├── README.md
├── data
    └── VOC0712
    │   └── labelmap_voc.prototxt
├── demo.py
├── image
    ├── result.jpeg
    ├── slam.jpeg
    ├── table0.png
    └── table1.jpg
├── model
    ├── _iter_180000.caffemodel
    ├── _iter_290000.caffemodel
    ├── tiny_ssd_voc_test.prototxt
    └── tiny_ssd_voc_train.prototxt
├── python_script
    ├── deploy.prototxt
    ├── generate.py
    └── train.prototxt
└── solver.prototxt


/README.md:
--------------------------------------------------------------------------------
 1 | # TinySSD Implementation @ ULSee Inc.
 2 | ### Prospective
 3 | We are looking forward to have a discussion with anyone who are interested in object detection and try to develop better faster algorithms. As the authors claimed, the tiny-ssd could achieve 61.3% on VOC2007, but our released version could only get an accuracy at 59.9%(we re-train and get 61.5%, may be released in future). Therefore, we are releasing our implementation and expect some common researches and discussions. 
 4 | ### Description  
 5 | This is an implementation of the following released paper by ULSee Inc.. It is a simplified structure for SSD(Single Shot MultiBox Detector). 
 6 | 
 7 | >@article{DBLP:journals/corr/abs-1802-06488, </br>
 8 |   author    = {Alexander Wong and  </br>
 9 |                Mohammad Javad Shafiee and </br>
10 |                Francis Li and </br>
11 |                Brendan Chwyl}, </br>
12 |   title     = {Tiny {SSD:} {A} Tiny Single-shot Detection Deep Convolutional Neural </br>
13 |                Network for Real-time Embedded Object Detection}, </br>
14 |   journal   = {CoRR}, </br>
15 |   volume    = {abs/1802.06488}, </br>
16 |   year      = {2018}, </br>
17 |   url       = {http://arxiv.org/abs/1802.06488}, </br>
18 |   archivePrefix = {arXiv}, </br>
19 |   eprint    = {1802.06488}, </br>
20 |   timestamp = {Thu, 01 Mar 2018 19:20:48 +0100}, </br>
21 |   biburl    = {https://dblp.org/rec/bib/journals/corr/abs-1802-06488}, </br>
22 |   bibsource = {dblp computer science bibliography, https://dblp.org} </br>
23 | }
24 | 
25 | **One Revision**
26 | Table II in the paper is not consistent with Table I and Figure 3. It should be revised as:
27 | 
28 | ![origin image](image/table0.png) => ![revision image](image/table1.jpg)
29 | 
30 | In our prototxt, we add every fire module id by 1. The fire1 in paper is our fire2 in protoxt. Therefore, the paper describes fire module from fire1, fire2, fire3 ... to fire10, and our supplied prototxt is from fire2, fire3, fire4 ... to fire11. 
31 | 
32 | ### Usage
33 |  1. Download ssd-caffe and compile the pycaffe
34 |  
35 |  > git clone -b ssd https://github.com/weiliu89/caffe.git 
36 |  2. Download our repository
37 |  
38 |  > git clone https://github.com/lampsonSong/tinySSD.git
39 |  
40 |  3. The hierarchy of tinySSD is:
41 | * **data** 
42 | * **model**
43 |     * *_iter_180000.caffemodel -> *pretrain model (in **our prototxt**, we use fire2 ~ fire9)*
44 |     * *_iter_290000.caffemodel -> *final model*
45 |     * tiny_ssd_voc_train.prototxt -> *train.prototxt, please modify data path*
46 |     * tiny_ssd_voc_test.prototxt -> *test.prototxt, please modify data path*
47 | * **python_script**
48 |     * *generate.py  -> *you can generate deploy.prototxt and train.prototxt using this. Custome the corresponding path including **caffe, your project, your train lmdb*** 
49 |     * *deploy.prototxt -> *our deploy prototxt*
50 |     * *train.prototxt -> *our train prototxt*
51 | * **image**
52 |     * *slam.jpeg -> *test image*
53 |     * *result.jpeg -> *test result*
54 |     * *table0.png -> *table from paper*
55 |     * *table1.jpg -> *table after revison*
56 | * **README.md**
57 | * **solver.prototxt** 
58 |  4. At the tinySSD directory, run
59 |  
60 | > python python_script/generate.py
61 | 
62 | 
63 | to get *deploy.txt* and *train.prototxt* under *python_script*
64 | 
65 |  5. revise the demo.py to check the result
66 | > python demo.py
67 | 
68 | ![origin image](image/slam.jpeg) => ![revision image](image/result.jpeg)
69 | 
70 | 
71 | 
72 | 
73 | 


--------------------------------------------------------------------------------
/data/VOC0712/labelmap_voc.prototxt:
--------------------------------------------------------------------------------
  1 | item {
  2 |   name: "none_of_the_above"
  3 |   label: 0
  4 |   display_name: "background"
  5 | }
  6 | item {
  7 |   name: "aeroplane"
  8 |   label: 1
  9 |   display_name: "aeroplane"
 10 | }
 11 | item {
 12 |   name: "bicycle"
 13 |   label: 2
 14 |   display_name: "bicycle"
 15 | }
 16 | item {
 17 |   name: "bird"
 18 |   label: 3
 19 |   display_name: "bird"
 20 | }
 21 | item {
 22 |   name: "boat"
 23 |   label: 4
 24 |   display_name: "boat"
 25 | }
 26 | item {
 27 |   name: "bottle"
 28 |   label: 5
 29 |   display_name: "bottle"
 30 | }
 31 | item {
 32 |   name: "bus"
 33 |   label: 6
 34 |   display_name: "bus"
 35 | }
 36 | item {
 37 |   name: "car"
 38 |   label: 7
 39 |   display_name: "car"
 40 | }
 41 | item {
 42 |   name: "cat"
 43 |   label: 8
 44 |   display_name: "cat"
 45 | }
 46 | item {
 47 |   name: "chair"
 48 |   label: 9
 49 |   display_name: "chair"
 50 | }
 51 | item {
 52 |   name: "cow"
 53 |   label: 10
 54 |   display_name: "cow"
 55 | }
 56 | item {
 57 |   name: "diningtable"
 58 |   label: 11
 59 |   display_name: "diningtable"
 60 | }
 61 | item {
 62 |   name: "dog"
 63 |   label: 12
 64 |   display_name: "dog"
 65 | }
 66 | item {
 67 |   name: "horse"
 68 |   label: 13
 69 |   display_name: "horse"
 70 | }
 71 | item {
 72 |   name: "motorbike"
 73 |   label: 14
 74 |   display_name: "motorbike"
 75 | }
 76 | item {
 77 |   name: "person"
 78 |   label: 15
 79 |   display_name: "person"
 80 | }
 81 | item {
 82 |   name: "pottedplant"
 83 |   label: 16
 84 |   display_name: "pottedplant"
 85 | }
 86 | item {
 87 |   name: "sheep"
 88 |   label: 17
 89 |   display_name: "sheep"
 90 | }
 91 | item {
 92 |   name: "sofa"
 93 |   label: 18
 94 |   display_name: "sofa"
 95 | }
 96 | item {
 97 |   name: "train"
 98 |   label: 19
 99 |   display_name: "train"
100 | }
101 | item {
102 |   name: "tvmonitor"
103 |   label: 20
104 |   display_name: "tvmonitor"
105 | }
106 | 


--------------------------------------------------------------------------------
/demo.py:
--------------------------------------------------------------------------------
 1 | # created by lampson.song @ 2018-07-18 
 2 | # a demo script of tinySSD
 3 | 
 4 | import numpy as np  
 5 | import sys,os  
 6 | import cv2
 7 | import time
 8 | caffe_root = '/home/lampson/1T_disk/workspace/objectDetection/ssd-caffe/'
 9 | sys.path.insert(0, caffe_root + 'python')  
10 | import caffe  
11 | from skimage import img_as_float
12 | 
13 | net_file= './python_script/deploy.prototxt'  
14 | if not os.path.exists(net_file):
15 |     print("prototxt does not exist, use generate.py.")
16 |     exit()
17 | 
18 | caffe_model='./model/_iter_290000.caffemodel'  
19 | if not os.path.exists(caffe_model):
20 |     print("model does not exist.")
21 |     exit()
22 | 
23 | CLASSES = ('background',
24 |            'aeroplane', 'bicycle', 'bird', 'boat',
25 |            'bottle', 'bus', 'car', 'cat', 'chair',
26 |            'cow', 'diningtable', 'dog', 'horse',
27 |            'motorbike', 'person', 'pottedplant',
28 |            'sheep', 'sofa', 'train', 'tvmonitor')
29 | 
30 | 
31 | def postprocess(img, out):
32 |     h = img.shape[0]
33 |     w = img.shape[1]
34 |     box = out['detection_out'][0,0,:,3:7] * np.array([w, h, w, h])
35 | 
36 |     cls = out['detection_out'][0,0,:,1]
37 |     conf = out['detection_out'][0,0,:,2]
38 |     return (box.astype(np.int32), conf, cls)
39 | 
40 | def detect(imgfile):
41 |     orig_img = cv2.imread(imgfile)
42 | 
43 |     img = orig_img[:,:,::-1]
44 |     img = img_as_float(img).astype(np.float32)
45 |     img = transformer.preprocess('data', img)
46 | 
47 |     net.blobs['data'].data[...] = img
48 |     out = net.forward()
49 | 
50 |     box, conf, cls = postprocess(orig_img, out)
51 | 
52 |     for i in range(len(box)):
53 |        p1 = (box[i][0], box[i][1])
54 |        p2 = (box[i][2], box[i][3])
55 |        p3 = (max(p1[0], 15), max(p1[1], 15))
56 | 
57 |        title = "%s:%.2f" % (CLASSES[int(cls[i])], conf[i])
58 |        cv2.putText(orig_img, title, p3, cv2.FONT_ITALIC, 0.6, (0, 255, 0), 1)
59 |        cv2.rectangle(orig_img, p1, p2, (0,255,0))
60 | 
61 |     cv2.imshow("tiny-SSD", orig_img)
62 |     cv2.waitKey(0)
63 | 
64 | 
65 | net = caffe.Net(net_file,caffe_model,caffe.TEST)  
66 | net.blobs['data'].reshape(1,3,300,300)
67 | 
68 | transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
69 | transformer.set_transpose('data', (2, 0, 1)) # H * W * C -> C * H * W
70 | transformer.set_mean('data', np.array([104, 117, 123]))
71 | transformer.set_raw_scale('data', 255) # [0, 1] -> [0, 255]
72 | transformer.set_channel_swap('data', (2, 1, 0)) # RGB -> BGR
73 | 
74 | detect("image/slam.jpeg")
75 | 


--------------------------------------------------------------------------------
/image/result.jpeg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/lampsonSong/tinySSD/b5b262e90c036e0cc535959efac73e83ff20f7fc/image/result.jpeg


--------------------------------------------------------------------------------
/image/slam.jpeg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/lampsonSong/tinySSD/b5b262e90c036e0cc535959efac73e83ff20f7fc/image/slam.jpeg


--------------------------------------------------------------------------------
/image/table0.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/lampsonSong/tinySSD/b5b262e90c036e0cc535959efac73e83ff20f7fc/image/table0.png


--------------------------------------------------------------------------------
/image/table1.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/lampsonSong/tinySSD/b5b262e90c036e0cc535959efac73e83ff20f7fc/image/table1.jpg


--------------------------------------------------------------------------------
/model/_iter_180000.caffemodel:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/lampsonSong/tinySSD/b5b262e90c036e0cc535959efac73e83ff20f7fc/model/_iter_180000.caffemodel


--------------------------------------------------------------------------------
/model/_iter_290000.caffemodel:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/lampsonSong/tinySSD/b5b262e90c036e0cc535959efac73e83ff20f7fc/model/_iter_290000.caffemodel


--------------------------------------------------------------------------------
/model/tiny_ssd_voc_test.prototxt:
--------------------------------------------------------------------------------
   1 | name: "tiny-ssd"
   2 | layer {
   3 |   name: "data"
   4 |   type: "AnnotatedData"
   5 |   top: "data"
   6 |   top: "label"
   7 |   include {
   8 |     phase: TEST
   9 |   }
  10 |   transform_param {
  11 |     mean_value: 104
  12 |     mean_value: 117
  13 |     mean_value: 123
  14 |     resize_param {
  15 |       prob: 1.0
  16 |       resize_mode: WARP
  17 |       height: 300
  18 |       width: 300
  19 |       interp_mode: LINEAR
  20 |     }
  21 |   }
  22 |   data_param {
  23 |     source: "examples/VOC0712/VOC0712_test_lmdb"
  24 |     batch_size: 8
  25 |     backend: LMDB
  26 |   }
  27 |   annotated_data_param {
  28 |     batch_sampler {
  29 |     }
  30 |     label_map_file: "data/VOC0712/labelmap.prototxt"
  31 |   }
  32 | }
  33 | layer {
  34 |   name: "conv1"
  35 |   type: "Convolution"
  36 |   bottom: "data"
  37 |   top: "conv1"
  38 |   param {
  39 |     lr_mult: 1.0
  40 |     decay_mult: 0.0
  41 |   }
  42 |   param {
  43 |     lr_mult: 1.0
  44 |     decay_mult: 0.0
  45 |   }
  46 |   convolution_param {
  47 |     num_output: 57
  48 |     kernel_size: 3
  49 |     stride: 2
  50 |     weight_filler {
  51 |       type: "xavier"
  52 |     }
  53 |   }
  54 | }
  55 | layer {
  56 |   name: "relu_conv1"
  57 |   type: "ReLU"
  58 |   bottom: "conv1"
  59 |   top: "conv1"
  60 | }
  61 | layer {
  62 |   name: "pool1"
  63 |   type: "Pooling"
  64 |   bottom: "conv1"
  65 |   top: "pool1"
  66 |   pooling_param {
  67 |     pool: MAX
  68 |     kernel_size: 3
  69 |     stride: 2
  70 |   }
  71 | }
  72 | layer {
  73 |   name: "fire2/squeeze1x1"
  74 |   type: "Convolution"
  75 |   bottom: "pool1"
  76 |   top: "fire2/squeeze1x1"
  77 |   param {
  78 |     lr_mult: 1.0
  79 |     decay_mult: 0.0
  80 |   }
  81 |   param {
  82 |     lr_mult: 1.0
  83 |     decay_mult: 0.0
  84 |   }
  85 |   convolution_param {
  86 |     num_output: 15
  87 |     kernel_size: 1
  88 |     weight_filler {
  89 |       type: "xavier"
  90 |     }
  91 |   }
  92 | }
  93 | layer {
  94 |   name: "fire2/expand1x1"
  95 |   type: "Convolution"
  96 |   bottom: "fire2/squeeze1x1"
  97 |   top: "fire2/expand1x1"
  98 |   param {
  99 |     lr_mult: 1.0
 100 |     decay_mult: 0.0
 101 |   }
 102 |   param {
 103 |     lr_mult: 1.0
 104 |     decay_mult: 0.0
 105 |   }
 106 |   convolution_param {
 107 |     num_output: 49
 108 |     kernel_size: 1
 109 |     weight_filler {
 110 |       type: "xavier"
 111 |     }
 112 |   }
 113 | }
 114 | layer {
 115 |   name: "fire2/relu_expand1x1"
 116 |   type: "ReLU"
 117 |   bottom: "fire2/expand1x1"
 118 |   top: "fire2/expand1x1"
 119 | }
 120 | layer {
 121 |   name: "fire2/expand3x3"
 122 |   type: "Convolution"
 123 |   bottom: "fire2/squeeze1x1"
 124 |   top: "fire2/expand3x3"
 125 |   param {
 126 |     lr_mult: 1.0
 127 |     decay_mult: 0.0
 128 |   }
 129 |   param {
 130 |     lr_mult: 1.0
 131 |     decay_mult: 0.0
 132 |   }
 133 |   convolution_param {
 134 |     num_output: 53
 135 |     pad: 1
 136 |     kernel_size: 3
 137 |     weight_filler {
 138 |       type: "xavier"
 139 |     }
 140 |   }
 141 | }
 142 | layer {
 143 |   name: "fire2/relu_expand3x3"
 144 |   type: "ReLU"
 145 |   bottom: "fire2/expand3x3"
 146 |   top: "fire2/expand3x3"
 147 | }
 148 | layer {
 149 |   name: "fire2/concat"
 150 |   type: "Concat"
 151 |   bottom: "fire2/expand1x1"
 152 |   bottom: "fire2/expand3x3"
 153 |   top: "fire2/concat"
 154 | }
 155 | layer {
 156 |   name: "fire3/squeeze1x1"
 157 |   type: "Convolution"
 158 |   bottom: "fire2/concat"
 159 |   top: "fire3/squeeze1x1"
 160 |   param {
 161 |     lr_mult: 1.0
 162 |     decay_mult: 0.0
 163 |   }
 164 |   param {
 165 |     lr_mult: 1.0
 166 |     decay_mult: 0.0
 167 |   }
 168 |   convolution_param {
 169 |     num_output: 15
 170 |     kernel_size: 1
 171 |     weight_filler {
 172 |       type: "xavier"
 173 |     }
 174 |   }
 175 | }
 176 | layer {
 177 |   name: "fire3/expand1x1"
 178 |   type: "Convolution"
 179 |   bottom: "fire3/squeeze1x1"
 180 |   top: "fire3/expand1x1"
 181 |   param {
 182 |     lr_mult: 1.0
 183 |     decay_mult: 0.0
 184 |   }
 185 |   param {
 186 |     lr_mult: 1.0
 187 |     decay_mult: 0.0
 188 |   }
 189 |   convolution_param {
 190 |     num_output: 54
 191 |     kernel_size: 1
 192 |     weight_filler {
 193 |       type: "xavier"
 194 |     }
 195 |   }
 196 | }
 197 | layer {
 198 |   name: "fire3/relu_expand1x1"
 199 |   type: "ReLU"
 200 |   bottom: "fire3/expand1x1"
 201 |   top: "fire3/expand1x1"
 202 | }
 203 | layer {
 204 |   name: "fire3/expand3x3"
 205 |   type: "Convolution"
 206 |   bottom: "fire3/squeeze1x1"
 207 |   top: "fire3/expand3x3"
 208 |   param {
 209 |     lr_mult: 1.0
 210 |     decay_mult: 0.0
 211 |   }
 212 |   param {
 213 |     lr_mult: 1.0
 214 |     decay_mult: 0.0
 215 |   }
 216 |   convolution_param {
 217 |     num_output: 52
 218 |     pad: 1
 219 |     kernel_size: 3
 220 |     weight_filler {
 221 |       type: "xavier"
 222 |     }
 223 |   }
 224 | }
 225 | layer {
 226 |   name: "fire3/relu_expand3x3"
 227 |   type: "ReLU"
 228 |   bottom: "fire3/expand3x3"
 229 |   top: "fire3/expand3x3"
 230 | }
 231 | layer {
 232 |   name: "fire3/concat"
 233 |   type: "Concat"
 234 |   bottom: "fire3/expand1x1"
 235 |   bottom: "fire3/expand3x3"
 236 |   top: "fire3/concat"
 237 | }
 238 | layer {
 239 |   name: "pool3"
 240 |   type: "Pooling"
 241 |   bottom: "fire3/concat"
 242 |   top: "pool3"
 243 |   pooling_param {
 244 |     pool: MAX
 245 |     kernel_size: 3
 246 |     stride: 2
 247 |   }
 248 | }
 249 | layer {
 250 |   name: "fire4/squeeze1x1"
 251 |   type: "Convolution"
 252 |   bottom: "pool3"
 253 |   top: "fire4/squeeze1x1"
 254 |   param {
 255 |     lr_mult: 1.0
 256 |     decay_mult: 0.0
 257 |   }
 258 |   param {
 259 |     lr_mult: 1.0
 260 |     decay_mult: 0.0
 261 |   }
 262 |   convolution_param {
 263 |     num_output: 29
 264 |     kernel_size: 1
 265 |     weight_filler {
 266 |       type: "xavier"
 267 |     }
 268 |   }
 269 | }
 270 | layer {
 271 |   name: "fire4/expand1x1"
 272 |   type: "Convolution"
 273 |   bottom: "fire4/squeeze1x1"
 274 |   top: "fire4/expand1x1"
 275 |   param {
 276 |     lr_mult: 1.0
 277 |     decay_mult: 0.0
 278 |   }
 279 |   param {
 280 |     lr_mult: 1.0
 281 |     decay_mult: 0.0
 282 |   }
 283 |   convolution_param {
 284 |     num_output: 92
 285 |     kernel_size: 1
 286 |     weight_filler {
 287 |       type: "xavier"
 288 |     }
 289 |   }
 290 | }
 291 | layer {
 292 |   name: "fire4/relu_expand1x1"
 293 |   type: "ReLU"
 294 |   bottom: "fire4/expand1x1"
 295 |   top: "fire4/expand1x1"
 296 | }
 297 | layer {
 298 |   name: "fire4/expand3x3"
 299 |   type: "Convolution"
 300 |   bottom: "fire4/squeeze1x1"
 301 |   top: "fire4/expand3x3"
 302 |   param {
 303 |     lr_mult: 1.0
 304 |     decay_mult: 0.0
 305 |   }
 306 |   param {
 307 |     lr_mult: 1.0
 308 |     decay_mult: 0.0
 309 |   }
 310 |   convolution_param {
 311 |     num_output: 94
 312 |     pad: 1
 313 |     kernel_size: 3
 314 |     weight_filler {
 315 |       type: "xavier"
 316 |     }
 317 |   }
 318 | }
 319 | layer {
 320 |   name: "fire4/relu_expand3x3"
 321 |   type: "ReLU"
 322 |   bottom: "fire4/expand3x3"
 323 |   top: "fire4/expand3x3"
 324 | }
 325 | layer {
 326 |   name: "fire4/concat"
 327 |   type: "Concat"
 328 |   bottom: "fire4/expand1x1"
 329 |   bottom: "fire4/expand3x3"
 330 |   top: "fire4/concat"
 331 | }
 332 | layer {
 333 |   name: "fire5/squeeze1x1"
 334 |   type: "Convolution"
 335 |   bottom: "fire4/concat"
 336 |   top: "fire5/squeeze1x1"
 337 |   param {
 338 |     lr_mult: 1.0
 339 |     decay_mult: 0.0
 340 |   }
 341 |   param {
 342 |     lr_mult: 1.0
 343 |     decay_mult: 0.0
 344 |   }
 345 |   convolution_param {
 346 |     num_output: 29
 347 |     kernel_size: 1
 348 |     weight_filler {
 349 |       type: "xavier"
 350 |     }
 351 |   }
 352 | }
 353 | layer {
 354 |   name: "fire5/expand1x1"
 355 |   type: "Convolution"
 356 |   bottom: "fire5/squeeze1x1"
 357 |   top: "fire5/expand1x1"
 358 |   param {
 359 |     lr_mult: 1.0
 360 |     decay_mult: 0.0
 361 |   }
 362 |   param {
 363 |     lr_mult: 1.0
 364 |     decay_mult: 0.0
 365 |   }
 366 |   convolution_param {
 367 |     num_output: 90
 368 |     kernel_size: 1
 369 |     weight_filler {
 370 |       type: "xavier"
 371 |     }
 372 |   }
 373 | }
 374 | layer {
 375 |   name: "fire5/relu_expand1x1"
 376 |   type: "ReLU"
 377 |   bottom: "fire5/expand1x1"
 378 |   top: "fire5/expand1x1"
 379 | }
 380 | layer {
 381 |   name: "fire5/expand3x3"
 382 |   type: "Convolution"
 383 |   bottom: "fire5/squeeze1x1"
 384 |   top: "fire5/expand3x3"
 385 |   param {
 386 |     lr_mult: 1.0
 387 |     decay_mult: 0.0
 388 |   }
 389 |   param {
 390 |     lr_mult: 1.0
 391 |     decay_mult: 0.0
 392 |   }
 393 |   convolution_param {
 394 |     num_output: 83
 395 |     pad: 1
 396 |     kernel_size: 3
 397 |     weight_filler {
 398 |       type: "xavier"
 399 |     }
 400 |   }
 401 | }
 402 | layer {
 403 |   name: "fire5/relu_expand3x3"
 404 |   type: "ReLU"
 405 |   bottom: "fire5/expand3x3"
 406 |   top: "fire5/expand3x3"
 407 | }
 408 | layer {
 409 |   name: "fire5/concat"
 410 |   type: "Concat"
 411 |   bottom: "fire5/expand1x1"
 412 |   bottom: "fire5/expand3x3"
 413 |   top: "fire5/concat"
 414 | }
 415 | layer {
 416 |   name: "pool5"
 417 |   type: "Pooling"
 418 |   bottom: "fire5/concat"
 419 |   top: "pool5"
 420 |   pooling_param {
 421 |     pool: MAX
 422 |     kernel_size: 3
 423 |     stride: 2
 424 |   }
 425 | }
 426 | layer {
 427 |   name: "fire6/squeeze1x1"
 428 |   type: "Convolution"
 429 |   bottom: "pool5"
 430 |   top: "fire6/squeeze1x1"
 431 |   param {
 432 |     lr_mult: 1.0
 433 |     decay_mult: 0.0
 434 |   }
 435 |   param {
 436 |     lr_mult: 1.0
 437 |     decay_mult: 0.0
 438 |   }
 439 |   convolution_param {
 440 |     num_output: 44
 441 |     kernel_size: 1
 442 |     weight_filler {
 443 |       type: "xavier"
 444 |     }
 445 |   }
 446 | }
 447 | layer {
 448 |   name: "fire6/expand1x1"
 449 |   type: "Convolution"
 450 |   bottom: "fire6/squeeze1x1"
 451 |   top: "fire6/expand1x1"
 452 |   param {
 453 |     lr_mult: 1.0
 454 |     decay_mult: 0.0
 455 |   }
 456 |   param {
 457 |     lr_mult: 1.0
 458 |     decay_mult: 0.0
 459 |   }
 460 |   convolution_param {
 461 |     num_output: 166
 462 |     kernel_size: 1
 463 |     weight_filler {
 464 |       type: "xavier"
 465 |     }
 466 |   }
 467 | }
 468 | layer {
 469 |   name: "fire6/relu_expand1x1"
 470 |   type: "ReLU"
 471 |   bottom: "fire6/expand1x1"
 472 |   top: "fire6/expand1x1"
 473 | }
 474 | layer {
 475 |   name: "fire6/expand3x3"
 476 |   type: "Convolution"
 477 |   bottom: "fire6/squeeze1x1"
 478 |   top: "fire6/expand3x3"
 479 |   param {
 480 |     lr_mult: 1.0
 481 |     decay_mult: 0.0
 482 |   }
 483 |   param {
 484 |     lr_mult: 1.0
 485 |     decay_mult: 0.0
 486 |   }
 487 |   convolution_param {
 488 |     num_output: 161
 489 |     pad: 1
 490 |     kernel_size: 3
 491 |     weight_filler {
 492 |       type: "xavier"
 493 |     }
 494 |   }
 495 | }
 496 | layer {
 497 |   name: "fire6/relu_expand3x3"
 498 |   type: "ReLU"
 499 |   bottom: "fire6/expand3x3"
 500 |   top: "fire6/expand3x3"
 501 | }
 502 | layer {
 503 |   name: "fire6/concat"
 504 |   type: "Concat"
 505 |   bottom: "fire6/expand1x1"
 506 |   bottom: "fire6/expand3x3"
 507 |   top: "fire6/concat"
 508 | }
 509 | layer {
 510 |   name: "fire7/squeeze1x1"
 511 |   type: "Convolution"
 512 |   bottom: "fire6/concat"
 513 |   top: "fire7/squeeze1x1"
 514 |   param {
 515 |     lr_mult: 1.0
 516 |     decay_mult: 0.0
 517 |   }
 518 |   param {
 519 |     lr_mult: 1.0
 520 |     decay_mult: 0.0
 521 |   }
 522 |   convolution_param {
 523 |     num_output: 45
 524 |     kernel_size: 1
 525 |     weight_filler {
 526 |       type: "xavier"
 527 |     }
 528 |   }
 529 | }
 530 | layer {
 531 |   name: "fire7/expand1x1"
 532 |   type: "Convolution"
 533 |   bottom: "fire7/squeeze1x1"
 534 |   top: "fire7/expand1x1"
 535 |   param {
 536 |     lr_mult: 1.0
 537 |     decay_mult: 0.0
 538 |   }
 539 |   param {
 540 |     lr_mult: 1.0
 541 |     decay_mult: 0.0
 542 |   }
 543 |   convolution_param {
 544 |     num_output: 155
 545 |     kernel_size: 1
 546 |     weight_filler {
 547 |       type: "xavier"
 548 |     }
 549 |   }
 550 | }
 551 | layer {
 552 |   name: "fire7/relu_expand1x1"
 553 |   type: "ReLU"
 554 |   bottom: "fire7/expand1x1"
 555 |   top: "fire7/expand1x1"
 556 | }
 557 | layer {
 558 |   name: "fire7/expand3x3"
 559 |   type: "Convolution"
 560 |   bottom: "fire7/squeeze1x1"
 561 |   top: "fire7/expand3x3"
 562 |   param {
 563 |     lr_mult: 1.0
 564 |     decay_mult: 0.0
 565 |   }
 566 |   param {
 567 |     lr_mult: 1.0
 568 |     decay_mult: 0.0
 569 |   }
 570 |   convolution_param {
 571 |     num_output: 146
 572 |     pad: 1
 573 |     kernel_size: 3
 574 |     weight_filler {
 575 |       type: "xavier"
 576 |     }
 577 |   }
 578 | }
 579 | layer {
 580 |   name: "fire7/relu_expand3x3"
 581 |   type: "ReLU"
 582 |   bottom: "fire7/expand3x3"
 583 |   top: "fire7/expand3x3"
 584 | }
 585 | layer {
 586 |   name: "fire7/concat"
 587 |   type: "Concat"
 588 |   bottom: "fire7/expand1x1"
 589 |   bottom: "fire7/expand3x3"
 590 |   top: "fire7/concat"
 591 | }
 592 | layer {
 593 |   name: "fire8/squeeze1x1"
 594 |   type: "Convolution"
 595 |   bottom: "fire7/concat"
 596 |   top: "fire8/squeeze1x1"
 597 |   param {
 598 |     lr_mult: 1.0
 599 |     decay_mult: 0.0
 600 |   }
 601 |   param {
 602 |     lr_mult: 1.0
 603 |     decay_mult: 0.0
 604 |   }
 605 |   convolution_param {
 606 |     num_output: 49
 607 |     kernel_size: 1
 608 |     weight_filler {
 609 |       type: "xavier"
 610 |     }
 611 |   }
 612 | }
 613 | layer {
 614 |   name: "fire8/expand1x1"
 615 |   type: "Convolution"
 616 |   bottom: "fire8/squeeze1x1"
 617 |   top: "fire8/expand1x1"
 618 |   param {
 619 |     lr_mult: 1.0
 620 |     decay_mult: 0.0
 621 |   }
 622 |   param {
 623 |     lr_mult: 1.0
 624 |     decay_mult: 0.0
 625 |   }
 626 |   convolution_param {
 627 |     num_output: 163
 628 |     kernel_size: 1
 629 |     weight_filler {
 630 |       type: "xavier"
 631 |     }
 632 |   }
 633 | }
 634 | layer {
 635 |   name: "fire8/relu_expand1x1"
 636 |   type: "ReLU"
 637 |   bottom: "fire8/expand1x1"
 638 |   top: "fire8/expand1x1"
 639 | }
 640 | layer {
 641 |   name: "fire8/expand3x3"
 642 |   type: "Convolution"
 643 |   bottom: "fire8/squeeze1x1"
 644 |   top: "fire8/expand3x3"
 645 |   param {
 646 |     lr_mult: 1.0
 647 |     decay_mult: 0.0
 648 |   }
 649 |   param {
 650 |     lr_mult: 1.0
 651 |     decay_mult: 0.0
 652 |   }
 653 |   convolution_param {
 654 |     num_output: 171
 655 |     pad: 1
 656 |     kernel_size: 3
 657 |     weight_filler {
 658 |       type: "xavier"
 659 |     }
 660 |   }
 661 | }
 662 | layer {
 663 |   name: "fire8/relu_expand3x3"
 664 |   type: "ReLU"
 665 |   bottom: "fire8/expand3x3"
 666 |   top: "fire8/expand3x3"
 667 | }
 668 | layer {
 669 |   name: "fire8/concat"
 670 |   type: "Concat"
 671 |   bottom: "fire8/expand1x1"
 672 |   bottom: "fire8/expand3x3"
 673 |   top: "fire8/concat"
 674 | }
 675 | layer {
 676 |   name: "fire9/squeeze1x1"
 677 |   type: "Convolution"
 678 |   bottom: "fire8/concat"
 679 |   top: "fire9/squeeze1x1"
 680 |   param {
 681 |     lr_mult: 1.0
 682 |     decay_mult: 0.0
 683 |   }
 684 |   param {
 685 |     lr_mult: 1.0
 686 |     decay_mult: 0.0
 687 |   }
 688 |   convolution_param {
 689 |     num_output: 25
 690 |     kernel_size: 1
 691 |     weight_filler {
 692 |       type: "xavier"
 693 |     }
 694 |   }
 695 | }
 696 | layer {
 697 |   name: "fire9/expand1x1"
 698 |   type: "Convolution"
 699 |   bottom: "fire9/squeeze1x1"
 700 |   top: "fire9/expand1x1"
 701 |   param {
 702 |     lr_mult: 1.0
 703 |     decay_mult: 0.0
 704 |   }
 705 |   param {
 706 |     lr_mult: 1.0
 707 |     decay_mult: 0.0
 708 |   }
 709 |   convolution_param {
 710 |     num_output: 29
 711 |     kernel_size: 1
 712 |     weight_filler {
 713 |       type: "xavier"
 714 |     }
 715 |   }
 716 | }
 717 | layer {
 718 |   name: "fire9/expand1x1/bn"
 719 |   type: "BatchNorm"
 720 |   bottom: "fire9/expand1x1"
 721 |   top: "fire9/expand1x1"
 722 |   param {
 723 |     lr_mult: 0
 724 |     decay_mult: 0
 725 |   }
 726 |   param {
 727 |     lr_mult: 0
 728 |     decay_mult: 0
 729 |   }
 730 |   param {
 731 |     lr_mult: 0
 732 |     decay_mult: 0
 733 |   }
 734 | }
 735 | layer {
 736 |   name: "fire9/expand1x1/scale"
 737 |   type: "Scale"
 738 |   bottom: "fire9/expand1x1"
 739 |   top: "fire9/expand1x1"
 740 |   param {
 741 |     lr_mult: 1.0
 742 |     decay_mult: 0.0
 743 |   }
 744 |   param {
 745 |     lr_mult: 2.0
 746 |     decay_mult: 0.0
 747 |   }
 748 |   scale_param {
 749 |     filler {
 750 |       value: 1
 751 |     }
 752 |     bias_term: true
 753 |     bias_filler {
 754 |       value: 0
 755 |     }
 756 |   }
 757 | }
 758 | layer {
 759 |   name: "fire9/relu_expand1x1"
 760 |   type: "ReLU"
 761 |   bottom: "fire9/expand1x1"
 762 |   top: "fire9/expand1x1"
 763 | }
 764 | layer {
 765 |   name: "fire9/expand3x3"
 766 |   type: "Convolution"
 767 |   bottom: "fire9/squeeze1x1"
 768 |   top: "fire9/expand3x3"
 769 |   param {
 770 |     lr_mult: 1.0
 771 |     decay_mult: 0.0
 772 |   }
 773 |   param {
 774 |     lr_mult: 1.0
 775 |     decay_mult: 0.0
 776 |   }
 777 |   convolution_param {
 778 |     num_output: 54
 779 |     pad: 1
 780 |     kernel_size: 3
 781 |     weight_filler {
 782 |       type: "xavier"
 783 |     }
 784 |   }
 785 | }
 786 | layer {
 787 |   name: "fire9/expand3x3/bn"
 788 |   type: "BatchNorm"
 789 |   bottom: "fire9/expand3x3"
 790 |   top: "fire9/expand3x3"
 791 |   param {
 792 |     lr_mult: 0
 793 |     decay_mult: 0
 794 |   }
 795 |   param {
 796 |     lr_mult: 0
 797 |     decay_mult: 0
 798 |   }
 799 |   param {
 800 |     lr_mult: 0
 801 |     decay_mult: 0
 802 |   }
 803 | }
 804 | layer {
 805 |   name: "fire9/expand3x3/scale"
 806 |   type: "Scale"
 807 |   bottom: "fire9/expand3x3"
 808 |   top: "fire9/expand3x3"
 809 |   param {
 810 |     lr_mult: 1.0
 811 |     decay_mult: 0.0
 812 |   }
 813 |   param {
 814 |     lr_mult: 2.0
 815 |     decay_mult: 0.0
 816 |   }
 817 |   scale_param {
 818 |     filler {
 819 |       value: 1
 820 |     }
 821 |     bias_term: true
 822 |     bias_filler {
 823 |       value: 0
 824 |     }
 825 |   }
 826 | }
 827 | layer {
 828 |   name: "fire9/relu_expand3x3"
 829 |   type: "ReLU"
 830 |   bottom: "fire9/expand3x3"
 831 |   top: "fire9/expand3x3"
 832 | }
 833 | layer {
 834 |   name: "fire9/concat"
 835 |   type: "Concat"
 836 |   bottom: "fire9/expand1x1"
 837 |   bottom: "fire9/expand3x3"
 838 |   top: "fire9/concat"
 839 | }
 840 | layer {
 841 |   name: "pool9"
 842 |   type: "Pooling"
 843 |   bottom: "fire9/concat"
 844 |   top: "pool9"
 845 |   pooling_param {
 846 |     pool: MAX
 847 |     kernel_size: 3
 848 |     stride: 2
 849 |   }
 850 | }
 851 | #================ssd append layer fire10, out 9 * 9
 852 | layer {
 853 |   name: "fire10/squeeze1x1"
 854 |   type: "Convolution"
 855 |   bottom: "pool9"
 856 |   top: "fire10/squeeze1x1"
 857 |   convolution_param {
 858 |     num_output: 37
 859 |     kernel_size: 1
 860 |     weight_filler {
 861 |       type: "xavier"
 862 |     }
 863 |   }
 864 | }
 865 | layer {
 866 |   name: "fire10/squeeze1x1/bn"
 867 |   type: "BatchNorm"
 868 |   top: "fire10/squeeze1x1"
 869 |   bottom: "fire10/squeeze1x1"
 870 |   param {
 871 |     lr_mult: 0
 872 |     decay_mult: 0
 873 |   }
 874 |   param {
 875 |     lr_mult: 0
 876 |     decay_mult: 0
 877 |   }
 878 |   param {
 879 |     lr_mult: 0
 880 |     decay_mult: 0
 881 |   }
 882 | }
 883 | layer {
 884 |   name: "fire10/squeeze1x1/scale"
 885 |   type: "Scale"
 886 |   top: "fire10/squeeze1x1"
 887 |   bottom: "fire10/squeeze1x1"
 888 |   param {
 889 |     lr_mult: 1.0
 890 |     decay_mult: 0.0
 891 |   }
 892 |   param {
 893 |     lr_mult: 2.0
 894 |     decay_mult: 0.0
 895 |   }
 896 |   scale_param {
 897 |     filler {
 898 |       value: 1
 899 |     }
 900 |     bias_term: true
 901 |     bias_filler {
 902 |       value: 0
 903 |     }
 904 |   }
 905 | }
 906 | layer {
 907 |   name: "fire10/expand1x1"
 908 |   type: "Convolution"
 909 |   bottom: "fire10/squeeze1x1"
 910 |   top: "fire10/expand1x1"
 911 |   convolution_param {
 912 |     num_output: 45
 913 |     kernel_size: 1
 914 |     weight_filler {
 915 |       type: "xavier"
 916 |     }
 917 |   }
 918 | }
 919 | layer {
 920 |   name: "fire10/expand1x1/bn"
 921 |   type: "BatchNorm"
 922 |   bottom: "fire10/expand1x1"
 923 |   top: "fire10/expand1x1"
 924 |   param {
 925 |     lr_mult: 0
 926 |     decay_mult: 0
 927 |   }
 928 |   param {
 929 |     lr_mult: 0
 930 |     decay_mult: 0
 931 |   }
 932 |   param {
 933 |     lr_mult: 0
 934 |     decay_mult: 0
 935 |   }
 936 | }
 937 | layer {
 938 |   name: "fire10/expand1x1/scale"
 939 |   type: "Scale"
 940 |   bottom: "fire10/expand1x1"
 941 |   top: "fire10/expand1x1"
 942 |   param {
 943 |     lr_mult: 1.0
 944 |     decay_mult: 0.0
 945 |   }
 946 |   param {
 947 |     lr_mult: 2.0
 948 |     decay_mult: 0.0
 949 |   }
 950 |   scale_param {
 951 |     filler {
 952 |       value: 1
 953 |     }
 954 |     bias_term: true
 955 |     bias_filler {
 956 |       value: 0
 957 |     }
 958 |   }
 959 | }
 960 | layer {
 961 |   name: "fire10/relu_expand1x1"
 962 |   type: "ReLU"
 963 |   bottom: "fire10/expand1x1"
 964 |   top: "fire10/expand1x1"
 965 | }
 966 | layer {
 967 |   name: "fire10/expand3x3"
 968 |   type: "Convolution"
 969 |   bottom: "fire10/squeeze1x1"
 970 |   top: "fire10/expand3x3"
 971 |   convolution_param {
 972 |     num_output: 56
 973 |     pad: 1
 974 |     kernel_size: 3
 975 |     weight_filler {
 976 |       type: "xavier"
 977 |     }
 978 |   }
 979 | }
 980 | layer {
 981 |   name: "fire10/expand3x3/bn"
 982 |   type: "BatchNorm"
 983 |   bottom: "fire10/expand3x3"
 984 |   top: "fire10/expand3x3"
 985 |   param {
 986 |     lr_mult: 0
 987 |     decay_mult: 0
 988 |   }
 989 |   param {
 990 |     lr_mult: 0
 991 |     decay_mult: 0
 992 |   }
 993 |   param {
 994 |     lr_mult: 0
 995 |     decay_mult: 0
 996 |   }
 997 | }
 998 | layer {
 999 |   name: "fire10/expand3x3/scale"
1000 |   type: "Scale"
1001 |   bottom: "fire10/expand3x3"
1002 |   top: "fire10/expand3x3"
1003 |   param {
1004 |     lr_mult: 1.0
1005 |     decay_mult: 0.0
1006 |   }
1007 |   param {
1008 |     lr_mult: 2.0
1009 |     decay_mult: 0.0
1010 |   }
1011 |   scale_param {
1012 |     filler {
1013 |       value: 1
1014 |     }
1015 |     bias_term: true
1016 |     bias_filler {
1017 |       value: 0
1018 |     }
1019 |   }
1020 | }
1021 | layer {
1022 |   name: "fire10/relu_expand3x3"
1023 |   type: "ReLU"
1024 |   bottom: "fire10/expand3x3"
1025 |   top: "fire10/expand3x3"
1026 | }
1027 | layer {
1028 |   name: "fire10/concat"
1029 |   type: "Concat"
1030 |   bottom: "fire10/expand1x1"
1031 |   bottom: "fire10/expand3x3"
1032 |   top: "fire10/concat"
1033 | }
1034 | #================ssd append layer fire11, out 4 * 4
1035 | layer {
1036 |   name: "pool10"
1037 |   type: "Pooling"
1038 |   bottom: "fire10/concat"
1039 |   top: "pool10"
1040 |   pooling_param {
1041 |     pool: MAX
1042 |     kernel_size: 3
1043 |     stride: 2
1044 |   }
1045 | }
1046 | layer {
1047 |   name: "fire11/squeeze1x1"
1048 |   type: "Convolution"
1049 |   bottom: "pool10"
1050 |   top: "fire11/squeeze1x1"
1051 |   convolution_param {
1052 |     num_output: 38
1053 |     kernel_size: 1
1054 |     weight_filler {
1055 |       type: "xavier"
1056 |     }
1057 |   }
1058 | }
1059 | layer {
1060 |   name: "fire11/squeeze1x1/bn"
1061 |   type: "BatchNorm"
1062 |   top: "fire11/squeeze1x1"
1063 |   bottom: "fire11/squeeze1x1"
1064 |   param {
1065 |     lr_mult: 0
1066 |     decay_mult: 0
1067 |   }
1068 |   param {
1069 |     lr_mult: 0
1070 |     decay_mult: 0
1071 |   }
1072 |   param {
1073 |     lr_mult: 0
1074 |     decay_mult: 0
1075 |   }
1076 | }
1077 | layer {
1078 |   name: "fire11/squeeze1x1/scale"
1079 |   type: "Scale"
1080 |   top: "fire11/squeeze1x1"
1081 |   bottom: "fire11/squeeze1x1"
1082 |   param {
1083 |     lr_mult: 1.0
1084 |     decay_mult: 0.0
1085 |   }
1086 |   param {
1087 |     lr_mult: 2.0
1088 |     decay_mult: 0.0
1089 |   }
1090 |   scale_param {
1091 |     filler {
1092 |       value: 1
1093 |     }
1094 |     bias_term: true
1095 |     bias_filler {
1096 |       value: 0
1097 |     }
1098 |   }
1099 | }
1100 | layer {
1101 |   name: "fire11/expand1x1"
1102 |   type: "Convolution"
1103 |   bottom: "fire11/squeeze1x1"
1104 |   top: "fire11/expand1x1"
1105 |   convolution_param {
1106 |     num_output: 41
1107 |     kernel_size: 1
1108 |     weight_filler {
1109 |       type: "xavier"
1110 |     }
1111 |   }
1112 | }
1113 | layer {
1114 |   name: "fire11/expand1x1/bn"
1115 |   type: "BatchNorm"
1116 |   bottom: "fire11/expand1x1"
1117 |   top: "fire11/expand1x1"
1118 |   param {
1119 |     lr_mult: 0
1120 |     decay_mult: 0
1121 |   }
1122 |   param {
1123 |     lr_mult: 0
1124 |     decay_mult: 0
1125 |   }
1126 |   param {
1127 |     lr_mult: 0
1128 |     decay_mult: 0
1129 |   }
1130 | }
1131 | layer {
1132 |   name: "fire11/expand1x1/scale"
1133 |   type: "Scale"
1134 |   bottom: "fire11/expand1x1"
1135 |   top: "fire11/expand1x1"
1136 |   param {
1137 |     lr_mult: 1.0
1138 |     decay_mult: 0.0
1139 |   }
1140 |   param {
1141 |     lr_mult: 2.0
1142 |     decay_mult: 0.0
1143 |   }
1144 |   scale_param {
1145 |     filler {
1146 |       value: 1
1147 |     }
1148 |     bias_term: true
1149 |     bias_filler {
1150 |       value: 0
1151 |     }
1152 |   }
1153 | }
1154 | layer {
1155 |   name: "fire11/relu_expand1x1"
1156 |   type: "ReLU"
1157 |   bottom: "fire11/expand1x1"
1158 |   top: "fire11/expand1x1"
1159 | }
1160 | layer {
1161 |   name: "fire11/expand3x3"
1162 |   type: "Convolution"
1163 |   bottom: "fire11/squeeze1x1"
1164 |   top: "fire11/expand3x3"
1165 |   convolution_param {
1166 |     num_output: 44
1167 |     pad: 1
1168 |     kernel_size: 3
1169 |     weight_filler {
1170 |       type: "xavier"
1171 |     }
1172 |   }
1173 | }
1174 | layer {
1175 |   name: "fire11/expand3x3/bn"
1176 |   type: "BatchNorm"
1177 |   bottom: "fire11/expand3x3"
1178 |   top: "fire11/expand3x3"
1179 |   param {
1180 |     lr_mult: 0
1181 |     decay_mult: 0
1182 |   }
1183 |   param {
1184 |     lr_mult: 0
1185 |     decay_mult: 0
1186 |   }
1187 |   param {
1188 |     lr_mult: 0
1189 |     decay_mult: 0
1190 |   }
1191 | }
1192 | layer {
1193 |   name: "fire11/expand3x3/scale"
1194 |   type: "Scale"
1195 |   bottom: "fire11/expand3x3"
1196 |   top: "fire11/expand3x3"
1197 |   param {
1198 |     lr_mult: 1.0
1199 |     decay_mult: 0.0
1200 |   }
1201 |   param {
1202 |     lr_mult: 2.0
1203 |     decay_mult: 0.0
1204 |   }
1205 |   scale_param {
1206 |     filler {
1207 |       value: 1
1208 |     }
1209 |     bias_term: true
1210 |     bias_filler {
1211 |       value: 0
1212 |     }
1213 |   }
1214 | }
1215 | layer {
1216 |   name: "fire11/relu_expand3x3"
1217 |   type: "ReLU"
1218 |   bottom: "fire11/expand3x3"
1219 |   top: "fire11/expand3x3"
1220 | }
1221 | layer {
1222 |   name: "fire11/concat"
1223 |   type: "Concat"
1224 |   bottom: "fire11/expand1x1"
1225 |   bottom: "fire11/expand3x3"
1226 |   top: "fire11/concat"
1227 | }
1228 | #================ssd append layer conv12, out 2 * 2
1229 | layer {
1230 |   name: "conv12_1"
1231 |   type: "Convolution"
1232 |   bottom: "fire11/concat"
1233 |   top: "conv12_1"
1234 |   param {
1235 |     lr_mult: 1.0
1236 |     decay_mult: 1.0
1237 |   }
1238 |   convolution_param {
1239 |     num_output: 51
1240 |     bias_term: false
1241 |     kernel_size: 1
1242 |     weight_filler {
1243 |       type: "msra"
1244 |     }
1245 |   }
1246 | }
1247 | layer {
1248 |   name: "conv12_1/bn"
1249 |   type: "BatchNorm"
1250 |   bottom: "conv12_1"
1251 |   top: "conv12_1"
1252 |   param {
1253 |     lr_mult: 0
1254 |     decay_mult: 0
1255 |   }
1256 |   param {
1257 |     lr_mult: 0
1258 |     decay_mult: 0
1259 |   }
1260 |   param {
1261 |     lr_mult: 0
1262 |     decay_mult: 0
1263 |   }
1264 | }
1265 | layer {
1266 |   name: "conv12_1/scale"
1267 |   type: "Scale"
1268 |   bottom: "conv12_1"
1269 |   top: "conv12_1"
1270 |   param {
1271 |     lr_mult: 1.0
1272 |     decay_mult: 0.0
1273 |   }
1274 |   param {
1275 |     lr_mult: 2.0
1276 |     decay_mult: 0.0
1277 |   }
1278 |   scale_param {
1279 |     filler {
1280 |       value: 1
1281 |     }
1282 |     bias_term: true
1283 |     bias_filler {
1284 |       value: 0
1285 |     }
1286 |   }
1287 | }
1288 | layer {
1289 |   name: "conv12_1/relu"
1290 |   type: "ReLU"
1291 |   bottom: "conv12_1"
1292 |   top: "conv12_1"
1293 | }
1294 | layer {
1295 |   name: "conv12_2"
1296 |   type: "Convolution"
1297 |   bottom: "conv12_1"
1298 |   top: "conv12_2"
1299 |   param {
1300 |     lr_mult: 1.0
1301 |     decay_mult: 1.0
1302 |   }
1303 |   convolution_param {
1304 |     num_output: 46
1305 |     bias_term: false
1306 |     pad: 1
1307 |     kernel_size: 3
1308 |     stride: 2
1309 |     weight_filler {
1310 |       type: "msra"
1311 |     }
1312 |   }
1313 | }
1314 | layer {
1315 |   name: "conv12_2/bn"
1316 |   type: "BatchNorm"
1317 |   bottom: "conv12_2"
1318 |   top: "conv12_2"
1319 |   param {
1320 |     lr_mult: 0
1321 |     decay_mult: 0
1322 |   }
1323 |   param {
1324 |     lr_mult: 0
1325 |     decay_mult: 0
1326 |   }
1327 |   param {
1328 |     lr_mult: 0
1329 |     decay_mult: 0
1330 |   }
1331 | }
1332 | layer {
1333 |   name: "conv12_2/scale"
1334 |   type: "Scale"
1335 |   bottom: "conv12_2"
1336 |   top: "conv12_2"
1337 |   param {
1338 |     lr_mult: 1.0
1339 |     decay_mult: 0.0
1340 |   }
1341 |   param {
1342 |     lr_mult: 2.0
1343 |     decay_mult: 0.0
1344 |   }
1345 |   scale_param {
1346 |     filler {
1347 |       value: 1
1348 |     }
1349 |     bias_term: true
1350 |     bias_filler {
1351 |       value: 0
1352 |     }
1353 |   }
1354 | }
1355 | layer {
1356 |   name: "conv12_2/relu"
1357 |   type: "ReLU"
1358 |   bottom: "conv12_2"
1359 |   top: "conv12_2"
1360 | }
1361 | #================ssd append layer conv13, out 1 * 1
1362 | layer {
1363 |   name: "conv13_1"
1364 |   type: "Convolution"
1365 |   bottom: "conv12_2"
1366 |   top: "conv13_1"
1367 |   param {
1368 |     lr_mult: 1.0
1369 |     decay_mult: 1.0
1370 |   }
1371 |   convolution_param {
1372 |     num_output: 55
1373 |     bias_term: false
1374 |     kernel_size: 1
1375 |     weight_filler {
1376 |       type: "msra"
1377 |     }
1378 |   }
1379 | }
1380 | layer {
1381 |   name: "conv13_1/bn"
1382 |   type: "BatchNorm"
1383 |   bottom: "conv13_1"
1384 |   top: "conv13_1"
1385 |   param {
1386 |     lr_mult: 0
1387 |     decay_mult: 0
1388 |   }
1389 |   param {
1390 |     lr_mult: 0
1391 |     decay_mult: 0
1392 |   }
1393 |   param {
1394 |     lr_mult: 0
1395 |     decay_mult: 0
1396 |   }
1397 | }
1398 | layer {
1399 |   name: "conv13_1/scale"
1400 |   type: "Scale"
1401 |   bottom: "conv13_1"
1402 |   top: "conv13_1"
1403 |   param {
1404 |     lr_mult: 1.0
1405 |     decay_mult: 0.0
1406 |   }
1407 |   param {
1408 |     lr_mult: 2.0
1409 |     decay_mult: 0.0
1410 |   }
1411 |   scale_param {
1412 |     filler {
1413 |       value: 1
1414 |     }
1415 |     bias_term: true
1416 |     bias_filler {
1417 |       value: 0
1418 |     }
1419 |   }
1420 | }
1421 | layer {
1422 |   name: "conv13_1/relu"
1423 |   type: "ReLU"
1424 |   bottom: "conv13_1"
1425 |   top: "conv13_1"
1426 | }
1427 | layer {
1428 |   name: "conv13_2"
1429 |   type: "Convolution"
1430 |   bottom: "conv13_1"
1431 |   top: "conv13_2"
1432 |   param {
1433 |     lr_mult: 1.0
1434 |     decay_mult: 1.0
1435 |   }
1436 |   convolution_param {
1437 |     num_output: 85
1438 |     bias_term: false
1439 |     pad: 1
1440 |     kernel_size: 3
1441 |     stride: 2
1442 |     weight_filler {
1443 |       type: "msra"
1444 |     }
1445 |   }
1446 | }
1447 | layer {
1448 |   name: "conv13_2/bn"
1449 |   type: "BatchNorm"
1450 |   bottom: "conv13_2"
1451 |   top: "conv13_2"
1452 |   param {
1453 |     lr_mult: 0
1454 |     decay_mult: 0
1455 |   }
1456 |   param {
1457 |     lr_mult: 0
1458 |     decay_mult: 0
1459 |   }
1460 |   param {
1461 |     lr_mult: 0
1462 |     decay_mult: 0
1463 |   }
1464 | }
1465 | layer {
1466 |   name: "conv13_2/scale"
1467 |   type: "Scale"
1468 |   bottom: "conv13_2"
1469 |   top: "conv13_2"
1470 |   param {
1471 |     lr_mult: 1.0
1472 |     decay_mult: 0.0
1473 |   }
1474 |   param {
1475 |     lr_mult: 2.0
1476 |     decay_mult: 0.0
1477 |   }
1478 |   scale_param {
1479 |     filler {
1480 |       value: 1
1481 |     }
1482 |     bias_term: true
1483 |     bias_filler {
1484 |       value: 0
1485 |     }
1486 |   }
1487 | }
1488 | layer {
1489 |   name: "conv13_2/relu"
1490 |   type: "ReLU"
1491 |   bottom: "conv13_2"
1492 |   top: "conv13_2"
1493 | }
1494 | #================ssd mbox layer
1495 | layer {
1496 |   name: "fire5/bn"
1497 |   type: "BatchNorm"
1498 |   bottom: "fire5/concat"
1499 |   top: "fire5/normal"
1500 |   param {
1501 |     lr_mult: 0
1502 |     decay_mult: 0
1503 |   }
1504 |   param {
1505 |     lr_mult: 0
1506 |     decay_mult: 0
1507 |   }
1508 |   param {
1509 |     lr_mult: 0
1510 |     decay_mult: 0
1511 |   }
1512 | }
1513 | layer {
1514 |   name: "fire5/scale"
1515 |   type: "Scale"
1516 |   bottom: "fire5/normal"
1517 |   top: "fire5/normal"
1518 |   param {
1519 |     lr_mult: 1.0
1520 |     decay_mult: 0.0
1521 |   }
1522 |   param {
1523 |     lr_mult: 2.0
1524 |     decay_mult: 0.0
1525 |   }
1526 |   scale_param {
1527 |     filler {
1528 |       value: 1
1529 |     }
1530 |     bias_term: true
1531 |     bias_filler {
1532 |       value: 0
1533 |     }
1534 |   }
1535 | }
1536 | layer {
1537 |   name: "fire5_mbox_loc"
1538 |   type: "Convolution"
1539 |   bottom: "fire5/normal"
1540 |   top: "fire5_mbox_loc"
1541 |   param {
1542 |     lr_mult: 1.0
1543 |     decay_mult: 1.0
1544 |   }
1545 |   param {
1546 |     lr_mult: 2.0
1547 |     decay_mult: 0.0
1548 |   }
1549 |   convolution_param {
1550 |     num_output: 16
1551 |     pad: 1
1552 |     kernel_size: 3
1553 |     stride: 1
1554 |     weight_filler {
1555 |       type: "msra"
1556 |     }
1557 |     bias_filler {
1558 |       type: "constant"
1559 |       value: 0.0
1560 |     }
1561 |   }
1562 | }
1563 | layer {
1564 |   name: "fire5_mbox_loc_perm"
1565 |   type: "Permute"
1566 |   bottom: "fire5_mbox_loc"
1567 |   top: "fire5_mbox_loc_perm"
1568 |   permute_param {
1569 |     order: 0
1570 |     order: 2
1571 |     order: 3
1572 |     order: 1
1573 |   }
1574 | }
1575 | layer {
1576 |   name: "fire5_mbox_loc_flat"
1577 |   type: "Flatten"
1578 |   bottom: "fire5_mbox_loc_perm"
1579 |   top: "fire5_mbox_loc_flat"
1580 |   flatten_param {
1581 |     axis: 1
1582 |   }
1583 | }
1584 | layer {
1585 |   name: "fire5_mbox_conf"
1586 |   type: "Convolution"
1587 |   bottom: "fire5/normal"
1588 |   top: "fire5_mbox_conf"
1589 |   param {
1590 |     lr_mult: 1.0
1591 |     decay_mult: 1.0
1592 |   }
1593 |   param {
1594 |     lr_mult: 2.0
1595 |     decay_mult: 0.0
1596 |   }
1597 |   convolution_param {
1598 |     num_output: 84
1599 |     pad: 1
1600 |     kernel_size: 3
1601 |     stride: 1
1602 |     weight_filler {
1603 |       type: "msra"
1604 |     }
1605 |     bias_filler {
1606 |       type: "constant"
1607 |       value: 0.0
1608 |     }
1609 |   }
1610 | }
1611 | layer {
1612 |   name: "fire5_mbox_conf_perm"
1613 |   type: "Permute"
1614 |   bottom: "fire5_mbox_conf"
1615 |   top: "fire5_mbox_conf_perm"
1616 |   permute_param {
1617 |     order: 0
1618 |     order: 2
1619 |     order: 3
1620 |     order: 1
1621 |   }
1622 | }
1623 | layer {
1624 |   name: "fire5_mbox_conf_flat"
1625 |   type: "Flatten"
1626 |   bottom: "fire5_mbox_conf_perm"
1627 |   top: "fire5_mbox_conf_flat"
1628 |   flatten_param {
1629 |     axis: 1
1630 |   }
1631 | }
1632 | layer {
1633 |   name: "fire5_mbox_priorbox"
1634 |   type: "PriorBox"
1635 |   bottom: "fire5/normal"
1636 |   bottom: "data"
1637 |   top: "fire5_mbox_priorbox"
1638 |   prior_box_param {
1639 |     min_size: 21.0
1640 |     max_size: 45.0
1641 |     aspect_ratio: 2.0
1642 |     flip: true
1643 |     clip: false
1644 |     variance: 0.1
1645 |     variance: 0.1
1646 |     variance: 0.2
1647 |     variance: 0.2
1648 |     step: 8
1649 |   }
1650 | }
1651 | #================mbox fire9
1652 | layer {
1653 |   name: "fire9_mbox_loc"
1654 |   type: "Convolution"
1655 |   bottom: "fire9/concat"
1656 |   top: "fire9_mbox_loc"
1657 |   param {
1658 |     lr_mult: 1.0
1659 |     decay_mult: 1.0
1660 |   }
1661 |   param {
1662 |     lr_mult: 2.0
1663 |     decay_mult: 0.0
1664 |   }
1665 |   convolution_param {
1666 |     num_output: 24
1667 |     pad: 1
1668 |     kernel_size: 3
1669 |     stride: 1
1670 |     weight_filler {
1671 |       type: "msra"
1672 |     }
1673 |     bias_filler {
1674 |       type: "constant"
1675 |       value: 0.0
1676 |     }
1677 |   }
1678 | }
1679 | layer {
1680 |   name: "fire9_mbox_loc_perm"
1681 |   type: "Permute"
1682 |   bottom: "fire9_mbox_loc"
1683 |   top: "fire9_mbox_loc_perm"
1684 |   permute_param {
1685 |     order: 0
1686 |     order: 2
1687 |     order: 3
1688 |     order: 1
1689 |   }
1690 | }
1691 | layer {
1692 |   name: "fire9_mbox_loc_flat"
1693 |   type: "Flatten"
1694 |   bottom: "fire9_mbox_loc_perm"
1695 |   top: "fire9_mbox_loc_flat"
1696 |   flatten_param {
1697 |     axis: 1
1698 |   }
1699 | }
1700 | layer {
1701 |   name: "fire9_mbox_conf"
1702 |   type: "Convolution"
1703 |   bottom: "fire9/concat"
1704 |   top: "fire9_mbox_conf"
1705 |   param {
1706 |     lr_mult: 1.0
1707 |     decay_mult: 1.0
1708 |   }
1709 |   param {
1710 |     lr_mult: 2.0
1711 |     decay_mult: 0.0
1712 |   }
1713 |   convolution_param {
1714 |     num_output: 126
1715 |     pad: 1
1716 |     kernel_size: 3
1717 |     stride: 1
1718 |     weight_filler {
1719 |       type: "msra"
1720 |     }
1721 |     bias_filler {
1722 |       type: "constant"
1723 |       value: 0.0
1724 |     }
1725 |   }
1726 | }
1727 | layer {
1728 |   name: "fire9_mbox_conf_perm"
1729 |   type: "Permute"
1730 |   bottom: "fire9_mbox_conf"
1731 |   top: "fire9_mbox_conf_perm"
1732 |   permute_param {
1733 |     order: 0
1734 |     order: 2
1735 |     order: 3
1736 |     order: 1
1737 |   }
1738 | }
1739 | layer {
1740 |   name: "fire9_mbox_conf_flat"
1741 |   type: "Flatten"
1742 |   bottom: "fire9_mbox_conf_perm"
1743 |   top: "fire9_mbox_conf_flat"
1744 |   flatten_param {
1745 |     axis: 1
1746 |   }
1747 | }
1748 | layer {
1749 |   name: "fire9_mbox_priorbox"
1750 |   type: "PriorBox"
1751 |   bottom: "fire9/concat"
1752 |   bottom: "data"
1753 |   top: "fire9_mbox_priorbox"
1754 |   prior_box_param {
1755 |     min_size: 45.0
1756 |     max_size: 99.0
1757 |     aspect_ratio: 2.0
1758 |     aspect_ratio: 3.0
1759 |     flip: true
1760 |     clip: false
1761 |     variance: 0.1
1762 |     variance: 0.1
1763 |     variance: 0.2
1764 |     variance: 0.2
1765 |     step: 16
1766 |   }
1767 | }
1768 | #=====fire10 mbox
1769 | layer {
1770 |   name: "fire10_mbox_loc"
1771 |   type: "Convolution"
1772 |   bottom: "fire10/concat"
1773 |   top: "fire10_mbox_loc"
1774 |   param {
1775 |     lr_mult: 1.0
1776 |     decay_mult: 1.0
1777 |   }
1778 |   param {
1779 |     lr_mult: 2.0
1780 |     decay_mult: 0.0
1781 |   }
1782 |   convolution_param {
1783 |     num_output: 24
1784 |     pad: 1
1785 |     kernel_size: 3
1786 |     stride: 1
1787 |     weight_filler {
1788 |       type: "msra"
1789 |     }
1790 |     bias_filler {
1791 |       type: "constant"
1792 |       value: 0.0
1793 |     }
1794 |   }
1795 | }
1796 | layer {
1797 |   name: "fire10_mbox_loc_perm"
1798 |   type: "Permute"
1799 |   bottom: "fire10_mbox_loc"
1800 |   top: "fire10_mbox_loc_perm"
1801 |   permute_param {
1802 |     order: 0
1803 |     order: 2
1804 |     order: 3
1805 |     order: 1
1806 |   }
1807 | }
1808 | layer {
1809 |   name: "fire10_mbox_loc_flat"
1810 |   type: "Flatten"
1811 |   bottom: "fire10_mbox_loc_perm"
1812 |   top: "fire10_mbox_loc_flat"
1813 |   flatten_param {
1814 |     axis: 1
1815 |   }
1816 | }
1817 | layer {
1818 |   name: "fire10_mbox_conf"
1819 |   type: "Convolution"
1820 |   bottom: "fire10/concat"
1821 |   top: "fire10_mbox_conf"
1822 |   param {
1823 |     lr_mult: 1.0
1824 |     decay_mult: 1.0
1825 |   }
1826 |   param {
1827 |     lr_mult: 2.0
1828 |     decay_mult: 0.0
1829 |   }
1830 |   convolution_param {
1831 |     num_output: 126
1832 |     pad: 1
1833 |     kernel_size: 3
1834 |     stride: 1
1835 |     weight_filler {
1836 |       type: "msra"
1837 |     }
1838 |     bias_filler {
1839 |       type: "constant"
1840 |       value: 0.0
1841 |     }
1842 |   }
1843 | }
1844 | layer {
1845 |   name: "fire10_mbox_conf_perm"
1846 |   type: "Permute"
1847 |   bottom: "fire10_mbox_conf"
1848 |   top: "fire10_mbox_conf_perm"
1849 |   permute_param {
1850 |     order: 0
1851 |     order: 2
1852 |     order: 3
1853 |     order: 1
1854 |   }
1855 | }
1856 | layer {
1857 |   name: "fire10_mbox_conf_flat"
1858 |   type: "Flatten"
1859 |   bottom: "fire10_mbox_conf_perm"
1860 |   top: "fire10_mbox_conf_flat"
1861 |   flatten_param {
1862 |     axis: 1
1863 |   }
1864 | }
1865 | layer {
1866 |   name: "fire10_mbox_priorbox"
1867 |   type: "PriorBox"
1868 |   bottom: "fire10/concat"
1869 |   bottom: "data"
1870 |   top: "fire10_mbox_priorbox"
1871 |   prior_box_param {
1872 |     min_size: 99.0
1873 |     max_size: 153.0
1874 |     aspect_ratio: 2.0
1875 |     aspect_ratio: 3.0
1876 |     flip: true
1877 |     clip: false
1878 |     variance: 0.1
1879 |     variance: 0.1
1880 |     variance: 0.2
1881 |     variance: 0.2
1882 |     step: 32
1883 |   }
1884 | }
1885 | 
1886 | #=====conv11 mbox
1887 | layer {
1888 |   name: "fire11_mbox_loc"
1889 |   type: "Convolution"
1890 |   bottom: "fire11/concat"
1891 |   top: "fire11_mbox_loc"
1892 |   param {
1893 |     lr_mult: 1.0
1894 |     decay_mult: 1.0
1895 |   }
1896 |   param {
1897 |     lr_mult: 2.0
1898 |     decay_mult: 0.0
1899 |   }
1900 |   convolution_param {
1901 |     num_output: 24
1902 |     pad: 1
1903 |     kernel_size: 3
1904 |     stride: 1
1905 |     weight_filler {
1906 |       type: "msra"
1907 |     }
1908 |     bias_filler {
1909 |       type: "constant"
1910 |       value: 0.0
1911 |     }
1912 |   }
1913 | }
1914 | layer {
1915 |   name: "fire11_mbox_loc_perm"
1916 |   type: "Permute"
1917 |   bottom: "fire11_mbox_loc"
1918 |   top: "fire11_mbox_loc_perm"
1919 |   permute_param {
1920 |     order: 0
1921 |     order: 2
1922 |     order: 3
1923 |     order: 1
1924 |   }
1925 | }
1926 | layer {
1927 |   name: "fire11_mbox_loc_flat"
1928 |   type: "Flatten"
1929 |   bottom: "fire11_mbox_loc_perm"
1930 |   top: "fire11_mbox_loc_flat"
1931 |   flatten_param {
1932 |     axis: 1
1933 |   }
1934 | }
1935 | layer {
1936 |   name: "fire11_mbox_conf"
1937 |   type: "Convolution"
1938 |   bottom: "fire11/concat"
1939 |   top: "fire11_mbox_conf"
1940 |   param {
1941 |     lr_mult: 1.0
1942 |     decay_mult: 1.0
1943 |   }
1944 |   param {
1945 |     lr_mult: 2.0
1946 |     decay_mult: 0.0
1947 |   }
1948 |   convolution_param {
1949 |     num_output: 126
1950 |     pad: 1
1951 |     kernel_size: 3
1952 |     stride: 1
1953 |     weight_filler {
1954 |       type: "msra"
1955 |     }
1956 |     bias_filler {
1957 |       type: "constant"
1958 |       value: 0.0
1959 |     }
1960 |   }
1961 | }
1962 | layer {
1963 |   name: "fire11_mbox_conf_perm"
1964 |   type: "Permute"
1965 |   bottom: "fire11_mbox_conf"
1966 |   top: "fire11_mbox_conf_perm"
1967 |   permute_param {
1968 |     order: 0
1969 |     order: 2
1970 |     order: 3
1971 |     order: 1
1972 |   }
1973 | }
1974 | layer {
1975 |   name: "fire11_mbox_conf_flat"
1976 |   type: "Flatten"
1977 |   bottom: "fire11_mbox_conf_perm"
1978 |   top: "fire11_mbox_conf_flat"
1979 |   flatten_param {
1980 |     axis: 1
1981 |   }
1982 | }
1983 | layer {
1984 |   name: "fire11_mbox_priorbox"
1985 |   type: "PriorBox"
1986 |   bottom: "fire11/concat"
1987 |   bottom: "data"
1988 |   top: "fire11_mbox_priorbox"
1989 |   prior_box_param {
1990 |     min_size: 153.0
1991 |     max_size: 207.0
1992 |     aspect_ratio: 2.0
1993 |     aspect_ratio: 3.0
1994 |     flip: true
1995 |     clip: false
1996 |     variance: 0.1
1997 |     variance: 0.1
1998 |     variance: 0.2
1999 |     variance: 0.2
2000 |     step: 64
2001 |   }
2002 | }
2003 | #=====conv12 mbox
2004 | layer {
2005 |   name: "conv12_2_mbox_loc"
2006 |   type: "Convolution"
2007 |   bottom: "conv12_2"
2008 |   top: "conv12_2_mbox_loc"
2009 |   param {
2010 |     lr_mult: 1.0
2011 |     decay_mult: 1.0
2012 |   }
2013 |   param {
2014 |     lr_mult: 2.0
2015 |     decay_mult: 0.0
2016 |   }
2017 |   convolution_param {
2018 |     num_output: 24
2019 |     pad: 1
2020 |     kernel_size: 3
2021 |     stride: 1
2022 |     weight_filler {
2023 |       type: "msra"
2024 |     }
2025 |     bias_filler {
2026 |       type: "constant"
2027 |       value: 0.0
2028 |     }
2029 |   }
2030 | }
2031 | layer {
2032 |   name: "conv12_2_mbox_loc_perm"
2033 |   type: "Permute"
2034 |   bottom: "conv12_2_mbox_loc"
2035 |   top: "conv12_2_mbox_loc_perm"
2036 |   permute_param {
2037 |     order: 0
2038 |     order: 2
2039 |     order: 3
2040 |     order: 1
2041 |   }
2042 | }
2043 | layer {
2044 |   name: "conv12_2_mbox_loc_flat"
2045 |   type: "Flatten"
2046 |   bottom: "conv12_2_mbox_loc_perm"
2047 |   top: "conv12_2_mbox_loc_flat"
2048 |   flatten_param {
2049 |     axis: 1
2050 |   }
2051 | }
2052 | layer {
2053 |   name: "conv12_2_mbox_conf"
2054 |   type: "Convolution"
2055 |   bottom: "conv12_2"
2056 |   top: "conv12_2_mbox_conf"
2057 |   param {
2058 |     lr_mult: 1.0
2059 |     decay_mult: 1.0
2060 |   }
2061 |   param {
2062 |     lr_mult: 2.0
2063 |     decay_mult: 0.0
2064 |   }
2065 |   convolution_param {
2066 |     num_output: 126
2067 |     pad: 1
2068 |     kernel_size: 3
2069 |     stride: 1
2070 |     weight_filler {
2071 |       type: "msra"
2072 |     }
2073 |     bias_filler {
2074 |       type: "constant"
2075 |       value: 0.0
2076 |     }
2077 |   }
2078 | }
2079 | layer {
2080 |   name: "conv12_2_mbox_conf_perm"
2081 |   type: "Permute"
2082 |   bottom: "conv12_2_mbox_conf"
2083 |   top: "conv12_2_mbox_conf_perm"
2084 |   permute_param {
2085 |     order: 0
2086 |     order: 2
2087 |     order: 3
2088 |     order: 1
2089 |   }
2090 | }
2091 | layer {
2092 |   name: "conv12_2_mbox_conf_flat"
2093 |   type: "Flatten"
2094 |   bottom: "conv12_2_mbox_conf_perm"
2095 |   top: "conv12_2_mbox_conf_flat"
2096 |   flatten_param {
2097 |     axis: 1
2098 |   }
2099 | }
2100 | layer {
2101 |   name: "conv12_2_mbox_priorbox"
2102 |   type: "PriorBox"
2103 |   bottom: "conv12_2"
2104 |   bottom: "data"
2105 |   top: "conv12_2_mbox_priorbox"
2106 |   prior_box_param {
2107 |     min_size: 207.0
2108 |     max_size: 261.0
2109 |     aspect_ratio: 2.0
2110 |     aspect_ratio: 3.0
2111 |     flip: true
2112 |     clip: false
2113 |     variance: 0.1
2114 |     variance: 0.1
2115 |     variance: 0.2
2116 |     variance: 0.2
2117 |     step: 100
2118 |   }
2119 | }
2120 | #=====conv13 mbox
2121 | layer {
2122 |   name: "conv13_2_mbox_loc"
2123 |   type: "Convolution"
2124 |   bottom: "conv13_2"
2125 |   top: "conv13_2_mbox_loc"
2126 |   param {
2127 |     lr_mult: 1.0
2128 |     decay_mult: 1.0
2129 |   }
2130 |   param {
2131 |     lr_mult: 2.0
2132 |     decay_mult: 0.0
2133 |   }
2134 |   convolution_param {
2135 |     num_output: 16
2136 |     pad: 1
2137 |     kernel_size: 3
2138 |     stride: 1
2139 |     weight_filler {
2140 |       type: "msra"
2141 |     }
2142 |     bias_filler {
2143 |       type: "constant"
2144 |       value: 0.0
2145 |     }
2146 |   }
2147 | }
2148 | layer {
2149 |   name: "conv13_2_mbox_loc_perm"
2150 |   type: "Permute"
2151 |   bottom: "conv13_2_mbox_loc"
2152 |   top: "conv13_2_mbox_loc_perm"
2153 |   permute_param {
2154 |     order: 0
2155 |     order: 2
2156 |     order: 3
2157 |     order: 1
2158 |   }
2159 | }
2160 | layer {
2161 |   name: "conv13_2_mbox_loc_flat"
2162 |   type: "Flatten"
2163 |   bottom: "conv13_2_mbox_loc_perm"
2164 |   top: "conv13_2_mbox_loc_flat"
2165 |   flatten_param {
2166 |     axis: 1
2167 |   }
2168 | }
2169 | layer {
2170 |   name: "conv13_2_mbox_conf"
2171 |   type: "Convolution"
2172 |   bottom: "conv13_2"
2173 |   top: "conv13_2_mbox_conf"
2174 |   param {
2175 |     lr_mult: 1.0
2176 |     decay_mult: 1.0
2177 |   }
2178 |   param {
2179 |     lr_mult: 2.0
2180 |     decay_mult: 0.0
2181 |   }
2182 |   convolution_param {
2183 |     num_output: 84
2184 |     pad: 1
2185 |     kernel_size: 3
2186 |     stride: 1
2187 |     weight_filler {
2188 |       type: "msra"
2189 |     }
2190 |     bias_filler {
2191 |       type: "constant"
2192 |       value: 0.0
2193 |     }
2194 |   }
2195 | }
2196 | layer {
2197 |   name: "conv13_2_mbox_conf_perm"
2198 |   type: "Permute"
2199 |   bottom: "conv13_2_mbox_conf"
2200 |   top: "conv13_2_mbox_conf_perm"
2201 |   permute_param {
2202 |     order: 0
2203 |     order: 2
2204 |     order: 3
2205 |     order: 1
2206 |   }
2207 | }
2208 | layer {
2209 |   name: "conv13_2_mbox_conf_flat"
2210 |   type: "Flatten"
2211 |   bottom: "conv13_2_mbox_conf_perm"
2212 |   top: "conv13_2_mbox_conf_flat"
2213 |   flatten_param {
2214 |     axis: 1
2215 |   }
2216 | }
2217 | layer {
2218 |   name: "conv13_2_mbox_priorbox"
2219 |   type: "PriorBox"
2220 |   bottom: "conv13_2"
2221 |   bottom: "data"
2222 |   top: "conv13_2_mbox_priorbox"
2223 |   prior_box_param {
2224 |     min_size: 261.0
2225 |     max_size: 315.0
2226 |     aspect_ratio: 2.0
2227 |     flip: true
2228 |     clip: false
2229 |     variance: 0.1
2230 |     variance: 0.1
2231 |     variance: 0.2
2232 |     variance: 0.2
2233 |     step: 300
2234 |   }
2235 | }
2236 | layer {
2237 |   name: "mbox_loc"
2238 |   type: "Concat"
2239 |   bottom: "fire5_mbox_loc_flat"
2240 |   bottom: "fire9_mbox_loc_flat"
2241 |   bottom: "fire10_mbox_loc_flat"
2242 |   bottom: "fire11_mbox_loc_flat"
2243 |   bottom: "conv12_2_mbox_loc_flat"
2244 |   bottom: "conv13_2_mbox_loc_flat"
2245 |   top: "mbox_loc"
2246 |   concat_param {
2247 |     axis: 1
2248 |   }
2249 | }
2250 | layer {
2251 |   name: "mbox_conf"
2252 |   type: "Concat"
2253 |   bottom: "fire5_mbox_conf_flat"
2254 |   bottom: "fire9_mbox_conf_flat"
2255 |   bottom: "fire10_mbox_conf_flat"
2256 |   bottom: "fire11_mbox_conf_flat"
2257 |   bottom: "conv12_2_mbox_conf_flat"
2258 |   bottom: "conv13_2_mbox_conf_flat"
2259 |   top: "mbox_conf"
2260 |   concat_param {
2261 |     axis: 1
2262 |   }
2263 | }
2264 | layer {
2265 |   name: "mbox_priorbox"
2266 |   type: "Concat"
2267 |   bottom: "fire5_mbox_priorbox"
2268 |   bottom: "fire9_mbox_priorbox"
2269 |   bottom: "fire10_mbox_priorbox"
2270 |   bottom: "fire11_mbox_priorbox"
2271 |   bottom: "conv12_2_mbox_priorbox"
2272 |   bottom: "conv13_2_mbox_priorbox"
2273 |   top: "mbox_priorbox"
2274 |   concat_param {
2275 |     axis: 2
2276 |   }
2277 | }
2278 | layer {
2279 |   name: "mbox_conf_reshape"
2280 |   type: "Reshape"
2281 |   bottom: "mbox_conf"
2282 |   top: "mbox_conf_reshape"
2283 |   reshape_param {
2284 |     shape {
2285 |       dim: 0
2286 |       dim: -1
2287 |       dim: 21
2288 |     }
2289 |   }
2290 | }
2291 | layer {
2292 |   name: "mbox_conf_softmax"
2293 |   type: "Softmax"
2294 |   bottom: "mbox_conf_reshape"
2295 |   top: "mbox_conf_softmax"
2296 |   softmax_param {
2297 |     axis: 2
2298 |   }
2299 | }
2300 | layer {
2301 |   name: "mbox_conf_flatten"
2302 |   type: "Flatten"
2303 |   bottom: "mbox_conf_softmax"
2304 |   top: "mbox_conf_flatten"
2305 |   flatten_param {
2306 |     axis: 1
2307 |   }
2308 | }
2309 | layer {
2310 |   name: "detection_out"
2311 |   type: "DetectionOutput"
2312 |   bottom: "mbox_loc"
2313 |   bottom: "mbox_conf_flatten"
2314 |   bottom: "mbox_priorbox"
2315 |   top: "detection_out"
2316 |   include {
2317 |     phase: TEST
2318 |   }
2319 |   detection_output_param {
2320 |     num_classes: 21
2321 |     share_location: true
2322 |     background_label_id: 0
2323 |     nms_param {
2324 |       nms_threshold: 0.45
2325 |       top_k: 400
2326 |     }
2327 |     code_type: CENTER_SIZE
2328 |     keep_top_k: 200
2329 |     confidence_threshold: 0.01
2330 |   }
2331 | }
2332 | layer {
2333 |   name: "detection_eval"
2334 |   type: "DetectionEvaluate"
2335 |   bottom: "detection_out"
2336 |   bottom: "label"
2337 |   top: "detection_eval"
2338 |   include {
2339 |     phase: TEST
2340 |   }
2341 |   detection_evaluate_param {
2342 |     num_classes: 21
2343 |     background_label_id: 0
2344 |     overlap_threshold: 0.5
2345 |     evaluate_difficult_gt: false
2346 |   }
2347 | }
2348 | 


--------------------------------------------------------------------------------
/model/tiny_ssd_voc_train.prototxt:
--------------------------------------------------------------------------------
   1 | name: "tiny-ssd"
   2 | layer {
   3 |   name: "data"
   4 |   type: "AnnotatedData"
   5 |   top: "data"
   6 |   top: "label"
   7 |   include {
   8 |     phase: TRAIN
   9 |   }
  10 |   transform_param {
  11 |     mirror: true
  12 |     mean_value: 104
  13 |     mean_value: 117
  14 |     mean_value: 123
  15 |     resize_param {
  16 |       prob: 1.0
  17 |       resize_mode: WARP
  18 |       height: 300
  19 |       width: 300
  20 |       interp_mode: LINEAR
  21 |       interp_mode: AREA
  22 |       interp_mode: NEAREST
  23 |       interp_mode: CUBIC
  24 |       interp_mode: LANCZOS4
  25 |     }
  26 |     emit_constraint {
  27 |       emit_type: CENTER
  28 |     }
  29 |     distort_param {
  30 |       brightness_prob: 0.5
  31 |       brightness_delta: 32.0
  32 |       contrast_prob: 0.5
  33 |       contrast_lower: 0.5
  34 |       contrast_upper: 1.5
  35 |       hue_prob: 0.5
  36 |       hue_delta: 18.0
  37 |       saturation_prob: 0.5
  38 |       saturation_lower: 0.5
  39 |       saturation_upper: 1.5
  40 |       random_order_prob: 0.0
  41 |     }
  42 |     expand_param {
  43 |       prob: 0.5
  44 |       max_expand_ratio: 4.0
  45 |     }
  46 |   }
  47 |   data_param {
  48 |     source: "examples/VOC0712/VOC0712_trainval_lmdb"
  49 |     batch_size: 32
  50 |     backend: LMDB
  51 |   }
  52 |   annotated_data_param {
  53 |     batch_sampler {
  54 |       max_sample: 1
  55 |       max_trials: 1
  56 |     }
  57 |     batch_sampler {
  58 |       sampler {
  59 |         min_scale: 0.3
  60 |         max_scale: 1.0
  61 |         min_aspect_ratio: 0.5
  62 |         max_aspect_ratio: 2.0
  63 |       }
  64 |       sample_constraint {
  65 |         min_jaccard_overlap: 0.1
  66 |       }
  67 |       max_sample: 1
  68 |       max_trials: 50
  69 |     }
  70 |     batch_sampler {
  71 |       sampler {
  72 |         min_scale: 0.3
  73 |         max_scale: 1.0
  74 |         min_aspect_ratio: 0.5
  75 |         max_aspect_ratio: 2.0
  76 |       }
  77 |       sample_constraint {
  78 |         min_jaccard_overlap: 0.3
  79 |       }
  80 |       max_sample: 1
  81 |       max_trials: 50
  82 |     }
  83 |     batch_sampler {
  84 |       sampler {
  85 |         min_scale: 0.3
  86 |         max_scale: 1.0
  87 |         min_aspect_ratio: 0.5
  88 |         max_aspect_ratio: 2.0
  89 |       }
  90 |       sample_constraint {
  91 |         min_jaccard_overlap: 0.5
  92 |       }
  93 |       max_sample: 1
  94 |       max_trials: 50
  95 |     }
  96 |     batch_sampler {
  97 |       sampler {
  98 |         min_scale: 0.3
  99 |         max_scale: 1.0
 100 |         min_aspect_ratio: 0.5
 101 |         max_aspect_ratio: 2.0
 102 |       }
 103 |       sample_constraint {
 104 |         min_jaccard_overlap: 0.7
 105 |       }
 106 |       max_sample: 1
 107 |       max_trials: 50
 108 |     }
 109 |     batch_sampler {
 110 |       sampler {
 111 |         min_scale: 0.3
 112 |         max_scale: 1.0
 113 |         min_aspect_ratio: 0.5
 114 |         max_aspect_ratio: 2.0
 115 |       }
 116 |       sample_constraint {
 117 |         min_jaccard_overlap: 0.9
 118 |       }
 119 |       max_sample: 1
 120 |       max_trials: 50
 121 |     }
 122 |     batch_sampler {
 123 |       sampler {
 124 |         min_scale: 0.3
 125 |         max_scale: 1.0
 126 |         min_aspect_ratio: 0.5
 127 |         max_aspect_ratio: 2.0
 128 |       }
 129 |       sample_constraint {
 130 |         max_jaccard_overlap: 1.0
 131 |       }
 132 |       max_sample: 1
 133 |       max_trials: 50
 134 |     }
 135 |     label_map_file: "data/VOC0712/labelmap.prototxt"
 136 |   }
 137 | }
 138 | layer {
 139 |   name: "conv1"
 140 |   type: "Convolution"
 141 |   bottom: "data"
 142 |   top: "conv1"
 143 |   param {
 144 |     lr_mult: 1.0
 145 |     decay_mult: 0.0
 146 |   }
 147 |   param {
 148 |     lr_mult: 1.0
 149 |     decay_mult: 0.0
 150 |   }
 151 |   convolution_param {
 152 |     num_output: 57
 153 |     kernel_size: 3
 154 |     stride: 2
 155 |     weight_filler {
 156 |       type: "xavier"
 157 |     }
 158 |   }
 159 | }
 160 | layer {
 161 |   name: "relu_conv1"
 162 |   type: "ReLU"
 163 |   bottom: "conv1"
 164 |   top: "conv1"
 165 | }
 166 | layer {
 167 |   name: "pool1"
 168 |   type: "Pooling"
 169 |   bottom: "conv1"
 170 |   top: "pool1"
 171 |   pooling_param {
 172 |     pool: MAX
 173 |     kernel_size: 3
 174 |     stride: 2
 175 |   }
 176 | }
 177 | layer {
 178 |   name: "fire2/squeeze1x1"
 179 |   type: "Convolution"
 180 |   bottom: "pool1"
 181 |   top: "fire2/squeeze1x1"
 182 |   param {
 183 |     lr_mult: 1.0
 184 |     decay_mult: 0.0
 185 |   }
 186 |   param {
 187 |     lr_mult: 1.0
 188 |     decay_mult: 0.0
 189 |   }
 190 |   convolution_param {
 191 |     num_output: 15
 192 |     kernel_size: 1
 193 |     weight_filler {
 194 |       type: "xavier"
 195 |     }
 196 |   }
 197 | }
 198 | layer {
 199 |   name: "fire2/expand1x1"
 200 |   type: "Convolution"
 201 |   bottom: "fire2/squeeze1x1"
 202 |   top: "fire2/expand1x1"
 203 |   param {
 204 |     lr_mult: 1.0
 205 |     decay_mult: 0.0
 206 |   }
 207 |   param {
 208 |     lr_mult: 1.0
 209 |     decay_mult: 0.0
 210 |   }
 211 |   convolution_param {
 212 |     num_output: 49
 213 |     kernel_size: 1
 214 |     weight_filler {
 215 |       type: "xavier"
 216 |     }
 217 |   }
 218 | }
 219 | layer {
 220 |   name: "fire2/relu_expand1x1"
 221 |   type: "ReLU"
 222 |   bottom: "fire2/expand1x1"
 223 |   top: "fire2/expand1x1"
 224 | }
 225 | layer {
 226 |   name: "fire2/expand3x3"
 227 |   type: "Convolution"
 228 |   bottom: "fire2/squeeze1x1"
 229 |   top: "fire2/expand3x3"
 230 |   param {
 231 |     lr_mult: 1.0
 232 |     decay_mult: 0.0
 233 |   }
 234 |   param {
 235 |     lr_mult: 1.0
 236 |     decay_mult: 0.0
 237 |   }
 238 |   convolution_param {
 239 |     num_output: 53
 240 |     pad: 1
 241 |     kernel_size: 3
 242 |     weight_filler {
 243 |       type: "xavier"
 244 |     }
 245 |   }
 246 | }
 247 | layer {
 248 |   name: "fire2/relu_expand3x3"
 249 |   type: "ReLU"
 250 |   bottom: "fire2/expand3x3"
 251 |   top: "fire2/expand3x3"
 252 | }
 253 | layer {
 254 |   name: "fire2/concat"
 255 |   type: "Concat"
 256 |   bottom: "fire2/expand1x1"
 257 |   bottom: "fire2/expand3x3"
 258 |   top: "fire2/concat"
 259 | }
 260 | layer {
 261 |   name: "fire3/squeeze1x1"
 262 |   type: "Convolution"
 263 |   bottom: "fire2/concat"
 264 |   top: "fire3/squeeze1x1"
 265 |   param {
 266 |     lr_mult: 1.0
 267 |     decay_mult: 0.0
 268 |   }
 269 |   param {
 270 |     lr_mult: 1.0
 271 |     decay_mult: 0.0
 272 |   }
 273 |   convolution_param {
 274 |     num_output: 15
 275 |     kernel_size: 1
 276 |     weight_filler {
 277 |       type: "xavier"
 278 |     }
 279 |   }
 280 | }
 281 | layer {
 282 |   name: "fire3/expand1x1"
 283 |   type: "Convolution"
 284 |   bottom: "fire3/squeeze1x1"
 285 |   top: "fire3/expand1x1"
 286 |   param {
 287 |     lr_mult: 1.0
 288 |     decay_mult: 0.0
 289 |   }
 290 |   param {
 291 |     lr_mult: 1.0
 292 |     decay_mult: 0.0
 293 |   }
 294 |   convolution_param {
 295 |     num_output: 54
 296 |     kernel_size: 1
 297 |     weight_filler {
 298 |       type: "xavier"
 299 |     }
 300 |   }
 301 | }
 302 | layer {
 303 |   name: "fire3/relu_expand1x1"
 304 |   type: "ReLU"
 305 |   bottom: "fire3/expand1x1"
 306 |   top: "fire3/expand1x1"
 307 | }
 308 | layer {
 309 |   name: "fire3/expand3x3"
 310 |   type: "Convolution"
 311 |   bottom: "fire3/squeeze1x1"
 312 |   top: "fire3/expand3x3"
 313 |   param {
 314 |     lr_mult: 1.0
 315 |     decay_mult: 0.0
 316 |   }
 317 |   param {
 318 |     lr_mult: 1.0
 319 |     decay_mult: 0.0
 320 |   }
 321 |   convolution_param {
 322 |     num_output: 52
 323 |     pad: 1
 324 |     kernel_size: 3
 325 |     weight_filler {
 326 |       type: "xavier"
 327 |     }
 328 |   }
 329 | }
 330 | layer {
 331 |   name: "fire3/relu_expand3x3"
 332 |   type: "ReLU"
 333 |   bottom: "fire3/expand3x3"
 334 |   top: "fire3/expand3x3"
 335 | }
 336 | layer {
 337 |   name: "fire3/concat"
 338 |   type: "Concat"
 339 |   bottom: "fire3/expand1x1"
 340 |   bottom: "fire3/expand3x3"
 341 |   top: "fire3/concat"
 342 | }
 343 | layer {
 344 |   name: "pool3"
 345 |   type: "Pooling"
 346 |   bottom: "fire3/concat"
 347 |   top: "pool3"
 348 |   pooling_param {
 349 |     pool: MAX
 350 |     kernel_size: 3
 351 |     stride: 2
 352 |   }
 353 | }
 354 | layer {
 355 |   name: "fire4/squeeze1x1"
 356 |   type: "Convolution"
 357 |   bottom: "pool3"
 358 |   top: "fire4/squeeze1x1"
 359 |   param {
 360 |     lr_mult: 1.0
 361 |     decay_mult: 0.0
 362 |   }
 363 |   param {
 364 |     lr_mult: 1.0
 365 |     decay_mult: 0.0
 366 |   }
 367 |   convolution_param {
 368 |     num_output: 29
 369 |     kernel_size: 1
 370 |     weight_filler {
 371 |       type: "xavier"
 372 |     }
 373 |   }
 374 | }
 375 | layer {
 376 |   name: "fire4/expand1x1"
 377 |   type: "Convolution"
 378 |   bottom: "fire4/squeeze1x1"
 379 |   top: "fire4/expand1x1"
 380 |   param {
 381 |     lr_mult: 1.0
 382 |     decay_mult: 0.0
 383 |   }
 384 |   param {
 385 |     lr_mult: 1.0
 386 |     decay_mult: 0.0
 387 |   }
 388 |   convolution_param {
 389 |     num_output: 92
 390 |     kernel_size: 1
 391 |     weight_filler {
 392 |       type: "xavier"
 393 |     }
 394 |   }
 395 | }
 396 | layer {
 397 |   name: "fire4/relu_expand1x1"
 398 |   type: "ReLU"
 399 |   bottom: "fire4/expand1x1"
 400 |   top: "fire4/expand1x1"
 401 | }
 402 | layer {
 403 |   name: "fire4/expand3x3"
 404 |   type: "Convolution"
 405 |   bottom: "fire4/squeeze1x1"
 406 |   top: "fire4/expand3x3"
 407 |   param {
 408 |     lr_mult: 1.0
 409 |     decay_mult: 0.0
 410 |   }
 411 |   param {
 412 |     lr_mult: 1.0
 413 |     decay_mult: 0.0
 414 |   }
 415 |   convolution_param {
 416 |     num_output: 94
 417 |     pad: 1
 418 |     kernel_size: 3
 419 |     weight_filler {
 420 |       type: "xavier"
 421 |     }
 422 |   }
 423 | }
 424 | layer {
 425 |   name: "fire4/relu_expand3x3"
 426 |   type: "ReLU"
 427 |   bottom: "fire4/expand3x3"
 428 |   top: "fire4/expand3x3"
 429 | }
 430 | layer {
 431 |   name: "fire4/concat"
 432 |   type: "Concat"
 433 |   bottom: "fire4/expand1x1"
 434 |   bottom: "fire4/expand3x3"
 435 |   top: "fire4/concat"
 436 | }
 437 | layer {
 438 |   name: "fire5/squeeze1x1"
 439 |   type: "Convolution"
 440 |   bottom: "fire4/concat"
 441 |   top: "fire5/squeeze1x1"
 442 |   param {
 443 |     lr_mult: 1.0
 444 |     decay_mult: 0.0
 445 |   }
 446 |   param {
 447 |     lr_mult: 1.0
 448 |     decay_mult: 0.0
 449 |   }
 450 |   convolution_param {
 451 |     num_output: 29
 452 |     kernel_size: 1
 453 |     weight_filler {
 454 |       type: "xavier"
 455 |     }
 456 |   }
 457 | }
 458 | layer {
 459 |   name: "fire5/expand1x1"
 460 |   type: "Convolution"
 461 |   bottom: "fire5/squeeze1x1"
 462 |   top: "fire5/expand1x1"
 463 |   param {
 464 |     lr_mult: 1.0
 465 |     decay_mult: 0.0
 466 |   }
 467 |   param {
 468 |     lr_mult: 1.0
 469 |     decay_mult: 0.0
 470 |   }
 471 |   convolution_param {
 472 |     num_output: 90
 473 |     kernel_size: 1
 474 |     weight_filler {
 475 |       type: "xavier"
 476 |     }
 477 |   }
 478 | }
 479 | layer {
 480 |   name: "fire5/relu_expand1x1"
 481 |   type: "ReLU"
 482 |   bottom: "fire5/expand1x1"
 483 |   top: "fire5/expand1x1"
 484 | }
 485 | layer {
 486 |   name: "fire5/expand3x3"
 487 |   type: "Convolution"
 488 |   bottom: "fire5/squeeze1x1"
 489 |   top: "fire5/expand3x3"
 490 |   param {
 491 |     lr_mult: 1.0
 492 |     decay_mult: 0.0
 493 |   }
 494 |   param {
 495 |     lr_mult: 1.0
 496 |     decay_mult: 0.0
 497 |   }
 498 |   convolution_param {
 499 |     num_output: 83
 500 |     pad: 1
 501 |     kernel_size: 3
 502 |     weight_filler {
 503 |       type: "xavier"
 504 |     }
 505 |   }
 506 | }
 507 | layer {
 508 |   name: "fire5/relu_expand3x3"
 509 |   type: "ReLU"
 510 |   bottom: "fire5/expand3x3"
 511 |   top: "fire5/expand3x3"
 512 | }
 513 | layer {
 514 |   name: "fire5/concat"
 515 |   type: "Concat"
 516 |   bottom: "fire5/expand1x1"
 517 |   bottom: "fire5/expand3x3"
 518 |   top: "fire5/concat"
 519 | }
 520 | layer {
 521 |   name: "pool5"
 522 |   type: "Pooling"
 523 |   bottom: "fire5/concat"
 524 |   top: "pool5"
 525 |   pooling_param {
 526 |     pool: MAX
 527 |     kernel_size: 3
 528 |     stride: 2
 529 |   }
 530 | }
 531 | layer {
 532 |   name: "fire6/squeeze1x1"
 533 |   type: "Convolution"
 534 |   bottom: "pool5"
 535 |   top: "fire6/squeeze1x1"
 536 |   param {
 537 |     lr_mult: 1.0
 538 |     decay_mult: 0.0
 539 |   }
 540 |   param {
 541 |     lr_mult: 1.0
 542 |     decay_mult: 0.0
 543 |   }
 544 |   convolution_param {
 545 |     num_output: 44
 546 |     kernel_size: 1
 547 |     weight_filler {
 548 |       type: "xavier"
 549 |     }
 550 |   }
 551 | }
 552 | layer {
 553 |   name: "fire6/expand1x1"
 554 |   type: "Convolution"
 555 |   bottom: "fire6/squeeze1x1"
 556 |   top: "fire6/expand1x1"
 557 |   param {
 558 |     lr_mult: 1.0
 559 |     decay_mult: 0.0
 560 |   }
 561 |   param {
 562 |     lr_mult: 1.0
 563 |     decay_mult: 0.0
 564 |   }
 565 |   convolution_param {
 566 |     num_output: 166
 567 |     kernel_size: 1
 568 |     weight_filler {
 569 |       type: "xavier"
 570 |     }
 571 |   }
 572 | }
 573 | layer {
 574 |   name: "fire6/relu_expand1x1"
 575 |   type: "ReLU"
 576 |   bottom: "fire6/expand1x1"
 577 |   top: "fire6/expand1x1"
 578 | }
 579 | layer {
 580 |   name: "fire6/expand3x3"
 581 |   type: "Convolution"
 582 |   bottom: "fire6/squeeze1x1"
 583 |   top: "fire6/expand3x3"
 584 |   param {
 585 |     lr_mult: 1.0
 586 |     decay_mult: 0.0
 587 |   }
 588 |   param {
 589 |     lr_mult: 1.0
 590 |     decay_mult: 0.0
 591 |   }
 592 |   convolution_param {
 593 |     num_output: 161
 594 |     pad: 1
 595 |     kernel_size: 3
 596 |     weight_filler {
 597 |       type: "xavier"
 598 |     }
 599 |   }
 600 | }
 601 | layer {
 602 |   name: "fire6/relu_expand3x3"
 603 |   type: "ReLU"
 604 |   bottom: "fire6/expand3x3"
 605 |   top: "fire6/expand3x3"
 606 | }
 607 | layer {
 608 |   name: "fire6/concat"
 609 |   type: "Concat"
 610 |   bottom: "fire6/expand1x1"
 611 |   bottom: "fire6/expand3x3"
 612 |   top: "fire6/concat"
 613 | }
 614 | layer {
 615 |   name: "fire7/squeeze1x1"
 616 |   type: "Convolution"
 617 |   bottom: "fire6/concat"
 618 |   top: "fire7/squeeze1x1"
 619 |   param {
 620 |     lr_mult: 1.0
 621 |     decay_mult: 0.0
 622 |   }
 623 |   param {
 624 |     lr_mult: 1.0
 625 |     decay_mult: 0.0
 626 |   }
 627 |   convolution_param {
 628 |     num_output: 45
 629 |     kernel_size: 1
 630 |     weight_filler {
 631 |       type: "xavier"
 632 |     }
 633 |   }
 634 | }
 635 | layer {
 636 |   name: "fire7/expand1x1"
 637 |   type: "Convolution"
 638 |   bottom: "fire7/squeeze1x1"
 639 |   top: "fire7/expand1x1"
 640 |   param {
 641 |     lr_mult: 1.0
 642 |     decay_mult: 0.0
 643 |   }
 644 |   param {
 645 |     lr_mult: 1.0
 646 |     decay_mult: 0.0
 647 |   }
 648 |   convolution_param {
 649 |     num_output: 155
 650 |     kernel_size: 1
 651 |     weight_filler {
 652 |       type: "xavier"
 653 |     }
 654 |   }
 655 | }
 656 | layer {
 657 |   name: "fire7/relu_expand1x1"
 658 |   type: "ReLU"
 659 |   bottom: "fire7/expand1x1"
 660 |   top: "fire7/expand1x1"
 661 | }
 662 | layer {
 663 |   name: "fire7/expand3x3"
 664 |   type: "Convolution"
 665 |   bottom: "fire7/squeeze1x1"
 666 |   top: "fire7/expand3x3"
 667 |   param {
 668 |     lr_mult: 1.0
 669 |     decay_mult: 0.0
 670 |   }
 671 |   param {
 672 |     lr_mult: 1.0
 673 |     decay_mult: 0.0
 674 |   }
 675 |   convolution_param {
 676 |     num_output: 146
 677 |     pad: 1
 678 |     kernel_size: 3
 679 |     weight_filler {
 680 |       type: "xavier"
 681 |     }
 682 |   }
 683 | }
 684 | layer {
 685 |   name: "fire7/relu_expand3x3"
 686 |   type: "ReLU"
 687 |   bottom: "fire7/expand3x3"
 688 |   top: "fire7/expand3x3"
 689 | }
 690 | layer {
 691 |   name: "fire7/concat"
 692 |   type: "Concat"
 693 |   bottom: "fire7/expand1x1"
 694 |   bottom: "fire7/expand3x3"
 695 |   top: "fire7/concat"
 696 | }
 697 | layer {
 698 |   name: "fire8/squeeze1x1"
 699 |   type: "Convolution"
 700 |   bottom: "fire7/concat"
 701 |   top: "fire8/squeeze1x1"
 702 |   param {
 703 |     lr_mult: 1.0
 704 |     decay_mult: 0.0
 705 |   }
 706 |   param {
 707 |     lr_mult: 1.0
 708 |     decay_mult: 0.0
 709 |   }
 710 |   convolution_param {
 711 |     num_output: 49
 712 |     kernel_size: 1
 713 |     weight_filler {
 714 |       type: "xavier"
 715 |     }
 716 |   }
 717 | }
 718 | layer {
 719 |   name: "fire8/expand1x1"
 720 |   type: "Convolution"
 721 |   bottom: "fire8/squeeze1x1"
 722 |   top: "fire8/expand1x1"
 723 |   param {
 724 |     lr_mult: 1.0
 725 |     decay_mult: 0.0
 726 |   }
 727 |   param {
 728 |     lr_mult: 1.0
 729 |     decay_mult: 0.0
 730 |   }
 731 |   convolution_param {
 732 |     num_output: 163
 733 |     kernel_size: 1
 734 |     weight_filler {
 735 |       type: "xavier"
 736 |     }
 737 |   }
 738 | }
 739 | layer {
 740 |   name: "fire8/relu_expand1x1"
 741 |   type: "ReLU"
 742 |   bottom: "fire8/expand1x1"
 743 |   top: "fire8/expand1x1"
 744 | }
 745 | layer {
 746 |   name: "fire8/expand3x3"
 747 |   type: "Convolution"
 748 |   bottom: "fire8/squeeze1x1"
 749 |   top: "fire8/expand3x3"
 750 |   param {
 751 |     lr_mult: 1.0
 752 |     decay_mult: 0.0
 753 |   }
 754 |   param {
 755 |     lr_mult: 1.0
 756 |     decay_mult: 0.0
 757 |   }
 758 |   convolution_param {
 759 |     num_output: 171
 760 |     pad: 1
 761 |     kernel_size: 3
 762 |     weight_filler {
 763 |       type: "xavier"
 764 |     }
 765 |   }
 766 | }
 767 | layer {
 768 |   name: "fire8/relu_expand3x3"
 769 |   type: "ReLU"
 770 |   bottom: "fire8/expand3x3"
 771 |   top: "fire8/expand3x3"
 772 | }
 773 | layer {
 774 |   name: "fire8/concat"
 775 |   type: "Concat"
 776 |   bottom: "fire8/expand1x1"
 777 |   bottom: "fire8/expand3x3"
 778 |   top: "fire8/concat"
 779 | }
 780 | layer {
 781 |   name: "fire9/squeeze1x1"
 782 |   type: "Convolution"
 783 |   bottom: "fire8/concat"
 784 |   top: "fire9/squeeze1x1"
 785 |   param {
 786 |     lr_mult: 1.0
 787 |     decay_mult: 0.0
 788 |   }
 789 |   param {
 790 |     lr_mult: 1.0
 791 |     decay_mult: 0.0
 792 |   }
 793 |   convolution_param {
 794 |     num_output: 25
 795 |     kernel_size: 1
 796 |     weight_filler {
 797 |       type: "xavier"
 798 |     }
 799 |   }
 800 | }
 801 | layer {
 802 |   name: "fire9/expand1x1"
 803 |   type: "Convolution"
 804 |   bottom: "fire9/squeeze1x1"
 805 |   top: "fire9/expand1x1"
 806 |   param {
 807 |     lr_mult: 1.0
 808 |     decay_mult: 0.0
 809 |   }
 810 |   param {
 811 |     lr_mult: 1.0
 812 |     decay_mult: 0.0
 813 |   }
 814 |   convolution_param {
 815 |     num_output: 29
 816 |     kernel_size: 1
 817 |     weight_filler {
 818 |       type: "xavier"
 819 |     }
 820 |   }
 821 | }
 822 | layer {
 823 |   name: "fire9/expand1x1/bn"
 824 |   type: "BatchNorm"
 825 |   bottom: "fire9/expand1x1"
 826 |   top: "fire9/expand1x1"
 827 |   param {
 828 |     lr_mult: 0
 829 |     decay_mult: 0
 830 |   }
 831 |   param {
 832 |     lr_mult: 0
 833 |     decay_mult: 0
 834 |   }
 835 |   param {
 836 |     lr_mult: 0
 837 |     decay_mult: 0
 838 |   }
 839 | }
 840 | layer {
 841 |   name: "fire9/expand1x1/scale"
 842 |   type: "Scale"
 843 |   bottom: "fire9/expand1x1"
 844 |   top: "fire9/expand1x1"
 845 |   param {
 846 |     lr_mult: 1.0
 847 |     decay_mult: 0.0
 848 |   }
 849 |   param {
 850 |     lr_mult: 2.0
 851 |     decay_mult: 0.0
 852 |   }
 853 |   scale_param {
 854 |     filler {
 855 |       value: 1
 856 |     }
 857 |     bias_term: true
 858 |     bias_filler {
 859 |       value: 0
 860 |     }
 861 |   }
 862 | }
 863 | layer {
 864 |   name: "fire9/relu_expand1x1"
 865 |   type: "ReLU"
 866 |   bottom: "fire9/expand1x1"
 867 |   top: "fire9/expand1x1"
 868 | }
 869 | layer {
 870 |   name: "fire9/expand3x3"
 871 |   type: "Convolution"
 872 |   bottom: "fire9/squeeze1x1"
 873 |   top: "fire9/expand3x3"
 874 |   param {
 875 |     lr_mult: 1.0
 876 |     decay_mult: 0.0
 877 |   }
 878 |   param {
 879 |     lr_mult: 1.0
 880 |     decay_mult: 0.0
 881 |   }
 882 |   convolution_param {
 883 |     num_output: 54
 884 |     pad: 1
 885 |     kernel_size: 3
 886 |     weight_filler {
 887 |       type: "xavier"
 888 |     }
 889 |   }
 890 | }
 891 | layer {
 892 |   name: "fire9/expand3x3/bn"
 893 |   type: "BatchNorm"
 894 |   bottom: "fire9/expand3x3"
 895 |   top: "fire9/expand3x3"
 896 |   param {
 897 |     lr_mult: 0
 898 |     decay_mult: 0
 899 |   }
 900 |   param {
 901 |     lr_mult: 0
 902 |     decay_mult: 0
 903 |   }
 904 |   param {
 905 |     lr_mult: 0
 906 |     decay_mult: 0
 907 |   }
 908 | }
 909 | layer {
 910 |   name: "fire9/expand3x3/scale"
 911 |   type: "Scale"
 912 |   bottom: "fire9/expand3x3"
 913 |   top: "fire9/expand3x3"
 914 |   param {
 915 |     lr_mult: 1.0
 916 |     decay_mult: 0.0
 917 |   }
 918 |   param {
 919 |     lr_mult: 2.0
 920 |     decay_mult: 0.0
 921 |   }
 922 |   scale_param {
 923 |     filler {
 924 |       value: 1
 925 |     }
 926 |     bias_term: true
 927 |     bias_filler {
 928 |       value: 0
 929 |     }
 930 |   }
 931 | }
 932 | layer {
 933 |   name: "fire9/relu_expand3x3"
 934 |   type: "ReLU"
 935 |   bottom: "fire9/expand3x3"
 936 |   top: "fire9/expand3x3"
 937 | }
 938 | layer {
 939 |   name: "fire9/concat"
 940 |   type: "Concat"
 941 |   bottom: "fire9/expand1x1"
 942 |   bottom: "fire9/expand3x3"
 943 |   top: "fire9/concat"
 944 | }
 945 | layer {
 946 |   name: "pool9"
 947 |   type: "Pooling"
 948 |   bottom: "fire9/concat"
 949 |   top: "pool9"
 950 |   pooling_param {
 951 |     pool: MAX
 952 |     kernel_size: 3
 953 |     stride: 2
 954 |   }
 955 | }
 956 | #================ssd append layer fire10, out 9 * 9
 957 | layer {
 958 |   name: "fire10/squeeze1x1"
 959 |   type: "Convolution"
 960 |   bottom: "pool9"
 961 |   top: "fire10/squeeze1x1"
 962 |   convolution_param {
 963 |     num_output: 37
 964 |     kernel_size: 1
 965 |     weight_filler {
 966 |       type: "xavier"
 967 |     }
 968 |   }
 969 | }
 970 | layer {
 971 |   name: "fire10/squeeze1x1/bn"
 972 |   type: "BatchNorm"
 973 |   top: "fire10/squeeze1x1"
 974 |   bottom: "fire10/squeeze1x1"
 975 |   param {
 976 |     lr_mult: 0
 977 |     decay_mult: 0
 978 |   }
 979 |   param {
 980 |     lr_mult: 0
 981 |     decay_mult: 0
 982 |   }
 983 |   param {
 984 |     lr_mult: 0
 985 |     decay_mult: 0
 986 |   }
 987 | }
 988 | layer {
 989 |   name: "fire10/squeeze1x1/scale"
 990 |   type: "Scale"
 991 |   top: "fire10/squeeze1x1"
 992 |   bottom: "fire10/squeeze1x1"
 993 |   param {
 994 |     lr_mult: 1.0
 995 |     decay_mult: 0.0
 996 |   }
 997 |   param {
 998 |     lr_mult: 2.0
 999 |     decay_mult: 0.0
1000 |   }
1001 |   scale_param {
1002 |     filler {
1003 |       value: 1
1004 |     }
1005 |     bias_term: true
1006 |     bias_filler {
1007 |       value: 0
1008 |     }
1009 |   }
1010 | }
1011 | layer {
1012 |   name: "fire10/expand1x1"
1013 |   type: "Convolution"
1014 |   bottom: "fire10/squeeze1x1"
1015 |   top: "fire10/expand1x1"
1016 |   convolution_param {
1017 |     num_output: 45
1018 |     kernel_size: 1
1019 |     weight_filler {
1020 |       type: "xavier"
1021 |     }
1022 |   }
1023 | }
1024 | layer {
1025 |   name: "fire10/expand1x1/bn"
1026 |   type: "BatchNorm"
1027 |   bottom: "fire10/expand1x1"
1028 |   top: "fire10/expand1x1"
1029 |   param {
1030 |     lr_mult: 0
1031 |     decay_mult: 0
1032 |   }
1033 |   param {
1034 |     lr_mult: 0
1035 |     decay_mult: 0
1036 |   }
1037 |   param {
1038 |     lr_mult: 0
1039 |     decay_mult: 0
1040 |   }
1041 | }
1042 | layer {
1043 |   name: "fire10/expand1x1/scale"
1044 |   type: "Scale"
1045 |   bottom: "fire10/expand1x1"
1046 |   top: "fire10/expand1x1"
1047 |   param {
1048 |     lr_mult: 1.0
1049 |     decay_mult: 0.0
1050 |   }
1051 |   param {
1052 |     lr_mult: 2.0
1053 |     decay_mult: 0.0
1054 |   }
1055 |   scale_param {
1056 |     filler {
1057 |       value: 1
1058 |     }
1059 |     bias_term: true
1060 |     bias_filler {
1061 |       value: 0
1062 |     }
1063 |   }
1064 | }
1065 | layer {
1066 |   name: "fire10/relu_expand1x1"
1067 |   type: "ReLU"
1068 |   bottom: "fire10/expand1x1"
1069 |   top: "fire10/expand1x1"
1070 | }
1071 | layer {
1072 |   name: "fire10/expand3x3"
1073 |   type: "Convolution"
1074 |   bottom: "fire10/squeeze1x1"
1075 |   top: "fire10/expand3x3"
1076 |   convolution_param {
1077 |     num_output: 56
1078 |     pad: 1
1079 |     kernel_size: 3
1080 |     weight_filler {
1081 |       type: "xavier"
1082 |     }
1083 |   }
1084 | }
1085 | layer {
1086 |   name: "fire10/expand3x3/bn"
1087 |   type: "BatchNorm"
1088 |   bottom: "fire10/expand3x3"
1089 |   top: "fire10/expand3x3"
1090 |   param {
1091 |     lr_mult: 0
1092 |     decay_mult: 0
1093 |   }
1094 |   param {
1095 |     lr_mult: 0
1096 |     decay_mult: 0
1097 |   }
1098 |   param {
1099 |     lr_mult: 0
1100 |     decay_mult: 0
1101 |   }
1102 | }
1103 | layer {
1104 |   name: "fire10/expand3x3/scale"
1105 |   type: "Scale"
1106 |   bottom: "fire10/expand3x3"
1107 |   top: "fire10/expand3x3"
1108 |   param {
1109 |     lr_mult: 1.0
1110 |     decay_mult: 0.0
1111 |   }
1112 |   param {
1113 |     lr_mult: 2.0
1114 |     decay_mult: 0.0
1115 |   }
1116 |   scale_param {
1117 |     filler {
1118 |       value: 1
1119 |     }
1120 |     bias_term: true
1121 |     bias_filler {
1122 |       value: 0
1123 |     }
1124 |   }
1125 | }
1126 | layer {
1127 |   name: "fire10/relu_expand3x3"
1128 |   type: "ReLU"
1129 |   bottom: "fire10/expand3x3"
1130 |   top: "fire10/expand3x3"
1131 | }
1132 | layer {
1133 |   name: "fire10/concat"
1134 |   type: "Concat"
1135 |   bottom: "fire10/expand1x1"
1136 |   bottom: "fire10/expand3x3"
1137 |   top: "fire10/concat"
1138 | }
1139 | #================ssd append layer fire11, out 4 * 4
1140 | layer {
1141 |   name: "pool10"
1142 |   type: "Pooling"
1143 |   bottom: "fire10/concat"
1144 |   top: "pool10"
1145 |   pooling_param {
1146 |     pool: MAX
1147 |     kernel_size: 3
1148 |     stride: 2
1149 |   }
1150 | }
1151 | layer {
1152 |   name: "fire11/squeeze1x1"
1153 |   type: "Convolution"
1154 |   bottom: "pool10"
1155 |   top: "fire11/squeeze1x1"
1156 |   convolution_param {
1157 |     num_output: 38
1158 |     kernel_size: 1
1159 |     weight_filler {
1160 |       type: "xavier"
1161 |     }
1162 |   }
1163 | }
1164 | layer {
1165 |   name: "fire11/squeeze1x1/bn"
1166 |   type: "BatchNorm"
1167 |   top: "fire11/squeeze1x1"
1168 |   bottom: "fire11/squeeze1x1"
1169 |   param {
1170 |     lr_mult: 0
1171 |     decay_mult: 0
1172 |   }
1173 |   param {
1174 |     lr_mult: 0
1175 |     decay_mult: 0
1176 |   }
1177 |   param {
1178 |     lr_mult: 0
1179 |     decay_mult: 0
1180 |   }
1181 | }
1182 | layer {
1183 |   name: "fire11/squeeze1x1/scale"
1184 |   type: "Scale"
1185 |   top: "fire11/squeeze1x1"
1186 |   bottom: "fire11/squeeze1x1"
1187 |   param {
1188 |     lr_mult: 1.0
1189 |     decay_mult: 0.0
1190 |   }
1191 |   param {
1192 |     lr_mult: 2.0
1193 |     decay_mult: 0.0
1194 |   }
1195 |   scale_param {
1196 |     filler {
1197 |       value: 1
1198 |     }
1199 |     bias_term: true
1200 |     bias_filler {
1201 |       value: 0
1202 |     }
1203 |   }
1204 | }
1205 | layer {
1206 |   name: "fire11/expand1x1"
1207 |   type: "Convolution"
1208 |   bottom: "fire11/squeeze1x1"
1209 |   top: "fire11/expand1x1"
1210 |   convolution_param {
1211 |     num_output: 41
1212 |     kernel_size: 1
1213 |     weight_filler {
1214 |       type: "xavier"
1215 |     }
1216 |   }
1217 | }
1218 | layer {
1219 |   name: "fire11/expand1x1/bn"
1220 |   type: "BatchNorm"
1221 |   bottom: "fire11/expand1x1"
1222 |   top: "fire11/expand1x1"
1223 |   param {
1224 |     lr_mult: 0
1225 |     decay_mult: 0
1226 |   }
1227 |   param {
1228 |     lr_mult: 0
1229 |     decay_mult: 0
1230 |   }
1231 |   param {
1232 |     lr_mult: 0
1233 |     decay_mult: 0
1234 |   }
1235 | }
1236 | layer {
1237 |   name: "fire11/expand1x1/scale"
1238 |   type: "Scale"
1239 |   bottom: "fire11/expand1x1"
1240 |   top: "fire11/expand1x1"
1241 |   param {
1242 |     lr_mult: 1.0
1243 |     decay_mult: 0.0
1244 |   }
1245 |   param {
1246 |     lr_mult: 2.0
1247 |     decay_mult: 0.0
1248 |   }
1249 |   scale_param {
1250 |     filler {
1251 |       value: 1
1252 |     }
1253 |     bias_term: true
1254 |     bias_filler {
1255 |       value: 0
1256 |     }
1257 |   }
1258 | }
1259 | layer {
1260 |   name: "fire11/relu_expand1x1"
1261 |   type: "ReLU"
1262 |   bottom: "fire11/expand1x1"
1263 |   top: "fire11/expand1x1"
1264 | }
1265 | layer {
1266 |   name: "fire11/expand3x3"
1267 |   type: "Convolution"
1268 |   bottom: "fire11/squeeze1x1"
1269 |   top: "fire11/expand3x3"
1270 |   convolution_param {
1271 |     num_output: 44
1272 |     pad: 1
1273 |     kernel_size: 3
1274 |     weight_filler {
1275 |       type: "xavier"
1276 |     }
1277 |   }
1278 | }
1279 | layer {
1280 |   name: "fire11/expand3x3/bn"
1281 |   type: "BatchNorm"
1282 |   bottom: "fire11/expand3x3"
1283 |   top: "fire11/expand3x3"
1284 |   param {
1285 |     lr_mult: 0
1286 |     decay_mult: 0
1287 |   }
1288 |   param {
1289 |     lr_mult: 0
1290 |     decay_mult: 0
1291 |   }
1292 |   param {
1293 |     lr_mult: 0
1294 |     decay_mult: 0
1295 |   }
1296 | }
1297 | layer {
1298 |   name: "fire11/expand3x3/scale"
1299 |   type: "Scale"
1300 |   bottom: "fire11/expand3x3"
1301 |   top: "fire11/expand3x3"
1302 |   param {
1303 |     lr_mult: 1.0
1304 |     decay_mult: 0.0
1305 |   }
1306 |   param {
1307 |     lr_mult: 2.0
1308 |     decay_mult: 0.0
1309 |   }
1310 |   scale_param {
1311 |     filler {
1312 |       value: 1
1313 |     }
1314 |     bias_term: true
1315 |     bias_filler {
1316 |       value: 0
1317 |     }
1318 |   }
1319 | }
1320 | layer {
1321 |   name: "fire11/relu_expand3x3"
1322 |   type: "ReLU"
1323 |   bottom: "fire11/expand3x3"
1324 |   top: "fire11/expand3x3"
1325 | }
1326 | layer {
1327 |   name: "fire11/concat"
1328 |   type: "Concat"
1329 |   bottom: "fire11/expand1x1"
1330 |   bottom: "fire11/expand3x3"
1331 |   top: "fire11/concat"
1332 | }
1333 | #================ssd append layer conv12, out 2 * 2
1334 | layer {
1335 |   name: "conv12_1"
1336 |   type: "Convolution"
1337 |   bottom: "fire11/concat"
1338 |   top: "conv12_1"
1339 |   param {
1340 |     lr_mult: 1.0
1341 |     decay_mult: 1.0
1342 |   }
1343 |   convolution_param {
1344 |     num_output: 51
1345 |     bias_term: false
1346 |     kernel_size: 1
1347 |     weight_filler {
1348 |       type: "msra"
1349 |     }
1350 |   }
1351 | }
1352 | layer {
1353 |   name: "conv12_1/bn"
1354 |   type: "BatchNorm"
1355 |   bottom: "conv12_1"
1356 |   top: "conv12_1"
1357 |   param {
1358 |     lr_mult: 0
1359 |     decay_mult: 0
1360 |   }
1361 |   param {
1362 |     lr_mult: 0
1363 |     decay_mult: 0
1364 |   }
1365 |   param {
1366 |     lr_mult: 0
1367 |     decay_mult: 0
1368 |   }
1369 | }
1370 | layer {
1371 |   name: "conv12_1/scale"
1372 |   type: "Scale"
1373 |   bottom: "conv12_1"
1374 |   top: "conv12_1"
1375 |   param {
1376 |     lr_mult: 1.0
1377 |     decay_mult: 0.0
1378 |   }
1379 |   param {
1380 |     lr_mult: 2.0
1381 |     decay_mult: 0.0
1382 |   }
1383 |   scale_param {
1384 |     filler {
1385 |       value: 1
1386 |     }
1387 |     bias_term: true
1388 |     bias_filler {
1389 |       value: 0
1390 |     }
1391 |   }
1392 | }
1393 | layer {
1394 |   name: "conv12_1/relu"
1395 |   type: "ReLU"
1396 |   bottom: "conv12_1"
1397 |   top: "conv12_1"
1398 | }
1399 | layer {
1400 |   name: "conv12_2"
1401 |   type: "Convolution"
1402 |   bottom: "conv12_1"
1403 |   top: "conv12_2"
1404 |   param {
1405 |     lr_mult: 1.0
1406 |     decay_mult: 1.0
1407 |   }
1408 |   convolution_param {
1409 |     num_output: 46
1410 |     bias_term: false
1411 |     pad: 1
1412 |     kernel_size: 3
1413 |     stride: 2
1414 |     weight_filler {
1415 |       type: "msra"
1416 |     }
1417 |   }
1418 | }
1419 | layer {
1420 |   name: "conv12_2/bn"
1421 |   type: "BatchNorm"
1422 |   bottom: "conv12_2"
1423 |   top: "conv12_2"
1424 |   param {
1425 |     lr_mult: 0
1426 |     decay_mult: 0
1427 |   }
1428 |   param {
1429 |     lr_mult: 0
1430 |     decay_mult: 0
1431 |   }
1432 |   param {
1433 |     lr_mult: 0
1434 |     decay_mult: 0
1435 |   }
1436 | }
1437 | layer {
1438 |   name: "conv12_2/scale"
1439 |   type: "Scale"
1440 |   bottom: "conv12_2"
1441 |   top: "conv12_2"
1442 |   param {
1443 |     lr_mult: 1.0
1444 |     decay_mult: 0.0
1445 |   }
1446 |   param {
1447 |     lr_mult: 2.0
1448 |     decay_mult: 0.0
1449 |   }
1450 |   scale_param {
1451 |     filler {
1452 |       value: 1
1453 |     }
1454 |     bias_term: true
1455 |     bias_filler {
1456 |       value: 0
1457 |     }
1458 |   }
1459 | }
1460 | layer {
1461 |   name: "conv12_2/relu"
1462 |   type: "ReLU"
1463 |   bottom: "conv12_2"
1464 |   top: "conv12_2"
1465 | }
1466 | #================ssd append layer conv13, out 1 * 1
1467 | layer {
1468 |   name: "conv13_1"
1469 |   type: "Convolution"
1470 |   bottom: "conv12_2"
1471 |   top: "conv13_1"
1472 |   param {
1473 |     lr_mult: 1.0
1474 |     decay_mult: 1.0
1475 |   }
1476 |   convolution_param {
1477 |     num_output: 55
1478 |     bias_term: false
1479 |     kernel_size: 1
1480 |     weight_filler {
1481 |       type: "msra"
1482 |     }
1483 |   }
1484 | }
1485 | layer {
1486 |   name: "conv13_1/bn"
1487 |   type: "BatchNorm"
1488 |   bottom: "conv13_1"
1489 |   top: "conv13_1"
1490 |   param {
1491 |     lr_mult: 0
1492 |     decay_mult: 0
1493 |   }
1494 |   param {
1495 |     lr_mult: 0
1496 |     decay_mult: 0
1497 |   }
1498 |   param {
1499 |     lr_mult: 0
1500 |     decay_mult: 0
1501 |   }
1502 | }
1503 | layer {
1504 |   name: "conv13_1/scale"
1505 |   type: "Scale"
1506 |   bottom: "conv13_1"
1507 |   top: "conv13_1"
1508 |   param {
1509 |     lr_mult: 1.0
1510 |     decay_mult: 0.0
1511 |   }
1512 |   param {
1513 |     lr_mult: 2.0
1514 |     decay_mult: 0.0
1515 |   }
1516 |   scale_param {
1517 |     filler {
1518 |       value: 1
1519 |     }
1520 |     bias_term: true
1521 |     bias_filler {
1522 |       value: 0
1523 |     }
1524 |   }
1525 | }
1526 | layer {
1527 |   name: "conv13_1/relu"
1528 |   type: "ReLU"
1529 |   bottom: "conv13_1"
1530 |   top: "conv13_1"
1531 | }
1532 | layer {
1533 |   name: "conv13_2"
1534 |   type: "Convolution"
1535 |   bottom: "conv13_1"
1536 |   top: "conv13_2"
1537 |   param {
1538 |     lr_mult: 1.0
1539 |     decay_mult: 1.0
1540 |   }
1541 |   convolution_param {
1542 |     num_output: 85
1543 |     bias_term: false
1544 |     pad: 1
1545 |     kernel_size: 3
1546 |     stride: 2
1547 |     weight_filler {
1548 |       type: "msra"
1549 |     }
1550 |   }
1551 | }
1552 | layer {
1553 |   name: "conv13_2/bn"
1554 |   type: "BatchNorm"
1555 |   bottom: "conv13_2"
1556 |   top: "conv13_2"
1557 |   param {
1558 |     lr_mult: 0
1559 |     decay_mult: 0
1560 |   }
1561 |   param {
1562 |     lr_mult: 0
1563 |     decay_mult: 0
1564 |   }
1565 |   param {
1566 |     lr_mult: 0
1567 |     decay_mult: 0
1568 |   }
1569 | }
1570 | layer {
1571 |   name: "conv13_2/scale"
1572 |   type: "Scale"
1573 |   bottom: "conv13_2"
1574 |   top: "conv13_2"
1575 |   param {
1576 |     lr_mult: 1.0
1577 |     decay_mult: 0.0
1578 |   }
1579 |   param {
1580 |     lr_mult: 2.0
1581 |     decay_mult: 0.0
1582 |   }
1583 |   scale_param {
1584 |     filler {
1585 |       value: 1
1586 |     }
1587 |     bias_term: true
1588 |     bias_filler {
1589 |       value: 0
1590 |     }
1591 |   }
1592 | }
1593 | layer {
1594 |   name: "conv13_2/relu"
1595 |   type: "ReLU"
1596 |   bottom: "conv13_2"
1597 |   top: "conv13_2"
1598 | }
1599 | #================ssd mbox layer
1600 | layer {
1601 |   name: "fire5/bn"
1602 |   type: "BatchNorm"
1603 |   bottom: "fire5/concat"
1604 |   top: "fire5/normal"
1605 |   param {
1606 |     lr_mult: 0
1607 |     decay_mult: 0
1608 |   }
1609 |   param {
1610 |     lr_mult: 0
1611 |     decay_mult: 0
1612 |   }
1613 |   param {
1614 |     lr_mult: 0
1615 |     decay_mult: 0
1616 |   }
1617 | }
1618 | layer {
1619 |   name: "fire5/scale"
1620 |   type: "Scale"
1621 |   bottom: "fire5/normal"
1622 |   top: "fire5/normal"
1623 |   param {
1624 |     lr_mult: 1.0
1625 |     decay_mult: 0.0
1626 |   }
1627 |   param {
1628 |     lr_mult: 2.0
1629 |     decay_mult: 0.0
1630 |   }
1631 |   scale_param {
1632 |     filler {
1633 |       value: 1
1634 |     }
1635 |     bias_term: true
1636 |     bias_filler {
1637 |       value: 0
1638 |     }
1639 |   }
1640 | }
1641 | layer {
1642 |   name: "fire5_mbox_loc"
1643 |   type: "Convolution"
1644 |   bottom: "fire5/normal"
1645 |   top: "fire5_mbox_loc"
1646 |   param {
1647 |     lr_mult: 1.0
1648 |     decay_mult: 1.0
1649 |   }
1650 |   param {
1651 |     lr_mult: 2.0
1652 |     decay_mult: 0.0
1653 |   }
1654 |   convolution_param {
1655 |     num_output: 16
1656 |     pad: 1
1657 |     kernel_size: 3
1658 |     stride: 1
1659 |     weight_filler {
1660 |       type: "msra"
1661 |     }
1662 |     bias_filler {
1663 |       type: "constant"
1664 |       value: 0.0
1665 |     }
1666 |   }
1667 | }
1668 | layer {
1669 |   name: "fire5_mbox_loc_perm"
1670 |   type: "Permute"
1671 |   bottom: "fire5_mbox_loc"
1672 |   top: "fire5_mbox_loc_perm"
1673 |   permute_param {
1674 |     order: 0
1675 |     order: 2
1676 |     order: 3
1677 |     order: 1
1678 |   }
1679 | }
1680 | layer {
1681 |   name: "fire5_mbox_loc_flat"
1682 |   type: "Flatten"
1683 |   bottom: "fire5_mbox_loc_perm"
1684 |   top: "fire5_mbox_loc_flat"
1685 |   flatten_param {
1686 |     axis: 1
1687 |   }
1688 | }
1689 | layer {
1690 |   name: "fire5_mbox_conf"
1691 |   type: "Convolution"
1692 |   bottom: "fire5/normal"
1693 |   top: "fire5_mbox_conf"
1694 |   param {
1695 |     lr_mult: 1.0
1696 |     decay_mult: 1.0
1697 |   }
1698 |   param {
1699 |     lr_mult: 2.0
1700 |     decay_mult: 0.0
1701 |   }
1702 |   convolution_param {
1703 |     num_output: 84
1704 |     pad: 1
1705 |     kernel_size: 3
1706 |     stride: 1
1707 |     weight_filler {
1708 |       type: "msra"
1709 |     }
1710 |     bias_filler {
1711 |       type: "constant"
1712 |       value: 0.0
1713 |     }
1714 |   }
1715 | }
1716 | layer {
1717 |   name: "fire5_mbox_conf_perm"
1718 |   type: "Permute"
1719 |   bottom: "fire5_mbox_conf"
1720 |   top: "fire5_mbox_conf_perm"
1721 |   permute_param {
1722 |     order: 0
1723 |     order: 2
1724 |     order: 3
1725 |     order: 1
1726 |   }
1727 | }
1728 | layer {
1729 |   name: "fire5_mbox_conf_flat"
1730 |   type: "Flatten"
1731 |   bottom: "fire5_mbox_conf_perm"
1732 |   top: "fire5_mbox_conf_flat"
1733 |   flatten_param {
1734 |     axis: 1
1735 |   }
1736 | }
1737 | layer {
1738 |   name: "fire5_mbox_priorbox"
1739 |   type: "PriorBox"
1740 |   bottom: "fire5/normal"
1741 |   bottom: "data"
1742 |   top: "fire5_mbox_priorbox"
1743 |   prior_box_param {
1744 |     min_size: 21.0
1745 |     max_size: 45.0
1746 |     aspect_ratio: 2.0
1747 |     flip: true
1748 |     clip: false
1749 |     variance: 0.1
1750 |     variance: 0.1
1751 |     variance: 0.2
1752 |     variance: 0.2
1753 |     step: 8
1754 |   }
1755 | }
1756 | #================mbox fire9
1757 | layer {
1758 |   name: "fire9_mbox_loc"
1759 |   type: "Convolution"
1760 |   bottom: "fire9/concat"
1761 |   top: "fire9_mbox_loc"
1762 |   param {
1763 |     lr_mult: 1.0
1764 |     decay_mult: 1.0
1765 |   }
1766 |   param {
1767 |     lr_mult: 2.0
1768 |     decay_mult: 0.0
1769 |   }
1770 |   convolution_param {
1771 |     num_output: 24
1772 |     pad: 1
1773 |     kernel_size: 3
1774 |     stride: 1
1775 |     weight_filler {
1776 |       type: "msra"
1777 |     }
1778 |     bias_filler {
1779 |       type: "constant"
1780 |       value: 0.0
1781 |     }
1782 |   }
1783 | }
1784 | layer {
1785 |   name: "fire9_mbox_loc_perm"
1786 |   type: "Permute"
1787 |   bottom: "fire9_mbox_loc"
1788 |   top: "fire9_mbox_loc_perm"
1789 |   permute_param {
1790 |     order: 0
1791 |     order: 2
1792 |     order: 3
1793 |     order: 1
1794 |   }
1795 | }
1796 | layer {
1797 |   name: "fire9_mbox_loc_flat"
1798 |   type: "Flatten"
1799 |   bottom: "fire9_mbox_loc_perm"
1800 |   top: "fire9_mbox_loc_flat"
1801 |   flatten_param {
1802 |     axis: 1
1803 |   }
1804 | }
1805 | layer {
1806 |   name: "fire9_mbox_conf"
1807 |   type: "Convolution"
1808 |   bottom: "fire9/concat"
1809 |   top: "fire9_mbox_conf"
1810 |   param {
1811 |     lr_mult: 1.0
1812 |     decay_mult: 1.0
1813 |   }
1814 |   param {
1815 |     lr_mult: 2.0
1816 |     decay_mult: 0.0
1817 |   }
1818 |   convolution_param {
1819 |     num_output: 126
1820 |     pad: 1
1821 |     kernel_size: 3
1822 |     stride: 1
1823 |     weight_filler {
1824 |       type: "msra"
1825 |     }
1826 |     bias_filler {
1827 |       type: "constant"
1828 |       value: 0.0
1829 |     }
1830 |   }
1831 | }
1832 | layer {
1833 |   name: "fire9_mbox_conf_perm"
1834 |   type: "Permute"
1835 |   bottom: "fire9_mbox_conf"
1836 |   top: "fire9_mbox_conf_perm"
1837 |   permute_param {
1838 |     order: 0
1839 |     order: 2
1840 |     order: 3
1841 |     order: 1
1842 |   }
1843 | }
1844 | layer {
1845 |   name: "fire9_mbox_conf_flat"
1846 |   type: "Flatten"
1847 |   bottom: "fire9_mbox_conf_perm"
1848 |   top: "fire9_mbox_conf_flat"
1849 |   flatten_param {
1850 |     axis: 1
1851 |   }
1852 | }
1853 | layer {
1854 |   name: "fire9_mbox_priorbox"
1855 |   type: "PriorBox"
1856 |   bottom: "fire9/concat"
1857 |   bottom: "data"
1858 |   top: "fire9_mbox_priorbox"
1859 |   prior_box_param {
1860 |     min_size: 45.0
1861 |     max_size: 99.0
1862 |     aspect_ratio: 2.0
1863 |     aspect_ratio: 3.0
1864 |     flip: true
1865 |     clip: false
1866 |     variance: 0.1
1867 |     variance: 0.1
1868 |     variance: 0.2
1869 |     variance: 0.2
1870 |     step: 16
1871 |   }
1872 | }
1873 | #=====fire10 mbox
1874 | layer {
1875 |   name: "fire10_mbox_loc"
1876 |   type: "Convolution"
1877 |   bottom: "fire10/concat"
1878 |   top: "fire10_mbox_loc"
1879 |   param {
1880 |     lr_mult: 1.0
1881 |     decay_mult: 1.0
1882 |   }
1883 |   param {
1884 |     lr_mult: 2.0
1885 |     decay_mult: 0.0
1886 |   }
1887 |   convolution_param {
1888 |     num_output: 24
1889 |     pad: 1
1890 |     kernel_size: 3
1891 |     stride: 1
1892 |     weight_filler {
1893 |       type: "msra"
1894 |     }
1895 |     bias_filler {
1896 |       type: "constant"
1897 |       value: 0.0
1898 |     }
1899 |   }
1900 | }
1901 | layer {
1902 |   name: "fire10_mbox_loc_perm"
1903 |   type: "Permute"
1904 |   bottom: "fire10_mbox_loc"
1905 |   top: "fire10_mbox_loc_perm"
1906 |   permute_param {
1907 |     order: 0
1908 |     order: 2
1909 |     order: 3
1910 |     order: 1
1911 |   }
1912 | }
1913 | layer {
1914 |   name: "fire10_mbox_loc_flat"
1915 |   type: "Flatten"
1916 |   bottom: "fire10_mbox_loc_perm"
1917 |   top: "fire10_mbox_loc_flat"
1918 |   flatten_param {
1919 |     axis: 1
1920 |   }
1921 | }
1922 | layer {
1923 |   name: "fire10_mbox_conf"
1924 |   type: "Convolution"
1925 |   bottom: "fire10/concat"
1926 |   top: "fire10_mbox_conf"
1927 |   param {
1928 |     lr_mult: 1.0
1929 |     decay_mult: 1.0
1930 |   }
1931 |   param {
1932 |     lr_mult: 2.0
1933 |     decay_mult: 0.0
1934 |   }
1935 |   convolution_param {
1936 |     num_output: 126
1937 |     pad: 1
1938 |     kernel_size: 3
1939 |     stride: 1
1940 |     weight_filler {
1941 |       type: "msra"
1942 |     }
1943 |     bias_filler {
1944 |       type: "constant"
1945 |       value: 0.0
1946 |     }
1947 |   }
1948 | }
1949 | layer {
1950 |   name: "fire10_mbox_conf_perm"
1951 |   type: "Permute"
1952 |   bottom: "fire10_mbox_conf"
1953 |   top: "fire10_mbox_conf_perm"
1954 |   permute_param {
1955 |     order: 0
1956 |     order: 2
1957 |     order: 3
1958 |     order: 1
1959 |   }
1960 | }
1961 | layer {
1962 |   name: "fire10_mbox_conf_flat"
1963 |   type: "Flatten"
1964 |   bottom: "fire10_mbox_conf_perm"
1965 |   top: "fire10_mbox_conf_flat"
1966 |   flatten_param {
1967 |     axis: 1
1968 |   }
1969 | }
1970 | layer {
1971 |   name: "fire10_mbox_priorbox"
1972 |   type: "PriorBox"
1973 |   bottom: "fire10/concat"
1974 |   bottom: "data"
1975 |   top: "fire10_mbox_priorbox"
1976 |   prior_box_param {
1977 |     min_size: 99.0
1978 |     max_size: 153.0
1979 |     aspect_ratio: 2.0
1980 |     aspect_ratio: 3.0
1981 |     flip: true
1982 |     clip: false
1983 |     variance: 0.1
1984 |     variance: 0.1
1985 |     variance: 0.2
1986 |     variance: 0.2
1987 |     step: 32
1988 |   }
1989 | }
1990 | 
1991 | #=====conv11 mbox
1992 | layer {
1993 |   name: "fire11_mbox_loc"
1994 |   type: "Convolution"
1995 |   bottom: "fire11/concat"
1996 |   top: "fire11_mbox_loc"
1997 |   param {
1998 |     lr_mult: 1.0
1999 |     decay_mult: 1.0
2000 |   }
2001 |   param {
2002 |     lr_mult: 2.0
2003 |     decay_mult: 0.0
2004 |   }
2005 |   convolution_param {
2006 |     num_output: 24
2007 |     pad: 1
2008 |     kernel_size: 3
2009 |     stride: 1
2010 |     weight_filler {
2011 |       type: "msra"
2012 |     }
2013 |     bias_filler {
2014 |       type: "constant"
2015 |       value: 0.0
2016 |     }
2017 |   }
2018 | }
2019 | layer {
2020 |   name: "fire11_mbox_loc_perm"
2021 |   type: "Permute"
2022 |   bottom: "fire11_mbox_loc"
2023 |   top: "fire11_mbox_loc_perm"
2024 |   permute_param {
2025 |     order: 0
2026 |     order: 2
2027 |     order: 3
2028 |     order: 1
2029 |   }
2030 | }
2031 | layer {
2032 |   name: "fire11_mbox_loc_flat"
2033 |   type: "Flatten"
2034 |   bottom: "fire11_mbox_loc_perm"
2035 |   top: "fire11_mbox_loc_flat"
2036 |   flatten_param {
2037 |     axis: 1
2038 |   }
2039 | }
2040 | layer {
2041 |   name: "fire11_mbox_conf"
2042 |   type: "Convolution"
2043 |   bottom: "fire11/concat"
2044 |   top: "fire11_mbox_conf"
2045 |   param {
2046 |     lr_mult: 1.0
2047 |     decay_mult: 1.0
2048 |   }
2049 |   param {
2050 |     lr_mult: 2.0
2051 |     decay_mult: 0.0
2052 |   }
2053 |   convolution_param {
2054 |     num_output: 126
2055 |     pad: 1
2056 |     kernel_size: 3
2057 |     stride: 1
2058 |     weight_filler {
2059 |       type: "msra"
2060 |     }
2061 |     bias_filler {
2062 |       type: "constant"
2063 |       value: 0.0
2064 |     }
2065 |   }
2066 | }
2067 | layer {
2068 |   name: "fire11_mbox_conf_perm"
2069 |   type: "Permute"
2070 |   bottom: "fire11_mbox_conf"
2071 |   top: "fire11_mbox_conf_perm"
2072 |   permute_param {
2073 |     order: 0
2074 |     order: 2
2075 |     order: 3
2076 |     order: 1
2077 |   }
2078 | }
2079 | layer {
2080 |   name: "fire11_mbox_conf_flat"
2081 |   type: "Flatten"
2082 |   bottom: "fire11_mbox_conf_perm"
2083 |   top: "fire11_mbox_conf_flat"
2084 |   flatten_param {
2085 |     axis: 1
2086 |   }
2087 | }
2088 | layer {
2089 |   name: "fire11_mbox_priorbox"
2090 |   type: "PriorBox"
2091 |   bottom: "fire11/concat"
2092 |   bottom: "data"
2093 |   top: "fire11_mbox_priorbox"
2094 |   prior_box_param {
2095 |     min_size: 153.0
2096 |     max_size: 207.0
2097 |     aspect_ratio: 2.0
2098 |     aspect_ratio: 3.0
2099 |     flip: true
2100 |     clip: false
2101 |     variance: 0.1
2102 |     variance: 0.1
2103 |     variance: 0.2
2104 |     variance: 0.2
2105 |     step: 64
2106 |   }
2107 | }
2108 | #=====conv12 mbox
2109 | layer {
2110 |   name: "conv12_2_mbox_loc"
2111 |   type: "Convolution"
2112 |   bottom: "conv12_2"
2113 |   top: "conv12_2_mbox_loc"
2114 |   param {
2115 |     lr_mult: 1.0
2116 |     decay_mult: 1.0
2117 |   }
2118 |   param {
2119 |     lr_mult: 2.0
2120 |     decay_mult: 0.0
2121 |   }
2122 |   convolution_param {
2123 |     num_output: 24
2124 |     pad: 1
2125 |     kernel_size: 3
2126 |     stride: 1
2127 |     weight_filler {
2128 |       type: "msra"
2129 |     }
2130 |     bias_filler {
2131 |       type: "constant"
2132 |       value: 0.0
2133 |     }
2134 |   }
2135 | }
2136 | layer {
2137 |   name: "conv12_2_mbox_loc_perm"
2138 |   type: "Permute"
2139 |   bottom: "conv12_2_mbox_loc"
2140 |   top: "conv12_2_mbox_loc_perm"
2141 |   permute_param {
2142 |     order: 0
2143 |     order: 2
2144 |     order: 3
2145 |     order: 1
2146 |   }
2147 | }
2148 | layer {
2149 |   name: "conv12_2_mbox_loc_flat"
2150 |   type: "Flatten"
2151 |   bottom: "conv12_2_mbox_loc_perm"
2152 |   top: "conv12_2_mbox_loc_flat"
2153 |   flatten_param {
2154 |     axis: 1
2155 |   }
2156 | }
2157 | layer {
2158 |   name: "conv12_2_mbox_conf"
2159 |   type: "Convolution"
2160 |   bottom: "conv12_2"
2161 |   top: "conv12_2_mbox_conf"
2162 |   param {
2163 |     lr_mult: 1.0
2164 |     decay_mult: 1.0
2165 |   }
2166 |   param {
2167 |     lr_mult: 2.0
2168 |     decay_mult: 0.0
2169 |   }
2170 |   convolution_param {
2171 |     num_output: 126
2172 |     pad: 1
2173 |     kernel_size: 3
2174 |     stride: 1
2175 |     weight_filler {
2176 |       type: "msra"
2177 |     }
2178 |     bias_filler {
2179 |       type: "constant"
2180 |       value: 0.0
2181 |     }
2182 |   }
2183 | }
2184 | layer {
2185 |   name: "conv12_2_mbox_conf_perm"
2186 |   type: "Permute"
2187 |   bottom: "conv12_2_mbox_conf"
2188 |   top: "conv12_2_mbox_conf_perm"
2189 |   permute_param {
2190 |     order: 0
2191 |     order: 2
2192 |     order: 3
2193 |     order: 1
2194 |   }
2195 | }
2196 | layer {
2197 |   name: "conv12_2_mbox_conf_flat"
2198 |   type: "Flatten"
2199 |   bottom: "conv12_2_mbox_conf_perm"
2200 |   top: "conv12_2_mbox_conf_flat"
2201 |   flatten_param {
2202 |     axis: 1
2203 |   }
2204 | }
2205 | layer {
2206 |   name: "conv12_2_mbox_priorbox"
2207 |   type: "PriorBox"
2208 |   bottom: "conv12_2"
2209 |   bottom: "data"
2210 |   top: "conv12_2_mbox_priorbox"
2211 |   prior_box_param {
2212 |     min_size: 207.0
2213 |     max_size: 261.0
2214 |     aspect_ratio: 2.0
2215 |     aspect_ratio: 3.0
2216 |     flip: true
2217 |     clip: false
2218 |     variance: 0.1
2219 |     variance: 0.1
2220 |     variance: 0.2
2221 |     variance: 0.2
2222 |     step: 100
2223 |   }
2224 | }
2225 | #=====conv13 mbox
2226 | layer {
2227 |   name: "conv13_2_mbox_loc"
2228 |   type: "Convolution"
2229 |   bottom: "conv13_2"
2230 |   top: "conv13_2_mbox_loc"
2231 |   param {
2232 |     lr_mult: 1.0
2233 |     decay_mult: 1.0
2234 |   }
2235 |   param {
2236 |     lr_mult: 2.0
2237 |     decay_mult: 0.0
2238 |   }
2239 |   convolution_param {
2240 |     num_output: 16
2241 |     pad: 1
2242 |     kernel_size: 3
2243 |     stride: 1
2244 |     weight_filler {
2245 |       type: "msra"
2246 |     }
2247 |     bias_filler {
2248 |       type: "constant"
2249 |       value: 0.0
2250 |     }
2251 |   }
2252 | }
2253 | layer {
2254 |   name: "conv13_2_mbox_loc_perm"
2255 |   type: "Permute"
2256 |   bottom: "conv13_2_mbox_loc"
2257 |   top: "conv13_2_mbox_loc_perm"
2258 |   permute_param {
2259 |     order: 0
2260 |     order: 2
2261 |     order: 3
2262 |     order: 1
2263 |   }
2264 | }
2265 | layer {
2266 |   name: "conv13_2_mbox_loc_flat"
2267 |   type: "Flatten"
2268 |   bottom: "conv13_2_mbox_loc_perm"
2269 |   top: "conv13_2_mbox_loc_flat"
2270 |   flatten_param {
2271 |     axis: 1
2272 |   }
2273 | }
2274 | layer {
2275 |   name: "conv13_2_mbox_conf"
2276 |   type: "Convolution"
2277 |   bottom: "conv13_2"
2278 |   top: "conv13_2_mbox_conf"
2279 |   param {
2280 |     lr_mult: 1.0
2281 |     decay_mult: 1.0
2282 |   }
2283 |   param {
2284 |     lr_mult: 2.0
2285 |     decay_mult: 0.0
2286 |   }
2287 |   convolution_param {
2288 |     num_output: 84
2289 |     pad: 1
2290 |     kernel_size: 3
2291 |     stride: 1
2292 |     weight_filler {
2293 |       type: "msra"
2294 |     }
2295 |     bias_filler {
2296 |       type: "constant"
2297 |       value: 0.0
2298 |     }
2299 |   }
2300 | }
2301 | layer {
2302 |   name: "conv13_2_mbox_conf_perm"
2303 |   type: "Permute"
2304 |   bottom: "conv13_2_mbox_conf"
2305 |   top: "conv13_2_mbox_conf_perm"
2306 |   permute_param {
2307 |     order: 0
2308 |     order: 2
2309 |     order: 3
2310 |     order: 1
2311 |   }
2312 | }
2313 | layer {
2314 |   name: "conv13_2_mbox_conf_flat"
2315 |   type: "Flatten"
2316 |   bottom: "conv13_2_mbox_conf_perm"
2317 |   top: "conv13_2_mbox_conf_flat"
2318 |   flatten_param {
2319 |     axis: 1
2320 |   }
2321 | }
2322 | layer {
2323 |   name: "conv13_2_mbox_priorbox"
2324 |   type: "PriorBox"
2325 |   bottom: "conv13_2"
2326 |   bottom: "data"
2327 |   top: "conv13_2_mbox_priorbox"
2328 |   prior_box_param {
2329 |     min_size: 261.0
2330 |     max_size: 315.0
2331 |     aspect_ratio: 2.0
2332 |     flip: true
2333 |     clip: false
2334 |     variance: 0.1
2335 |     variance: 0.1
2336 |     variance: 0.2
2337 |     variance: 0.2
2338 |     step: 300
2339 |   }
2340 | }
2341 | layer {
2342 |   name: "mbox_loc"
2343 |   type: "Concat"
2344 |   bottom: "fire5_mbox_loc_flat"
2345 |   bottom: "fire9_mbox_loc_flat"
2346 |   bottom: "fire10_mbox_loc_flat"
2347 |   bottom: "fire11_mbox_loc_flat"
2348 |   bottom: "conv12_2_mbox_loc_flat"
2349 |   bottom: "conv13_2_mbox_loc_flat"
2350 |   top: "mbox_loc"
2351 |   concat_param {
2352 |     axis: 1
2353 |   }
2354 | }
2355 | layer {
2356 |   name: "mbox_conf"
2357 |   type: "Concat"
2358 |   bottom: "fire5_mbox_conf_flat"
2359 |   bottom: "fire9_mbox_conf_flat"
2360 |   bottom: "fire10_mbox_conf_flat"
2361 |   bottom: "fire11_mbox_conf_flat"
2362 |   bottom: "conv12_2_mbox_conf_flat"
2363 |   bottom: "conv13_2_mbox_conf_flat"
2364 |   top: "mbox_conf"
2365 |   concat_param {
2366 |     axis: 1
2367 |   }
2368 | }
2369 | layer {
2370 |   name: "mbox_priorbox"
2371 |   type: "Concat"
2372 |   bottom: "fire5_mbox_priorbox"
2373 |   bottom: "fire9_mbox_priorbox"
2374 |   bottom: "fire10_mbox_priorbox"
2375 |   bottom: "fire11_mbox_priorbox"
2376 |   bottom: "conv12_2_mbox_priorbox"
2377 |   bottom: "conv13_2_mbox_priorbox"
2378 |   top: "mbox_priorbox"
2379 |   concat_param {
2380 |     axis: 2
2381 |   }
2382 | }
2383 | layer {
2384 |   name: "mbox_loss"
2385 |   type: "MultiBoxLoss"
2386 |   bottom: "mbox_loc"
2387 |   bottom: "mbox_conf"
2388 |   bottom: "mbox_priorbox"
2389 |   bottom: "label"
2390 |   top: "mbox_loss"
2391 |   include {
2392 |     phase: TRAIN
2393 |   }
2394 |   propagate_down: true
2395 |   propagate_down: true
2396 |   propagate_down: false
2397 |   propagate_down: false
2398 |   loss_param {
2399 |     normalization: VALID
2400 |   }
2401 |   multibox_loss_param {
2402 |     loc_loss_type: SMOOTH_L1
2403 |     conf_loss_type: SOFTMAX
2404 |     loc_weight: 1.0
2405 |     num_classes: 21
2406 |     share_location: true
2407 |     match_type: PER_PREDICTION
2408 |     overlap_threshold: 0.5
2409 |     use_prior_for_matching: true
2410 |     background_label_id: 0
2411 |     use_difficult_gt: true
2412 |     neg_pos_ratio: 3.0
2413 |     neg_overlap: 0.5
2414 |     code_type: CENTER_SIZE
2415 |     ignore_cross_boundary_bbox: false
2416 |     mining_type: MAX_NEGATIVE
2417 |   }
2418 | }
2419 | 


--------------------------------------------------------------------------------
/python_script/deploy.prototxt:
--------------------------------------------------------------------------------
   1 | layer {
   2 |   name: "data"
   3 |   type: "Input"
   4 |   top: "data"
   5 |   input_param {
   6 |     shape {
   7 |       dim: 1
   8 |       dim: 3
   9 |       dim: 300
  10 |       dim: 300
  11 |     }
  12 |   }
  13 | }
  14 | layer {
  15 |   name: "conv1"
  16 |   type: "Convolution"
  17 |   bottom: "data"
  18 |   top: "conv1"
  19 |   param {
  20 |     lr_mult: 1.0
  21 |     decay_mult: 0.0
  22 |   }
  23 |   param {
  24 |     lr_mult: 1.0
  25 |     decay_mult: 0.0
  26 |   }
  27 |   convolution_param {
  28 |     num_output: 57
  29 |     kernel_size: 3
  30 |     stride: 2
  31 |     weight_filler {
  32 |       type: "xavier"
  33 |     }
  34 |   }
  35 | }
  36 | layer {
  37 |   name: "relu_conv1"
  38 |   type: "ReLU"
  39 |   bottom: "conv1"
  40 |   top: "conv1"
  41 | }
  42 | layer {
  43 |   name: "pool1"
  44 |   type: "Pooling"
  45 |   bottom: "conv1"
  46 |   top: "pool1"
  47 |   pooling_param {
  48 |     pool: MAX
  49 |     kernel_size: 3
  50 |     stride: 2
  51 |   }
  52 | }
  53 | layer {
  54 |   name: "fire2/squeeze1x1"
  55 |   type: "Convolution"
  56 |   bottom: "pool1"
  57 |   top: "fire2/squeeze1x1"
  58 |   param {
  59 |     lr_mult: 1.0
  60 |     decay_mult: 0.0
  61 |   }
  62 |   param {
  63 |     lr_mult: 1.0
  64 |     decay_mult: 0.0
  65 |   }
  66 |   convolution_param {
  67 |     num_output: 15
  68 |     kernel_size: 1
  69 |     weight_filler {
  70 |       type: "xavier"
  71 |     }
  72 |   }
  73 | }
  74 | layer {
  75 |   name: "fire2/expand1x1"
  76 |   type: "Convolution"
  77 |   bottom: "fire2/squeeze1x1"
  78 |   top: "fire2/expand1x1"
  79 |   param {
  80 |     lr_mult: 1.0
  81 |     decay_mult: 0.0
  82 |   }
  83 |   param {
  84 |     lr_mult: 1.0
  85 |     decay_mult: 0.0
  86 |   }
  87 |   convolution_param {
  88 |     num_output: 49
  89 |     kernel_size: 1
  90 |     weight_filler {
  91 |       type: "xavier"
  92 |     }
  93 |   }
  94 | }
  95 | layer {
  96 |   name: "fire2/relu_expand1x1"
  97 |   type: "ReLU"
  98 |   bottom: "fire2/expand1x1"
  99 |   top: "fire2/expand1x1"
 100 | }
 101 | layer {
 102 |   name: "fire2/expand3x3"
 103 |   type: "Convolution"
 104 |   bottom: "fire2/squeeze1x1"
 105 |   top: "fire2/expand3x3"
 106 |   param {
 107 |     lr_mult: 1.0
 108 |     decay_mult: 0.0
 109 |   }
 110 |   param {
 111 |     lr_mult: 1.0
 112 |     decay_mult: 0.0
 113 |   }
 114 |   convolution_param {
 115 |     num_output: 53
 116 |     pad: 1
 117 |     kernel_size: 3
 118 |     weight_filler {
 119 |       type: "xavier"
 120 |     }
 121 |   }
 122 | }
 123 | layer {
 124 |   name: "fire2/relu_expand3x3"
 125 |   type: "ReLU"
 126 |   bottom: "fire2/expand3x3"
 127 |   top: "fire2/expand3x3"
 128 | }
 129 | layer {
 130 |   name: "fire2/concat"
 131 |   type: "Concat"
 132 |   bottom: "fire2/expand1x1"
 133 |   bottom: "fire2/expand3x3"
 134 |   top: "fire2/concat"
 135 | }
 136 | layer {
 137 |   name: "fire3/squeeze1x1"
 138 |   type: "Convolution"
 139 |   bottom: "fire2/concat"
 140 |   top: "fire3/squeeze1x1"
 141 |   param {
 142 |     lr_mult: 1.0
 143 |     decay_mult: 0.0
 144 |   }
 145 |   param {
 146 |     lr_mult: 1.0
 147 |     decay_mult: 0.0
 148 |   }
 149 |   convolution_param {
 150 |     num_output: 15
 151 |     kernel_size: 1
 152 |     weight_filler {
 153 |       type: "xavier"
 154 |     }
 155 |   }
 156 | }
 157 | layer {
 158 |   name: "fire3/expand1x1"
 159 |   type: "Convolution"
 160 |   bottom: "fire3/squeeze1x1"
 161 |   top: "fire3/expand1x1"
 162 |   param {
 163 |     lr_mult: 1.0
 164 |     decay_mult: 0.0
 165 |   }
 166 |   param {
 167 |     lr_mult: 1.0
 168 |     decay_mult: 0.0
 169 |   }
 170 |   convolution_param {
 171 |     num_output: 54
 172 |     kernel_size: 1
 173 |     weight_filler {
 174 |       type: "xavier"
 175 |     }
 176 |   }
 177 | }
 178 | layer {
 179 |   name: "fire3/relu_expand1x1"
 180 |   type: "ReLU"
 181 |   bottom: "fire3/expand1x1"
 182 |   top: "fire3/expand1x1"
 183 | }
 184 | layer {
 185 |   name: "fire3/expand3x3"
 186 |   type: "Convolution"
 187 |   bottom: "fire3/squeeze1x1"
 188 |   top: "fire3/expand3x3"
 189 |   param {
 190 |     lr_mult: 1.0
 191 |     decay_mult: 0.0
 192 |   }
 193 |   param {
 194 |     lr_mult: 1.0
 195 |     decay_mult: 0.0
 196 |   }
 197 |   convolution_param {
 198 |     num_output: 52
 199 |     pad: 1
 200 |     kernel_size: 3
 201 |     weight_filler {
 202 |       type: "xavier"
 203 |     }
 204 |   }
 205 | }
 206 | layer {
 207 |   name: "fire3/relu_expand3x3"
 208 |   type: "ReLU"
 209 |   bottom: "fire3/expand3x3"
 210 |   top: "fire3/expand3x3"
 211 | }
 212 | layer {
 213 |   name: "fire3/concat"
 214 |   type: "Concat"
 215 |   bottom: "fire3/expand1x1"
 216 |   bottom: "fire3/expand3x3"
 217 |   top: "fire3/concat"
 218 | }
 219 | layer {
 220 |   name: "pool3"
 221 |   type: "Pooling"
 222 |   bottom: "fire3/concat"
 223 |   top: "pool3"
 224 |   pooling_param {
 225 |     pool: MAX
 226 |     kernel_size: 3
 227 |     stride: 2
 228 |   }
 229 | }
 230 | layer {
 231 |   name: "fire4/squeeze1x1"
 232 |   type: "Convolution"
 233 |   bottom: "pool3"
 234 |   top: "fire4/squeeze1x1"
 235 |   param {
 236 |     lr_mult: 1.0
 237 |     decay_mult: 0.0
 238 |   }
 239 |   param {
 240 |     lr_mult: 1.0
 241 |     decay_mult: 0.0
 242 |   }
 243 |   convolution_param {
 244 |     num_output: 29
 245 |     kernel_size: 1
 246 |     weight_filler {
 247 |       type: "xavier"
 248 |     }
 249 |   }
 250 | }
 251 | layer {
 252 |   name: "fire4/expand1x1"
 253 |   type: "Convolution"
 254 |   bottom: "fire4/squeeze1x1"
 255 |   top: "fire4/expand1x1"
 256 |   param {
 257 |     lr_mult: 1.0
 258 |     decay_mult: 0.0
 259 |   }
 260 |   param {
 261 |     lr_mult: 1.0
 262 |     decay_mult: 0.0
 263 |   }
 264 |   convolution_param {
 265 |     num_output: 92
 266 |     kernel_size: 1
 267 |     weight_filler {
 268 |       type: "xavier"
 269 |     }
 270 |   }
 271 | }
 272 | layer {
 273 |   name: "fire4/relu_expand1x1"
 274 |   type: "ReLU"
 275 |   bottom: "fire4/expand1x1"
 276 |   top: "fire4/expand1x1"
 277 | }
 278 | layer {
 279 |   name: "fire4/expand3x3"
 280 |   type: "Convolution"
 281 |   bottom: "fire4/squeeze1x1"
 282 |   top: "fire4/expand3x3"
 283 |   param {
 284 |     lr_mult: 1.0
 285 |     decay_mult: 0.0
 286 |   }
 287 |   param {
 288 |     lr_mult: 1.0
 289 |     decay_mult: 0.0
 290 |   }
 291 |   convolution_param {
 292 |     num_output: 94
 293 |     pad: 1
 294 |     kernel_size: 3
 295 |     weight_filler {
 296 |       type: "xavier"
 297 |     }
 298 |   }
 299 | }
 300 | layer {
 301 |   name: "fire4/relu_expand3x3"
 302 |   type: "ReLU"
 303 |   bottom: "fire4/expand3x3"
 304 |   top: "fire4/expand3x3"
 305 | }
 306 | layer {
 307 |   name: "fire4/concat"
 308 |   type: "Concat"
 309 |   bottom: "fire4/expand1x1"
 310 |   bottom: "fire4/expand3x3"
 311 |   top: "fire4/concat"
 312 | }
 313 | layer {
 314 |   name: "fire5/squeeze1x1"
 315 |   type: "Convolution"
 316 |   bottom: "fire4/concat"
 317 |   top: "fire5/squeeze1x1"
 318 |   param {
 319 |     lr_mult: 1.0
 320 |     decay_mult: 0.0
 321 |   }
 322 |   param {
 323 |     lr_mult: 1.0
 324 |     decay_mult: 0.0
 325 |   }
 326 |   convolution_param {
 327 |     num_output: 29
 328 |     kernel_size: 1
 329 |     weight_filler {
 330 |       type: "xavier"
 331 |     }
 332 |   }
 333 | }
 334 | layer {
 335 |   name: "fire5/expand1x1"
 336 |   type: "Convolution"
 337 |   bottom: "fire5/squeeze1x1"
 338 |   top: "fire5/expand1x1"
 339 |   param {
 340 |     lr_mult: 1.0
 341 |     decay_mult: 0.0
 342 |   }
 343 |   param {
 344 |     lr_mult: 1.0
 345 |     decay_mult: 0.0
 346 |   }
 347 |   convolution_param {
 348 |     num_output: 90
 349 |     kernel_size: 1
 350 |     weight_filler {
 351 |       type: "xavier"
 352 |     }
 353 |   }
 354 | }
 355 | layer {
 356 |   name: "fire5/relu_expand1x1"
 357 |   type: "ReLU"
 358 |   bottom: "fire5/expand1x1"
 359 |   top: "fire5/expand1x1"
 360 | }
 361 | layer {
 362 |   name: "fire5/expand3x3"
 363 |   type: "Convolution"
 364 |   bottom: "fire5/squeeze1x1"
 365 |   top: "fire5/expand3x3"
 366 |   param {
 367 |     lr_mult: 1.0
 368 |     decay_mult: 0.0
 369 |   }
 370 |   param {
 371 |     lr_mult: 1.0
 372 |     decay_mult: 0.0
 373 |   }
 374 |   convolution_param {
 375 |     num_output: 83
 376 |     pad: 1
 377 |     kernel_size: 3
 378 |     weight_filler {
 379 |       type: "xavier"
 380 |     }
 381 |   }
 382 | }
 383 | layer {
 384 |   name: "fire5/relu_expand3x3"
 385 |   type: "ReLU"
 386 |   bottom: "fire5/expand3x3"
 387 |   top: "fire5/expand3x3"
 388 | }
 389 | layer {
 390 |   name: "fire5/concat"
 391 |   type: "Concat"
 392 |   bottom: "fire5/expand1x1"
 393 |   bottom: "fire5/expand3x3"
 394 |   top: "fire5/concat"
 395 | }
 396 | layer {
 397 |   name: "pool5"
 398 |   type: "Pooling"
 399 |   bottom: "fire5/concat"
 400 |   top: "pool5"
 401 |   pooling_param {
 402 |     pool: MAX
 403 |     kernel_size: 3
 404 |     stride: 2
 405 |   }
 406 | }
 407 | layer {
 408 |   name: "fire6/squeeze1x1"
 409 |   type: "Convolution"
 410 |   bottom: "pool5"
 411 |   top: "fire6/squeeze1x1"
 412 |   param {
 413 |     lr_mult: 1.0
 414 |     decay_mult: 0.0
 415 |   }
 416 |   param {
 417 |     lr_mult: 1.0
 418 |     decay_mult: 0.0
 419 |   }
 420 |   convolution_param {
 421 |     num_output: 44
 422 |     kernel_size: 1
 423 |     weight_filler {
 424 |       type: "xavier"
 425 |     }
 426 |   }
 427 | }
 428 | layer {
 429 |   name: "fire6/expand1x1"
 430 |   type: "Convolution"
 431 |   bottom: "fire6/squeeze1x1"
 432 |   top: "fire6/expand1x1"
 433 |   param {
 434 |     lr_mult: 1.0
 435 |     decay_mult: 0.0
 436 |   }
 437 |   param {
 438 |     lr_mult: 1.0
 439 |     decay_mult: 0.0
 440 |   }
 441 |   convolution_param {
 442 |     num_output: 166
 443 |     kernel_size: 1
 444 |     weight_filler {
 445 |       type: "xavier"
 446 |     }
 447 |   }
 448 | }
 449 | layer {
 450 |   name: "fire6/relu_expand1x1"
 451 |   type: "ReLU"
 452 |   bottom: "fire6/expand1x1"
 453 |   top: "fire6/expand1x1"
 454 | }
 455 | layer {
 456 |   name: "fire6/expand3x3"
 457 |   type: "Convolution"
 458 |   bottom: "fire6/squeeze1x1"
 459 |   top: "fire6/expand3x3"
 460 |   param {
 461 |     lr_mult: 1.0
 462 |     decay_mult: 0.0
 463 |   }
 464 |   param {
 465 |     lr_mult: 1.0
 466 |     decay_mult: 0.0
 467 |   }
 468 |   convolution_param {
 469 |     num_output: 161
 470 |     pad: 1
 471 |     kernel_size: 3
 472 |     weight_filler {
 473 |       type: "xavier"
 474 |     }
 475 |   }
 476 | }
 477 | layer {
 478 |   name: "fire6/relu_expand3x3"
 479 |   type: "ReLU"
 480 |   bottom: "fire6/expand3x3"
 481 |   top: "fire6/expand3x3"
 482 | }
 483 | layer {
 484 |   name: "fire6/concat"
 485 |   type: "Concat"
 486 |   bottom: "fire6/expand1x1"
 487 |   bottom: "fire6/expand3x3"
 488 |   top: "fire6/concat"
 489 | }
 490 | layer {
 491 |   name: "fire7/squeeze1x1"
 492 |   type: "Convolution"
 493 |   bottom: "fire6/concat"
 494 |   top: "fire7/squeeze1x1"
 495 |   param {
 496 |     lr_mult: 1.0
 497 |     decay_mult: 0.0
 498 |   }
 499 |   param {
 500 |     lr_mult: 1.0
 501 |     decay_mult: 0.0
 502 |   }
 503 |   convolution_param {
 504 |     num_output: 45
 505 |     kernel_size: 1
 506 |     weight_filler {
 507 |       type: "xavier"
 508 |     }
 509 |   }
 510 | }
 511 | layer {
 512 |   name: "fire7/expand1x1"
 513 |   type: "Convolution"
 514 |   bottom: "fire7/squeeze1x1"
 515 |   top: "fire7/expand1x1"
 516 |   param {
 517 |     lr_mult: 1.0
 518 |     decay_mult: 0.0
 519 |   }
 520 |   param {
 521 |     lr_mult: 1.0
 522 |     decay_mult: 0.0
 523 |   }
 524 |   convolution_param {
 525 |     num_output: 155
 526 |     kernel_size: 1
 527 |     weight_filler {
 528 |       type: "xavier"
 529 |     }
 530 |   }
 531 | }
 532 | layer {
 533 |   name: "fire7/relu_expand1x1"
 534 |   type: "ReLU"
 535 |   bottom: "fire7/expand1x1"
 536 |   top: "fire7/expand1x1"
 537 | }
 538 | layer {
 539 |   name: "fire7/expand3x3"
 540 |   type: "Convolution"
 541 |   bottom: "fire7/squeeze1x1"
 542 |   top: "fire7/expand3x3"
 543 |   param {
 544 |     lr_mult: 1.0
 545 |     decay_mult: 0.0
 546 |   }
 547 |   param {
 548 |     lr_mult: 1.0
 549 |     decay_mult: 0.0
 550 |   }
 551 |   convolution_param {
 552 |     num_output: 146
 553 |     pad: 1
 554 |     kernel_size: 3
 555 |     weight_filler {
 556 |       type: "xavier"
 557 |     }
 558 |   }
 559 | }
 560 | layer {
 561 |   name: "fire7/relu_expand3x3"
 562 |   type: "ReLU"
 563 |   bottom: "fire7/expand3x3"
 564 |   top: "fire7/expand3x3"
 565 | }
 566 | layer {
 567 |   name: "fire7/concat"
 568 |   type: "Concat"
 569 |   bottom: "fire7/expand1x1"
 570 |   bottom: "fire7/expand3x3"
 571 |   top: "fire7/concat"
 572 | }
 573 | layer {
 574 |   name: "fire8/squeeze1x1"
 575 |   type: "Convolution"
 576 |   bottom: "fire7/concat"
 577 |   top: "fire8/squeeze1x1"
 578 |   param {
 579 |     lr_mult: 1.0
 580 |     decay_mult: 0.0
 581 |   }
 582 |   param {
 583 |     lr_mult: 1.0
 584 |     decay_mult: 0.0
 585 |   }
 586 |   convolution_param {
 587 |     num_output: 49
 588 |     kernel_size: 1
 589 |     weight_filler {
 590 |       type: "xavier"
 591 |     }
 592 |   }
 593 | }
 594 | layer {
 595 |   name: "fire8/expand1x1"
 596 |   type: "Convolution"
 597 |   bottom: "fire8/squeeze1x1"
 598 |   top: "fire8/expand1x1"
 599 |   param {
 600 |     lr_mult: 1.0
 601 |     decay_mult: 0.0
 602 |   }
 603 |   param {
 604 |     lr_mult: 1.0
 605 |     decay_mult: 0.0
 606 |   }
 607 |   convolution_param {
 608 |     num_output: 163
 609 |     kernel_size: 1
 610 |     weight_filler {
 611 |       type: "xavier"
 612 |     }
 613 |   }
 614 | }
 615 | layer {
 616 |   name: "fire8/relu_expand1x1"
 617 |   type: "ReLU"
 618 |   bottom: "fire8/expand1x1"
 619 |   top: "fire8/expand1x1"
 620 | }
 621 | layer {
 622 |   name: "fire8/expand3x3"
 623 |   type: "Convolution"
 624 |   bottom: "fire8/squeeze1x1"
 625 |   top: "fire8/expand3x3"
 626 |   param {
 627 |     lr_mult: 1.0
 628 |     decay_mult: 0.0
 629 |   }
 630 |   param {
 631 |     lr_mult: 1.0
 632 |     decay_mult: 0.0
 633 |   }
 634 |   convolution_param {
 635 |     num_output: 171
 636 |     pad: 1
 637 |     kernel_size: 3
 638 |     weight_filler {
 639 |       type: "xavier"
 640 |     }
 641 |   }
 642 | }
 643 | layer {
 644 |   name: "fire8/relu_expand3x3"
 645 |   type: "ReLU"
 646 |   bottom: "fire8/expand3x3"
 647 |   top: "fire8/expand3x3"
 648 | }
 649 | layer {
 650 |   name: "fire8/concat"
 651 |   type: "Concat"
 652 |   bottom: "fire8/expand1x1"
 653 |   bottom: "fire8/expand3x3"
 654 |   top: "fire8/concat"
 655 | }
 656 | layer {
 657 |   name: "fire9/squeeze1x1"
 658 |   type: "Convolution"
 659 |   bottom: "fire8/concat"
 660 |   top: "fire9/squeeze1x1"
 661 |   param {
 662 |     lr_mult: 1.0
 663 |     decay_mult: 0.0
 664 |   }
 665 |   param {
 666 |     lr_mult: 1.0
 667 |     decay_mult: 0.0
 668 |   }
 669 |   convolution_param {
 670 |     num_output: 25
 671 |     kernel_size: 1
 672 |     weight_filler {
 673 |       type: "xavier"
 674 |     }
 675 |   }
 676 | }
 677 | layer {
 678 |   name: "fire9/expand1x1"
 679 |   type: "Convolution"
 680 |   bottom: "fire9/squeeze1x1"
 681 |   top: "fire9/expand1x1"
 682 |   param {
 683 |     lr_mult: 1.0
 684 |     decay_mult: 0.0
 685 |   }
 686 |   param {
 687 |     lr_mult: 1.0
 688 |     decay_mult: 0.0
 689 |   }
 690 |   convolution_param {
 691 |     num_output: 29
 692 |     kernel_size: 1
 693 |     weight_filler {
 694 |       type: "xavier"
 695 |     }
 696 |   }
 697 | }
 698 | layer {
 699 |   name: "fire9/expand1x1/bn"
 700 |   type: "BatchNorm"
 701 |   bottom: "fire9/expand1x1"
 702 |   top: "fire9/expand1x1"
 703 |   param {
 704 |     lr_mult: 0.0
 705 |     decay_mult: 0.0
 706 |   }
 707 |   param {
 708 |     lr_mult: 0.0
 709 |     decay_mult: 0.0
 710 |   }
 711 |   param {
 712 |     lr_mult: 0.0
 713 |     decay_mult: 0.0
 714 |   }
 715 | }
 716 | layer {
 717 |   name: "fire9/expand1x1/scale"
 718 |   type: "Scale"
 719 |   bottom: "fire9/expand1x1"
 720 |   top: "fire9/expand1x1"
 721 |   param {
 722 |     lr_mult: 1.0
 723 |     decay_mult: 0.0
 724 |   }
 725 |   param {
 726 |     lr_mult: 2.0
 727 |     decay_mult: 0.0
 728 |   }
 729 |   scale_param {
 730 |     filler {
 731 |       value: 1.0
 732 |     }
 733 |     bias_term: true
 734 |     bias_filler {
 735 |       value: 0.0
 736 |     }
 737 |   }
 738 | }
 739 | layer {
 740 |   name: "fire9/relu_expand1x1"
 741 |   type: "ReLU"
 742 |   bottom: "fire9/expand1x1"
 743 |   top: "fire9/expand1x1"
 744 | }
 745 | layer {
 746 |   name: "fire9/expand3x3"
 747 |   type: "Convolution"
 748 |   bottom: "fire9/squeeze1x1"
 749 |   top: "fire9/expand3x3"
 750 |   param {
 751 |     lr_mult: 1.0
 752 |     decay_mult: 0.0
 753 |   }
 754 |   param {
 755 |     lr_mult: 1.0
 756 |     decay_mult: 0.0
 757 |   }
 758 |   convolution_param {
 759 |     num_output: 54
 760 |     pad: 1
 761 |     kernel_size: 3
 762 |     weight_filler {
 763 |       type: "xavier"
 764 |     }
 765 |   }
 766 | }
 767 | layer {
 768 |   name: "fire9/expand3x3/bn"
 769 |   type: "BatchNorm"
 770 |   bottom: "fire9/expand3x3"
 771 |   top: "fire9/expand3x3"
 772 |   param {
 773 |     lr_mult: 0.0
 774 |     decay_mult: 0.0
 775 |   }
 776 |   param {
 777 |     lr_mult: 0.0
 778 |     decay_mult: 0.0
 779 |   }
 780 |   param {
 781 |     lr_mult: 0.0
 782 |     decay_mult: 0.0
 783 |   }
 784 | }
 785 | layer {
 786 |   name: "fire9/expand3x3/scale"
 787 |   type: "Scale"
 788 |   bottom: "fire9/expand3x3"
 789 |   top: "fire9/expand3x3"
 790 |   param {
 791 |     lr_mult: 1.0
 792 |     decay_mult: 0.0
 793 |   }
 794 |   param {
 795 |     lr_mult: 2.0
 796 |     decay_mult: 0.0
 797 |   }
 798 |   scale_param {
 799 |     filler {
 800 |       value: 1.0
 801 |     }
 802 |     bias_term: true
 803 |     bias_filler {
 804 |       value: 0.0
 805 |     }
 806 |   }
 807 | }
 808 | layer {
 809 |   name: "fire9/relu_expand3x3"
 810 |   type: "ReLU"
 811 |   bottom: "fire9/expand3x3"
 812 |   top: "fire9/expand3x3"
 813 | }
 814 | layer {
 815 |   name: "fire9/concat"
 816 |   type: "Concat"
 817 |   bottom: "fire9/expand1x1"
 818 |   bottom: "fire9/expand3x3"
 819 |   top: "fire9/concat"
 820 | }
 821 | layer {
 822 |   name: "pool9"
 823 |   type: "Pooling"
 824 |   bottom: "fire9/concat"
 825 |   top: "pool9"
 826 |   pooling_param {
 827 |     pool: MAX
 828 |     kernel_size: 3
 829 |     stride: 2
 830 |   }
 831 | }
 832 | layer {
 833 |   name: "fire10/squeeze1x1"
 834 |   type: "Convolution"
 835 |   bottom: "pool9"
 836 |   top: "fire10/squeeze1x1"
 837 |   convolution_param {
 838 |     num_output: 37
 839 |     kernel_size: 1
 840 |     weight_filler {
 841 |       type: "xavier"
 842 |     }
 843 |   }
 844 | }
 845 | layer {
 846 |   name: "fire10/squeeze1x1/bn"
 847 |   type: "BatchNorm"
 848 |   bottom: "fire10/squeeze1x1"
 849 |   top: "fire10/squeeze1x1"
 850 |   param {
 851 |     lr_mult: 0.0
 852 |     decay_mult: 0.0
 853 |   }
 854 |   param {
 855 |     lr_mult: 0.0
 856 |     decay_mult: 0.0
 857 |   }
 858 |   param {
 859 |     lr_mult: 0.0
 860 |     decay_mult: 0.0
 861 |   }
 862 | }
 863 | layer {
 864 |   name: "fire10/squeeze1x1/scale"
 865 |   type: "Scale"
 866 |   bottom: "fire10/squeeze1x1"
 867 |   top: "fire10/squeeze1x1"
 868 |   param {
 869 |     lr_mult: 1.0
 870 |     decay_mult: 0.0
 871 |   }
 872 |   param {
 873 |     lr_mult: 2.0
 874 |     decay_mult: 0.0
 875 |   }
 876 |   scale_param {
 877 |     filler {
 878 |       value: 1.0
 879 |     }
 880 |     bias_term: true
 881 |     bias_filler {
 882 |       value: 0.0
 883 |     }
 884 |   }
 885 | }
 886 | layer {
 887 |   name: "fire10/expand1x1"
 888 |   type: "Convolution"
 889 |   bottom: "fire10/squeeze1x1"
 890 |   top: "fire10/expand1x1"
 891 |   convolution_param {
 892 |     num_output: 45
 893 |     kernel_size: 1
 894 |     weight_filler {
 895 |       type: "xavier"
 896 |     }
 897 |   }
 898 | }
 899 | layer {
 900 |   name: "fire10/expand1x1/bn"
 901 |   type: "BatchNorm"
 902 |   bottom: "fire10/expand1x1"
 903 |   top: "fire10/expand1x1"
 904 |   param {
 905 |     lr_mult: 0.0
 906 |     decay_mult: 0.0
 907 |   }
 908 |   param {
 909 |     lr_mult: 0.0
 910 |     decay_mult: 0.0
 911 |   }
 912 |   param {
 913 |     lr_mult: 0.0
 914 |     decay_mult: 0.0
 915 |   }
 916 | }
 917 | layer {
 918 |   name: "fire10/expand1x1/scale"
 919 |   type: "Scale"
 920 |   bottom: "fire10/expand1x1"
 921 |   top: "fire10/expand1x1"
 922 |   param {
 923 |     lr_mult: 1.0
 924 |     decay_mult: 0.0
 925 |   }
 926 |   param {
 927 |     lr_mult: 2.0
 928 |     decay_mult: 0.0
 929 |   }
 930 |   scale_param {
 931 |     filler {
 932 |       value: 1.0
 933 |     }
 934 |     bias_term: true
 935 |     bias_filler {
 936 |       value: 0.0
 937 |     }
 938 |   }
 939 | }
 940 | layer {
 941 |   name: "fire10/relu_expand1x1"
 942 |   type: "ReLU"
 943 |   bottom: "fire10/expand1x1"
 944 |   top: "fire10/expand1x1"
 945 | }
 946 | layer {
 947 |   name: "fire10/expand3x3"
 948 |   type: "Convolution"
 949 |   bottom: "fire10/squeeze1x1"
 950 |   top: "fire10/expand3x3"
 951 |   convolution_param {
 952 |     num_output: 56
 953 |     pad: 1
 954 |     kernel_size: 3
 955 |     weight_filler {
 956 |       type: "xavier"
 957 |     }
 958 |   }
 959 | }
 960 | layer {
 961 |   name: "fire10/expand3x3/bn"
 962 |   type: "BatchNorm"
 963 |   bottom: "fire10/expand3x3"
 964 |   top: "fire10/expand3x3"
 965 |   param {
 966 |     lr_mult: 0.0
 967 |     decay_mult: 0.0
 968 |   }
 969 |   param {
 970 |     lr_mult: 0.0
 971 |     decay_mult: 0.0
 972 |   }
 973 |   param {
 974 |     lr_mult: 0.0
 975 |     decay_mult: 0.0
 976 |   }
 977 | }
 978 | layer {
 979 |   name: "fire10/expand3x3/scale"
 980 |   type: "Scale"
 981 |   bottom: "fire10/expand3x3"
 982 |   top: "fire10/expand3x3"
 983 |   param {
 984 |     lr_mult: 1.0
 985 |     decay_mult: 0.0
 986 |   }
 987 |   param {
 988 |     lr_mult: 2.0
 989 |     decay_mult: 0.0
 990 |   }
 991 |   scale_param {
 992 |     filler {
 993 |       value: 1.0
 994 |     }
 995 |     bias_term: true
 996 |     bias_filler {
 997 |       value: 0.0
 998 |     }
 999 |   }
1000 | }
1001 | layer {
1002 |   name: "fire10/relu_expand3x3"
1003 |   type: "ReLU"
1004 |   bottom: "fire10/expand3x3"
1005 |   top: "fire10/expand3x3"
1006 | }
1007 | layer {
1008 |   name: "fire10/concat"
1009 |   type: "Concat"
1010 |   bottom: "fire10/expand1x1"
1011 |   bottom: "fire10/expand3x3"
1012 |   top: "fire10/concat"
1013 | }
1014 | layer {
1015 |   name: "pool10"
1016 |   type: "Pooling"
1017 |   bottom: "fire10/concat"
1018 |   top: "pool10"
1019 |   pooling_param {
1020 |     pool: MAX
1021 |     kernel_size: 3
1022 |     stride: 2
1023 |   }
1024 | }
1025 | layer {
1026 |   name: "fire11/squeeze1x1"
1027 |   type: "Convolution"
1028 |   bottom: "pool10"
1029 |   top: "fire11/squeeze1x1"
1030 |   convolution_param {
1031 |     num_output: 38
1032 |     kernel_size: 1
1033 |     weight_filler {
1034 |       type: "xavier"
1035 |     }
1036 |   }
1037 | }
1038 | layer {
1039 |   name: "fire11/squeeze1x1/bn"
1040 |   type: "BatchNorm"
1041 |   bottom: "fire11/squeeze1x1"
1042 |   top: "fire11/squeeze1x1"
1043 |   param {
1044 |     lr_mult: 0.0
1045 |     decay_mult: 0.0
1046 |   }
1047 |   param {
1048 |     lr_mult: 0.0
1049 |     decay_mult: 0.0
1050 |   }
1051 |   param {
1052 |     lr_mult: 0.0
1053 |     decay_mult: 0.0
1054 |   }
1055 | }
1056 | layer {
1057 |   name: "fire11/squeeze1x1/scale"
1058 |   type: "Scale"
1059 |   bottom: "fire11/squeeze1x1"
1060 |   top: "fire11/squeeze1x1"
1061 |   param {
1062 |     lr_mult: 1.0
1063 |     decay_mult: 0.0
1064 |   }
1065 |   param {
1066 |     lr_mult: 2.0
1067 |     decay_mult: 0.0
1068 |   }
1069 |   scale_param {
1070 |     filler {
1071 |       value: 1.0
1072 |     }
1073 |     bias_term: true
1074 |     bias_filler {
1075 |       value: 0.0
1076 |     }
1077 |   }
1078 | }
1079 | layer {
1080 |   name: "fire11/expand1x1"
1081 |   type: "Convolution"
1082 |   bottom: "fire11/squeeze1x1"
1083 |   top: "fire11/expand1x1"
1084 |   convolution_param {
1085 |     num_output: 41
1086 |     kernel_size: 1
1087 |     weight_filler {
1088 |       type: "xavier"
1089 |     }
1090 |   }
1091 | }
1092 | layer {
1093 |   name: "fire11/expand1x1/bn"
1094 |   type: "BatchNorm"
1095 |   bottom: "fire11/expand1x1"
1096 |   top: "fire11/expand1x1"
1097 |   param {
1098 |     lr_mult: 0.0
1099 |     decay_mult: 0.0
1100 |   }
1101 |   param {
1102 |     lr_mult: 0.0
1103 |     decay_mult: 0.0
1104 |   }
1105 |   param {
1106 |     lr_mult: 0.0
1107 |     decay_mult: 0.0
1108 |   }
1109 | }
1110 | layer {
1111 |   name: "fire11/expand1x1/scale"
1112 |   type: "Scale"
1113 |   bottom: "fire11/expand1x1"
1114 |   top: "fire11/expand1x1"
1115 |   param {
1116 |     lr_mult: 1.0
1117 |     decay_mult: 0.0
1118 |   }
1119 |   param {
1120 |     lr_mult: 2.0
1121 |     decay_mult: 0.0
1122 |   }
1123 |   scale_param {
1124 |     filler {
1125 |       value: 1.0
1126 |     }
1127 |     bias_term: true
1128 |     bias_filler {
1129 |       value: 0.0
1130 |     }
1131 |   }
1132 | }
1133 | layer {
1134 |   name: "fire11/relu_expand1x1"
1135 |   type: "ReLU"
1136 |   bottom: "fire11/expand1x1"
1137 |   top: "fire11/expand1x1"
1138 | }
1139 | layer {
1140 |   name: "fire11/expand3x3"
1141 |   type: "Convolution"
1142 |   bottom: "fire11/squeeze1x1"
1143 |   top: "fire11/expand3x3"
1144 |   convolution_param {
1145 |     num_output: 44
1146 |     pad: 1
1147 |     kernel_size: 3
1148 |     weight_filler {
1149 |       type: "xavier"
1150 |     }
1151 |   }
1152 | }
1153 | layer {
1154 |   name: "fire11/expand3x3/bn"
1155 |   type: "BatchNorm"
1156 |   bottom: "fire11/expand3x3"
1157 |   top: "fire11/expand3x3"
1158 |   param {
1159 |     lr_mult: 0.0
1160 |     decay_mult: 0.0
1161 |   }
1162 |   param {
1163 |     lr_mult: 0.0
1164 |     decay_mult: 0.0
1165 |   }
1166 |   param {
1167 |     lr_mult: 0.0
1168 |     decay_mult: 0.0
1169 |   }
1170 | }
1171 | layer {
1172 |   name: "fire11/expand3x3/scale"
1173 |   type: "Scale"
1174 |   bottom: "fire11/expand3x3"
1175 |   top: "fire11/expand3x3"
1176 |   param {
1177 |     lr_mult: 1.0
1178 |     decay_mult: 0.0
1179 |   }
1180 |   param {
1181 |     lr_mult: 2.0
1182 |     decay_mult: 0.0
1183 |   }
1184 |   scale_param {
1185 |     filler {
1186 |       value: 1.0
1187 |     }
1188 |     bias_term: true
1189 |     bias_filler {
1190 |       value: 0.0
1191 |     }
1192 |   }
1193 | }
1194 | layer {
1195 |   name: "fire11/relu_expand3x3"
1196 |   type: "ReLU"
1197 |   bottom: "fire11/expand3x3"
1198 |   top: "fire11/expand3x3"
1199 | }
1200 | layer {
1201 |   name: "fire11/concat"
1202 |   type: "Concat"
1203 |   bottom: "fire11/expand1x1"
1204 |   bottom: "fire11/expand3x3"
1205 |   top: "fire11/concat"
1206 | }
1207 | layer {
1208 |   name: "conv12_1"
1209 |   type: "Convolution"
1210 |   bottom: "fire11/concat"
1211 |   top: "conv12_1"
1212 |   param {
1213 |     lr_mult: 1.0
1214 |     decay_mult: 1.0
1215 |   }
1216 |   convolution_param {
1217 |     num_output: 51
1218 |     bias_term: false
1219 |     kernel_size: 1
1220 |     weight_filler {
1221 |       type: "msra"
1222 |     }
1223 |   }
1224 | }
1225 | layer {
1226 |   name: "conv12_1/bn"
1227 |   type: "BatchNorm"
1228 |   bottom: "conv12_1"
1229 |   top: "conv12_1"
1230 |   param {
1231 |     lr_mult: 0.0
1232 |     decay_mult: 0.0
1233 |   }
1234 |   param {
1235 |     lr_mult: 0.0
1236 |     decay_mult: 0.0
1237 |   }
1238 |   param {
1239 |     lr_mult: 0.0
1240 |     decay_mult: 0.0
1241 |   }
1242 | }
1243 | layer {
1244 |   name: "conv12_1/scale"
1245 |   type: "Scale"
1246 |   bottom: "conv12_1"
1247 |   top: "conv12_1"
1248 |   param {
1249 |     lr_mult: 1.0
1250 |     decay_mult: 0.0
1251 |   }
1252 |   param {
1253 |     lr_mult: 2.0
1254 |     decay_mult: 0.0
1255 |   }
1256 |   scale_param {
1257 |     filler {
1258 |       value: 1.0
1259 |     }
1260 |     bias_term: true
1261 |     bias_filler {
1262 |       value: 0.0
1263 |     }
1264 |   }
1265 | }
1266 | layer {
1267 |   name: "conv12_1/relu"
1268 |   type: "ReLU"
1269 |   bottom: "conv12_1"
1270 |   top: "conv12_1"
1271 | }
1272 | layer {
1273 |   name: "conv12_2"
1274 |   type: "Convolution"
1275 |   bottom: "conv12_1"
1276 |   top: "conv12_2"
1277 |   param {
1278 |     lr_mult: 1.0
1279 |     decay_mult: 1.0
1280 |   }
1281 |   convolution_param {
1282 |     num_output: 46
1283 |     bias_term: false
1284 |     pad: 1
1285 |     kernel_size: 3
1286 |     stride: 2
1287 |     weight_filler {
1288 |       type: "msra"
1289 |     }
1290 |   }
1291 | }
1292 | layer {
1293 |   name: "conv12_2/bn"
1294 |   type: "BatchNorm"
1295 |   bottom: "conv12_2"
1296 |   top: "conv12_2"
1297 |   param {
1298 |     lr_mult: 0.0
1299 |     decay_mult: 0.0
1300 |   }
1301 |   param {
1302 |     lr_mult: 0.0
1303 |     decay_mult: 0.0
1304 |   }
1305 |   param {
1306 |     lr_mult: 0.0
1307 |     decay_mult: 0.0
1308 |   }
1309 | }
1310 | layer {
1311 |   name: "conv12_2/scale"
1312 |   type: "Scale"
1313 |   bottom: "conv12_2"
1314 |   top: "conv12_2"
1315 |   param {
1316 |     lr_mult: 1.0
1317 |     decay_mult: 0.0
1318 |   }
1319 |   param {
1320 |     lr_mult: 2.0
1321 |     decay_mult: 0.0
1322 |   }
1323 |   scale_param {
1324 |     filler {
1325 |       value: 1.0
1326 |     }
1327 |     bias_term: true
1328 |     bias_filler {
1329 |       value: 0.0
1330 |     }
1331 |   }
1332 | }
1333 | layer {
1334 |   name: "conv12_2/relu"
1335 |   type: "ReLU"
1336 |   bottom: "conv12_2"
1337 |   top: "conv12_2"
1338 | }
1339 | layer {
1340 |   name: "conv13_1"
1341 |   type: "Convolution"
1342 |   bottom: "conv12_2"
1343 |   top: "conv13_1"
1344 |   param {
1345 |     lr_mult: 1.0
1346 |     decay_mult: 1.0
1347 |   }
1348 |   convolution_param {
1349 |     num_output: 55
1350 |     bias_term: false
1351 |     kernel_size: 1
1352 |     weight_filler {
1353 |       type: "msra"
1354 |     }
1355 |   }
1356 | }
1357 | layer {
1358 |   name: "conv13_1/bn"
1359 |   type: "BatchNorm"
1360 |   bottom: "conv13_1"
1361 |   top: "conv13_1"
1362 |   param {
1363 |     lr_mult: 0.0
1364 |     decay_mult: 0.0
1365 |   }
1366 |   param {
1367 |     lr_mult: 0.0
1368 |     decay_mult: 0.0
1369 |   }
1370 |   param {
1371 |     lr_mult: 0.0
1372 |     decay_mult: 0.0
1373 |   }
1374 | }
1375 | layer {
1376 |   name: "conv13_1/scale"
1377 |   type: "Scale"
1378 |   bottom: "conv13_1"
1379 |   top: "conv13_1"
1380 |   param {
1381 |     lr_mult: 1.0
1382 |     decay_mult: 0.0
1383 |   }
1384 |   param {
1385 |     lr_mult: 2.0
1386 |     decay_mult: 0.0
1387 |   }
1388 |   scale_param {
1389 |     filler {
1390 |       value: 1.0
1391 |     }
1392 |     bias_term: true
1393 |     bias_filler {
1394 |       value: 0.0
1395 |     }
1396 |   }
1397 | }
1398 | layer {
1399 |   name: "conv13_1/relu"
1400 |   type: "ReLU"
1401 |   bottom: "conv13_1"
1402 |   top: "conv13_1"
1403 | }
1404 | layer {
1405 |   name: "conv13_2"
1406 |   type: "Convolution"
1407 |   bottom: "conv13_1"
1408 |   top: "conv13_2"
1409 |   param {
1410 |     lr_mult: 1.0
1411 |     decay_mult: 1.0
1412 |   }
1413 |   convolution_param {
1414 |     num_output: 85
1415 |     bias_term: false
1416 |     pad: 1
1417 |     kernel_size: 3
1418 |     stride: 2
1419 |     weight_filler {
1420 |       type: "msra"
1421 |     }
1422 |   }
1423 | }
1424 | layer {
1425 |   name: "conv13_2/bn"
1426 |   type: "BatchNorm"
1427 |   bottom: "conv13_2"
1428 |   top: "conv13_2"
1429 |   param {
1430 |     lr_mult: 0.0
1431 |     decay_mult: 0.0
1432 |   }
1433 |   param {
1434 |     lr_mult: 0.0
1435 |     decay_mult: 0.0
1436 |   }
1437 |   param {
1438 |     lr_mult: 0.0
1439 |     decay_mult: 0.0
1440 |   }
1441 | }
1442 | layer {
1443 |   name: "conv13_2/scale"
1444 |   type: "Scale"
1445 |   bottom: "conv13_2"
1446 |   top: "conv13_2"
1447 |   param {
1448 |     lr_mult: 1.0
1449 |     decay_mult: 0.0
1450 |   }
1451 |   param {
1452 |     lr_mult: 2.0
1453 |     decay_mult: 0.0
1454 |   }
1455 |   scale_param {
1456 |     filler {
1457 |       value: 1.0
1458 |     }
1459 |     bias_term: true
1460 |     bias_filler {
1461 |       value: 0.0
1462 |     }
1463 |   }
1464 | }
1465 | layer {
1466 |   name: "conv13_2/relu"
1467 |   type: "ReLU"
1468 |   bottom: "conv13_2"
1469 |   top: "conv13_2"
1470 | }
1471 | layer {
1472 |   name: "fire5/bn"
1473 |   type: "BatchNorm"
1474 |   bottom: "fire5/concat"
1475 |   top: "fire5/normal"
1476 |   param {
1477 |     lr_mult: 0.0
1478 |     decay_mult: 0.0
1479 |   }
1480 |   param {
1481 |     lr_mult: 0.0
1482 |     decay_mult: 0.0
1483 |   }
1484 |   param {
1485 |     lr_mult: 0.0
1486 |     decay_mult: 0.0
1487 |   }
1488 | }
1489 | layer {
1490 |   name: "fire5/scale"
1491 |   type: "Scale"
1492 |   bottom: "fire5/normal"
1493 |   top: "fire5/normal"
1494 |   param {
1495 |     lr_mult: 1.0
1496 |     decay_mult: 0.0
1497 |   }
1498 |   param {
1499 |     lr_mult: 2.0
1500 |     decay_mult: 0.0
1501 |   }
1502 |   scale_param {
1503 |     filler {
1504 |       value: 1.0
1505 |     }
1506 |     bias_term: true
1507 |     bias_filler {
1508 |       value: 0.0
1509 |     }
1510 |   }
1511 | }
1512 | layer {
1513 |   name: "fire5_mbox_loc"
1514 |   type: "Convolution"
1515 |   bottom: "fire5/normal"
1516 |   top: "fire5_mbox_loc"
1517 |   param {
1518 |     lr_mult: 1.0
1519 |     decay_mult: 1.0
1520 |   }
1521 |   param {
1522 |     lr_mult: 2.0
1523 |     decay_mult: 0.0
1524 |   }
1525 |   convolution_param {
1526 |     num_output: 16
1527 |     pad: 1
1528 |     kernel_size: 3
1529 |     stride: 1
1530 |     weight_filler {
1531 |       type: "msra"
1532 |     }
1533 |     bias_filler {
1534 |       type: "constant"
1535 |       value: 0.0
1536 |     }
1537 |   }
1538 | }
1539 | layer {
1540 |   name: "fire5_mbox_loc_perm"
1541 |   type: "Permute"
1542 |   bottom: "fire5_mbox_loc"
1543 |   top: "fire5_mbox_loc_perm"
1544 |   permute_param {
1545 |     order: 0
1546 |     order: 2
1547 |     order: 3
1548 |     order: 1
1549 |   }
1550 | }
1551 | layer {
1552 |   name: "fire5_mbox_loc_flat"
1553 |   type: "Flatten"
1554 |   bottom: "fire5_mbox_loc_perm"
1555 |   top: "fire5_mbox_loc_flat"
1556 |   flatten_param {
1557 |     axis: 1
1558 |   }
1559 | }
1560 | layer {
1561 |   name: "fire5_mbox_conf"
1562 |   type: "Convolution"
1563 |   bottom: "fire5/normal"
1564 |   top: "fire5_mbox_conf"
1565 |   param {
1566 |     lr_mult: 1.0
1567 |     decay_mult: 1.0
1568 |   }
1569 |   param {
1570 |     lr_mult: 2.0
1571 |     decay_mult: 0.0
1572 |   }
1573 |   convolution_param {
1574 |     num_output: 84
1575 |     pad: 1
1576 |     kernel_size: 3
1577 |     stride: 1
1578 |     weight_filler {
1579 |       type: "msra"
1580 |     }
1581 |     bias_filler {
1582 |       type: "constant"
1583 |       value: 0.0
1584 |     }
1585 |   }
1586 | }
1587 | layer {
1588 |   name: "fire5_mbox_conf_perm"
1589 |   type: "Permute"
1590 |   bottom: "fire5_mbox_conf"
1591 |   top: "fire5_mbox_conf_perm"
1592 |   permute_param {
1593 |     order: 0
1594 |     order: 2
1595 |     order: 3
1596 |     order: 1
1597 |   }
1598 | }
1599 | layer {
1600 |   name: "fire5_mbox_conf_flat"
1601 |   type: "Flatten"
1602 |   bottom: "fire5_mbox_conf_perm"
1603 |   top: "fire5_mbox_conf_flat"
1604 |   flatten_param {
1605 |     axis: 1
1606 |   }
1607 | }
1608 | layer {
1609 |   name: "fire5_mbox_priorbox"
1610 |   type: "PriorBox"
1611 |   bottom: "fire5/normal"
1612 |   bottom: "data"
1613 |   top: "fire5_mbox_priorbox"
1614 |   prior_box_param {
1615 |     min_size: 21.0
1616 |     max_size: 45.0
1617 |     aspect_ratio: 2.0
1618 |     flip: true
1619 |     clip: false
1620 |     variance: 0.10000000149
1621 |     variance: 0.10000000149
1622 |     variance: 0.20000000298
1623 |     variance: 0.20000000298
1624 |     step: 8.0
1625 |   }
1626 | }
1627 | layer {
1628 |   name: "fire9_mbox_loc"
1629 |   type: "Convolution"
1630 |   bottom: "fire9/concat"
1631 |   top: "fire9_mbox_loc"
1632 |   param {
1633 |     lr_mult: 1.0
1634 |     decay_mult: 1.0
1635 |   }
1636 |   param {
1637 |     lr_mult: 2.0
1638 |     decay_mult: 0.0
1639 |   }
1640 |   convolution_param {
1641 |     num_output: 24
1642 |     pad: 1
1643 |     kernel_size: 3
1644 |     stride: 1
1645 |     weight_filler {
1646 |       type: "msra"
1647 |     }
1648 |     bias_filler {
1649 |       type: "constant"
1650 |       value: 0.0
1651 |     }
1652 |   }
1653 | }
1654 | layer {
1655 |   name: "fire9_mbox_loc_perm"
1656 |   type: "Permute"
1657 |   bottom: "fire9_mbox_loc"
1658 |   top: "fire9_mbox_loc_perm"
1659 |   permute_param {
1660 |     order: 0
1661 |     order: 2
1662 |     order: 3
1663 |     order: 1
1664 |   }
1665 | }
1666 | layer {
1667 |   name: "fire9_mbox_loc_flat"
1668 |   type: "Flatten"
1669 |   bottom: "fire9_mbox_loc_perm"
1670 |   top: "fire9_mbox_loc_flat"
1671 |   flatten_param {
1672 |     axis: 1
1673 |   }
1674 | }
1675 | layer {
1676 |   name: "fire9_mbox_conf"
1677 |   type: "Convolution"
1678 |   bottom: "fire9/concat"
1679 |   top: "fire9_mbox_conf"
1680 |   param {
1681 |     lr_mult: 1.0
1682 |     decay_mult: 1.0
1683 |   }
1684 |   param {
1685 |     lr_mult: 2.0
1686 |     decay_mult: 0.0
1687 |   }
1688 |   convolution_param {
1689 |     num_output: 126
1690 |     pad: 1
1691 |     kernel_size: 3
1692 |     stride: 1
1693 |     weight_filler {
1694 |       type: "msra"
1695 |     }
1696 |     bias_filler {
1697 |       type: "constant"
1698 |       value: 0.0
1699 |     }
1700 |   }
1701 | }
1702 | layer {
1703 |   name: "fire9_mbox_conf_perm"
1704 |   type: "Permute"
1705 |   bottom: "fire9_mbox_conf"
1706 |   top: "fire9_mbox_conf_perm"
1707 |   permute_param {
1708 |     order: 0
1709 |     order: 2
1710 |     order: 3
1711 |     order: 1
1712 |   }
1713 | }
1714 | layer {
1715 |   name: "fire9_mbox_conf_flat"
1716 |   type: "Flatten"
1717 |   bottom: "fire9_mbox_conf_perm"
1718 |   top: "fire9_mbox_conf_flat"
1719 |   flatten_param {
1720 |     axis: 1
1721 |   }
1722 | }
1723 | layer {
1724 |   name: "fire9_mbox_priorbox"
1725 |   type: "PriorBox"
1726 |   bottom: "fire9/concat"
1727 |   bottom: "data"
1728 |   top: "fire9_mbox_priorbox"
1729 |   prior_box_param {
1730 |     min_size: 45.0
1731 |     max_size: 99.0
1732 |     aspect_ratio: 2.0
1733 |     aspect_ratio: 3.0
1734 |     flip: true
1735 |     clip: false
1736 |     variance: 0.10000000149
1737 |     variance: 0.10000000149
1738 |     variance: 0.20000000298
1739 |     variance: 0.20000000298
1740 |     step: 16.0
1741 |   }
1742 | }
1743 | layer {
1744 |   name: "fire10_mbox_loc"
1745 |   type: "Convolution"
1746 |   bottom: "fire10/concat"
1747 |   top: "fire10_mbox_loc"
1748 |   param {
1749 |     lr_mult: 1.0
1750 |     decay_mult: 1.0
1751 |   }
1752 |   param {
1753 |     lr_mult: 2.0
1754 |     decay_mult: 0.0
1755 |   }
1756 |   convolution_param {
1757 |     num_output: 24
1758 |     pad: 1
1759 |     kernel_size: 3
1760 |     stride: 1
1761 |     weight_filler {
1762 |       type: "msra"
1763 |     }
1764 |     bias_filler {
1765 |       type: "constant"
1766 |       value: 0.0
1767 |     }
1768 |   }
1769 | }
1770 | layer {
1771 |   name: "fire10_mbox_loc_perm"
1772 |   type: "Permute"
1773 |   bottom: "fire10_mbox_loc"
1774 |   top: "fire10_mbox_loc_perm"
1775 |   permute_param {
1776 |     order: 0
1777 |     order: 2
1778 |     order: 3
1779 |     order: 1
1780 |   }
1781 | }
1782 | layer {
1783 |   name: "fire10_mbox_loc_flat"
1784 |   type: "Flatten"
1785 |   bottom: "fire10_mbox_loc_perm"
1786 |   top: "fire10_mbox_loc_flat"
1787 |   flatten_param {
1788 |     axis: 1
1789 |   }
1790 | }
1791 | layer {
1792 |   name: "fire10_mbox_conf"
1793 |   type: "Convolution"
1794 |   bottom: "fire10/concat"
1795 |   top: "fire10_mbox_conf"
1796 |   param {
1797 |     lr_mult: 1.0
1798 |     decay_mult: 1.0
1799 |   }
1800 |   param {
1801 |     lr_mult: 2.0
1802 |     decay_mult: 0.0
1803 |   }
1804 |   convolution_param {
1805 |     num_output: 126
1806 |     pad: 1
1807 |     kernel_size: 3
1808 |     stride: 1
1809 |     weight_filler {
1810 |       type: "msra"
1811 |     }
1812 |     bias_filler {
1813 |       type: "constant"
1814 |       value: 0.0
1815 |     }
1816 |   }
1817 | }
1818 | layer {
1819 |   name: "fire10_mbox_conf_perm"
1820 |   type: "Permute"
1821 |   bottom: "fire10_mbox_conf"
1822 |   top: "fire10_mbox_conf_perm"
1823 |   permute_param {
1824 |     order: 0
1825 |     order: 2
1826 |     order: 3
1827 |     order: 1
1828 |   }
1829 | }
1830 | layer {
1831 |   name: "fire10_mbox_conf_flat"
1832 |   type: "Flatten"
1833 |   bottom: "fire10_mbox_conf_perm"
1834 |   top: "fire10_mbox_conf_flat"
1835 |   flatten_param {
1836 |     axis: 1
1837 |   }
1838 | }
1839 | layer {
1840 |   name: "fire10_mbox_priorbox"
1841 |   type: "PriorBox"
1842 |   bottom: "fire10/concat"
1843 |   bottom: "data"
1844 |   top: "fire10_mbox_priorbox"
1845 |   prior_box_param {
1846 |     min_size: 99.0
1847 |     max_size: 153.0
1848 |     aspect_ratio: 2.0
1849 |     aspect_ratio: 3.0
1850 |     flip: true
1851 |     clip: false
1852 |     variance: 0.10000000149
1853 |     variance: 0.10000000149
1854 |     variance: 0.20000000298
1855 |     variance: 0.20000000298
1856 |     step: 32.0
1857 |   }
1858 | }
1859 | layer {
1860 |   name: "fire11_mbox_loc"
1861 |   type: "Convolution"
1862 |   bottom: "fire11/concat"
1863 |   top: "fire11_mbox_loc"
1864 |   param {
1865 |     lr_mult: 1.0
1866 |     decay_mult: 1.0
1867 |   }
1868 |   param {
1869 |     lr_mult: 2.0
1870 |     decay_mult: 0.0
1871 |   }
1872 |   convolution_param {
1873 |     num_output: 24
1874 |     pad: 1
1875 |     kernel_size: 3
1876 |     stride: 1
1877 |     weight_filler {
1878 |       type: "msra"
1879 |     }
1880 |     bias_filler {
1881 |       type: "constant"
1882 |       value: 0.0
1883 |     }
1884 |   }
1885 | }
1886 | layer {
1887 |   name: "fire11_mbox_loc_perm"
1888 |   type: "Permute"
1889 |   bottom: "fire11_mbox_loc"
1890 |   top: "fire11_mbox_loc_perm"
1891 |   permute_param {
1892 |     order: 0
1893 |     order: 2
1894 |     order: 3
1895 |     order: 1
1896 |   }
1897 | }
1898 | layer {
1899 |   name: "fire11_mbox_loc_flat"
1900 |   type: "Flatten"
1901 |   bottom: "fire11_mbox_loc_perm"
1902 |   top: "fire11_mbox_loc_flat"
1903 |   flatten_param {
1904 |     axis: 1
1905 |   }
1906 | }
1907 | layer {
1908 |   name: "fire11_mbox_conf"
1909 |   type: "Convolution"
1910 |   bottom: "fire11/concat"
1911 |   top: "fire11_mbox_conf"
1912 |   param {
1913 |     lr_mult: 1.0
1914 |     decay_mult: 1.0
1915 |   }
1916 |   param {
1917 |     lr_mult: 2.0
1918 |     decay_mult: 0.0
1919 |   }
1920 |   convolution_param {
1921 |     num_output: 126
1922 |     pad: 1
1923 |     kernel_size: 3
1924 |     stride: 1
1925 |     weight_filler {
1926 |       type: "msra"
1927 |     }
1928 |     bias_filler {
1929 |       type: "constant"
1930 |       value: 0.0
1931 |     }
1932 |   }
1933 | }
1934 | layer {
1935 |   name: "fire11_mbox_conf_perm"
1936 |   type: "Permute"
1937 |   bottom: "fire11_mbox_conf"
1938 |   top: "fire11_mbox_conf_perm"
1939 |   permute_param {
1940 |     order: 0
1941 |     order: 2
1942 |     order: 3
1943 |     order: 1
1944 |   }
1945 | }
1946 | layer {
1947 |   name: "fire11_mbox_conf_flat"
1948 |   type: "Flatten"
1949 |   bottom: "fire11_mbox_conf_perm"
1950 |   top: "fire11_mbox_conf_flat"
1951 |   flatten_param {
1952 |     axis: 1
1953 |   }
1954 | }
1955 | layer {
1956 |   name: "fire11_mbox_priorbox"
1957 |   type: "PriorBox"
1958 |   bottom: "fire11/concat"
1959 |   bottom: "data"
1960 |   top: "fire11_mbox_priorbox"
1961 |   prior_box_param {
1962 |     min_size: 153.0
1963 |     max_size: 207.0
1964 |     aspect_ratio: 2.0
1965 |     aspect_ratio: 3.0
1966 |     flip: true
1967 |     clip: false
1968 |     variance: 0.10000000149
1969 |     variance: 0.10000000149
1970 |     variance: 0.20000000298
1971 |     variance: 0.20000000298
1972 |     step: 64.0
1973 |   }
1974 | }
1975 | layer {
1976 |   name: "conv12_2_mbox_loc"
1977 |   type: "Convolution"
1978 |   bottom: "conv12_2"
1979 |   top: "conv12_2_mbox_loc"
1980 |   param {
1981 |     lr_mult: 1.0
1982 |     decay_mult: 1.0
1983 |   }
1984 |   param {
1985 |     lr_mult: 2.0
1986 |     decay_mult: 0.0
1987 |   }
1988 |   convolution_param {
1989 |     num_output: 24
1990 |     pad: 1
1991 |     kernel_size: 3
1992 |     stride: 1
1993 |     weight_filler {
1994 |       type: "msra"
1995 |     }
1996 |     bias_filler {
1997 |       type: "constant"
1998 |       value: 0.0
1999 |     }
2000 |   }
2001 | }
2002 | layer {
2003 |   name: "conv12_2_mbox_loc_perm"
2004 |   type: "Permute"
2005 |   bottom: "conv12_2_mbox_loc"
2006 |   top: "conv12_2_mbox_loc_perm"
2007 |   permute_param {
2008 |     order: 0
2009 |     order: 2
2010 |     order: 3
2011 |     order: 1
2012 |   }
2013 | }
2014 | layer {
2015 |   name: "conv12_2_mbox_loc_flat"
2016 |   type: "Flatten"
2017 |   bottom: "conv12_2_mbox_loc_perm"
2018 |   top: "conv12_2_mbox_loc_flat"
2019 |   flatten_param {
2020 |     axis: 1
2021 |   }
2022 | }
2023 | layer {
2024 |   name: "conv12_2_mbox_conf"
2025 |   type: "Convolution"
2026 |   bottom: "conv12_2"
2027 |   top: "conv12_2_mbox_conf"
2028 |   param {
2029 |     lr_mult: 1.0
2030 |     decay_mult: 1.0
2031 |   }
2032 |   param {
2033 |     lr_mult: 2.0
2034 |     decay_mult: 0.0
2035 |   }
2036 |   convolution_param {
2037 |     num_output: 126
2038 |     pad: 1
2039 |     kernel_size: 3
2040 |     stride: 1
2041 |     weight_filler {
2042 |       type: "msra"
2043 |     }
2044 |     bias_filler {
2045 |       type: "constant"
2046 |       value: 0.0
2047 |     }
2048 |   }
2049 | }
2050 | layer {
2051 |   name: "conv12_2_mbox_conf_perm"
2052 |   type: "Permute"
2053 |   bottom: "conv12_2_mbox_conf"
2054 |   top: "conv12_2_mbox_conf_perm"
2055 |   permute_param {
2056 |     order: 0
2057 |     order: 2
2058 |     order: 3
2059 |     order: 1
2060 |   }
2061 | }
2062 | layer {
2063 |   name: "conv12_2_mbox_conf_flat"
2064 |   type: "Flatten"
2065 |   bottom: "conv12_2_mbox_conf_perm"
2066 |   top: "conv12_2_mbox_conf_flat"
2067 |   flatten_param {
2068 |     axis: 1
2069 |   }
2070 | }
2071 | layer {
2072 |   name: "conv12_2_mbox_priorbox"
2073 |   type: "PriorBox"
2074 |   bottom: "conv12_2"
2075 |   bottom: "data"
2076 |   top: "conv12_2_mbox_priorbox"
2077 |   prior_box_param {
2078 |     min_size: 207.0
2079 |     max_size: 261.0
2080 |     aspect_ratio: 2.0
2081 |     aspect_ratio: 3.0
2082 |     flip: true
2083 |     clip: false
2084 |     variance: 0.10000000149
2085 |     variance: 0.10000000149
2086 |     variance: 0.20000000298
2087 |     variance: 0.20000000298
2088 |     step: 100.0
2089 |   }
2090 | }
2091 | layer {
2092 |   name: "conv13_2_mbox_loc"
2093 |   type: "Convolution"
2094 |   bottom: "conv13_2"
2095 |   top: "conv13_2_mbox_loc"
2096 |   param {
2097 |     lr_mult: 1.0
2098 |     decay_mult: 1.0
2099 |   }
2100 |   param {
2101 |     lr_mult: 2.0
2102 |     decay_mult: 0.0
2103 |   }
2104 |   convolution_param {
2105 |     num_output: 16
2106 |     pad: 1
2107 |     kernel_size: 3
2108 |     stride: 1
2109 |     weight_filler {
2110 |       type: "msra"
2111 |     }
2112 |     bias_filler {
2113 |       type: "constant"
2114 |       value: 0.0
2115 |     }
2116 |   }
2117 | }
2118 | layer {
2119 |   name: "conv13_2_mbox_loc_perm"
2120 |   type: "Permute"
2121 |   bottom: "conv13_2_mbox_loc"
2122 |   top: "conv13_2_mbox_loc_perm"
2123 |   permute_param {
2124 |     order: 0
2125 |     order: 2
2126 |     order: 3
2127 |     order: 1
2128 |   }
2129 | }
2130 | layer {
2131 |   name: "conv13_2_mbox_loc_flat"
2132 |   type: "Flatten"
2133 |   bottom: "conv13_2_mbox_loc_perm"
2134 |   top: "conv13_2_mbox_loc_flat"
2135 |   flatten_param {
2136 |     axis: 1
2137 |   }
2138 | }
2139 | layer {
2140 |   name: "conv13_2_mbox_conf"
2141 |   type: "Convolution"
2142 |   bottom: "conv13_2"
2143 |   top: "conv13_2_mbox_conf"
2144 |   param {
2145 |     lr_mult: 1.0
2146 |     decay_mult: 1.0
2147 |   }
2148 |   param {
2149 |     lr_mult: 2.0
2150 |     decay_mult: 0.0
2151 |   }
2152 |   convolution_param {
2153 |     num_output: 84
2154 |     pad: 1
2155 |     kernel_size: 3
2156 |     stride: 1
2157 |     weight_filler {
2158 |       type: "msra"
2159 |     }
2160 |     bias_filler {
2161 |       type: "constant"
2162 |       value: 0.0
2163 |     }
2164 |   }
2165 | }
2166 | layer {
2167 |   name: "conv13_2_mbox_conf_perm"
2168 |   type: "Permute"
2169 |   bottom: "conv13_2_mbox_conf"
2170 |   top: "conv13_2_mbox_conf_perm"
2171 |   permute_param {
2172 |     order: 0
2173 |     order: 2
2174 |     order: 3
2175 |     order: 1
2176 |   }
2177 | }
2178 | layer {
2179 |   name: "conv13_2_mbox_conf_flat"
2180 |   type: "Flatten"
2181 |   bottom: "conv13_2_mbox_conf_perm"
2182 |   top: "conv13_2_mbox_conf_flat"
2183 |   flatten_param {
2184 |     axis: 1
2185 |   }
2186 | }
2187 | layer {
2188 |   name: "conv13_2_mbox_priorbox"
2189 |   type: "PriorBox"
2190 |   bottom: "conv13_2"
2191 |   bottom: "data"
2192 |   top: "conv13_2_mbox_priorbox"
2193 |   prior_box_param {
2194 |     min_size: 261.0
2195 |     max_size: 315.0
2196 |     aspect_ratio: 2.0
2197 |     flip: true
2198 |     clip: false
2199 |     variance: 0.10000000149
2200 |     variance: 0.10000000149
2201 |     variance: 0.20000000298
2202 |     variance: 0.20000000298
2203 |     step: 300.0
2204 |   }
2205 | }
2206 | layer {
2207 |   name: "mbox_loc"
2208 |   type: "Concat"
2209 |   bottom: "fire5_mbox_loc_flat"
2210 |   bottom: "fire9_mbox_loc_flat"
2211 |   bottom: "fire10_mbox_loc_flat"
2212 |   bottom: "fire11_mbox_loc_flat"
2213 |   bottom: "conv12_2_mbox_loc_flat"
2214 |   bottom: "conv13_2_mbox_loc_flat"
2215 |   top: "mbox_loc"
2216 |   concat_param {
2217 |     axis: 1
2218 |   }
2219 | }
2220 | layer {
2221 |   name: "mbox_conf"
2222 |   type: "Concat"
2223 |   bottom: "fire5_mbox_conf_flat"
2224 |   bottom: "fire9_mbox_conf_flat"
2225 |   bottom: "fire10_mbox_conf_flat"
2226 |   bottom: "fire11_mbox_conf_flat"
2227 |   bottom: "conv12_2_mbox_conf_flat"
2228 |   bottom: "conv13_2_mbox_conf_flat"
2229 |   top: "mbox_conf"
2230 |   concat_param {
2231 |     axis: 1
2232 |   }
2233 | }
2234 | layer {
2235 |   name: "mbox_priorbox"
2236 |   type: "Concat"
2237 |   bottom: "fire5_mbox_priorbox"
2238 |   bottom: "fire9_mbox_priorbox"
2239 |   bottom: "fire10_mbox_priorbox"
2240 |   bottom: "fire11_mbox_priorbox"
2241 |   bottom: "conv12_2_mbox_priorbox"
2242 |   bottom: "conv13_2_mbox_priorbox"
2243 |   top: "mbox_priorbox"
2244 |   concat_param {
2245 |     axis: 2
2246 |   }
2247 | }
2248 | layer {
2249 |   name: "mbox_conf_reshape"
2250 |   type: "Reshape"
2251 |   bottom: "mbox_conf"
2252 |   top: "mbox_conf_reshape"
2253 |   reshape_param {
2254 |     shape {
2255 |       dim: 0
2256 |       dim: -1
2257 |       dim: 21
2258 |     }
2259 |   }
2260 | }
2261 | layer {
2262 |   name: "mbox_conf_softmax"
2263 |   type: "Softmax"
2264 |   bottom: "mbox_conf_reshape"
2265 |   top: "mbox_conf_softmax"
2266 |   softmax_param {
2267 |     axis: 2
2268 |   }
2269 | }
2270 | layer {
2271 |   name: "mbox_conf_flatten"
2272 |   type: "Flatten"
2273 |   bottom: "mbox_conf_softmax"
2274 |   top: "mbox_conf_flatten"
2275 |   flatten_param {
2276 |     axis: 1
2277 |   }
2278 | }
2279 | layer {
2280 |   name: "detection_out"
2281 |   type: "DetectionOutput"
2282 |   bottom: "mbox_loc"
2283 |   bottom: "mbox_conf_flatten"
2284 |   bottom: "mbox_priorbox"
2285 |   top: "detection_out"
2286 |   include {
2287 |     phase: TEST
2288 |   }
2289 |   detection_output_param {
2290 |     num_classes: 21
2291 |     share_location: true
2292 |     background_label_id: 0
2293 |     nms_param {
2294 |       nms_threshold: 0.449999988079
2295 |       top_k: 100
2296 |     }
2297 |     code_type: CENTER_SIZE
2298 |     keep_top_k: 100
2299 |     confidence_threshold: 0.25
2300 |   }
2301 | }
2302 | 


--------------------------------------------------------------------------------
/python_script/generate.py:
--------------------------------------------------------------------------------
  1 | # created by lampson.song @ 2018-07-14
  2 | # generate tiny-ssd prototxt
  3 | 
  4 | import sys, os
  5 | sys.path.insert(0,'/home/lampson/1T_disk/workspace/objectDetection/ssd-caffe/python') # your caffe path
  6 | import caffe
  7 | 
  8 | def fire(net, net_bottom, prefix, out1, out2, out3):
  9 |     if(prefix[0:6] == "fire10" or prefix[0:6] == "fire11"):
 10 |         net.tops[prefix+'/squeeze1x1'] = caffe.layers.Convolution(net_bottom, num_output=out1, 
 11 |                 kernel_size=1, weight_filler={"type":"xavier"})
 12 |         net.tops[prefix+'/squeeze1x1/bn'] = caffe.layers.BatchNorm(net.tops[prefix+'/squeeze1x1'], param=[dict(lr_mult=0, decay_mult=0), 
 13 |             dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
 14 |         net.tops[prefix+'/squeeze1x1/scale'] = caffe.layers.Scale(net.tops[prefix+'/squeeze1x1/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0),
 15 |             dict(lr_mult=2.0, decay_mult=0.0)], scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
 16 |     else:
 17 |         net.tops[prefix+'/squeeze1x1'] = caffe.layers.Convolution(net_bottom, num_output=out1, 
 18 |                 kernel_size=1, weight_filler={"type":"xavier"}, param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=1.0, decay_mult=0.0)])
 19 |    
 20 |     if(prefix[0:6] == "fire10" or prefix[0:6] == "fire11"):
 21 |         net.tops[prefix+'/expand1x1'] = caffe.layers.Convolution(net.tops[prefix+'/squeeze1x1'], num_output=out2, 
 22 |                 kernel_size=1, weight_filler={"type":"xavier"})
 23 |     else:
 24 |         net.tops[prefix+'/expand1x1'] = caffe.layers.Convolution(net.tops[prefix+'/squeeze1x1'], num_output=out2, 
 25 |                 kernel_size=1, weight_filler={"type":"xavier"}, param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=1.0, decay_mult=0.0)])
 26 |    
 27 |     next_top = net.tops[prefix+'/expand1x1']
 28 |     if(prefix[0:5] == "fire9" or prefix[0:6] == "fire10" or prefix[0:6] == "fire11"):
 29 |         net.tops[prefix+'/expand1x1/bn'] = caffe.layers.BatchNorm(next_top, param=[dict(lr_mult=0, decay_mult=0), 
 30 |             dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
 31 |         net.tops[prefix+'/expand1x1/scale'] = caffe.layers.Scale(net.tops[prefix+'/expand1x1/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0),
 32 |             dict(lr_mult=2.0, decay_mult=0.0)], scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
 33 |         
 34 |         next_top = net.tops[prefix+'/expand1x1/scale']
 35 |     
 36 |     net.tops[prefix+'/relu_expand1x1'] = caffe.layers.ReLU(next_top, in_place=True)
 37 | 
 38 |     if(prefix[0:6] == "fire10" or prefix[0:6] == "fire11"):
 39 |         net.tops[prefix+'/expand3x3'] = caffe.layers.Convolution(net.tops[prefix+'/squeeze1x1'], num_output=out3, pad=1,
 40 |                 kernel_size=3, weight_filler={'type':'xavier'})
 41 |     else:
 42 |         net.tops[prefix+'/expand3x3'] = caffe.layers.Convolution(net.tops[prefix+'/squeeze1x1'], num_output=out3, pad=1,
 43 |                 kernel_size=3, weight_filler={'type':'xavier'}, param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=1.0, decay_mult=0.0)])
 44 | 
 45 |     next_top = net.tops[prefix+'/expand3x3']
 46 |     if(prefix[0:5] == "fire9" or prefix[0:6] == "fire10" or prefix[0:6] == "fire11"):
 47 |         net.tops[prefix+'/expand3x3/bn'] = caffe.layers.BatchNorm(next_top, param=[dict(lr_mult=0, decay_mult=0), 
 48 |             dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
 49 |         net.tops[prefix+'/expand3x3/scale'] = caffe.layers.Scale(net.tops[prefix+'/expand3x3/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0),
 50 |             dict(lr_mult=2.0, decay_mult=0.0)], scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
 51 |         
 52 |         next_top = net.tops[prefix+'/expand3x3/scale']
 53 |     
 54 |     net.tops[prefix+'/relu_expand3x3'] = caffe.layers.ReLU(next_top, in_place=True)
 55 |     net.tops[prefix+'/concat'] = caffe.layers.Concat(net.tops[prefix+'/expand1x1'], net.tops[prefix+'/expand3x3'])
 56 | 
 57 |     return net.tops[prefix+'/concat'] 
 58 | 
 59 | 
 60 | def prior_box(net, net_bottom, prefix, out1, out2, min_size, max_size, aspect_ratio, step):
 61 |     # prior box
 62 |     if(prefix[0:5] == "fire5"):
 63 |         net.tops[prefix+'/normal'] = caffe.layers.BatchNorm(net_bottom, name=prefix+'/bn', param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)])
 64 |         net.tops[prefix+'/scale'] = caffe.layers.Scale(net.tops[prefix+'/normal'], name=prefix+'/scale', param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=2.0, decay_mult=0.0)],
 65 |                 scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
 66 |         net_bottom = net.tops[prefix+'/scale']
 67 |     net.tops[prefix+'_mbox_loc'] = caffe.layers.Convolution(net_bottom, param=[dict(lr_mult=1.0, decay_mult=1.0), dict(lr_mult=2.0, decay_mult=0.0)],
 68 |             convolution_param={'num_output':out1, 'pad':1, 'kernel_size':3, 'stride':1, 'weight_filler':{'type':'msra'},'bias_filler':{'type':'constant', 'value':0.0}})
 69 |     net.tops[prefix+'_mbox_loc_perm'] = caffe.layers.Permute(net.tops[prefix+'_mbox_loc'], permute_param={'order':[0,2,3,1]})
 70 |     net.tops[prefix+'_mbox_loc_flat'] = caffe.layers.Flatten(net.tops[prefix+'_mbox_loc_perm'], flatten_param={'axis':1})
 71 |     
 72 |     net.tops[prefix+'_mbox_conf'] = caffe.layers.Convolution(net_bottom, param=[dict(lr_mult=1.0, decay_mult=1.0), dict(lr_mult=2.0, decay_mult=0.0)],
 73 |             convolution_param={'num_output':out2, 'pad':1, 'kernel_size':3, 'stride':1, 'weight_filler':{'type':'msra'}, 'bias_filler':{'type':'constant', 'value':0.0}})
 74 |     net.tops[prefix+'_mbox_conf_perm'] = caffe.layers.Permute(net.tops[prefix+'_mbox_conf'], permute_param={'order':[0,2,3,1]})
 75 |     net.tops[prefix+'_mbox_conf_flat'] = caffe.layers.Flatten(net.tops[prefix+'_mbox_conf_perm'], flatten_param={'axis':1})
 76 | 
 77 |     net.tops[prefix+'_mbox_priorbox'] = caffe.layers.PriorBox(net_bottom, net.data, prior_box_param={'min_size':min_size, 'max_size':max_size, 'aspect_ratio':aspect_ratio, 'flip':True,
 78 |         'clip':False, 'variance':[0.1, 0.1, 0.2, 0.2], 'step':step})
 79 | 
 80 | def generate_net(lmdb, label_file, PHASE, batch_size):
 81 |     net = caffe.NetSpec()
 82 | 
 83 |     if(PHASE=="TRAIN"):
 84 |         # data layer
 85 |         net.data, net.label = caffe.layers.AnnotatedData(ntop=2, include={'phase':caffe.TRAIN}, 
 86 |                 transform_param=dict(mirror=True, mean_value=[104, 117, 123],
 87 |                     resize_param=dict(prob=1.0, resize_mode=caffe.params.Resize.WARP, height=300, width=300, 
 88 |                     interp_mode=[caffe.params.Resize.LINEAR,caffe.params.Resize.AREA,caffe.params.Resize.NEAREST,caffe.params.Resize.CUBIC,caffe.params.Resize.LANCZOS4]),
 89 |                     emit_constraint=dict(emit_type=0), 
 90 |                     distort_param=dict(brightness_prob=0.5, brightness_delta=32.0,
 91 |                             contrast_prob=0.5, contrast_lower=0.5, contrast_upper=1.5, hue_prob=0.5, hue_delta=18.0,
 92 |                             saturation_prob=0.5, saturation_lower=0.5, saturation_upper=1.5, random_order_prob=0.0),
 93 |                     expand_param=dict(prob=0.5, max_expand_ratio=4.0)),
 94 | 
 95 |                 data_param=dict(source=lmdb, batch_size=batch_size, backend=caffe.params.Data.LMDB), 
 96 |                
 97 |                 annotated_data_param=dict(
 98 |                 batch_sampler=[dict(max_sample=1, max_trials=1),
 99 |                 dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0), 
100 |                     sample_constraint=dict(min_jaccard_overlap=0.1), max_sample=1, max_trials=50),
101 |                 dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0), 
102 |                     sample_constraint=dict(min_jaccard_overlap=0.3), max_sample=1, max_trials=50),
103 |                 dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0), 
104 |                     sample_constraint=dict(min_jaccard_overlap=0.5), max_sample=1, max_trials=50),
105 |                 dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0), 
106 |                     sample_constraint=dict(min_jaccard_overlap=0.7), max_sample=1, max_trials=50),
107 |                 dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0), 
108 |                     sample_constraint=dict(min_jaccard_overlap=0.9), max_sample=1, max_trials=50),
109 |                 dict(sampler=dict(min_scale=0.3, max_scale=1.0,min_aspect_ratio=0.5, max_aspect_ratio=2.0), 
110 |                     sample_constraint=dict(min_jaccard_overlap=1.0), max_sample=1, max_trials=50)],
111 |                 label_map_file=label_file))
112 |     elif(PHASE=="DEPLOY"):
113 |         net.data = caffe.layers.Input(shape={'dim':[1,3,300,300]})
114 |    
115 |     # bone
116 |     net.conv1 = caffe.layers.Convolution(net.data, num_output=57, kernel_size=3, stride=2, weight_filler={"type":"xavier"},
117 |             param=[dict(lr_mult=1.0,decay_mult=0.0),dict(lr_mult=1.0,decay_mult=0.0)])
118 |     net.relu_conv1 = caffe.layers.ReLU(net.conv1,in_place=True)
119 |     net.pool1 = caffe.layers.Pooling(net.relu_conv1, pool=caffe.params.Pooling.MAX, kernel_size=3, stride=2)
120 |    
121 |     # fire1
122 |     net.tops['fire2/concat'] = fire(net, net.pool1, 'fire2', 15,49, 53)
123 |     # fire2
124 |     net.tops['fire3/concat'] = fire(net, net.tops['fire2/concat'], 'fire3', 15, 54, 52)
125 |     net.pool3 = caffe.layers.Pooling(net.tops['fire3/concat'], pool=caffe.params.Pooling.MAX, kernel_size=3, stride=2)
126 |     # fire3
127 |     net.tops['fire4/concat'] = fire(net, net.pool3, 'fire4', 29, 92, 94)
128 |     # fire4
129 |     net.tops['fire5/concat'] = fire(net, net.tops['fire4/concat'], 'fire5', 29, 90, 83)
130 |     net.pool5 = caffe.layers.Pooling(net.tops['fire5/concat'], pool=caffe.params.Pooling.MAX, kernel_size=3, stride=2)
131 |     # fire5
132 |     net.tops['fire6/concat'] = fire(net, net.pool5, 'fire6', 44, 166, 161)
133 |     # fire6
134 |     net.tops['fire7/concat'] = fire(net, net.tops['fire6/concat'], 'fire7', 45, 155, 146)
135 |     # fire7
136 |     net.tops['fire8/concat'] = fire(net, net.tops['fire7/concat'], 'fire8', 49, 163, 171)
137 |     # fire8
138 |     net.tops['fire9/concat'] = fire(net, net.tops['fire8/concat'], 'fire9', 25, 29, 54)
139 |     net.pool9 = caffe.layers.Pooling(net.tops['fire9/concat'], pool=caffe.params.Pooling.MAX, kernel_size=3, stride=2)
140 |     # fire9
141 |     net.tops['fire10/concat'] = fire(net, net.pool9, 'fire10', 37, 45, 56)
142 |     net.pool10 = caffe.layers.Pooling(net.tops['fire10/concat'], pool=caffe.params.Pooling.MAX, kernel_size=3, stride=2)
143 |     # fire10
144 |     net.tops['fire11/concat'] = fire(net, net.pool10, 'fire11', 38, 41, 44)
145 | 
146 |     # conv12
147 |     net.conv12_1 = caffe.layers.Convolution(net.tops['fire11/concat'], param=[dict(lr_mult=1.0, decay_mult=1.0)],
148 |             convolution_param={'num_output':51, 'bias_term':False, 'kernel_size':1, 'weight_filler':{'type':'msra'}})
149 |     net.tops['conv12_1/bn'] = caffe.layers.BatchNorm(net.conv12_1, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
150 |     net.tops['conv12_1/scale'] = caffe.layers.Scale(net.tops['conv12_1/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=2.0, decay_mult=0.0)], 
151 |             scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
152 |     net.tops['conv12_1/relu'] = caffe.layers.ReLU(net.tops['conv12_1/scale'], in_place=True)
153 |     net.conv12_2 = caffe.layers.Convolution(net.tops['conv12_1/relu'], param=[dict(lr_mult=1.0, decay_mult=1.0)],
154 |             convolution_param={'num_output':46, 'bias_term':False, 'pad':1, 'kernel_size':3, 'stride':2, 'weight_filler':{'type':'msra'}})
155 |     net.tops['conv12_2/bn'] = caffe.layers.BatchNorm(net.conv12_2, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
156 |     net.tops['conv12_2/scale'] = caffe.layers.Scale(net.tops['conv12_2/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=2.0, decay_mult=0.0)], 
157 |             scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
158 |     net.tops['conv12_2/relu'] = caffe.layers.ReLU(net.tops['conv12_2/scale'], in_place=True)
159 | 
160 |     # conv13
161 |     net.conv13_1 = caffe.layers.Convolution(net.tops['conv12_2/relu'], param=[dict(lr_mult=1.0, decay_mult=1.0)],
162 |             convolution_param={'num_output':55, 'bias_term':False, 'kernel_size':1, 'weight_filler':{'type':'msra'}})
163 |     net.tops['conv13_1/bn'] = caffe.layers.BatchNorm(net.conv13_1, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
164 |     net.tops['conv13_1/scale'] = caffe.layers.Scale(net.tops['conv13_1/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=2.0, decay_mult=0.0)], 
165 |             scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
166 |     net.tops['conv13_1/relu'] = caffe.layers.ReLU(net.conv13_1, in_place=True)
167 |     net.conv13_2 = caffe.layers.Convolution(net.tops['conv13_1/relu'], param=[dict(lr_mult=1.0, decay_mult=1.0)],
168 |             convolution_param={'num_output':85, 'bias_term':False, 'pad':1, 'kernel_size':3, 'stride':2, 'weight_filler':{'type':'msra'}})
169 |     net.tops['conv13_2/bn'] = caffe.layers.BatchNorm(net.conv13_2, param=[dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)], in_place=True)
170 |     net.tops['conv13_2/scale'] = caffe.layers.Scale(net.tops['conv13_2/bn'], param=[dict(lr_mult=1.0, decay_mult=0.0), dict(lr_mult=2.0, decay_mult=0.0)], 
171 |             scale_param={'filler':{'value':1}, 'bias_term':True, 'bias_filler':{'value':0}}, in_place=True)
172 |     net.tops['conv13_2/relu'] = caffe.layers.ReLU(net.tops['conv13_2/scale'], in_place=True)
173 | 
174 | 
175 |     # fire 5 prior box
176 |     prior_box(net, net.tops['fire5/concat'], 'fire5', 16, 84, 21.0, 45.0, [2.0], 8)
177 |     # fire 9 prior box
178 |     prior_box(net, net.tops['fire9/concat'], 'fire9', 24, 126, 45.0, 99.0, [2.0, 3.0], 16)
179 |     # fire 10 prior box
180 |     prior_box(net, net.tops['fire10/concat'], 'fire10', 24, 126, 99.0, 153.0, [2.0, 3.0], 32)
181 |     # fire 11 prior box
182 |     prior_box(net, net.tops['fire11/concat'], 'fire11', 24, 126, 153.0, 207.0, [2.0, 3.0], 64)
183 |     # conv12_2 prior box
184 |     prior_box(net, net.tops['conv12_2'], 'conv12_2', 24, 126, 207.0, 261.0, [2.0, 3.0], 100)
185 |     # conv13_2 prior box
186 |     prior_box(net, net.tops['conv13_2'], 'conv13_2', 16, 84, 261.0, 315.0, [2.0], 300)
187 | 
188 |     # last process
189 |     net.tops['mbox_loc'] = caffe.layers.Concat(net.tops['fire5_mbox_loc_flat'], net.tops['fire9_mbox_loc_flat'], net.tops['fire10_mbox_loc_flat'], net.tops['fire11_mbox_loc_flat'],
190 |             net.tops['conv12_2_mbox_loc_flat'], net.tops['conv13_2_mbox_loc_flat'], concat_param={'axis':1})
191 |     net.tops['mbox_conf'] = caffe.layers.Concat(net.tops['fire5_mbox_conf_flat'], net.tops['fire9_mbox_conf_flat'], net.tops['fire10_mbox_conf_flat'], net.tops['fire11_mbox_conf_flat'],
192 |             net.tops['conv12_2_mbox_conf_flat'], net.tops['conv13_2_mbox_conf_flat'], concat_param={'axis':1})
193 |     net.tops['mbox_priorbox'] = caffe.layers.Concat(net.tops['fire5_mbox_priorbox'], net.tops['fire9_mbox_priorbox'], net.tops['fire10_mbox_priorbox'], net.tops['fire11_mbox_priorbox'],
194 |             net.tops['conv12_2_mbox_priorbox'], net.tops['conv13_2_mbox_priorbox'], concat_param={'axis':2})
195 |     if(PHASE=='TRAIN'):
196 |         net.tops['mbox_loss'] = caffe.layers.MultiBoxLoss(net.tops['mbox_loc'], net.tops['mbox_conf'], net.tops['mbox_priorbox'], net.label, include={'phase':caffe.TRAIN},
197 |                 propagate_down=[True, True, False, False], loss_param={'normalization':caffe.params.Loss.VALID}, multibox_loss_param={'loc_loss_type':caffe.params.MultiBoxLoss.SMOOTH_L1, 
198 |                     'conf_loss_type':caffe.params.MultiBoxLoss.SOFTMAX, 'loc_weight':1.0, 
199 |                     'num_classes':21, 'share_location':True, 'match_type':caffe.params.MultiBoxLoss.PER_PREDICTION, 'overlap_threshold':0.5, 'use_prior_for_matching':True, 
200 |                     'background_label_id':0, 'use_difficult_gt':True, 'neg_pos_ratio':3.0, 'neg_overlap':0.5, 
201 |                     'code_type':caffe.params.PriorBox.CENTER_SIZE, 'ignore_cross_boundary_bbox':False, 'mining_type':caffe.params.MultiBoxLoss.MAX_NEGATIVE})
202 |     elif(PHASE=='DEPLOY'):
203 |         net.tops['mbox_conf_reshape'] = caffe.layers.Reshape(net.tops['mbox_conf'], reshape_param={'shape':{'dim':[0,-1,21]}})
204 |         net.tops['mbox_conf_softmax'] = caffe.layers.Softmax(net.tops['mbox_conf_reshape'], softmax_param={'axis':2})
205 |         net.tops['mbox_conf_flatten'] = caffe.layers.Flatten(net.tops['mbox_conf_softmax'], flatten_param={'axis':1})
206 |         net.tops['detection_out'] = caffe.layers.DetectionOutput(net.tops['mbox_loc'], net.tops['mbox_conf_flatten'], net.tops['mbox_priorbox'], include={
207 |             'phase':caffe.TEST}, detection_output_param={'num_classes':21, 'share_location':True, 'background_label_id':0, 
208 |             'nms_param':{'nms_threshold':0.45, 'top_k':100}, 'code_type':caffe.params.PriorBox.CENTER_SIZE, 'keep_top_k':100, 'confidence_threshold':0.25})
209 | 
210 |     return str(net.to_proto())
211 | 
212 | def write_net(train_proto, train_lmdb, deploy_proto, label_file):
213 |     with open(train_proto, 'w') as f:
214 |         f.write(str(generate_net(train_lmdb, label_file, "TRAIN", batch_size=64)))
215 |         f.close()
216 |     with open(deploy_proto, 'w') as f:
217 |         f.write(str(generate_net(train_lmdb, label_file, "DEPLOY", batch_size=64)))
218 |         f.close()
219 | 
220 | 
221 | if __name__ == '__main__':
222 |     project_dir = "./" # add your directory absolute path here
223 | 
224 |     label_file = project_dir + "data/VOC0712/labelmap_voc.prototxt"
225 |     train_lmdb = project_dir + "-- your lmdb path -- /trian.lmdb" # your lmdb path
226 |     train_proto = project_dir + "python_script/train.prototxt"
227 |     deploy_proto = project_dir + "python_script/deploy.prototxt"
228 | 
229 |     write_net(train_proto, train_lmdb, deploy_proto, label_file)
230 | 


--------------------------------------------------------------------------------
/python_script/train.prototxt:
--------------------------------------------------------------------------------
   1 | layer {
   2 |   name: "data"
   3 |   type: "AnnotatedData"
   4 |   top: "data"
   5 |   top: "label"
   6 |   include {
   7 |     phase: TRAIN
   8 |   }
   9 |   transform_param {
  10 |     mirror: true
  11 |     mean_value: 104.0
  12 |     mean_value: 117.0
  13 |     mean_value: 123.0
  14 |     resize_param {
  15 |       prob: 1.0
  16 |       resize_mode: WARP
  17 |       height: 300
  18 |       width: 300
  19 |       interp_mode: LINEAR
  20 |       interp_mode: AREA
  21 |       interp_mode: NEAREST
  22 |       interp_mode: CUBIC
  23 |       interp_mode: LANCZOS4
  24 |     }
  25 |     emit_constraint {
  26 |       emit_type: CENTER
  27 |     }
  28 |     distort_param {
  29 |       brightness_prob: 0.5
  30 |       brightness_delta: 32.0
  31 |       contrast_prob: 0.5
  32 |       contrast_lower: 0.5
  33 |       contrast_upper: 1.5
  34 |       hue_prob: 0.5
  35 |       hue_delta: 18.0
  36 |       saturation_prob: 0.5
  37 |       saturation_lower: 0.5
  38 |       saturation_upper: 1.5
  39 |       random_order_prob: 0.0
  40 |     }
  41 |     expand_param {
  42 |       prob: 0.5
  43 |       max_expand_ratio: 4.0
  44 |     }
  45 |   }
  46 |   data_param {
  47 |     source: "./-- your lmdb path -- /trian.lmdb"
  48 |     batch_size: 64
  49 |     backend: LMDB
  50 |   }
  51 |   annotated_data_param {
  52 |     batch_sampler {
  53 |       max_sample: 1
  54 |       max_trials: 1
  55 |     }
  56 |     batch_sampler {
  57 |       sampler {
  58 |         min_scale: 0.300000011921
  59 |         max_scale: 1.0
  60 |         min_aspect_ratio: 0.5
  61 |         max_aspect_ratio: 2.0
  62 |       }
  63 |       sample_constraint {
  64 |         min_jaccard_overlap: 0.10000000149
  65 |       }
  66 |       max_sample: 1
  67 |       max_trials: 50
  68 |     }
  69 |     batch_sampler {
  70 |       sampler {
  71 |         min_scale: 0.300000011921
  72 |         max_scale: 1.0
  73 |         min_aspect_ratio: 0.5
  74 |         max_aspect_ratio: 2.0
  75 |       }
  76 |       sample_constraint {
  77 |         min_jaccard_overlap: 0.300000011921
  78 |       }
  79 |       max_sample: 1
  80 |       max_trials: 50
  81 |     }
  82 |     batch_sampler {
  83 |       sampler {
  84 |         min_scale: 0.300000011921
  85 |         max_scale: 1.0
  86 |         min_aspect_ratio: 0.5
  87 |         max_aspect_ratio: 2.0
  88 |       }
  89 |       sample_constraint {
  90 |         min_jaccard_overlap: 0.5
  91 |       }
  92 |       max_sample: 1
  93 |       max_trials: 50
  94 |     }
  95 |     batch_sampler {
  96 |       sampler {
  97 |         min_scale: 0.300000011921
  98 |         max_scale: 1.0
  99 |         min_aspect_ratio: 0.5
 100 |         max_aspect_ratio: 2.0
 101 |       }
 102 |       sample_constraint {
 103 |         min_jaccard_overlap: 0.699999988079
 104 |       }
 105 |       max_sample: 1
 106 |       max_trials: 50
 107 |     }
 108 |     batch_sampler {
 109 |       sampler {
 110 |         min_scale: 0.300000011921
 111 |         max_scale: 1.0
 112 |         min_aspect_ratio: 0.5
 113 |         max_aspect_ratio: 2.0
 114 |       }
 115 |       sample_constraint {
 116 |         min_jaccard_overlap: 0.899999976158
 117 |       }
 118 |       max_sample: 1
 119 |       max_trials: 50
 120 |     }
 121 |     batch_sampler {
 122 |       sampler {
 123 |         min_scale: 0.300000011921
 124 |         max_scale: 1.0
 125 |         min_aspect_ratio: 0.5
 126 |         max_aspect_ratio: 2.0
 127 |       }
 128 |       sample_constraint {
 129 |         min_jaccard_overlap: 1.0
 130 |       }
 131 |       max_sample: 1
 132 |       max_trials: 50
 133 |     }
 134 |     label_map_file: "./data/VOC0712/labelmap_voc.prototxt"
 135 |   }
 136 | }
 137 | layer {
 138 |   name: "conv1"
 139 |   type: "Convolution"
 140 |   bottom: "data"
 141 |   top: "conv1"
 142 |   param {
 143 |     lr_mult: 1.0
 144 |     decay_mult: 0.0
 145 |   }
 146 |   param {
 147 |     lr_mult: 1.0
 148 |     decay_mult: 0.0
 149 |   }
 150 |   convolution_param {
 151 |     num_output: 57
 152 |     kernel_size: 3
 153 |     stride: 2
 154 |     weight_filler {
 155 |       type: "xavier"
 156 |     }
 157 |   }
 158 | }
 159 | layer {
 160 |   name: "relu_conv1"
 161 |   type: "ReLU"
 162 |   bottom: "conv1"
 163 |   top: "conv1"
 164 | }
 165 | layer {
 166 |   name: "pool1"
 167 |   type: "Pooling"
 168 |   bottom: "conv1"
 169 |   top: "pool1"
 170 |   pooling_param {
 171 |     pool: MAX
 172 |     kernel_size: 3
 173 |     stride: 2
 174 |   }
 175 | }
 176 | layer {
 177 |   name: "fire2/squeeze1x1"
 178 |   type: "Convolution"
 179 |   bottom: "pool1"
 180 |   top: "fire2/squeeze1x1"
 181 |   param {
 182 |     lr_mult: 1.0
 183 |     decay_mult: 0.0
 184 |   }
 185 |   param {
 186 |     lr_mult: 1.0
 187 |     decay_mult: 0.0
 188 |   }
 189 |   convolution_param {
 190 |     num_output: 15
 191 |     kernel_size: 1
 192 |     weight_filler {
 193 |       type: "xavier"
 194 |     }
 195 |   }
 196 | }
 197 | layer {
 198 |   name: "fire2/expand1x1"
 199 |   type: "Convolution"
 200 |   bottom: "fire2/squeeze1x1"
 201 |   top: "fire2/expand1x1"
 202 |   param {
 203 |     lr_mult: 1.0
 204 |     decay_mult: 0.0
 205 |   }
 206 |   param {
 207 |     lr_mult: 1.0
 208 |     decay_mult: 0.0
 209 |   }
 210 |   convolution_param {
 211 |     num_output: 49
 212 |     kernel_size: 1
 213 |     weight_filler {
 214 |       type: "xavier"
 215 |     }
 216 |   }
 217 | }
 218 | layer {
 219 |   name: "fire2/relu_expand1x1"
 220 |   type: "ReLU"
 221 |   bottom: "fire2/expand1x1"
 222 |   top: "fire2/expand1x1"
 223 | }
 224 | layer {
 225 |   name: "fire2/expand3x3"
 226 |   type: "Convolution"
 227 |   bottom: "fire2/squeeze1x1"
 228 |   top: "fire2/expand3x3"
 229 |   param {
 230 |     lr_mult: 1.0
 231 |     decay_mult: 0.0
 232 |   }
 233 |   param {
 234 |     lr_mult: 1.0
 235 |     decay_mult: 0.0
 236 |   }
 237 |   convolution_param {
 238 |     num_output: 53
 239 |     pad: 1
 240 |     kernel_size: 3
 241 |     weight_filler {
 242 |       type: "xavier"
 243 |     }
 244 |   }
 245 | }
 246 | layer {
 247 |   name: "fire2/relu_expand3x3"
 248 |   type: "ReLU"
 249 |   bottom: "fire2/expand3x3"
 250 |   top: "fire2/expand3x3"
 251 | }
 252 | layer {
 253 |   name: "fire2/concat"
 254 |   type: "Concat"
 255 |   bottom: "fire2/expand1x1"
 256 |   bottom: "fire2/expand3x3"
 257 |   top: "fire2/concat"
 258 | }
 259 | layer {
 260 |   name: "fire3/squeeze1x1"
 261 |   type: "Convolution"
 262 |   bottom: "fire2/concat"
 263 |   top: "fire3/squeeze1x1"
 264 |   param {
 265 |     lr_mult: 1.0
 266 |     decay_mult: 0.0
 267 |   }
 268 |   param {
 269 |     lr_mult: 1.0
 270 |     decay_mult: 0.0
 271 |   }
 272 |   convolution_param {
 273 |     num_output: 15
 274 |     kernel_size: 1
 275 |     weight_filler {
 276 |       type: "xavier"
 277 |     }
 278 |   }
 279 | }
 280 | layer {
 281 |   name: "fire3/expand1x1"
 282 |   type: "Convolution"
 283 |   bottom: "fire3/squeeze1x1"
 284 |   top: "fire3/expand1x1"
 285 |   param {
 286 |     lr_mult: 1.0
 287 |     decay_mult: 0.0
 288 |   }
 289 |   param {
 290 |     lr_mult: 1.0
 291 |     decay_mult: 0.0
 292 |   }
 293 |   convolution_param {
 294 |     num_output: 54
 295 |     kernel_size: 1
 296 |     weight_filler {
 297 |       type: "xavier"
 298 |     }
 299 |   }
 300 | }
 301 | layer {
 302 |   name: "fire3/relu_expand1x1"
 303 |   type: "ReLU"
 304 |   bottom: "fire3/expand1x1"
 305 |   top: "fire3/expand1x1"
 306 | }
 307 | layer {
 308 |   name: "fire3/expand3x3"
 309 |   type: "Convolution"
 310 |   bottom: "fire3/squeeze1x1"
 311 |   top: "fire3/expand3x3"
 312 |   param {
 313 |     lr_mult: 1.0
 314 |     decay_mult: 0.0
 315 |   }
 316 |   param {
 317 |     lr_mult: 1.0
 318 |     decay_mult: 0.0
 319 |   }
 320 |   convolution_param {
 321 |     num_output: 52
 322 |     pad: 1
 323 |     kernel_size: 3
 324 |     weight_filler {
 325 |       type: "xavier"
 326 |     }
 327 |   }
 328 | }
 329 | layer {
 330 |   name: "fire3/relu_expand3x3"
 331 |   type: "ReLU"
 332 |   bottom: "fire3/expand3x3"
 333 |   top: "fire3/expand3x3"
 334 | }
 335 | layer {
 336 |   name: "fire3/concat"
 337 |   type: "Concat"
 338 |   bottom: "fire3/expand1x1"
 339 |   bottom: "fire3/expand3x3"
 340 |   top: "fire3/concat"
 341 | }
 342 | layer {
 343 |   name: "pool3"
 344 |   type: "Pooling"
 345 |   bottom: "fire3/concat"
 346 |   top: "pool3"
 347 |   pooling_param {
 348 |     pool: MAX
 349 |     kernel_size: 3
 350 |     stride: 2
 351 |   }
 352 | }
 353 | layer {
 354 |   name: "fire4/squeeze1x1"
 355 |   type: "Convolution"
 356 |   bottom: "pool3"
 357 |   top: "fire4/squeeze1x1"
 358 |   param {
 359 |     lr_mult: 1.0
 360 |     decay_mult: 0.0
 361 |   }
 362 |   param {
 363 |     lr_mult: 1.0
 364 |     decay_mult: 0.0
 365 |   }
 366 |   convolution_param {
 367 |     num_output: 29
 368 |     kernel_size: 1
 369 |     weight_filler {
 370 |       type: "xavier"
 371 |     }
 372 |   }
 373 | }
 374 | layer {
 375 |   name: "fire4/expand1x1"
 376 |   type: "Convolution"
 377 |   bottom: "fire4/squeeze1x1"
 378 |   top: "fire4/expand1x1"
 379 |   param {
 380 |     lr_mult: 1.0
 381 |     decay_mult: 0.0
 382 |   }
 383 |   param {
 384 |     lr_mult: 1.0
 385 |     decay_mult: 0.0
 386 |   }
 387 |   convolution_param {
 388 |     num_output: 92
 389 |     kernel_size: 1
 390 |     weight_filler {
 391 |       type: "xavier"
 392 |     }
 393 |   }
 394 | }
 395 | layer {
 396 |   name: "fire4/relu_expand1x1"
 397 |   type: "ReLU"
 398 |   bottom: "fire4/expand1x1"
 399 |   top: "fire4/expand1x1"
 400 | }
 401 | layer {
 402 |   name: "fire4/expand3x3"
 403 |   type: "Convolution"
 404 |   bottom: "fire4/squeeze1x1"
 405 |   top: "fire4/expand3x3"
 406 |   param {
 407 |     lr_mult: 1.0
 408 |     decay_mult: 0.0
 409 |   }
 410 |   param {
 411 |     lr_mult: 1.0
 412 |     decay_mult: 0.0
 413 |   }
 414 |   convolution_param {
 415 |     num_output: 94
 416 |     pad: 1
 417 |     kernel_size: 3
 418 |     weight_filler {
 419 |       type: "xavier"
 420 |     }
 421 |   }
 422 | }
 423 | layer {
 424 |   name: "fire4/relu_expand3x3"
 425 |   type: "ReLU"
 426 |   bottom: "fire4/expand3x3"
 427 |   top: "fire4/expand3x3"
 428 | }
 429 | layer {
 430 |   name: "fire4/concat"
 431 |   type: "Concat"
 432 |   bottom: "fire4/expand1x1"
 433 |   bottom: "fire4/expand3x3"
 434 |   top: "fire4/concat"
 435 | }
 436 | layer {
 437 |   name: "fire5/squeeze1x1"
 438 |   type: "Convolution"
 439 |   bottom: "fire4/concat"
 440 |   top: "fire5/squeeze1x1"
 441 |   param {
 442 |     lr_mult: 1.0
 443 |     decay_mult: 0.0
 444 |   }
 445 |   param {
 446 |     lr_mult: 1.0
 447 |     decay_mult: 0.0
 448 |   }
 449 |   convolution_param {
 450 |     num_output: 29
 451 |     kernel_size: 1
 452 |     weight_filler {
 453 |       type: "xavier"
 454 |     }
 455 |   }
 456 | }
 457 | layer {
 458 |   name: "fire5/expand1x1"
 459 |   type: "Convolution"
 460 |   bottom: "fire5/squeeze1x1"
 461 |   top: "fire5/expand1x1"
 462 |   param {
 463 |     lr_mult: 1.0
 464 |     decay_mult: 0.0
 465 |   }
 466 |   param {
 467 |     lr_mult: 1.0
 468 |     decay_mult: 0.0
 469 |   }
 470 |   convolution_param {
 471 |     num_output: 90
 472 |     kernel_size: 1
 473 |     weight_filler {
 474 |       type: "xavier"
 475 |     }
 476 |   }
 477 | }
 478 | layer {
 479 |   name: "fire5/relu_expand1x1"
 480 |   type: "ReLU"
 481 |   bottom: "fire5/expand1x1"
 482 |   top: "fire5/expand1x1"
 483 | }
 484 | layer {
 485 |   name: "fire5/expand3x3"
 486 |   type: "Convolution"
 487 |   bottom: "fire5/squeeze1x1"
 488 |   top: "fire5/expand3x3"
 489 |   param {
 490 |     lr_mult: 1.0
 491 |     decay_mult: 0.0
 492 |   }
 493 |   param {
 494 |     lr_mult: 1.0
 495 |     decay_mult: 0.0
 496 |   }
 497 |   convolution_param {
 498 |     num_output: 83
 499 |     pad: 1
 500 |     kernel_size: 3
 501 |     weight_filler {
 502 |       type: "xavier"
 503 |     }
 504 |   }
 505 | }
 506 | layer {
 507 |   name: "fire5/relu_expand3x3"
 508 |   type: "ReLU"
 509 |   bottom: "fire5/expand3x3"
 510 |   top: "fire5/expand3x3"
 511 | }
 512 | layer {
 513 |   name: "fire5/concat"
 514 |   type: "Concat"
 515 |   bottom: "fire5/expand1x1"
 516 |   bottom: "fire5/expand3x3"
 517 |   top: "fire5/concat"
 518 | }
 519 | layer {
 520 |   name: "pool5"
 521 |   type: "Pooling"
 522 |   bottom: "fire5/concat"
 523 |   top: "pool5"
 524 |   pooling_param {
 525 |     pool: MAX
 526 |     kernel_size: 3
 527 |     stride: 2
 528 |   }
 529 | }
 530 | layer {
 531 |   name: "fire6/squeeze1x1"
 532 |   type: "Convolution"
 533 |   bottom: "pool5"
 534 |   top: "fire6/squeeze1x1"
 535 |   param {
 536 |     lr_mult: 1.0
 537 |     decay_mult: 0.0
 538 |   }
 539 |   param {
 540 |     lr_mult: 1.0
 541 |     decay_mult: 0.0
 542 |   }
 543 |   convolution_param {
 544 |     num_output: 44
 545 |     kernel_size: 1
 546 |     weight_filler {
 547 |       type: "xavier"
 548 |     }
 549 |   }
 550 | }
 551 | layer {
 552 |   name: "fire6/expand1x1"
 553 |   type: "Convolution"
 554 |   bottom: "fire6/squeeze1x1"
 555 |   top: "fire6/expand1x1"
 556 |   param {
 557 |     lr_mult: 1.0
 558 |     decay_mult: 0.0
 559 |   }
 560 |   param {
 561 |     lr_mult: 1.0
 562 |     decay_mult: 0.0
 563 |   }
 564 |   convolution_param {
 565 |     num_output: 166
 566 |     kernel_size: 1
 567 |     weight_filler {
 568 |       type: "xavier"
 569 |     }
 570 |   }
 571 | }
 572 | layer {
 573 |   name: "fire6/relu_expand1x1"
 574 |   type: "ReLU"
 575 |   bottom: "fire6/expand1x1"
 576 |   top: "fire6/expand1x1"
 577 | }
 578 | layer {
 579 |   name: "fire6/expand3x3"
 580 |   type: "Convolution"
 581 |   bottom: "fire6/squeeze1x1"
 582 |   top: "fire6/expand3x3"
 583 |   param {
 584 |     lr_mult: 1.0
 585 |     decay_mult: 0.0
 586 |   }
 587 |   param {
 588 |     lr_mult: 1.0
 589 |     decay_mult: 0.0
 590 |   }
 591 |   convolution_param {
 592 |     num_output: 161
 593 |     pad: 1
 594 |     kernel_size: 3
 595 |     weight_filler {
 596 |       type: "xavier"
 597 |     }
 598 |   }
 599 | }
 600 | layer {
 601 |   name: "fire6/relu_expand3x3"
 602 |   type: "ReLU"
 603 |   bottom: "fire6/expand3x3"
 604 |   top: "fire6/expand3x3"
 605 | }
 606 | layer {
 607 |   name: "fire6/concat"
 608 |   type: "Concat"
 609 |   bottom: "fire6/expand1x1"
 610 |   bottom: "fire6/expand3x3"
 611 |   top: "fire6/concat"
 612 | }
 613 | layer {
 614 |   name: "fire7/squeeze1x1"
 615 |   type: "Convolution"
 616 |   bottom: "fire6/concat"
 617 |   top: "fire7/squeeze1x1"
 618 |   param {
 619 |     lr_mult: 1.0
 620 |     decay_mult: 0.0
 621 |   }
 622 |   param {
 623 |     lr_mult: 1.0
 624 |     decay_mult: 0.0
 625 |   }
 626 |   convolution_param {
 627 |     num_output: 45
 628 |     kernel_size: 1
 629 |     weight_filler {
 630 |       type: "xavier"
 631 |     }
 632 |   }
 633 | }
 634 | layer {
 635 |   name: "fire7/expand1x1"
 636 |   type: "Convolution"
 637 |   bottom: "fire7/squeeze1x1"
 638 |   top: "fire7/expand1x1"
 639 |   param {
 640 |     lr_mult: 1.0
 641 |     decay_mult: 0.0
 642 |   }
 643 |   param {
 644 |     lr_mult: 1.0
 645 |     decay_mult: 0.0
 646 |   }
 647 |   convolution_param {
 648 |     num_output: 155
 649 |     kernel_size: 1
 650 |     weight_filler {
 651 |       type: "xavier"
 652 |     }
 653 |   }
 654 | }
 655 | layer {
 656 |   name: "fire7/relu_expand1x1"
 657 |   type: "ReLU"
 658 |   bottom: "fire7/expand1x1"
 659 |   top: "fire7/expand1x1"
 660 | }
 661 | layer {
 662 |   name: "fire7/expand3x3"
 663 |   type: "Convolution"
 664 |   bottom: "fire7/squeeze1x1"
 665 |   top: "fire7/expand3x3"
 666 |   param {
 667 |     lr_mult: 1.0
 668 |     decay_mult: 0.0
 669 |   }
 670 |   param {
 671 |     lr_mult: 1.0
 672 |     decay_mult: 0.0
 673 |   }
 674 |   convolution_param {
 675 |     num_output: 146
 676 |     pad: 1
 677 |     kernel_size: 3
 678 |     weight_filler {
 679 |       type: "xavier"
 680 |     }
 681 |   }
 682 | }
 683 | layer {
 684 |   name: "fire7/relu_expand3x3"
 685 |   type: "ReLU"
 686 |   bottom: "fire7/expand3x3"
 687 |   top: "fire7/expand3x3"
 688 | }
 689 | layer {
 690 |   name: "fire7/concat"
 691 |   type: "Concat"
 692 |   bottom: "fire7/expand1x1"
 693 |   bottom: "fire7/expand3x3"
 694 |   top: "fire7/concat"
 695 | }
 696 | layer {
 697 |   name: "fire8/squeeze1x1"
 698 |   type: "Convolution"
 699 |   bottom: "fire7/concat"
 700 |   top: "fire8/squeeze1x1"
 701 |   param {
 702 |     lr_mult: 1.0
 703 |     decay_mult: 0.0
 704 |   }
 705 |   param {
 706 |     lr_mult: 1.0
 707 |     decay_mult: 0.0
 708 |   }
 709 |   convolution_param {
 710 |     num_output: 49
 711 |     kernel_size: 1
 712 |     weight_filler {
 713 |       type: "xavier"
 714 |     }
 715 |   }
 716 | }
 717 | layer {
 718 |   name: "fire8/expand1x1"
 719 |   type: "Convolution"
 720 |   bottom: "fire8/squeeze1x1"
 721 |   top: "fire8/expand1x1"
 722 |   param {
 723 |     lr_mult: 1.0
 724 |     decay_mult: 0.0
 725 |   }
 726 |   param {
 727 |     lr_mult: 1.0
 728 |     decay_mult: 0.0
 729 |   }
 730 |   convolution_param {
 731 |     num_output: 163
 732 |     kernel_size: 1
 733 |     weight_filler {
 734 |       type: "xavier"
 735 |     }
 736 |   }
 737 | }
 738 | layer {
 739 |   name: "fire8/relu_expand1x1"
 740 |   type: "ReLU"
 741 |   bottom: "fire8/expand1x1"
 742 |   top: "fire8/expand1x1"
 743 | }
 744 | layer {
 745 |   name: "fire8/expand3x3"
 746 |   type: "Convolution"
 747 |   bottom: "fire8/squeeze1x1"
 748 |   top: "fire8/expand3x3"
 749 |   param {
 750 |     lr_mult: 1.0
 751 |     decay_mult: 0.0
 752 |   }
 753 |   param {
 754 |     lr_mult: 1.0
 755 |     decay_mult: 0.0
 756 |   }
 757 |   convolution_param {
 758 |     num_output: 171
 759 |     pad: 1
 760 |     kernel_size: 3
 761 |     weight_filler {
 762 |       type: "xavier"
 763 |     }
 764 |   }
 765 | }
 766 | layer {
 767 |   name: "fire8/relu_expand3x3"
 768 |   type: "ReLU"
 769 |   bottom: "fire8/expand3x3"
 770 |   top: "fire8/expand3x3"
 771 | }
 772 | layer {
 773 |   name: "fire8/concat"
 774 |   type: "Concat"
 775 |   bottom: "fire8/expand1x1"
 776 |   bottom: "fire8/expand3x3"
 777 |   top: "fire8/concat"
 778 | }
 779 | layer {
 780 |   name: "fire9/squeeze1x1"
 781 |   type: "Convolution"
 782 |   bottom: "fire8/concat"
 783 |   top: "fire9/squeeze1x1"
 784 |   param {
 785 |     lr_mult: 1.0
 786 |     decay_mult: 0.0
 787 |   }
 788 |   param {
 789 |     lr_mult: 1.0
 790 |     decay_mult: 0.0
 791 |   }
 792 |   convolution_param {
 793 |     num_output: 25
 794 |     kernel_size: 1
 795 |     weight_filler {
 796 |       type: "xavier"
 797 |     }
 798 |   }
 799 | }
 800 | layer {
 801 |   name: "fire9/expand1x1"
 802 |   type: "Convolution"
 803 |   bottom: "fire9/squeeze1x1"
 804 |   top: "fire9/expand1x1"
 805 |   param {
 806 |     lr_mult: 1.0
 807 |     decay_mult: 0.0
 808 |   }
 809 |   param {
 810 |     lr_mult: 1.0
 811 |     decay_mult: 0.0
 812 |   }
 813 |   convolution_param {
 814 |     num_output: 29
 815 |     kernel_size: 1
 816 |     weight_filler {
 817 |       type: "xavier"
 818 |     }
 819 |   }
 820 | }
 821 | layer {
 822 |   name: "fire9/expand1x1/bn"
 823 |   type: "BatchNorm"
 824 |   bottom: "fire9/expand1x1"
 825 |   top: "fire9/expand1x1"
 826 |   param {
 827 |     lr_mult: 0.0
 828 |     decay_mult: 0.0
 829 |   }
 830 |   param {
 831 |     lr_mult: 0.0
 832 |     decay_mult: 0.0
 833 |   }
 834 |   param {
 835 |     lr_mult: 0.0
 836 |     decay_mult: 0.0
 837 |   }
 838 | }
 839 | layer {
 840 |   name: "fire9/expand1x1/scale"
 841 |   type: "Scale"
 842 |   bottom: "fire9/expand1x1"
 843 |   top: "fire9/expand1x1"
 844 |   param {
 845 |     lr_mult: 1.0
 846 |     decay_mult: 0.0
 847 |   }
 848 |   param {
 849 |     lr_mult: 2.0
 850 |     decay_mult: 0.0
 851 |   }
 852 |   scale_param {
 853 |     filler {
 854 |       value: 1.0
 855 |     }
 856 |     bias_term: true
 857 |     bias_filler {
 858 |       value: 0.0
 859 |     }
 860 |   }
 861 | }
 862 | layer {
 863 |   name: "fire9/relu_expand1x1"
 864 |   type: "ReLU"
 865 |   bottom: "fire9/expand1x1"
 866 |   top: "fire9/expand1x1"
 867 | }
 868 | layer {
 869 |   name: "fire9/expand3x3"
 870 |   type: "Convolution"
 871 |   bottom: "fire9/squeeze1x1"
 872 |   top: "fire9/expand3x3"
 873 |   param {
 874 |     lr_mult: 1.0
 875 |     decay_mult: 0.0
 876 |   }
 877 |   param {
 878 |     lr_mult: 1.0
 879 |     decay_mult: 0.0
 880 |   }
 881 |   convolution_param {
 882 |     num_output: 54
 883 |     pad: 1
 884 |     kernel_size: 3
 885 |     weight_filler {
 886 |       type: "xavier"
 887 |     }
 888 |   }
 889 | }
 890 | layer {
 891 |   name: "fire9/expand3x3/bn"
 892 |   type: "BatchNorm"
 893 |   bottom: "fire9/expand3x3"
 894 |   top: "fire9/expand3x3"
 895 |   param {
 896 |     lr_mult: 0.0
 897 |     decay_mult: 0.0
 898 |   }
 899 |   param {
 900 |     lr_mult: 0.0
 901 |     decay_mult: 0.0
 902 |   }
 903 |   param {
 904 |     lr_mult: 0.0
 905 |     decay_mult: 0.0
 906 |   }
 907 | }
 908 | layer {
 909 |   name: "fire9/expand3x3/scale"
 910 |   type: "Scale"
 911 |   bottom: "fire9/expand3x3"
 912 |   top: "fire9/expand3x3"
 913 |   param {
 914 |     lr_mult: 1.0
 915 |     decay_mult: 0.0
 916 |   }
 917 |   param {
 918 |     lr_mult: 2.0
 919 |     decay_mult: 0.0
 920 |   }
 921 |   scale_param {
 922 |     filler {
 923 |       value: 1.0
 924 |     }
 925 |     bias_term: true
 926 |     bias_filler {
 927 |       value: 0.0
 928 |     }
 929 |   }
 930 | }
 931 | layer {
 932 |   name: "fire9/relu_expand3x3"
 933 |   type: "ReLU"
 934 |   bottom: "fire9/expand3x3"
 935 |   top: "fire9/expand3x3"
 936 | }
 937 | layer {
 938 |   name: "fire9/concat"
 939 |   type: "Concat"
 940 |   bottom: "fire9/expand1x1"
 941 |   bottom: "fire9/expand3x3"
 942 |   top: "fire9/concat"
 943 | }
 944 | layer {
 945 |   name: "pool9"
 946 |   type: "Pooling"
 947 |   bottom: "fire9/concat"
 948 |   top: "pool9"
 949 |   pooling_param {
 950 |     pool: MAX
 951 |     kernel_size: 3
 952 |     stride: 2
 953 |   }
 954 | }
 955 | layer {
 956 |   name: "fire10/squeeze1x1"
 957 |   type: "Convolution"
 958 |   bottom: "pool9"
 959 |   top: "fire10/squeeze1x1"
 960 |   convolution_param {
 961 |     num_output: 37
 962 |     kernel_size: 1
 963 |     weight_filler {
 964 |       type: "xavier"
 965 |     }
 966 |   }
 967 | }
 968 | layer {
 969 |   name: "fire10/squeeze1x1/bn"
 970 |   type: "BatchNorm"
 971 |   bottom: "fire10/squeeze1x1"
 972 |   top: "fire10/squeeze1x1"
 973 |   param {
 974 |     lr_mult: 0.0
 975 |     decay_mult: 0.0
 976 |   }
 977 |   param {
 978 |     lr_mult: 0.0
 979 |     decay_mult: 0.0
 980 |   }
 981 |   param {
 982 |     lr_mult: 0.0
 983 |     decay_mult: 0.0
 984 |   }
 985 | }
 986 | layer {
 987 |   name: "fire10/squeeze1x1/scale"
 988 |   type: "Scale"
 989 |   bottom: "fire10/squeeze1x1"
 990 |   top: "fire10/squeeze1x1"
 991 |   param {
 992 |     lr_mult: 1.0
 993 |     decay_mult: 0.0
 994 |   }
 995 |   param {
 996 |     lr_mult: 2.0
 997 |     decay_mult: 0.0
 998 |   }
 999 |   scale_param {
1000 |     filler {
1001 |       value: 1.0
1002 |     }
1003 |     bias_term: true
1004 |     bias_filler {
1005 |       value: 0.0
1006 |     }
1007 |   }
1008 | }
1009 | layer {
1010 |   name: "fire10/expand1x1"
1011 |   type: "Convolution"
1012 |   bottom: "fire10/squeeze1x1"
1013 |   top: "fire10/expand1x1"
1014 |   convolution_param {
1015 |     num_output: 45
1016 |     kernel_size: 1
1017 |     weight_filler {
1018 |       type: "xavier"
1019 |     }
1020 |   }
1021 | }
1022 | layer {
1023 |   name: "fire10/expand1x1/bn"
1024 |   type: "BatchNorm"
1025 |   bottom: "fire10/expand1x1"
1026 |   top: "fire10/expand1x1"
1027 |   param {
1028 |     lr_mult: 0.0
1029 |     decay_mult: 0.0
1030 |   }
1031 |   param {
1032 |     lr_mult: 0.0
1033 |     decay_mult: 0.0
1034 |   }
1035 |   param {
1036 |     lr_mult: 0.0
1037 |     decay_mult: 0.0
1038 |   }
1039 | }
1040 | layer {
1041 |   name: "fire10/expand1x1/scale"
1042 |   type: "Scale"
1043 |   bottom: "fire10/expand1x1"
1044 |   top: "fire10/expand1x1"
1045 |   param {
1046 |     lr_mult: 1.0
1047 |     decay_mult: 0.0
1048 |   }
1049 |   param {
1050 |     lr_mult: 2.0
1051 |     decay_mult: 0.0
1052 |   }
1053 |   scale_param {
1054 |     filler {
1055 |       value: 1.0
1056 |     }
1057 |     bias_term: true
1058 |     bias_filler {
1059 |       value: 0.0
1060 |     }
1061 |   }
1062 | }
1063 | layer {
1064 |   name: "fire10/relu_expand1x1"
1065 |   type: "ReLU"
1066 |   bottom: "fire10/expand1x1"
1067 |   top: "fire10/expand1x1"
1068 | }
1069 | layer {
1070 |   name: "fire10/expand3x3"
1071 |   type: "Convolution"
1072 |   bottom: "fire10/squeeze1x1"
1073 |   top: "fire10/expand3x3"
1074 |   convolution_param {
1075 |     num_output: 56
1076 |     pad: 1
1077 |     kernel_size: 3
1078 |     weight_filler {
1079 |       type: "xavier"
1080 |     }
1081 |   }
1082 | }
1083 | layer {
1084 |   name: "fire10/expand3x3/bn"
1085 |   type: "BatchNorm"
1086 |   bottom: "fire10/expand3x3"
1087 |   top: "fire10/expand3x3"
1088 |   param {
1089 |     lr_mult: 0.0
1090 |     decay_mult: 0.0
1091 |   }
1092 |   param {
1093 |     lr_mult: 0.0
1094 |     decay_mult: 0.0
1095 |   }
1096 |   param {
1097 |     lr_mult: 0.0
1098 |     decay_mult: 0.0
1099 |   }
1100 | }
1101 | layer {
1102 |   name: "fire10/expand3x3/scale"
1103 |   type: "Scale"
1104 |   bottom: "fire10/expand3x3"
1105 |   top: "fire10/expand3x3"
1106 |   param {
1107 |     lr_mult: 1.0
1108 |     decay_mult: 0.0
1109 |   }
1110 |   param {
1111 |     lr_mult: 2.0
1112 |     decay_mult: 0.0
1113 |   }
1114 |   scale_param {
1115 |     filler {
1116 |       value: 1.0
1117 |     }
1118 |     bias_term: true
1119 |     bias_filler {
1120 |       value: 0.0
1121 |     }
1122 |   }
1123 | }
1124 | layer {
1125 |   name: "fire10/relu_expand3x3"
1126 |   type: "ReLU"
1127 |   bottom: "fire10/expand3x3"
1128 |   top: "fire10/expand3x3"
1129 | }
1130 | layer {
1131 |   name: "fire10/concat"
1132 |   type: "Concat"
1133 |   bottom: "fire10/expand1x1"
1134 |   bottom: "fire10/expand3x3"
1135 |   top: "fire10/concat"
1136 | }
1137 | layer {
1138 |   name: "pool10"
1139 |   type: "Pooling"
1140 |   bottom: "fire10/concat"
1141 |   top: "pool10"
1142 |   pooling_param {
1143 |     pool: MAX
1144 |     kernel_size: 3
1145 |     stride: 2
1146 |   }
1147 | }
1148 | layer {
1149 |   name: "fire11/squeeze1x1"
1150 |   type: "Convolution"
1151 |   bottom: "pool10"
1152 |   top: "fire11/squeeze1x1"
1153 |   convolution_param {
1154 |     num_output: 38
1155 |     kernel_size: 1
1156 |     weight_filler {
1157 |       type: "xavier"
1158 |     }
1159 |   }
1160 | }
1161 | layer {
1162 |   name: "fire11/squeeze1x1/bn"
1163 |   type: "BatchNorm"
1164 |   bottom: "fire11/squeeze1x1"
1165 |   top: "fire11/squeeze1x1"
1166 |   param {
1167 |     lr_mult: 0.0
1168 |     decay_mult: 0.0
1169 |   }
1170 |   param {
1171 |     lr_mult: 0.0
1172 |     decay_mult: 0.0
1173 |   }
1174 |   param {
1175 |     lr_mult: 0.0
1176 |     decay_mult: 0.0
1177 |   }
1178 | }
1179 | layer {
1180 |   name: "fire11/squeeze1x1/scale"
1181 |   type: "Scale"
1182 |   bottom: "fire11/squeeze1x1"
1183 |   top: "fire11/squeeze1x1"
1184 |   param {
1185 |     lr_mult: 1.0
1186 |     decay_mult: 0.0
1187 |   }
1188 |   param {
1189 |     lr_mult: 2.0
1190 |     decay_mult: 0.0
1191 |   }
1192 |   scale_param {
1193 |     filler {
1194 |       value: 1.0
1195 |     }
1196 |     bias_term: true
1197 |     bias_filler {
1198 |       value: 0.0
1199 |     }
1200 |   }
1201 | }
1202 | layer {
1203 |   name: "fire11/expand1x1"
1204 |   type: "Convolution"
1205 |   bottom: "fire11/squeeze1x1"
1206 |   top: "fire11/expand1x1"
1207 |   convolution_param {
1208 |     num_output: 41
1209 |     kernel_size: 1
1210 |     weight_filler {
1211 |       type: "xavier"
1212 |     }
1213 |   }
1214 | }
1215 | layer {
1216 |   name: "fire11/expand1x1/bn"
1217 |   type: "BatchNorm"
1218 |   bottom: "fire11/expand1x1"
1219 |   top: "fire11/expand1x1"
1220 |   param {
1221 |     lr_mult: 0.0
1222 |     decay_mult: 0.0
1223 |   }
1224 |   param {
1225 |     lr_mult: 0.0
1226 |     decay_mult: 0.0
1227 |   }
1228 |   param {
1229 |     lr_mult: 0.0
1230 |     decay_mult: 0.0
1231 |   }
1232 | }
1233 | layer {
1234 |   name: "fire11/expand1x1/scale"
1235 |   type: "Scale"
1236 |   bottom: "fire11/expand1x1"
1237 |   top: "fire11/expand1x1"
1238 |   param {
1239 |     lr_mult: 1.0
1240 |     decay_mult: 0.0
1241 |   }
1242 |   param {
1243 |     lr_mult: 2.0
1244 |     decay_mult: 0.0
1245 |   }
1246 |   scale_param {
1247 |     filler {
1248 |       value: 1.0
1249 |     }
1250 |     bias_term: true
1251 |     bias_filler {
1252 |       value: 0.0
1253 |     }
1254 |   }
1255 | }
1256 | layer {
1257 |   name: "fire11/relu_expand1x1"
1258 |   type: "ReLU"
1259 |   bottom: "fire11/expand1x1"
1260 |   top: "fire11/expand1x1"
1261 | }
1262 | layer {
1263 |   name: "fire11/expand3x3"
1264 |   type: "Convolution"
1265 |   bottom: "fire11/squeeze1x1"
1266 |   top: "fire11/expand3x3"
1267 |   convolution_param {
1268 |     num_output: 44
1269 |     pad: 1
1270 |     kernel_size: 3
1271 |     weight_filler {
1272 |       type: "xavier"
1273 |     }
1274 |   }
1275 | }
1276 | layer {
1277 |   name: "fire11/expand3x3/bn"
1278 |   type: "BatchNorm"
1279 |   bottom: "fire11/expand3x3"
1280 |   top: "fire11/expand3x3"
1281 |   param {
1282 |     lr_mult: 0.0
1283 |     decay_mult: 0.0
1284 |   }
1285 |   param {
1286 |     lr_mult: 0.0
1287 |     decay_mult: 0.0
1288 |   }
1289 |   param {
1290 |     lr_mult: 0.0
1291 |     decay_mult: 0.0
1292 |   }
1293 | }
1294 | layer {
1295 |   name: "fire11/expand3x3/scale"
1296 |   type: "Scale"
1297 |   bottom: "fire11/expand3x3"
1298 |   top: "fire11/expand3x3"
1299 |   param {
1300 |     lr_mult: 1.0
1301 |     decay_mult: 0.0
1302 |   }
1303 |   param {
1304 |     lr_mult: 2.0
1305 |     decay_mult: 0.0
1306 |   }
1307 |   scale_param {
1308 |     filler {
1309 |       value: 1.0
1310 |     }
1311 |     bias_term: true
1312 |     bias_filler {
1313 |       value: 0.0
1314 |     }
1315 |   }
1316 | }
1317 | layer {
1318 |   name: "fire11/relu_expand3x3"
1319 |   type: "ReLU"
1320 |   bottom: "fire11/expand3x3"
1321 |   top: "fire11/expand3x3"
1322 | }
1323 | layer {
1324 |   name: "fire11/concat"
1325 |   type: "Concat"
1326 |   bottom: "fire11/expand1x1"
1327 |   bottom: "fire11/expand3x3"
1328 |   top: "fire11/concat"
1329 | }
1330 | layer {
1331 |   name: "conv12_1"
1332 |   type: "Convolution"
1333 |   bottom: "fire11/concat"
1334 |   top: "conv12_1"
1335 |   param {
1336 |     lr_mult: 1.0
1337 |     decay_mult: 1.0
1338 |   }
1339 |   convolution_param {
1340 |     num_output: 51
1341 |     bias_term: false
1342 |     kernel_size: 1
1343 |     weight_filler {
1344 |       type: "msra"
1345 |     }
1346 |   }
1347 | }
1348 | layer {
1349 |   name: "conv12_1/bn"
1350 |   type: "BatchNorm"
1351 |   bottom: "conv12_1"
1352 |   top: "conv12_1"
1353 |   param {
1354 |     lr_mult: 0.0
1355 |     decay_mult: 0.0
1356 |   }
1357 |   param {
1358 |     lr_mult: 0.0
1359 |     decay_mult: 0.0
1360 |   }
1361 |   param {
1362 |     lr_mult: 0.0
1363 |     decay_mult: 0.0
1364 |   }
1365 | }
1366 | layer {
1367 |   name: "conv12_1/scale"
1368 |   type: "Scale"
1369 |   bottom: "conv12_1"
1370 |   top: "conv12_1"
1371 |   param {
1372 |     lr_mult: 1.0
1373 |     decay_mult: 0.0
1374 |   }
1375 |   param {
1376 |     lr_mult: 2.0
1377 |     decay_mult: 0.0
1378 |   }
1379 |   scale_param {
1380 |     filler {
1381 |       value: 1.0
1382 |     }
1383 |     bias_term: true
1384 |     bias_filler {
1385 |       value: 0.0
1386 |     }
1387 |   }
1388 | }
1389 | layer {
1390 |   name: "conv12_1/relu"
1391 |   type: "ReLU"
1392 |   bottom: "conv12_1"
1393 |   top: "conv12_1"
1394 | }
1395 | layer {
1396 |   name: "conv12_2"
1397 |   type: "Convolution"
1398 |   bottom: "conv12_1"
1399 |   top: "conv12_2"
1400 |   param {
1401 |     lr_mult: 1.0
1402 |     decay_mult: 1.0
1403 |   }
1404 |   convolution_param {
1405 |     num_output: 46
1406 |     bias_term: false
1407 |     pad: 1
1408 |     kernel_size: 3
1409 |     stride: 2
1410 |     weight_filler {
1411 |       type: "msra"
1412 |     }
1413 |   }
1414 | }
1415 | layer {
1416 |   name: "conv12_2/bn"
1417 |   type: "BatchNorm"
1418 |   bottom: "conv12_2"
1419 |   top: "conv12_2"
1420 |   param {
1421 |     lr_mult: 0.0
1422 |     decay_mult: 0.0
1423 |   }
1424 |   param {
1425 |     lr_mult: 0.0
1426 |     decay_mult: 0.0
1427 |   }
1428 |   param {
1429 |     lr_mult: 0.0
1430 |     decay_mult: 0.0
1431 |   }
1432 | }
1433 | layer {
1434 |   name: "conv12_2/scale"
1435 |   type: "Scale"
1436 |   bottom: "conv12_2"
1437 |   top: "conv12_2"
1438 |   param {
1439 |     lr_mult: 1.0
1440 |     decay_mult: 0.0
1441 |   }
1442 |   param {
1443 |     lr_mult: 2.0
1444 |     decay_mult: 0.0
1445 |   }
1446 |   scale_param {
1447 |     filler {
1448 |       value: 1.0
1449 |     }
1450 |     bias_term: true
1451 |     bias_filler {
1452 |       value: 0.0
1453 |     }
1454 |   }
1455 | }
1456 | layer {
1457 |   name: "conv12_2/relu"
1458 |   type: "ReLU"
1459 |   bottom: "conv12_2"
1460 |   top: "conv12_2"
1461 | }
1462 | layer {
1463 |   name: "conv13_1"
1464 |   type: "Convolution"
1465 |   bottom: "conv12_2"
1466 |   top: "conv13_1"
1467 |   param {
1468 |     lr_mult: 1.0
1469 |     decay_mult: 1.0
1470 |   }
1471 |   convolution_param {
1472 |     num_output: 55
1473 |     bias_term: false
1474 |     kernel_size: 1
1475 |     weight_filler {
1476 |       type: "msra"
1477 |     }
1478 |   }
1479 | }
1480 | layer {
1481 |   name: "conv13_1/bn"
1482 |   type: "BatchNorm"
1483 |   bottom: "conv13_1"
1484 |   top: "conv13_1"
1485 |   param {
1486 |     lr_mult: 0.0
1487 |     decay_mult: 0.0
1488 |   }
1489 |   param {
1490 |     lr_mult: 0.0
1491 |     decay_mult: 0.0
1492 |   }
1493 |   param {
1494 |     lr_mult: 0.0
1495 |     decay_mult: 0.0
1496 |   }
1497 | }
1498 | layer {
1499 |   name: "conv13_1/scale"
1500 |   type: "Scale"
1501 |   bottom: "conv13_1"
1502 |   top: "conv13_1"
1503 |   param {
1504 |     lr_mult: 1.0
1505 |     decay_mult: 0.0
1506 |   }
1507 |   param {
1508 |     lr_mult: 2.0
1509 |     decay_mult: 0.0
1510 |   }
1511 |   scale_param {
1512 |     filler {
1513 |       value: 1.0
1514 |     }
1515 |     bias_term: true
1516 |     bias_filler {
1517 |       value: 0.0
1518 |     }
1519 |   }
1520 | }
1521 | layer {
1522 |   name: "conv13_1/relu"
1523 |   type: "ReLU"
1524 |   bottom: "conv13_1"
1525 |   top: "conv13_1"
1526 | }
1527 | layer {
1528 |   name: "conv13_2"
1529 |   type: "Convolution"
1530 |   bottom: "conv13_1"
1531 |   top: "conv13_2"
1532 |   param {
1533 |     lr_mult: 1.0
1534 |     decay_mult: 1.0
1535 |   }
1536 |   convolution_param {
1537 |     num_output: 85
1538 |     bias_term: false
1539 |     pad: 1
1540 |     kernel_size: 3
1541 |     stride: 2
1542 |     weight_filler {
1543 |       type: "msra"
1544 |     }
1545 |   }
1546 | }
1547 | layer {
1548 |   name: "conv13_2/bn"
1549 |   type: "BatchNorm"
1550 |   bottom: "conv13_2"
1551 |   top: "conv13_2"
1552 |   param {
1553 |     lr_mult: 0.0
1554 |     decay_mult: 0.0
1555 |   }
1556 |   param {
1557 |     lr_mult: 0.0
1558 |     decay_mult: 0.0
1559 |   }
1560 |   param {
1561 |     lr_mult: 0.0
1562 |     decay_mult: 0.0
1563 |   }
1564 | }
1565 | layer {
1566 |   name: "conv13_2/scale"
1567 |   type: "Scale"
1568 |   bottom: "conv13_2"
1569 |   top: "conv13_2"
1570 |   param {
1571 |     lr_mult: 1.0
1572 |     decay_mult: 0.0
1573 |   }
1574 |   param {
1575 |     lr_mult: 2.0
1576 |     decay_mult: 0.0
1577 |   }
1578 |   scale_param {
1579 |     filler {
1580 |       value: 1.0
1581 |     }
1582 |     bias_term: true
1583 |     bias_filler {
1584 |       value: 0.0
1585 |     }
1586 |   }
1587 | }
1588 | layer {
1589 |   name: "conv13_2/relu"
1590 |   type: "ReLU"
1591 |   bottom: "conv13_2"
1592 |   top: "conv13_2"
1593 | }
1594 | layer {
1595 |   name: "fire5/bn"
1596 |   type: "BatchNorm"
1597 |   bottom: "fire5/concat"
1598 |   top: "fire5/normal"
1599 |   param {
1600 |     lr_mult: 0.0
1601 |     decay_mult: 0.0
1602 |   }
1603 |   param {
1604 |     lr_mult: 0.0
1605 |     decay_mult: 0.0
1606 |   }
1607 |   param {
1608 |     lr_mult: 0.0
1609 |     decay_mult: 0.0
1610 |   }
1611 | }
1612 | layer {
1613 |   name: "fire5/scale"
1614 |   type: "Scale"
1615 |   bottom: "fire5/normal"
1616 |   top: "fire5/normal"
1617 |   param {
1618 |     lr_mult: 1.0
1619 |     decay_mult: 0.0
1620 |   }
1621 |   param {
1622 |     lr_mult: 2.0
1623 |     decay_mult: 0.0
1624 |   }
1625 |   scale_param {
1626 |     filler {
1627 |       value: 1.0
1628 |     }
1629 |     bias_term: true
1630 |     bias_filler {
1631 |       value: 0.0
1632 |     }
1633 |   }
1634 | }
1635 | layer {
1636 |   name: "fire5_mbox_loc"
1637 |   type: "Convolution"
1638 |   bottom: "fire5/normal"
1639 |   top: "fire5_mbox_loc"
1640 |   param {
1641 |     lr_mult: 1.0
1642 |     decay_mult: 1.0
1643 |   }
1644 |   param {
1645 |     lr_mult: 2.0
1646 |     decay_mult: 0.0
1647 |   }
1648 |   convolution_param {
1649 |     num_output: 16
1650 |     pad: 1
1651 |     kernel_size: 3
1652 |     stride: 1
1653 |     weight_filler {
1654 |       type: "msra"
1655 |     }
1656 |     bias_filler {
1657 |       type: "constant"
1658 |       value: 0.0
1659 |     }
1660 |   }
1661 | }
1662 | layer {
1663 |   name: "fire5_mbox_loc_perm"
1664 |   type: "Permute"
1665 |   bottom: "fire5_mbox_loc"
1666 |   top: "fire5_mbox_loc_perm"
1667 |   permute_param {
1668 |     order: 0
1669 |     order: 2
1670 |     order: 3
1671 |     order: 1
1672 |   }
1673 | }
1674 | layer {
1675 |   name: "fire5_mbox_loc_flat"
1676 |   type: "Flatten"
1677 |   bottom: "fire5_mbox_loc_perm"
1678 |   top: "fire5_mbox_loc_flat"
1679 |   flatten_param {
1680 |     axis: 1
1681 |   }
1682 | }
1683 | layer {
1684 |   name: "fire5_mbox_conf"
1685 |   type: "Convolution"
1686 |   bottom: "fire5/normal"
1687 |   top: "fire5_mbox_conf"
1688 |   param {
1689 |     lr_mult: 1.0
1690 |     decay_mult: 1.0
1691 |   }
1692 |   param {
1693 |     lr_mult: 2.0
1694 |     decay_mult: 0.0
1695 |   }
1696 |   convolution_param {
1697 |     num_output: 84
1698 |     pad: 1
1699 |     kernel_size: 3
1700 |     stride: 1
1701 |     weight_filler {
1702 |       type: "msra"
1703 |     }
1704 |     bias_filler {
1705 |       type: "constant"
1706 |       value: 0.0
1707 |     }
1708 |   }
1709 | }
1710 | layer {
1711 |   name: "fire5_mbox_conf_perm"
1712 |   type: "Permute"
1713 |   bottom: "fire5_mbox_conf"
1714 |   top: "fire5_mbox_conf_perm"
1715 |   permute_param {
1716 |     order: 0
1717 |     order: 2
1718 |     order: 3
1719 |     order: 1
1720 |   }
1721 | }
1722 | layer {
1723 |   name: "fire5_mbox_conf_flat"
1724 |   type: "Flatten"
1725 |   bottom: "fire5_mbox_conf_perm"
1726 |   top: "fire5_mbox_conf_flat"
1727 |   flatten_param {
1728 |     axis: 1
1729 |   }
1730 | }
1731 | layer {
1732 |   name: "fire5_mbox_priorbox"
1733 |   type: "PriorBox"
1734 |   bottom: "fire5/normal"
1735 |   bottom: "data"
1736 |   top: "fire5_mbox_priorbox"
1737 |   prior_box_param {
1738 |     min_size: 21.0
1739 |     max_size: 45.0
1740 |     aspect_ratio: 2.0
1741 |     flip: true
1742 |     clip: false
1743 |     variance: 0.10000000149
1744 |     variance: 0.10000000149
1745 |     variance: 0.20000000298
1746 |     variance: 0.20000000298
1747 |     step: 8.0
1748 |   }
1749 | }
1750 | layer {
1751 |   name: "fire9_mbox_loc"
1752 |   type: "Convolution"
1753 |   bottom: "fire9/concat"
1754 |   top: "fire9_mbox_loc"
1755 |   param {
1756 |     lr_mult: 1.0
1757 |     decay_mult: 1.0
1758 |   }
1759 |   param {
1760 |     lr_mult: 2.0
1761 |     decay_mult: 0.0
1762 |   }
1763 |   convolution_param {
1764 |     num_output: 24
1765 |     pad: 1
1766 |     kernel_size: 3
1767 |     stride: 1
1768 |     weight_filler {
1769 |       type: "msra"
1770 |     }
1771 |     bias_filler {
1772 |       type: "constant"
1773 |       value: 0.0
1774 |     }
1775 |   }
1776 | }
1777 | layer {
1778 |   name: "fire9_mbox_loc_perm"
1779 |   type: "Permute"
1780 |   bottom: "fire9_mbox_loc"
1781 |   top: "fire9_mbox_loc_perm"
1782 |   permute_param {
1783 |     order: 0
1784 |     order: 2
1785 |     order: 3
1786 |     order: 1
1787 |   }
1788 | }
1789 | layer {
1790 |   name: "fire9_mbox_loc_flat"
1791 |   type: "Flatten"
1792 |   bottom: "fire9_mbox_loc_perm"
1793 |   top: "fire9_mbox_loc_flat"
1794 |   flatten_param {
1795 |     axis: 1
1796 |   }
1797 | }
1798 | layer {
1799 |   name: "fire9_mbox_conf"
1800 |   type: "Convolution"
1801 |   bottom: "fire9/concat"
1802 |   top: "fire9_mbox_conf"
1803 |   param {
1804 |     lr_mult: 1.0
1805 |     decay_mult: 1.0
1806 |   }
1807 |   param {
1808 |     lr_mult: 2.0
1809 |     decay_mult: 0.0
1810 |   }
1811 |   convolution_param {
1812 |     num_output: 126
1813 |     pad: 1
1814 |     kernel_size: 3
1815 |     stride: 1
1816 |     weight_filler {
1817 |       type: "msra"
1818 |     }
1819 |     bias_filler {
1820 |       type: "constant"
1821 |       value: 0.0
1822 |     }
1823 |   }
1824 | }
1825 | layer {
1826 |   name: "fire9_mbox_conf_perm"
1827 |   type: "Permute"
1828 |   bottom: "fire9_mbox_conf"
1829 |   top: "fire9_mbox_conf_perm"
1830 |   permute_param {
1831 |     order: 0
1832 |     order: 2
1833 |     order: 3
1834 |     order: 1
1835 |   }
1836 | }
1837 | layer {
1838 |   name: "fire9_mbox_conf_flat"
1839 |   type: "Flatten"
1840 |   bottom: "fire9_mbox_conf_perm"
1841 |   top: "fire9_mbox_conf_flat"
1842 |   flatten_param {
1843 |     axis: 1
1844 |   }
1845 | }
1846 | layer {
1847 |   name: "fire9_mbox_priorbox"
1848 |   type: "PriorBox"
1849 |   bottom: "fire9/concat"
1850 |   bottom: "data"
1851 |   top: "fire9_mbox_priorbox"
1852 |   prior_box_param {
1853 |     min_size: 45.0
1854 |     max_size: 99.0
1855 |     aspect_ratio: 2.0
1856 |     aspect_ratio: 3.0
1857 |     flip: true
1858 |     clip: false
1859 |     variance: 0.10000000149
1860 |     variance: 0.10000000149
1861 |     variance: 0.20000000298
1862 |     variance: 0.20000000298
1863 |     step: 16.0
1864 |   }
1865 | }
1866 | layer {
1867 |   name: "fire10_mbox_loc"
1868 |   type: "Convolution"
1869 |   bottom: "fire10/concat"
1870 |   top: "fire10_mbox_loc"
1871 |   param {
1872 |     lr_mult: 1.0
1873 |     decay_mult: 1.0
1874 |   }
1875 |   param {
1876 |     lr_mult: 2.0
1877 |     decay_mult: 0.0
1878 |   }
1879 |   convolution_param {
1880 |     num_output: 24
1881 |     pad: 1
1882 |     kernel_size: 3
1883 |     stride: 1
1884 |     weight_filler {
1885 |       type: "msra"
1886 |     }
1887 |     bias_filler {
1888 |       type: "constant"
1889 |       value: 0.0
1890 |     }
1891 |   }
1892 | }
1893 | layer {
1894 |   name: "fire10_mbox_loc_perm"
1895 |   type: "Permute"
1896 |   bottom: "fire10_mbox_loc"
1897 |   top: "fire10_mbox_loc_perm"
1898 |   permute_param {
1899 |     order: 0
1900 |     order: 2
1901 |     order: 3
1902 |     order: 1
1903 |   }
1904 | }
1905 | layer {
1906 |   name: "fire10_mbox_loc_flat"
1907 |   type: "Flatten"
1908 |   bottom: "fire10_mbox_loc_perm"
1909 |   top: "fire10_mbox_loc_flat"
1910 |   flatten_param {
1911 |     axis: 1
1912 |   }
1913 | }
1914 | layer {
1915 |   name: "fire10_mbox_conf"
1916 |   type: "Convolution"
1917 |   bottom: "fire10/concat"
1918 |   top: "fire10_mbox_conf"
1919 |   param {
1920 |     lr_mult: 1.0
1921 |     decay_mult: 1.0
1922 |   }
1923 |   param {
1924 |     lr_mult: 2.0
1925 |     decay_mult: 0.0
1926 |   }
1927 |   convolution_param {
1928 |     num_output: 126
1929 |     pad: 1
1930 |     kernel_size: 3
1931 |     stride: 1
1932 |     weight_filler {
1933 |       type: "msra"
1934 |     }
1935 |     bias_filler {
1936 |       type: "constant"
1937 |       value: 0.0
1938 |     }
1939 |   }
1940 | }
1941 | layer {
1942 |   name: "fire10_mbox_conf_perm"
1943 |   type: "Permute"
1944 |   bottom: "fire10_mbox_conf"
1945 |   top: "fire10_mbox_conf_perm"
1946 |   permute_param {
1947 |     order: 0
1948 |     order: 2
1949 |     order: 3
1950 |     order: 1
1951 |   }
1952 | }
1953 | layer {
1954 |   name: "fire10_mbox_conf_flat"
1955 |   type: "Flatten"
1956 |   bottom: "fire10_mbox_conf_perm"
1957 |   top: "fire10_mbox_conf_flat"
1958 |   flatten_param {
1959 |     axis: 1
1960 |   }
1961 | }
1962 | layer {
1963 |   name: "fire10_mbox_priorbox"
1964 |   type: "PriorBox"
1965 |   bottom: "fire10/concat"
1966 |   bottom: "data"
1967 |   top: "fire10_mbox_priorbox"
1968 |   prior_box_param {
1969 |     min_size: 99.0
1970 |     max_size: 153.0
1971 |     aspect_ratio: 2.0
1972 |     aspect_ratio: 3.0
1973 |     flip: true
1974 |     clip: false
1975 |     variance: 0.10000000149
1976 |     variance: 0.10000000149
1977 |     variance: 0.20000000298
1978 |     variance: 0.20000000298
1979 |     step: 32.0
1980 |   }
1981 | }
1982 | layer {
1983 |   name: "fire11_mbox_loc"
1984 |   type: "Convolution"
1985 |   bottom: "fire11/concat"
1986 |   top: "fire11_mbox_loc"
1987 |   param {
1988 |     lr_mult: 1.0
1989 |     decay_mult: 1.0
1990 |   }
1991 |   param {
1992 |     lr_mult: 2.0
1993 |     decay_mult: 0.0
1994 |   }
1995 |   convolution_param {
1996 |     num_output: 24
1997 |     pad: 1
1998 |     kernel_size: 3
1999 |     stride: 1
2000 |     weight_filler {
2001 |       type: "msra"
2002 |     }
2003 |     bias_filler {
2004 |       type: "constant"
2005 |       value: 0.0
2006 |     }
2007 |   }
2008 | }
2009 | layer {
2010 |   name: "fire11_mbox_loc_perm"
2011 |   type: "Permute"
2012 |   bottom: "fire11_mbox_loc"
2013 |   top: "fire11_mbox_loc_perm"
2014 |   permute_param {
2015 |     order: 0
2016 |     order: 2
2017 |     order: 3
2018 |     order: 1
2019 |   }
2020 | }
2021 | layer {
2022 |   name: "fire11_mbox_loc_flat"
2023 |   type: "Flatten"
2024 |   bottom: "fire11_mbox_loc_perm"
2025 |   top: "fire11_mbox_loc_flat"
2026 |   flatten_param {
2027 |     axis: 1
2028 |   }
2029 | }
2030 | layer {
2031 |   name: "fire11_mbox_conf"
2032 |   type: "Convolution"
2033 |   bottom: "fire11/concat"
2034 |   top: "fire11_mbox_conf"
2035 |   param {
2036 |     lr_mult: 1.0
2037 |     decay_mult: 1.0
2038 |   }
2039 |   param {
2040 |     lr_mult: 2.0
2041 |     decay_mult: 0.0
2042 |   }
2043 |   convolution_param {
2044 |     num_output: 126
2045 |     pad: 1
2046 |     kernel_size: 3
2047 |     stride: 1
2048 |     weight_filler {
2049 |       type: "msra"
2050 |     }
2051 |     bias_filler {
2052 |       type: "constant"
2053 |       value: 0.0
2054 |     }
2055 |   }
2056 | }
2057 | layer {
2058 |   name: "fire11_mbox_conf_perm"
2059 |   type: "Permute"
2060 |   bottom: "fire11_mbox_conf"
2061 |   top: "fire11_mbox_conf_perm"
2062 |   permute_param {
2063 |     order: 0
2064 |     order: 2
2065 |     order: 3
2066 |     order: 1
2067 |   }
2068 | }
2069 | layer {
2070 |   name: "fire11_mbox_conf_flat"
2071 |   type: "Flatten"
2072 |   bottom: "fire11_mbox_conf_perm"
2073 |   top: "fire11_mbox_conf_flat"
2074 |   flatten_param {
2075 |     axis: 1
2076 |   }
2077 | }
2078 | layer {
2079 |   name: "fire11_mbox_priorbox"
2080 |   type: "PriorBox"
2081 |   bottom: "fire11/concat"
2082 |   bottom: "data"
2083 |   top: "fire11_mbox_priorbox"
2084 |   prior_box_param {
2085 |     min_size: 153.0
2086 |     max_size: 207.0
2087 |     aspect_ratio: 2.0
2088 |     aspect_ratio: 3.0
2089 |     flip: true
2090 |     clip: false
2091 |     variance: 0.10000000149
2092 |     variance: 0.10000000149
2093 |     variance: 0.20000000298
2094 |     variance: 0.20000000298
2095 |     step: 64.0
2096 |   }
2097 | }
2098 | layer {
2099 |   name: "conv12_2_mbox_loc"
2100 |   type: "Convolution"
2101 |   bottom: "conv12_2"
2102 |   top: "conv12_2_mbox_loc"
2103 |   param {
2104 |     lr_mult: 1.0
2105 |     decay_mult: 1.0
2106 |   }
2107 |   param {
2108 |     lr_mult: 2.0
2109 |     decay_mult: 0.0
2110 |   }
2111 |   convolution_param {
2112 |     num_output: 24
2113 |     pad: 1
2114 |     kernel_size: 3
2115 |     stride: 1
2116 |     weight_filler {
2117 |       type: "msra"
2118 |     }
2119 |     bias_filler {
2120 |       type: "constant"
2121 |       value: 0.0
2122 |     }
2123 |   }
2124 | }
2125 | layer {
2126 |   name: "conv12_2_mbox_loc_perm"
2127 |   type: "Permute"
2128 |   bottom: "conv12_2_mbox_loc"
2129 |   top: "conv12_2_mbox_loc_perm"
2130 |   permute_param {
2131 |     order: 0
2132 |     order: 2
2133 |     order: 3
2134 |     order: 1
2135 |   }
2136 | }
2137 | layer {
2138 |   name: "conv12_2_mbox_loc_flat"
2139 |   type: "Flatten"
2140 |   bottom: "conv12_2_mbox_loc_perm"
2141 |   top: "conv12_2_mbox_loc_flat"
2142 |   flatten_param {
2143 |     axis: 1
2144 |   }
2145 | }
2146 | layer {
2147 |   name: "conv12_2_mbox_conf"
2148 |   type: "Convolution"
2149 |   bottom: "conv12_2"
2150 |   top: "conv12_2_mbox_conf"
2151 |   param {
2152 |     lr_mult: 1.0
2153 |     decay_mult: 1.0
2154 |   }
2155 |   param {
2156 |     lr_mult: 2.0
2157 |     decay_mult: 0.0
2158 |   }
2159 |   convolution_param {
2160 |     num_output: 126
2161 |     pad: 1
2162 |     kernel_size: 3
2163 |     stride: 1
2164 |     weight_filler {
2165 |       type: "msra"
2166 |     }
2167 |     bias_filler {
2168 |       type: "constant"
2169 |       value: 0.0
2170 |     }
2171 |   }
2172 | }
2173 | layer {
2174 |   name: "conv12_2_mbox_conf_perm"
2175 |   type: "Permute"
2176 |   bottom: "conv12_2_mbox_conf"
2177 |   top: "conv12_2_mbox_conf_perm"
2178 |   permute_param {
2179 |     order: 0
2180 |     order: 2
2181 |     order: 3
2182 |     order: 1
2183 |   }
2184 | }
2185 | layer {
2186 |   name: "conv12_2_mbox_conf_flat"
2187 |   type: "Flatten"
2188 |   bottom: "conv12_2_mbox_conf_perm"
2189 |   top: "conv12_2_mbox_conf_flat"
2190 |   flatten_param {
2191 |     axis: 1
2192 |   }
2193 | }
2194 | layer {
2195 |   name: "conv12_2_mbox_priorbox"
2196 |   type: "PriorBox"
2197 |   bottom: "conv12_2"
2198 |   bottom: "data"
2199 |   top: "conv12_2_mbox_priorbox"
2200 |   prior_box_param {
2201 |     min_size: 207.0
2202 |     max_size: 261.0
2203 |     aspect_ratio: 2.0
2204 |     aspect_ratio: 3.0
2205 |     flip: true
2206 |     clip: false
2207 |     variance: 0.10000000149
2208 |     variance: 0.10000000149
2209 |     variance: 0.20000000298
2210 |     variance: 0.20000000298
2211 |     step: 100.0
2212 |   }
2213 | }
2214 | layer {
2215 |   name: "conv13_2_mbox_loc"
2216 |   type: "Convolution"
2217 |   bottom: "conv13_2"
2218 |   top: "conv13_2_mbox_loc"
2219 |   param {
2220 |     lr_mult: 1.0
2221 |     decay_mult: 1.0
2222 |   }
2223 |   param {
2224 |     lr_mult: 2.0
2225 |     decay_mult: 0.0
2226 |   }
2227 |   convolution_param {
2228 |     num_output: 16
2229 |     pad: 1
2230 |     kernel_size: 3
2231 |     stride: 1
2232 |     weight_filler {
2233 |       type: "msra"
2234 |     }
2235 |     bias_filler {
2236 |       type: "constant"
2237 |       value: 0.0
2238 |     }
2239 |   }
2240 | }
2241 | layer {
2242 |   name: "conv13_2_mbox_loc_perm"
2243 |   type: "Permute"
2244 |   bottom: "conv13_2_mbox_loc"
2245 |   top: "conv13_2_mbox_loc_perm"
2246 |   permute_param {
2247 |     order: 0
2248 |     order: 2
2249 |     order: 3
2250 |     order: 1
2251 |   }
2252 | }
2253 | layer {
2254 |   name: "conv13_2_mbox_loc_flat"
2255 |   type: "Flatten"
2256 |   bottom: "conv13_2_mbox_loc_perm"
2257 |   top: "conv13_2_mbox_loc_flat"
2258 |   flatten_param {
2259 |     axis: 1
2260 |   }
2261 | }
2262 | layer {
2263 |   name: "conv13_2_mbox_conf"
2264 |   type: "Convolution"
2265 |   bottom: "conv13_2"
2266 |   top: "conv13_2_mbox_conf"
2267 |   param {
2268 |     lr_mult: 1.0
2269 |     decay_mult: 1.0
2270 |   }
2271 |   param {
2272 |     lr_mult: 2.0
2273 |     decay_mult: 0.0
2274 |   }
2275 |   convolution_param {
2276 |     num_output: 84
2277 |     pad: 1
2278 |     kernel_size: 3
2279 |     stride: 1
2280 |     weight_filler {
2281 |       type: "msra"
2282 |     }
2283 |     bias_filler {
2284 |       type: "constant"
2285 |       value: 0.0
2286 |     }
2287 |   }
2288 | }
2289 | layer {
2290 |   name: "conv13_2_mbox_conf_perm"
2291 |   type: "Permute"
2292 |   bottom: "conv13_2_mbox_conf"
2293 |   top: "conv13_2_mbox_conf_perm"
2294 |   permute_param {
2295 |     order: 0
2296 |     order: 2
2297 |     order: 3
2298 |     order: 1
2299 |   }
2300 | }
2301 | layer {
2302 |   name: "conv13_2_mbox_conf_flat"
2303 |   type: "Flatten"
2304 |   bottom: "conv13_2_mbox_conf_perm"
2305 |   top: "conv13_2_mbox_conf_flat"
2306 |   flatten_param {
2307 |     axis: 1
2308 |   }
2309 | }
2310 | layer {
2311 |   name: "conv13_2_mbox_priorbox"
2312 |   type: "PriorBox"
2313 |   bottom: "conv13_2"
2314 |   bottom: "data"
2315 |   top: "conv13_2_mbox_priorbox"
2316 |   prior_box_param {
2317 |     min_size: 261.0
2318 |     max_size: 315.0
2319 |     aspect_ratio: 2.0
2320 |     flip: true
2321 |     clip: false
2322 |     variance: 0.10000000149
2323 |     variance: 0.10000000149
2324 |     variance: 0.20000000298
2325 |     variance: 0.20000000298
2326 |     step: 300.0
2327 |   }
2328 | }
2329 | layer {
2330 |   name: "mbox_loc"
2331 |   type: "Concat"
2332 |   bottom: "fire5_mbox_loc_flat"
2333 |   bottom: "fire9_mbox_loc_flat"
2334 |   bottom: "fire10_mbox_loc_flat"
2335 |   bottom: "fire11_mbox_loc_flat"
2336 |   bottom: "conv12_2_mbox_loc_flat"
2337 |   bottom: "conv13_2_mbox_loc_flat"
2338 |   top: "mbox_loc"
2339 |   concat_param {
2340 |     axis: 1
2341 |   }
2342 | }
2343 | layer {
2344 |   name: "mbox_conf"
2345 |   type: "Concat"
2346 |   bottom: "fire5_mbox_conf_flat"
2347 |   bottom: "fire9_mbox_conf_flat"
2348 |   bottom: "fire10_mbox_conf_flat"
2349 |   bottom: "fire11_mbox_conf_flat"
2350 |   bottom: "conv12_2_mbox_conf_flat"
2351 |   bottom: "conv13_2_mbox_conf_flat"
2352 |   top: "mbox_conf"
2353 |   concat_param {
2354 |     axis: 1
2355 |   }
2356 | }
2357 | layer {
2358 |   name: "mbox_priorbox"
2359 |   type: "Concat"
2360 |   bottom: "fire5_mbox_priorbox"
2361 |   bottom: "fire9_mbox_priorbox"
2362 |   bottom: "fire10_mbox_priorbox"
2363 |   bottom: "fire11_mbox_priorbox"
2364 |   bottom: "conv12_2_mbox_priorbox"
2365 |   bottom: "conv13_2_mbox_priorbox"
2366 |   top: "mbox_priorbox"
2367 |   concat_param {
2368 |     axis: 2
2369 |   }
2370 | }
2371 | layer {
2372 |   name: "mbox_loss"
2373 |   type: "MultiBoxLoss"
2374 |   bottom: "mbox_loc"
2375 |   bottom: "mbox_conf"
2376 |   bottom: "mbox_priorbox"
2377 |   bottom: "label"
2378 |   top: "mbox_loss"
2379 |   include {
2380 |     phase: TRAIN
2381 |   }
2382 |   propagate_down: true
2383 |   propagate_down: true
2384 |   propagate_down: false
2385 |   propagate_down: false
2386 |   loss_param {
2387 |     normalization: VALID
2388 |   }
2389 |   multibox_loss_param {
2390 |     loc_loss_type: SMOOTH_L1
2391 |     conf_loss_type: SOFTMAX
2392 |     loc_weight: 1.0
2393 |     num_classes: 21
2394 |     share_location: true
2395 |     match_type: PER_PREDICTION
2396 |     overlap_threshold: 0.5
2397 |     use_prior_for_matching: true
2398 |     background_label_id: 0
2399 |     use_difficult_gt: true
2400 |     neg_pos_ratio: 3.0
2401 |     neg_overlap: 0.5
2402 |     code_type: CENTER_SIZE
2403 |     ignore_cross_boundary_bbox: false
2404 |     mining_type: MAX_NEGATIVE
2405 |   }
2406 | }
2407 | 


--------------------------------------------------------------------------------
/solver.prototxt:
--------------------------------------------------------------------------------
 1 | train_net: "models/tiny_ssd_squeezenet/tiny_ssd_voc_train.prototxt"
 2 | test_net: "models/tiny_ssd_squeezenet/tiny_ssd_voc_test.prototxt"
 3 | test_iter: 619
 4 | test_interval: 10000
 5 | base_lr: 0.0001
 6 | display: 100
 7 | max_iter: 400000
 8 | lr_policy: "multistep"
 9 | gamma: 0.5
10 | weight_decay: 0.0
11 | snapshot: 10000
12 | snapshot_prefix: "models/tiny_ssd_squeezenet/model/"
13 | solver_mode: GPU
14 | device_id: 0
15 | debug_info: false
16 | snapshot_after_train: true
17 | test_initialization: false
18 | average_loss: 10
19 | stepvalue: 190000
20 | stepvalue: 250000
21 | stepvalue: 280000
22 | iter_size: 1
23 | type: "RMSProp"
24 | eval_type: "detection"
25 | ap_version: "11point"
26 | 


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