├── LICENSE
├── README.md
├── mobilenet_v3_large_1.0.prototxt
└── mobilenet_v3_small_1.0.prototxt


/LICENSE:
--------------------------------------------------------------------------------
 1 | MIT License
 2 | 
 3 | Copyright (c) 2019 jixing0415
 4 | 
 5 | Permission is hereby granted, free of charge, to any person obtaining a copy
 6 | of this software and associated documentation files (the "Software"), to deal
 7 | in the Software without restriction, including without limitation the rights
 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 9 | copies of the Software, and to permit persons to whom the Software is
10 | furnished to do so, subject to the following conditions:
11 | 
12 | The above copyright notice and this permission notice shall be included in all
13 | copies or substantial portions of the Software.
14 | 
15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 | SOFTWARE.
22 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # caffe-mobilenet-v3
 2 | ### Introduction
 3 | This is a personal Caffe implementation of MobileNetV3. For details, please read the original papers: [Searching for MobileNetV3](https://arxiv.org/pdf/1905.02244.pdf).
 4 | 
 5 | ### How to use
 6 | 1. Requirements for `Caffe` (see: [Caffe installation instructions](http://caffe.berkeleyvision.org/installation.html))
 7 | 2. Add new caffe layers and rebuild the caffe:
 8 | 	- **Depthwise Convolutional Layer** from [yonghenglh6/DepthwiseConvolution](https://github.com/yonghenglh6/DepthwiseConvolution) 
 9 | 	- **ReLU6 Layer** from [RuiminChen/Caffe-MobileNetV2-ReLU6](https://github.com/RuiminChen/Caffe-MobileNetV2-ReLU6) 
10 | 3. Run test
11 | 	```Shell
12 | 	CPU:
13 | 	$CAFFE_ROOT/build/tools/caffe time -model mobilenet_v3_large_1.0.prototxt
14 | 	GPU:
15 | 	$CAFFE_ROOT/build/tools/caffe time -model mobilenet_v3_large_1.0.prototxt -gpu 0
16 | 	```
17 | 
18 | ### Performance on MobileNetV3
19 | | Backbone            | CPU-Forward | CPU-Backward | GPU-Forward | GPU-Backward |
20 | | ------------------- |:---------:  | :---------:  | :---------: | :---------:  |
21 | | V3-Large 1.0        |  134.55 ms  |  140.23 ms   |  15.44 ms   |   21.79 ms   |
22 | | V3-Small 1.0        |  58.64  ms  |  59.30  ms   |  11.49 ms   |   12.58 ms   |
23 | 
24 | 
25 | ### TODO
26 |  - More MobileNetV3 architectures.
27 |  - Traning & validation.
28 | 
29 | 
30 | 
31 | 
32 | 


--------------------------------------------------------------------------------
/mobilenet_v3_large_1.0.prototxt:
--------------------------------------------------------------------------------
   1 | name: "mobilenet v3"
   2 | input: "data"
   3 | input_dim: 1
   4 | input_dim: 3
   5 | input_dim: 224
   6 | input_dim: 224
   7 | #################### stage 1 ####################
   8 | layer {
   9 |   name: "conv1"
  10 |   type: "Convolution"
  11 |   bottom: "data"
  12 |   top: "conv1"
  13 |   param {
  14 |     lr_mult: 1
  15 |     decay_mult: 1
  16 |   }
  17 |   convolution_param {
  18 |     num_output: 16
  19 |     pad: 1
  20 |     bias_term: false
  21 |     kernel_size: 3
  22 |     stride: 2
  23 |     weight_filler {
  24 |       type: "msra"
  25 |     }
  26 |   }
  27 | }
  28 | layer {
  29 |   name: "conv1-bn"
  30 |   type: "BatchNorm"
  31 |   bottom: "conv1"
  32 |   top: "conv1"
  33 | }
  34 | layer {
  35 |   name: "conv1-bn-scale"
  36 |   type: "Scale"
  37 |   bottom: "conv1"
  38 |   top: "conv1"
  39 |   scale_param {
  40 |     bias_term: true
  41 |   }
  42 | }
  43 | # H-swish
  44 | layer {
  45 |   name: "conv1/shift"
  46 |   type: "Power"
  47 |   bottom: "conv1"
  48 |   top: "conv1/shift"
  49 |   power_param {
  50 |     power: 1
  51 |     scale: 1
  52 |     shift: 3
  53 |   }
  54 | }
  55 | layer {
  56 |   name: "conv1/shift-relu"
  57 |   type: "ReLU6"
  58 |   bottom: "conv1/shift"
  59 |   top: "conv1/shift"
  60 | }
  61 | layer {
  62 |   name: "conv1/shift-div"
  63 |   type: "Power"
  64 |   bottom: "conv1/shift"
  65 |   top: "conv1/shift-div"
  66 |   power_param {
  67 |     power: 1
  68 |     scale: 0.1666666667
  69 |     shift: 0
  70 |   }
  71 | }
  72 | layer {
  73 |   name: "conv1/hswish"
  74 |   type: "Eltwise"
  75 |   bottom: "conv1"
  76 |   bottom: "conv1/shift-div"
  77 |   top: "conv1/hswish"
  78 |   eltwise_param {
  79 |     operation: PROD
  80 |   }
  81 | }
  82 | 
  83 | #################### stage 2-1 ####################
  84 | layer {
  85 |   name: "conv2-1/expand"
  86 |   type: "Convolution"
  87 |   bottom: "conv1/hswish"
  88 |   top: "conv2-1/expand"
  89 |   param {
  90 |     lr_mult: 1
  91 |     decay_mult: 1
  92 |   }
  93 |   convolution_param {
  94 |     num_output: 16
  95 |     pad: 0
  96 |     bias_term: false
  97 |     kernel_size: 1
  98 |     stride: 1
  99 |     weight_filler {
 100 |       type: "msra"
 101 |     }
 102 |   }
 103 | }
 104 | layer {
 105 |   name: "conv2-1/expand-bn"
 106 |   type: "BatchNorm"
 107 |   bottom: "conv2-1/expand"
 108 |   top: "conv2-1/expand"
 109 | }
 110 | layer {
 111 |   name: "conv2-1/expand-bn-scale"
 112 |   type: "Scale"
 113 |   bottom: "conv2-1/expand"
 114 |   top: "conv2-1/expand"
 115 |   scale_param {
 116 |     bias_term: true
 117 |   }
 118 | }
 119 | layer {
 120 |   name: "conv2-1/expand-relu"
 121 |   type: "ReLU6"
 122 |   bottom: "conv2-1/expand"
 123 |   top: "conv2-1/expand"
 124 | }
 125 | 
 126 | layer {
 127 |   name: "conv2-1/dwise"
 128 |   #type: "Convolution"
 129 |   type: "DepthwiseConvolution"
 130 |   bottom: "conv2-1/expand"
 131 |   top: "conv2-1/dwise"
 132 |   param {
 133 |     lr_mult: 1
 134 |     decay_mult: 1
 135 |   }
 136 |   convolution_param {
 137 |     num_output: 16
 138 |     group: 16
 139 |     pad: 1
 140 |     bias_term: false
 141 |     kernel_size: 3
 142 |     stride: 1
 143 |     weight_filler {
 144 |       type: "msra"
 145 |     }
 146 |     engine: CAFFE
 147 |   }
 148 | }
 149 | layer {
 150 |   name: "conv2-1/dwise-bn"
 151 |   type: "BatchNorm"
 152 |   bottom: "conv2-1/dwise"
 153 |   top: "conv2-1/dwise"
 154 | }
 155 | layer {
 156 |   name: "conv2-1/dwise-bn-scale"
 157 |   type: "Scale"
 158 |   bottom: "conv2-1/dwise"
 159 |   top: "conv2-1/dwise"
 160 |   scale_param {
 161 |     bias_term: true
 162 |   }
 163 | }
 164 | layer {
 165 |   name: "conv2-1/dwise-relu"
 166 |   type: "ReLU6"
 167 |   bottom: "conv2-1/dwise"
 168 |   top: "conv2-1/dwise"
 169 | }
 170 | 
 171 | layer {
 172 |   name: "conv2-1/linear"
 173 |   type: "Convolution"
 174 |   bottom: "conv2-1/dwise"
 175 |   top: "conv2-1/linear"
 176 |   param {
 177 |     lr_mult: 1
 178 |     decay_mult: 1
 179 |   }
 180 |   convolution_param {
 181 |     num_output: 16
 182 |     pad: 0
 183 |     bias_term: false
 184 |     kernel_size: 1
 185 |     stride: 1
 186 |     weight_filler {
 187 |       type: "msra"
 188 |     }
 189 |   }
 190 | }
 191 | layer {
 192 |   name: "conv2-1/linear-bn"
 193 |   type: "BatchNorm"
 194 |   bottom: "conv2-1/linear"
 195 |   top: "conv2-1/linear"
 196 | }
 197 | layer {
 198 |   name: "conv2-1/linear-bn-scale"
 199 |   type: "Scale"
 200 |   bottom: "conv2-1/linear"
 201 |   top: "conv2-1/linear"
 202 |   scale_param {
 203 |     bias_term: true
 204 |   }
 205 | }
 206 | 
 207 | layer {
 208 |   name: "block2-1"
 209 |   type: "Eltwise"
 210 |   bottom: "conv1/hswish"
 211 |   bottom: "conv2-1/linear"
 212 |   top: "block2-1"
 213 | }
 214 | 
 215 | #################### stage 2-2 ####################
 216 | layer {
 217 |   name: "conv2-2/expand"
 218 |   type: "Convolution"
 219 |   bottom: "block2-1"
 220 |   top: "conv2-2/expand"
 221 |   param {
 222 |     lr_mult: 1
 223 |     decay_mult: 1
 224 |   }
 225 |   convolution_param {
 226 |     num_output: 64
 227 |     pad: 0
 228 |     bias_term: false
 229 |     kernel_size: 1
 230 |     stride: 1
 231 |     weight_filler {
 232 |       type: "msra"
 233 |     }
 234 |   }
 235 | }
 236 | layer {
 237 |   name: "conv2-2/expand-bn"
 238 |   type: "BatchNorm"
 239 |   bottom: "conv2-2/expand"
 240 |   top: "conv2-2/expand"
 241 | }
 242 | layer {
 243 |   name: "conv2-2/expand-bn-scale"
 244 |   type: "Scale"
 245 |   bottom: "conv2-2/expand"
 246 |   top: "conv2-2/expand"
 247 |   scale_param {
 248 |     bias_term: true
 249 |   }
 250 | }
 251 | layer {
 252 |   name: "conv2-2/expand-relu"
 253 |   type: "ReLU6"
 254 |   bottom: "conv2-2/expand"
 255 |   top: "conv2-2/expand"
 256 | }
 257 | 
 258 | layer {
 259 |   name: "conv2-2/dwise"
 260 |   #type: "Convolution"
 261 |   type: "DepthwiseConvolution"
 262 |   bottom: "conv2-2/expand"
 263 |   top: "conv2-2/dwise"
 264 |   param {
 265 |     lr_mult: 1
 266 |     decay_mult: 1
 267 |   }
 268 |   convolution_param {
 269 |     num_output: 64
 270 |     group: 64
 271 |     pad: 1
 272 |     bias_term: false
 273 |     kernel_size: 3
 274 |     stride: 2
 275 |     weight_filler {
 276 |       type: "msra"
 277 |     }
 278 |     engine: CAFFE
 279 |   }
 280 | }
 281 | layer {
 282 |   name: "conv2-2/dwise-bn"
 283 |   type: "BatchNorm"
 284 |   bottom: "conv2-2/dwise"
 285 |   top: "conv2-2/dwise"
 286 | }
 287 | layer {
 288 |   name: "conv2-2/dwise-bn-scale"
 289 |   type: "Scale"
 290 |   bottom: "conv2-2/dwise"
 291 |   top: "conv2-2/dwise"
 292 |   scale_param {
 293 |     bias_term: true
 294 |   }
 295 | }
 296 | layer {
 297 |   name: "conv2-2/dwise-relu"
 298 |   type: "ReLU6"
 299 |   bottom: "conv2-2/dwise"
 300 |   top: "conv2-2/dwise"
 301 | }
 302 | 
 303 | layer {
 304 |   name: "conv2-2/linear"
 305 |   type: "Convolution"
 306 |   bottom: "conv2-2/dwise"
 307 |   top: "conv2-2/linear"
 308 |   param {
 309 |     lr_mult: 1
 310 |     decay_mult: 1
 311 |   }
 312 |   convolution_param {
 313 |     num_output: 24
 314 |     pad: 0
 315 |     bias_term: false
 316 |     kernel_size: 1
 317 |     stride: 1
 318 |     weight_filler {
 319 |       type: "msra"
 320 |     }
 321 |   }
 322 | }
 323 | layer {
 324 |   name: "conv2-2/linear-bn"
 325 |   type: "BatchNorm"
 326 |   bottom: "conv2-2/linear"
 327 |   top: "conv2-2/linear"
 328 | }
 329 | layer {
 330 |   name: "conv2-2/linear-bn-scale"
 331 |   type: "Scale"
 332 |   bottom: "conv2-2/linear"
 333 |   top: "conv2-2/linear"
 334 |   scale_param {
 335 |     bias_term: true
 336 |   }
 337 | }
 338 | 
 339 | #################### stage 3-1 ####################
 340 | layer {
 341 |   name: "conv3-1/expand"
 342 |   type: "Convolution"
 343 |   bottom: "conv2-2/linear"
 344 |   top: "conv3-1/expand"
 345 |   param {
 346 |     lr_mult: 1
 347 |     decay_mult: 1
 348 |   }
 349 |   convolution_param {
 350 |     num_output: 72
 351 |     pad: 0
 352 |     bias_term: false
 353 |     kernel_size: 1
 354 |     stride: 1
 355 |     weight_filler {
 356 |       type: "msra"
 357 |     }
 358 |   }
 359 | }
 360 | layer {
 361 |   name: "conv3-1/expand-bn"
 362 |   type: "BatchNorm"
 363 |   bottom: "conv3-1/expand"
 364 |   top: "conv3-1/expand"
 365 | }
 366 | layer {
 367 |   name: "conv3-1/expand-bn-scale"
 368 |   type: "Scale"
 369 |   bottom: "conv3-1/expand"
 370 |   top: "conv3-1/expand"
 371 |   scale_param {
 372 |     bias_term: true
 373 |   }
 374 | }
 375 | layer {
 376 |   name: "conv3-1/expand-relu"
 377 |   type: "ReLU6"
 378 |   bottom: "conv3-1/expand"
 379 |   top: "conv3-1/expand"
 380 | }
 381 | 
 382 | layer {
 383 |   name: "conv3-1/dwise"
 384 |   #type: "Convolution"
 385 |   type: "DepthwiseConvolution"
 386 |   bottom: "conv3-1/expand"
 387 |   top: "conv3-1/dwise"
 388 |   param {
 389 |     lr_mult: 1
 390 |     decay_mult: 1
 391 |   }
 392 |   convolution_param {
 393 |     num_output: 72
 394 |     group: 72
 395 |     pad: 1
 396 |     bias_term: false
 397 |     kernel_size: 3
 398 |     stride: 1
 399 |     weight_filler {
 400 |       type: "msra"
 401 |     }
 402 |     engine: CAFFE
 403 |   }
 404 | }
 405 | layer {
 406 |   name: "conv3-1/dwise-bn"
 407 |   type: "BatchNorm"
 408 |   bottom: "conv3-1/dwise"
 409 |   top: "conv3-1/dwise"
 410 | }
 411 | layer {
 412 |   name: "conv3-1/dwise-bn-scale"
 413 |   type: "Scale"
 414 |   bottom: "conv3-1/dwise"
 415 |   top: "conv3-1/dwise"
 416 |   scale_param {
 417 |     bias_term: true
 418 |   }
 419 | }
 420 | layer {
 421 |   name: "conv3-1/dwise-relu"
 422 |   type: "ReLU6"
 423 |   bottom: "conv3-1/dwise"
 424 |   top: "conv3-1/dwise"
 425 | }
 426 | 
 427 | layer {
 428 |   name: "conv3-1/linear"
 429 |   type: "Convolution"
 430 |   bottom: "conv3-1/dwise"
 431 |   top: "conv3-1/linear"
 432 |   param {
 433 |     lr_mult: 1
 434 |     decay_mult: 1
 435 |   }
 436 |   convolution_param {
 437 |     num_output: 24
 438 |     pad: 0
 439 |     bias_term: false
 440 |     kernel_size: 1
 441 |     stride: 1
 442 |     weight_filler {
 443 |       type: "msra"
 444 |     }
 445 |   }
 446 | }
 447 | layer {
 448 |   name: "conv3-1/linear-bn"
 449 |   type: "BatchNorm"
 450 |   bottom: "conv3-1/linear"
 451 |   top: "conv3-1/linear"
 452 | }
 453 | layer {
 454 |   name: "conv3-1/linear-bn-scale"
 455 |   type: "Scale"
 456 |   bottom: "conv3-1/linear"
 457 |   top: "conv3-1/linear"
 458 |   scale_param {
 459 |     bias_term: true
 460 |   }
 461 | }
 462 | 
 463 | layer {
 464 |   name: "block3-1"
 465 |   type: "Eltwise"
 466 |   bottom: "conv2-2/linear"
 467 |   bottom: "conv3-1/linear"
 468 |   top: "block3-1"
 469 | }
 470 | 
 471 | #################### stage 3-2 ####################
 472 | layer {
 473 |   name: "conv3-2/expand"
 474 |   type: "Convolution"
 475 |   bottom: "block3-1"
 476 |   top: "conv3-2/expand"
 477 |   param {
 478 |     lr_mult: 1
 479 |     decay_mult: 1
 480 |   }
 481 |   convolution_param {
 482 |     num_output: 72
 483 |     pad: 0
 484 |     bias_term: false
 485 |     kernel_size: 1
 486 |     stride: 1
 487 |     weight_filler {
 488 |       type: "msra"
 489 |     }
 490 |   }
 491 | }
 492 | layer {
 493 |   name: "conv3-2/expand-bn"
 494 |   type: "BatchNorm"
 495 |   bottom: "conv3-2/expand"
 496 |   top: "conv3-2/expand"
 497 | }
 498 | layer {
 499 |   name: "conv3-2/expand-bn-scale"
 500 |   type: "Scale"
 501 |   bottom: "conv3-2/expand"
 502 |   top: "conv3-2/expand"
 503 |   scale_param {
 504 |     bias_term: true
 505 |   }
 506 | }
 507 | layer {
 508 |   name: "conv3-2/expand-relu"
 509 |   type: "ReLU6"
 510 |   bottom: "conv3-2/expand"
 511 |   top: "conv3-2/expand"
 512 | }
 513 | 
 514 | layer {
 515 |   name: "conv3-2/dwise"
 516 |   #type: "Convolution"
 517 |   type: "DepthwiseConvolution"
 518 |   bottom: "conv3-2/expand"
 519 |   top: "conv3-2/dwise"
 520 |   param {
 521 |     lr_mult: 1
 522 |     decay_mult: 1
 523 |   }
 524 |   convolution_param {
 525 |     num_output: 72
 526 |     group: 72
 527 |     pad: 2
 528 |     bias_term: false
 529 |     kernel_size: 5
 530 |     stride: 2
 531 |     weight_filler {
 532 |       type: "msra"
 533 |     }
 534 |     engine: CAFFE
 535 |   }
 536 | }
 537 | layer {
 538 |   name: "conv3-2/dwise-bn"
 539 |   type: "BatchNorm"
 540 |   bottom: "conv3-2/dwise"
 541 |   top: "conv3-2/dwise"
 542 | }
 543 | layer {
 544 |   name: "conv3-2/dwise-bn-scale"
 545 |   type: "Scale"
 546 |   bottom: "conv3-2/dwise"
 547 |   top: "conv3-2/dwise"
 548 |   scale_param {
 549 |     bias_term: true
 550 |   }
 551 | }
 552 | 
 553 | # se
 554 | layer {
 555 |   name: "conv3-2/dwise-se-pool"
 556 |   type: "Pooling"
 557 |   bottom: "conv3-2/dwise"
 558 |   top: "conv3-2/dwise-se-pool"
 559 |   pooling_param {
 560 |     pool: AVE
 561 |     global_pooling: true
 562 |   }
 563 | }
 564 | layer {
 565 |   name: "conv3-2/dwise-se-fc1"
 566 |   type: "InnerProduct"
 567 |   bottom: "conv3-2/dwise-se-pool"
 568 |   top:"conv3-2/dwise-se-fc1"
 569 |   param {
 570 |     lr_mult: 1
 571 |     decay_mult: 1
 572 |   }
 573 |   param {
 574 |     lr_mult: 2
 575 |     decay_mult: 0
 576 |   }
 577 |   inner_product_param {
 578 |     num_output: 18
 579 |     weight_filler {
 580 |       type: "msra"
 581 |     }
 582 |     bias_filler {
 583 |       type: "constant"
 584 |       value: 0
 585 |     }
 586 |   }
 587 | }
 588 | layer {
 589 |   name: "conv3-2/dwise-se-fc1-relu"
 590 |   type: "ReLU6"
 591 |   bottom: "conv3-2/dwise-se-fc1"
 592 |   top: "conv3-2/dwise-se-fc1"
 593 | }
 594 | layer {
 595 |   name: "conv3-2/dwise-se-fc2"
 596 |   type: "InnerProduct"
 597 |   bottom: "conv3-2/dwise-se-fc1"
 598 |   top:"conv3-2/dwise-se-fc2"
 599 |   param {
 600 |     lr_mult: 1
 601 |     decay_mult: 1
 602 |   }
 603 |   param {
 604 |     lr_mult: 2
 605 |     decay_mult: 0
 606 |   }
 607 |   inner_product_param {
 608 |     num_output: 72
 609 |     weight_filler {
 610 |       type: "msra"
 611 |     }
 612 |     bias_filler {
 613 |       type: "constant"
 614 |       value: 0
 615 |     }
 616 |   }
 617 | }
 618 | # H-sigmoid
 619 | layer {
 620 |   name: "conv3-2/dwise-se-fc2/shift"
 621 |   type: "Power"
 622 |   bottom: "conv3-2/dwise-se-fc2"
 623 |   top: "conv3-2/dwise-se-fc2/shift"
 624 |   power_param {
 625 |     power: 1
 626 |     scale: 1
 627 |     shift: 3
 628 |   }
 629 | }
 630 | layer {
 631 |   name: "conv3-2/dwise-se-fc2/shift-relu"
 632 |   type: "ReLU6"
 633 |   bottom: "conv3-2/dwise-se-fc2/shift"
 634 |   top: "conv3-2/dwise-se-fc2/shift"
 635 | }
 636 | layer {
 637 |   name: "conv3-2/dwise-se-fc2/shift-div"
 638 |   type: "Power"
 639 |   bottom: "conv3-2/dwise-se-fc2/shift"
 640 |   top: "conv3-2/dwise-se-fc2/shift-div"
 641 |   power_param {
 642 |     power: 1
 643 |     scale: 0.1666666667
 644 |     shift: 0
 645 |   }
 646 | }
 647 | layer {
 648 |   name: "conv3-2/dwise/scale"
 649 |   type: "Scale"
 650 |   bottom: "conv3-2/dwise"
 651 |   bottom: "conv3-2/dwise-se-fc2/shift-div"
 652 |   top: "conv3-2/dwise/scale"
 653 |   scale_param{
 654 |     axis: 0
 655 |   }
 656 | }
 657 | layer {
 658 |   name: "conv3-2/dwise/scale-relu"
 659 |   type: "ReLU6"
 660 |   bottom: "conv3-2/dwise/scale"
 661 |   top: "conv3-2/dwise/scale"
 662 | }
 663 | 
 664 | layer {
 665 |   name: "conv3-2/linear"
 666 |   type: "Convolution"
 667 |   bottom: "conv3-2/dwise/scale"
 668 |   top: "conv3-2/linear"
 669 |   param {
 670 |     lr_mult: 1
 671 |     decay_mult: 1
 672 |   }
 673 |   convolution_param {
 674 |     num_output: 40
 675 |     pad: 0
 676 |     bias_term: false
 677 |     kernel_size: 1
 678 |     stride: 1
 679 |     weight_filler {
 680 |       type: "msra"
 681 |     }
 682 |   }
 683 | }
 684 | layer {
 685 |   name: "conv3-2/linear-bn"
 686 |   type: "BatchNorm"
 687 |   bottom: "conv3-2/linear"
 688 |   top: "conv3-2/linear"
 689 | }
 690 | layer {
 691 |   name: "conv3-2/linear-bn-scale"
 692 |   type: "Scale"
 693 |   bottom: "conv3-2/linear"
 694 |   top: "conv3-2/linear"
 695 |   scale_param {
 696 |     bias_term: true
 697 |   }
 698 | }
 699 | 
 700 | #################### stage 4-1 ####################
 701 | layer {
 702 |   name: "conv4-1/expand"
 703 |   type: "Convolution"
 704 |   bottom: "conv3-2/linear"
 705 |   top: "conv4-1/expand"
 706 |   param {
 707 |     lr_mult: 1
 708 |     decay_mult: 1
 709 |   }
 710 |   convolution_param {
 711 |     num_output: 120
 712 |     pad: 0
 713 |     bias_term: false
 714 |     kernel_size: 1
 715 |     stride: 1
 716 |     weight_filler {
 717 |       type: "msra"
 718 |     }
 719 |   }
 720 | }
 721 | layer {
 722 |   name: "conv4-1/expand-bn"
 723 |   type: "BatchNorm"
 724 |   bottom: "conv4-1/expand"
 725 |   top: "conv4-1/expand"
 726 | }
 727 | layer {
 728 |   name: "conv4-1/expand-bn-scale"
 729 |   type: "Scale"
 730 |   bottom: "conv4-1/expand"
 731 |   top: "conv4-1/expand"
 732 |   scale_param {
 733 |     bias_term: true
 734 |   }
 735 | }
 736 | layer {
 737 |   name: "conv4-1/expand-relu"
 738 |   type: "ReLU6"
 739 |   bottom: "conv4-1/expand"
 740 |   top: "conv4-1/expand"
 741 | }
 742 | 
 743 | layer {
 744 |   name: "conv4-1/dwise"
 745 |   #type: "Convolution"
 746 |   type: "DepthwiseConvolution"
 747 |   bottom: "conv4-1/expand"
 748 |   top: "conv4-1/dwise"
 749 |   param {
 750 |     lr_mult: 1
 751 |     decay_mult: 1
 752 |   }
 753 |   convolution_param {
 754 |     num_output: 120
 755 |     group: 120
 756 |     pad: 2
 757 |     bias_term: false
 758 |     kernel_size: 5
 759 |     stride: 1
 760 |     weight_filler {
 761 |       type: "msra"
 762 |     }
 763 |     engine: CAFFE
 764 |   }
 765 | }
 766 | layer {
 767 |   name: "conv4-1/dwise-bn"
 768 |   type: "BatchNorm"
 769 |   bottom: "conv4-1/dwise"
 770 |   top: "conv4-1/dwise"
 771 | }
 772 | layer {
 773 |   name: "conv4-1/dwise-bn-scale"
 774 |   type: "Scale"
 775 |   bottom: "conv4-1/dwise"
 776 |   top: "conv4-1/dwise"
 777 |   scale_param {
 778 |     bias_term: true
 779 |   }
 780 | }
 781 | 
 782 | # se
 783 | layer {
 784 |   name: "conv4-1/dwise-se-pool"
 785 |   type: "Pooling"
 786 |   bottom: "conv4-1/dwise"
 787 |   top: "conv4-1/dwise-se-pool"
 788 |   pooling_param {
 789 |     pool: AVE
 790 |     global_pooling: true
 791 |   }
 792 | }
 793 | layer {
 794 |   name: "conv4-1/dwise-se-fc1"
 795 |   type: "InnerProduct"
 796 |   bottom: "conv4-1/dwise-se-pool"
 797 |   top:"conv4-1/dwise-se-fc1"
 798 |   param {
 799 |     lr_mult: 1
 800 |     decay_mult: 1
 801 |   }
 802 |   param {
 803 |     lr_mult: 2
 804 |     decay_mult: 0
 805 |   }
 806 |   inner_product_param {
 807 |     num_output: 30
 808 |     weight_filler {
 809 |       type: "msra"
 810 |     }
 811 |     bias_filler {
 812 |       type: "constant"
 813 |       value: 0
 814 |     }
 815 |   }
 816 | }
 817 | layer {
 818 |   name: "conv4-1/dwise-se-fc1-relu"
 819 |   type: "ReLU6"
 820 |   bottom: "conv4-1/dwise-se-fc1"
 821 |   top: "conv4-1/dwise-se-fc1"
 822 | }
 823 | layer {
 824 |   name: "conv4-1/dwise-se-fc2"
 825 |   type: "InnerProduct"
 826 |   bottom: "conv4-1/dwise-se-fc1"
 827 |   top:"conv4-1/dwise-se-fc2"
 828 |   param {
 829 |     lr_mult: 1
 830 |     decay_mult: 1
 831 |   }
 832 |   param {
 833 |     lr_mult: 2
 834 |     decay_mult: 0
 835 |   }
 836 |   inner_product_param {
 837 |     num_output: 120
 838 |     weight_filler {
 839 |       type: "msra"
 840 |     }
 841 |     bias_filler {
 842 |       type: "constant"
 843 |       value: 0
 844 |     }
 845 |   }
 846 | }
 847 | # H-sigmoid
 848 | layer {
 849 |   name: "conv4-1/dwise-se-fc2/shift"
 850 |   type: "Power"
 851 |   bottom: "conv4-1/dwise-se-fc2"
 852 |   top: "conv4-1/dwise-se-fc2/shift"
 853 |   power_param {
 854 |     power: 1
 855 |     scale: 1
 856 |     shift: 3
 857 |   }
 858 | }
 859 | layer {
 860 |   name: "conv4-1/dwise-se-fc2/shift-relu"
 861 |   type: "ReLU6"
 862 |   bottom: "conv4-1/dwise-se-fc2/shift"
 863 |   top: "conv4-1/dwise-se-fc2/shift"
 864 | }
 865 | layer {
 866 |   name: "conv4-1/dwise-se-fc2/shift-div"
 867 |   type: "Power"
 868 |   bottom: "conv4-1/dwise-se-fc2/shift"
 869 |   top: "conv4-1/dwise-se-fc2/shift-div"
 870 |   power_param {
 871 |     power: 1
 872 |     scale: 0.1666666667
 873 |     shift: 0
 874 |   }
 875 | }
 876 | layer {
 877 |   name: "conv4-1/dwise/scale"
 878 |   type: "Scale"
 879 |   bottom: "conv4-1/dwise"
 880 |   bottom: "conv4-1/dwise-se-fc2/shift-div"
 881 |   top: "conv4-1/dwise/scale"
 882 |   scale_param{
 883 |     axis: 0
 884 |   }
 885 | }
 886 | layer {
 887 |   name: "conv4-1/dwise/scale-relu"
 888 |   type: "ReLU6"
 889 |   bottom: "conv4-1/dwise/scale"
 890 |   top: "conv4-1/dwise/scale"
 891 | }
 892 | 
 893 | layer {
 894 |   name: "conv4-1/linear"
 895 |   type: "Convolution"
 896 |   bottom: "conv4-1/dwise/scale"
 897 |   top: "conv4-1/linear"
 898 |   param {
 899 |     lr_mult: 1
 900 |     decay_mult: 1
 901 |   }
 902 |   convolution_param {
 903 |     num_output: 40
 904 |     pad: 0
 905 |     bias_term: false
 906 |     kernel_size: 1
 907 |     stride: 1
 908 |     weight_filler {
 909 |       type: "msra"
 910 |     }
 911 |   }
 912 | }
 913 | layer {
 914 |   name: "conv4-1/linear-bn"
 915 |   type: "BatchNorm"
 916 |   bottom: "conv4-1/linear"
 917 |   top: "conv4-1/linear"
 918 | }
 919 | layer {
 920 |   name: "conv4-1/linear-bn-scale"
 921 |   type: "Scale"
 922 |   bottom: "conv4-1/linear"
 923 |   top: "conv4-1/linear"
 924 |   scale_param {
 925 |     bias_term: true
 926 |   }
 927 | }
 928 | 
 929 | layer {
 930 |   name: "block4-1"
 931 |   type: "Eltwise"
 932 |   bottom: "conv3-2/linear"
 933 |   bottom: "conv4-1/linear"
 934 |   top: "block4-1"
 935 | }
 936 | 
 937 | #################### stage 4-2 ####################
 938 | layer {
 939 |   name: "conv4-2/expand"
 940 |   type: "Convolution"
 941 |   bottom: "block4-1"
 942 |   top: "conv4-2/expand"
 943 |   param {
 944 |     lr_mult: 1
 945 |     decay_mult: 1
 946 |   }
 947 |   convolution_param {
 948 |     num_output: 120
 949 |     pad: 0
 950 |     bias_term: false
 951 |     kernel_size: 1
 952 |     stride: 1
 953 |     weight_filler {
 954 |       type: "msra"
 955 |     }
 956 |   }
 957 | }
 958 | layer {
 959 |   name: "conv4-2/expand-bn"
 960 |   type: "BatchNorm"
 961 |   bottom: "conv4-2/expand"
 962 |   top: "conv4-2/expand"
 963 | }
 964 | layer {
 965 |   name: "conv4-2/expand-bn-scale"
 966 |   type: "Scale"
 967 |   bottom: "conv4-2/expand"
 968 |   top: "conv4-2/expand"
 969 |   scale_param {
 970 |     bias_term: true
 971 |   }
 972 | }
 973 | layer {
 974 |   name: "conv4-2/expand-relu"
 975 |   type: "ReLU6"
 976 |   bottom: "conv4-2/expand"
 977 |   top: "conv4-2/expand"
 978 | }
 979 | 
 980 | layer {
 981 |   name: "conv4-2/dwise"
 982 |   #type: "Convolution"
 983 |   type: "DepthwiseConvolution"
 984 |   bottom: "conv4-2/expand"
 985 |   top: "conv4-2/dwise"
 986 |   param {
 987 |     lr_mult: 1
 988 |     decay_mult: 1
 989 |   }
 990 |   convolution_param {
 991 |     num_output: 120
 992 |     group: 120
 993 |     pad: 2
 994 |     bias_term: false
 995 |     kernel_size: 5
 996 |     stride: 1
 997 |     weight_filler {
 998 |       type: "msra"
 999 |     }
1000 |     engine: CAFFE
1001 |   }
1002 | }
1003 | layer {
1004 |   name: "conv4-2/dwise-bn"
1005 |   type: "BatchNorm"
1006 |   bottom: "conv4-2/dwise"
1007 |   top: "conv4-2/dwise"
1008 | }
1009 | layer {
1010 |   name: "conv4-2/dwise-bn-scale"
1011 |   type: "Scale"
1012 |   bottom: "conv4-2/dwise"
1013 |   top: "conv4-2/dwise"
1014 |   scale_param {
1015 |     bias_term: true
1016 |   }
1017 | }
1018 | 
1019 | # se
1020 | layer {
1021 |   name: "conv4-2/dwise-se-pool"
1022 |   type: "Pooling"
1023 |   bottom: "conv4-2/dwise"
1024 |   top: "conv4-2/dwise-se-pool"
1025 |   pooling_param {
1026 |     pool: AVE
1027 |     global_pooling: true
1028 |   }
1029 | }
1030 | layer {
1031 |   name: "conv4-2/dwise-se-fc1"
1032 |   type: "InnerProduct"
1033 |   bottom: "conv4-2/dwise-se-pool"
1034 |   top:"conv4-2/dwise-se-fc1"
1035 |   param {
1036 |     lr_mult: 1
1037 |     decay_mult: 1
1038 |   }
1039 |   param {
1040 |     lr_mult: 2
1041 |     decay_mult: 0
1042 |   }
1043 |   inner_product_param {
1044 |     num_output: 30
1045 |     weight_filler {
1046 |       type: "msra"
1047 |     }
1048 |     bias_filler {
1049 |       type: "constant"
1050 |       value: 0
1051 |     }
1052 |   }
1053 | }
1054 | layer {
1055 |   name: "conv4-2/dwise-se-fc1-relu"
1056 |   type: "ReLU6"
1057 |   bottom: "conv4-2/dwise-se-fc1"
1058 |   top: "conv4-2/dwise-se-fc1"
1059 | }
1060 | layer {
1061 |   name: "conv4-2/dwise-se-fc2"
1062 |   type: "InnerProduct"
1063 |   bottom: "conv4-2/dwise-se-fc1"
1064 |   top:"conv4-2/dwise-se-fc2"
1065 |   param {
1066 |     lr_mult: 1
1067 |     decay_mult: 1
1068 |   }
1069 |   param {
1070 |     lr_mult: 2
1071 |     decay_mult: 0
1072 |   }
1073 |   inner_product_param {
1074 |     num_output: 120
1075 |     weight_filler {
1076 |       type: "msra"
1077 |     }
1078 |     bias_filler {
1079 |       type: "constant"
1080 |       value: 0
1081 |     }
1082 |   }
1083 | }
1084 | # H-sigmoid
1085 | layer {
1086 |   name: "conv4-2/dwise-se-fc2/shift"
1087 |   type: "Power"
1088 |   bottom: "conv4-2/dwise-se-fc2"
1089 |   top: "conv4-2/dwise-se-fc2/shift"
1090 |   power_param {
1091 |     power: 1
1092 |     scale: 1
1093 |     shift: 3
1094 |   }
1095 | }
1096 | layer {
1097 |   name: "conv4-2/dwise-se-fc2/shift-relu"
1098 |   type: "ReLU6"
1099 |   bottom: "conv4-2/dwise-se-fc2/shift"
1100 |   top: "conv4-2/dwise-se-fc2/shift"
1101 | }
1102 | layer {
1103 |   name: "conv4-2/dwise-se-fc2/shift-div"
1104 |   type: "Power"
1105 |   bottom: "conv4-2/dwise-se-fc2/shift"
1106 |   top: "conv4-2/dwise-se-fc2/shift-div"
1107 |   power_param {
1108 |     power: 1
1109 |     scale: 0.1666666667
1110 |     shift: 0
1111 |   }
1112 | }
1113 | layer {
1114 |   name: "conv4-2/dwise/scale"
1115 |   type: "Scale"
1116 |   bottom: "conv4-2/dwise"
1117 |   bottom: "conv4-2/dwise-se-fc2/shift-div"
1118 |   top: "conv4-2/dwise/scale"
1119 |   scale_param{
1120 |     axis: 0
1121 |   }
1122 | }
1123 | layer {
1124 |   name: "conv4-2/dwise/scale-relu"
1125 |   type: "ReLU6"
1126 |   bottom: "conv4-2/dwise/scale"
1127 |   top: "conv4-2/dwise/scale"
1128 | }
1129 | 
1130 | layer {
1131 |   name: "conv4-2/linear"
1132 |   type: "Convolution"
1133 |   bottom: "conv4-2/dwise/scale"
1134 |   top: "conv4-2/linear"
1135 |   param {
1136 |     lr_mult: 1
1137 |     decay_mult: 1
1138 |   }
1139 |   convolution_param {
1140 |     num_output: 40
1141 |     pad: 0
1142 |     bias_term: false
1143 |     kernel_size: 1
1144 |     stride: 1
1145 |     weight_filler {
1146 |       type: "msra"
1147 |     }
1148 |   }
1149 | }
1150 | layer {
1151 |   name: "conv4-2/linear-bn"
1152 |   type: "BatchNorm"
1153 |   bottom: "conv4-2/linear"
1154 |   top: "conv4-2/linear"
1155 | }
1156 | layer {
1157 |   name: "conv4-2/linear-bn-scale"
1158 |   type: "Scale"
1159 |   bottom: "conv4-2/linear"
1160 |   top: "conv4-2/linear"
1161 |   scale_param {
1162 |     bias_term: true
1163 |   }
1164 | }
1165 | 
1166 | layer {
1167 |   name: "block4-2"
1168 |   type: "Eltwise"
1169 |   bottom: "block4-1"
1170 |   bottom: "conv4-2/linear"
1171 |   top: "block4-2"
1172 | }
1173 | 
1174 | #################### stage 4-3 ####################
1175 | layer {
1176 |   name: "conv4-3/expand"
1177 |   type: "Convolution"
1178 |   bottom: "block4-2"
1179 |   top: "conv4-3/expand"
1180 |   param {
1181 |     lr_mult: 1
1182 |     decay_mult: 1
1183 |   }
1184 |   convolution_param {
1185 |     num_output: 240
1186 |     pad: 0
1187 |     bias_term: false
1188 |     kernel_size: 1
1189 |     stride: 1
1190 |     weight_filler {
1191 |       type: "msra"
1192 |     }
1193 |   }
1194 | }
1195 | layer {
1196 |   name: "conv4-3/expand-bn"
1197 |   type: "BatchNorm"
1198 |   bottom: "conv4-3/expand"
1199 |   top: "conv4-3/expand"
1200 | }
1201 | layer {
1202 |   name: "conv4-3/expand-bn-scale"
1203 |   type: "Scale"
1204 |   bottom: "conv4-3/expand"
1205 |   top: "conv4-3/expand"
1206 |   scale_param {
1207 |     bias_term: true
1208 |   }
1209 | }
1210 | # H-swish
1211 | layer {
1212 |   name: "conv4-3/expand/shift"
1213 |   type: "Power"
1214 |   bottom: "conv4-3/expand"
1215 |   top: "conv4-3/expand/shift"
1216 |   power_param {
1217 |     power: 1
1218 |     scale: 1
1219 |     shift: 3
1220 |   }
1221 | }
1222 | layer {
1223 |   name: "conv4-3/expand/shift-relu"
1224 |   type: "ReLU6"
1225 |   bottom: "conv4-3/expand/shift"
1226 |   top: "conv4-3/expand/shift"
1227 | }
1228 | layer {
1229 |   name: "conv4-3/expand/shift-div"
1230 |   type: "Power"
1231 |   bottom: "conv4-3/expand/shift"
1232 |   top: "conv4-3/expand/shift-div"
1233 |   power_param {
1234 |     power: 1
1235 |     scale: 0.1666666667
1236 |     shift: 0
1237 |   }
1238 | }
1239 | layer {
1240 |   name: "conv4-3/expand/hswish"
1241 |   type: "Eltwise"
1242 |   bottom: "conv4-3/expand"
1243 |   bottom: "conv4-3/expand/shift-div"
1244 |   top: "conv4-3/expand/hswish"
1245 |   eltwise_param {
1246 |     operation: PROD
1247 |   }
1248 | }
1249 | 
1250 | layer {
1251 |   name: "conv4-3/dwise"
1252 |   #type: "Convolution"
1253 |   type: "DepthwiseConvolution"
1254 |   bottom: "conv4-3/expand/hswish"
1255 |   top: "conv4-3/dwise"
1256 |   param {
1257 |     lr_mult: 1
1258 |     decay_mult: 1
1259 |   }
1260 |   convolution_param {
1261 |     num_output: 240
1262 |     group: 240
1263 |     pad: 1
1264 |     bias_term: false
1265 |     kernel_size: 3
1266 |     stride: 2
1267 |     weight_filler {
1268 |       type: "msra"
1269 |     }
1270 |     engine: CAFFE
1271 |   }
1272 | }
1273 | layer {
1274 |   name: "conv4-3/dwise-bn"
1275 |   type: "BatchNorm"
1276 |   bottom: "conv4-3/dwise"
1277 |   top: "conv4-3/dwise"
1278 | }
1279 | layer {
1280 |   name: "conv4-3/dwise-bn-scale"
1281 |   type: "Scale"
1282 |   bottom: "conv4-3/dwise"
1283 |   top: "conv4-3/dwise"
1284 |   scale_param {
1285 |     bias_term: true
1286 |   }
1287 | }
1288 | # H-swish
1289 | layer {
1290 |   name: "conv4-3/dwise/shift"
1291 |   type: "Power"
1292 |   bottom: "conv4-3/dwise"
1293 |   top: "conv4-3/dwise/shift"
1294 |   power_param {
1295 |     power: 1
1296 |     scale: 1
1297 |     shift: 3
1298 |   }
1299 | }
1300 | layer {
1301 |   name: "conv4-3/dwise/shift-relu"
1302 |   type: "ReLU6"
1303 |   bottom: "conv4-3/dwise/shift"
1304 |   top: "conv4-3/dwise/shift"
1305 | }
1306 | layer {
1307 |   name: "conv4-3/dwise/shift-div"
1308 |   type: "Power"
1309 |   bottom: "conv4-3/dwise/shift"
1310 |   top: "conv4-3/dwise/shift-div"
1311 |   power_param {
1312 |     power: 1
1313 |     scale: 0.1666666667
1314 |     shift: 0
1315 |   }
1316 | }
1317 | layer {
1318 |   name: "conv4-3/dwise/hswish"
1319 |   type: "Eltwise"
1320 |   bottom: "conv4-3/dwise"
1321 |   bottom: "conv4-3/dwise/shift-div"
1322 |   top: "conv4-3/dwise/hswish"
1323 |   eltwise_param {
1324 |     operation: PROD
1325 |   }
1326 | }
1327 | 
1328 | layer {
1329 |   name: "conv4-3/linear"
1330 |   type: "Convolution"
1331 |   bottom: "conv4-3/dwise/hswish"
1332 |   top: "conv4-3/linear"
1333 |   param {
1334 |     lr_mult: 1
1335 |     decay_mult: 1
1336 |   }
1337 |   convolution_param {
1338 |     num_output: 80
1339 |     pad: 0
1340 |     bias_term: false
1341 |     kernel_size: 1
1342 |     stride: 1
1343 |     weight_filler {
1344 |       type: "msra"
1345 |     }
1346 |   }
1347 | }
1348 | layer {
1349 |   name: "conv4-3/linear-bn"
1350 |   type: "BatchNorm"
1351 |   bottom: "conv4-3/linear"
1352 |   top: "conv4-3/linear"
1353 | }
1354 | layer {
1355 |   name: "conv4-3/linear-bn-scale"
1356 |   type: "Scale"
1357 |   bottom: "conv4-3/linear"
1358 |   top: "conv4-3/linear"
1359 |   scale_param {
1360 |     bias_term: true
1361 |   }
1362 | }
1363 | 
1364 | #################### stage 5-1 ####################
1365 | layer {
1366 |   name: "conv5-1/expand"
1367 |   type: "Convolution"
1368 |   bottom: "conv4-3/linear"
1369 |   top: "conv5-1/expand"
1370 |   param {
1371 |     lr_mult: 1
1372 |     decay_mult: 1
1373 |   }
1374 |   convolution_param {
1375 |     num_output: 200
1376 |     pad: 0
1377 |     bias_term: false
1378 |     kernel_size: 1
1379 |     stride: 1
1380 |     weight_filler {
1381 |       type: "msra"
1382 |     }
1383 |   }
1384 | }
1385 | layer {
1386 |   name: "conv5-1/expand-bn"
1387 |   type: "BatchNorm"
1388 |   bottom: "conv5-1/expand"
1389 |   top: "conv5-1/expand"
1390 | }
1391 | layer {
1392 |   name: "conv5-1/expand-bn-scale"
1393 |   type: "Scale"
1394 |   bottom: "conv5-1/expand"
1395 |   top: "conv5-1/expand"
1396 |   scale_param {
1397 |     bias_term: true
1398 |   }
1399 | }
1400 | # H-swish
1401 | layer {
1402 |   name: "conv5-1/expand/shift"
1403 |   type: "Power"
1404 |   bottom: "conv5-1/expand"
1405 |   top: "conv5-1/expand/shift"
1406 |   power_param {
1407 |     power: 1
1408 |     scale: 1
1409 |     shift: 3
1410 |   }
1411 | }
1412 | layer {
1413 |   name: "conv5-1/expand/shift-relu"
1414 |   type: "ReLU6"
1415 |   bottom: "conv5-1/expand/shift"
1416 |   top: "conv5-1/expand/shift"
1417 | }
1418 | layer {
1419 |   name: "conv5-1/expand/shift-div"
1420 |   type: "Power"
1421 |   bottom: "conv5-1/expand/shift"
1422 |   top: "conv5-1/expand/shift-div"
1423 |   power_param {
1424 |     power: 1
1425 |     scale: 0.1666666667
1426 |     shift: 0
1427 |   }
1428 | }
1429 | layer {
1430 |   name: "conv5-1/expand/hswish"
1431 |   type: "Eltwise"
1432 |   bottom: "conv5-1/expand"
1433 |   bottom: "conv5-1/expand/shift-div"
1434 |   top: "conv5-1/expand/hswish"
1435 |   eltwise_param {
1436 |     operation: PROD
1437 |   }
1438 | }
1439 | 
1440 | layer {
1441 |   name: "conv5-1/dwise"
1442 |   #type: "Convolution"
1443 |   type: "DepthwiseConvolution"
1444 |   bottom: "conv5-1/expand/hswish"
1445 |   top: "conv5-1/dwise"
1446 |   param {
1447 |     lr_mult: 1
1448 |     decay_mult: 1
1449 |   }
1450 |   convolution_param {
1451 |     num_output: 200
1452 |     group: 200
1453 |     pad: 1
1454 |     bias_term: false
1455 |     kernel_size: 3
1456 |     stride: 1
1457 |     weight_filler {
1458 |       type: "msra"
1459 |     }
1460 |     engine: CAFFE
1461 |   }
1462 | }
1463 | layer {
1464 |   name: "conv5-1/dwise-bn"
1465 |   type: "BatchNorm"
1466 |   bottom: "conv5-1/dwise"
1467 |   top: "conv5-1/dwise"
1468 | }
1469 | layer {
1470 |   name: "conv5-1/dwise-bn-scale"
1471 |   type: "Scale"
1472 |   bottom: "conv5-1/dwise"
1473 |   top: "conv5-1/dwise"
1474 |   scale_param {
1475 |     bias_term: true
1476 |   }
1477 | }
1478 | # H-swish
1479 | layer {
1480 |   name: "conv5-1/dwise/shift"
1481 |   type: "Power"
1482 |   bottom: "conv5-1/dwise"
1483 |   top: "conv5-1/dwise/shift"
1484 |   power_param {
1485 |     power: 1
1486 |     scale: 1
1487 |     shift: 3
1488 |   }
1489 | }
1490 | layer {
1491 |   name: "conv5-1/dwise/shift-relu"
1492 |   type: "ReLU6"
1493 |   bottom: "conv5-1/dwise/shift"
1494 |   top: "conv5-1/dwise/shift"
1495 | }
1496 | layer {
1497 |   name: "conv5-1/dwise/shift-div"
1498 |   type: "Power"
1499 |   bottom: "conv5-1/dwise/shift"
1500 |   top: "conv5-1/dwise/shift-div"
1501 |   power_param {
1502 |     power: 1
1503 |     scale: 0.1666666667
1504 |     shift: 0
1505 |   }
1506 | }
1507 | layer {
1508 |   name: "conv5-1/dwise/hswish"
1509 |   type: "Eltwise"
1510 |   bottom: "conv5-1/dwise"
1511 |   bottom: "conv5-1/dwise/shift-div"
1512 |   top: "conv5-1/dwise/hswish"
1513 |   eltwise_param {
1514 |     operation: PROD
1515 |   }
1516 | }
1517 | 
1518 | layer {
1519 |   name: "conv5-1/linear"
1520 |   type: "Convolution"
1521 |   bottom: "conv5-1/dwise/hswish"
1522 |   top: "conv5-1/linear"
1523 |   param {
1524 |     lr_mult: 1
1525 |     decay_mult: 1
1526 |   }
1527 |   convolution_param {
1528 |     num_output: 80
1529 |     pad: 0
1530 |     bias_term: false
1531 |     kernel_size: 1
1532 |     stride: 1
1533 |     weight_filler {
1534 |       type: "msra"
1535 |     }
1536 |   }
1537 | }
1538 | layer {
1539 |   name: "conv5-1/linear-bn"
1540 |   type: "BatchNorm"
1541 |   bottom: "conv5-1/linear"
1542 |   top: "conv5-1/linear"
1543 | }
1544 | layer {
1545 |   name: "conv5-1/linear-bn-scale"
1546 |   type: "Scale"
1547 |   bottom: "conv5-1/linear"
1548 |   top: "conv5-1/linear"
1549 |   scale_param {
1550 |     bias_term: true
1551 |   }
1552 | }
1553 | 
1554 | layer {
1555 |   name: "block5-1"
1556 |   type: "Eltwise"
1557 |   bottom: "conv4-3/linear"
1558 |   bottom: "conv5-1/linear"
1559 |   top: "block5-1"
1560 | }
1561 | 
1562 | #################### stage 5-2 ####################
1563 | layer {
1564 |   name: "conv5-2/expand"
1565 |   type: "Convolution"
1566 |   bottom: "block5-1"
1567 |   top: "conv5-2/expand"
1568 |   param {
1569 |     lr_mult: 1
1570 |     decay_mult: 1
1571 |   }
1572 |   convolution_param {
1573 |     num_output: 184
1574 |     pad: 0
1575 |     bias_term: false
1576 |     kernel_size: 1
1577 |     stride: 1
1578 |     weight_filler {
1579 |       type: "msra"
1580 |     }
1581 |   }
1582 | }
1583 | layer {
1584 |   name: "conv5-2/expand-bn"
1585 |   type: "BatchNorm"
1586 |   bottom: "conv5-2/expand"
1587 |   top: "conv5-2/expand"
1588 | }
1589 | layer {
1590 |   name: "conv5-2/expand-bn-scale"
1591 |   type: "Scale"
1592 |   bottom: "conv5-2/expand"
1593 |   top: "conv5-2/expand"
1594 |   scale_param {
1595 |     bias_term: true
1596 |   }
1597 | }
1598 | # H-swish
1599 | layer {
1600 |   name: "conv5-2/expand/shift"
1601 |   type: "Power"
1602 |   bottom: "conv5-2/expand"
1603 |   top: "conv5-2/expand/shift"
1604 |   power_param {
1605 |     power: 1
1606 |     scale: 1
1607 |     shift: 3
1608 |   }
1609 | }
1610 | layer {
1611 |   name: "conv5-2/expand/shift-relu"
1612 |   type: "ReLU6"
1613 |   bottom: "conv5-2/expand/shift"
1614 |   top: "conv5-2/expand/shift"
1615 | }
1616 | layer {
1617 |   name: "conv5-2/expand/shift-div"
1618 |   type: "Power"
1619 |   bottom: "conv5-2/expand/shift"
1620 |   top: "conv5-2/expand/shift-div"
1621 |   power_param {
1622 |     power: 1
1623 |     scale: 0.1666666667
1624 |     shift: 0
1625 |   }
1626 | }
1627 | layer {
1628 |   name: "conv5-2/expand/hswish"
1629 |   type: "Eltwise"
1630 |   bottom: "conv5-2/expand"
1631 |   bottom: "conv5-2/expand/shift-div"
1632 |   top: "conv5-2/expand/hswish"
1633 |   eltwise_param {
1634 |     operation: PROD
1635 |   }
1636 | }
1637 | 
1638 | layer {
1639 |   name: "conv5-2/dwise"
1640 |   #type: "Convolution"
1641 |   type: "DepthwiseConvolution"
1642 |   bottom: "conv5-2/expand/hswish"
1643 |   top: "conv5-2/dwise"
1644 |   param {
1645 |     lr_mult: 1
1646 |     decay_mult: 1
1647 |   }
1648 |   convolution_param {
1649 |     num_output: 184
1650 |     group: 184
1651 |     pad: 1
1652 |     bias_term: false
1653 |     kernel_size: 3
1654 |     stride: 1
1655 |     weight_filler {
1656 |       type: "msra"
1657 |     }
1658 |     engine: CAFFE
1659 |   }
1660 | }
1661 | layer {
1662 |   name: "conv5-2/dwise-bn"
1663 |   type: "BatchNorm"
1664 |   bottom: "conv5-2/dwise"
1665 |   top: "conv5-2/dwise"
1666 | }
1667 | layer {
1668 |   name: "conv5-2/dwise-bn-scale"
1669 |   type: "Scale"
1670 |   bottom: "conv5-2/dwise"
1671 |   top: "conv5-2/dwise"
1672 |   scale_param {
1673 |     bias_term: true
1674 |   }
1675 | }
1676 | # H-swish
1677 | layer {
1678 |   name: "conv5-2/dwise/shift"
1679 |   type: "Power"
1680 |   bottom: "conv5-2/dwise"
1681 |   top: "conv5-2/dwise/shift"
1682 |   power_param {
1683 |     power: 1
1684 |     scale: 1
1685 |     shift: 3
1686 |   }
1687 | }
1688 | layer {
1689 |   name: "conv5-2/dwise/shift-relu"
1690 |   type: "ReLU6"
1691 |   bottom: "conv5-2/dwise/shift"
1692 |   top: "conv5-2/dwise/shift"
1693 | }
1694 | layer {
1695 |   name: "conv5-2/dwise/shift-div"
1696 |   type: "Power"
1697 |   bottom: "conv5-2/dwise/shift"
1698 |   top: "conv5-2/dwise/shift-div"
1699 |   power_param {
1700 |     power: 1
1701 |     scale: 0.1666666667
1702 |     shift: 0
1703 |   }
1704 | }
1705 | layer {
1706 |   name: "conv5-2/dwise/hswish"
1707 |   type: "Eltwise"
1708 |   bottom: "conv5-2/dwise"
1709 |   bottom: "conv5-2/dwise/shift-div"
1710 |   top: "conv5-2/dwise/hswish"
1711 |   eltwise_param {
1712 |     operation: PROD
1713 |   }
1714 | }
1715 | 
1716 | layer {
1717 |   name: "conv5-2/linear"
1718 |   type: "Convolution"
1719 |   bottom: "conv5-2/dwise/hswish"
1720 |   top: "conv5-2/linear"
1721 |   param {
1722 |     lr_mult: 1
1723 |     decay_mult: 1
1724 |   }
1725 |   convolution_param {
1726 |     num_output: 80
1727 |     pad: 0
1728 |     bias_term: false
1729 |     kernel_size: 1
1730 |     stride: 1
1731 |     weight_filler {
1732 |       type: "msra"
1733 |     }
1734 |   }
1735 | }
1736 | layer {
1737 |   name: "conv5-2/linear-bn"
1738 |   type: "BatchNorm"
1739 |   bottom: "conv5-2/linear"
1740 |   top: "conv5-2/linear"
1741 | }
1742 | layer {
1743 |   name: "conv5-2/linear-bn-scale"
1744 |   type: "Scale"
1745 |   bottom: "conv5-2/linear"
1746 |   top: "conv5-2/linear"
1747 |   scale_param {
1748 |     bias_term: true
1749 |   }
1750 | }
1751 | 
1752 | layer {
1753 |   name: "block5-2"
1754 |   type: "Eltwise"
1755 |   bottom: "block5-1"
1756 |   bottom: "conv5-2/linear"
1757 |   top: "block5-2"
1758 | }
1759 | 
1760 | #################### stage 5-3 ####################
1761 | layer {
1762 |   name: "conv5-3/expand"
1763 |   type: "Convolution"
1764 |   bottom: "block5-2"
1765 |   top: "conv5-3/expand"
1766 |   param {
1767 |     lr_mult: 1
1768 |     decay_mult: 1
1769 |   }
1770 |   convolution_param {
1771 |     num_output: 184
1772 |     pad: 0
1773 |     bias_term: false
1774 |     kernel_size: 1
1775 |     stride: 1
1776 |     weight_filler {
1777 |       type: "msra"
1778 |     }
1779 |   }
1780 | }
1781 | layer {
1782 |   name: "conv5-3/expand-bn"
1783 |   type: "BatchNorm"
1784 |   bottom: "conv5-3/expand"
1785 |   top: "conv5-3/expand"
1786 | }
1787 | layer {
1788 |   name: "conv5-3/expand-bn-scale"
1789 |   type: "Scale"
1790 |   bottom: "conv5-3/expand"
1791 |   top: "conv5-3/expand"
1792 |   scale_param {
1793 |     bias_term: true
1794 |   }
1795 | }
1796 | # H-swish
1797 | layer {
1798 |   name: "conv5-3/expand/shift"
1799 |   type: "Power"
1800 |   bottom: "conv5-3/expand"
1801 |   top: "conv5-3/expand/shift"
1802 |   power_param {
1803 |     power: 1
1804 |     scale: 1
1805 |     shift: 3
1806 |   }
1807 | }
1808 | layer {
1809 |   name: "conv5-3/expand/shift-relu"
1810 |   type: "ReLU6"
1811 |   bottom: "conv5-3/expand/shift"
1812 |   top: "conv5-3/expand/shift"
1813 | }
1814 | layer {
1815 |   name: "conv5-3/expand/shift-div"
1816 |   type: "Power"
1817 |   bottom: "conv5-3/expand/shift"
1818 |   top: "conv5-3/expand/shift-div"
1819 |   power_param {
1820 |     power: 1
1821 |     scale: 0.1666666667
1822 |     shift: 0
1823 |   }
1824 | }
1825 | layer {
1826 |   name: "conv5-3/expand/hswish"
1827 |   type: "Eltwise"
1828 |   bottom: "conv5-3/expand"
1829 |   bottom: "conv5-3/expand/shift-div"
1830 |   top: "conv5-3/expand/hswish"
1831 |   eltwise_param {
1832 |     operation: PROD
1833 |   }
1834 | }
1835 | 
1836 | layer {
1837 |   name: "conv5-3/dwise"
1838 |   #type: "Convolution"
1839 |   type: "DepthwiseConvolution"
1840 |   bottom: "conv5-3/expand/hswish"
1841 |   top: "conv5-3/dwise"
1842 |   param {
1843 |     lr_mult: 1
1844 |     decay_mult: 1
1845 |   }
1846 |   convolution_param {
1847 |     num_output: 184
1848 |     group: 184
1849 |     pad: 1
1850 |     bias_term: false
1851 |     kernel_size: 3
1852 |     stride: 1
1853 |     weight_filler {
1854 |       type: "msra"
1855 |     }
1856 |     engine: CAFFE
1857 |   }
1858 | }
1859 | layer {
1860 |   name: "conv5-3/dwise-bn"
1861 |   type: "BatchNorm"
1862 |   bottom: "conv5-3/dwise"
1863 |   top: "conv5-3/dwise"
1864 | }
1865 | layer {
1866 |   name: "conv5-3/dwise-bn-scale"
1867 |   type: "Scale"
1868 |   bottom: "conv5-3/dwise"
1869 |   top: "conv5-3/dwise"
1870 |   scale_param {
1871 |     bias_term: true
1872 |   }
1873 | }
1874 | # H-swish
1875 | layer {
1876 |   name: "conv5-3/dwise/shift"
1877 |   type: "Power"
1878 |   bottom: "conv5-3/dwise"
1879 |   top: "conv5-3/dwise/shift"
1880 |   power_param {
1881 |     power: 1
1882 |     scale: 1
1883 |     shift: 3
1884 |   }
1885 | }
1886 | layer {
1887 |   name: "conv5-3/dwise/shift-relu"
1888 |   type: "ReLU6"
1889 |   bottom: "conv5-3/dwise/shift"
1890 |   top: "conv5-3/dwise/shift"
1891 | }
1892 | layer {
1893 |   name: "conv5-3/dwise/shift-div"
1894 |   type: "Power"
1895 |   bottom: "conv5-3/dwise/shift"
1896 |   top: "conv5-3/dwise/shift-div"
1897 |   power_param {
1898 |     power: 1
1899 |     scale: 0.1666666667
1900 |     shift: 0
1901 |   }
1902 | }
1903 | layer {
1904 |   name: "conv5-3/dwise/hswish"
1905 |   type: "Eltwise"
1906 |   bottom: "conv5-3/dwise"
1907 |   bottom: "conv5-3/dwise/shift-div"
1908 |   top: "conv5-3/dwise/hswish"
1909 |   eltwise_param {
1910 |     operation: PROD
1911 |   }
1912 | }
1913 | 
1914 | layer {
1915 |   name: "conv5-3/linear"
1916 |   type: "Convolution"
1917 |   bottom: "conv5-3/dwise/hswish"
1918 |   top: "conv5-3/linear"
1919 |   param {
1920 |     lr_mult: 1
1921 |     decay_mult: 1
1922 |   }
1923 |   convolution_param {
1924 |     num_output: 80
1925 |     pad: 0
1926 |     bias_term: false
1927 |     kernel_size: 1
1928 |     stride: 1
1929 |     weight_filler {
1930 |       type: "msra"
1931 |     }
1932 |   }
1933 | }
1934 | layer {
1935 |   name: "conv5-3/linear-bn"
1936 |   type: "BatchNorm"
1937 |   bottom: "conv5-3/linear"
1938 |   top: "conv5-3/linear"
1939 | }
1940 | layer {
1941 |   name: "conv5-3/linear-bn-scale"
1942 |   type: "Scale"
1943 |   bottom: "conv5-3/linear"
1944 |   top: "conv5-3/linear"
1945 |   scale_param {
1946 |     bias_term: true
1947 |   }
1948 | }
1949 | 
1950 | layer {
1951 |   name: "block5-3"
1952 |   type: "Eltwise"
1953 |   bottom: "block5-2"
1954 |   bottom: "conv5-3/linear"
1955 |   top: "block5-3"
1956 | }
1957 | 
1958 | #################### stage 5-4 ####################
1959 | layer {
1960 |   name: "conv5-4/expand"
1961 |   type: "Convolution"
1962 |   bottom: "block5-3"
1963 |   top: "conv5-4/expand"
1964 |   param {
1965 |     lr_mult: 1
1966 |     decay_mult: 1
1967 |   }
1968 |   convolution_param {
1969 |     num_output: 480
1970 |     pad: 0
1971 |     bias_term: false
1972 |     kernel_size: 1
1973 |     stride: 1
1974 |     weight_filler {
1975 |       type: "msra"
1976 |     }
1977 |   }
1978 | }
1979 | layer {
1980 |   name: "conv5-4/expand-bn"
1981 |   type: "BatchNorm"
1982 |   bottom: "conv5-4/expand"
1983 |   top: "conv5-4/expand"
1984 | }
1985 | layer {
1986 |   name: "conv5-4/expand-bn-scale"
1987 |   type: "Scale"
1988 |   bottom: "conv5-4/expand"
1989 |   top: "conv5-4/expand"
1990 |   scale_param {
1991 |     bias_term: true
1992 |   }
1993 | }
1994 | # H-swish
1995 | layer {
1996 |   name: "conv5-4/expand/shift"
1997 |   type: "Power"
1998 |   bottom: "conv5-4/expand"
1999 |   top: "conv5-4/expand/shift"
2000 |   power_param {
2001 |     power: 1
2002 |     scale: 1
2003 |     shift: 3
2004 |   }
2005 | }
2006 | layer {
2007 |   name: "conv5-4/expand/shift-relu"
2008 |   type: "ReLU6"
2009 |   bottom: "conv5-4/expand/shift"
2010 |   top: "conv5-4/expand/shift"
2011 | }
2012 | layer {
2013 |   name: "conv5-4/expand/shift-div"
2014 |   type: "Power"
2015 |   bottom: "conv5-4/expand/shift"
2016 |   top: "conv5-4/expand/shift-div"
2017 |   power_param {
2018 |     power: 1
2019 |     scale: 0.1666666667
2020 |     shift: 0
2021 |   }
2022 | }
2023 | layer {
2024 |   name: "conv5-4/expand/hswish"
2025 |   type: "Eltwise"
2026 |   bottom: "conv5-4/expand"
2027 |   bottom: "conv5-4/expand/shift-div"
2028 |   top: "conv5-4/expand/hswish"
2029 |   eltwise_param {
2030 |     operation: PROD
2031 |   }
2032 | }
2033 | 
2034 | layer {
2035 |   name: "conv5-4/dwise"
2036 |   #type: "Convolution"
2037 |   type: "DepthwiseConvolution"
2038 |   bottom: "conv5-4/expand/hswish"
2039 |   top: "conv5-4/dwise"
2040 |   param {
2041 |     lr_mult: 1
2042 |     decay_mult: 1
2043 |   }
2044 |   convolution_param {
2045 |     num_output: 480
2046 |     group: 480
2047 |     pad: 1
2048 |     bias_term: false
2049 |     kernel_size: 3
2050 |     stride: 1
2051 |     weight_filler {
2052 |       type: "msra"
2053 |     }
2054 |     engine: CAFFE
2055 |   }
2056 | }
2057 | layer {
2058 |   name: "conv5-4/dwise-bn"
2059 |   type: "BatchNorm"
2060 |   bottom: "conv5-4/dwise"
2061 |   top: "conv5-4/dwise"
2062 | }
2063 | layer {
2064 |   name: "conv5-4/dwise-bn-scale"
2065 |   type: "Scale"
2066 |   bottom: "conv5-4/dwise"
2067 |   top: "conv5-4/dwise"
2068 |   scale_param {
2069 |     bias_term: true
2070 |   }
2071 | }
2072 | 
2073 | # se
2074 | layer {
2075 |   name: "conv5-4/dwise-se-pool"
2076 |   type: "Pooling"
2077 |   bottom: "conv5-4/dwise"
2078 |   top: "conv5-4/dwise-se-pool"
2079 |   pooling_param {
2080 |     pool: AVE
2081 |     global_pooling: true
2082 |   }
2083 | }
2084 | layer {
2085 |   name: "conv5-4/dwise-se-fc1"
2086 |   type: "InnerProduct"
2087 |   bottom: "conv5-4/dwise-se-pool"
2088 |   top:"conv5-4/dwise-se-fc1"
2089 |   param {
2090 |     lr_mult: 1
2091 |     decay_mult: 1
2092 |   }
2093 |   param {
2094 |     lr_mult: 2
2095 |     decay_mult: 0
2096 |   }
2097 |   inner_product_param {
2098 |     num_output: 120
2099 |     weight_filler {
2100 |       type: "msra"
2101 |     }
2102 |     bias_filler {
2103 |       type: "constant"
2104 |       value: 0
2105 |     }
2106 |   }
2107 | }
2108 | layer {
2109 |   name: "conv5-4/dwise-se-fc1-relu"
2110 |   type: "ReLU6"
2111 |   bottom: "conv5-4/dwise-se-fc1"
2112 |   top: "conv5-4/dwise-se-fc1"
2113 | }
2114 | layer {
2115 |   name: "conv5-4/dwise-se-fc2"
2116 |   type: "InnerProduct"
2117 |   bottom: "conv5-4/dwise-se-fc1"
2118 |   top:"conv5-4/dwise-se-fc2"
2119 |   param {
2120 |     lr_mult: 1
2121 |     decay_mult: 1
2122 |   }
2123 |   param {
2124 |     lr_mult: 2
2125 |     decay_mult: 0
2126 |   }
2127 |   inner_product_param {
2128 |     num_output: 480
2129 |     weight_filler {
2130 |       type: "msra"
2131 |     }
2132 |     bias_filler {
2133 |       type: "constant"
2134 |       value: 0
2135 |     }
2136 |   }
2137 | }
2138 | # H-sigmoid
2139 | layer {
2140 |   name: "conv5-4/dwise-se-fc2/shift"
2141 |   type: "Power"
2142 |   bottom: "conv5-4/dwise-se-fc2"
2143 |   top: "conv5-4/dwise-se-fc2/shift"
2144 |   power_param {
2145 |     power: 1
2146 |     scale: 1
2147 |     shift: 3
2148 |   }
2149 | }
2150 | layer {
2151 |   name: "conv5-4/dwise-se-fc2/shift-relu"
2152 |   type: "ReLU6"
2153 |   bottom: "conv5-4/dwise-se-fc2/shift"
2154 |   top: "conv5-4/dwise-se-fc2/shift"
2155 | }
2156 | layer {
2157 |   name: "conv5-4/dwise-se-fc2/shift-div"
2158 |   type: "Power"
2159 |   bottom: "conv5-4/dwise-se-fc2/shift"
2160 |   top: "conv5-4/dwise-se-fc2/shift-div"
2161 |   power_param {
2162 |     power: 1
2163 |     scale: 0.1666666667
2164 |     shift: 0
2165 |   }
2166 | }
2167 | layer {
2168 |   name: "conv5-4/dwise/scale"
2169 |   type: "Scale"
2170 |   bottom: "conv5-4/dwise"
2171 |   bottom: "conv5-4/dwise-se-fc2/shift-div"
2172 |   top: "conv5-4/dwise/scale"
2173 |   scale_param{
2174 |     axis: 0
2175 |   }
2176 | }
2177 | # H-swish
2178 | layer {
2179 |   name: "conv5-4/dwise/shift"
2180 |   type: "Power"
2181 |   bottom: "conv5-4/dwise/scale"
2182 |   top: "conv5-4/dwise/shift"
2183 |   power_param {
2184 |     power: 1
2185 |     scale: 1
2186 |     shift: 3
2187 |   }
2188 | }
2189 | layer {
2190 |   name: "conv5-4/dwise/shift-relu"
2191 |   type: "ReLU6"
2192 |   bottom: "conv5-4/dwise/shift"
2193 |   top: "conv5-4/dwise/shift"
2194 | }
2195 | layer {
2196 |   name: "conv5-4/dwise/shift-div"
2197 |   type: "Power"
2198 |   bottom: "conv5-4/dwise/shift"
2199 |   top: "conv5-4/dwise/shift-div"
2200 |   power_param {
2201 |     power: 1
2202 |     scale: 0.1666666667
2203 |     shift: 0
2204 |   }
2205 | }
2206 | layer {
2207 |   name: "conv5-4/dwise/scale/hswish"
2208 |   type: "Eltwise"
2209 |   bottom: "conv5-4/dwise/scale"
2210 |   bottom: "conv5-4/dwise/shift-div"
2211 |   top: "conv5-4/dwise/scale/hswish"
2212 |   eltwise_param {
2213 |     operation: PROD
2214 |   }
2215 | }
2216 | 
2217 | layer {
2218 |   name: "conv5-4/linear"
2219 |   type: "Convolution"
2220 |   bottom: "conv5-4/dwise/scale/hswish"
2221 |   top: "conv5-4/linear"
2222 |   param {
2223 |     lr_mult: 1
2224 |     decay_mult: 1
2225 |   }
2226 |   convolution_param {
2227 |     num_output: 112
2228 |     pad: 0
2229 |     bias_term: false
2230 |     kernel_size: 1
2231 |     stride: 1
2232 |     weight_filler {
2233 |       type: "msra"
2234 |     }
2235 |   }
2236 | }
2237 | layer {
2238 |   name: "conv5-4/linear-bn"
2239 |   type: "BatchNorm"
2240 |   bottom: "conv5-4/linear"
2241 |   top: "conv5-4/linear"
2242 | }
2243 | layer {
2244 |   name: "conv5-4/linear-bn-scale"
2245 |   type: "Scale"
2246 |   bottom: "conv5-4/linear"
2247 |   top: "conv5-4/linear"
2248 |   scale_param {
2249 |     bias_term: true
2250 |   }
2251 | }
2252 | 
2253 | #################### stage 5-5 ####################
2254 | layer {
2255 |   name: "conv5-5/expand"
2256 |   type: "Convolution"
2257 |   bottom: "conv5-4/linear"
2258 |   top: "conv5-5/expand"
2259 |   param {
2260 |     lr_mult: 1
2261 |     decay_mult: 1
2262 |   }
2263 |   convolution_param {
2264 |     num_output: 672
2265 |     pad: 0
2266 |     bias_term: false
2267 |     kernel_size: 1
2268 |     stride: 1
2269 |     weight_filler {
2270 |       type: "msra"
2271 |     }
2272 |   }
2273 | }
2274 | layer {
2275 |   name: "conv5-5/expand-bn"
2276 |   type: "BatchNorm"
2277 |   bottom: "conv5-5/expand"
2278 |   top: "conv5-5/expand"
2279 | }
2280 | layer {
2281 |   name: "conv5-5/expand-bn-scale"
2282 |   type: "Scale"
2283 |   bottom: "conv5-5/expand"
2284 |   top: "conv5-5/expand"
2285 |   scale_param {
2286 |     bias_term: true
2287 |   }
2288 | }
2289 | # H-swish
2290 | layer {
2291 |   name: "conv5-5/expand/shift"
2292 |   type: "Power"
2293 |   bottom: "conv5-5/expand"
2294 |   top: "conv5-5/expand/shift"
2295 |   power_param {
2296 |     power: 1
2297 |     scale: 1
2298 |     shift: 3
2299 |   }
2300 | }
2301 | layer {
2302 |   name: "conv5-5/expand/shift-relu"
2303 |   type: "ReLU6"
2304 |   bottom: "conv5-5/expand/shift"
2305 |   top: "conv5-5/expand/shift"
2306 | }
2307 | layer {
2308 |   name: "conv5-5/expand/shift-div"
2309 |   type: "Power"
2310 |   bottom: "conv5-5/expand/shift"
2311 |   top: "conv5-5/expand/shift-div"
2312 |   power_param {
2313 |     power: 1
2314 |     scale: 0.1666666667
2315 |     shift: 0
2316 |   }
2317 | }
2318 | layer {
2319 |   name: "conv5-5/expand/hswish"
2320 |   type: "Eltwise"
2321 |   bottom: "conv5-5/expand"
2322 |   bottom: "conv5-5/expand/shift-div"
2323 |   top: "conv5-5/expand/hswish"
2324 |   eltwise_param {
2325 |     operation: PROD
2326 |   }
2327 | }
2328 | 
2329 | layer {
2330 |   name: "conv5-5/dwise"
2331 |   #type: "Convolution"
2332 |   type: "DepthwiseConvolution"
2333 |   bottom: "conv5-5/expand/hswish"
2334 |   top: "conv5-5/dwise"
2335 |   param {
2336 |     lr_mult: 1
2337 |     decay_mult: 1
2338 |   }
2339 |   convolution_param {
2340 |     num_output: 672
2341 |     group: 672
2342 |     pad: 1
2343 |     bias_term: false
2344 |     kernel_size: 3
2345 |     stride: 1
2346 |     weight_filler {
2347 |       type: "msra"
2348 |     }
2349 |     engine: CAFFE
2350 |   }
2351 | }
2352 | layer {
2353 |   name: "conv5-5/dwise-bn"
2354 |   type: "BatchNorm"
2355 |   bottom: "conv5-5/dwise"
2356 |   top: "conv5-5/dwise"
2357 | }
2358 | layer {
2359 |   name: "conv5-5/dwise-bn-scale"
2360 |   type: "Scale"
2361 |   bottom: "conv5-5/dwise"
2362 |   top: "conv5-5/dwise"
2363 |   scale_param {
2364 |     bias_term: true
2365 |   }
2366 | }
2367 | 
2368 | # se
2369 | layer {
2370 |   name: "conv5-5/dwise-se-pool"
2371 |   type: "Pooling"
2372 |   bottom: "conv5-5/dwise"
2373 |   top: "conv5-5/dwise-se-pool"
2374 |   pooling_param {
2375 |     pool: AVE
2376 |     global_pooling: true
2377 |   }
2378 | }
2379 | layer {
2380 |   name: "conv5-5/dwise-se-fc1"
2381 |   type: "InnerProduct"
2382 |   bottom: "conv5-5/dwise-se-pool"
2383 |   top:"conv5-5/dwise-se-fc1"
2384 |   param {
2385 |     lr_mult: 1
2386 |     decay_mult: 1
2387 |   }
2388 |   param {
2389 |     lr_mult: 2
2390 |     decay_mult: 0
2391 |   }
2392 |   inner_product_param {
2393 |     num_output: 168
2394 |     weight_filler {
2395 |       type: "msra"
2396 |     }
2397 |     bias_filler {
2398 |       type: "constant"
2399 |       value: 0
2400 |     }
2401 |   }
2402 | }
2403 | layer {
2404 |   name: "conv5-5/dwise-se-fc1-relu"
2405 |   type: "ReLU6"
2406 |   bottom: "conv5-5/dwise-se-fc1"
2407 |   top: "conv5-5/dwise-se-fc1"
2408 | }
2409 | layer {
2410 |   name: "conv5-5/dwise-se-fc2"
2411 |   type: "InnerProduct"
2412 |   bottom: "conv5-5/dwise-se-fc1"
2413 |   top:"conv5-5/dwise-se-fc2"
2414 |   param {
2415 |     lr_mult: 1
2416 |     decay_mult: 1
2417 |   }
2418 |   param {
2419 |     lr_mult: 2
2420 |     decay_mult: 0
2421 |   }
2422 |   inner_product_param {
2423 |     num_output: 672
2424 |     weight_filler {
2425 |       type: "msra"
2426 |     }
2427 |     bias_filler {
2428 |       type: "constant"
2429 |       value: 0
2430 |     }
2431 |   }
2432 | }
2433 | # H-sigmoid
2434 | layer {
2435 |   name: "conv5-5/dwise-se-fc2/shift"
2436 |   type: "Power"
2437 |   bottom: "conv5-5/dwise-se-fc2"
2438 |   top: "conv5-5/dwise-se-fc2/shift"
2439 |   power_param {
2440 |     power: 1
2441 |     scale: 1
2442 |     shift: 3
2443 |   }
2444 | }
2445 | layer {
2446 |   name: "conv5-5/dwise-se-fc2/shift-relu"
2447 |   type: "ReLU6"
2448 |   bottom: "conv5-5/dwise-se-fc2/shift"
2449 |   top: "conv5-5/dwise-se-fc2/shift"
2450 | }
2451 | layer {
2452 |   name: "conv5-5/dwise-se-fc2/shift-div"
2453 |   type: "Power"
2454 |   bottom: "conv5-5/dwise-se-fc2/shift"
2455 |   top: "conv5-5/dwise-se-fc2/shift-div"
2456 |   power_param {
2457 |     power: 1
2458 |     scale: 0.1666666667
2459 |     shift: 0
2460 |   }
2461 | }
2462 | layer {
2463 |   name: "conv5-5/dwise/scale"
2464 |   type: "Scale"
2465 |   bottom: "conv5-5/dwise"
2466 |   bottom: "conv5-5/dwise-se-fc2/shift-div"
2467 |   top: "conv5-5/dwise/scale"
2468 |   scale_param{
2469 |     axis: 0
2470 |   }
2471 | }
2472 | # H-swish
2473 | layer {
2474 |   name: "conv5-5/dwise/shift"
2475 |   type: "Power"
2476 |   bottom: "conv5-5/dwise/scale"
2477 |   top: "conv5-5/dwise/shift"
2478 |   power_param {
2479 |     power: 1
2480 |     scale: 1
2481 |     shift: 3
2482 |   }
2483 | }
2484 | layer {
2485 |   name: "conv5-5/dwise/shift-relu"
2486 |   type: "ReLU6"
2487 |   bottom: "conv5-5/dwise/shift"
2488 |   top: "conv5-5/dwise/shift"
2489 | }
2490 | layer {
2491 |   name: "conv5-5/dwise/shift-div"
2492 |   type: "Power"
2493 |   bottom: "conv5-5/dwise/shift"
2494 |   top: "conv5-5/dwise/shift-div"
2495 |   power_param {
2496 |     power: 1
2497 |     scale: 0.1666666667
2498 |     shift: 0
2499 |   }
2500 | }
2501 | layer {
2502 |   name: "conv5-5/dwise/hswish"
2503 |   type: "Eltwise"
2504 |   bottom: "conv5-5/dwise/scale"
2505 |   bottom: "conv5-5/dwise/shift-div"
2506 |   top: "conv5-5/dwise/hswish"
2507 |   eltwise_param {
2508 |     operation: PROD
2509 |   }
2510 | }
2511 | 
2512 | layer {
2513 |   name: "conv5-5/linear"
2514 |   type: "Convolution"
2515 |   bottom: "conv5-5/dwise/hswish"
2516 |   top: "conv5-5/linear"
2517 |   param {
2518 |     lr_mult: 1
2519 |     decay_mult: 1
2520 |   }
2521 |   convolution_param {
2522 |     num_output: 112
2523 |     pad: 0
2524 |     bias_term: false
2525 |     kernel_size: 1
2526 |     stride: 1
2527 |     weight_filler {
2528 |       type: "msra"
2529 |     }
2530 |   }
2531 | }
2532 | layer {
2533 |   name: "conv5-5/linear-bn"
2534 |   type: "BatchNorm"
2535 |   bottom: "conv5-5/linear"
2536 |   top: "conv5-5/linear"
2537 | }
2538 | layer {
2539 |   name: "conv5-5/linear-bn-scale"
2540 |   type: "Scale"
2541 |   bottom: "conv5-5/linear"
2542 |   top: "conv5-5/linear"
2543 |   scale_param {
2544 |     bias_term: true
2545 |   }
2546 | }
2547 | 
2548 | layer {
2549 |   name: "block5-5"
2550 |   type: "Eltwise"
2551 |   bottom: "conv5-4/linear"
2552 |   bottom: "conv5-5/linear"
2553 |   top: "block5-5"
2554 | }
2555 | 
2556 | #################### stage 5-6 ####################
2557 | layer {
2558 |   name: "conv5-6/expand"
2559 |   type: "Convolution"
2560 |   bottom: "block5-5"
2561 |   top: "conv5-6/expand"
2562 |   param {
2563 |     lr_mult: 1
2564 |     decay_mult: 1
2565 |   }
2566 |   convolution_param {
2567 |     num_output: 672
2568 |     pad: 0
2569 |     bias_term: false
2570 |     kernel_size: 1
2571 |     stride: 1
2572 |     weight_filler {
2573 |       type: "msra"
2574 |     }
2575 |   }
2576 | }
2577 | layer {
2578 |   name: "conv5-6/expand-bn"
2579 |   type: "BatchNorm"
2580 |   bottom: "conv5-6/expand"
2581 |   top: "conv5-6/expand"
2582 | }
2583 | layer {
2584 |   name: "conv5-6/expand-bn-scale"
2585 |   type: "Scale"
2586 |   bottom: "conv5-6/expand"
2587 |   top: "conv5-6/expand"
2588 |   scale_param {
2589 |     bias_term: true
2590 |   }
2591 | }
2592 | # H-swish
2593 | layer {
2594 |   name: "conv5-6/expand/shift"
2595 |   type: "Power"
2596 |   bottom: "conv5-6/expand"
2597 |   top: "conv5-6/expand/shift"
2598 |   power_param {
2599 |     power: 1
2600 |     scale: 1
2601 |     shift: 3
2602 |   }
2603 | }
2604 | layer {
2605 |   name: "conv5-6/expand/shift-relu"
2606 |   type: "ReLU6"
2607 |   bottom: "conv5-6/expand/shift"
2608 |   top: "conv5-6/expand/shift"
2609 | }
2610 | layer {
2611 |   name: "conv5-6/expand/shift-div"
2612 |   type: "Power"
2613 |   bottom: "conv5-6/expand/shift"
2614 |   top: "conv5-6/expand/shift-div"
2615 |   power_param {
2616 |     power: 1
2617 |     scale: 0.1666666667
2618 |     shift: 0
2619 |   }
2620 | }
2621 | layer {
2622 |   name: "conv5-6/expand/hswish"
2623 |   type: "Eltwise"
2624 |   bottom: "conv5-6/expand"
2625 |   bottom: "conv5-6/expand/shift-div"
2626 |   top: "conv5-6/expand/hswish"
2627 |   eltwise_param {
2628 |     operation: PROD
2629 |   }
2630 | }
2631 | 
2632 | layer {
2633 |   name: "conv5-6/dwise"
2634 |   #type: "Convolution"
2635 |   type: "DepthwiseConvolution"
2636 |   bottom: "conv5-6/expand/hswish"
2637 |   top: "conv5-6/dwise"
2638 |   param {
2639 |     lr_mult: 1
2640 |     decay_mult: 1
2641 |   }
2642 |   convolution_param {
2643 |     num_output: 672
2644 |     group: 672
2645 |     pad: 2
2646 |     bias_term: false
2647 |     kernel_size: 5
2648 |     stride: 1
2649 |     weight_filler {
2650 |       type: "msra"
2651 |     }
2652 |     engine: CAFFE
2653 |   }
2654 | }
2655 | layer {
2656 |   name: "conv5-6/dwise-bn"
2657 |   type: "BatchNorm"
2658 |   bottom: "conv5-6/dwise"
2659 |   top: "conv5-6/dwise"
2660 | }
2661 | layer {
2662 |   name: "conv5-6/dwise-bn-scale"
2663 |   type: "Scale"
2664 |   bottom: "conv5-6/dwise"
2665 |   top: "conv5-6/dwise"
2666 |   scale_param {
2667 |     bias_term: true
2668 |   }
2669 | }
2670 | 
2671 | # se
2672 | layer {
2673 |   name: "conv5-6/dwise-se-pool"
2674 |   type: "Pooling"
2675 |   bottom: "conv5-6/dwise"
2676 |   top: "conv5-6/dwise-se-pool"
2677 |   pooling_param {
2678 |     pool: AVE
2679 |     global_pooling: true
2680 |   }
2681 | }
2682 | layer {
2683 |   name: "conv5-6/dwise-se-fc1"
2684 |   type: "InnerProduct"
2685 |   bottom: "conv5-6/dwise-se-pool"
2686 |   top:"conv5-6/dwise-se-fc1"
2687 |   param {
2688 |     lr_mult: 1
2689 |     decay_mult: 1
2690 |   }
2691 |   param {
2692 |     lr_mult: 2
2693 |     decay_mult: 0
2694 |   }
2695 |   inner_product_param {
2696 |     num_output: 168
2697 |     weight_filler {
2698 |       type: "msra"
2699 |     }
2700 |     bias_filler {
2701 |       type: "constant"
2702 |       value: 0
2703 |     }
2704 |   }
2705 | }
2706 | layer {
2707 |   name: "conv5-6/dwise-se-fc1-relu"
2708 |   type: "ReLU6"
2709 |   bottom: "conv5-6/dwise-se-fc1"
2710 |   top: "conv5-6/dwise-se-fc1"
2711 | }
2712 | layer {
2713 |   name: "conv5-6/dwise-se-fc2"
2714 |   type: "InnerProduct"
2715 |   bottom: "conv5-6/dwise-se-fc1"
2716 |   top:"conv5-6/dwise-se-fc2"
2717 |   param {
2718 |     lr_mult: 1
2719 |     decay_mult: 1
2720 |   }
2721 |   param {
2722 |     lr_mult: 2
2723 |     decay_mult: 0
2724 |   }
2725 |   inner_product_param {
2726 |     num_output: 672
2727 |     weight_filler {
2728 |       type: "msra"
2729 |     }
2730 |     bias_filler {
2731 |       type: "constant"
2732 |       value: 0
2733 |     }
2734 |   }
2735 | }
2736 | # H-sigmoid
2737 | layer {
2738 |   name: "conv5-6/dwise-se-fc2/shift"
2739 |   type: "Power"
2740 |   bottom: "conv5-6/dwise-se-fc2"
2741 |   top: "conv5-6/dwise-se-fc2/shift"
2742 |   power_param {
2743 |     power: 1
2744 |     scale: 1
2745 |     shift: 3
2746 |   }
2747 | }
2748 | layer {
2749 |   name: "conv5-6/dwise-se-fc2/shift-relu"
2750 |   type: "ReLU6"
2751 |   bottom: "conv5-6/dwise-se-fc2/shift"
2752 |   top: "conv5-6/dwise-se-fc2/shift"
2753 | }
2754 | layer {
2755 |   name: "conv5-6/dwise-se-fc2/shift-div"
2756 |   type: "Power"
2757 |   bottom: "conv5-6/dwise-se-fc2/shift"
2758 |   top: "conv5-6/dwise-se-fc2/shift-div"
2759 |   power_param {
2760 |     power: 1
2761 |     scale: 0.1666666667
2762 |     shift: 0
2763 |   }
2764 | }
2765 | layer {
2766 |   name: "conv5-6/dwise/scale"
2767 |   type: "Scale"
2768 |   bottom: "conv5-6/dwise"
2769 |   bottom: "conv5-6/dwise-se-fc2/shift-div"
2770 |   top: "conv5-6/dwise/scale"
2771 |   scale_param{
2772 |     axis: 0
2773 |   }
2774 | }
2775 | # H-swish
2776 | layer {
2777 |   name: "conv5-6/dwise/shift"
2778 |   type: "Power"
2779 |   bottom: "conv5-6/dwise/scale"
2780 |   top: "conv5-6/dwise/shift"
2781 |   power_param {
2782 |     power: 1
2783 |     scale: 1
2784 |     shift: 3
2785 |   }
2786 | }
2787 | layer {
2788 |   name: "conv5-6/dwise/shift-relu"
2789 |   type: "ReLU6"
2790 |   bottom: "conv5-6/dwise/shift"
2791 |   top: "conv5-6/dwise/shift"
2792 | }
2793 | layer {
2794 |   name: "conv5-6/dwise/shift-div"
2795 |   type: "Power"
2796 |   bottom: "conv5-6/dwise/shift"
2797 |   top: "conv5-6/dwise/shift-div"
2798 |   power_param {
2799 |     power: 1
2800 |     scale: 0.1666666667
2801 |     shift: 0
2802 |   }
2803 | }
2804 | layer {
2805 |   name: "conv5-6/dwise/hswish"
2806 |   type: "Eltwise"
2807 |   bottom: "conv5-6/dwise/scale"
2808 |   bottom: "conv5-6/dwise/shift-div"
2809 |   top: "conv5-6/dwise/hswish"
2810 |   eltwise_param {
2811 |     operation: PROD
2812 |   }
2813 | }
2814 | 
2815 | layer {
2816 |   name: "conv5-6/linear"
2817 |   type: "Convolution"
2818 |   bottom: "conv5-6/dwise/hswish"
2819 |   top: "conv5-6/linear"
2820 |   param {
2821 |     lr_mult: 1
2822 |     decay_mult: 1
2823 |   }
2824 |   convolution_param {
2825 |     num_output: 112
2826 |     pad: 0
2827 |     bias_term: false
2828 |     kernel_size: 1
2829 |     stride: 1
2830 |     weight_filler {
2831 |       type: "msra"
2832 |     }
2833 |   }
2834 | }
2835 | layer {
2836 |   name: "conv5-6/linear-bn"
2837 |   type: "BatchNorm"
2838 |   bottom: "conv5-6/linear"
2839 |   top: "conv5-6/linear"
2840 | }
2841 | layer {
2842 |   name: "conv5-6/linear-bn-scale"
2843 |   type: "Scale"
2844 |   bottom: "conv5-6/linear"
2845 |   top: "conv5-6/linear"
2846 |   scale_param {
2847 |     bias_term: true
2848 |   }
2849 | }
2850 | 
2851 | layer {
2852 |   name: "block5-6"
2853 |   type: "Eltwise"
2854 |   bottom: "block5-5"
2855 |   bottom: "conv5-6/linear"
2856 |   top: "block5-6"
2857 | }
2858 | 
2859 | #################### stage 5-7 ####################
2860 | layer {
2861 |   name: "conv5-7/expand"
2862 |   type: "Convolution"
2863 |   bottom: "block5-6"
2864 |   top: "conv5-7/expand"
2865 |   param {
2866 |     lr_mult: 1
2867 |     decay_mult: 1
2868 |   }
2869 |   convolution_param {
2870 |     num_output: 672
2871 |     pad: 0
2872 |     bias_term: false
2873 |     kernel_size: 1
2874 |     stride: 1
2875 |     weight_filler {
2876 |       type: "msra"
2877 |     }
2878 |   }
2879 | }
2880 | layer {
2881 |   name: "conv5-7/expand-bn"
2882 |   type: "BatchNorm"
2883 |   bottom: "conv5-7/expand"
2884 |   top: "conv5-7/expand"
2885 | }
2886 | layer {
2887 |   name: "conv5-7/expand-bn-scale"
2888 |   type: "Scale"
2889 |   bottom: "conv5-7/expand"
2890 |   top: "conv5-7/expand"
2891 |   scale_param {
2892 |     bias_term: true
2893 |   }
2894 | }
2895 | # H-swish
2896 | layer {
2897 |   name: "conv5-7/expand/shift"
2898 |   type: "Power"
2899 |   bottom: "conv5-7/expand"
2900 |   top: "conv5-7/expand/shift"
2901 |   power_param {
2902 |     power: 1
2903 |     scale: 1
2904 |     shift: 3
2905 |   }
2906 | }
2907 | layer {
2908 |   name: "conv5-7/expand/shift-relu"
2909 |   type: "ReLU6"
2910 |   bottom: "conv5-7/expand/shift"
2911 |   top: "conv5-7/expand/shift"
2912 | }
2913 | layer {
2914 |   name: "conv5-7/expand/shift-div"
2915 |   type: "Power"
2916 |   bottom: "conv5-7/expand/shift"
2917 |   top: "conv5-7/expand/shift-div"
2918 |   power_param {
2919 |     power: 1
2920 |     scale: 0.1666666667
2921 |     shift: 0
2922 |   }
2923 | }
2924 | layer {
2925 |   name: "conv5-7/expand/hswish"
2926 |   type: "Eltwise"
2927 |   bottom: "conv5-7/expand"
2928 |   bottom: "conv5-7/expand/shift-div"
2929 |   top: "conv5-7/expand/hswish"
2930 |   eltwise_param {
2931 |     operation: PROD
2932 |   }
2933 | }
2934 | 
2935 | layer {
2936 |   name: "conv5-7/dwise"
2937 |   #type: "Convolution"
2938 |   type: "DepthwiseConvolution"
2939 |   bottom: "conv5-7/expand/hswish"
2940 |   top: "conv5-7/dwise"
2941 |   param {
2942 |     lr_mult: 1
2943 |     decay_mult: 1
2944 |   }
2945 |   convolution_param {
2946 |     num_output: 672
2947 |     group: 672
2948 |     pad: 2
2949 |     bias_term: false
2950 |     kernel_size: 5
2951 |     stride: 2
2952 |     weight_filler {
2953 |       type: "msra"
2954 |     }
2955 |     engine: CAFFE
2956 |   }
2957 | }
2958 | layer {
2959 |   name: "conv5-7/dwise-bn"
2960 |   type: "BatchNorm"
2961 |   bottom: "conv5-7/dwise"
2962 |   top: "conv5-7/dwise"
2963 | }
2964 | layer {
2965 |   name: "conv5-7/dwise-bn-scale"
2966 |   type: "Scale"
2967 |   bottom: "conv5-7/dwise"
2968 |   top: "conv5-7/dwise"
2969 |   scale_param {
2970 |     bias_term: true
2971 |   }
2972 | }
2973 | 
2974 | # se
2975 | layer {
2976 |   name: "conv5-7/dwise-se-pool"
2977 |   type: "Pooling"
2978 |   bottom: "conv5-7/dwise"
2979 |   top: "conv5-7/dwise-se-pool"
2980 |   pooling_param {
2981 |     pool: AVE
2982 |     global_pooling: true
2983 |   }
2984 | }
2985 | layer {
2986 |   name: "conv5-7/dwise-se-fc1"
2987 |   type: "InnerProduct"
2988 |   bottom: "conv5-7/dwise-se-pool"
2989 |   top:"conv5-7/dwise-se-fc1"
2990 |   param {
2991 |     lr_mult: 1
2992 |     decay_mult: 1
2993 |   }
2994 |   param {
2995 |     lr_mult: 2
2996 |     decay_mult: 0
2997 |   }
2998 |   inner_product_param {
2999 |     num_output: 168
3000 |     weight_filler {
3001 |       type: "msra"
3002 |     }
3003 |     bias_filler {
3004 |       type: "constant"
3005 |       value: 0
3006 |     }
3007 |   }
3008 | }
3009 | layer {
3010 |   name: "conv5-7/dwise-se-fc1-relu"
3011 |   type: "ReLU6"
3012 |   bottom: "conv5-7/dwise-se-fc1"
3013 |   top: "conv5-7/dwise-se-fc1"
3014 | }
3015 | layer {
3016 |   name: "conv5-7/dwise-se-fc2"
3017 |   type: "InnerProduct"
3018 |   bottom: "conv5-7/dwise-se-fc1"
3019 |   top:"conv5-7/dwise-se-fc2"
3020 |   param {
3021 |     lr_mult: 1
3022 |     decay_mult: 1
3023 |   }
3024 |   param {
3025 |     lr_mult: 2
3026 |     decay_mult: 0
3027 |   }
3028 |   inner_product_param {
3029 |     num_output: 672
3030 |     weight_filler {
3031 |       type: "msra"
3032 |     }
3033 |     bias_filler {
3034 |       type: "constant"
3035 |       value: 0
3036 |     }
3037 |   }
3038 | }
3039 | # H-sigmoid
3040 | layer {
3041 |   name: "conv5-7/dwise-se-fc2/shift"
3042 |   type: "Power"
3043 |   bottom: "conv5-7/dwise-se-fc2"
3044 |   top: "conv5-7/dwise-se-fc2/shift"
3045 |   power_param {
3046 |     power: 1
3047 |     scale: 1
3048 |     shift: 3
3049 |   }
3050 | }
3051 | layer {
3052 |   name: "conv5-7/dwise-se-fc2/shift-relu"
3053 |   type: "ReLU6"
3054 |   bottom: "conv5-7/dwise-se-fc2/shift"
3055 |   top: "conv5-7/dwise-se-fc2/shift"
3056 | }
3057 | layer {
3058 |   name: "conv5-7/dwise-se-fc2/shift-div"
3059 |   type: "Power"
3060 |   bottom: "conv5-7/dwise-se-fc2/shift"
3061 |   top: "conv5-7/dwise-se-fc2/shift-div"
3062 |   power_param {
3063 |     power: 1
3064 |     scale: 0.1666666667
3065 |     shift: 0
3066 |   }
3067 | }
3068 | layer {
3069 |   name: "conv5-7/dwise/scale"
3070 |   type: "Scale"
3071 |   bottom: "conv5-7/dwise"
3072 |   bottom: "conv5-7/dwise-se-fc2/shift-div"
3073 |   top: "conv5-7/dwise/scale"
3074 |   scale_param{
3075 |     axis: 0
3076 |   }
3077 | }
3078 | # H-swish
3079 | layer {
3080 |   name: "conv5-7/dwise/shift"
3081 |   type: "Power"
3082 |   bottom: "conv5-7/dwise/scale"
3083 |   top: "conv5-7/dwise/shift"
3084 |   power_param {
3085 |     power: 1
3086 |     scale: 1
3087 |     shift: 3
3088 |   }
3089 | }
3090 | layer {
3091 |   name: "conv5-7/dwise/shift-relu"
3092 |   type: "ReLU6"
3093 |   bottom: "conv5-7/dwise/shift"
3094 |   top: "conv5-7/dwise/shift"
3095 | }
3096 | layer {
3097 |   name: "conv5-7/dwise/shift-div"
3098 |   type: "Power"
3099 |   bottom: "conv5-7/dwise/shift"
3100 |   top: "conv5-7/dwise/shift-div"
3101 |   power_param {
3102 |     power: 1
3103 |     scale: 0.1666666667
3104 |     shift: 0
3105 |   }
3106 | }
3107 | layer {
3108 |   name: "conv5-7/dwise/hswish"
3109 |   type: "Eltwise"
3110 |   bottom: "conv5-7/dwise/scale"
3111 |   bottom: "conv5-7/dwise/shift-div"
3112 |   top: "conv5-7/dwise/hswish"
3113 |   eltwise_param {
3114 |     operation: PROD
3115 |   }
3116 | }
3117 | 
3118 | layer {
3119 |   name: "conv5-7/linear"
3120 |   type: "Convolution"
3121 |   bottom: "conv5-7/dwise/hswish"
3122 |   top: "conv5-7/linear"
3123 |   param {
3124 |     lr_mult: 1
3125 |     decay_mult: 1
3126 |   }
3127 |   convolution_param {
3128 |     num_output: 160
3129 |     pad: 0
3130 |     bias_term: false
3131 |     kernel_size: 1
3132 |     stride: 1
3133 |     weight_filler {
3134 |       type: "msra"
3135 |     }
3136 |   }
3137 | }
3138 | layer {
3139 |   name: "conv5-7/linear-bn"
3140 |   type: "BatchNorm"
3141 |   bottom: "conv5-7/linear"
3142 |   top: "conv5-7/linear"
3143 | }
3144 | layer {
3145 |   name: "conv5-7/linear-bn-scale"
3146 |   type: "Scale"
3147 |   bottom: "conv5-7/linear"
3148 |   top: "conv5-7/linear"
3149 |   scale_param {
3150 |     bias_term: true
3151 |   }
3152 | }
3153 | 
3154 | #################### stage 6-1 ####################
3155 | layer {
3156 |   name: "conv6-1/expand"
3157 |   type: "Convolution"
3158 |   bottom: "conv5-7/linear"
3159 |   top: "conv6-1/expand"
3160 |   param {
3161 |     lr_mult: 1
3162 |     decay_mult: 1
3163 |   }
3164 |   convolution_param {
3165 |     num_output: 960
3166 |     pad: 0
3167 |     bias_term: false
3168 |     kernel_size: 1
3169 |     stride: 1
3170 |     weight_filler {
3171 |       type: "msra"
3172 |     }
3173 |   }
3174 | }
3175 | layer {
3176 |   name: "conv6-1/expand-bn"
3177 |   type: "BatchNorm"
3178 |   bottom: "conv6-1/expand"
3179 |   top: "conv6-1/expand"
3180 | }
3181 | layer {
3182 |   name: "conv6-1/expand-bn-scale"
3183 |   type: "Scale"
3184 |   bottom: "conv6-1/expand"
3185 |   top: "conv6-1/expand"
3186 |   scale_param {
3187 |     bias_term: true
3188 |   }
3189 | }
3190 | # H-swish
3191 | layer {
3192 |   name: "conv6-1/expand/shift"
3193 |   type: "Power"
3194 |   bottom: "conv6-1/expand"
3195 |   top: "conv6-1/expand/shift"
3196 |   power_param {
3197 |     power: 1
3198 |     scale: 1
3199 |     shift: 3
3200 |   }
3201 | }
3202 | layer {
3203 |   name: "conv6-1/expand/shift-relu"
3204 |   type: "ReLU6"
3205 |   bottom: "conv6-1/expand/shift"
3206 |   top: "conv6-1/expand/shift"
3207 | }
3208 | layer {
3209 |   name: "conv6-1/expand/shift-div"
3210 |   type: "Power"
3211 |   bottom: "conv6-1/expand/shift"
3212 |   top: "conv6-1/expand/shift-div"
3213 |   power_param {
3214 |     power: 1
3215 |     scale: 0.1666666667
3216 |     shift: 0
3217 |   }
3218 | }
3219 | layer {
3220 |   name: "conv6-1/expand/hswish"
3221 |   type: "Eltwise"
3222 |   bottom: "conv6-1/expand"
3223 |   bottom: "conv6-1/expand/shift-div"
3224 |   top: "conv6-1/expand/hswish"
3225 |   eltwise_param {
3226 |     operation: PROD
3227 |   }
3228 | }
3229 | 
3230 | layer {
3231 |   name: "conv6-1/dwise"
3232 |   #type: "Convolution"
3233 |   type: "DepthwiseConvolution"
3234 |   bottom: "conv6-1/expand/hswish"
3235 |   top: "conv6-1/dwise"
3236 |   param {
3237 |     lr_mult: 1
3238 |     decay_mult: 1
3239 |   }
3240 |   convolution_param {
3241 |     num_output: 960
3242 |     group: 960
3243 |     pad: 2
3244 |     bias_term: false
3245 |     kernel_size: 5
3246 |     stride: 1
3247 |     weight_filler {
3248 |       type: "msra"
3249 |     }
3250 |     engine: CAFFE
3251 |   }
3252 | }
3253 | layer {
3254 |   name: "conv6-1/dwise-bn"
3255 |   type: "BatchNorm"
3256 |   bottom: "conv6-1/dwise"
3257 |   top: "conv6-1/dwise"
3258 | }
3259 | layer {
3260 |   name: "conv6-1/dwise-bn-scale"
3261 |   type: "Scale"
3262 |   bottom: "conv6-1/dwise"
3263 |   top: "conv6-1/dwise"
3264 |   scale_param {
3265 |     bias_term: true
3266 |   }
3267 | }
3268 | 
3269 | # se
3270 | layer {
3271 |   name: "conv6-1/dwise-se-pool"
3272 |   type: "Pooling"
3273 |   bottom: "conv6-1/dwise"
3274 |   top: "conv6-1/dwise-se-pool"
3275 |   pooling_param {
3276 |     pool: AVE
3277 |     global_pooling: true
3278 |   }
3279 | }
3280 | layer {
3281 |   name: "conv6-1/dwise-se-fc1"
3282 |   type: "InnerProduct"
3283 |   bottom: "conv6-1/dwise-se-pool"
3284 |   top:"conv6-1/dwise-se-fc1"
3285 |   param {
3286 |     lr_mult: 1
3287 |     decay_mult: 1
3288 |   }
3289 |   param {
3290 |     lr_mult: 2
3291 |     decay_mult: 0
3292 |   }
3293 |   inner_product_param {
3294 |     num_output: 240
3295 |     weight_filler {
3296 |       type: "msra"
3297 |     }
3298 |     bias_filler {
3299 |       type: "constant"
3300 |       value: 0
3301 |     }
3302 |   }
3303 | }
3304 | layer {
3305 |   name: "conv6-1/dwise-se-fc1-relu"
3306 |   type: "ReLU6"
3307 |   bottom: "conv6-1/dwise-se-fc1"
3308 |   top: "conv6-1/dwise-se-fc1"
3309 | }
3310 | layer {
3311 |   name: "conv6-1/dwise-se-fc2"
3312 |   type: "InnerProduct"
3313 |   bottom: "conv6-1/dwise-se-fc1"
3314 |   top:"conv6-1/dwise-se-fc2"
3315 |   param {
3316 |     lr_mult: 1
3317 |     decay_mult: 1
3318 |   }
3319 |   param {
3320 |     lr_mult: 2
3321 |     decay_mult: 0
3322 |   }
3323 |   inner_product_param {
3324 |     num_output: 960
3325 |     weight_filler {
3326 |       type: "msra"
3327 |     }
3328 |     bias_filler {
3329 |       type: "constant"
3330 |       value: 0
3331 |     }
3332 |   }
3333 | }
3334 | # H-sigmoid
3335 | layer {
3336 |   name: "conv6-1/dwise-se-fc2/shift"
3337 |   type: "Power"
3338 |   bottom: "conv6-1/dwise-se-fc2"
3339 |   top: "conv6-1/dwise-se-fc2/shift"
3340 |   power_param {
3341 |     power: 1
3342 |     scale: 1
3343 |     shift: 3
3344 |   }
3345 | }
3346 | layer {
3347 |   name: "conv6-1/dwise-se-fc2/shift-relu"
3348 |   type: "ReLU6"
3349 |   bottom: "conv6-1/dwise-se-fc2/shift"
3350 |   top: "conv6-1/dwise-se-fc2/shift"
3351 | }
3352 | layer {
3353 |   name: "conv6-1/dwise-se-fc2/shift-div"
3354 |   type: "Power"
3355 |   bottom: "conv6-1/dwise-se-fc2/shift"
3356 |   top: "conv6-1/dwise-se-fc2/shift-div"
3357 |   power_param {
3358 |     power: 1
3359 |     scale: 0.1666666667
3360 |     shift: 0
3361 |   }
3362 | }
3363 | layer {
3364 |   name: "conv6-1/dwise/scale"
3365 |   type: "Scale"
3366 |   bottom: "conv6-1/dwise"
3367 |   bottom: "conv6-1/dwise-se-fc2/shift-div"
3368 |   top: "conv6-1/dwise/scale"
3369 |   scale_param{
3370 |     axis: 0
3371 |   }
3372 | }
3373 | # H-swish
3374 | layer {
3375 |   name: "conv6-1/dwise/shift"
3376 |   type: "Power"
3377 |   bottom: "conv6-1/dwise/scale"
3378 |   top: "conv6-1/dwise/shift"
3379 |   power_param {
3380 |     power: 1
3381 |     scale: 1
3382 |     shift: 3
3383 |   }
3384 | }
3385 | layer {
3386 |   name: "conv6-1/dwise/shift-relu"
3387 |   type: "ReLU6"
3388 |   bottom: "conv6-1/dwise/shift"
3389 |   top: "conv6-1/dwise/shift"
3390 | }
3391 | layer {
3392 |   name: "conv6-1/dwise/shift-div"
3393 |   type: "Power"
3394 |   bottom: "conv6-1/dwise/shift"
3395 |   top: "conv6-1/dwise/shift-div"
3396 |   power_param {
3397 |     power: 1
3398 |     scale: 0.1666666667
3399 |     shift: 0
3400 |   }
3401 | }
3402 | layer {
3403 |   name: "conv6-1/dwise/hswish"
3404 |   type: "Eltwise"
3405 |   bottom: "conv6-1/dwise/scale"
3406 |   bottom: "conv6-1/dwise/shift-div"
3407 |   top: "conv6-1/dwise/hswish"
3408 |   eltwise_param {
3409 |     operation: PROD
3410 |   }
3411 | }
3412 | 
3413 | layer {
3414 |   name: "conv6-1/linear"
3415 |   type: "Convolution"
3416 |   bottom: "conv6-1/dwise/hswish"
3417 |   top: "conv6-1/linear"
3418 |   param {
3419 |     lr_mult: 1
3420 |     decay_mult: 1
3421 |   }
3422 |   convolution_param {
3423 |     num_output: 160
3424 |     pad: 0
3425 |     bias_term: false
3426 |     kernel_size: 1
3427 |     stride: 1
3428 |     weight_filler {
3429 |       type: "msra"
3430 |     }
3431 |   }
3432 | }
3433 | layer {
3434 |   name: "conv6-1/linear-bn"
3435 |   type: "BatchNorm"
3436 |   bottom: "conv6-1/linear"
3437 |   top: "conv6-1/linear"
3438 | }
3439 | layer {
3440 |   name: "conv6-1/linear-bn-scale"
3441 |   type: "Scale"
3442 |   bottom: "conv6-1/linear"
3443 |   top: "conv6-1/linear"
3444 |   scale_param {
3445 |     bias_term: true
3446 |   }
3447 | }
3448 | 
3449 | layer {
3450 |   name: "block6-1"
3451 |   type: "Eltwise"
3452 |   bottom: "conv5-7/linear"
3453 |   bottom: "conv6-1/linear"
3454 |   top: "block6-1"
3455 | }
3456 | 
3457 | #################### stage 6-2 ####################
3458 | layer {
3459 |   name: "conv6-2"
3460 |   type: "Convolution"
3461 |   bottom: "block6-1"
3462 |   top: "conv6-2"
3463 |   param {
3464 |     lr_mult: 1
3465 |     decay_mult: 1
3466 |   }
3467 |   convolution_param {
3468 |     num_output: 960
3469 |     pad: 0
3470 |     bias_term: false
3471 |     kernel_size: 1
3472 |     stride: 1
3473 |     weight_filler {
3474 |       type: "msra"
3475 |     }
3476 |   }
3477 | }
3478 | layer {
3479 |   name: "conv6-2-bn"
3480 |   type: "BatchNorm"
3481 |   bottom: "conv6-2"
3482 |   top: "conv6-2"
3483 | }
3484 | layer {
3485 |   name: "conv6-2-bn-scale"
3486 |   type: "Scale"
3487 |   bottom: "conv6-2"
3488 |   top: "conv6-2"
3489 |   scale_param {
3490 |     bias_term: true
3491 |   }
3492 | }
3493 | # H-swish
3494 | layer {
3495 |   name: "conv6-2/shift"
3496 |   type: "Power"
3497 |   bottom: "conv6-2"
3498 |   top: "conv6-2/shift"
3499 |   power_param {
3500 |     power: 1
3501 |     scale: 1
3502 |     shift: 3
3503 |   }
3504 | }
3505 | layer {
3506 |   name: "conv6-2/shift-relu"
3507 |   type: "ReLU6"
3508 |   bottom: "conv6-2/shift"
3509 |   top: "conv6-2/shift"
3510 | }
3511 | layer {
3512 |   name: "conv6-2/shift-div"
3513 |   type: "Power"
3514 |   bottom: "conv6-2/shift"
3515 |   top: "conv6-2/shift-div"
3516 |   power_param {
3517 |     power: 1
3518 |     scale: 0.1666666667
3519 |     shift: 0
3520 |   }
3521 | }
3522 | layer {
3523 |   name: "conv6-2/hswish"
3524 |   type: "Eltwise"
3525 |   bottom: "conv6-2"
3526 |   bottom: "conv6-2/shift-div"
3527 |   top: "conv6-2/hswish"
3528 |   eltwise_param {
3529 |     operation: PROD
3530 |   }
3531 | }
3532 | 
3533 | #################### stage 7-1 ####################
3534 | layer {
3535 |   name: "pool7-1"
3536 |   type: "Pooling"
3537 |   bottom: "conv6-2/hswish"
3538 |   top: "pool7-1"
3539 |   pooling_param {
3540 |     pool: AVE
3541 |     global_pooling: true
3542 |   }
3543 | }
3544 | # H-swish
3545 | layer {
3546 |   name: "pool7-1/shift"
3547 |   type: "Power"
3548 |   bottom: "pool7-1"
3549 |   top: "pool7-1/shift"
3550 |   power_param {
3551 |     power: 1
3552 |     scale: 1
3553 |     shift: 3
3554 |   }
3555 | }
3556 | layer {
3557 |   name: "pool7-1/shift-relu"
3558 |   type: "ReLU6"
3559 |   bottom: "pool7-1/shift"
3560 |   top: "pool7-1/shift"
3561 | }
3562 | layer {
3563 |   name: "pool7-1/shift-div"
3564 |   type: "Power"
3565 |   bottom: "pool7-1/shift"
3566 |   top: "pool7-1/shift-div"
3567 |   power_param {
3568 |     power: 1
3569 |     scale: 0.1666666667
3570 |     shift: 0
3571 |   }
3572 | }
3573 | layer {
3574 |   name: "pool7-1/hswish"
3575 |   type: "Eltwise"
3576 |   bottom: "pool7-1"
3577 |   bottom: "pool7-1/shift-div"
3578 |   top: "pool7-1/hswish"
3579 |   eltwise_param {
3580 |     operation: PROD
3581 |   }
3582 | }
3583 | 
3584 | #################### stage 7-2 ####################
3585 | layer {
3586 |   name: "conv7-2"
3587 |   type: "Convolution"
3588 |   bottom: "pool7-1/hswish"
3589 |   top: "conv7-2"
3590 |   param {
3591 |     lr_mult: 1
3592 |     decay_mult: 1
3593 |   }
3594 |   param {
3595 |     lr_mult: 2
3596 |     decay_mult: 0
3597 |   }
3598 |   convolution_param {
3599 |     num_output: 1280
3600 |     pad: 0
3601 |     kernel_size: 1
3602 |     stride: 1
3603 |     weight_filler {
3604 |       type: "msra"
3605 |     }
3606 |     bias_filler {
3607 |       type: "constant"
3608 |       value: 0
3609 |     }
3610 |   }
3611 | }
3612 | # H-swish
3613 | layer {
3614 |   name: "conv7-2/shift"
3615 |   type: "Power"
3616 |   bottom: "conv7-2"
3617 |   top: "conv7-2/shift"
3618 |   power_param {
3619 |     power: 1
3620 |     scale: 1
3621 |     shift: 3
3622 |   }
3623 | }
3624 | layer {
3625 |   name: "conv7-2/shift-relu"
3626 |   type: "ReLU6"
3627 |   bottom: "conv7-2/shift"
3628 |   top: "conv7-2/shift"
3629 | }
3630 | layer {
3631 |   name: "conv7-2/shift-div"
3632 |   type: "Power"
3633 |   bottom: "conv7-2/shift"
3634 |   top: "conv7-2/shift-div"
3635 |   power_param {
3636 |     power: 1
3637 |     scale: 0.1666666667
3638 |     shift: 0
3639 |   }
3640 | }
3641 | layer {
3642 |   name: "conv7-2/hswish"
3643 |   type: "Eltwise"
3644 |   bottom: "conv7-2"
3645 |   bottom: "conv7-2/shift-div"
3646 |   top: "conv7-2/hswish"
3647 |   eltwise_param {
3648 |     operation: PROD
3649 |   }
3650 | }
3651 | 
3652 | #################### output ####################
3653 | layer {
3654 |   name: "conv7-3"
3655 |   type: "Convolution"
3656 |   bottom: "conv7-2/hswish"
3657 |   top: "conv7-3"
3658 |   param {
3659 |     lr_mult: 1
3660 |     decay_mult: 1
3661 |   }
3662 |   param {
3663 |     lr_mult: 2
3664 |     decay_mult: 0
3665 |   }
3666 |   convolution_param {
3667 |     num_output: 1000
3668 |     pad: 0
3669 |     kernel_size: 1
3670 |     stride: 1
3671 |     weight_filler {
3672 |       type: "msra"
3673 |     }
3674 |     bias_filler {
3675 |       type: "constant"
3676 |       value: 0
3677 |     }
3678 |   }
3679 | }
3680 | 
3681 | layer {
3682 |   name: "prob"
3683 |   type: "Softmax"
3684 |   bottom: "conv7-3"
3685 |   top: "prob"
3686 | }
3687 | 


--------------------------------------------------------------------------------
/mobilenet_v3_small_1.0.prototxt:
--------------------------------------------------------------------------------
   1 | name: "mobilenet v3"
   2 | input: "data"
   3 | input_dim: 1
   4 | input_dim: 3
   5 | input_dim: 224
   6 | input_dim: 224
   7 | #################### stage 1 ####################
   8 | layer {
   9 |   name: "conv1"
  10 |   type: "Convolution"
  11 |   bottom: "data"
  12 |   top: "conv1"
  13 |   param {
  14 |     lr_mult: 1
  15 |     decay_mult: 1
  16 |   }
  17 |   convolution_param {
  18 |     num_output: 16
  19 |     pad: 1
  20 |     bias_term: false
  21 |     kernel_size: 3
  22 |     stride: 2
  23 |     weight_filler {
  24 |       type: "msra"
  25 |     }
  26 |   }
  27 | }
  28 | layer {
  29 |   name: "conv1-bn"
  30 |   type: "BatchNorm"
  31 |   bottom: "conv1"
  32 |   top: "conv1"
  33 | }
  34 | layer {
  35 |   name: "conv1-bn-scale"
  36 |   type: "Scale"
  37 |   bottom: "conv1"
  38 |   top: "conv1"
  39 |   scale_param {
  40 |     bias_term: true
  41 |   }
  42 | }
  43 | # H-swish
  44 | layer {
  45 |   name: "conv1/shift"
  46 |   type: "Power"
  47 |   bottom: "conv1"
  48 |   top: "conv1/shift"
  49 |   power_param {
  50 |     power: 1
  51 |     scale: 1
  52 |     shift: 3
  53 |   }
  54 | }
  55 | layer {
  56 |   name: "conv1/shift-relu"
  57 |   type: "ReLU6"
  58 |   bottom: "conv1/shift"
  59 |   top: "conv1/shift"
  60 | }
  61 | layer {
  62 |   name: "conv1/shift-div"
  63 |   type: "Power"
  64 |   bottom: "conv1/shift"
  65 |   top: "conv1/shift-div"
  66 |   power_param {
  67 |     power: 1
  68 |     scale: 0.1666666667
  69 |     shift: 0
  70 |   }
  71 | }
  72 | layer {
  73 |   name: "conv1/hswish"
  74 |   type: "Eltwise"
  75 |   bottom: "conv1"
  76 |   bottom: "conv1/shift-div"
  77 |   top: "conv1/hswish"
  78 |   eltwise_param {
  79 |     operation: PROD
  80 |   }
  81 | }
  82 | 
  83 | #################### stage 2-1 ####################
  84 | layer {
  85 |   name: "conv2-1/expand"
  86 |   type: "Convolution"
  87 |   bottom: "conv1/hswish"
  88 |   top: "conv2-1/expand"
  89 |   param {
  90 |     lr_mult: 1
  91 |     decay_mult: 1
  92 |   }
  93 |   convolution_param {
  94 |     num_output: 16
  95 |     pad: 0
  96 |     bias_term: false
  97 |     kernel_size: 1
  98 |     stride: 1
  99 |     weight_filler {
 100 |       type: "msra"
 101 |     }
 102 |   }
 103 | }
 104 | layer {
 105 |   name: "conv2-1/expand-bn"
 106 |   type: "BatchNorm"
 107 |   bottom: "conv2-1/expand"
 108 |   top: "conv2-1/expand"
 109 | }
 110 | layer {
 111 |   name: "conv2-1/expand-bn-scale"
 112 |   type: "Scale"
 113 |   bottom: "conv2-1/expand"
 114 |   top: "conv2-1/expand"
 115 |   scale_param {
 116 |     bias_term: true
 117 |   }
 118 | }
 119 | layer {
 120 |   name: "conv2-1/expand-relu"
 121 |   type: "ReLU6"
 122 |   bottom: "conv2-1/expand"
 123 |   top: "conv2-1/expand"
 124 | }
 125 | 
 126 | layer {
 127 |   name: "conv2-1/dwise"
 128 |   #type: "Convolution"
 129 |   type: "DepthwiseConvolution"
 130 |   bottom: "conv2-1/expand"
 131 |   top: "conv2-1/dwise"
 132 |   param {
 133 |     lr_mult: 1
 134 |     decay_mult: 1
 135 |   }
 136 |   convolution_param {
 137 |     num_output: 16
 138 |     group: 16
 139 |     pad: 1
 140 |     bias_term: false
 141 |     kernel_size: 3
 142 |     stride: 2
 143 |     weight_filler {
 144 |       type: "msra"
 145 |     }
 146 |     engine: CAFFE
 147 |   }
 148 | }
 149 | layer {
 150 |   name: "conv2-1/dwise-bn"
 151 |   type: "BatchNorm"
 152 |   bottom: "conv2-1/dwise"
 153 |   top: "conv2-1/dwise"
 154 | }
 155 | layer {
 156 |   name: "conv2-1/dwise-bn-scale"
 157 |   type: "Scale"
 158 |   bottom: "conv2-1/dwise"
 159 |   top: "conv2-1/dwise"
 160 |   scale_param {
 161 |     bias_term: true
 162 |   }
 163 | }
 164 | 
 165 | # se
 166 | layer {
 167 |   name: "conv2-1/dwise-se-pool"
 168 |   type: "Pooling"
 169 |   bottom: "conv2-1/dwise"
 170 |   top: "conv2-1/dwise-se-pool"
 171 |   pooling_param {
 172 |     pool: AVE
 173 |     global_pooling: true
 174 |   }
 175 | }
 176 | layer {
 177 |   name: "conv2-1/dwise-se-fc1"
 178 |   type: "InnerProduct"
 179 |   bottom: "conv2-1/dwise-se-pool"
 180 |   top:"conv2-1/dwise-se-fc1"
 181 |   param {
 182 |     lr_mult: 1
 183 |     decay_mult: 1
 184 |   }
 185 |   param {
 186 |     lr_mult: 2
 187 |     decay_mult: 0
 188 |   }
 189 |   inner_product_param {
 190 |     num_output: 4
 191 |     weight_filler {
 192 |       type: "msra"
 193 |     }
 194 |     bias_filler {
 195 |       type: "constant"
 196 |       value: 0
 197 |     }
 198 |   }
 199 | }
 200 | layer {
 201 |   name: "conv2-1/dwise-se-fc1-relu"
 202 |   type: "ReLU6"
 203 |   bottom: "conv2-1/dwise-se-fc1"
 204 |   top: "conv2-1/dwise-se-fc1"
 205 | }
 206 | layer {
 207 |   name: "conv2-1/dwise-se-fc2"
 208 |   type: "InnerProduct"
 209 |   bottom: "conv2-1/dwise-se-fc1"
 210 |   top:"conv2-1/dwise-se-fc2"
 211 |   param {
 212 |     lr_mult: 1
 213 |     decay_mult: 1
 214 |   }
 215 |   param {
 216 |     lr_mult: 2
 217 |     decay_mult: 0
 218 |   }
 219 |   inner_product_param {
 220 |     num_output: 16
 221 |     weight_filler {
 222 |       type: "msra"
 223 |     }
 224 |     bias_filler {
 225 |       type: "constant"
 226 |       value: 0
 227 |     }
 228 |   }
 229 | }
 230 | # H-sigmoid
 231 | layer {
 232 |   name: "conv2-1/dwise-se-fc2/shift"
 233 |   type: "Power"
 234 |   bottom: "conv2-1/dwise-se-fc2"
 235 |   top: "conv2-1/dwise-se-fc2/shift"
 236 |   power_param {
 237 |     power: 1
 238 |     scale: 1
 239 |     shift: 3
 240 |   }
 241 | }
 242 | layer {
 243 |   name: "conv2-1/dwise-se-fc2/shift-relu"
 244 |   type: "ReLU6"
 245 |   bottom: "conv2-1/dwise-se-fc2/shift"
 246 |   top: "conv2-1/dwise-se-fc2/shift"
 247 | }
 248 | layer {
 249 |   name: "conv2-1/dwise-se-fc2/shift-div"
 250 |   type: "Power"
 251 |   bottom: "conv2-1/dwise-se-fc2/shift"
 252 |   top: "conv2-1/dwise-se-fc2/shift-div"
 253 |   power_param {
 254 |     power: 1
 255 |     scale: 0.1666666667
 256 |     shift: 0
 257 |   }
 258 | }
 259 | layer {
 260 |   name: "conv2-1/dwise/scale"
 261 |   type: "Scale"
 262 |   bottom: "conv2-1/dwise"
 263 |   bottom: "conv2-1/dwise-se-fc2/shift-div"
 264 |   top: "conv2-1/dwise/scale"
 265 |   scale_param{
 266 |     axis: 0
 267 |   }
 268 | }
 269 | layer {
 270 |   name: "conv2-1/dwise/scale-relu"
 271 |   type: "ReLU6"
 272 |   bottom: "conv2-1/dwise/scale"
 273 |   top: "conv2-1/dwise/scale"
 274 | }
 275 | 
 276 | layer {
 277 |   name: "conv2-1/linear"
 278 |   type: "Convolution"
 279 |   bottom: "conv2-1/dwise/scale"
 280 |   top: "conv2-1/linear"
 281 |   param {
 282 |     lr_mult: 1
 283 |     decay_mult: 1
 284 |   }
 285 |   convolution_param {
 286 |     num_output: 16
 287 |     pad: 0
 288 |     bias_term: false
 289 |     kernel_size: 1
 290 |     stride: 1
 291 |     weight_filler {
 292 |       type: "msra"
 293 |     }
 294 |   }
 295 | }
 296 | layer {
 297 |   name: "conv2-1/linear-bn"
 298 |   type: "BatchNorm"
 299 |   bottom: "conv2-1/linear"
 300 |   top: "conv2-1/linear"
 301 | }
 302 | layer {
 303 |   name: "conv2-1/linear-bn-scale"
 304 |   type: "Scale"
 305 |   bottom: "conv2-1/linear"
 306 |   top: "conv2-1/linear"
 307 |   scale_param {
 308 |     bias_term: true
 309 |   }
 310 | }
 311 | 
 312 | #################### stage 3-1 ####################
 313 | layer {
 314 |   name: "conv3-1/expand"
 315 |   type: "Convolution"
 316 |   bottom: "conv2-1/linear"
 317 |   top: "conv3-1/expand"
 318 |   param {
 319 |     lr_mult: 1
 320 |     decay_mult: 1
 321 |   }
 322 |   convolution_param {
 323 |     num_output: 72
 324 |     pad: 0
 325 |     bias_term: false
 326 |     kernel_size: 1
 327 |     stride: 1
 328 |     weight_filler {
 329 |       type: "msra"
 330 |     }
 331 |   }
 332 | }
 333 | layer {
 334 |   name: "conv3-1/expand-bn"
 335 |   type: "BatchNorm"
 336 |   bottom: "conv3-1/expand"
 337 |   top: "conv3-1/expand"
 338 | }
 339 | layer {
 340 |   name: "conv3-1/expand-bn-scale"
 341 |   type: "Scale"
 342 |   bottom: "conv3-1/expand"
 343 |   top: "conv3-1/expand"
 344 |   scale_param {
 345 |     bias_term: true
 346 |   }
 347 | }
 348 | layer {
 349 |   name: "conv3-1/expand-relu"
 350 |   type: "ReLU6"
 351 |   bottom: "conv3-1/expand"
 352 |   top: "conv3-1/expand"
 353 | }
 354 | 
 355 | layer {
 356 |   name: "conv3-1/dwise"
 357 |   #type: "Convolution"
 358 |   type: "DepthwiseConvolution"
 359 |   bottom: "conv3-1/expand"
 360 |   top: "conv3-1/dwise"
 361 |   param {
 362 |     lr_mult: 1
 363 |     decay_mult: 1
 364 |   }
 365 |   convolution_param {
 366 |     num_output: 72
 367 |     group: 72
 368 |     pad: 1
 369 |     bias_term: false
 370 |     kernel_size: 3
 371 |     stride: 2
 372 |     weight_filler {
 373 |       type: "msra"
 374 |     }
 375 |     engine: CAFFE
 376 |   }
 377 | }
 378 | layer {
 379 |   name: "conv3-1/dwise-bn"
 380 |   type: "BatchNorm"
 381 |   bottom: "conv3-1/dwise"
 382 |   top: "conv3-1/dwise"
 383 | }
 384 | layer {
 385 |   name: "conv3-1/dwise-bn-scale"
 386 |   type: "Scale"
 387 |   bottom: "conv3-1/dwise"
 388 |   top: "conv3-1/dwise"
 389 |   scale_param {
 390 |     bias_term: true
 391 |   }
 392 | }
 393 | layer {
 394 |   name: "conv3-1/dwise-relu"
 395 |   type: "ReLU6"
 396 |   bottom: "conv3-1/dwise"
 397 |   top: "conv3-1/dwise"
 398 | }
 399 | 
 400 | layer {
 401 |   name: "conv3-1/linear"
 402 |   type: "Convolution"
 403 |   bottom: "conv3-1/dwise"
 404 |   top: "conv3-1/linear"
 405 |   param {
 406 |     lr_mult: 1
 407 |     decay_mult: 1
 408 |   }
 409 |   convolution_param {
 410 |     num_output: 24
 411 |     pad: 0
 412 |     bias_term: false
 413 |     kernel_size: 1
 414 |     stride: 1
 415 |     weight_filler {
 416 |       type: "msra"
 417 |     }
 418 |   }
 419 | }
 420 | layer {
 421 |   name: "conv3-1/linear-bn"
 422 |   type: "BatchNorm"
 423 |   bottom: "conv3-1/linear"
 424 |   top: "conv3-1/linear"
 425 | }
 426 | layer {
 427 |   name: "conv3-1/linear-bn-scale"
 428 |   type: "Scale"
 429 |   bottom: "conv3-1/linear"
 430 |   top: "conv3-1/linear"
 431 |   scale_param {
 432 |     bias_term: true
 433 |   }
 434 | }
 435 | 
 436 | #################### stage 4-1 ####################
 437 | layer {
 438 |   name: "conv4-1/expand"
 439 |   type: "Convolution"
 440 |   bottom: "conv3-1/linear"
 441 |   top: "conv4-1/expand"
 442 |   param {
 443 |     lr_mult: 1
 444 |     decay_mult: 1
 445 |   }
 446 |   convolution_param {
 447 |     num_output: 88
 448 |     pad: 0
 449 |     bias_term: false
 450 |     kernel_size: 1
 451 |     stride: 1
 452 |     weight_filler {
 453 |       type: "msra"
 454 |     }
 455 |   }
 456 | }
 457 | layer {
 458 |   name: "conv4-1/expand-bn"
 459 |   type: "BatchNorm"
 460 |   bottom: "conv4-1/expand"
 461 |   top: "conv4-1/expand"
 462 | }
 463 | layer {
 464 |   name: "conv4-1/expand-bn-scale"
 465 |   type: "Scale"
 466 |   bottom: "conv4-1/expand"
 467 |   top: "conv4-1/expand"
 468 |   scale_param {
 469 |     bias_term: true
 470 |   }
 471 | }
 472 | layer {
 473 |   name: "conv4-1/expand-relu"
 474 |   type: "ReLU6"
 475 |   bottom: "conv4-1/expand"
 476 |   top: "conv4-1/expand"
 477 | }
 478 | 
 479 | layer {
 480 |   name: "conv4-1/dwise"
 481 |   #type: "Convolution"
 482 |   type: "DepthwiseConvolution"
 483 |   bottom: "conv4-1/expand"
 484 |   top: "conv4-1/dwise"
 485 |   param {
 486 |     lr_mult: 1
 487 |     decay_mult: 1
 488 |   }
 489 |   convolution_param {
 490 |     num_output: 88
 491 |     group: 88
 492 |     pad: 1
 493 |     bias_term: false
 494 |     kernel_size: 3
 495 |     stride: 1
 496 |     weight_filler {
 497 |       type: "msra"
 498 |     }
 499 |     engine: CAFFE
 500 |   }
 501 | }
 502 | layer {
 503 |   name: "conv4-1/dwise-bn"
 504 |   type: "BatchNorm"
 505 |   bottom: "conv4-1/dwise"
 506 |   top: "conv4-1/dwise"
 507 | }
 508 | layer {
 509 |   name: "conv4-1/dwise-bn-scale"
 510 |   type: "Scale"
 511 |   bottom: "conv4-1/dwise"
 512 |   top: "conv4-1/dwise"
 513 |   scale_param {
 514 |     bias_term: true
 515 |   }
 516 | }
 517 | layer {
 518 |   name: "conv4-1/dwise-relu"
 519 |   type: "ReLU6"
 520 |   bottom: "conv4-1/dwise"
 521 |   top: "conv4-1/dwise"
 522 | }
 523 | 
 524 | layer {
 525 |   name: "conv4-1/linear"
 526 |   type: "Convolution"
 527 |   bottom: "conv4-1/dwise"
 528 |   top: "conv4-1/linear"
 529 |   param {
 530 |     lr_mult: 1
 531 |     decay_mult: 1
 532 |   }
 533 |   convolution_param {
 534 |     num_output: 24
 535 |     pad: 0
 536 |     bias_term: false
 537 |     kernel_size: 1
 538 |     stride: 1
 539 |     weight_filler {
 540 |       type: "msra"
 541 |     }
 542 |   }
 543 | }
 544 | layer {
 545 |   name: "conv4-1/linear-bn"
 546 |   type: "BatchNorm"
 547 |   bottom: "conv4-1/linear"
 548 |   top: "conv4-1/linear"
 549 | }
 550 | layer {
 551 |   name: "conv4-1/linear-bn-scale"
 552 |   type: "Scale"
 553 |   bottom: "conv4-1/linear"
 554 |   top: "conv4-1/linear"
 555 |   scale_param {
 556 |     bias_term: true
 557 |   }
 558 | }
 559 | 
 560 | layer {
 561 |   name: "block4-1"
 562 |   type: "Eltwise"
 563 |   bottom: "conv3-1/linear"
 564 |   bottom: "conv4-1/linear"
 565 |   top: "block4-1"
 566 | }
 567 | 
 568 | #################### stage 4-2 ####################
 569 | layer {
 570 |   name: "conv4-2/expand"
 571 |   type: "Convolution"
 572 |   bottom: "block4-1"
 573 |   top: "conv4-2/expand"
 574 |   param {
 575 |     lr_mult: 1
 576 |     decay_mult: 1
 577 |   }
 578 |   convolution_param {
 579 |     num_output: 96
 580 |     pad: 0
 581 |     bias_term: false
 582 |     kernel_size: 1
 583 |     stride: 1
 584 |     weight_filler {
 585 |       type: "msra"
 586 |     }
 587 |   }
 588 | }
 589 | layer {
 590 |   name: "conv4-2/expand-bn"
 591 |   type: "BatchNorm"
 592 |   bottom: "conv4-2/expand"
 593 |   top: "conv4-2/expand"
 594 | }
 595 | layer {
 596 |   name: "conv4-2/expand-bn-scale"
 597 |   type: "Scale"
 598 |   bottom: "conv4-2/expand"
 599 |   top: "conv4-2/expand"
 600 |   scale_param {
 601 |     bias_term: true
 602 |   }
 603 | }
 604 | # H-swish
 605 | layer {
 606 |   name: "conv4-2/expand/shift"
 607 |   type: "Power"
 608 |   bottom: "conv4-2/expand"
 609 |   top: "conv4-2/expand/shift"
 610 |   power_param {
 611 |     power: 1
 612 |     scale: 1
 613 |     shift: 3
 614 |   }
 615 | }
 616 | layer {
 617 |   name: "conv4-2/expand/shift-relu"
 618 |   type: "ReLU6"
 619 |   bottom: "conv4-2/expand/shift"
 620 |   top: "conv4-2/expand/shift"
 621 | }
 622 | layer {
 623 |   name: "conv4-2/expand/shift-div"
 624 |   type: "Power"
 625 |   bottom: "conv4-2/expand/shift"
 626 |   top: "conv4-2/expand/shift-div"
 627 |   power_param {
 628 |     power: 1
 629 |     scale: 0.1666666667
 630 |     shift: 0
 631 |   }
 632 | }
 633 | layer {
 634 |   name: "conv4-2/expand/hswish"
 635 |   type: "Eltwise"
 636 |   bottom: "conv4-2/expand"
 637 |   bottom: "conv4-2/expand/shift-div"
 638 |   top: "conv4-2/expand/hswish"
 639 |   eltwise_param {
 640 |     operation: PROD
 641 |   }
 642 | }
 643 | 
 644 | layer {
 645 |   name: "conv4-2/dwise"
 646 |   #type: "Convolution"
 647 |   type: "DepthwiseConvolution"
 648 |   bottom: "conv4-2/expand/hswish"
 649 |   top: "conv4-2/dwise"
 650 |   param {
 651 |     lr_mult: 1
 652 |     decay_mult: 1
 653 |   }
 654 |   convolution_param {
 655 |     num_output: 96
 656 |     group: 96
 657 |     pad: 2
 658 |     bias_term: false
 659 |     kernel_size: 5
 660 |     stride: 2
 661 |     weight_filler {
 662 |       type: "msra"
 663 |     }
 664 |     engine: CAFFE
 665 |   }
 666 | }
 667 | layer {
 668 |   name: "conv4-2/dwise-bn"
 669 |   type: "BatchNorm"
 670 |   bottom: "conv4-2/dwise"
 671 |   top: "conv4-2/dwise"
 672 | }
 673 | layer {
 674 |   name: "conv4-2/dwise-bn-scale"
 675 |   type: "Scale"
 676 |   bottom: "conv4-2/dwise"
 677 |   top: "conv4-2/dwise"
 678 |   scale_param {
 679 |     bias_term: true
 680 |   }
 681 | }
 682 | 
 683 | # se
 684 | layer {
 685 |   name: "conv4-2/dwise-se-pool"
 686 |   type: "Pooling"
 687 |   bottom: "conv4-2/dwise"
 688 |   top: "conv4-2/dwise-se-pool"
 689 |   pooling_param {
 690 |     pool: AVE
 691 |     global_pooling: true
 692 |   }
 693 | }
 694 | layer {
 695 |   name: "conv4-2/dwise-se-fc1"
 696 |   type: "InnerProduct"
 697 |   bottom: "conv4-2/dwise-se-pool"
 698 |   top:"conv4-2/dwise-se-fc1"
 699 |   param {
 700 |     lr_mult: 1
 701 |     decay_mult: 1
 702 |   }
 703 |   param {
 704 |     lr_mult: 2
 705 |     decay_mult: 0
 706 |   }
 707 |   inner_product_param {
 708 |     num_output: 24
 709 |     weight_filler {
 710 |       type: "msra"
 711 |     }
 712 |     bias_filler {
 713 |       type: "constant"
 714 |       value: 0
 715 |     }
 716 |   }
 717 | }
 718 | layer {
 719 |   name: "conv4-2/dwise-se-fc1-relu"
 720 |   type: "ReLU6"
 721 |   bottom: "conv4-2/dwise-se-fc1"
 722 |   top: "conv4-2/dwise-se-fc1"
 723 | }
 724 | layer {
 725 |   name: "conv4-2/dwise-se-fc2"
 726 |   type: "InnerProduct"
 727 |   bottom: "conv4-2/dwise-se-fc1"
 728 |   top:"conv4-2/dwise-se-fc2"
 729 |   param {
 730 |     lr_mult: 1
 731 |     decay_mult: 1
 732 |   }
 733 |   param {
 734 |     lr_mult: 2
 735 |     decay_mult: 0
 736 |   }
 737 |   inner_product_param {
 738 |     num_output: 96
 739 |     weight_filler {
 740 |       type: "msra"
 741 |     }
 742 |     bias_filler {
 743 |       type: "constant"
 744 |       value: 0
 745 |     }
 746 |   }
 747 | }
 748 | # H-sigmoid
 749 | layer {
 750 |   name: "conv4-2/dwise-se-fc2/shift"
 751 |   type: "Power"
 752 |   bottom: "conv4-2/dwise-se-fc2"
 753 |   top: "conv4-2/dwise-se-fc2/shift"
 754 |   power_param {
 755 |     power: 1
 756 |     scale: 1
 757 |     shift: 3
 758 |   }
 759 | }
 760 | layer {
 761 |   name: "conv4-2/dwise-se-fc2/shift-relu"
 762 |   type: "ReLU6"
 763 |   bottom: "conv4-2/dwise-se-fc2/shift"
 764 |   top: "conv4-2/dwise-se-fc2/shift"
 765 | }
 766 | layer {
 767 |   name: "conv4-2/dwise-se-fc2/shift-div"
 768 |   type: "Power"
 769 |   bottom: "conv4-2/dwise-se-fc2/shift"
 770 |   top: "conv4-2/dwise-se-fc2/shift-div"
 771 |   power_param {
 772 |     power: 1
 773 |     scale: 0.1666666667
 774 |     shift: 0
 775 |   }
 776 | }
 777 | layer {
 778 |   name: "conv4-2/dwise/scale"
 779 |   type: "Scale"
 780 |   bottom: "conv4-2/dwise"
 781 |   bottom: "conv4-2/dwise-se-fc2/shift-div"
 782 |   top: "conv4-2/dwise/scale"
 783 |   scale_param{
 784 |     axis: 0
 785 |   }
 786 | }
 787 | # H-swish
 788 | layer {
 789 |   name: "conv4-2/dwise/shift"
 790 |   type: "Power"
 791 |   bottom: "conv4-2/dwise/scale"
 792 |   top: "conv4-2/dwise/shift"
 793 |   power_param {
 794 |     power: 1
 795 |     scale: 1
 796 |     shift: 3
 797 |   }
 798 | }
 799 | layer {
 800 |   name: "conv4-2/dwise/shift-relu"
 801 |   type: "ReLU6"
 802 |   bottom: "conv4-2/dwise/shift"
 803 |   top: "conv4-2/dwise/shift"
 804 | }
 805 | layer {
 806 |   name: "conv4-2/dwise/shift-div"
 807 |   type: "Power"
 808 |   bottom: "conv4-2/dwise/shift"
 809 |   top: "conv4-2/dwise/shift-div"
 810 |   power_param {
 811 |     power: 1
 812 |     scale: 0.1666666667
 813 |     shift: 0
 814 |   }
 815 | }
 816 | layer {
 817 |   name: "conv4-2/dwise/hswish"
 818 |   type: "Eltwise"
 819 |   bottom: "conv4-2/dwise/scale"
 820 |   bottom: "conv4-2/dwise/shift-div"
 821 |   top: "conv4-2/dwise/hswish"
 822 |   eltwise_param {
 823 |     operation: PROD
 824 |   }
 825 | }
 826 | 
 827 | layer {
 828 |   name: "conv4-2/linear"
 829 |   type: "Convolution"
 830 |   bottom: "conv4-2/dwise/hswish"
 831 |   top: "conv4-2/linear"
 832 |   param {
 833 |     lr_mult: 1
 834 |     decay_mult: 1
 835 |   }
 836 |   convolution_param {
 837 |     num_output: 40
 838 |     pad: 0
 839 |     bias_term: false
 840 |     kernel_size: 1
 841 |     stride: 1
 842 |     weight_filler {
 843 |       type: "msra"
 844 |     }
 845 |   }
 846 | }
 847 | layer {
 848 |   name: "conv4-2/linear-bn"
 849 |   type: "BatchNorm"
 850 |   bottom: "conv4-2/linear"
 851 |   top: "conv4-2/linear"
 852 | }
 853 | layer {
 854 |   name: "conv4-2/linear-bn-scale"
 855 |   type: "Scale"
 856 |   bottom: "conv4-2/linear"
 857 |   top: "conv4-2/linear"
 858 |   scale_param {
 859 |     bias_term: true
 860 |   }
 861 | }
 862 | 
 863 | #################### stage 5-1 ####################
 864 | layer {
 865 |   name: "conv5-1/expand"
 866 |   type: "Convolution"
 867 |   bottom: "conv4-2/linear"
 868 |   top: "conv5-1/expand"
 869 |   param {
 870 |     lr_mult: 1
 871 |     decay_mult: 1
 872 |   }
 873 |   convolution_param {
 874 |     num_output: 240
 875 |     pad: 0
 876 |     bias_term: false
 877 |     kernel_size: 1
 878 |     stride: 1
 879 |     weight_filler {
 880 |       type: "msra"
 881 |     }
 882 |   }
 883 | }
 884 | layer {
 885 |   name: "conv5-1/expand-bn"
 886 |   type: "BatchNorm"
 887 |   bottom: "conv5-1/expand"
 888 |   top: "conv5-1/expand"
 889 | }
 890 | layer {
 891 |   name: "conv5-1/expand-bn-scale"
 892 |   type: "Scale"
 893 |   bottom: "conv5-1/expand"
 894 |   top: "conv5-1/expand"
 895 |   scale_param {
 896 |     bias_term: true
 897 |   }
 898 | }
 899 | # H-swish
 900 | layer {
 901 |   name: "conv5-1/expand/shift"
 902 |   type: "Power"
 903 |   bottom: "conv5-1/expand"
 904 |   top: "conv5-1/expand/shift"
 905 |   power_param {
 906 |     power: 1
 907 |     scale: 1
 908 |     shift: 3
 909 |   }
 910 | }
 911 | layer {
 912 |   name: "conv5-1/expand/shift-relu"
 913 |   type: "ReLU6"
 914 |   bottom: "conv5-1/expand/shift"
 915 |   top: "conv5-1/expand/shift"
 916 | }
 917 | layer {
 918 |   name: "conv5-1/expand/shift-div"
 919 |   type: "Power"
 920 |   bottom: "conv5-1/expand/shift"
 921 |   top: "conv5-1/expand/shift-div"
 922 |   power_param {
 923 |     power: 1
 924 |     scale: 0.1666666667
 925 |     shift: 0
 926 |   }
 927 | }
 928 | layer {
 929 |   name: "conv5-1/expand/hswish"
 930 |   type: "Eltwise"
 931 |   bottom: "conv5-1/expand"
 932 |   bottom: "conv5-1/expand/shift-div"
 933 |   top: "conv5-1/expand/hswish"
 934 |   eltwise_param {
 935 |     operation: PROD
 936 |   }
 937 | }
 938 | 
 939 | layer {
 940 |   name: "conv5-1/dwise"
 941 |   #type: "Convolution"
 942 |   type: "DepthwiseConvolution"
 943 |   bottom: "conv5-1/expand/hswish"
 944 |   top: "conv5-1/dwise"
 945 |   param {
 946 |     lr_mult: 1
 947 |     decay_mult: 1
 948 |   }
 949 |   convolution_param {
 950 |     num_output: 240
 951 |     group: 240
 952 |     pad: 2
 953 |     bias_term: false
 954 |     kernel_size: 5
 955 |     stride: 1
 956 |     weight_filler {
 957 |       type: "msra"
 958 |     }
 959 |     engine: CAFFE
 960 |   }
 961 | }
 962 | layer {
 963 |   name: "conv5-1/dwise-bn"
 964 |   type: "BatchNorm"
 965 |   bottom: "conv5-1/dwise"
 966 |   top: "conv5-1/dwise"
 967 | }
 968 | layer {
 969 |   name: "conv5-1/dwise-bn-scale"
 970 |   type: "Scale"
 971 |   bottom: "conv5-1/dwise"
 972 |   top: "conv5-1/dwise"
 973 |   scale_param {
 974 |     bias_term: true
 975 |   }
 976 | }
 977 | 
 978 | # se
 979 | layer {
 980 |   name: "conv5-1/dwise-se-pool"
 981 |   type: "Pooling"
 982 |   bottom: "conv5-1/dwise"
 983 |   top: "conv5-1/dwise-se-pool"
 984 |   pooling_param {
 985 |     pool: AVE
 986 |     global_pooling: true
 987 |   }
 988 | }
 989 | layer {
 990 |   name: "conv5-1/dwise-se-fc1"
 991 |   type: "InnerProduct"
 992 |   bottom: "conv5-1/dwise-se-pool"
 993 |   top:"conv5-1/dwise-se-fc1"
 994 |   param {
 995 |     lr_mult: 1
 996 |     decay_mult: 1
 997 |   }
 998 |   param {
 999 |     lr_mult: 2
1000 |     decay_mult: 0
1001 |   }
1002 |   inner_product_param {
1003 |     num_output: 60
1004 |     weight_filler {
1005 |       type: "msra"
1006 |     }
1007 |     bias_filler {
1008 |       type: "constant"
1009 |       value: 0
1010 |     }
1011 |   }
1012 | }
1013 | layer {
1014 |   name: "conv5-1/dwise-se-fc1-relu"
1015 |   type: "ReLU6"
1016 |   bottom: "conv5-1/dwise-se-fc1"
1017 |   top: "conv5-1/dwise-se-fc1"
1018 | }
1019 | layer {
1020 |   name: "conv5-1/dwise-se-fc2"
1021 |   type: "InnerProduct"
1022 |   bottom: "conv5-1/dwise-se-fc1"
1023 |   top:"conv5-1/dwise-se-fc2"
1024 |   param {
1025 |     lr_mult: 1
1026 |     decay_mult: 1
1027 |   }
1028 |   param {
1029 |     lr_mult: 2
1030 |     decay_mult: 0
1031 |   }
1032 |   inner_product_param {
1033 |     num_output: 240
1034 |     weight_filler {
1035 |       type: "msra"
1036 |     }
1037 |     bias_filler {
1038 |       type: "constant"
1039 |       value: 0
1040 |     }
1041 |   }
1042 | }
1043 | # H-sigmoid
1044 | layer {
1045 |   name: "conv5-1/dwise-se-fc2/shift"
1046 |   type: "Power"
1047 |   bottom: "conv5-1/dwise-se-fc2"
1048 |   top: "conv5-1/dwise-se-fc2/shift"
1049 |   power_param {
1050 |     power: 1
1051 |     scale: 1
1052 |     shift: 3
1053 |   }
1054 | }
1055 | layer {
1056 |   name: "conv5-1/dwise-se-fc2/shift-relu"
1057 |   type: "ReLU6"
1058 |   bottom: "conv5-1/dwise-se-fc2/shift"
1059 |   top: "conv5-1/dwise-se-fc2/shift"
1060 | }
1061 | layer {
1062 |   name: "conv5-1/dwise-se-fc2/shift-div"
1063 |   type: "Power"
1064 |   bottom: "conv5-1/dwise-se-fc2/shift"
1065 |   top: "conv5-1/dwise-se-fc2/shift-div"
1066 |   power_param {
1067 |     power: 1
1068 |     scale: 0.1666666667
1069 |     shift: 0
1070 |   }
1071 | }
1072 | layer {
1073 |   name: "conv5-1/dwise/scale"
1074 |   type: "Scale"
1075 |   bottom: "conv5-1/dwise"
1076 |   bottom: "conv5-1/dwise-se-fc2/shift-div"
1077 |   top: "conv5-1/dwise/scale"
1078 |   scale_param{
1079 |     axis: 0
1080 |   }
1081 | }
1082 | # H-swish
1083 | layer {
1084 |   name: "conv5-1/dwise/shift"
1085 |   type: "Power"
1086 |   bottom: "conv5-1/dwise/scale"
1087 |   top: "conv5-1/dwise/shift"
1088 |   power_param {
1089 |     power: 1
1090 |     scale: 1
1091 |     shift: 3
1092 |   }
1093 | }
1094 | layer {
1095 |   name: "conv5-1/dwise/shift-relu"
1096 |   type: "ReLU6"
1097 |   bottom: "conv5-1/dwise/shift"
1098 |   top: "conv5-1/dwise/shift"
1099 | }
1100 | layer {
1101 |   name: "conv5-1/dwise/shift-div"
1102 |   type: "Power"
1103 |   bottom: "conv5-1/dwise/shift"
1104 |   top: "conv5-1/dwise/shift-div"
1105 |   power_param {
1106 |     power: 1
1107 |     scale: 0.1666666667
1108 |     shift: 0
1109 |   }
1110 | }
1111 | layer {
1112 |   name: "conv5-1/dwise/hswish"
1113 |   type: "Eltwise"
1114 |   bottom: "conv5-1/dwise/scale"
1115 |   bottom: "conv5-1/dwise/shift-div"
1116 |   top: "conv5-1/dwise/hswish"
1117 |   eltwise_param {
1118 |     operation: PROD
1119 |   }
1120 | }
1121 | 
1122 | layer {
1123 |   name: "conv5-1/linear"
1124 |   type: "Convolution"
1125 |   bottom: "conv5-1/dwise/hswish"
1126 |   top: "conv5-1/linear"
1127 |   param {
1128 |     lr_mult: 1
1129 |     decay_mult: 1
1130 |   }
1131 |   convolution_param {
1132 |     num_output: 40
1133 |     pad: 0
1134 |     bias_term: false
1135 |     kernel_size: 1
1136 |     stride: 1
1137 |     weight_filler {
1138 |       type: "msra"
1139 |     }
1140 |   }
1141 | }
1142 | layer {
1143 |   name: "conv5-1/linear-bn"
1144 |   type: "BatchNorm"
1145 |   bottom: "conv5-1/linear"
1146 |   top: "conv5-1/linear"
1147 | }
1148 | layer {
1149 |   name: "conv5-1/linear-bn-scale"
1150 |   type: "Scale"
1151 |   bottom: "conv5-1/linear"
1152 |   top: "conv5-1/linear"
1153 |   scale_param {
1154 |     bias_term: true
1155 |   }
1156 | }
1157 | 
1158 | layer {
1159 |   name: "block5-1"
1160 |   type: "Eltwise"
1161 |   bottom: "conv4-2/linear"
1162 |   bottom: "conv5-1/linear"
1163 |   top: "block5-1"
1164 | }
1165 | 
1166 | #################### stage 5-2 ####################
1167 | layer {
1168 |   name: "conv5-2/expand"
1169 |   type: "Convolution"
1170 |   bottom: "block5-1"
1171 |   top: "conv5-2/expand"
1172 |   param {
1173 |     lr_mult: 1
1174 |     decay_mult: 1
1175 |   }
1176 |   convolution_param {
1177 |     num_output: 240
1178 |     pad: 0
1179 |     bias_term: false
1180 |     kernel_size: 1
1181 |     stride: 1
1182 |     weight_filler {
1183 |       type: "msra"
1184 |     }
1185 |   }
1186 | }
1187 | layer {
1188 |   name: "conv5-2/expand-bn"
1189 |   type: "BatchNorm"
1190 |   bottom: "conv5-2/expand"
1191 |   top: "conv5-2/expand"
1192 | }
1193 | layer {
1194 |   name: "conv5-2/expand-bn-scale"
1195 |   type: "Scale"
1196 |   bottom: "conv5-2/expand"
1197 |   top: "conv5-2/expand"
1198 |   scale_param {
1199 |     bias_term: true
1200 |   }
1201 | }
1202 | # H-swish
1203 | layer {
1204 |   name: "conv5-2/expand/shift"
1205 |   type: "Power"
1206 |   bottom: "conv5-2/expand"
1207 |   top: "conv5-2/expand/shift"
1208 |   power_param {
1209 |     power: 1
1210 |     scale: 1
1211 |     shift: 3
1212 |   }
1213 | }
1214 | layer {
1215 |   name: "conv5-2/expand/shift-relu"
1216 |   type: "ReLU6"
1217 |   bottom: "conv5-2/expand/shift"
1218 |   top: "conv5-2/expand/shift"
1219 | }
1220 | layer {
1221 |   name: "conv5-2/expand/shift-div"
1222 |   type: "Power"
1223 |   bottom: "conv5-2/expand/shift"
1224 |   top: "conv5-2/expand/shift-div"
1225 |   power_param {
1226 |     power: 1
1227 |     scale: 0.1666666667
1228 |     shift: 0
1229 |   }
1230 | }
1231 | layer {
1232 |   name: "conv5-2/expand/hswish"
1233 |   type: "Eltwise"
1234 |   bottom: "conv5-2/expand"
1235 |   bottom: "conv5-2/expand/shift-div"
1236 |   top: "conv5-2/expand/hswish"
1237 |   eltwise_param {
1238 |     operation: PROD
1239 |   }
1240 | }
1241 | 
1242 | layer {
1243 |   name: "conv5-2/dwise"
1244 |   #type: "Convolution"
1245 |   type: "DepthwiseConvolution"
1246 |   bottom: "conv5-2/expand/hswish"
1247 |   top: "conv5-2/dwise"
1248 |   param {
1249 |     lr_mult: 1
1250 |     decay_mult: 1
1251 |   }
1252 |   convolution_param {
1253 |     num_output: 240
1254 |     group: 240
1255 |     pad: 2
1256 |     bias_term: false
1257 |     kernel_size: 5
1258 |     stride: 1
1259 |     weight_filler {
1260 |       type: "msra"
1261 |     }
1262 |     engine: CAFFE
1263 |   }
1264 | }
1265 | layer {
1266 |   name: "conv5-2/dwise-bn"
1267 |   type: "BatchNorm"
1268 |   bottom: "conv5-2/dwise"
1269 |   top: "conv5-2/dwise"
1270 | }
1271 | layer {
1272 |   name: "conv5-2/dwise-bn-scale"
1273 |   type: "Scale"
1274 |   bottom: "conv5-2/dwise"
1275 |   top: "conv5-2/dwise"
1276 |   scale_param {
1277 |     bias_term: true
1278 |   }
1279 | }
1280 | 
1281 | # se
1282 | layer {
1283 |   name: "conv5-2/dwise-se-pool"
1284 |   type: "Pooling"
1285 |   bottom: "conv5-2/dwise"
1286 |   top: "conv5-2/dwise-se-pool"
1287 |   pooling_param {
1288 |     pool: AVE
1289 |     global_pooling: true
1290 |   }
1291 | }
1292 | layer {
1293 |   name: "conv5-2/dwise-se-fc1"
1294 |   type: "InnerProduct"
1295 |   bottom: "conv5-2/dwise-se-pool"
1296 |   top:"conv5-2/dwise-se-fc1"
1297 |   param {
1298 |     lr_mult: 1
1299 |     decay_mult: 1
1300 |   }
1301 |   param {
1302 |     lr_mult: 2
1303 |     decay_mult: 0
1304 |   }
1305 |   inner_product_param {
1306 |     num_output: 60
1307 |     weight_filler {
1308 |       type: "msra"
1309 |     }
1310 |     bias_filler {
1311 |       type: "constant"
1312 |       value: 0
1313 |     }
1314 |   }
1315 | }
1316 | layer {
1317 |   name: "conv5-2/dwise-se-fc1-relu"
1318 |   type: "ReLU6"
1319 |   bottom: "conv5-2/dwise-se-fc1"
1320 |   top: "conv5-2/dwise-se-fc1"
1321 | }
1322 | layer {
1323 |   name: "conv5-2/dwise-se-fc2"
1324 |   type: "InnerProduct"
1325 |   bottom: "conv5-2/dwise-se-fc1"
1326 |   top:"conv5-2/dwise-se-fc2"
1327 |   param {
1328 |     lr_mult: 1
1329 |     decay_mult: 1
1330 |   }
1331 |   param {
1332 |     lr_mult: 2
1333 |     decay_mult: 0
1334 |   }
1335 |   inner_product_param {
1336 |     num_output: 240
1337 |     weight_filler {
1338 |       type: "msra"
1339 |     }
1340 |     bias_filler {
1341 |       type: "constant"
1342 |       value: 0
1343 |     }
1344 |   }
1345 | }
1346 | # H-sigmoid
1347 | layer {
1348 |   name: "conv5-2/dwise-se-fc2/shift"
1349 |   type: "Power"
1350 |   bottom: "conv5-2/dwise-se-fc2"
1351 |   top: "conv5-2/dwise-se-fc2/shift"
1352 |   power_param {
1353 |     power: 1
1354 |     scale: 1
1355 |     shift: 3
1356 |   }
1357 | }
1358 | layer {
1359 |   name: "conv5-2/dwise-se-fc2/shift-relu"
1360 |   type: "ReLU6"
1361 |   bottom: "conv5-2/dwise-se-fc2/shift"
1362 |   top: "conv5-2/dwise-se-fc2/shift"
1363 | }
1364 | layer {
1365 |   name: "conv5-2/dwise-se-fc2/shift-div"
1366 |   type: "Power"
1367 |   bottom: "conv5-2/dwise-se-fc2/shift"
1368 |   top: "conv5-2/dwise-se-fc2/shift-div"
1369 |   power_param {
1370 |     power: 1
1371 |     scale: 0.1666666667
1372 |     shift: 0
1373 |   }
1374 | }
1375 | layer {
1376 |   name: "conv5-2/dwise/scale"
1377 |   type: "Scale"
1378 |   bottom: "conv5-2/dwise"
1379 |   bottom: "conv5-2/dwise-se-fc2/shift-div"
1380 |   top: "conv5-2/dwise/scale"
1381 |   scale_param{
1382 |     axis: 0
1383 |   }
1384 | }
1385 | # H-swish
1386 | layer {
1387 |   name: "conv5-2/dwise/shift"
1388 |   type: "Power"
1389 |   bottom: "conv5-2/dwise/scale"
1390 |   top: "conv5-2/dwise/shift"
1391 |   power_param {
1392 |     power: 1
1393 |     scale: 1
1394 |     shift: 3
1395 |   }
1396 | }
1397 | layer {
1398 |   name: "conv5-2/dwise/shift-relu"
1399 |   type: "ReLU6"
1400 |   bottom: "conv5-2/dwise/shift"
1401 |   top: "conv5-2/dwise/shift"
1402 | }
1403 | layer {
1404 |   name: "conv5-2/dwise/shift-div"
1405 |   type: "Power"
1406 |   bottom: "conv5-2/dwise/shift"
1407 |   top: "conv5-2/dwise/shift-div"
1408 |   power_param {
1409 |     power: 1
1410 |     scale: 0.1666666667
1411 |     shift: 0
1412 |   }
1413 | }
1414 | layer {
1415 |   name: "conv5-2/dwise/hswish"
1416 |   type: "Eltwise"
1417 |   bottom: "conv5-2/dwise/scale"
1418 |   bottom: "conv5-2/dwise/shift-div"
1419 |   top: "conv5-2/dwise/hswish"
1420 |   eltwise_param {
1421 |     operation: PROD
1422 |   }
1423 | }
1424 | 
1425 | layer {
1426 |   name: "conv5-2/linear"
1427 |   type: "Convolution"
1428 |   bottom: "conv5-2/dwise/hswish"
1429 |   top: "conv5-2/linear"
1430 |   param {
1431 |     lr_mult: 1
1432 |     decay_mult: 1
1433 |   }
1434 |   convolution_param {
1435 |     num_output: 40
1436 |     pad: 0
1437 |     bias_term: false
1438 |     kernel_size: 1
1439 |     stride: 1
1440 |     weight_filler {
1441 |       type: "msra"
1442 |     }
1443 |   }
1444 | }
1445 | layer {
1446 |   name: "conv5-2/linear-bn"
1447 |   type: "BatchNorm"
1448 |   bottom: "conv5-2/linear"
1449 |   top: "conv5-2/linear"
1450 | }
1451 | layer {
1452 |   name: "conv5-2/linear-bn-scale"
1453 |   type: "Scale"
1454 |   bottom: "conv5-2/linear"
1455 |   top: "conv5-2/linear"
1456 |   scale_param {
1457 |     bias_term: true
1458 |   }
1459 | }
1460 | 
1461 | layer {
1462 |   name: "block5-2"
1463 |   type: "Eltwise"
1464 |   bottom: "block5-1"
1465 |   bottom: "conv5-2/linear"
1466 |   top: "block5-2"
1467 | }
1468 | 
1469 | #################### stage 5-3 ####################
1470 | layer {
1471 |   name: "conv5-3/expand"
1472 |   type: "Convolution"
1473 |   bottom: "block5-2"
1474 |   top: "conv5-3/expand"
1475 |   param {
1476 |     lr_mult: 1
1477 |     decay_mult: 1
1478 |   }
1479 |   convolution_param {
1480 |     num_output: 120
1481 |     pad: 0
1482 |     bias_term: false
1483 |     kernel_size: 1
1484 |     stride: 1
1485 |     weight_filler {
1486 |       type: "msra"
1487 |     }
1488 |   }
1489 | }
1490 | layer {
1491 |   name: "conv5-3/expand-bn"
1492 |   type: "BatchNorm"
1493 |   bottom: "conv5-3/expand"
1494 |   top: "conv5-3/expand"
1495 | }
1496 | layer {
1497 |   name: "conv5-3/expand-bn-scale"
1498 |   type: "Scale"
1499 |   bottom: "conv5-3/expand"
1500 |   top: "conv5-3/expand"
1501 |   scale_param {
1502 |     bias_term: true
1503 |   }
1504 | }
1505 | # H-swish
1506 | layer {
1507 |   name: "conv5-3/expand/shift"
1508 |   type: "Power"
1509 |   bottom: "conv5-3/expand"
1510 |   top: "conv5-3/expand/shift"
1511 |   power_param {
1512 |     power: 1
1513 |     scale: 1
1514 |     shift: 3
1515 |   }
1516 | }
1517 | layer {
1518 |   name: "conv5-3/expand/shift-relu"
1519 |   type: "ReLU6"
1520 |   bottom: "conv5-3/expand/shift"
1521 |   top: "conv5-3/expand/shift"
1522 | }
1523 | layer {
1524 |   name: "conv5-3/expand/shift-div"
1525 |   type: "Power"
1526 |   bottom: "conv5-3/expand/shift"
1527 |   top: "conv5-3/expand/shift-div"
1528 |   power_param {
1529 |     power: 1
1530 |     scale: 0.1666666667
1531 |     shift: 0
1532 |   }
1533 | }
1534 | layer {
1535 |   name: "conv5-3/expand/hswish"
1536 |   type: "Eltwise"
1537 |   bottom: "conv5-3/expand"
1538 |   bottom: "conv5-3/expand/shift-div"
1539 |   top: "conv5-3/expand/hswish"
1540 |   eltwise_param {
1541 |     operation: PROD
1542 |   }
1543 | }
1544 | 
1545 | layer {
1546 |   name: "conv5-3/dwise"
1547 |   #type: "Convolution"
1548 |   type: "DepthwiseConvolution"
1549 |   bottom: "conv5-3/expand/hswish"
1550 |   top: "conv5-3/dwise"
1551 |   param {
1552 |     lr_mult: 1
1553 |     decay_mult: 1
1554 |   }
1555 |   convolution_param {
1556 |     num_output: 120
1557 |     group: 120
1558 |     pad: 2
1559 |     bias_term: false
1560 |     kernel_size: 5
1561 |     stride: 1
1562 |     weight_filler {
1563 |       type: "msra"
1564 |     }
1565 |     engine: CAFFE
1566 |   }
1567 | }
1568 | layer {
1569 |   name: "conv5-3/dwise-bn"
1570 |   type: "BatchNorm"
1571 |   bottom: "conv5-3/dwise"
1572 |   top: "conv5-3/dwise"
1573 | }
1574 | layer {
1575 |   name: "conv5-3/dwise-bn-scale"
1576 |   type: "Scale"
1577 |   bottom: "conv5-3/dwise"
1578 |   top: "conv5-3/dwise"
1579 |   scale_param {
1580 |     bias_term: true
1581 |   }
1582 | }
1583 | 
1584 | # se
1585 | layer {
1586 |   name: "conv5-3/dwise-se-pool"
1587 |   type: "Pooling"
1588 |   bottom: "conv5-3/dwise"
1589 |   top: "conv5-3/dwise-se-pool"
1590 |   pooling_param {
1591 |     pool: AVE
1592 |     global_pooling: true
1593 |   }
1594 | }
1595 | layer {
1596 |   name: "conv5-3/dwise-se-fc1"
1597 |   type: "InnerProduct"
1598 |   bottom: "conv5-3/dwise-se-pool"
1599 |   top:"conv5-3/dwise-se-fc1"
1600 |   param {
1601 |     lr_mult: 1
1602 |     decay_mult: 1
1603 |   }
1604 |   param {
1605 |     lr_mult: 2
1606 |     decay_mult: 0
1607 |   }
1608 |   inner_product_param {
1609 |     num_output: 30
1610 |     weight_filler {
1611 |       type: "msra"
1612 |     }
1613 |     bias_filler {
1614 |       type: "constant"
1615 |       value: 0
1616 |     }
1617 |   }
1618 | }
1619 | layer {
1620 |   name: "conv5-3/dwise-se-fc1-relu"
1621 |   type: "ReLU6"
1622 |   bottom: "conv5-3/dwise-se-fc1"
1623 |   top: "conv5-3/dwise-se-fc1"
1624 | }
1625 | layer {
1626 |   name: "conv5-3/dwise-se-fc2"
1627 |   type: "InnerProduct"
1628 |   bottom: "conv5-3/dwise-se-fc1"
1629 |   top:"conv5-3/dwise-se-fc2"
1630 |   param {
1631 |     lr_mult: 1
1632 |     decay_mult: 1
1633 |   }
1634 |   param {
1635 |     lr_mult: 2
1636 |     decay_mult: 0
1637 |   }
1638 |   inner_product_param {
1639 |     num_output: 120
1640 |     weight_filler {
1641 |       type: "msra"
1642 |     }
1643 |     bias_filler {
1644 |       type: "constant"
1645 |       value: 0
1646 |     }
1647 |   }
1648 | }
1649 | # H-sigmoid
1650 | layer {
1651 |   name: "conv5-3/dwise-se-fc2/shift"
1652 |   type: "Power"
1653 |   bottom: "conv5-3/dwise-se-fc2"
1654 |   top: "conv5-3/dwise-se-fc2/shift"
1655 |   power_param {
1656 |     power: 1
1657 |     scale: 1
1658 |     shift: 3
1659 |   }
1660 | }
1661 | layer {
1662 |   name: "conv5-3/dwise-se-fc2/shift-relu"
1663 |   type: "ReLU6"
1664 |   bottom: "conv5-3/dwise-se-fc2/shift"
1665 |   top: "conv5-3/dwise-se-fc2/shift"
1666 | }
1667 | layer {
1668 |   name: "conv5-3/dwise-se-fc2/shift-div"
1669 |   type: "Power"
1670 |   bottom: "conv5-3/dwise-se-fc2/shift"
1671 |   top: "conv5-3/dwise-se-fc2/shift-div"
1672 |   power_param {
1673 |     power: 1
1674 |     scale: 0.1666666667
1675 |     shift: 0
1676 |   }
1677 | }
1678 | layer {
1679 |   name: "conv5-3/dwise/scale"
1680 |   type: "Scale"
1681 |   bottom: "conv5-3/dwise"
1682 |   bottom: "conv5-3/dwise-se-fc2/shift-div"
1683 |   top: "conv5-3/dwise/scale"
1684 |   scale_param{
1685 |     axis: 0
1686 |   }
1687 | }
1688 | # H-swish
1689 | layer {
1690 |   name: "conv5-3/dwise/shift"
1691 |   type: "Power"
1692 |   bottom: "conv5-3/dwise/scale"
1693 |   top: "conv5-3/dwise/shift"
1694 |   power_param {
1695 |     power: 1
1696 |     scale: 1
1697 |     shift: 3
1698 |   }
1699 | }
1700 | layer {
1701 |   name: "conv5-3/dwise/shift-relu"
1702 |   type: "ReLU6"
1703 |   bottom: "conv5-3/dwise/shift"
1704 |   top: "conv5-3/dwise/shift"
1705 | }
1706 | layer {
1707 |   name: "conv5-3/dwise/shift-div"
1708 |   type: "Power"
1709 |   bottom: "conv5-3/dwise/shift"
1710 |   top: "conv5-3/dwise/shift-div"
1711 |   power_param {
1712 |     power: 1
1713 |     scale: 0.1666666667
1714 |     shift: 0
1715 |   }
1716 | }
1717 | layer {
1718 |   name: "conv5-3/dwise/hswish"
1719 |   type: "Eltwise"
1720 |   bottom: "conv5-3/dwise/scale"
1721 |   bottom: "conv5-3/dwise/shift-div"
1722 |   top: "conv5-3/dwise/hswish"
1723 |   eltwise_param {
1724 |     operation: PROD
1725 |   }
1726 | }
1727 | 
1728 | layer {
1729 |   name: "conv5-3/linear"
1730 |   type: "Convolution"
1731 |   bottom: "conv5-3/dwise/hswish"
1732 |   top: "conv5-3/linear"
1733 |   param {
1734 |     lr_mult: 1
1735 |     decay_mult: 1
1736 |   }
1737 |   convolution_param {
1738 |     num_output: 48
1739 |     pad: 0
1740 |     bias_term: false
1741 |     kernel_size: 1
1742 |     stride: 1
1743 |     weight_filler {
1744 |       type: "msra"
1745 |     }
1746 |   }
1747 | }
1748 | layer {
1749 |   name: "conv5-3/linear-bn"
1750 |   type: "BatchNorm"
1751 |   bottom: "conv5-3/linear"
1752 |   top: "conv5-3/linear"
1753 | }
1754 | layer {
1755 |   name: "conv5-3/linear-bn-scale"
1756 |   type: "Scale"
1757 |   bottom: "conv5-3/linear"
1758 |   top: "conv5-3/linear"
1759 |   scale_param {
1760 |     bias_term: true
1761 |   }
1762 | }
1763 | 
1764 | #################### stage 5-4 ####################
1765 | layer {
1766 |   name: "conv5-4/expand"
1767 |   type: "Convolution"
1768 |   bottom: "conv5-3/linear"
1769 |   top: "conv5-4/expand"
1770 |   param {
1771 |     lr_mult: 1
1772 |     decay_mult: 1
1773 |   }
1774 |   convolution_param {
1775 |     num_output: 144
1776 |     pad: 0
1777 |     bias_term: false
1778 |     kernel_size: 1
1779 |     stride: 1
1780 |     weight_filler {
1781 |       type: "msra"
1782 |     }
1783 |   }
1784 | }
1785 | layer {
1786 |   name: "conv5-4/expand-bn"
1787 |   type: "BatchNorm"
1788 |   bottom: "conv5-4/expand"
1789 |   top: "conv5-4/expand"
1790 | }
1791 | layer {
1792 |   name: "conv5-4/expand-bn-scale"
1793 |   type: "Scale"
1794 |   bottom: "conv5-4/expand"
1795 |   top: "conv5-4/expand"
1796 |   scale_param {
1797 |     bias_term: true
1798 |   }
1799 | }
1800 | # H-swish
1801 | layer {
1802 |   name: "conv5-4/expand/shift"
1803 |   type: "Power"
1804 |   bottom: "conv5-4/expand"
1805 |   top: "conv5-4/expand/shift"
1806 |   power_param {
1807 |     power: 1
1808 |     scale: 1
1809 |     shift: 3
1810 |   }
1811 | }
1812 | layer {
1813 |   name: "conv5-4/expand/shift-relu"
1814 |   type: "ReLU6"
1815 |   bottom: "conv5-4/expand/shift"
1816 |   top: "conv5-4/expand/shift"
1817 | }
1818 | layer {
1819 |   name: "conv5-4/expand/shift-div"
1820 |   type: "Power"
1821 |   bottom: "conv5-4/expand/shift"
1822 |   top: "conv5-4/expand/shift-div"
1823 |   power_param {
1824 |     power: 1
1825 |     scale: 0.1666666667
1826 |     shift: 0
1827 |   }
1828 | }
1829 | layer {
1830 |   name: "conv5-4/expand/hswish"
1831 |   type: "Eltwise"
1832 |   bottom: "conv5-4/expand"
1833 |   bottom: "conv5-4/expand/shift-div"
1834 |   top: "conv5-4/expand/hswish"
1835 |   eltwise_param {
1836 |     operation: PROD
1837 |   }
1838 | }
1839 | 
1840 | layer {
1841 |   name: "conv5-4/dwise"
1842 |   #type: "Convolution"
1843 |   type: "DepthwiseConvolution"
1844 |   bottom: "conv5-4/expand/hswish"
1845 |   top: "conv5-4/dwise"
1846 |   param {
1847 |     lr_mult: 1
1848 |     decay_mult: 1
1849 |   }
1850 |   convolution_param {
1851 |     num_output: 144
1852 |     group: 144
1853 |     pad: 2
1854 |     bias_term: false
1855 |     kernel_size: 5
1856 |     stride: 1
1857 |     weight_filler {
1858 |       type: "msra"
1859 |     }
1860 |     engine: CAFFE
1861 |   }
1862 | }
1863 | layer {
1864 |   name: "conv5-4/dwise-bn"
1865 |   type: "BatchNorm"
1866 |   bottom: "conv5-4/dwise"
1867 |   top: "conv5-4/dwise"
1868 | }
1869 | layer {
1870 |   name: "conv5-4/dwise-bn-scale"
1871 |   type: "Scale"
1872 |   bottom: "conv5-4/dwise"
1873 |   top: "conv5-4/dwise"
1874 |   scale_param {
1875 |     bias_term: true
1876 |   }
1877 | }
1878 | 
1879 | # se
1880 | layer {
1881 |   name: "conv5-4/dwise-se-pool"
1882 |   type: "Pooling"
1883 |   bottom: "conv5-4/dwise"
1884 |   top: "conv5-4/dwise-se-pool"
1885 |   pooling_param {
1886 |     pool: AVE
1887 |     global_pooling: true
1888 |   }
1889 | }
1890 | layer {
1891 |   name: "conv5-4/dwise-se-fc1"
1892 |   type: "InnerProduct"
1893 |   bottom: "conv5-4/dwise-se-pool"
1894 |   top:"conv5-4/dwise-se-fc1"
1895 |   param {
1896 |     lr_mult: 1
1897 |     decay_mult: 1
1898 |   }
1899 |   param {
1900 |     lr_mult: 2
1901 |     decay_mult: 0
1902 |   }
1903 |   inner_product_param {
1904 |     num_output: 36
1905 |     weight_filler {
1906 |       type: "msra"
1907 |     }
1908 |     bias_filler {
1909 |       type: "constant"
1910 |       value: 0
1911 |     }
1912 |   }
1913 | }
1914 | layer {
1915 |   name: "conv5-4/dwise-se-fc1-relu"
1916 |   type: "ReLU6"
1917 |   bottom: "conv5-4/dwise-se-fc1"
1918 |   top: "conv5-4/dwise-se-fc1"
1919 | }
1920 | layer {
1921 |   name: "conv5-4/dwise-se-fc2"
1922 |   type: "InnerProduct"
1923 |   bottom: "conv5-4/dwise-se-fc1"
1924 |   top:"conv5-4/dwise-se-fc2"
1925 |   param {
1926 |     lr_mult: 1
1927 |     decay_mult: 1
1928 |   }
1929 |   param {
1930 |     lr_mult: 2
1931 |     decay_mult: 0
1932 |   }
1933 |   inner_product_param {
1934 |     num_output: 144
1935 |     weight_filler {
1936 |       type: "msra"
1937 |     }
1938 |     bias_filler {
1939 |       type: "constant"
1940 |       value: 0
1941 |     }
1942 |   }
1943 | }
1944 | # H-sigmoid
1945 | layer {
1946 |   name: "conv5-4/dwise-se-fc2/shift"
1947 |   type: "Power"
1948 |   bottom: "conv5-4/dwise-se-fc2"
1949 |   top: "conv5-4/dwise-se-fc2/shift"
1950 |   power_param {
1951 |     power: 1
1952 |     scale: 1
1953 |     shift: 3
1954 |   }
1955 | }
1956 | layer {
1957 |   name: "conv5-4/dwise-se-fc2/shift-relu"
1958 |   type: "ReLU6"
1959 |   bottom: "conv5-4/dwise-se-fc2/shift"
1960 |   top: "conv5-4/dwise-se-fc2/shift"
1961 | }
1962 | layer {
1963 |   name: "conv5-4/dwise-se-fc2/shift-div"
1964 |   type: "Power"
1965 |   bottom: "conv5-4/dwise-se-fc2/shift"
1966 |   top: "conv5-4/dwise-se-fc2/shift-div"
1967 |   power_param {
1968 |     power: 1
1969 |     scale: 0.1666666667
1970 |     shift: 0
1971 |   }
1972 | }
1973 | layer {
1974 |   name: "conv5-4/dwise/scale"
1975 |   type: "Scale"
1976 |   bottom: "conv5-4/dwise"
1977 |   bottom: "conv5-4/dwise-se-fc2/shift-div"
1978 |   top: "conv5-4/dwise/scale"
1979 |   scale_param{
1980 |     axis: 0
1981 |   }
1982 | }
1983 | # H-swish
1984 | layer {
1985 |   name: "conv5-4/dwise/shift"
1986 |   type: "Power"
1987 |   bottom: "conv5-4/dwise/scale"
1988 |   top: "conv5-4/dwise/shift"
1989 |   power_param {
1990 |     power: 1
1991 |     scale: 1
1992 |     shift: 3
1993 |   }
1994 | }
1995 | layer {
1996 |   name: "conv5-4/dwise/shift-relu"
1997 |   type: "ReLU6"
1998 |   bottom: "conv5-4/dwise/shift"
1999 |   top: "conv5-4/dwise/shift"
2000 | }
2001 | layer {
2002 |   name: "conv5-4/dwise/shift-div"
2003 |   type: "Power"
2004 |   bottom: "conv5-4/dwise/shift"
2005 |   top: "conv5-4/dwise/shift-div"
2006 |   power_param {
2007 |     power: 1
2008 |     scale: 0.1666666667
2009 |     shift: 0
2010 |   }
2011 | }
2012 | layer {
2013 |   name: "conv5-4/dwise/hswish"
2014 |   type: "Eltwise"
2015 |   bottom: "conv5-4/dwise/scale"
2016 |   bottom: "conv5-4/dwise/shift-div"
2017 |   top: "conv5-4/dwise/hswish"
2018 |   eltwise_param {
2019 |     operation: PROD
2020 |   }
2021 | }
2022 | 
2023 | layer {
2024 |   name: "conv5-4/linear"
2025 |   type: "Convolution"
2026 |   bottom: "conv5-4/dwise/hswish"
2027 |   top: "conv5-4/linear"
2028 |   param {
2029 |     lr_mult: 1
2030 |     decay_mult: 1
2031 |   }
2032 |   convolution_param {
2033 |     num_output: 48
2034 |     pad: 0
2035 |     bias_term: false
2036 |     kernel_size: 1
2037 |     stride: 1
2038 |     weight_filler {
2039 |       type: "msra"
2040 |     }
2041 |   }
2042 | }
2043 | layer {
2044 |   name: "conv5-4/linear-bn"
2045 |   type: "BatchNorm"
2046 |   bottom: "conv5-4/linear"
2047 |   top: "conv5-4/linear"
2048 | }
2049 | layer {
2050 |   name: "conv5-4/linear-bn-scale"
2051 |   type: "Scale"
2052 |   bottom: "conv5-4/linear"
2053 |   top: "conv5-4/linear"
2054 |   scale_param {
2055 |     bias_term: true
2056 |   }
2057 | }
2058 | 
2059 | layer {
2060 |   name: "block5-4"
2061 |   type: "Eltwise"
2062 |   bottom: "conv5-3/linear"
2063 |   bottom: "conv5-4/linear"
2064 |   top: "block5-4"
2065 | }
2066 | 
2067 | #################### stage 5-5 ####################
2068 | layer {
2069 |   name: "conv5-5/expand"
2070 |   type: "Convolution"
2071 |   bottom: "block5-4"
2072 |   top: "conv5-5/expand"
2073 |   param {
2074 |     lr_mult: 1
2075 |     decay_mult: 1
2076 |   }
2077 |   convolution_param {
2078 |     num_output: 288
2079 |     pad: 0
2080 |     bias_term: false
2081 |     kernel_size: 1
2082 |     stride: 1
2083 |     weight_filler {
2084 |       type: "msra"
2085 |     }
2086 |   }
2087 | }
2088 | layer {
2089 |   name: "conv5-5/expand-bn"
2090 |   type: "BatchNorm"
2091 |   bottom: "conv5-5/expand"
2092 |   top: "conv5-5/expand"
2093 | }
2094 | layer {
2095 |   name: "conv5-5/expand-bn-scale"
2096 |   type: "Scale"
2097 |   bottom: "conv5-5/expand"
2098 |   top: "conv5-5/expand"
2099 |   scale_param {
2100 |     bias_term: true
2101 |   }
2102 | }
2103 | # H-swish
2104 | layer {
2105 |   name: "conv5-5/expand/shift"
2106 |   type: "Power"
2107 |   bottom: "conv5-5/expand"
2108 |   top: "conv5-5/expand/shift"
2109 |   power_param {
2110 |     power: 1
2111 |     scale: 1
2112 |     shift: 3
2113 |   }
2114 | }
2115 | layer {
2116 |   name: "conv5-5/expand/shift-relu"
2117 |   type: "ReLU6"
2118 |   bottom: "conv5-5/expand/shift"
2119 |   top: "conv5-5/expand/shift"
2120 | }
2121 | layer {
2122 |   name: "conv5-5/expand/shift-div"
2123 |   type: "Power"
2124 |   bottom: "conv5-5/expand/shift"
2125 |   top: "conv5-5/expand/shift-div"
2126 |   power_param {
2127 |     power: 1
2128 |     scale: 0.1666666667
2129 |     shift: 0
2130 |   }
2131 | }
2132 | layer {
2133 |   name: "conv5-5/expand/hswish"
2134 |   type: "Eltwise"
2135 |   bottom: "conv5-5/expand"
2136 |   bottom: "conv5-5/expand/shift-div"
2137 |   top: "conv5-5/expand/hswish"
2138 |   eltwise_param {
2139 |     operation: PROD
2140 |   }
2141 | }
2142 | 
2143 | layer {
2144 |   name: "conv5-5/dwise"
2145 |   #type: "Convolution"
2146 |   type: "DepthwiseConvolution"
2147 |   bottom: "conv5-5/expand/hswish"
2148 |   top: "conv5-5/dwise"
2149 |   param {
2150 |     lr_mult: 1
2151 |     decay_mult: 1
2152 |   }
2153 |   convolution_param {
2154 |     num_output: 288
2155 |     group: 288
2156 |     pad: 2
2157 |     bias_term: false
2158 |     kernel_size: 5
2159 |     stride: 2
2160 |     weight_filler {
2161 |       type: "msra"
2162 |     }
2163 |     engine: CAFFE
2164 |   }
2165 | }
2166 | layer {
2167 |   name: "conv5-5/dwise-bn"
2168 |   type: "BatchNorm"
2169 |   bottom: "conv5-5/dwise"
2170 |   top: "conv5-5/dwise"
2171 | }
2172 | layer {
2173 |   name: "conv5-5/dwise-bn-scale"
2174 |   type: "Scale"
2175 |   bottom: "conv5-5/dwise"
2176 |   top: "conv5-5/dwise"
2177 |   scale_param {
2178 |     bias_term: true
2179 |   }
2180 | }
2181 | 
2182 | # se
2183 | layer {
2184 |   name: "conv5-5/dwise-se-pool"
2185 |   type: "Pooling"
2186 |   bottom: "conv5-5/dwise"
2187 |   top: "conv5-5/dwise-se-pool"
2188 |   pooling_param {
2189 |     pool: AVE
2190 |     global_pooling: true
2191 |   }
2192 | }
2193 | layer {
2194 |   name: "conv5-5/dwise-se-fc1"
2195 |   type: "InnerProduct"
2196 |   bottom: "conv5-5/dwise-se-pool"
2197 |   top:"conv5-5/dwise-se-fc1"
2198 |   param {
2199 |     lr_mult: 1
2200 |     decay_mult: 1
2201 |   }
2202 |   param {
2203 |     lr_mult: 2
2204 |     decay_mult: 0
2205 |   }
2206 |   inner_product_param {
2207 |     num_output: 72
2208 |     weight_filler {
2209 |       type: "msra"
2210 |     }
2211 |     bias_filler {
2212 |       type: "constant"
2213 |       value: 0
2214 |     }
2215 |   }
2216 | }
2217 | layer {
2218 |   name: "conv5-5/dwise-se-fc1-relu"
2219 |   type: "ReLU6"
2220 |   bottom: "conv5-5/dwise-se-fc1"
2221 |   top: "conv5-5/dwise-se-fc1"
2222 | }
2223 | layer {
2224 |   name: "conv5-5/dwise-se-fc2"
2225 |   type: "InnerProduct"
2226 |   bottom: "conv5-5/dwise-se-fc1"
2227 |   top:"conv5-5/dwise-se-fc2"
2228 |   param {
2229 |     lr_mult: 1
2230 |     decay_mult: 1
2231 |   }
2232 |   param {
2233 |     lr_mult: 2
2234 |     decay_mult: 0
2235 |   }
2236 |   inner_product_param {
2237 |     num_output: 288
2238 |     weight_filler {
2239 |       type: "msra"
2240 |     }
2241 |     bias_filler {
2242 |       type: "constant"
2243 |       value: 0
2244 |     }
2245 |   }
2246 | }
2247 | # H-sigmoid
2248 | layer {
2249 |   name: "conv5-5/dwise-se-fc2/shift"
2250 |   type: "Power"
2251 |   bottom: "conv5-5/dwise-se-fc2"
2252 |   top: "conv5-5/dwise-se-fc2/shift"
2253 |   power_param {
2254 |     power: 1
2255 |     scale: 1
2256 |     shift: 3
2257 |   }
2258 | }
2259 | layer {
2260 |   name: "conv5-5/dwise-se-fc2/shift-relu"
2261 |   type: "ReLU6"
2262 |   bottom: "conv5-5/dwise-se-fc2/shift"
2263 |   top: "conv5-5/dwise-se-fc2/shift"
2264 | }
2265 | layer {
2266 |   name: "conv5-5/dwise-se-fc2/shift-div"
2267 |   type: "Power"
2268 |   bottom: "conv5-5/dwise-se-fc2/shift"
2269 |   top: "conv5-5/dwise-se-fc2/shift-div"
2270 |   power_param {
2271 |     power: 1
2272 |     scale: 0.1666666667
2273 |     shift: 0
2274 |   }
2275 | }
2276 | layer {
2277 |   name: "conv5-5/dwise/scale"
2278 |   type: "Scale"
2279 |   bottom: "conv5-5/dwise"
2280 |   bottom: "conv5-5/dwise-se-fc2/shift-div"
2281 |   top: "conv5-5/dwise/scale"
2282 |   scale_param{
2283 |     axis: 0
2284 |   }
2285 | }
2286 | # H-swish
2287 | layer {
2288 |   name: "conv5-5/dwise/shift"
2289 |   type: "Power"
2290 |   bottom: "conv5-5/dwise/scale"
2291 |   top: "conv5-5/dwise/shift"
2292 |   power_param {
2293 |     power: 1
2294 |     scale: 1
2295 |     shift: 3
2296 |   }
2297 | }
2298 | layer {
2299 |   name: "conv5-5/dwise/shift-relu"
2300 |   type: "ReLU6"
2301 |   bottom: "conv5-5/dwise/shift"
2302 |   top: "conv5-5/dwise/shift"
2303 | }
2304 | layer {
2305 |   name: "conv5-5/dwise/shift-div"
2306 |   type: "Power"
2307 |   bottom: "conv5-5/dwise/shift"
2308 |   top: "conv5-5/dwise/shift-div"
2309 |   power_param {
2310 |     power: 1
2311 |     scale: 0.1666666667
2312 |     shift: 0
2313 |   }
2314 | }
2315 | layer {
2316 |   name: "conv5-5/dwise/hswish"
2317 |   type: "Eltwise"
2318 |   bottom: "conv5-5/dwise/scale"
2319 |   bottom: "conv5-5/dwise/shift-div"
2320 |   top: "conv5-5/dwise/hswish"
2321 |   eltwise_param {
2322 |     operation: PROD
2323 |   }
2324 | }
2325 | 
2326 | layer {
2327 |   name: "conv5-5/linear"
2328 |   type: "Convolution"
2329 |   bottom: "conv5-5/dwise/hswish"
2330 |   top: "conv5-5/linear"
2331 |   param {
2332 |     lr_mult: 1
2333 |     decay_mult: 1
2334 |   }
2335 |   convolution_param {
2336 |     num_output: 96
2337 |     pad: 0
2338 |     bias_term: false
2339 |     kernel_size: 1
2340 |     stride: 1
2341 |     weight_filler {
2342 |       type: "msra"
2343 |     }
2344 |   }
2345 | }
2346 | layer {
2347 |   name: "conv5-5/linear-bn"
2348 |   type: "BatchNorm"
2349 |   bottom: "conv5-5/linear"
2350 |   top: "conv5-5/linear"
2351 | }
2352 | layer {
2353 |   name: "conv5-5/linear-bn-scale"
2354 |   type: "Scale"
2355 |   bottom: "conv5-5/linear"
2356 |   top: "conv5-5/linear"
2357 |   scale_param {
2358 |     bias_term: true
2359 |   }
2360 | }
2361 | 
2362 | #################### stage 6-1 ####################
2363 | layer {
2364 |   name: "conv6-1/expand"
2365 |   type: "Convolution"
2366 |   bottom: "conv5-5/linear"
2367 |   top: "conv6-1/expand"
2368 |   param {
2369 |     lr_mult: 1
2370 |     decay_mult: 1
2371 |   }
2372 |   convolution_param {
2373 |     num_output: 576
2374 |     pad: 0
2375 |     bias_term: false
2376 |     kernel_size: 1
2377 |     stride: 1
2378 |     weight_filler {
2379 |       type: "msra"
2380 |     }
2381 |   }
2382 | }
2383 | layer {
2384 |   name: "conv6-1/expand-bn"
2385 |   type: "BatchNorm"
2386 |   bottom: "conv6-1/expand"
2387 |   top: "conv6-1/expand"
2388 | }
2389 | layer {
2390 |   name: "conv6-1/expand-bn-scale"
2391 |   type: "Scale"
2392 |   bottom: "conv6-1/expand"
2393 |   top: "conv6-1/expand"
2394 |   scale_param {
2395 |     bias_term: true
2396 |   }
2397 | }
2398 | # H-swish
2399 | layer {
2400 |   name: "conv6-1/expand/shift"
2401 |   type: "Power"
2402 |   bottom: "conv6-1/expand"
2403 |   top: "conv6-1/expand/shift"
2404 |   power_param {
2405 |     power: 1
2406 |     scale: 1
2407 |     shift: 3
2408 |   }
2409 | }
2410 | layer {
2411 |   name: "conv6-1/expand/shift-relu"
2412 |   type: "ReLU6"
2413 |   bottom: "conv6-1/expand/shift"
2414 |   top: "conv6-1/expand/shift"
2415 | }
2416 | layer {
2417 |   name: "conv6-1/expand/shift-div"
2418 |   type: "Power"
2419 |   bottom: "conv6-1/expand/shift"
2420 |   top: "conv6-1/expand/shift-div"
2421 |   power_param {
2422 |     power: 1
2423 |     scale: 0.1666666667
2424 |     shift: 0
2425 |   }
2426 | }
2427 | layer {
2428 |   name: "conv6-1/expand/hswish"
2429 |   type: "Eltwise"
2430 |   bottom: "conv6-1/expand"
2431 |   bottom: "conv6-1/expand/shift-div"
2432 |   top: "conv6-1/expand/hswish"
2433 |   eltwise_param {
2434 |     operation: PROD
2435 |   }
2436 | }
2437 | 
2438 | layer {
2439 |   name: "conv6-1/dwise"
2440 |   #type: "Convolution"
2441 |   type: "DepthwiseConvolution"
2442 |   bottom: "conv6-1/expand/hswish"
2443 |   top: "conv6-1/dwise"
2444 |   param {
2445 |     lr_mult: 1
2446 |     decay_mult: 1
2447 |   }
2448 |   convolution_param {
2449 |     num_output: 576
2450 |     group: 576
2451 |     pad: 2
2452 |     bias_term: false
2453 |     kernel_size: 5
2454 |     stride: 1
2455 |     weight_filler {
2456 |       type: "msra"
2457 |     }
2458 |     engine: CAFFE
2459 |   }
2460 | }
2461 | layer {
2462 |   name: "conv6-1/dwise-bn"
2463 |   type: "BatchNorm"
2464 |   bottom: "conv6-1/dwise"
2465 |   top: "conv6-1/dwise"
2466 | }
2467 | layer {
2468 |   name: "conv6-1/dwise-bn-scale"
2469 |   type: "Scale"
2470 |   bottom: "conv6-1/dwise"
2471 |   top: "conv6-1/dwise"
2472 |   scale_param {
2473 |     bias_term: true
2474 |   }
2475 | }
2476 | 
2477 | # se
2478 | layer {
2479 |   name: "conv6-1/dwise-se-pool"
2480 |   type: "Pooling"
2481 |   bottom: "conv6-1/dwise"
2482 |   top: "conv6-1/dwise-se-pool"
2483 |   pooling_param {
2484 |     pool: AVE
2485 |     global_pooling: true
2486 |   }
2487 | }
2488 | layer {
2489 |   name: "conv6-1/dwise-se-fc1"
2490 |   type: "InnerProduct"
2491 |   bottom: "conv6-1/dwise-se-pool"
2492 |   top:"conv6-1/dwise-se-fc1"
2493 |   param {
2494 |     lr_mult: 1
2495 |     decay_mult: 1
2496 |   }
2497 |   param {
2498 |     lr_mult: 2
2499 |     decay_mult: 0
2500 |   }
2501 |   inner_product_param {
2502 |     num_output: 144
2503 |     weight_filler {
2504 |       type: "msra"
2505 |     }
2506 |     bias_filler {
2507 |       type: "constant"
2508 |       value: 0
2509 |     }
2510 |   }
2511 | }
2512 | layer {
2513 |   name: "conv6-1/dwise-se-fc1-relu"
2514 |   type: "ReLU6"
2515 |   bottom: "conv6-1/dwise-se-fc1"
2516 |   top: "conv6-1/dwise-se-fc1"
2517 | }
2518 | layer {
2519 |   name: "conv6-1/dwise-se-fc2"
2520 |   type: "InnerProduct"
2521 |   bottom: "conv6-1/dwise-se-fc1"
2522 |   top:"conv6-1/dwise-se-fc2"
2523 |   param {
2524 |     lr_mult: 1
2525 |     decay_mult: 1
2526 |   }
2527 |   param {
2528 |     lr_mult: 2
2529 |     decay_mult: 0
2530 |   }
2531 |   inner_product_param {
2532 |     num_output: 576
2533 |     weight_filler {
2534 |       type: "msra"
2535 |     }
2536 |     bias_filler {
2537 |       type: "constant"
2538 |       value: 0
2539 |     }
2540 |   }
2541 | }
2542 | # H-sigmoid
2543 | layer {
2544 |   name: "conv6-1/dwise-se-fc2/shift"
2545 |   type: "Power"
2546 |   bottom: "conv6-1/dwise-se-fc2"
2547 |   top: "conv6-1/dwise-se-fc2/shift"
2548 |   power_param {
2549 |     power: 1
2550 |     scale: 1
2551 |     shift: 3
2552 |   }
2553 | }
2554 | layer {
2555 |   name: "conv6-1/dwise-se-fc2/shift-relu"
2556 |   type: "ReLU6"
2557 |   bottom: "conv6-1/dwise-se-fc2/shift"
2558 |   top: "conv6-1/dwise-se-fc2/shift"
2559 | }
2560 | layer {
2561 |   name: "conv6-1/dwise-se-fc2/shift-div"
2562 |   type: "Power"
2563 |   bottom: "conv6-1/dwise-se-fc2/shift"
2564 |   top: "conv6-1/dwise-se-fc2/shift-div"
2565 |   power_param {
2566 |     power: 1
2567 |     scale: 0.1666666667
2568 |     shift: 0
2569 |   }
2570 | }
2571 | layer {
2572 |   name: "conv6-1/dwise/scale"
2573 |   type: "Scale"
2574 |   bottom: "conv6-1/dwise"
2575 |   bottom: "conv6-1/dwise-se-fc2/shift-div"
2576 |   top: "conv6-1/dwise/scale"
2577 |   scale_param{
2578 |     axis: 0
2579 |   }
2580 | }
2581 | # H-swish
2582 | layer {
2583 |   name: "conv6-1/dwise/shift"
2584 |   type: "Power"
2585 |   bottom: "conv6-1/dwise/scale"
2586 |   top: "conv6-1/dwise/shift"
2587 |   power_param {
2588 |     power: 1
2589 |     scale: 1
2590 |     shift: 3
2591 |   }
2592 | }
2593 | layer {
2594 |   name: "conv6-1/dwise/shift-relu"
2595 |   type: "ReLU6"
2596 |   bottom: "conv6-1/dwise/shift"
2597 |   top: "conv6-1/dwise/shift"
2598 | }
2599 | layer {
2600 |   name: "conv6-1/dwise/shift-div"
2601 |   type: "Power"
2602 |   bottom: "conv6-1/dwise/shift"
2603 |   top: "conv6-1/dwise/shift-div"
2604 |   power_param {
2605 |     power: 1
2606 |     scale: 0.1666666667
2607 |     shift: 0
2608 |   }
2609 | }
2610 | layer {
2611 |   name: "conv6-1/dwise/hswish"
2612 |   type: "Eltwise"
2613 |   bottom: "conv6-1/dwise/scale"
2614 |   bottom: "conv6-1/dwise/shift-div"
2615 |   top: "conv6-1/dwise/hswish"
2616 |   eltwise_param {
2617 |     operation: PROD
2618 |   }
2619 | }
2620 | 
2621 | layer {
2622 |   name: "conv6-1/linear"
2623 |   type: "Convolution"
2624 |   bottom: "conv6-1/dwise/hswish"
2625 |   top: "conv6-1/linear"
2626 |   param {
2627 |     lr_mult: 1
2628 |     decay_mult: 1
2629 |   }
2630 |   convolution_param {
2631 |     num_output: 96
2632 |     pad: 0
2633 |     bias_term: false
2634 |     kernel_size: 1
2635 |     stride: 1
2636 |     weight_filler {
2637 |       type: "msra"
2638 |     }
2639 |   }
2640 | }
2641 | layer {
2642 |   name: "conv6-1/linear-bn"
2643 |   type: "BatchNorm"
2644 |   bottom: "conv6-1/linear"
2645 |   top: "conv6-1/linear"
2646 | }
2647 | layer {
2648 |   name: "conv6-1/linear-bn-scale"
2649 |   type: "Scale"
2650 |   bottom: "conv6-1/linear"
2651 |   top: "conv6-1/linear"
2652 |   scale_param {
2653 |     bias_term: true
2654 |   }
2655 | }
2656 | 
2657 | layer {
2658 |   name: "block6-1"
2659 |   type: "Eltwise"
2660 |   bottom: "conv5-5/linear"
2661 |   bottom: "conv6-1/linear"
2662 |   top: "block6-1"
2663 | }
2664 | 
2665 | #################### stage 6-2 ####################
2666 | layer {
2667 |   name: "conv6-2/expand"
2668 |   type: "Convolution"
2669 |   bottom: "block6-1"
2670 |   top: "conv6-2/expand"
2671 |   param {
2672 |     lr_mult: 1
2673 |     decay_mult: 1
2674 |   }
2675 |   convolution_param {
2676 |     num_output: 576
2677 |     pad: 0
2678 |     bias_term: false
2679 |     kernel_size: 1
2680 |     stride: 1
2681 |     weight_filler {
2682 |       type: "msra"
2683 |     }
2684 |   }
2685 | }
2686 | layer {
2687 |   name: "conv6-2/expand-bn"
2688 |   type: "BatchNorm"
2689 |   bottom: "conv6-2/expand"
2690 |   top: "conv6-2/expand"
2691 | }
2692 | layer {
2693 |   name: "conv6-2/expand-bn-scale"
2694 |   type: "Scale"
2695 |   bottom: "conv6-2/expand"
2696 |   top: "conv6-2/expand"
2697 |   scale_param {
2698 |     bias_term: true
2699 |   }
2700 | }
2701 | # H-swish
2702 | layer {
2703 |   name: "conv6-2/expand/shift"
2704 |   type: "Power"
2705 |   bottom: "conv6-2/expand"
2706 |   top: "conv6-2/expand/shift"
2707 |   power_param {
2708 |     power: 1
2709 |     scale: 1
2710 |     shift: 3
2711 |   }
2712 | }
2713 | layer {
2714 |   name: "conv6-2/expand/shift-relu"
2715 |   type: "ReLU6"
2716 |   bottom: "conv6-2/expand/shift"
2717 |   top: "conv6-2/expand/shift"
2718 | }
2719 | layer {
2720 |   name: "conv6-2/expand/shift-div"
2721 |   type: "Power"
2722 |   bottom: "conv6-2/expand/shift"
2723 |   top: "conv6-2/expand/shift-div"
2724 |   power_param {
2725 |     power: 1
2726 |     scale: 0.1666666667
2727 |     shift: 0
2728 |   }
2729 | }
2730 | layer {
2731 |   name: "conv6-2/expand/hswish"
2732 |   type: "Eltwise"
2733 |   bottom: "conv6-2/expand"
2734 |   bottom: "conv6-2/expand/shift-div"
2735 |   top: "conv6-2/expand/hswish"
2736 |   eltwise_param {
2737 |     operation: PROD
2738 |   }
2739 | }
2740 | 
2741 | layer {
2742 |   name: "conv6-2/dwise"
2743 |   #type: "Convolution"
2744 |   type: "DepthwiseConvolution"
2745 |   bottom: "conv6-2/expand/hswish"
2746 |   top: "conv6-2/dwise"
2747 |   param {
2748 |     lr_mult: 1
2749 |     decay_mult: 1
2750 |   }
2751 |   convolution_param {
2752 |     num_output: 576
2753 |     group: 576
2754 |     pad: 2
2755 |     bias_term: false
2756 |     kernel_size: 5
2757 |     stride: 1
2758 |     weight_filler {
2759 |       type: "msra"
2760 |     }
2761 |     engine: CAFFE
2762 |   }
2763 | }
2764 | layer {
2765 |   name: "conv6-2/dwise-bn"
2766 |   type: "BatchNorm"
2767 |   bottom: "conv6-2/dwise"
2768 |   top: "conv6-2/dwise"
2769 | }
2770 | layer {
2771 |   name: "conv6-2/dwise-bn-scale"
2772 |   type: "Scale"
2773 |   bottom: "conv6-2/dwise"
2774 |   top: "conv6-2/dwise"
2775 |   scale_param {
2776 |     bias_term: true
2777 |   }
2778 | }
2779 | 
2780 | # se
2781 | layer {
2782 |   name: "conv6-2/dwise-se-pool"
2783 |   type: "Pooling"
2784 |   bottom: "conv6-2/dwise"
2785 |   top: "conv6-2/dwise-se-pool"
2786 |   pooling_param {
2787 |     pool: AVE
2788 |     global_pooling: true
2789 |   }
2790 | }
2791 | layer {
2792 |   name: "conv6-2/dwise-se-fc1"
2793 |   type: "InnerProduct"
2794 |   bottom: "conv6-2/dwise-se-pool"
2795 |   top:"conv6-2/dwise-se-fc1"
2796 |   param {
2797 |     lr_mult: 1
2798 |     decay_mult: 1
2799 |   }
2800 |   param {
2801 |     lr_mult: 2
2802 |     decay_mult: 0
2803 |   }
2804 |   inner_product_param {
2805 |     num_output: 144
2806 |     weight_filler {
2807 |       type: "msra"
2808 |     }
2809 |     bias_filler {
2810 |       type: "constant"
2811 |       value: 0
2812 |     }
2813 |   }
2814 | }
2815 | layer {
2816 |   name: "conv6-2/dwise-se-fc1-relu"
2817 |   type: "ReLU6"
2818 |   bottom: "conv6-2/dwise-se-fc1"
2819 |   top: "conv6-2/dwise-se-fc1"
2820 | }
2821 | layer {
2822 |   name: "conv6-2/dwise-se-fc2"
2823 |   type: "InnerProduct"
2824 |   bottom: "conv6-2/dwise-se-fc1"
2825 |   top:"conv6-2/dwise-se-fc2"
2826 |   param {
2827 |     lr_mult: 1
2828 |     decay_mult: 1
2829 |   }
2830 |   param {
2831 |     lr_mult: 2
2832 |     decay_mult: 0
2833 |   }
2834 |   inner_product_param {
2835 |     num_output: 576
2836 |     weight_filler {
2837 |       type: "msra"
2838 |     }
2839 |     bias_filler {
2840 |       type: "constant"
2841 |       value: 0
2842 |     }
2843 |   }
2844 | }
2845 | # H-sigmoid
2846 | layer {
2847 |   name: "conv6-2/dwise-se-fc2/shift"
2848 |   type: "Power"
2849 |   bottom: "conv6-2/dwise-se-fc2"
2850 |   top: "conv6-2/dwise-se-fc2/shift"
2851 |   power_param {
2852 |     power: 1
2853 |     scale: 1
2854 |     shift: 3
2855 |   }
2856 | }
2857 | layer {
2858 |   name: "conv6-2/dwise-se-fc2/shift-relu"
2859 |   type: "ReLU6"
2860 |   bottom: "conv6-2/dwise-se-fc2/shift"
2861 |   top: "conv6-2/dwise-se-fc2/shift"
2862 | }
2863 | layer {
2864 |   name: "conv6-2/dwise-se-fc2/shift-div"
2865 |   type: "Power"
2866 |   bottom: "conv6-2/dwise-se-fc2/shift"
2867 |   top: "conv6-2/dwise-se-fc2/shift-div"
2868 |   power_param {
2869 |     power: 1
2870 |     scale: 0.1666666667
2871 |     shift: 0
2872 |   }
2873 | }
2874 | layer {
2875 |   name: "conv6-2/dwise/scale"
2876 |   type: "Scale"
2877 |   bottom: "conv6-2/dwise"
2878 |   bottom: "conv6-2/dwise-se-fc2/shift-div"
2879 |   top: "conv6-2/dwise/scale"
2880 |   scale_param{
2881 |     axis: 0
2882 |   }
2883 | }
2884 | # H-swish
2885 | layer {
2886 |   name: "conv6-2/dwise/shift"
2887 |   type: "Power"
2888 |   bottom: "conv6-2/dwise/scale"
2889 |   top: "conv6-2/dwise/shift"
2890 |   power_param {
2891 |     power: 1
2892 |     scale: 1
2893 |     shift: 3
2894 |   }
2895 | }
2896 | layer {
2897 |   name: "conv6-2/dwise/shift-relu"
2898 |   type: "ReLU6"
2899 |   bottom: "conv6-2/dwise/shift"
2900 |   top: "conv6-2/dwise/shift"
2901 | }
2902 | layer {
2903 |   name: "conv6-2/dwise/shift-div"
2904 |   type: "Power"
2905 |   bottom: "conv6-2/dwise/shift"
2906 |   top: "conv6-2/dwise/shift-div"
2907 |   power_param {
2908 |     power: 1
2909 |     scale: 0.1666666667
2910 |     shift: 0
2911 |   }
2912 | }
2913 | layer {
2914 |   name: "conv6-2/dwise/hswish"
2915 |   type: "Eltwise"
2916 |   bottom: "conv6-2/dwise/scale"
2917 |   bottom: "conv6-2/dwise/shift-div"
2918 |   top: "conv6-2/dwise/hswish"
2919 |   eltwise_param {
2920 |     operation: PROD
2921 |   }
2922 | }
2923 | 
2924 | layer {
2925 |   name: "conv6-2/linear"
2926 |   type: "Convolution"
2927 |   bottom: "conv6-2/dwise/hswish"
2928 |   top: "conv6-2/linear"
2929 |   param {
2930 |     lr_mult: 1
2931 |     decay_mult: 1
2932 |   }
2933 |   convolution_param {
2934 |     num_output: 96
2935 |     pad: 0
2936 |     bias_term: false
2937 |     kernel_size: 1
2938 |     stride: 1
2939 |     weight_filler {
2940 |       type: "msra"
2941 |     }
2942 |   }
2943 | }
2944 | layer {
2945 |   name: "conv6-2/linear-bn"
2946 |   type: "BatchNorm"
2947 |   bottom: "conv6-2/linear"
2948 |   top: "conv6-2/linear"
2949 | }
2950 | layer {
2951 |   name: "conv6-2/linear-bn-scale"
2952 |   type: "Scale"
2953 |   bottom: "conv6-2/linear"
2954 |   top: "conv6-2/linear"
2955 |   scale_param {
2956 |     bias_term: true
2957 |   }
2958 | }
2959 | 
2960 | layer {
2961 |   name: "block6-2"
2962 |   type: "Eltwise"
2963 |   bottom: "block6-1"
2964 |   bottom: "conv6-2/linear"
2965 |   top: "block6-2"
2966 | }
2967 | 
2968 | #################### stage 6-3 ####################
2969 | layer {
2970 |   name: "conv6-3"
2971 |   type: "Convolution"
2972 |   bottom: "block6-2"
2973 |   top: "conv6-3"
2974 |   param {
2975 |     lr_mult: 1
2976 |     decay_mult: 1
2977 |   }
2978 |   convolution_param {
2979 |     num_output: 576
2980 |     pad: 0
2981 |     bias_term: false
2982 |     kernel_size: 1
2983 |     stride: 1
2984 |     weight_filler {
2985 |       type: "msra"
2986 |     }
2987 |   }
2988 | }
2989 | layer {
2990 |   name: "conv6-3-bn"
2991 |   type: "BatchNorm"
2992 |   bottom: "conv6-3"
2993 |   top: "conv6-3"
2994 | }
2995 | layer {
2996 |   name: "conv6-3-bn-scale"
2997 |   type: "Scale"
2998 |   bottom: "conv6-3"
2999 |   top: "conv6-3"
3000 |   scale_param {
3001 |     bias_term: true
3002 |   }
3003 | }
3004 | 
3005 | # se
3006 | layer {
3007 |   name: "conv6-3-se-pool"
3008 |   type: "Pooling"
3009 |   bottom: "conv6-3"
3010 |   top: "conv6-3-se-pool"
3011 |   pooling_param {
3012 |     pool: AVE
3013 |     global_pooling: true
3014 |   }
3015 | }
3016 | layer {
3017 |   name: "conv6-3-se-fc1"
3018 |   type: "InnerProduct"
3019 |   bottom: "conv6-3-se-pool"
3020 |   top:"conv6-3-se-fc1"
3021 |   param {
3022 |     lr_mult: 1
3023 |     decay_mult: 1
3024 |   }
3025 |   param {
3026 |     lr_mult: 2
3027 |     decay_mult: 0
3028 |   }
3029 |   inner_product_param {
3030 |     num_output: 144
3031 |     weight_filler {
3032 |       type: "msra"
3033 |     }
3034 |     bias_filler {
3035 |       type: "constant"
3036 |       value: 0
3037 |     }
3038 |   }
3039 | }
3040 | layer {
3041 |   name: "conv6-3-se-fc1-relu"
3042 |   type: "ReLU6"
3043 |   bottom: "conv6-3-se-fc1"
3044 |   top: "conv6-3-se-fc1"
3045 | }
3046 | layer {
3047 |   name: "conv6-3-se-fc2"
3048 |   type: "InnerProduct"
3049 |   bottom: "conv6-3-se-fc1"
3050 |   top:"conv6-3-se-fc2"
3051 |   param {
3052 |     lr_mult: 1
3053 |     decay_mult: 1
3054 |   }
3055 |   param {
3056 |     lr_mult: 2
3057 |     decay_mult: 0
3058 |   }
3059 |   inner_product_param {
3060 |     num_output: 576
3061 |     weight_filler {
3062 |       type: "msra"
3063 |     }
3064 |     bias_filler {
3065 |       type: "constant"
3066 |       value: 0
3067 |     }
3068 |   }
3069 | }
3070 | # H-sigmoid
3071 | layer {
3072 |   name: "conv6-3-se-fc2/shift"
3073 |   type: "Power"
3074 |   bottom: "conv6-3-se-fc2"
3075 |   top: "conv6-3-se-fc2/shift"
3076 |   power_param {
3077 |     power: 1
3078 |     scale: 1
3079 |     shift: 3
3080 |   }
3081 | }
3082 | layer {
3083 |   name: "conv6-3-se-fc2/shift-relu"
3084 |   type: "ReLU6"
3085 |   bottom: "conv6-3-se-fc2/shift"
3086 |   top: "conv6-3-se-fc2/shift"
3087 | }
3088 | layer {
3089 |   name: "conv6-3-se-fc2/shift-div"
3090 |   type: "Power"
3091 |   bottom: "conv6-3-se-fc2/shift"
3092 |   top: "conv6-3-se-fc2/shift-div"
3093 |   power_param {
3094 |     power: 1
3095 |     scale: 0.1666666667
3096 |     shift: 0
3097 |   }
3098 | }
3099 | layer {
3100 |   name: "conv6-3/scale"
3101 |   type: "Scale"
3102 |   bottom: "conv6-3"
3103 |   bottom: "conv6-3-se-fc2/shift-div"
3104 |   top: "conv6-3/scale"
3105 |   scale_param{
3106 |     axis: 0
3107 |   }
3108 | }
3109 | # H-swish
3110 | layer {
3111 |   name: "conv6-3/shift"
3112 |   type: "Power"
3113 |   bottom: "conv6-3/scale"
3114 |   top: "conv6-3/shift"
3115 |   power_param {
3116 |     power: 1
3117 |     scale: 1
3118 |     shift: 3
3119 |   }
3120 | }
3121 | layer {
3122 |   name: "conv6-3/shift-relu"
3123 |   type: "ReLU6"
3124 |   bottom: "conv6-3/shift"
3125 |   top: "conv6-3/shift"
3126 | }
3127 | layer {
3128 |   name: "conv6-3/shift-div"
3129 |   type: "Power"
3130 |   bottom: "conv6-3/shift"
3131 |   top: "conv6-3/shift-div"
3132 |   power_param {
3133 |     power: 1
3134 |     scale: 0.1666666667
3135 |     shift: 0
3136 |   }
3137 | }
3138 | layer {
3139 |   name: "conv6-3/hswish"
3140 |   type: "Eltwise"
3141 |   bottom: "conv6-3/scale"
3142 |   bottom: "conv6-3/shift-div"
3143 |   top: "conv6-3/hswish"
3144 |   eltwise_param {
3145 |     operation: PROD
3146 |   }
3147 | }
3148 | 
3149 | #################### stage 7-1 ####################
3150 | layer {
3151 |   name: "pool7-1"
3152 |   type: "Pooling"
3153 |   bottom: "conv6-3/hswish"
3154 |   top: "pool7-1"
3155 |   pooling_param {
3156 |     pool: AVE
3157 |     global_pooling: true
3158 |   }
3159 | }
3160 | # H-swish
3161 | layer {
3162 |   name: "pool7-1/shift"
3163 |   type: "Power"
3164 |   bottom: "pool7-1"
3165 |   top: "pool7-1/shift"
3166 |   power_param {
3167 |     power: 1
3168 |     scale: 1
3169 |     shift: 3
3170 |   }
3171 | }
3172 | layer {
3173 |   name: "pool7-1/shift-relu"
3174 |   type: "ReLU6"
3175 |   bottom: "pool7-1/shift"
3176 |   top: "pool7-1/shift"
3177 | }
3178 | layer {
3179 |   name: "pool7-1/shift-div"
3180 |   type: "Power"
3181 |   bottom: "pool7-1/shift"
3182 |   top: "pool7-1/shift-div"
3183 |   power_param {
3184 |     power: 1
3185 |     scale: 0.1666666667
3186 |     shift: 0
3187 |   }
3188 | }
3189 | layer {
3190 |   name: "pool7-1/hswish"
3191 |   type: "Eltwise"
3192 |   bottom: "pool7-1"
3193 |   bottom: "pool7-1/shift-div"
3194 |   top: "pool7-1/hswish"
3195 |   eltwise_param {
3196 |     operation: PROD
3197 |   }
3198 | }
3199 | 
3200 | #################### stage 7-2 ####################
3201 | layer {
3202 |   name: "conv7-2"
3203 |   type: "Convolution"
3204 |   bottom: "pool7-1/hswish"
3205 |   top: "conv7-2"
3206 |   param {
3207 |     lr_mult: 1
3208 |     decay_mult: 1
3209 |   }
3210 |   convolution_param {
3211 |     num_output: 1280
3212 |     pad: 0
3213 |     bias_term: false
3214 |     kernel_size: 1
3215 |     stride: 1
3216 |     weight_filler {
3217 |       type: "msra"
3218 |     }
3219 |   }
3220 | }
3221 | layer {
3222 |   name: "conv7-2-bn"
3223 |   type: "BatchNorm"
3224 |   bottom: "conv7-2"
3225 |   top: "conv7-2"
3226 | }
3227 | layer {
3228 |   name: "conv7-2-bn-scale"
3229 |   type: "Scale"
3230 |   bottom: "conv7-2"
3231 |   top: "conv7-2"
3232 |   scale_param {
3233 |     bias_term: true
3234 |   }
3235 | }
3236 | # H-swish
3237 | layer {
3238 |   name: "conv7-2/shift"
3239 |   type: "Power"
3240 |   bottom: "conv7-2"
3241 |   top: "conv7-2/shift"
3242 |   power_param {
3243 |     power: 1
3244 |     scale: 1
3245 |     shift: 3
3246 |   }
3247 | }
3248 | layer {
3249 |   name: "conv7-2/shift-relu"
3250 |   type: "ReLU6"
3251 |   bottom: "conv7-2/shift"
3252 |   top: "conv7-2/shift"
3253 | }
3254 | layer {
3255 |   name: "conv7-2/shift-div"
3256 |   type: "Power"
3257 |   bottom: "conv7-2/shift"
3258 |   top: "conv7-2/shift-div"
3259 |   power_param {
3260 |     power: 1
3261 |     scale: 0.1666666667
3262 |     shift: 0
3263 |   }
3264 | }
3265 | layer {
3266 |   name: "conv7-2/hswish"
3267 |   type: "Eltwise"
3268 |   bottom: "conv7-2"
3269 |   bottom: "conv7-2/shift-div"
3270 |   top: "conv7-2/hswish"
3271 |   eltwise_param {
3272 |     operation: PROD
3273 |   }
3274 | }
3275 | 
3276 | #################### output ####################
3277 | layer {
3278 |   name: "conv7-3"
3279 |   type: "Convolution"
3280 |   bottom: "conv7-2/hswish"
3281 |   top: "conv7-3"
3282 |   param {
3283 |     lr_mult: 1
3284 |     decay_mult: 1
3285 |   }
3286 |   convolution_param {
3287 |     num_output: 1000
3288 |     pad: 0
3289 |     bias_term: false
3290 |     kernel_size: 1
3291 |     stride: 1
3292 |     weight_filler {
3293 |       type: "msra"
3294 |     }
3295 |   }
3296 | }
3297 | layer {
3298 |   name: "conv7-3-bn"
3299 |   type: "BatchNorm"
3300 |   bottom: "conv7-3"
3301 |   top: "conv7-3"
3302 | }
3303 | layer {
3304 |   name: "conv7-3-bn-scale"
3305 |   type: "Scale"
3306 |   bottom: "conv7-3"
3307 |   top: "conv7-3"
3308 |   scale_param {
3309 |     bias_term: true
3310 |   }
3311 | }
3312 | # H-swish
3313 | layer {
3314 |   name: "conv7-3/shift"
3315 |   type: "Power"
3316 |   bottom: "conv7-3"
3317 |   top: "conv7-3/shift"
3318 |   power_param {
3319 |     power: 1
3320 |     scale: 1
3321 |     shift: 3
3322 |   }
3323 | }
3324 | layer {
3325 |   name: "conv7-3/shift-relu"
3326 |   type: "ReLU6"
3327 |   bottom: "conv7-3/shift"
3328 |   top: "conv7-3/shift"
3329 | }
3330 | layer {
3331 |   name: "conv7-3/shift-div"
3332 |   type: "Power"
3333 |   bottom: "conv7-3/shift"
3334 |   top: "conv7-3/shift-div"
3335 |   power_param {
3336 |     power: 1
3337 |     scale: 0.1666666667
3338 |     shift: 0
3339 |   }
3340 | }
3341 | layer {
3342 |   name: "conv7-3/hswish"
3343 |   type: "Eltwise"
3344 |   bottom: "conv7-3"
3345 |   bottom: "conv7-3/shift-div"
3346 |   top: "conv7-3/hswish"
3347 |   eltwise_param {
3348 |     operation: PROD
3349 |   }
3350 | }
3351 | 
3352 | layer {
3353 |   name: "prob"
3354 |   type: "Softmax"
3355 |   bottom: "conv7-3/hswish"
3356 |   top: "prob"
3357 | }
3358 | 


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