr_mult: 0
1201 |     decay_mult: 0
1202 |   }
1203 |   param {
1204 |     lr_mult: 0
1205 |     decay_mult: 0
1206 |   }
1207 |   batch_norm_param {
1208 |     use_global_stats: true
1209 |   }
1210 | }
1211 | layer {
1212 |   name: "Scale16"
1213 |   type: "Scale"
1214 |   bottom: "Convolution16"
1215 |   top: "Convolution16"
1216 |   param {
1217 |     lr_mult: 0
1218 |     decay_mult: 0
1219 |   }
1220 |   param {
1221 |     lr_mult: 0
1222 |     decay_mult: 0
1223 |   }
1224 |   scale_param {
1225 |     bias_term: true
1226 |   }
1227 | }
1228 | layer {
1229 |   name: "ReLU14"
1230 |   type: "ReLU"
1231 |   bottom: "Convolution16"
1232 |   top: "Convolution16"
1233 | }
1234 | layer {
1235 |   name: "Convolution17"
1236 |   type: "Convolution"
1237 |   bottom: "Convolution16"
1238 |   top: "Convolution17"
1239 |   param {
1240 |     lr_mult: 0
1241 |     decay_mult: 0
1242 |   }
1243 |   param {
1244 |     lr_mult: 0
1245 |     decay_mult: 0
1246 |   }
1247 |   convolution_param {
1248 |     num_output: 64
1249 |     pad: 1
1250 |     kernel_size: 3
1251 |     stride: 1
1252 |     weight_filler {
1253 |       type: "msra"
1254 |     }
1255 |     bias_filler {
1256 |       type: "constant"
1257 |       value: 0
1258 |     }
1259 |   }
1260 | }
1261 | layer {
1262 |   name: "BatchNorm17"
1263 |   type: "BatchNorm"
1264 |   bottom: "Convolution17"
1265 |   top: "Convolution17"
1266 |   param {
1267 |     lr_mult: 0
1268 |     decay_mult: 0
1269 |   }
1270 |   param {
1271 |     lr_mult: 0
1272 |     decay_mult: 0
1273 |   }
1274 |   param {
1275 |     lr_mult: 0
1276 |     decay_mult: 0
1277 |   }
1278 |   batch_norm_param {
1279 |     use_global_stats: true
1280 |   }
1281 | }
1282 | layer {
1283 |   name: "Scale17"
1284 |   type: "Scale"
1285 |   bottom: "Convolution17"
1286 |   top: "Convolution17"
1287 |   param {
1288 |     lr_mult: 0
1289 |     decay_mult: 0
1290 |   }
1291 |   param {
1292 |     lr_mult: 0
1293 |     decay_mult: 0
1294 |   }
1295 |   scale_param {
1296 |     bias_term: true
1297 |   }
1298 | }
1299 | layer {
1300 |   name: "Eltwise7"
1301 |   type: "Eltwise"
1302 |   bottom: "Convolution15"
1303 |   bottom: "Convolution17"
1304 |   top: "Eltwise7"
1305 |   eltwise_param {
1306 |     operation: SUM
1307 |   }
1308 | }
1309 | layer {
1310 |   name: "ReLU15"
1311 |   type: "ReLU"
1312 |   bottom: "Eltwise7"
1313 |   top: "Eltwise7"
1314 | }
1315 | layer {
1316 |   name: "Convolution18"
1317 |   type: "Convolution"
1318 |   bottom: "Eltwise7"
1319 |   top: "Convolution18"
1320 |   param {
1321 |     lr_mult: 0
1322 |     decay_mult: 0
1323 |   }
1324 |   param {
1325 |     lr_mult: 0
1326 |     decay_mult: 0
1327 |   }
1328 |   convolution_param {
1329 |     num_output: 64
1330 |     pad: 1
1331 |     kernel_size: 3
1332 |     stride: 1
1333 |     weight_filler {
1334 |       type: "msra"
1335 |     }
1336 |     bias_filler {
1337 |       type: "constant"
1338 |       value: 0
1339 |     }
1340 |   }
1341 | }
1342 | layer {
1343 |   name: "BatchNorm18"
1344 |   type: "BatchNorm"
1345 |   bottom: "Convolution18"
1346 |   top: "Convolution18"
1347 |   param {
1348 |     lr_mult: 0
1349 |     decay_mult: 0
1350 |   }
1351 |   param {
1352 |     lr_mult: 0
1353 |     decay_mult: 0
1354 |   }
1355 |   param {
1356 |     lr_mult: 0
1357 |     decay_mult: 0
1358 |   }
1359 |   batch_norm_param {
1360 |     use_global_stats: true
1361 |   }
1362 | }
1363 | layer {
1364 |   name: "Scale18"
1365 |   type: "Scale"
1366 |   bottom: "Convolution18"
1367 |   top: "Convolution18"
1368 |   param {
1369 |     lr_mult: 0
1370 |     decay_mult: 0
1371 |   }
1372 |   param {
1373 |     lr_mult: 0
1374 |     decay_mult: 0
1375 |   }
1376 |   scale_param {
1377 |     bias_term: true
1378 |   }
1379 | }
1380 | layer {
1381 |   name: "ReLU16"
1382 |   type: "ReLU"
1383 |   bottom: "Convolution18"
1384 |   top: "Convolution18"
1385 | }
1386 | layer {
1387 |   name: "Convolution19"
1388 |   type: "Convolution"
1389 |   bottom: "Convolution18"
1390 |   top: "Convolution19"
1391 |   param {
1392 |     lr_mult: 0
1393 |     decay_mult: 0
1394 |   }
1395 |   param {
1396 |     lr_mult: 0
1397 |     decay_mult: 0
1398 |   }
1399 |   convolution_param {
1400 |     num_output: 64
1401 |     pad: 1
1402 |     kernel_size: 3
1403 |     stride: 1
1404 |     weight_filler {
1405 |       type: "msra"
1406 |     }
1407 |     bias_filler {
1408 |       type: "constant"
1409 |       value: 0
1410 |     }
1411 |   }
1412 | }
1413 | layer {
1414 |   name: "BatchNorm19"
1415 |   type: "BatchNorm"
1416 |   bottom: "Convolution19"
1417 |   top: "Convolution19"
1418 |   param {
1419 |     lr_mult: 0
1420 |     decay_mult: 0
1421 |   }
1422 |   param {
1423 |     lr_mult: 0
1424 |     decay_mult: 0
1425 |   }
1426 |   param {
1427 |     lr_mult: 0
1428 |     decay_mult: 0
1429 |   }
1430 |   batch_norm_param {
1431 |     use_global_stats: true
1432 |   }
1433 | }
1434 | layer {
1435 |   name: "Scale19"
1436 |   type: "Scale"
1437 |   bottom: "Convolution19"
1438 |   top: "Convolution19"
1439 |   param {
1440 |     lr_mult: 0
1441 |     decay_mult: 0
1442 |   }
1443 |   param {
1444 |     lr_mult: 0
1445 |     decay_mult: 0
1446 |   }
1447 |   scale_param {
1448 |     bias_term: true
1449 |   }
1450 | }
1451 | layer {
1452 |   name: "Eltwise8"
1453 |   type: "Eltwise"
1454 |   bottom: "Eltwise7"
1455 |   bottom: "Convolution19"
1456 |   top: "Eltwise8"
1457 |   eltwise_param {
1458 |     operation: SUM
1459 |   }
1460 | }
1461 | layer {
1462 |   name: "ReLU17"
1463 |   type: "ReLU"
1464 |   bottom: "Eltwise8"
1465 |   top: "Eltwise8"
1466 | }
1467 | layer {
1468 |   name: "Convolution20"
1469 |   type: "Convolution"
1470 |   bottom: "Eltwise8"
1471 |   top: "Convolution20"
1472 |   param {
1473 |     lr_mult: 0
1474 |     decay_mult: 0
1475 |   }
1476 |   param {
1477 |     lr_mult: 0
1478 |     decay_mult: 0
1479 |   }
1480 |   convolution_param {
1481 |     num_output: 64
1482 |     pad: 1
1483 |     kernel_size: 3
1484 |     stride: 1
1485 |     weight_filler {
1486 |       type: "msra"
1487 |     }
1488 |     bias_filler {
1489 |       type: "constant"
1490 |       value: 0
1491 |     }
1492 |   }
1493 | }
1494 | layer {
1495 |   name: "BatchNorm20"
1496 |   type: "BatchNorm"
1497 |   bottom: "Convolution20"
1498 |   top: "Convolution20"
1499 |   param {
1500 |     lr_mult: 0
1501 |     decay_mult: 0
1502 |   }
1503 |   param {
1504 |     lr_mult: 0
1505 |     decay_mult: 0
1506 |   }
1507 |   param {
1508 |     lr_mult: 0
1509 |     decay_mult: 0
1510 |   }
1511 |   batch_norm_param {
1512 |     use_global_stats: true
1513 |   }
1514 | }
1515 | layer {
1516 |   name: "Scale20"
1517 |   type: "Scale"
1518 |   bottom: "Convolution20"
1519 |   top: "Convolution20"
1520 |   param {
1521 |     lr_mult: 0
1522 |     decay_mult: 0
1523 |   }
1524 |   param {
1525 |     lr_mult: 0
1526 |     decay_mult: 0
1527 |   }
1528 |   scale_param {
1529 |     bias_term: true
1530 |   }
1531 | }
1532 | layer {
1533 |   name: "ReLU18"
1534 |   type: "ReLU"
1535 |   bottom: "Convolution20"
1536 |   top: "Convolution20"
1537 | }
1538 | layer {
1539 |   name: "Convolution21"
1540 |   type: "Convolution"
1541 |   bottom: "Convolution20"
1542 |   top: "Convolution21"
1543 |   param {
1544 |     lr_mult: 0
1545 |     decay_mult: 0
1546 |   }
1547 |   param {
1548 |     lr_mult: 0
1549 |     decay_mult: 0
1550 |   }
1551 |   convolution_param {
1552 |     num_output: 64
1553 |     pad: 1
1554 |     kernel_size: 3
1555 |     stride: 1
1556 |     weight_filler {
1557 |       type: "msra"
1558 |     }
1559 |     bias_filler {
1560 |       type: "constant"
1561 |       value: 0
1562 |     }
1563 |   }
1564 | }
1565 | layer {
1566 |   name: "BatchNorm21"
1567 |   type: "BatchNorm"
1568 |   bottom: "Convolution21"
1569 |   top: "Convolution21"
1570 |   param {
1571 |     lr_mult: 0
1572 |     decay_mult: 0
1573 |   }
1574 |   param {
1575 |     lr_mult: 0
1576 |     decay_mult: 0
1577 |   }
1578 |   param {
1579 |     lr_mult: 0
1580 |     decay_mult: 0
1581 |   }
1582 |   batch_norm_param {
1583 |     use_global_stats: true
1584 |   }
1585 | }
1586 | layer {
1587 |   name: "Scale21"
1588 |   type: "Scale"
1589 |   bottom: "Convolution21"
1590 |   top: "Convolution21"
1591 |   param {
1592 |     lr_mult: 0
1593 |     decay_mult: 0
1594 |   }
1595 |   param {
1596 |     lr_mult: 0
1597 |     decay_mult: 0
1598 |   }
1599 |   scale_param {
1600 |     bias_term: true
1601 |   }
1602 | }
1603 | layer {
1604 |   name: "Eltwise9"
1605 |   type: "Eltwise"
1606 |   bottom: "Eltwise8"
1607 |   bottom: "Convolution21"
1608 |   top: "Eltwise9"
1609 |   eltwise_param {
1610 |     operation: SUM
1611 |   }
1612 | }
1613 | layer {
1614 |   name: "ReLU19"
1615 |   type: "ReLU"
1616 |   bottom: "Eltwise9"
1617 |   top: "Eltwise9"
1618 | }
1619 | layer {
1620 |   name: "Pooling1"
1621 |   type: "Pooling"
1622 |   bottom: "Eltwise9"
1623 |   top: "Pooling1"
1624 |   pooling_param {
1625 |     pool: AVE
1626 |     global_pooling: true
1627 |   }
1628 | }
1629 | layer {
1630 |   name: "InnerProduct1"
1631 |   type: "InnerProduct"
1632 |   bottom: "Pooling1"
1633 |   top: "InnerProduct1"
1634 |   param {
1635 |     lr_mult: 0
1636 |     decay_mult: 0
1637 |   }
1638 |   param {
1639 |     lr_mult: 0
1640 |     decay_mult: 0
1641 |   }
1642 |   inner_product_param {
1643 |     num_output: 10
1644 |     weight_filler {
1645 |       type: "msra"
1646 |     }
1647 |     bias_filler {
1648 |       type: "constant"
1649 |       value: 0
1650 |     }
1651 |   }
1652 | }
1653 | layer {
1654 |   name: "SoftmaxWithLoss1"
1655 |   type: "SoftmaxWithLoss"
1656 |   bottom: "InnerProduct1"
1657 |   bottom: "Data2"
1658 |   top: "SoftmaxWithLoss1"
1659 | }
1660 | layer {
1661 |   name: "Accuracy1"
1662 |   type: "Accuracy"
1663 |   bottom: "InnerProduct1"
1664 |   bottom: "Data2"
1665 |   top: "Accuracy1"
1666 |   include {
1667 |     phase: TEST
1668 |   }
1669 | }
1670 | 


--------------------------------------------------------------------------------
/batch_euclidean_map_layer.cpp:
--------------------------------------------------------------------------------
 1 | #include <algorithm>
 2 | #include <vector>
 3 | 
 4 | #include "caffe/layer.hpp"
 5 | #include "caffe/util/io.hpp"
 6 | #include "caffe/util/math_functions.hpp"
 7 | #include "caffe/layers/batch_euclidean_map_layer.hpp"
 8 | 
 9 | namespace caffe {
10 |   
11 |   template <typename Dtype>
12 |   void BatchEuclideanMapLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
13 |       const vector<Blob<Dtype>*>& top){
14 |   	num_ = bottom[0]->num();
15 |   	channels_ = bottom[0]->channels();
16 |   	CHECK_EQ(bottom[0]->height(), 1);
17 |   	CHECK_EQ(bottom[0]->width(), 1);
18 |   }
19 | 
20 |   template <typename Dtype>
21 |   void BatchEuclideanMapLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
22 |       const vector<Blob<Dtype>*>& top){
23 |   	diff_feat.Reshape(1, bottom[0]->channels(), 1, 1);
24 |   	top[0]->Reshape(bottom[0]->num(), bottom[0]->num(), 1, 1);
25 |   }
26 | 
27 |   template <typename Dtype>
28 |   void BatchEuclideanMapLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
29 |       const vector<Blob<Dtype>*>& top){
30 |     const Dtype* bottom_data = bottom[0]->cpu_data();
31 |     Dtype* top_data = top[0]->mutable_cpu_data();
32 |     Dtype* diff_feat_data = diff_feat.mutable_cpu_data();
33 | 	max_d = Dtype(0.0);
34 | 
35 |     for (int n = 0; n < num_; ++n){
36 |       for (int nn = 0; nn < num_; ++nn){
37 |       	//diff_feat = x_n - x_nn
38 |       	caffe_sub(channels_, bottom_data+n*channels_, bottom_data+nn*channels_, diff_feat_data);
39 |       	//sim = diff_feat * diff_feat
40 |       	Dtype distance = caffe_cpu_dot(channels_, diff_feat_data, diff_feat_data);
41 | 		if (distance > max_d){
42 | 			max_d = distance;
43 | 		}
44 |       	//top[n, nn] = sim
45 |       	caffe_set(1, distance, top_data+n*num_+nn);
46 |       }
47 |     }
48 | 	caffe_scal(num_*num_, Dtype(1.0) / max_d, top_data);
49 |   }
50 | 
51 |   template <typename Dtype>
52 |   void BatchEuclideanMapLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
53 |       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom){
54 |   	if (!propagate_down[0]){return;}
55 |   	const Dtype* top_diff = top[0]->cpu_diff();
56 |   	Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
57 |   	caffe_set(num_*channels_, Dtype(0.0), bottom_diff);
58 |   	Dtype* diff_feat_data = diff_feat.mutable_cpu_data();
59 |   	const Dtype* bottom_data = bottom[0]->cpu_data();
60 |   	Dtype scale = Dtype(0.0);
61 |   	for (int n = 0; n < num_; ++n){
62 |       for (int nn = 0; nn < num_; ++nn){
63 |       	caffe_sub(channels_, bottom_data+n*channels_, bottom_data+nn*channels_, diff_feat_data);
64 |         caffe_copy(1, top_diff+n*num_+nn, &scale);
65 | 		//const Dtype* diff_feat_data2 = diff_feat.cpu_data();
66 |         caffe_axpy(channels_, scale*Dtype(2.0) / max_d, diff_feat_data, bottom_diff+n*channels_);
67 |       }
68 |   	}
69 |   }
70 | 
71 | 
72 |   #ifdef CPU_ONLY
73 |     STUB_GPU(BatchEuclideanMapLayer);
74 |   #endif
75 |   INSTANTIATE_CLASS(BatchEuclideanMapLayer);
76 |   REGISTER_LAYER_CLASS(BatchEuclideanMap);
77 | } // namespace caffe
78 | 


--------------------------------------------------------------------------------
/batch_euclidean_map_layer.cu:
--------------------------------------------------------------------------------
 1 | #include <algorithm>
 2 | #include <vector>
 3 | 
 4 | #include "caffe/layers/batch_euclidean_map_layer.hpp"
 5 | 
 6 | namespace caffe {
 7 |   template <typename Dtype>
 8 |   void BatchEuclideanMapLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top){
 9 |   	const Dtype* bottom_data = bottom[0]->gpu_data();
10 |     Dtype* top_data = top[0]->mutable_gpu_data();
11 |     Dtype* diff_feat_data = diff_feat.mutable_gpu_data();
12 | 
13 |     max_d = Dtype(0.0);
14 |     for (int n = 0; n < num_; ++n){
15 |       for (int nn = 0; nn < num_; ++nn){
16 |       	//diff_feat = x_n - x_nn
17 |       	caffe_gpu_sub(channels_, bottom_data+n*channels_, bottom_data+nn*channels_, diff_feat_data);
18 |       	//sim = diff_feat * diff_feat
19 |       	Dtype distance = Dtype(0.0);
20 | 		caffe_gpu_dot(channels_, diff_feat_data, diff_feat_data, &distance);
21 | 		if (distance > max_d){
22 | 			max_d = distance;
23 | 		}
24 |       	//top[n, nn] = sim
25 |       	caffe_gpu_set(1, distance, top_data+n*num_+nn);
26 |       }
27 |     }
28 | 	caffe_gpu_scal(num_*num_, Dtype(1.0) / max_d, top_data);
29 |   }
30 | 
31 |   template <typename Dtype>
32 |   void BatchEuclideanMapLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom){
33 |   	if (!propagate_down[0]){return;}
34 |   	const Dtype* top_diff = top[0]->gpu_diff();
35 |   	Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
36 |   	caffe_gpu_set(num_*channels_, Dtype(0.0), bottom_diff);
37 |   	Dtype* diff_feat_data = diff_feat.mutable_gpu_data();
38 |   	const Dtype* bottom_data = bottom[0]->gpu_data();
39 |   	Dtype scale = Dtype(0.0);
40 |   	for (int n = 0; n < num_; ++n){
41 |       for (int nn = 0; nn < num_; ++nn){
42 |       	caffe_gpu_sub(channels_, bottom_data+n*channels_, bottom_data+nn*channels_, diff_feat_data);
43 |         caffe_copy(1, top_diff+n*num_+nn, &scale);
44 | 		//const Dtype* diff_feat_data2 = diff_feat.gpu_data();
45 |         caffe_gpu_axpy(channels_, scale*Dtype(2.0) / max_d, diff_feat_data, bottom_diff+n*channels_);
46 |       }
47 |   	}
48 |   }
49 |   INSTANTIATE_LAYER_GPU_FUNCS(BatchEuclideanMapLayer);
50 | }
51 | 


--------------------------------------------------------------------------------
/batch_euclidean_map_layer.hpp:
--------------------------------------------------------------------------------
 1 | #ifndef CAFFE_BATCH_EUCLIDEAN_MAP_LAYER_HPP_
 2 | #define CAFFE_BATCH_EUCLIDEAN_MAP_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 | namespace caffe {
11 |     template <typename Dtype>
12 |     class BatchEuclideanMapLayer : public Layer<Dtype> {
13 |     public:
14 |         explicit BatchEuclideanMapLayer(const LayerParameter& param) : Layer<Dtype>(param) {}
15 |         virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
16 | 		virtual void Reshape(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
17 |         virtual inline const char* type() const { return "BatchEuclideanMap"; }
18 |         virtual inline int ExactNumBottomBlobs() const { return 1; }
19 |         virtual inline int MinTopBlobs() const { return 1; }
20 |         virtual inline int MaxTopBlobs() const { return 1; }
21 |     protected:
22 |         virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
23 |         virtual void Backward_cpu(const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
24 |         virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
25 |         virtual void Backward_gpu(const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
26 |         int channels_;
27 |         int num_;
28 |         Blob<Dtype> diff_feat;
29 | 		Dtype max_d;
30 | 
31 | 	};
32 | }  // namespace caffe
33 | 
34 | #endif //CAFFE_BatchEuclideanMap_LAYER_HPP_
35 | 


--------------------------------------------------------------------------------
/batch_euclidean_vector_layer.cpp:
--------------------------------------------------------------------------------
 1 | #include <algorithm>
 2 | #include <vector>
 3 | 
 4 | #include "caffe/layer.hpp"
 5 | #include "caffe/util/io.hpp"
 6 | #include "caffe/util/math_functions.hpp"
 7 | #include "caffe/layers/batch_euclidean_vector_layer.hpp"
 8 | 
 9 | namespace caffe {
10 |   
11 |   template <typename Dtype>
12 |   void BatchEuclideanVectorLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
13 |       const vector<Blob<Dtype>*>& top){
14 |   	num_ = bottom[0]->num();
15 |   	channels_ = bottom[0]->channels();
16 |   	CHECK_EQ(bottom[0]->height(), 1);
17 |   	CHECK_EQ(bottom[0]->width(), 1);
18 |     CHECK_EQ(bottom[1]->height(), 1);
19 |     CHECK_EQ(bottom[1]->width(), 1);
20 |     CHECK_EQ(channels_, bottom[1]->channels());
21 |   }
22 | 
23 |   template <typename Dtype>
24 |   void BatchEuclideanVectorLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
25 |       const vector<Blob<Dtype>*>& top){
26 |   	diff_feat.Reshape(1, channels_, 1, 1);
27 |     top[0]->Reshape(1, 1, 1, num_);
28 |   }
29 | 
30 |   template <typename Dtype>
31 |   void BatchEuclideanVectorLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
32 |       const vector<Blob<Dtype>*>& top){
33 |     const Dtype* bottom_data1 = bottom[0]->cpu_data();
34 |     const Dtype* bottom_data2 = bottom[1]->cpu_data();
35 |     Dtype* top_data = top[0]->mutable_cpu_data();
36 |     Dtype* diff_feat_data = diff_feat.mutable_cpu_data();
37 |     max_d = Dtype(0.0);
38 |     for(int n = 0; n < num_; ++n){
39 |       caffe_sub(channels_, bottom_data1+n*channels_, bottom_data2+n*channels_, diff_feat_data);
40 |       Dtype distance = caffe_cpu_dot(channels_, diff_feat_data, diff_feat_data);
41 |       if (distance > max_d){
42 |         max_d = distance;
43 |       }
44 |       caffe_set(1, distance, top_data+n);
45 |     }
46 |     caffe_scal(num_, Dtype(1.0) / max_d, top_data);
47 |   }
48 | 
49 |   template <typename Dtype>
50 |   void BatchEuclideanVectorLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
51 |       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom){
52 |   	if (!propagate_down[0]){return;}
53 |   	const Dtype* top_diff = top[0]->cpu_diff();
54 |   	Dtype* bottom_diff1 = bottom[0]->mutable_cpu_diff();
55 |     Dtype* bottom_diff2 = bottom[1]->mutable_cpu_diff();
56 |   	caffe_set(num_*channels_, Dtype(0.0), bottom_diff1);
57 |     caffe_set(num_*channels_, Dtype(0.0), bottom_diff2);
58 |   	Dtype* diff_feat_data = diff_feat.mutable_cpu_data();
59 |   	const Dtype* bottom_data1 = bottom[0]->cpu_data();
60 |     const Dtype* bottom_data2 = bottom[1]->cpu_data();
61 |   	Dtype scale = Dtype(0.0);
62 |       for (int n = 0; n < num_; ++n){
63 |         caffe_sub(channels_, bottom_data1+n*channels_, bottom_data2+n*channels_, diff_feat_data);
64 |         caffe_copy(1, top_diff+n, &scale);
65 |     //const Dtype* diff_feat_data2 = diff_feat.cpu_data();
66 |         caffe_axpy(channels_, scale*Dtype(2.0) / max_d, diff_feat_data, bottom_diff1+n*channels_);
67 |         caffe_axpy(channels_, scale*Dtype(-2.0) / max_d, diff_feat_data, bottom_diff2+n*channels_);
68 |     }
69 |   }
70 | 
71 | 
72 |   #ifdef CPU_ONLY
73 |     STUB_GPU(BatchEuclideanVectorLayer);
74 |   #endif
75 |   INSTANTIATE_CLASS(BatchEuclideanVectorLayer);
76 |   REGISTER_LAYER_CLASS(BatchEuclideanVector);
77 | } // namespace caffe
78 | 


--------------------------------------------------------------------------------
/batch_euclidean_vector_layer.cu:
--------------------------------------------------------------------------------
 1 | #include <algorithm>
 2 | #include <vector>
 3 | 
 4 | #include "caffe/layers/batch_euclidean_vector_layer.hpp"
 5 | 
 6 | namespace caffe {
 7 |   template <typename Dtype>
 8 |   void BatchEuclideanVectorLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top){
 9 |   	const Dtype* bottom_data1 = bottom[0]->gpu_data();
10 |     const Dtype* bottom_data2 = bottom[1]->gpu_data();
11 |     Dtype* top_data = top[0]->mutable_gpu_data();
12 |     Dtype* diff_feat_data = diff_feat.mutable_gpu_data();
13 | 
14 |     max_d = Dtype(0.0);
15 |   for (int n = 0; n < num_; ++n){
16 |         //diff_feat = x_n - x_nn
17 |         caffe_gpu_sub(channels_, bottom_data1+n*channels_, bottom_data2+n*channels_, diff_feat_data);
18 |         //sim = diff_feat * diff_feat
19 |         Dtype distance = Dtype(0.0);
20 |         caffe_gpu_dot(channels_, diff_feat_data, diff_feat_data, &distance);
21 |         if (distance > max_d){
22 |           max_d = distance;
23 |         }
24 |         //top[n, nn] = sim
25 |         caffe_gpu_set(1, distance, top_data+n);
26 |     }
27 |     caffe_gpu_scal(num_, Dtype(1.0) / max_d, top_data);  
28 |   }
29 | 
30 |   template <typename Dtype>
31 |   void BatchEuclideanVectorLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom){
32 |   	if (!propagate_down[0]){return;}
33 |   	const Dtype* top_diff = top[0]->gpu_diff();
34 |   	Dtype* bottom_diff1 = bottom[0]->mutable_gpu_diff();
35 |     Dtype* bottom_diff2 = bottom[1]->mutable_gpu_diff();
36 |   	caffe_gpu_set(num_*channels_, Dtype(0.0), bottom_diff1);
37 |     caffe_gpu_set(num_*channels_, Dtype(0.0), bottom_diff2);
38 |   	Dtype* diff_feat_data = diff_feat.mutable_gpu_data();
39 |   	const Dtype* bottom_data1 = bottom[0]->gpu_data();
40 |     const Dtype* bottom_data2 = bottom[1]->gpu_data();
41 |   	Dtype scale = Dtype(0.0);
42 |     for (int n = 0; n < num_; ++n){
43 |         caffe_gpu_sub(channels_, bottom_data1+n*channels_, bottom_data2+n*channels_, diff_feat_data);
44 |         caffe_copy(1, top_diff+n, &scale);
45 |     //const Dtype* diff_feat_data2 = diff_feat.gpu_data();
46 |         caffe_gpu_axpy(channels_, scale*Dtype(2.0) / max_d, diff_feat_data, bottom_diff1+n*channels_);
47 |         caffe_gpu_axpy(channels_, scale*Dtype(-2.0) / max_d, diff_feat_data, bottom_diff2+n*channels_);
48 |       }
49 |   }
50 |   INSTANTIATE_LAYER_GPU_FUNCS(BatchEuclideanVectorLayer);
51 | }
52 | 


--------------------------------------------------------------------------------
/batch_euclidean_vector_layer.hpp:
--------------------------------------------------------------------------------
 1 | #ifndef CAFFE_BATCH_EUCLIDEAN_VECTOR_LAYER_HPP_
 2 | #define CAFFE_BATCH_EUCLIDEAN_VECTOR_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 | namespace caffe {
11 |     template <typename Dtype>
12 |     class BatchEuclideanVectorLayer : public Layer<Dtype> {
13 |     public:
14 |         explicit BatchEuclideanVectorLayer(const LayerParameter& param) : Layer<Dtype>(param) {}
15 |         virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
16 | 		virtual void Reshape(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
17 |         virtual inline const char* type() const { return "BatchEuclideanVector"; }
18 |         virtual inline int ExactNumBottomBlobs() const { return 2; }
19 |         virtual inline int MinTopBlobs() const { return 1; }
20 |         virtual inline int MaxTopBlobs() const { return 1; }
21 |     protected:
22 |         virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
23 |         virtual void Backward_cpu(const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
24 |         virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
25 |         virtual void Backward_gpu(const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
26 |         int channels_;
27 |         int num_;
28 |         Blob<Dtype> diff_feat;
29 | 		Dtype max_d;
30 | 	};
31 | }  // namespace caffe
32 | 
33 | #endif //CAFFE_BatchEuclideanVector_LAYER_HPP_
34 | 


--------------------------------------------------------------------------------
/img/Exp_results.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yufanLIU/IRG/0a6837c923cf3a814a4c15022c11a658905297a5/img/Exp_results.png


--------------------------------------------------------------------------------
/img/framework.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yufanLIU/IRG/0a6837c923cf3a814a4c15022c11a658905297a5/img/framework.png


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