├── .gitignore
├── demo.sh
├── vgg16.py
├── utils.py
├── convert_vgg16.py
├── convert_resnet152.py
├── VGG-Places365.prototxt
├── Places2_365_CNN.py
└── Places2-365-CNN.prototxt


/.gitignore:
--------------------------------------------------------------------------------
1 | *.pyc
2 | 


--------------------------------------------------------------------------------
/demo.sh:
--------------------------------------------------------------------------------
 1 | export LD_PRELOAD=libmkl_rt.so
 2 | export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/nccl/build/lib
 3 | 
 4 | CAFFE_ROOT=/home/lab-qiu.suo/devel/caffe
 5 | export PYTHONPATH=$CAFFE_ROOT/python:$PYTHONPATH
 6 | 
 7 | set -x
 8 | set -e
 9 | 
10 | LOG=convert.log
11 | python convert_resnet152.py \
12 | 2>&1 | tee -i $LOG


--------------------------------------------------------------------------------
/vgg16.py:
--------------------------------------------------------------------------------
 1 | import torch.nn as nn
 2 | 
 3 | cfg = [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512]#, 'M']
 4 | 
 5 | class VGG16(nn.Module):
 6 |     def __init__(self):
 7 |         super(VGG16, self).__init__()
 8 |         self.features = self.make_layers()
 9 | 
10 |     def make_layers(self):
11 |         features = nn.Sequential()
12 |         in_channels = 3
13 |         ith_block = 1
14 |         ith_sublayer = 1
15 |         for v in cfg:
16 |             if v == 'M':
17 |                 features.add_module('pool%d'%ith_block, nn.MaxPool2d(kernel_size=2, stride=2))
18 |                 ith_block += 1
19 |                 ith_sublayer = 1
20 |             else:
21 |                 features.add_module('conv%d_%d'%(ith_block,ith_sublayer), 
22 |                                     nn.Conv2d(in_channels, v, kernel_size=3, padding=1))
23 |                 features.add_module('relu%d_%d'%(ith_block,ith_sublayer), nn.ReLU(inplace=True))
24 |                 ith_sublayer += 1
25 |                 in_channels = v
26 |         return features
27 | 
28 |     def forward(self, x):
29 |         x = self.features(x)
30 |         return x
31 | 
32 |     


--------------------------------------------------------------------------------
/utils.py:
--------------------------------------------------------------------------------
 1 | import caffe
 2 | 
 3 | def load_caffe(proto, weight):
 4 |     caffe.set_mode_cpu()
 5 |     net = caffe.Net(proto, weight, caffe.TEST)
 6 |     return net
 7 | 
 8 | class CaffeParamProvider():
 9 |     def __init__(self, caffe_net):
10 |         self.caffe_net = caffe_net
11 | 
12 |     def conv_kernel(self, name):
13 |         k = self.caffe_net.params[name][0].data
14 |         return k
15 | 
16 |     def conv_biases(self, name):
17 |         k = self.caffe_net.params[name][1].data
18 |         return k
19 | 
20 |     def bn_gamma(self, name):
21 |         return self.caffe_net.params[name][0].data
22 | 
23 |     def bn_beta(self, name):
24 |         return self.caffe_net.params[name][1].data
25 | 
26 |     def bn_mean(self, name):
27 |         return (self.caffe_net.params[name][0].data/self.caffe_net.params[name][2].data)
28 | 
29 |     def bn_variance(self, name):
30 |         return (self.caffe_net.params[name][1].data/self.caffe_net.params[name][2].data)
31 | 
32 |     def fc_weights(self, name):
33 |         w = self.caffe_net.params[name][0].data
34 |         #w = w.transpose((1, 0))
35 |         return w
36 | 
37 |     def fc_biases(self, name):
38 |         b = self.caffe_net.params[name][1].data
39 |         return b


--------------------------------------------------------------------------------
/convert_vgg16.py:
--------------------------------------------------------------------------------
 1 | from utils import *
 2 | import torch
 3 | from collections import OrderedDict
 4 | 
 5 | def parse_param_name(p, var_name):
 6 |     a = var_name.strip().split('.')
 7 |     name = a[-2]
 8 |     if a[-1] == 'weight':
 9 |         data = p.conv_kernel(name)
10 |     if a[-1] == 'bias':
11 |     	data = p.conv_biases(name)
12 |     return name, data
13 | 
14 | if __name__ == '__main__':
15 |     
16 |     prototxt = 'VGG-Places365.prototxt'
17 |     caffemodel = 'vgg16_places365.caffemodel'
18 |     net = load_caffe(prototxt, caffemodel)
19 |     caffe_keys = net.params.keys()
20 |     print(caffe_keys)
21 |     param_provider = CaffeParamProvider(net)
22 |     record_caffe = dict.fromkeys(caffe_keys, 0)
23 | 
24 |     from vgg16 import VGG16
25 |     model = VGG16()
26 |     py_keys = model.state_dict().keys()
27 |     print(py_keys)
28 | 
29 |     print('caffe: {} keys, pytoch: {} keys'.format(len(caffe_keys),len(py_keys)))
30 | 
31 |     new_state_dict = OrderedDict()
32 |     for var_name in py_keys:
33 |         name, data = parse_param_name(param_provider, var_name)
34 |         new_state_dict[var_name] = torch.from_numpy(data).float()
35 |         print('copy param from {} to {}'.format(name, var_name))
36 |         record_caffe[name] = record_caffe[name] + 1
37 | 
38 |     for key, val in record_caffe.items():
39 |         if val == 0:
40 |             print('warning: ignoring {}'.format(key))
41 | 
42 |     model_dict = model.state_dict()
43 |     model_dict.update(new_state_dict)
44 |     model.load_state_dict(model_dict)
45 |     torch.save(model.state_dict(), 'VGG-Places365.pth')


--------------------------------------------------------------------------------
/convert_resnet152.py:
--------------------------------------------------------------------------------
 1 | from utils import *
 2 | import torch
 3 | from collections import OrderedDict
 4 | 
 5 | def parse_param_name(p, var_name):
 6 |     name = 'conv%d_%d_%dx%d'
 7 | 
 8 |     a = var_name.strip().split('.')
 9 |     if len(a) == 2:
10 |         subith_conv = int(int(a[0]) / 3) + 1
11 |         name = name % (1, subith_conv, 3, 3)
12 |         if subith_conv == 1:
13 |             name = name + '_s2'
14 |         if a[-1] == 'weight':
15 |             if int(a[0]) % 3 == 0:
16 |                 data = p.conv_kernel(name)
17 |             else:
18 |                 name = name + '/scale'
19 |                 data = p.bn_gamma(name)
20 |     else:
21 |         ith_conv = int(a[0]) - 8
22 |         subith_conv = int(a[1]) + 1
23 |         if int(a[3]) == 1:
24 |             name = name % (ith_conv, subith_conv, 1, 1) + '_proj'
25 |         else:
26 |             subblock = int(int(a[4]) / 3)
27 |             if subblock == 0:
28 |                 name = name % (ith_conv, subith_conv, 1, 1) + '_reduce'
29 |             elif subblock == 1:
30 |                 name = name % (ith_conv, subith_conv, 3, 3)
31 |             elif subblock == 2:
32 |                 name = name % (ith_conv, subith_conv, 1, 1) + '_increase'
33 |         if a[-1] == 'weight':
34 |             if int(a[4]) % 3 == 0:
35 |                 data = p.conv_kernel(name)
36 |             else:
37 |                 name = name + '/scale'
38 |                 data = p.bn_gamma(name)
39 |     if a[-1] == 'bias':
40 |         name = name + '/scale'
41 |         data = p.bn_beta(name)
42 |     elif a[-1] == 'running_mean':
43 |         name = name + '/bn'
44 |         data = p.bn_mean(name)
45 |     elif a[-1] == 'running_var':
46 |         name = name + '/bn'
47 |         data = p.bn_variance(name)
48 | 
49 |     return name, data
50 | 
51 | if __name__ == '__main__':
52 |     
53 |     prototxt = 'Places2-365-CNN.prototxt'
54 |     caffemodel = 'Places2-365-CNN.caffemodel'
55 |     net = load_caffe(prototxt, caffemodel)
56 |     caffe_keys = net.params.keys()
57 |     print(caffe_keys)
58 |     param_provider = CaffeParamProvider(net)
59 |     record_caffe = dict.fromkeys(caffe_keys, 0)
60 | 
61 |     import Places2_365_CNN
62 |     model = Places2_365_CNN.resnet152_places365
63 |     py_keys = model.state_dict().keys()
64 |     print(py_keys)
65 | 
66 |     print('caffe: {} keys, pytoch: {} keys'.format(len(caffe_keys),len(py_keys)))
67 | 
68 |     new_state_dict = OrderedDict()
69 |     for var_name in py_keys:
70 |         if var_name == '16.1.weight':
71 |             break
72 |         name, data = parse_param_name(param_provider, var_name)
73 |         new_state_dict[var_name] = torch.from_numpy(data).float()
74 |         print('copy param from {} to {}'.format(name, var_name))
75 |         record_caffe[name] = record_caffe[name] + 1
76 | 
77 |     for key, val in record_caffe.items():
78 |         if val == 0:
79 |             print('warning: ignoring {}'.format(key))
80 | 
81 |     model_dict = model.state_dict()
82 |     model_dict.update(new_state_dict)
83 |     model.load_state_dict(model_dict)
84 |     torch.save(model.state_dict(), 'Places2_365_CNN.pth')


--------------------------------------------------------------------------------
/VGG-Places365.prototxt:
--------------------------------------------------------------------------------
  1 | name: "VGG-Places365"
  2 | input: "data"
  3 | input_dim: 10
  4 | input_dim: 3
  5 | input_dim: 224
  6 | input_dim: 224
  7 | layer {
  8 |   name: "conv1_1"
  9 |   type: "Convolution"
 10 |   bottom: "data"
 11 |   top: "conv1_1"
 12 |   param {
 13 |     lr_mult: 1.0
 14 |     decay_mult: 1.0
 15 |   }
 16 |   param {
 17 |     lr_mult: 2.0
 18 |     decay_mult: 0.0
 19 |   }
 20 |   convolution_param {
 21 |     num_output: 64
 22 |     pad: 1
 23 |     kernel_size: 3
 24 |     weight_filler {
 25 |       type: "gaussian"
 26 |       std: 0.01
 27 |     }
 28 |     bias_filler {
 29 |       type: "constant"
 30 |       value: 0.0
 31 |     }
 32 |   }
 33 | }
 34 | layer {
 35 |   name: "relu1_1"
 36 |   type: "ReLU"
 37 |   bottom: "conv1_1"
 38 |   top: "conv1_1"
 39 | }
 40 | layer {
 41 |   name: "conv1_2"
 42 |   type: "Convolution"
 43 |   bottom: "conv1_1"
 44 |   top: "conv1_2"
 45 |   param {
 46 |     lr_mult: 1.0
 47 |     decay_mult: 1.0
 48 |   }
 49 |   param {
 50 |     lr_mult: 2.0
 51 |     decay_mult: 0.0
 52 |   }
 53 |   convolution_param {
 54 |     num_output: 64
 55 |     pad: 1
 56 |     kernel_size: 3
 57 |     weight_filler {
 58 |       type: "gaussian"
 59 |       std: 0.01
 60 |     }
 61 |     bias_filler {
 62 |       type: "constant"
 63 |       value: 0.0
 64 |     }
 65 |   }
 66 | }
 67 | layer {
 68 |   name: "relu1_2"
 69 |   type: "ReLU"
 70 |   bottom: "conv1_2"
 71 |   top: "conv1_2"
 72 | }
 73 | layer {
 74 |   name: "pool1"
 75 |   type: "Pooling"
 76 |   bottom: "conv1_2"
 77 |   top: "pool1"
 78 |   pooling_param {
 79 |     pool: MAX
 80 |     kernel_size: 2
 81 |     stride: 2
 82 |   }
 83 | }
 84 | layer {
 85 |   name: "conv2_1"
 86 |   type: "Convolution"
 87 |   bottom: "pool1"
 88 |   top: "conv2_1"
 89 |   param {
 90 |     lr_mult: 1.0
 91 |     decay_mult: 1.0
 92 |   }
 93 |   param {
 94 |     lr_mult: 2.0
 95 |     decay_mult: 0.0
 96 |   }
 97 |   convolution_param {
 98 |     num_output: 128
 99 |     pad: 1
100 |     kernel_size: 3
101 |     weight_filler {
102 |       type: "gaussian"
103 |       std: 0.01
104 |     }
105 |     bias_filler {
106 |       type: "constant"
107 |       value: 0.0
108 |     }
109 |   }
110 | }
111 | layer {
112 |   name: "relu2_1"
113 |   type: "ReLU"
114 |   bottom: "conv2_1"
115 |   top: "conv2_1"
116 | }
117 | layer {
118 |   name: "conv2_2"
119 |   type: "Convolution"
120 |   bottom: "conv2_1"
121 |   top: "conv2_2"
122 |   param {
123 |     lr_mult: 1.0
124 |     decay_mult: 1.0
125 |   }
126 |   param {
127 |     lr_mult: 2.0
128 |     decay_mult: 0.0
129 |   }
130 |   convolution_param {
131 |     num_output: 128
132 |     pad: 1
133 |     kernel_size: 3
134 |     weight_filler {
135 |       type: "gaussian"
136 |       std: 0.01
137 |     }
138 |     bias_filler {
139 |       type: "constant"
140 |       value: 0.0
141 |     }
142 |   }
143 | }
144 | layer {
145 |   name: "relu2_2"
146 |   type: "ReLU"
147 |   bottom: "conv2_2"
148 |   top: "conv2_2"
149 | }
150 | layer {
151 |   name: "pool2"
152 |   type: "Pooling"
153 |   bottom: "conv2_2"
154 |   top: "pool2"
155 |   pooling_param {
156 |     pool: MAX
157 |     kernel_size: 2
158 |     stride: 2
159 |   }
160 | }
161 | layer {
162 |   name: "conv3_1"
163 |   type: "Convolution"
164 |   bottom: "pool2"
165 |   top: "conv3_1"
166 |   param {
167 |     lr_mult: 1.0
168 |     decay_mult: 1.0
169 |   }
170 |   param {
171 |     lr_mult: 2.0
172 |     decay_mult: 0.0
173 |   }
174 |   convolution_param {
175 |     num_output: 256
176 |     pad: 1
177 |     kernel_size: 3
178 |     weight_filler {
179 |       type: "gaussian"
180 |       std: 0.01
181 |     }
182 |     bias_filler {
183 |       type: "constant"
184 |       value: 0.0
185 |     }
186 |   }
187 | }
188 | layer {
189 |   name: "relu3_1"
190 |   type: "ReLU"
191 |   bottom: "conv3_1"
192 |   top: "conv3_1"
193 | }
194 | layer {
195 |   name: "conv3_2"
196 |   type: "Convolution"
197 |   bottom: "conv3_1"
198 |   top: "conv3_2"
199 |   param {
200 |     lr_mult: 1.0
201 |     decay_mult: 1.0
202 |   }
203 |   param {
204 |     lr_mult: 2.0
205 |     decay_mult: 0.0
206 |   }
207 |   convolution_param {
208 |     num_output: 256
209 |     pad: 1
210 |     kernel_size: 3
211 |     weight_filler {
212 |       type: "gaussian"
213 |       std: 0.01
214 |     }
215 |     bias_filler {
216 |       type: "constant"
217 |       value: 0.0
218 |     }
219 |   }
220 | }
221 | layer {
222 |   name: "relu3_2"
223 |   type: "ReLU"
224 |   bottom: "conv3_2"
225 |   top: "conv3_2"
226 | }
227 | layer {
228 |   name: "conv3_3"
229 |   type: "Convolution"
230 |   bottom: "conv3_2"
231 |   top: "conv3_3"
232 |   param {
233 |     lr_mult: 1.0
234 |     decay_mult: 1.0
235 |   }
236 |   param {
237 |     lr_mult: 2.0
238 |     decay_mult: 0.0
239 |   }
240 |   convolution_param {
241 |     num_output: 256
242 |     pad: 1
243 |     kernel_size: 3
244 |     weight_filler {
245 |       type: "gaussian"
246 |       std: 0.01
247 |     }
248 |     bias_filler {
249 |       type: "constant"
250 |       value: 0.0
251 |     }
252 |   }
253 | }
254 | layer {
255 |   name: "relu3_3"
256 |   type: "ReLU"
257 |   bottom: "conv3_3"
258 |   top: "conv3_3"
259 | }
260 | layer {
261 |   name: "pool3"
262 |   type: "Pooling"
263 |   bottom: "conv3_3"
264 |   top: "pool3"
265 |   pooling_param {
266 |     pool: MAX
267 |     kernel_size: 2
268 |     stride: 2
269 |   }
270 | }
271 | layer {
272 |   name: "conv4_1"
273 |   type: "Convolution"
274 |   bottom: "pool3"
275 |   top: "conv4_1"
276 |   param {
277 |     lr_mult: 1.0
278 |     decay_mult: 1.0
279 |   }
280 |   param {
281 |     lr_mult: 2.0
282 |     decay_mult: 0.0
283 |   }
284 |   convolution_param {
285 |     num_output: 512
286 |     pad: 1
287 |     kernel_size: 3
288 |     weight_filler {
289 |       type: "gaussian"
290 |       std: 0.01
291 |     }
292 |     bias_filler {
293 |       type: "constant"
294 |       value: 0.0
295 |     }
296 |   }
297 | }
298 | layer {
299 |   name: "relu4_1"
300 |   type: "ReLU"
301 |   bottom: "conv4_1"
302 |   top: "conv4_1"
303 | }
304 | layer {
305 |   name: "conv4_2"
306 |   type: "Convolution"
307 |   bottom: "conv4_1"
308 |   top: "conv4_2"
309 |   param {
310 |     lr_mult: 1.0
311 |     decay_mult: 1.0
312 |   }
313 |   param {
314 |     lr_mult: 2.0
315 |     decay_mult: 0.0
316 |   }
317 |   convolution_param {
318 |     num_output: 512
319 |     pad: 1
320 |     kernel_size: 3
321 |     weight_filler {
322 |       type: "gaussian"
323 |       std: 0.01
324 |     }
325 |     bias_filler {
326 |       type: "constant"
327 |       value: 0.0
328 |     }
329 |   }
330 | }
331 | layer {
332 |   name: "relu4_2"
333 |   type: "ReLU"
334 |   bottom: "conv4_2"
335 |   top: "conv4_2"
336 | }
337 | layer {
338 |   name: "conv4_3"
339 |   type: "Convolution"
340 |   bottom: "conv4_2"
341 |   top: "conv4_3"
342 |   param {
343 |     lr_mult: 1.0
344 |     decay_mult: 1.0
345 |   }
346 |   param {
347 |     lr_mult: 2.0
348 |     decay_mult: 0.0
349 |   }
350 |   convolution_param {
351 |     num_output: 512
352 |     pad: 1
353 |     kernel_size: 3
354 |     weight_filler {
355 |       type: "gaussian"
356 |       std: 0.01
357 |     }
358 |     bias_filler {
359 |       type: "constant"
360 |       value: 0.0
361 |     }
362 |   }
363 | }
364 | layer {
365 |   name: "relu4_3"
366 |   type: "ReLU"
367 |   bottom: "conv4_3"
368 |   top: "conv4_3"
369 | }
370 | layer {
371 |   name: "pool4"
372 |   type: "Pooling"
373 |   bottom: "conv4_3"
374 |   top: "pool4"
375 |   pooling_param {
376 |     pool: MAX
377 |     kernel_size: 2
378 |     stride: 2
379 |   }
380 | }
381 | layer {
382 |   name: "conv5_1"
383 |   type: "Convolution"
384 |   bottom: "pool4"
385 |   top: "conv5_1"
386 |   param {
387 |     lr_mult: 1.0
388 |     decay_mult: 1.0
389 |   }
390 |   param {
391 |     lr_mult: 2.0
392 |     decay_mult: 0.0
393 |   }
394 |   convolution_param {
395 |     num_output: 512
396 |     pad: 1
397 |     kernel_size: 3
398 |     weight_filler {
399 |       type: "gaussian"
400 |       std: 0.01
401 |     }
402 |     bias_filler {
403 |       type: "constant"
404 |       value: 0.0
405 |     }
406 |   }
407 | }
408 | layer {
409 |   name: "relu5_1"
410 |   type: "ReLU"
411 |   bottom: "conv5_1"
412 |   top: "conv5_1"
413 | }
414 | layer {
415 |   name: "conv5_2"
416 |   type: "Convolution"
417 |   bottom: "conv5_1"
418 |   top: "conv5_2"
419 |   param {
420 |     lr_mult: 1.0
421 |     decay_mult: 1.0
422 |   }
423 |   param {
424 |     lr_mult: 2.0
425 |     decay_mult: 0.0
426 |   }
427 |   convolution_param {
428 |     num_output: 512
429 |     pad: 1
430 |     kernel_size: 3
431 |     weight_filler {
432 |       type: "gaussian"
433 |       std: 0.01
434 |     }
435 |     bias_filler {
436 |       type: "constant"
437 |       value: 0.0
438 |     }
439 |   }
440 | }
441 | layer {
442 |   name: "relu5_2"
443 |   type: "ReLU"
444 |   bottom: "conv5_2"
445 |   top: "conv5_2"
446 | }
447 | layer {
448 |   name: "conv5_3"
449 |   type: "Convolution"
450 |   bottom: "conv5_2"
451 |   top: "conv5_3"
452 |   param {
453 |     lr_mult: 1.0
454 |     decay_mult: 1.0
455 |   }
456 |   param {
457 |     lr_mult: 2.0
458 |     decay_mult: 0.0
459 |   }
460 |   convolution_param {
461 |     num_output: 512
462 |     pad: 1
463 |     kernel_size: 3
464 |     weight_filler {
465 |       type: "gaussian"
466 |       std: 0.01
467 |     }
468 |     bias_filler {
469 |       type: "constant"
470 |       value: 0.0
471 |     }
472 |   }
473 | }
474 | layer {
475 |   name: "relu5_3"
476 |   type: "ReLU"
477 |   bottom: "conv5_3"
478 |   top: "conv5_3"
479 | }
480 | layer {
481 |   name: "pool5"
482 |   type: "Pooling"
483 |   bottom: "conv5_3"
484 |   top: "pool5"
485 |   pooling_param {
486 |     pool: MAX
487 |     kernel_size: 2
488 |     stride: 2
489 |   }
490 | }
491 | layer {
492 |   name: "fc6"
493 |   type: "InnerProduct"
494 |   bottom: "pool5"
495 |   top: "fc6"
496 |   param {
497 |     lr_mult: 1.0
498 |     decay_mult: 1.0
499 |   }
500 |   param {
501 |     lr_mult: 2.0
502 |     decay_mult: 0.0
503 |   }
504 |   inner_product_param {
505 |     num_output: 4096
506 |     weight_filler {
507 |       type: "gaussian"
508 |       std: 0.01
509 |     }
510 |     bias_filler {
511 |       type: "constant"
512 |       value: 0.0
513 |     }
514 |   }
515 | }
516 | layer {
517 |   name: "relu6"
518 |   type: "ReLU"
519 |   bottom: "fc6"
520 |   top: "fc6"
521 | }
522 | layer {
523 |   name: "drop6"
524 |   type: "Dropout"
525 |   bottom: "fc6"
526 |   top: "fc6"
527 |   dropout_param {
528 |     dropout_ratio: 0.5
529 |   }
530 | }
531 | layer {
532 |   name: "fc7"
533 |   type: "InnerProduct"
534 |   bottom: "fc6"
535 |   top: "fc7"
536 |   param {
537 |     lr_mult: 1.0
538 |     decay_mult: 1.0
539 |   }
540 |   param {
541 |     lr_mult: 2.0
542 |     decay_mult: 0.0
543 |   }
544 |   inner_product_param {
545 |     num_output: 4096
546 |     weight_filler {
547 |       type: "gaussian"
548 |       std: 0.01
549 |     }
550 |     bias_filler {
551 |       type: "constant"
552 |       value: 0.0
553 |     }
554 |   }
555 | }
556 | layer {
557 |   name: "relu7"
558 |   type: "ReLU"
559 |   bottom: "fc7"
560 |   top: "fc7"
561 | }
562 | layer {
563 |   name: "drop7"
564 |   type: "Dropout"
565 |   bottom: "fc7"
566 |   top: "fc7"
567 |   dropout_param {
568 |     dropout_ratio: 0.5
569 |   }
570 | }
571 | layer {
572 |   name: "fc8a"
573 |   type: "InnerProduct"
574 |   bottom: "fc7"
575 |   top: "fc8a"
576 |   param {
577 |     lr_mult: 1.0
578 |     decay_mult: 1.0
579 |   }
580 |   param {
581 |     lr_mult: 2.0
582 |     decay_mult: 0.0
583 |   }
584 |   inner_product_param {
585 |     num_output: 365
586 |   }
587 | }
588 | layer {
589 |   name: "prob"
590 |   type: "Softmax"
591 |   bottom: "fc8a"
592 |   top: "prob"
593 | }


--------------------------------------------------------------------------------
/Places2_365_CNN.py:
--------------------------------------------------------------------------------
  1 | 
  2 | import torch
  3 | import torch.nn as nn
  4 | from torch.autograd import Variable
  5 | from functools import reduce
  6 | 
  7 | class LambdaBase(nn.Sequential):
  8 |     def __init__(self, fn, *args):
  9 |         super(LambdaBase, self).__init__(*args)
 10 |         self.lambda_func = fn
 11 | 
 12 |     def forward_prepare(self, input):
 13 |         output = []
 14 |         for module in self._modules.values():
 15 |             output.append(module(input))
 16 |         return output if output else input
 17 | 
 18 | class Lambda(LambdaBase):
 19 |     def forward(self, input):
 20 |         return self.lambda_func(self.forward_prepare(input))
 21 | 
 22 | class LambdaMap(LambdaBase):
 23 |     def forward(self, input):
 24 |         return list(map(self.lambda_func,self.forward_prepare(input)))
 25 | 
 26 | class LambdaReduce(LambdaBase):
 27 |     def forward(self, input):
 28 |         return reduce(self.lambda_func,self.forward_prepare(input))
 29 | 
 30 | 
 31 | resnet152_places365 = nn.Sequential( # Sequential,
 32 |     # --------------------- Conv1 1:3
 33 |     nn.Conv2d(3,64,(3, 3),(2, 2),(1, 1),1,1,bias=False),
 34 |     nn.BatchNorm2d(64),
 35 |     nn.ReLU(),
 36 |     nn.Conv2d(64,64,(3, 3),(1, 1),(1, 1),1,1,bias=False),
 37 |     nn.BatchNorm2d(64),
 38 |     nn.ReLU(),
 39 |     nn.Conv2d(64,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
 40 |     nn.BatchNorm2d(128),
 41 |     nn.ReLU(),
 42 |     nn.MaxPool2d((3, 3),(2, 2),(1, 1)),
 43 |     # --------------------- Conv2 1:8
 44 |     nn.Sequential( # Sequential,
 45 |         nn.Sequential( # Sequential,
 46 |             LambdaMap(lambda x: x, # ConcatTable,
 47 |                 nn.Sequential( # Sequential,
 48 |                     nn.Conv2d(128,64,(1, 1),(1, 1),(0, 0),1,1,bias=False),
 49 |                     nn.BatchNorm2d(64),
 50 |                     nn.ReLU(),
 51 |                     nn.Conv2d(64,64,(3, 3),(1, 1),(1, 1),1,1,bias=False),
 52 |                     nn.BatchNorm2d(64),
 53 |                     nn.ReLU(),
 54 |                     nn.Conv2d(64,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
 55 |                     nn.BatchNorm2d(256),
 56 |                 ),
 57 |                 nn.Sequential( # Sequential,
 58 |                     nn.Conv2d(128,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
 59 |                     nn.BatchNorm2d(256),
 60 |                 ),
 61 |             ),
 62 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
 63 |             nn.ReLU(),
 64 |         ),
 65 |         nn.Sequential( # Sequential,
 66 |             LambdaMap(lambda x: x, # ConcatTable,
 67 |                 nn.Sequential( # Sequential,
 68 |                     nn.Conv2d(256,64,(1, 1),(1, 1),(0, 0),1,1,bias=False),
 69 |                     nn.BatchNorm2d(64),
 70 |                     nn.ReLU(),
 71 |                     nn.Conv2d(64,64,(3, 3),(1, 1),(1, 1),1,1,bias=False),
 72 |                     nn.BatchNorm2d(64),
 73 |                     nn.ReLU(),
 74 |                     nn.Conv2d(64,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
 75 |                     nn.BatchNorm2d(256),
 76 |                 ),
 77 |                 Lambda(lambda x: x), # Identity,
 78 |             ),
 79 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
 80 |             nn.ReLU(),
 81 |         ),
 82 |         nn.Sequential( # Sequential,
 83 |             LambdaMap(lambda x: x, # ConcatTable,
 84 |                 nn.Sequential( # Sequential,
 85 |                     nn.Conv2d(256,64,(1, 1),(1, 1),(0, 0),1,1,bias=False),
 86 |                     nn.BatchNorm2d(64),
 87 |                     nn.ReLU(),
 88 |                     nn.Conv2d(64,64,(3, 3),(1, 1),(1, 1),1,1,bias=False),
 89 |                     nn.BatchNorm2d(64),
 90 |                     nn.ReLU(),
 91 |                     nn.Conv2d(64,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
 92 |                     nn.BatchNorm2d(256),
 93 |                 ),
 94 |                 Lambda(lambda x: x), # Identity,
 95 |             ),
 96 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
 97 |             nn.ReLU(),
 98 |         ),
 99 |         nn.Sequential( # Sequential,
100 |             LambdaMap(lambda x: x, # ConcatTable,
101 |                 nn.Sequential( # Sequential,
102 |                     nn.Conv2d(256,64,(1, 1),(1, 1),(0, 0),1,1,bias=False),
103 |                     nn.BatchNorm2d(64),
104 |                     nn.ReLU(),
105 |                     nn.Conv2d(64,64,(3, 3),(1, 1),(1, 1),1,1,bias=False),
106 |                     nn.BatchNorm2d(64),
107 |                     nn.ReLU(),
108 |                     nn.Conv2d(64,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
109 |                     nn.BatchNorm2d(256),
110 |                 ),
111 |                 Lambda(lambda x: x), # Identity,
112 |             ),
113 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
114 |             nn.ReLU(),
115 |         ),
116 |         nn.Sequential( # Sequential,
117 |             LambdaMap(lambda x: x, # ConcatTable,
118 |                 nn.Sequential( # Sequential,
119 |                     nn.Conv2d(256,64,(1, 1),(1, 1),(0, 0),1,1,bias=False),
120 |                     nn.BatchNorm2d(64),
121 |                     nn.ReLU(),
122 |                     nn.Conv2d(64,64,(3, 3),(1, 1),(1, 1),1,1,bias=False),
123 |                     nn.BatchNorm2d(64),
124 |                     nn.ReLU(),
125 |                     nn.Conv2d(64,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
126 |                     nn.BatchNorm2d(256),
127 |                 ),
128 |                 Lambda(lambda x: x), # Identity,
129 |             ),
130 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
131 |             nn.ReLU(),
132 |         ),
133 |         nn.Sequential( # Sequential,
134 |             LambdaMap(lambda x: x, # ConcatTable,
135 |                 nn.Sequential( # Sequential,
136 |                     nn.Conv2d(256,64,(1, 1),(1, 1),(0, 0),1,1,bias=False),
137 |                     nn.BatchNorm2d(64),
138 |                     nn.ReLU(),
139 |                     nn.Conv2d(64,64,(3, 3),(1, 1),(1, 1),1,1,bias=False),
140 |                     nn.BatchNorm2d(64),
141 |                     nn.ReLU(),
142 |                     nn.Conv2d(64,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
143 |                     nn.BatchNorm2d(256),
144 |                 ),
145 |                 Lambda(lambda x: x), # Identity,
146 |             ),
147 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
148 |             nn.ReLU(),
149 |         ),
150 |         nn.Sequential( # Sequential,
151 |             LambdaMap(lambda x: x, # ConcatTable,
152 |                 nn.Sequential( # Sequential,
153 |                     nn.Conv2d(256,64,(1, 1),(1, 1),(0, 0),1,1,bias=False),
154 |                     nn.BatchNorm2d(64),
155 |                     nn.ReLU(),
156 |                     nn.Conv2d(64,64,(3, 3),(1, 1),(1, 1),1,1,bias=False),
157 |                     nn.BatchNorm2d(64),
158 |                     nn.ReLU(),
159 |                     nn.Conv2d(64,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
160 |                     nn.BatchNorm2d(256),
161 |                 ),
162 |                 Lambda(lambda x: x), # Identity,
163 |             ),
164 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
165 |             nn.ReLU(),
166 |         ),
167 |         nn.Sequential( # Sequential,
168 |             LambdaMap(lambda x: x, # ConcatTable,
169 |                 nn.Sequential( # Sequential,
170 |                     nn.Conv2d(256,64,(1, 1),(1, 1),(0, 0),1,1,bias=False),
171 |                     nn.BatchNorm2d(64),
172 |                     nn.ReLU(),
173 |                     nn.Conv2d(64,64,(3, 3),(1, 1),(1, 1),1,1,bias=False),
174 |                     nn.BatchNorm2d(64),
175 |                     nn.ReLU(),
176 |                     nn.Conv2d(64,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
177 |                     nn.BatchNorm2d(256),
178 |                 ),
179 |                 Lambda(lambda x: x), # Identity,
180 |             ),
181 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
182 |             nn.ReLU(),
183 |         ),
184 |     ),
185 |     # --------------------- Conv3 1:12
186 |     nn.Sequential( # Sequential,
187 |         nn.Sequential( # Sequential,
188 |             LambdaMap(lambda x: x, # ConcatTable,
189 |                 nn.Sequential( # Sequential,
190 |                     nn.Conv2d(256,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
191 |                     nn.BatchNorm2d(128),
192 |                     nn.ReLU(),
193 |                     nn.Conv2d(128,128,(3, 3),(2, 2),(1, 1),1,1,bias=False),
194 |                     nn.BatchNorm2d(128),
195 |                     nn.ReLU(),
196 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
197 |                     nn.BatchNorm2d(512),
198 |                 ),
199 |                 nn.Sequential( # Sequential,
200 |                     nn.Conv2d(256,512,(1, 1),(2, 2),(0, 0),1,1,bias=False),
201 |                     nn.BatchNorm2d(512),
202 |                 ),
203 |             ),
204 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
205 |             nn.ReLU(),
206 |         ),
207 |         nn.Sequential( # Sequential,
208 |             LambdaMap(lambda x: x, # ConcatTable,
209 |                 nn.Sequential( # Sequential,
210 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
211 |                     nn.BatchNorm2d(128),
212 |                     nn.ReLU(),
213 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
214 |                     nn.BatchNorm2d(128),
215 |                     nn.ReLU(),
216 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
217 |                     nn.BatchNorm2d(512),
218 |                 ),
219 |                 Lambda(lambda x: x), # Identity,
220 |             ),
221 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
222 |             nn.ReLU(),
223 |         ),
224 |         nn.Sequential( # Sequential,
225 |             LambdaMap(lambda x: x, # ConcatTable,
226 |                 nn.Sequential( # Sequential,
227 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
228 |                     nn.BatchNorm2d(128),
229 |                     nn.ReLU(),
230 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
231 |                     nn.BatchNorm2d(128),
232 |                     nn.ReLU(),
233 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
234 |                     nn.BatchNorm2d(512),
235 |                 ),
236 |                 Lambda(lambda x: x), # Identity,
237 |             ),
238 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
239 |             nn.ReLU(),
240 |         ),
241 |         nn.Sequential( # Sequential,
242 |             LambdaMap(lambda x: x, # ConcatTable,
243 |                 nn.Sequential( # Sequential,
244 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
245 |                     nn.BatchNorm2d(128),
246 |                     nn.ReLU(),
247 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
248 |                     nn.BatchNorm2d(128),
249 |                     nn.ReLU(),
250 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
251 |                     nn.BatchNorm2d(512),
252 |                 ),
253 |                 Lambda(lambda x: x), # Identity,
254 |             ),
255 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
256 |             nn.ReLU(),
257 |         ),
258 |         nn.Sequential( # Sequential,
259 |             LambdaMap(lambda x: x, # ConcatTable,
260 |                 nn.Sequential( # Sequential,
261 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
262 |                     nn.BatchNorm2d(128),
263 |                     nn.ReLU(),
264 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
265 |                     nn.BatchNorm2d(128),
266 |                     nn.ReLU(),
267 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
268 |                     nn.BatchNorm2d(512),
269 |                 ),
270 |                 Lambda(lambda x: x), # Identity,
271 |             ),
272 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
273 |             nn.ReLU(),
274 |         ),
275 |         nn.Sequential( # Sequential,
276 |             LambdaMap(lambda x: x, # ConcatTable,
277 |                 nn.Sequential( # Sequential,
278 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
279 |                     nn.BatchNorm2d(128),
280 |                     nn.ReLU(),
281 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
282 |                     nn.BatchNorm2d(128),
283 |                     nn.ReLU(),
284 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
285 |                     nn.BatchNorm2d(512),
286 |                 ),
287 |                 Lambda(lambda x: x), # Identity,
288 |             ),
289 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
290 |             nn.ReLU(),
291 |         ),
292 |         nn.Sequential( # Sequential,
293 |             LambdaMap(lambda x: x, # ConcatTable,
294 |                 nn.Sequential( # Sequential,
295 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
296 |                     nn.BatchNorm2d(128),
297 |                     nn.ReLU(),
298 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
299 |                     nn.BatchNorm2d(128),
300 |                     nn.ReLU(),
301 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
302 |                     nn.BatchNorm2d(512),
303 |                 ),
304 |                 Lambda(lambda x: x), # Identity,
305 |             ),
306 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
307 |             nn.ReLU(),
308 |         ),
309 |         nn.Sequential( # Sequential,
310 |             LambdaMap(lambda x: x, # ConcatTable,
311 |                 nn.Sequential( # Sequential,
312 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
313 |                     nn.BatchNorm2d(128),
314 |                     nn.ReLU(),
315 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
316 |                     nn.BatchNorm2d(128),
317 |                     nn.ReLU(),
318 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
319 |                     nn.BatchNorm2d(512),
320 |                 ),
321 |                 Lambda(lambda x: x), # Identity,
322 |             ),
323 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
324 |             nn.ReLU(),
325 |         ),
326 |         nn.Sequential( # Sequential,
327 |             LambdaMap(lambda x: x, # ConcatTable,
328 |                 nn.Sequential( # Sequential,
329 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
330 |                     nn.BatchNorm2d(128),
331 |                     nn.ReLU(),
332 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
333 |                     nn.BatchNorm2d(128),
334 |                     nn.ReLU(),
335 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
336 |                     nn.BatchNorm2d(512),
337 |                 ),
338 |                 Lambda(lambda x: x), # Identity,
339 |             ),
340 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
341 |             nn.ReLU(),
342 |         ),
343 |         nn.Sequential( # Sequential,
344 |             LambdaMap(lambda x: x, # ConcatTable,
345 |                 nn.Sequential( # Sequential,
346 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
347 |                     nn.BatchNorm2d(128),
348 |                     nn.ReLU(),
349 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
350 |                     nn.BatchNorm2d(128),
351 |                     nn.ReLU(),
352 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
353 |                     nn.BatchNorm2d(512),
354 |                 ),
355 |                 Lambda(lambda x: x), # Identity,
356 |             ),
357 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
358 |             nn.ReLU(),
359 |         ),
360 |         nn.Sequential( # Sequential,
361 |             LambdaMap(lambda x: x, # ConcatTable,
362 |                 nn.Sequential( # Sequential,
363 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
364 |                     nn.BatchNorm2d(128),
365 |                     nn.ReLU(),
366 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
367 |                     nn.BatchNorm2d(128),
368 |                     nn.ReLU(),
369 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
370 |                     nn.BatchNorm2d(512),
371 |                 ),
372 |                 Lambda(lambda x: x), # Identity,
373 |             ),
374 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
375 |             nn.ReLU(),
376 |         ),
377 |         nn.Sequential( # Sequential,
378 |             LambdaMap(lambda x: x, # ConcatTable,
379 |                 nn.Sequential( # Sequential,
380 |                     nn.Conv2d(512,128,(1, 1),(1, 1),(0, 0),1,1,bias=False),
381 |                     nn.BatchNorm2d(128),
382 |                     nn.ReLU(),
383 |                     nn.Conv2d(128,128,(3, 3),(1, 1),(1, 1),1,1,bias=False),
384 |                     nn.BatchNorm2d(128),
385 |                     nn.ReLU(),
386 |                     nn.Conv2d(128,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
387 |                     nn.BatchNorm2d(512),
388 |                 ),
389 |                 Lambda(lambda x: x), # Identity,
390 |             ),
391 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
392 |             nn.ReLU(),
393 |         ),
394 |     ),
395 |     # --------------------- Conv4 1:23
396 |     nn.Sequential( # Sequential,
397 |         nn.Sequential( # Sequential,
398 |             LambdaMap(lambda x: x, # ConcatTable,
399 |                 nn.Sequential( # Sequential,
400 |                     nn.Conv2d(512,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
401 |                     nn.BatchNorm2d(256),
402 |                     nn.ReLU(),
403 |                     nn.Conv2d(256,256,(3, 3),(2, 2),(1, 1),1,1,bias=False),
404 |                     nn.BatchNorm2d(256),
405 |                     nn.ReLU(),
406 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
407 |                     nn.BatchNorm2d(1024),
408 |                 ),
409 |                 nn.Sequential( # Sequential,
410 |                     nn.Conv2d(512,1024,(1, 1),(2, 2),(0, 0),1,1,bias=False),
411 |                     nn.BatchNorm2d(1024),
412 |                 ),
413 |             ),
414 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
415 |             nn.ReLU(),
416 |         ),
417 |         nn.Sequential( # Sequential,
418 |             LambdaMap(lambda x: x, # ConcatTable,
419 |                 nn.Sequential( # Sequential,
420 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
421 |                     nn.BatchNorm2d(256),
422 |                     nn.ReLU(),
423 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
424 |                     nn.BatchNorm2d(256),
425 |                     nn.ReLU(),
426 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
427 |                     nn.BatchNorm2d(1024),
428 |                 ),
429 |                 Lambda(lambda x: x), # Identity,
430 |             ),
431 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
432 |             nn.ReLU(),
433 |         ),
434 |         nn.Sequential( # Sequential,
435 |             LambdaMap(lambda x: x, # ConcatTable,
436 |                 nn.Sequential( # Sequential,
437 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
438 |                     nn.BatchNorm2d(256),
439 |                     nn.ReLU(),
440 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
441 |                     nn.BatchNorm2d(256),
442 |                     nn.ReLU(),
443 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
444 |                     nn.BatchNorm2d(1024),
445 |                 ),
446 |                 Lambda(lambda x: x), # Identity,
447 |             ),
448 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
449 |             nn.ReLU(),
450 |         ),
451 |         nn.Sequential( # Sequential,
452 |             LambdaMap(lambda x: x, # ConcatTable,
453 |                 nn.Sequential( # Sequential,
454 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
455 |                     nn.BatchNorm2d(256),
456 |                     nn.ReLU(),
457 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
458 |                     nn.BatchNorm2d(256),
459 |                     nn.ReLU(),
460 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
461 |                     nn.BatchNorm2d(1024),
462 |                 ),
463 |                 Lambda(lambda x: x), # Identity,
464 |             ),
465 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
466 |             nn.ReLU(),
467 |         ),
468 |         nn.Sequential( # Sequential,
469 |             LambdaMap(lambda x: x, # ConcatTable,
470 |                 nn.Sequential( # Sequential,
471 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
472 |                     nn.BatchNorm2d(256),
473 |                     nn.ReLU(),
474 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
475 |                     nn.BatchNorm2d(256),
476 |                     nn.ReLU(),
477 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
478 |                     nn.BatchNorm2d(1024),
479 |                 ),
480 |                 Lambda(lambda x: x), # Identity,
481 |             ),
482 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
483 |             nn.ReLU(),
484 |         ),
485 |         nn.Sequential( # Sequential,
486 |             LambdaMap(lambda x: x, # ConcatTable,
487 |                 nn.Sequential( # Sequential,
488 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
489 |                     nn.BatchNorm2d(256),
490 |                     nn.ReLU(),
491 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
492 |                     nn.BatchNorm2d(256),
493 |                     nn.ReLU(),
494 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
495 |                     nn.BatchNorm2d(1024),
496 |                 ),
497 |                 Lambda(lambda x: x), # Identity,
498 |             ),
499 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
500 |             nn.ReLU(),
501 |         ),
502 |         nn.Sequential( # Sequential,
503 |             LambdaMap(lambda x: x, # ConcatTable,
504 |                 nn.Sequential( # Sequential,
505 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
506 |                     nn.BatchNorm2d(256),
507 |                     nn.ReLU(),
508 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
509 |                     nn.BatchNorm2d(256),
510 |                     nn.ReLU(),
511 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
512 |                     nn.BatchNorm2d(1024),
513 |                 ),
514 |                 Lambda(lambda x: x), # Identity,
515 |             ),
516 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
517 |             nn.ReLU(),
518 |         ),
519 |         nn.Sequential( # Sequential,
520 |             LambdaMap(lambda x: x, # ConcatTable,
521 |                 nn.Sequential( # Sequential,
522 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
523 |                     nn.BatchNorm2d(256),
524 |                     nn.ReLU(),
525 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
526 |                     nn.BatchNorm2d(256),
527 |                     nn.ReLU(),
528 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
529 |                     nn.BatchNorm2d(1024),
530 |                 ),
531 |                 Lambda(lambda x: x), # Identity,
532 |             ),
533 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
534 |             nn.ReLU(),
535 |         ),
536 |         nn.Sequential( # Sequential,
537 |             LambdaMap(lambda x: x, # ConcatTable,
538 |                 nn.Sequential( # Sequential,
539 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
540 |                     nn.BatchNorm2d(256),
541 |                     nn.ReLU(),
542 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
543 |                     nn.BatchNorm2d(256),
544 |                     nn.ReLU(),
545 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
546 |                     nn.BatchNorm2d(1024),
547 |                 ),
548 |                 Lambda(lambda x: x), # Identity,
549 |             ),
550 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
551 |             nn.ReLU(),
552 |         ),
553 |         nn.Sequential( # Sequential,
554 |             LambdaMap(lambda x: x, # ConcatTable,
555 |                 nn.Sequential( # Sequential,
556 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
557 |                     nn.BatchNorm2d(256),
558 |                     nn.ReLU(),
559 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
560 |                     nn.BatchNorm2d(256),
561 |                     nn.ReLU(),
562 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
563 |                     nn.BatchNorm2d(1024),
564 |                 ),
565 |                 Lambda(lambda x: x), # Identity,
566 |             ),
567 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
568 |             nn.ReLU(),
569 |         ),
570 |         nn.Sequential( # Sequential,
571 |             LambdaMap(lambda x: x, # ConcatTable,
572 |                 nn.Sequential( # Sequential,
573 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
574 |                     nn.BatchNorm2d(256),
575 |                     nn.ReLU(),
576 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
577 |                     nn.BatchNorm2d(256),
578 |                     nn.ReLU(),
579 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
580 |                     nn.BatchNorm2d(1024),
581 |                 ),
582 |                 Lambda(lambda x: x), # Identity,
583 |             ),
584 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
585 |             nn.ReLU(),
586 |         ),
587 |         nn.Sequential( # Sequential,
588 |             LambdaMap(lambda x: x, # ConcatTable,
589 |                 nn.Sequential( # Sequential,
590 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
591 |                     nn.BatchNorm2d(256),
592 |                     nn.ReLU(),
593 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
594 |                     nn.BatchNorm2d(256),
595 |                     nn.ReLU(),
596 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
597 |                     nn.BatchNorm2d(1024),
598 |                 ),
599 |                 Lambda(lambda x: x), # Identity,
600 |             ),
601 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
602 |             nn.ReLU(),
603 |         ),
604 |         nn.Sequential( # Sequential,
605 |             LambdaMap(lambda x: x, # ConcatTable,
606 |                 nn.Sequential( # Sequential,
607 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
608 |                     nn.BatchNorm2d(256),
609 |                     nn.ReLU(),
610 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
611 |                     nn.BatchNorm2d(256),
612 |                     nn.ReLU(),
613 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
614 |                     nn.BatchNorm2d(1024),
615 |                 ),
616 |                 Lambda(lambda x: x), # Identity,
617 |             ),
618 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
619 |             nn.ReLU(),
620 |         ),
621 |         nn.Sequential( # Sequential,
622 |             LambdaMap(lambda x: x, # ConcatTable,
623 |                 nn.Sequential( # Sequential,
624 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
625 |                     nn.BatchNorm2d(256),
626 |                     nn.ReLU(),
627 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
628 |                     nn.BatchNorm2d(256),
629 |                     nn.ReLU(),
630 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
631 |                     nn.BatchNorm2d(1024),
632 |                 ),
633 |                 Lambda(lambda x: x), # Identity,
634 |             ),
635 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
636 |             nn.ReLU(),
637 |         ),
638 |         nn.Sequential( # Sequential,
639 |             LambdaMap(lambda x: x, # ConcatTable,
640 |                 nn.Sequential( # Sequential,
641 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
642 |                     nn.BatchNorm2d(256),
643 |                     nn.ReLU(),
644 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
645 |                     nn.BatchNorm2d(256),
646 |                     nn.ReLU(),
647 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
648 |                     nn.BatchNorm2d(1024),
649 |                 ),
650 |                 Lambda(lambda x: x), # Identity,
651 |             ),
652 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
653 |             nn.ReLU(),
654 |         ),
655 |         nn.Sequential( # Sequential,
656 |             LambdaMap(lambda x: x, # ConcatTable,
657 |                 nn.Sequential( # Sequential,
658 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
659 |                     nn.BatchNorm2d(256),
660 |                     nn.ReLU(),
661 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
662 |                     nn.BatchNorm2d(256),
663 |                     nn.ReLU(),
664 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
665 |                     nn.BatchNorm2d(1024),
666 |                 ),
667 |                 Lambda(lambda x: x), # Identity,
668 |             ),
669 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
670 |             nn.ReLU(),
671 |         ),
672 |         nn.Sequential( # Sequential,
673 |             LambdaMap(lambda x: x, # ConcatTable,
674 |                 nn.Sequential( # Sequential,
675 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
676 |                     nn.BatchNorm2d(256),
677 |                     nn.ReLU(),
678 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
679 |                     nn.BatchNorm2d(256),
680 |                     nn.ReLU(),
681 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
682 |                     nn.BatchNorm2d(1024),
683 |                 ),
684 |                 Lambda(lambda x: x), # Identity,
685 |             ),
686 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
687 |             nn.ReLU(),
688 |         ),
689 |         nn.Sequential( # Sequential,
690 |             LambdaMap(lambda x: x, # ConcatTable,
691 |                 nn.Sequential( # Sequential,
692 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
693 |                     nn.BatchNorm2d(256),
694 |                     nn.ReLU(),
695 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
696 |                     nn.BatchNorm2d(256),
697 |                     nn.ReLU(),
698 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
699 |                     nn.BatchNorm2d(1024),
700 |                 ),
701 |                 Lambda(lambda x: x), # Identity,
702 |             ),
703 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
704 |             nn.ReLU(),
705 |         ),
706 |         nn.Sequential( # Sequential,
707 |             LambdaMap(lambda x: x, # ConcatTable,
708 |                 nn.Sequential( # Sequential,
709 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
710 |                     nn.BatchNorm2d(256),
711 |                     nn.ReLU(),
712 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
713 |                     nn.BatchNorm2d(256),
714 |                     nn.ReLU(),
715 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
716 |                     nn.BatchNorm2d(1024),
717 |                 ),
718 |                 Lambda(lambda x: x), # Identity,
719 |             ),
720 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
721 |             nn.ReLU(),
722 |         ),
723 |         nn.Sequential( # Sequential,
724 |             LambdaMap(lambda x: x, # ConcatTable,
725 |                 nn.Sequential( # Sequential,
726 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
727 |                     nn.BatchNorm2d(256),
728 |                     nn.ReLU(),
729 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
730 |                     nn.BatchNorm2d(256),
731 |                     nn.ReLU(),
732 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
733 |                     nn.BatchNorm2d(1024),
734 |                 ),
735 |                 Lambda(lambda x: x), # Identity,
736 |             ),
737 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
738 |             nn.ReLU(),
739 |         ),
740 |         nn.Sequential( # Sequential,
741 |             LambdaMap(lambda x: x, # ConcatTable,
742 |                 nn.Sequential( # Sequential,
743 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
744 |                     nn.BatchNorm2d(256),
745 |                     nn.ReLU(),
746 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
747 |                     nn.BatchNorm2d(256),
748 |                     nn.ReLU(),
749 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
750 |                     nn.BatchNorm2d(1024),
751 |                 ),
752 |                 Lambda(lambda x: x), # Identity,
753 |             ),
754 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
755 |             nn.ReLU(),
756 |         ),
757 |         nn.Sequential( # Sequential,
758 |             LambdaMap(lambda x: x, # ConcatTable,
759 |                 nn.Sequential( # Sequential,
760 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
761 |                     nn.BatchNorm2d(256),
762 |                     nn.ReLU(),
763 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
764 |                     nn.BatchNorm2d(256),
765 |                     nn.ReLU(),
766 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
767 |                     nn.BatchNorm2d(1024),
768 |                 ),
769 |                 Lambda(lambda x: x), # Identity,
770 |             ),
771 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
772 |             nn.ReLU(),
773 |         ),
774 |         nn.Sequential( # Sequential,
775 |             LambdaMap(lambda x: x, # ConcatTable,
776 |                 nn.Sequential( # Sequential,
777 |                     nn.Conv2d(1024,256,(1, 1),(1, 1),(0, 0),1,1,bias=False),
778 |                     nn.BatchNorm2d(256),
779 |                     nn.ReLU(),
780 |                     nn.Conv2d(256,256,(3, 3),(1, 1),(1, 1),1,1,bias=False),
781 |                     nn.BatchNorm2d(256),
782 |                     nn.ReLU(),
783 |                     nn.Conv2d(256,1024,(1, 1),(1, 1),(0, 0),1,1,bias=False),
784 |                     nn.BatchNorm2d(1024),
785 |                 ),
786 |                 Lambda(lambda x: x), # Identity,
787 |             ),
788 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
789 |             nn.ReLU(),
790 |         ),
791 |     ),
792 |     # --------------------- Conv5 1:3
793 |     nn.Sequential( # Sequential,
794 |         nn.Sequential( # Sequential,
795 |             LambdaMap(lambda x: x, # ConcatTable,
796 |                 nn.Sequential( # Sequential,
797 |                     nn.Conv2d(1024,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
798 |                     nn.BatchNorm2d(512),
799 |                     nn.ReLU(),
800 |                     nn.Conv2d(512,512,(3, 3),(2, 2),(1, 1),1,1,bias=False),
801 |                     nn.BatchNorm2d(512),
802 |                     nn.ReLU(),
803 |                     nn.Conv2d(512,2048,(1, 1),(1, 1),(0, 0),1,1,bias=False),
804 |                     nn.BatchNorm2d(2048),
805 |                 ),
806 |                 nn.Sequential( # Sequential,
807 |                     nn.Conv2d(1024,2048,(1, 1),(2, 2),(0, 0),1,1,bias=False),
808 |                     nn.BatchNorm2d(2048),
809 |                 ),
810 |             ),
811 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
812 |             nn.ReLU(),
813 |         ),
814 |         nn.Sequential( # Sequential,
815 |             LambdaMap(lambda x: x, # ConcatTable,
816 |                 nn.Sequential( # Sequential,
817 |                     nn.Conv2d(2048,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
818 |                     nn.BatchNorm2d(512),
819 |                     nn.ReLU(),
820 |                     nn.Conv2d(512,512,(3, 3),(1, 1),(1, 1),1,1,bias=False),
821 |                     nn.BatchNorm2d(512),
822 |                     nn.ReLU(),
823 |                     nn.Conv2d(512,2048,(1, 1),(1, 1),(0, 0),1,1,bias=False),
824 |                     nn.BatchNorm2d(2048),
825 |                 ),
826 |                 Lambda(lambda x: x), # Identity,
827 |             ),
828 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
829 |             nn.ReLU(),
830 |         ),
831 |         nn.Sequential( # Sequential,
832 |             LambdaMap(lambda x: x, # ConcatTable,
833 |                 nn.Sequential( # Sequential,
834 |                     nn.Conv2d(2048,512,(1, 1),(1, 1),(0, 0),1,1,bias=False),
835 |                     nn.BatchNorm2d(512),
836 |                     nn.ReLU(),
837 |                     nn.Conv2d(512,512,(3, 3),(1, 1),(1, 1),1,1,bias=False),
838 |                     nn.BatchNorm2d(512),
839 |                     nn.ReLU(),
840 |                     nn.Conv2d(512,2048,(1, 1),(1, 1),(0, 0),1,1,bias=False),
841 |                     nn.BatchNorm2d(2048),
842 |                 ),
843 |                 Lambda(lambda x: x), # Identity,
844 |             ),
845 |             LambdaReduce(lambda x,y: x+y), # CAddTable,
846 |             nn.ReLU(),
847 |         ),
848 |     ),
849 |     nn.AvgPool2d((7, 7),(1, 1)),
850 |     Lambda(lambda x: x.view(x.size(0),-1)), # View,
851 |     nn.Sequential(
852 |         Lambda(lambda x: x.view(1,-1) if 1==len(x.size()) else x ),
853 |         nn.Linear(2048,80)), # Linear,
854 | )


--------------------------------------------------------------------------------
/Places2-365-CNN.prototxt:
--------------------------------------------------------------------------------
   1 | name: "Places2-365-CNN"
   2 | input: "data"
   3 | input_dim: 10
   4 | input_dim: 3
   5 | input_dim: 224
   6 | input_dim: 224
   7 | layer {
   8 |   name: "conv1_1_3x3_s2"
   9 |   type: "Convolution"
  10 |   bottom: "data"
  11 |   top: "conv1_1_3x3_s2"
  12 |   convolution_param {
  13 |     num_output: 64
  14 |     bias_term: false
  15 |     pad: 1
  16 |     kernel_size: 3
  17 |     stride: 2
  18 |   }
  19 | }
  20 | layer {
  21 |   name: "conv1_1_3x3_s2/bn"
  22 |   type: "BatchNorm"
  23 |   bottom: "conv1_1_3x3_s2"
  24 |   top: "conv1_1_3x3_s2"
  25 |   batch_norm_param {
  26 |     use_global_stats: true
  27 |   }
  28 | }
  29 | layer {
  30 |   name: "conv1_1_3x3_s2/scale"
  31 |   type: "Scale"
  32 |   bottom: "conv1_1_3x3_s2"
  33 |   top: "conv1_1_3x3_s2"
  34 |   scale_param {
  35 |     bias_term: true
  36 |   }
  37 | }
  38 | layer {
  39 |   name: "conv1_1_3x3_s2/relu"
  40 |   type: "ReLU"
  41 |   bottom: "conv1_1_3x3_s2"
  42 |   top: "conv1_1_3x3_s2"
  43 | }
  44 | layer {
  45 |   name: "conv1_2_3x3"
  46 |   type: "Convolution"
  47 |   bottom: "conv1_1_3x3_s2"
  48 |   top: "conv1_2_3x3"
  49 |   convolution_param {
  50 |     num_output: 64
  51 |     bias_term: false
  52 |     pad: 1
  53 |     kernel_size: 3
  54 |     stride: 1
  55 |   }
  56 | }
  57 | layer {
  58 |   name: "conv1_2_3x3/bn"
  59 |   type: "BatchNorm"
  60 |   bottom: "conv1_2_3x3"
  61 |   top: "conv1_2_3x3"
  62 |   batch_norm_param {
  63 |     use_global_stats: true
  64 |   }
  65 | }
  66 | layer {
  67 |   name: "conv1_2_3x3/scale"
  68 |   type: "Scale"
  69 |   bottom: "conv1_2_3x3"
  70 |   top: "conv1_2_3x3"
  71 |   scale_param {
  72 |     bias_term: true
  73 |   }
  74 | }
  75 | layer {
  76 |   name: "conv1_2_3x3/relu"
  77 |   type: "ReLU"
  78 |   bottom: "conv1_2_3x3"
  79 |   top: "conv1_2_3x3"
  80 | }
  81 | layer {
  82 |   name: "conv1_3_3x3"
  83 |   type: "Convolution"
  84 |   bottom: "conv1_2_3x3"
  85 |   top: "conv1_3_3x3"
  86 |   convolution_param {
  87 |     num_output: 128
  88 |     bias_term: false
  89 |     pad: 1
  90 |     kernel_size: 3
  91 |     stride: 1
  92 |   }
  93 | }
  94 | layer {
  95 |   name: "conv1_3_3x3/bn"
  96 |   type: "BatchNorm"
  97 |   bottom: "conv1_3_3x3"
  98 |   top: "conv1_3_3x3"
  99 |   batch_norm_param {
 100 |     use_global_stats: true
 101 |   }
 102 | }
 103 | layer {
 104 |   name: "conv1_3_3x3/scale"
 105 |   type: "Scale"
 106 |   bottom: "conv1_3_3x3"
 107 |   top: "conv1_3_3x3"
 108 |   scale_param {
 109 |     bias_term: true
 110 |   }
 111 | }
 112 | layer {
 113 |   name: "conv1_3_3x3/relu"
 114 |   type: "ReLU"
 115 |   bottom: "conv1_3_3x3"
 116 |   top: "conv1_3_3x3"
 117 | }
 118 | layer {
 119 |   name: "pool1_3x3_s2"
 120 |   type: "Pooling"
 121 |   bottom: "conv1_3_3x3"
 122 |   top: "pool1_3x3_s2"
 123 |   pooling_param {
 124 |     pool: MAX
 125 |     kernel_size: 3
 126 |     stride: 2
 127 |   }
 128 | }
 129 | layer {
 130 |   name: "conv2_1_1x1_reduce"
 131 |   type: "Convolution"
 132 |   bottom: "pool1_3x3_s2"
 133 |   top: "conv2_1_1x1_reduce"
 134 |   convolution_param {
 135 |     num_output: 64
 136 |     bias_term: false
 137 |     pad: 0
 138 |     kernel_size: 1
 139 |     stride: 1
 140 |   }
 141 | }
 142 | layer {
 143 |   name: "conv2_1_1x1_reduce/bn"
 144 |   type: "BatchNorm"
 145 |   bottom: "conv2_1_1x1_reduce"
 146 |   top: "conv2_1_1x1_reduce"
 147 |   batch_norm_param {
 148 |     use_global_stats: true
 149 |   }
 150 | }
 151 | layer {
 152 |   name: "conv2_1_1x1_reduce/scale"
 153 |   type: "Scale"
 154 |   bottom: "conv2_1_1x1_reduce"
 155 |   top: "conv2_1_1x1_reduce"
 156 |   scale_param {
 157 |     bias_term: true
 158 |   }
 159 | }
 160 | layer {
 161 |   name: "conv2_1_1x1_reduce/relu"
 162 |   type: "ReLU"
 163 |   bottom: "conv2_1_1x1_reduce"
 164 |   top: "conv2_1_1x1_reduce"
 165 | }
 166 | layer {
 167 |   name: "conv2_1_3x3"
 168 |   type: "Convolution"
 169 |   bottom: "conv2_1_1x1_reduce"
 170 |   top: "conv2_1_3x3"
 171 |   convolution_param {
 172 |     num_output: 64
 173 |     bias_term: false
 174 |     pad: 1
 175 |     kernel_size: 3
 176 |     stride: 1
 177 |   }
 178 | }
 179 | layer {
 180 |   name: "conv2_1_3x3/bn"
 181 |   type: "BatchNorm"
 182 |   bottom: "conv2_1_3x3"
 183 |   top: "conv2_1_3x3"
 184 |   batch_norm_param {
 185 |     use_global_stats: true
 186 |   }
 187 | }
 188 | layer {
 189 |   name: "conv2_1_3x3/scale"
 190 |   type: "Scale"
 191 |   bottom: "conv2_1_3x3"
 192 |   top: "conv2_1_3x3"
 193 |   scale_param {
 194 |     bias_term: true
 195 |   }
 196 | }
 197 | layer {
 198 |   name: "conv2_1_3x3/relu"
 199 |   type: "ReLU"
 200 |   bottom: "conv2_1_3x3"
 201 |   top: "conv2_1_3x3"
 202 | }
 203 | layer {
 204 |   name: "conv2_1_1x1_increase"
 205 |   type: "Convolution"
 206 |   bottom: "conv2_1_3x3"
 207 |   top: "conv2_1_1x1_increase"
 208 |   convolution_param {
 209 |     num_output: 256
 210 |     bias_term: false
 211 |     pad: 0
 212 |     kernel_size: 1
 213 |     stride: 1
 214 |   }
 215 | }
 216 | layer {
 217 |   name: "conv2_1_1x1_increase/bn"
 218 |   type: "BatchNorm"
 219 |   bottom: "conv2_1_1x1_increase"
 220 |   top: "conv2_1_1x1_increase"
 221 |   batch_norm_param {
 222 |     use_global_stats: true
 223 |   }
 224 | }
 225 | layer {
 226 |   name: "conv2_1_1x1_increase/scale"
 227 |   type: "Scale"
 228 |   bottom: "conv2_1_1x1_increase"
 229 |   top: "conv2_1_1x1_increase"
 230 |   scale_param {
 231 |     bias_term: true
 232 |   }
 233 | }
 234 | layer {
 235 |   name: "conv2_1_1x1_proj"
 236 |   type: "Convolution"
 237 |   bottom: "pool1_3x3_s2"
 238 |   top: "conv2_1_1x1_proj"
 239 |   convolution_param {
 240 |     num_output: 256
 241 |     bias_term: false
 242 |     pad: 0
 243 |     kernel_size: 1
 244 |     stride: 1
 245 |   }
 246 | }
 247 | layer {
 248 |   name: "conv2_1_1x1_proj/bn"
 249 |   type: "BatchNorm"
 250 |   bottom: "conv2_1_1x1_proj"
 251 |   top: "conv2_1_1x1_proj"
 252 |   batch_norm_param {
 253 |     use_global_stats: true
 254 |   }
 255 | }
 256 | layer {
 257 |   name: "conv2_1_1x1_proj/scale"
 258 |   type: "Scale"
 259 |   bottom: "conv2_1_1x1_proj"
 260 |   top: "conv2_1_1x1_proj"
 261 |   scale_param {
 262 |     bias_term: true
 263 |   }
 264 | }
 265 | layer {
 266 |   name: "conv2_1"
 267 |   type: "Eltwise"
 268 |   bottom: "conv2_1_1x1_proj"
 269 |   bottom: "conv2_1_1x1_increase"
 270 |   top: "conv2_1"
 271 |   eltwise_param {
 272 |     operation: SUM
 273 |   }
 274 | }
 275 | layer {
 276 |   name: "conv2_1/relu"
 277 |   type: "ReLU"
 278 |   bottom: "conv2_1"
 279 |   top: "conv2_1"
 280 | }
 281 | layer {
 282 |   name: "conv2_2_1x1_reduce"
 283 |   type: "Convolution"
 284 |   bottom: "conv2_1"
 285 |   top: "conv2_2_1x1_reduce"
 286 |   convolution_param {
 287 |     num_output: 64
 288 |     bias_term: false
 289 |     pad: 0
 290 |     kernel_size: 1
 291 |     stride: 1
 292 |   }
 293 | }
 294 | layer {
 295 |   name: "conv2_2_1x1_reduce/bn"
 296 |   type: "BatchNorm"
 297 |   bottom: "conv2_2_1x1_reduce"
 298 |   top: "conv2_2_1x1_reduce"
 299 |   batch_norm_param {
 300 |     use_global_stats: true
 301 |   }
 302 | }
 303 | layer {
 304 |   name: "conv2_2_1x1_reduce/scale"
 305 |   type: "Scale"
 306 |   bottom: "conv2_2_1x1_reduce"
 307 |   top: "conv2_2_1x1_reduce"
 308 |   scale_param {
 309 |     bias_term: true
 310 |   }
 311 | }
 312 | layer {
 313 |   name: "conv2_2_1x1_reduce/relu"
 314 |   type: "ReLU"
 315 |   bottom: "conv2_2_1x1_reduce"
 316 |   top: "conv2_2_1x1_reduce"
 317 | }
 318 | layer {
 319 |   name: "conv2_2_3x3"
 320 |   type: "Convolution"
 321 |   bottom: "conv2_2_1x1_reduce"
 322 |   top: "conv2_2_3x3"
 323 |   convolution_param {
 324 |     num_output: 64
 325 |     bias_term: false
 326 |     pad: 1
 327 |     kernel_size: 3
 328 |     stride: 1
 329 |   }
 330 | }
 331 | layer {
 332 |   name: "conv2_2_3x3/bn"
 333 |   type: "BatchNorm"
 334 |   bottom: "conv2_2_3x3"
 335 |   top: "conv2_2_3x3"
 336 |   batch_norm_param {
 337 |     use_global_stats: true
 338 |   }
 339 | }
 340 | layer {
 341 |   name: "conv2_2_3x3/scale"
 342 |   type: "Scale"
 343 |   bottom: "conv2_2_3x3"
 344 |   top: "conv2_2_3x3"
 345 |   scale_param {
 346 |     bias_term: true
 347 |   }
 348 | }
 349 | layer {
 350 |   name: "conv2_2_3x3/relu"
 351 |   type: "ReLU"
 352 |   bottom: "conv2_2_3x3"
 353 |   top: "conv2_2_3x3"
 354 | }
 355 | layer {
 356 |   name: "conv2_2_1x1_increase"
 357 |   type: "Convolution"
 358 |   bottom: "conv2_2_3x3"
 359 |   top: "conv2_2_1x1_increase"
 360 |   convolution_param {
 361 |     num_output: 256
 362 |     bias_term: false
 363 |     pad: 0
 364 |     kernel_size: 1
 365 |     stride: 1
 366 |   }
 367 | }
 368 | layer {
 369 |   name: "conv2_2_1x1_increase/bn"
 370 |   type: "BatchNorm"
 371 |   bottom: "conv2_2_1x1_increase"
 372 |   top: "conv2_2_1x1_increase"
 373 |   batch_norm_param {
 374 |     use_global_stats: true
 375 |   }
 376 | }
 377 | layer {
 378 |   name: "conv2_2_1x1_increase/scale"
 379 |   type: "Scale"
 380 |   bottom: "conv2_2_1x1_increase"
 381 |   top: "conv2_2_1x1_increase"
 382 |   scale_param {
 383 |     bias_term: true
 384 |   }
 385 | }
 386 | layer {
 387 |   name: "conv2_2"
 388 |   type: "Eltwise"
 389 |   bottom: "conv2_1"
 390 |   bottom: "conv2_2_1x1_increase"
 391 |   top: "conv2_2"
 392 |   eltwise_param {
 393 |     operation: SUM
 394 |   }
 395 | }
 396 | layer {
 397 |   name: "conv2_2/relu"
 398 |   type: "ReLU"
 399 |   bottom: "conv2_2"
 400 |   top: "conv2_2"
 401 | }
 402 | layer {
 403 |   name: "conv2_3_1x1_reduce"
 404 |   type: "Convolution"
 405 |   bottom: "conv2_2"
 406 |   top: "conv2_3_1x1_reduce"
 407 |   convolution_param {
 408 |     num_output: 64
 409 |     bias_term: false
 410 |     pad: 0
 411 |     kernel_size: 1
 412 |     stride: 1
 413 |   }
 414 | }
 415 | layer {
 416 |   name: "conv2_3_1x1_reduce/bn"
 417 |   type: "BatchNorm"
 418 |   bottom: "conv2_3_1x1_reduce"
 419 |   top: "conv2_3_1x1_reduce"
 420 |   batch_norm_param {
 421 |     use_global_stats: true
 422 |   }
 423 | }
 424 | layer {
 425 |   name: "conv2_3_1x1_reduce/scale"
 426 |   type: "Scale"
 427 |   bottom: "conv2_3_1x1_reduce"
 428 |   top: "conv2_3_1x1_reduce"
 429 |   scale_param {
 430 |     bias_term: true
 431 |   }
 432 | }
 433 | layer {
 434 |   name: "conv2_3_1x1_reduce/relu"
 435 |   type: "ReLU"
 436 |   bottom: "conv2_3_1x1_reduce"
 437 |   top: "conv2_3_1x1_reduce"
 438 | }
 439 | layer {
 440 |   name: "conv2_3_3x3"
 441 |   type: "Convolution"
 442 |   bottom: "conv2_3_1x1_reduce"
 443 |   top: "conv2_3_3x3"
 444 |   convolution_param {
 445 |     num_output: 64
 446 |     bias_term: false
 447 |     pad: 1
 448 |     kernel_size: 3
 449 |     stride: 1
 450 |   }
 451 | }
 452 | layer {
 453 |   name: "conv2_3_3x3/bn"
 454 |   type: "BatchNorm"
 455 |   bottom: "conv2_3_3x3"
 456 |   top: "conv2_3_3x3"
 457 |   batch_norm_param {
 458 |     use_global_stats: true
 459 |   }
 460 | }
 461 | layer {
 462 |   name: "conv2_3_3x3/scale"
 463 |   type: "Scale"
 464 |   bottom: "conv2_3_3x3"
 465 |   top: "conv2_3_3x3"
 466 |   scale_param {
 467 |     bias_term: true
 468 |   }
 469 | }
 470 | layer {
 471 |   name: "conv2_3_3x3/relu"
 472 |   type: "ReLU"
 473 |   bottom: "conv2_3_3x3"
 474 |   top: "conv2_3_3x3"
 475 | }
 476 | layer {
 477 |   name: "conv2_3_1x1_increase"
 478 |   type: "Convolution"
 479 |   bottom: "conv2_3_3x3"
 480 |   top: "conv2_3_1x1_increase"
 481 |   convolution_param {
 482 |     num_output: 256
 483 |     bias_term: false
 484 |     pad: 0
 485 |     kernel_size: 1
 486 |     stride: 1
 487 |   }
 488 | }
 489 | layer {
 490 |   name: "conv2_3_1x1_increase/bn"
 491 |   type: "BatchNorm"
 492 |   bottom: "conv2_3_1x1_increase"
 493 |   top: "conv2_3_1x1_increase"
 494 |   batch_norm_param {
 495 |     use_global_stats: true
 496 |   }
 497 | }
 498 | layer {
 499 |   name: "conv2_3_1x1_increase/scale"
 500 |   type: "Scale"
 501 |   bottom: "conv2_3_1x1_increase"
 502 |   top: "conv2_3_1x1_increase"
 503 |   scale_param {
 504 |     bias_term: true
 505 |   }
 506 | }
 507 | layer {
 508 |   name: "conv2_3"
 509 |   type: "Eltwise"
 510 |   bottom: "conv2_2"
 511 |   bottom: "conv2_3_1x1_increase"
 512 |   top: "conv2_3"
 513 |   eltwise_param {
 514 |     operation: SUM
 515 |   }
 516 | }
 517 | layer {
 518 |   name: "conv2_3/relu"
 519 |   type: "ReLU"
 520 |   bottom: "conv2_3"
 521 |   top: "conv2_3"
 522 | }
 523 | layer {
 524 |   name: "conv2_4_1x1_reduce"
 525 |   type: "Convolution"
 526 |   bottom: "conv2_3"
 527 |   top: "conv2_4_1x1_reduce"
 528 |   convolution_param {
 529 |     num_output: 64
 530 |     bias_term: false
 531 |     pad: 0
 532 |     kernel_size: 1
 533 |     stride: 1
 534 |   }
 535 | }
 536 | layer {
 537 |   name: "conv2_4_1x1_reduce/bn"
 538 |   type: "BatchNorm"
 539 |   bottom: "conv2_4_1x1_reduce"
 540 |   top: "conv2_4_1x1_reduce"
 541 |   batch_norm_param {
 542 |     use_global_stats: true
 543 |   }
 544 | }
 545 | layer {
 546 |   name: "conv2_4_1x1_reduce/scale"
 547 |   type: "Scale"
 548 |   bottom: "conv2_4_1x1_reduce"
 549 |   top: "conv2_4_1x1_reduce"
 550 |   scale_param {
 551 |     bias_term: true
 552 |   }
 553 | }
 554 | layer {
 555 |   name: "conv2_4_1x1_reduce/relu"
 556 |   type: "ReLU"
 557 |   bottom: "conv2_4_1x1_reduce"
 558 |   top: "conv2_4_1x1_reduce"
 559 | }
 560 | layer {
 561 |   name: "conv2_4_3x3"
 562 |   type: "Convolution"
 563 |   bottom: "conv2_4_1x1_reduce"
 564 |   top: "conv2_4_3x3"
 565 |   convolution_param {
 566 |     num_output: 64
 567 |     bias_term: false
 568 |     pad: 1
 569 |     kernel_size: 3
 570 |     stride: 1
 571 |   }
 572 | }
 573 | layer {
 574 |   name: "conv2_4_3x3/bn"
 575 |   type: "BatchNorm"
 576 |   bottom: "conv2_4_3x3"
 577 |   top: "conv2_4_3x3"
 578 |   batch_norm_param {
 579 |     use_global_stats: true
 580 |   }
 581 | }
 582 | layer {
 583 |   name: "conv2_4_3x3/scale"
 584 |   type: "Scale"
 585 |   bottom: "conv2_4_3x3"
 586 |   top: "conv2_4_3x3"
 587 |   scale_param {
 588 |     bias_term: true
 589 |   }
 590 | }
 591 | layer {
 592 |   name: "conv2_4_3x3/relu"
 593 |   type: "ReLU"
 594 |   bottom: "conv2_4_3x3"
 595 |   top: "conv2_4_3x3"
 596 | }
 597 | layer {
 598 |   name: "conv2_4_1x1_increase"
 599 |   type: "Convolution"
 600 |   bottom: "conv2_4_3x3"
 601 |   top: "conv2_4_1x1_increase"
 602 |   convolution_param {
 603 |     num_output: 256
 604 |     bias_term: false
 605 |     pad: 0
 606 |     kernel_size: 1
 607 |     stride: 1
 608 |   }
 609 | }
 610 | layer {
 611 |   name: "conv2_4_1x1_increase/bn"
 612 |   type: "BatchNorm"
 613 |   bottom: "conv2_4_1x1_increase"
 614 |   top: "conv2_4_1x1_increase"
 615 |   batch_norm_param {
 616 |     use_global_stats: true
 617 |   }
 618 | }
 619 | layer {
 620 |   name: "conv2_4_1x1_increase/scale"
 621 |   type: "Scale"
 622 |   bottom: "conv2_4_1x1_increase"
 623 |   top: "conv2_4_1x1_increase"
 624 |   scale_param {
 625 |     bias_term: true
 626 |   }
 627 | }
 628 | layer {
 629 |   name: "conv2_4"
 630 |   type: "Eltwise"
 631 |   bottom: "conv2_3"
 632 |   bottom: "conv2_4_1x1_increase"
 633 |   top: "conv2_4"
 634 |   eltwise_param {
 635 |     operation: SUM
 636 |   }
 637 | }
 638 | layer {
 639 |   name: "conv2_4/relu"
 640 |   type: "ReLU"
 641 |   bottom: "conv2_4"
 642 |   top: "conv2_4"
 643 | }
 644 | layer {
 645 |   name: "conv2_5_1x1_reduce"
 646 |   type: "Convolution"
 647 |   bottom: "conv2_4"
 648 |   top: "conv2_5_1x1_reduce"
 649 |   convolution_param {
 650 |     num_output: 64
 651 |     bias_term: false
 652 |     pad: 0
 653 |     kernel_size: 1
 654 |     stride: 1
 655 |   }
 656 | }
 657 | layer {
 658 |   name: "conv2_5_1x1_reduce/bn"
 659 |   type: "BatchNorm"
 660 |   bottom: "conv2_5_1x1_reduce"
 661 |   top: "conv2_5_1x1_reduce"
 662 |   batch_norm_param {
 663 |     use_global_stats: true
 664 |   }
 665 | }
 666 | layer {
 667 |   name: "conv2_5_1x1_reduce/scale"
 668 |   type: "Scale"
 669 |   bottom: "conv2_5_1x1_reduce"
 670 |   top: "conv2_5_1x1_reduce"
 671 |   scale_param {
 672 |     bias_term: true
 673 |   }
 674 | }
 675 | layer {
 676 |   name: "conv2_5_1x1_reduce/relu"
 677 |   type: "ReLU"
 678 |   bottom: "conv2_5_1x1_reduce"
 679 |   top: "conv2_5_1x1_reduce"
 680 | }
 681 | layer {
 682 |   name: "conv2_5_3x3"
 683 |   type: "Convolution"
 684 |   bottom: "conv2_5_1x1_reduce"
 685 |   top: "conv2_5_3x3"
 686 |   convolution_param {
 687 |     num_output: 64
 688 |     bias_term: false
 689 |     pad: 1
 690 |     kernel_size: 3
 691 |     stride: 1
 692 |   }
 693 | }
 694 | layer {
 695 |   name: "conv2_5_3x3/bn"
 696 |   type: "BatchNorm"
 697 |   bottom: "conv2_5_3x3"
 698 |   top: "conv2_5_3x3"
 699 |   batch_norm_param {
 700 |     use_global_stats: true
 701 |   }
 702 | }
 703 | layer {
 704 |   name: "conv2_5_3x3/scale"
 705 |   type: "Scale"
 706 |   bottom: "conv2_5_3x3"
 707 |   top: "conv2_5_3x3"
 708 |   scale_param {
 709 |     bias_term: true
 710 |   }
 711 | }
 712 | layer {
 713 |   name: "conv2_5_3x3/relu"
 714 |   type: "ReLU"
 715 |   bottom: "conv2_5_3x3"
 716 |   top: "conv2_5_3x3"
 717 | }
 718 | layer {
 719 |   name: "conv2_5_1x1_increase"
 720 |   type: "Convolution"
 721 |   bottom: "conv2_5_3x3"
 722 |   top: "conv2_5_1x1_increase"
 723 |   convolution_param {
 724 |     num_output: 256
 725 |     bias_term: false
 726 |     pad: 0
 727 |     kernel_size: 1
 728 |     stride: 1
 729 |   }
 730 | }
 731 | layer {
 732 |   name: "conv2_5_1x1_increase/bn"
 733 |   type: "BatchNorm"
 734 |   bottom: "conv2_5_1x1_increase"
 735 |   top: "conv2_5_1x1_increase"
 736 |   batch_norm_param {
 737 |     use_global_stats: true
 738 |   }
 739 | }
 740 | layer {
 741 |   name: "conv2_5_1x1_increase/scale"
 742 |   type: "Scale"
 743 |   bottom: "conv2_5_1x1_increase"
 744 |   top: "conv2_5_1x1_increase"
 745 |   scale_param {
 746 |     bias_term: true
 747 |   }
 748 | }
 749 | layer {
 750 |   name: "conv2_5"
 751 |   type: "Eltwise"
 752 |   bottom: "conv2_4"
 753 |   bottom: "conv2_5_1x1_increase"
 754 |   top: "conv2_5"
 755 |   eltwise_param {
 756 |     operation: SUM
 757 |   }
 758 | }
 759 | layer {
 760 |   name: "conv2_5/relu"
 761 |   type: "ReLU"
 762 |   bottom: "conv2_5"
 763 |   top: "conv2_5"
 764 | }
 765 | layer {
 766 |   name: "conv2_6_1x1_reduce"
 767 |   type: "Convolution"
 768 |   bottom: "conv2_5"
 769 |   top: "conv2_6_1x1_reduce"
 770 |   convolution_param {
 771 |     num_output: 64
 772 |     bias_term: false
 773 |     pad: 0
 774 |     kernel_size: 1
 775 |     stride: 1
 776 |   }
 777 | }
 778 | layer {
 779 |   name: "conv2_6_1x1_reduce/bn"
 780 |   type: "BatchNorm"
 781 |   bottom: "conv2_6_1x1_reduce"
 782 |   top: "conv2_6_1x1_reduce"
 783 |   batch_norm_param {
 784 |     use_global_stats: true
 785 |   }
 786 | }
 787 | layer {
 788 |   name: "conv2_6_1x1_reduce/scale"
 789 |   type: "Scale"
 790 |   bottom: "conv2_6_1x1_reduce"
 791 |   top: "conv2_6_1x1_reduce"
 792 |   scale_param {
 793 |     bias_term: true
 794 |   }
 795 | }
 796 | layer {
 797 |   name: "conv2_6_1x1_reduce/relu"
 798 |   type: "ReLU"
 799 |   bottom: "conv2_6_1x1_reduce"
 800 |   top: "conv2_6_1x1_reduce"
 801 | }
 802 | layer {
 803 |   name: "conv2_6_3x3"
 804 |   type: "Convolution"
 805 |   bottom: "conv2_6_1x1_reduce"
 806 |   top: "conv2_6_3x3"
 807 |   convolution_param {
 808 |     num_output: 64
 809 |     bias_term: false
 810 |     pad: 1
 811 |     kernel_size: 3
 812 |     stride: 1
 813 |   }
 814 | }
 815 | layer {
 816 |   name: "conv2_6_3x3/bn"
 817 |   type: "BatchNorm"
 818 |   bottom: "conv2_6_3x3"
 819 |   top: "conv2_6_3x3"
 820 |   batch_norm_param {
 821 |     use_global_stats: true
 822 |   }
 823 | }
 824 | layer {
 825 |   name: "conv2_6_3x3/scale"
 826 |   type: "Scale"
 827 |   bottom: "conv2_6_3x3"
 828 |   top: "conv2_6_3x3"
 829 |   scale_param {
 830 |     bias_term: true
 831 |   }
 832 | }
 833 | layer {
 834 |   name: "conv2_6_3x3/relu"
 835 |   type: "ReLU"
 836 |   bottom: "conv2_6_3x3"
 837 |   top: "conv2_6_3x3"
 838 | }
 839 | layer {
 840 |   name: "conv2_6_1x1_increase"
 841 |   type: "Convolution"
 842 |   bottom: "conv2_6_3x3"
 843 |   top: "conv2_6_1x1_increase"
 844 |   convolution_param {
 845 |     num_output: 256
 846 |     bias_term: false
 847 |     pad: 0
 848 |     kernel_size: 1
 849 |     stride: 1
 850 |   }
 851 | }
 852 | layer {
 853 |   name: "conv2_6_1x1_increase/bn"
 854 |   type: "BatchNorm"
 855 |   bottom: "conv2_6_1x1_increase"
 856 |   top: "conv2_6_1x1_increase"
 857 |   batch_norm_param {
 858 |     use_global_stats: true
 859 |   }
 860 | }
 861 | layer {
 862 |   name: "conv2_6_1x1_increase/scale"
 863 |   type: "Scale"
 864 |   bottom: "conv2_6_1x1_increase"
 865 |   top: "conv2_6_1x1_increase"
 866 |   scale_param {
 867 |     bias_term: true
 868 |   }
 869 | }
 870 | layer {
 871 |   name: "conv2_6"
 872 |   type: "Eltwise"
 873 |   bottom: "conv2_5"
 874 |   bottom: "conv2_6_1x1_increase"
 875 |   top: "conv2_6"
 876 |   eltwise_param {
 877 |     operation: SUM
 878 |   }
 879 | }
 880 | layer {
 881 |   name: "conv2_6/relu"
 882 |   type: "ReLU"
 883 |   bottom: "conv2_6"
 884 |   top: "conv2_6"
 885 | }
 886 | layer {
 887 |   name: "conv2_7_1x1_reduce"
 888 |   type: "Convolution"
 889 |   bottom: "conv2_6"
 890 |   top: "conv2_7_1x1_reduce"
 891 |   convolution_param {
 892 |     num_output: 64
 893 |     bias_term: false
 894 |     pad: 0
 895 |     kernel_size: 1
 896 |     stride: 1
 897 |   }
 898 | }
 899 | layer {
 900 |   name: "conv2_7_1x1_reduce/bn"
 901 |   type: "BatchNorm"
 902 |   bottom: "conv2_7_1x1_reduce"
 903 |   top: "conv2_7_1x1_reduce"
 904 |   batch_norm_param {
 905 |     use_global_stats: true
 906 |   }
 907 | }
 908 | layer {
 909 |   name: "conv2_7_1x1_reduce/scale"
 910 |   type: "Scale"
 911 |   bottom: "conv2_7_1x1_reduce"
 912 |   top: "conv2_7_1x1_reduce"
 913 |   scale_param {
 914 |     bias_term: true
 915 |   }
 916 | }
 917 | layer {
 918 |   name: "conv2_7_1x1_reduce/relu"
 919 |   type: "ReLU"
 920 |   bottom: "conv2_7_1x1_reduce"
 921 |   top: "conv2_7_1x1_reduce"
 922 | }
 923 | layer {
 924 |   name: "conv2_7_3x3"
 925 |   type: "Convolution"
 926 |   bottom: "conv2_7_1x1_reduce"
 927 |   top: "conv2_7_3x3"
 928 |   convolution_param {
 929 |     num_output: 64
 930 |     bias_term: false
 931 |     pad: 1
 932 |     kernel_size: 3
 933 |     stride: 1
 934 |   }
 935 | }
 936 | layer {
 937 |   name: "conv2_7_3x3/bn"
 938 |   type: "BatchNorm"
 939 |   bottom: "conv2_7_3x3"
 940 |   top: "conv2_7_3x3"
 941 |   batch_norm_param {
 942 |     use_global_stats: true
 943 |   }
 944 | }
 945 | layer {
 946 |   name: "conv2_7_3x3/scale"
 947 |   type: "Scale"
 948 |   bottom: "conv2_7_3x3"
 949 |   top: "conv2_7_3x3"
 950 |   scale_param {
 951 |     bias_term: true
 952 |   }
 953 | }
 954 | layer {
 955 |   name: "conv2_7_3x3/relu"
 956 |   type: "ReLU"
 957 |   bottom: "conv2_7_3x3"
 958 |   top: "conv2_7_3x3"
 959 | }
 960 | layer {
 961 |   name: "conv2_7_1x1_increase"
 962 |   type: "Convolution"
 963 |   bottom: "conv2_7_3x3"
 964 |   top: "conv2_7_1x1_increase"
 965 |   convolution_param {
 966 |     num_output: 256
 967 |     bias_term: false
 968 |     pad: 0
 969 |     kernel_size: 1
 970 |     stride: 1
 971 |   }
 972 | }
 973 | layer {
 974 |   name: "conv2_7_1x1_increase/bn"
 975 |   type: "BatchNorm"
 976 |   bottom: "conv2_7_1x1_increase"
 977 |   top: "conv2_7_1x1_increase"
 978 |   batch_norm_param {
 979 |     use_global_stats: true
 980 |   }
 981 | }
 982 | layer {
 983 |   name: "conv2_7_1x1_increase/scale"
 984 |   type: "Scale"
 985 |   bottom: "conv2_7_1x1_increase"
 986 |   top: "conv2_7_1x1_increase"
 987 |   scale_param {
 988 |     bias_term: true
 989 |   }
 990 | }
 991 | layer {
 992 |   name: "conv2_7"
 993 |   type: "Eltwise"
 994 |   bottom: "conv2_6"
 995 |   bottom: "conv2_7_1x1_increase"
 996 |   top: "conv2_7"
 997 |   eltwise_param {
 998 |     operation: SUM
 999 |   }
1000 | }
1001 | layer {
1002 |   name: "conv2_7/relu"
1003 |   type: "ReLU"
1004 |   bottom: "conv2_7"
1005 |   top: "conv2_7"
1006 | }
1007 | layer {
1008 |   name: "conv2_8_1x1_reduce"
1009 |   type: "Convolution"
1010 |   bottom: "conv2_7"
1011 |   top: "conv2_8_1x1_reduce"
1012 |   convolution_param {
1013 |     num_output: 64
1014 |     bias_term: false
1015 |     pad: 0
1016 |     kernel_size: 1
1017 |     stride: 1
1018 |   }
1019 | }
1020 | layer {
1021 |   name: "conv2_8_1x1_reduce/bn"
1022 |   type: "BatchNorm"
1023 |   bottom: "conv2_8_1x1_reduce"
1024 |   top: "conv2_8_1x1_reduce"
1025 |   batch_norm_param {
1026 |     use_global_stats: true
1027 |   }
1028 | }
1029 | layer {
1030 |   name: "conv2_8_1x1_reduce/scale"
1031 |   type: "Scale"
1032 |   bottom: "conv2_8_1x1_reduce"
1033 |   top: "conv2_8_1x1_reduce"
1034 |   scale_param {
1035 |     bias_term: true
1036 |   }
1037 | }
1038 | layer {
1039 |   name: "conv2_8_1x1_reduce/relu"
1040 |   type: "ReLU"
1041 |   bottom: "conv2_8_1x1_reduce"
1042 |   top: "conv2_8_1x1_reduce"
1043 | }
1044 | layer {
1045 |   name: "conv2_8_3x3"
1046 |   type: "Convolution"
1047 |   bottom: "conv2_8_1x1_reduce"
1048 |   top: "conv2_8_3x3"
1049 |   convolution_param {
1050 |     num_output: 64
1051 |     bias_term: false
1052 |     pad: 1
1053 |     kernel_size: 3
1054 |     stride: 1
1055 |   }
1056 | }
1057 | layer {
1058 |   name: "conv2_8_3x3/bn"
1059 |   type: "BatchNorm"
1060 |   bottom: "conv2_8_3x3"
1061 |   top: "conv2_8_3x3"
1062 |   batch_norm_param {
1063 |     use_global_stats: true
1064 |   }
1065 | }
1066 | layer {
1067 |   name: "conv2_8_3x3/scale"
1068 |   type: "Scale"
1069 |   bottom: "conv2_8_3x3"
1070 |   top: "conv2_8_3x3"
1071 |   scale_param {
1072 |     bias_term: true
1073 |   }
1074 | }
1075 | layer {
1076 |   name: "conv2_8_3x3/relu"
1077 |   type: "ReLU"
1078 |   bottom: "conv2_8_3x3"
1079 |   top: "conv2_8_3x3"
1080 | }
1081 | layer {
1082 |   name: "conv2_8_1x1_increase"
1083 |   type: "Convolution"
1084 |   bottom: "conv2_8_3x3"
1085 |   top: "conv2_8_1x1_increase"
1086 |   convolution_param {
1087 |     num_output: 256
1088 |     bias_term: false
1089 |     pad: 0
1090 |     kernel_size: 1
1091 |     stride: 1
1092 |   }
1093 | }
1094 | layer {
1095 |   name: "conv2_8_1x1_increase/bn"
1096 |   type: "BatchNorm"
1097 |   bottom: "conv2_8_1x1_increase"
1098 |   top: "conv2_8_1x1_increase"
1099 |   batch_norm_param {
1100 |     use_global_stats: true
1101 |   }
1102 | }
1103 | layer {
1104 |   name: "conv2_8_1x1_increase/scale"
1105 |   type: "Scale"
1106 |   bottom: "conv2_8_1x1_increase"
1107 |   top: "conv2_8_1x1_increase"
1108 |   scale_param {
1109 |     bias_term: true
1110 |   }
1111 | }
1112 | layer {
1113 |   name: "conv2_8"
1114 |   type: "Eltwise"
1115 |   bottom: "conv2_7"
1116 |   bottom: "conv2_8_1x1_increase"
1117 |   top: "conv2_8"
1118 |   eltwise_param {
1119 |     operation: SUM
1120 |   }
1121 | }
1122 | layer {
1123 |   name: "conv2_8/relu"
1124 |   type: "ReLU"
1125 |   bottom: "conv2_8"
1126 |   top: "conv2_8"
1127 | }
1128 | layer {
1129 |   name: "conv3_1_1x1_reduce"
1130 |   type: "Convolution"
1131 |   bottom: "conv2_8"
1132 |   top: "conv3_1_1x1_reduce"
1133 |   convolution_param {
1134 |     num_output: 128
1135 |     bias_term: false
1136 |     pad: 0
1137 |     kernel_size: 1
1138 |     stride: 1
1139 |   }
1140 | }
1141 | layer {
1142 |   name: "conv3_1_1x1_reduce/bn"
1143 |   type: "BatchNorm"
1144 |   bottom: "conv3_1_1x1_reduce"
1145 |   top: "conv3_1_1x1_reduce"
1146 |   batch_norm_param {
1147 |     use_global_stats: true
1148 |   }
1149 | }
1150 | layer {
1151 |   name: "conv3_1_1x1_reduce/scale"
1152 |   type: "Scale"
1153 |   bottom: "conv3_1_1x1_reduce"
1154 |   top: "conv3_1_1x1_reduce"
1155 |   scale_param {
1156 |     bias_term: true
1157 |   }
1158 | }
1159 | layer {
1160 |   name: "conv3_1_1x1_reduce/relu"
1161 |   type: "ReLU"
1162 |   bottom: "conv3_1_1x1_reduce"
1163 |   top: "conv3_1_1x1_reduce"
1164 | }
1165 | layer {
1166 |   name: "conv3_1_3x3"
1167 |   type: "Convolution"
1168 |   bottom: "conv3_1_1x1_reduce"
1169 |   top: "conv3_1_3x3"
1170 |   convolution_param {
1171 |     num_output: 128
1172 |     bias_term: false
1173 |     pad: 1
1174 |     kernel_size: 3
1175 |     stride: 2
1176 |   }
1177 | }
1178 | layer {
1179 |   name: "conv3_1_3x3/bn"
1180 |   type: "BatchNorm"
1181 |   bottom: "conv3_1_3x3"
1182 |   top: "conv3_1_3x3"
1183 |   batch_norm_param {
1184 |     use_global_stats: true
1185 |   }
1186 | }
1187 | layer {
1188 |   name: "conv3_1_3x3/scale"
1189 |   type: "Scale"
1190 |   bottom: "conv3_1_3x3"
1191 |   top: "conv3_1_3x3"
1192 |   scale_param {
1193 |     bias_term: true
1194 |   }
1195 | }
1196 | layer {
1197 |   name: "conv3_1_3x3/relu"
1198 |   type: "ReLU"
1199 |   bottom: "conv3_1_3x3"
1200 |   top: "conv3_1_3x3"
1201 | }
1202 | layer {
1203 |   name: "conv3_1_1x1_increase"
1204 |   type: "Convolution"
1205 |   bottom: "conv3_1_3x3"
1206 |   top: "conv3_1_1x1_increase"
1207 |   convolution_param {
1208 |     num_output: 512
1209 |     bias_term: false
1210 |     pad: 0
1211 |     kernel_size: 1
1212 |     stride: 1
1213 |   }
1214 | }
1215 | layer {
1216 |   name: "conv3_1_1x1_increase/bn"
1217 |   type: "BatchNorm"
1218 |   bottom: "conv3_1_1x1_increase"
1219 |   top: "conv3_1_1x1_increase"
1220 |   batch_norm_param {
1221 |     use_global_stats: true
1222 |   }
1223 | }
1224 | layer {
1225 |   name: "conv3_1_1x1_increase/scale"
1226 |   type: "Scale"
1227 |   bottom: "conv3_1_1x1_increase"
1228 |   top: "conv3_1_1x1_increase"
1229 |   scale_param {
1230 |     bias_term: true
1231 |   }
1232 | }
1233 | layer {
1234 |   name: "conv3_1_1x1_proj"
1235 |   type: "Convolution"
1236 |   bottom: "conv2_8"
1237 |   top: "conv3_1_1x1_proj"
1238 |   convolution_param {
1239 |     num_output: 512
1240 |     bias_term: false
1241 |     pad: 0
1242 |     kernel_size: 1
1243 |     stride: 2
1244 |   }
1245 | }
1246 | layer {
1247 |   name: "conv3_1_1x1_proj/bn"
1248 |   type: "BatchNorm"
1249 |   bottom: "conv3_1_1x1_proj"
1250 |   top: "conv3_1_1x1_proj"
1251 |   batch_norm_param {
1252 |     use_global_stats: true
1253 |   }
1254 | }
1255 | layer {
1256 |   name: "conv3_1_1x1_proj/scale"
1257 |   type: "Scale"
1258 |   bottom: "conv3_1_1x1_proj"
1259 |   top: "conv3_1_1x1_proj"
1260 |   scale_param {
1261 |     bias_term: true
1262 |   }
1263 | }
1264 | layer {
1265 |   name: "conv3_1"
1266 |   type: "Eltwise"
1267 |   bottom: "conv3_1_1x1_proj"
1268 |   bottom: "conv3_1_1x1_increase"
1269 |   top: "conv3_1"
1270 |   eltwise_param {
1271 |     operation: SUM
1272 |   }
1273 | }
1274 | layer {
1275 |   name: "conv3_1/relu"
1276 |   type: "ReLU"
1277 |   bottom: "conv3_1"
1278 |   top: "conv3_1"
1279 | }
1280 | layer {
1281 |   name: "conv3_2_1x1_reduce"
1282 |   type: "Convolution"
1283 |   bottom: "conv3_1"
1284 |   top: "conv3_2_1x1_reduce"
1285 |   convolution_param {
1286 |     num_output: 128
1287 |     bias_term: false
1288 |     pad: 0
1289 |     kernel_size: 1
1290 |     stride: 1
1291 |   }
1292 | }
1293 | layer {
1294 |   name: "conv3_2_1x1_reduce/bn"
1295 |   type: "BatchNorm"
1296 |   bottom: "conv3_2_1x1_reduce"
1297 |   top: "conv3_2_1x1_reduce"
1298 |   batch_norm_param {
1299 |     use_global_stats: true
1300 |   }
1301 | }
1302 | layer {
1303 |   name: "conv3_2_1x1_reduce/scale"
1304 |   type: "Scale"
1305 |   bottom: "conv3_2_1x1_reduce"
1306 |   top: "conv3_2_1x1_reduce"
1307 |   scale_param {
1308 |     bias_term: true
1309 |   }
1310 | }
1311 | layer {
1312 |   name: "conv3_2_1x1_reduce/relu"
1313 |   type: "ReLU"
1314 |   bottom: "conv3_2_1x1_reduce"
1315 |   top: "conv3_2_1x1_reduce"
1316 | }
1317 | layer {
1318 |   name: "conv3_2_3x3"
1319 |   type: "Convolution"
1320 |   bottom: "conv3_2_1x1_reduce"
1321 |   top: "conv3_2_3x3"
1322 |   convolution_param {
1323 |     num_output: 128
1324 |     bias_term: false
1325 |     pad: 1
1326 |     kernel_size: 3
1327 |     stride: 1
1328 |   }
1329 | }
1330 | layer {
1331 |   name: "conv3_2_3x3/bn"
1332 |   type: "BatchNorm"
1333 |   bottom: "conv3_2_3x3"
1334 |   top: "conv3_2_3x3"
1335 |   batch_norm_param {
1336 |     use_global_stats: true
1337 |   }
1338 | }
1339 | layer {
1340 |   name: "conv3_2_3x3/scale"
1341 |   type: "Scale"
1342 |   bottom: "conv3_2_3x3"
1343 |   top: "conv3_2_3x3"
1344 |   scale_param {
1345 |     bias_term: true
1346 |   }
1347 | }
1348 | layer {
1349 |   name: "conv3_2_3x3/relu"
1350 |   type: "ReLU"
1351 |   bottom: "conv3_2_3x3"
1352 |   top: "conv3_2_3x3"
1353 | }
1354 | layer {
1355 |   name: "conv3_2_1x1_increase"
1356 |   type: "Convolution"
1357 |   bottom: "conv3_2_3x3"
1358 |   top: "conv3_2_1x1_increase"
1359 |   convolution_param {
1360 |     num_output: 512
1361 |     bias_term: false
1362 |     pad: 0
1363 |     kernel_size: 1
1364 |     stride: 1
1365 |   }
1366 | }
1367 | layer {
1368 |   name: "conv3_2_1x1_increase/bn"
1369 |   type: "BatchNorm"
1370 |   bottom: "conv3_2_1x1_increase"
1371 |   top: "conv3_2_1x1_increase"
1372 |   batch_norm_param {
1373 |     use_global_stats: true
1374 |   }
1375 | }
1376 | layer {
1377 |   name: "conv3_2_1x1_increase/scale"
1378 |   type: "Scale"
1379 |   bottom: "conv3_2_1x1_increase"
1380 |   top: "conv3_2_1x1_increase"
1381 |   scale_param {
1382 |     bias_term: true
1383 |   }
1384 | }
1385 | layer {
1386 |   name: "conv3_2"
1387 |   type: "Eltwise"
1388 |   bottom: "conv3_1"
1389 |   bottom: "conv3_2_1x1_increase"
1390 |   top: "conv3_2"
1391 |   eltwise_param {
1392 |     operation: SUM
1393 |   }
1394 | }
1395 | layer {
1396 |   name: "conv3_2/relu"
1397 |   type: "ReLU"
1398 |   bottom: "conv3_2"
1399 |   top: "conv3_2"
1400 | }
1401 | layer {
1402 |   name: "conv3_3_1x1_reduce"
1403 |   type: "Convolution"
1404 |   bottom: "conv3_2"
1405 |   top: "conv3_3_1x1_reduce"
1406 |   convolution_param {
1407 |     num_output: 128
1408 |     bias_term: false
1409 |     pad: 0
1410 |     kernel_size: 1
1411 |     stride: 1
1412 |   }
1413 | }
1414 | layer {
1415 |   name: "conv3_3_1x1_reduce/bn"
1416 |   type: "BatchNorm"
1417 |   bottom: "conv3_3_1x1_reduce"
1418 |   top: "conv3_3_1x1_reduce"
1419 |   batch_norm_param {
1420 |     use_global_stats: true
1421 |   }
1422 | }
1423 | layer {
1424 |   name: "conv3_3_1x1_reduce/scale"
1425 |   type: "Scale"
1426 |   bottom: "conv3_3_1x1_reduce"
1427 |   top: "conv3_3_1x1_reduce"
1428 |   scale_param {
1429 |     bias_term: true
1430 |   }
1431 | }
1432 | layer {
1433 |   name: "conv3_3_1x1_reduce/relu"
1434 |   type: "ReLU"
1435 |   bottom: "conv3_3_1x1_reduce"
1436 |   top: "conv3_3_1x1_reduce"
1437 | }
1438 | layer {
1439 |   name: "conv3_3_3x3"
1440 |   type: "Convolution"
1441 |   bottom: "conv3_3_1x1_reduce"
1442 |   top: "conv3_3_3x3"
1443 |   convolution_param {
1444 |     num_output: 128
1445 |     bias_term: false
1446 |     pad: 1
1447 |     kernel_size: 3
1448 |     stride: 1
1449 |   }
1450 | }
1451 | layer {
1452 |   name: "conv3_3_3x3/bn"
1453 |   type: "BatchNorm"
1454 |   bottom: "conv3_3_3x3"
1455 |   top: "conv3_3_3x3"
1456 |   batch_norm_param {
1457 |     use_global_stats: true
1458 |   }
1459 | }
1460 | layer {
1461 |   name: "conv3_3_3x3/scale"
1462 |   type: "Scale"
1463 |   bottom: "conv3_3_3x3"
1464 |   top: "conv3_3_3x3"
1465 |   scale_param {
1466 |     bias_term: true
1467 |   }
1468 | }
1469 | layer {
1470 |   name: "conv3_3_3x3/relu"
1471 |   type: "ReLU"
1472 |   bottom: "conv3_3_3x3"
1473 |   top: "conv3_3_3x3"
1474 | }
1475 | layer {
1476 |   name: "conv3_3_1x1_increase"
1477 |   type: "Convolution"
1478 |   bottom: "conv3_3_3x3"
1479 |   top: "conv3_3_1x1_increase"
1480 |   convolution_param {
1481 |     num_output: 512
1482 |     bias_term: false
1483 |     pad: 0
1484 |     kernel_size: 1
1485 |     stride: 1
1486 |   }
1487 | }
1488 | layer {
1489 |   name: "conv3_3_1x1_increase/bn"
1490 |   type: "BatchNorm"
1491 |   bottom: "conv3_3_1x1_increase"
1492 |   top: "conv3_3_1x1_increase"
1493 |   batch_norm_param {
1494 |     use_global_stats: true
1495 |   }
1496 | }
1497 | layer {
1498 |   name: "conv3_3_1x1_increase/scale"
1499 |   type: "Scale"
1500 |   bottom: "conv3_3_1x1_increase"
1501 |   top: "conv3_3_1x1_increase"
1502 |   scale_param {
1503 |     bias_term: true
1504 |   }
1505 | }
1506 | layer {
1507 |   name: "conv3_3"
1508 |   type: "Eltwise"
1509 |   bottom: "conv3_2"
1510 |   bottom: "conv3_3_1x1_increase"
1511 |   top: "conv3_3"
1512 |   eltwise_param {
1513 |     operation: SUM
1514 |   }
1515 | }
1516 | layer {
1517 |   name: "conv3_3/relu"
1518 |   type: "ReLU"
1519 |   bottom: "conv3_3"
1520 |   top: "conv3_3"
1521 | }
1522 | layer {
1523 |   name: "conv3_4_1x1_reduce"
1524 |   type: "Convolution"
1525 |   bottom: "conv3_3"
1526 |   top: "conv3_4_1x1_reduce"
1527 |   convolution_param {
1528 |     num_output: 128
1529 |     bias_term: false
1530 |     pad: 0
1531 |     kernel_size: 1
1532 |     stride: 1
1533 |   }
1534 | }
1535 | layer {
1536 |   name: "conv3_4_1x1_reduce/bn"
1537 |   type: "BatchNorm"
1538 |   bottom: "conv3_4_1x1_reduce"
1539 |   top: "conv3_4_1x1_reduce"
1540 |   batch_norm_param {
1541 |     use_global_stats: true
1542 |   }
1543 | }
1544 | layer {
1545 |   name: "conv3_4_1x1_reduce/scale"
1546 |   type: "Scale"
1547 |   bottom: "conv3_4_1x1_reduce"
1548 |   top: "conv3_4_1x1_reduce"
1549 |   scale_param {
1550 |     bias_term: true
1551 |   }
1552 | }
1553 | layer {
1554 |   name: "conv3_4_1x1_reduce/relu"
1555 |   type: "ReLU"
1556 |   bottom: "conv3_4_1x1_reduce"
1557 |   top: "conv3_4_1x1_reduce"
1558 | }
1559 | layer {
1560 |   name: "conv3_4_3x3"
1561 |   type: "Convolution"
1562 |   bottom: "conv3_4_1x1_reduce"
1563 |   top: "conv3_4_3x3"
1564 |   convolution_param {
1565 |     num_output: 128
1566 |     bias_term: false
1567 |     pad: 1
1568 |     kernel_size: 3
1569 |     stride: 1
1570 |   }
1571 | }
1572 | layer {
1573 |   name: "conv3_4_3x3/bn"
1574 |   type: "BatchNorm"
1575 |   bottom: "conv3_4_3x3"
1576 |   top: "conv3_4_3x3"
1577 |   batch_norm_param {
1578 |     use_global_stats: true
1579 |   }
1580 | }
1581 | layer {
1582 |   name: "conv3_4_3x3/scale"
1583 |   type: "Scale"
1584 |   bottom: "conv3_4_3x3"
1585 |   top: "conv3_4_3x3"
1586 |   scale_param {
1587 |     bias_term: true
1588 |   }
1589 | }
1590 | layer {
1591 |   name: "conv3_4_3x3/relu"
1592 |   type: "ReLU"
1593 |   bottom: "conv3_4_3x3"
1594 |   top: "conv3_4_3x3"
1595 | }
1596 | layer {
1597 |   name: "conv3_4_1x1_increase"
1598 |   type: "Convolution"
1599 |   bottom: "conv3_4_3x3"
1600 |   top: "conv3_4_1x1_increase"
1601 |   convolution_param {
1602 |     num_output: 512
1603 |     bias_term: false
1604 |     pad: 0
1605 |     kernel_size: 1
1606 |     stride: 1
1607 |   }
1608 | }
1609 | layer {
1610 |   name: "conv3_4_1x1_increase/bn"
1611 |   type: "BatchNorm"
1612 |   bottom: "conv3_4_1x1_increase"
1613 |   top: "conv3_4_1x1_increase"
1614 |   batch_norm_param {
1615 |     use_global_stats: true
1616 |   }
1617 | }
1618 | layer {
1619 |   name: "conv3_4_1x1_increase/scale"
1620 |   type: "Scale"
1621 |   bottom: "conv3_4_1x1_increase"
1622 |   top: "conv3_4_1x1_increase"
1623 |   scale_param {
1624 |     bias_term: true
1625 |   }
1626 | }
1627 | layer {
1628 |   name: "conv3_4"
1629 |   type: "Eltwise"
1630 |   bottom: "conv3_3"
1631 |   bottom: "conv3_4_1x1_increase"
1632 |   top: "conv3_4"
1633 |   eltwise_param {
1634 |     operation: SUM
1635 |   }
1636 | }
1637 | layer {
1638 |   name: "conv3_4/relu"
1639 |   type: "ReLU"
1640 |   bottom: "conv3_4"
1641 |   top: "conv3_4"
1642 | }
1643 | layer {
1644 |   name: "conv3_5_1x1_reduce"
1645 |   type: "Convolution"
1646 |   bottom: "conv3_4"
1647 |   top: "conv3_5_1x1_reduce"
1648 |   convolution_param {
1649 |     num_output: 128
1650 |     bias_term: false
1651 |     pad: 0
1652 |     kernel_size: 1
1653 |     stride: 1
1654 |   }
1655 | }
1656 | layer {
1657 |   name: "conv3_5_1x1_reduce/bn"
1658 |   type: "BatchNorm"
1659 |   bottom: "conv3_5_1x1_reduce"
1660 |   top: "conv3_5_1x1_reduce"
1661 |   batch_norm_param {
1662 |     use_global_stats: true
1663 |   }
1664 | }
1665 | layer {
1666 |   name: "conv3_5_1x1_reduce/scale"
1667 |   type: "Scale"
1668 |   bottom: "conv3_5_1x1_reduce"
1669 |   top: "conv3_5_1x1_reduce"
1670 |   scale_param {
1671 |     bias_term: true
1672 |   }
1673 | }
1674 | layer {
1675 |   name: "conv3_5_1x1_reduce/relu"
1676 |   type: "ReLU"
1677 |   bottom: "conv3_5_1x1_reduce"
1678 |   top: "conv3_5_1x1_reduce"
1679 | }
1680 | layer {
1681 |   name: "conv3_5_3x3"
1682 |   type: "Convolution"
1683 |   bottom: "conv3_5_1x1_reduce"
1684 |   top: "conv3_5_3x3"
1685 |   convolution_param {
1686 |     num_output: 128
1687 |     bias_term: false
1688 |     pad: 1
1689 |     kernel_size: 3
1690 |     stride: 1
1691 |   }
1692 | }
1693 | layer {
1694 |   name: "conv3_5_3x3/bn"
1695 |   type: "BatchNorm"
1696 |   bottom: "conv3_5_3x3"
1697 |   top: "conv3_5_3x3"
1698 |   batch_norm_param {
1699 |     use_global_stats: true
1700 |   }
1701 | }
1702 | layer {
1703 |   name: "conv3_5_3x3/scale"
1704 |   type: "Scale"
1705 |   bottom: "conv3_5_3x3"
1706 |   top: "conv3_5_3x3"
1707 |   scale_param {
1708 |     bias_term: true
1709 |   }
1710 | }
1711 | layer {
1712 |   name: "conv3_5_3x3/relu"
1713 |   type: "ReLU"
1714 |   bottom: "conv3_5_3x3"
1715 |   top: "conv3_5_3x3"
1716 | }
1717 | layer {
1718 |   name: "conv3_5_1x1_increase"
1719 |   type: "Convolution"
1720 |   bottom: "conv3_5_3x3"
1721 |   top: "conv3_5_1x1_increase"
1722 |   convolution_param {
1723 |     num_output: 512
1724 |     bias_term: false
1725 |     pad: 0
1726 |     kernel_size: 1
1727 |     stride: 1
1728 |   }
1729 | }
1730 | layer {
1731 |   name: "conv3_5_1x1_increase/bn"
1732 |   type: "BatchNorm"
1733 |   bottom: "conv3_5_1x1_increase"
1734 |   top: "conv3_5_1x1_increase"
1735 |   batch_norm_param {
1736 |     use_global_stats: true
1737 |   }
1738 | }
1739 | layer {
1740 |   name: "conv3_5_1x1_increase/scale"
1741 |   type: "Scale"
1742 |   bottom: "conv3_5_1x1_increase"
1743 |   top: "conv3_5_1x1_increase"
1744 |   scale_param {
1745 |     bias_term: true
1746 |   }
1747 | }
1748 | layer {
1749 |   name: "conv3_5"
1750 |   type: "Eltwise"
1751 |   bottom: "conv3_4"
1752 |   bottom: "conv3_5_1x1_increase"
1753 |   top: "conv3_5"
1754 |   eltwise_param {
1755 |     operation: SUM
1756 |   }
1757 | }
1758 | layer {
1759 |   name: "conv3_5/relu"
1760 |   type: "ReLU"
1761 |   bottom: "conv3_5"
1762 |   top: "conv3_5"
1763 | }
1764 | layer {
1765 |   name: "conv3_6_1x1_reduce"
1766 |   type: "Convolution"
1767 |   bottom: "conv3_5"
1768 |   top: "conv3_6_1x1_reduce"
1769 |   convolution_param {
1770 |     num_output: 128
1771 |     bias_term: false
1772 |     pad: 0
1773 |     kernel_size: 1
1774 |     stride: 1
1775 |   }
1776 | }
1777 | layer {
1778 |   name: "conv3_6_1x1_reduce/bn"
1779 |   type: "BatchNorm"
1780 |   bottom: "conv3_6_1x1_reduce"
1781 |   top: "conv3_6_1x1_reduce"
1782 |   batch_norm_param {
1783 |     use_global_stats: true
1784 |   }
1785 | }
1786 | layer {
1787 |   name: "conv3_6_1x1_reduce/scale"
1788 |   type: "Scale"
1789 |   bottom: "conv3_6_1x1_reduce"
1790 |   top: "conv3_6_1x1_reduce"
1791 |   scale_param {
1792 |     bias_term: true
1793 |   }
1794 | }
1795 | layer {
1796 |   name: "conv3_6_1x1_reduce/relu"
1797 |   type: "ReLU"
1798 |   bottom: "conv3_6_1x1_reduce"
1799 |   top: "conv3_6_1x1_reduce"
1800 | }
1801 | layer {
1802 |   name: "conv3_6_3x3"
1803 |   type: "Convolution"
1804 |   bottom: "conv3_6_1x1_reduce"
1805 |   top: "conv3_6_3x3"
1806 |   convolution_param {
1807 |     num_output: 128
1808 |     bias_term: false
1809 |     pad: 1
1810 |     kernel_size: 3
1811 |     stride: 1
1812 |   }
1813 | }
1814 | layer {
1815 |   name: "conv3_6_3x3/bn"
1816 |   type: "BatchNorm"
1817 |   bottom: "conv3_6_3x3"
1818 |   top: "conv3_6_3x3"
1819 |   batch_norm_param {
1820 |     use_global_stats: true
1821 |   }
1822 | }
1823 | layer {
1824 |   name: "conv3_6_3x3/scale"
1825 |   type: "Scale"
1826 |   bottom: "conv3_6_3x3"
1827 |   top: "conv3_6_3x3"
1828 |   scale_param {
1829 |     bias_term: true
1830 |   }
1831 | }
1832 | layer {
1833 |   name: "conv3_6_3x3/relu"
1834 |   type: "ReLU"
1835 |   bottom: "conv3_6_3x3"
1836 |   top: "conv3_6_3x3"
1837 | }
1838 | layer {
1839 |   name: "conv3_6_1x1_increase"
1840 |   type: "Convolution"
1841 |   bottom: "conv3_6_3x3"
1842 |   top: "conv3_6_1x1_increase"
1843 |   convolution_param {
1844 |     num_output: 512
1845 |     bias_term: false
1846 |     pad: 0
1847 |     kernel_size: 1
1848 |     stride: 1
1849 |   }
1850 | }
1851 | layer {
1852 |   name: "conv3_6_1x1_increase/bn"
1853 |   type: "BatchNorm"
1854 |   bottom: "conv3_6_1x1_increase"
1855 |   top: "conv3_6_1x1_increase"
1856 |   batch_norm_param {
1857 |     use_global_stats: true
1858 |   }
1859 | }
1860 | layer {
1861 |   name: "conv3_6_1x1_increase/scale"
1862 |   type: "Scale"
1863 |   bottom: "conv3_6_1x1_increase"
1864 |   top: "conv3_6_1x1_increase"
1865 |   scale_param {
1866 |     bias_term: true
1867 |   }
1868 | }
1869 | layer {
1870 |   name: "conv3_6"
1871 |   type: "Eltwise"
1872 |   bottom: "conv3_5"
1873 |   bottom: "conv3_6_1x1_increase"
1874 |   top: "conv3_6"
1875 |   eltwise_param {
1876 |     operation: SUM
1877 |   }
1878 | }
1879 | layer {
1880 |   name: "conv3_6/relu"
1881 |   type: "ReLU"
1882 |   bottom: "conv3_6"
1883 |   top: "conv3_6"
1884 | }
1885 | layer {
1886 |   name: "conv3_7_1x1_reduce"
1887 |   type: "Convolution"
1888 |   bottom: "conv3_6"
1889 |   top: "conv3_7_1x1_reduce"
1890 |   convolution_param {
1891 |     num_output: 128
1892 |     bias_term: false
1893 |     pad: 0
1894 |     kernel_size: 1
1895 |     stride: 1
1896 |   }
1897 | }
1898 | layer {
1899 |   name: "conv3_7_1x1_reduce/bn"
1900 |   type: "BatchNorm"
1901 |   bottom: "conv3_7_1x1_reduce"
1902 |   top: "conv3_7_1x1_reduce"
1903 |   batch_norm_param {
1904 |     use_global_stats: true
1905 |   }
1906 | }
1907 | layer {
1908 |   name: "conv3_7_1x1_reduce/scale"
1909 |   type: "Scale"
1910 |   bottom: "conv3_7_1x1_reduce"
1911 |   top: "conv3_7_1x1_reduce"
1912 |   scale_param {
1913 |     bias_term: true
1914 |   }
1915 | }
1916 | layer {
1917 |   name: "conv3_7_1x1_reduce/relu"
1918 |   type: "ReLU"
1919 |   bottom: "conv3_7_1x1_reduce"
1920 |   top: "conv3_7_1x1_reduce"
1921 | }
1922 | layer {
1923 |   name: "conv3_7_3x3"
1924 |   type: "Convolution"
1925 |   bottom: "conv3_7_1x1_reduce"
1926 |   top: "conv3_7_3x3"
1927 |   convolution_param {
1928 |     num_output: 128
1929 |     bias_term: false
1930 |     pad: 1
1931 |     kernel_size: 3
1932 |     stride: 1
1933 |   }
1934 | }
1935 | layer {
1936 |   name: "conv3_7_3x3/bn"
1937 |   type: "BatchNorm"
1938 |   bottom: "conv3_7_3x3"
1939 |   top: "conv3_7_3x3"
1940 |   batch_norm_param {
1941 |     use_global_stats: true
1942 |   }
1943 | }
1944 | layer {
1945 |   name: "conv3_7_3x3/scale"
1946 |   type: "Scale"
1947 |   bottom: "conv3_7_3x3"
1948 |   top: "conv3_7_3x3"
1949 |   scale_param {
1950 |     bias_term: true
1951 |   }
1952 | }
1953 | layer {
1954 |   name: "conv3_7_3x3/relu"
1955 |   type: "ReLU"
1956 |   bottom: "conv3_7_3x3"
1957 |   top: "conv3_7_3x3"
1958 | }
1959 | layer {
1960 |   name: "conv3_7_1x1_increase"
1961 |   type: "Convolution"
1962 |   bottom: "conv3_7_3x3"
1963 |   top: "conv3_7_1x1_increase"
1964 |   convolution_param {
1965 |     num_output: 512
1966 |     bias_term: false
1967 |     pad: 0
1968 |     kernel_size: 1
1969 |     stride: 1
1970 |   }
1971 | }
1972 | layer {
1973 |   name: "conv3_7_1x1_increase/bn"
1974 |   type: "BatchNorm"
1975 |   bottom: "conv3_7_1x1_increase"
1976 |   top: "conv3_7_1x1_increase"
1977 |   batch_norm_param {
1978 |     use_global_stats: true
1979 |   }
1980 | }
1981 | layer {
1982 |   name: "conv3_7_1x1_increase/scale"
1983 |   type: "Scale"
1984 |   bottom: "conv3_7_1x1_increase"
1985 |   top: "conv3_7_1x1_increase"
1986 |   scale_param {
1987 |     bias_term: true
1988 |   }
1989 | }
1990 | layer {
1991 |   name: "conv3_7"
1992 |   type: "Eltwise"
1993 |   bottom: "conv3_6"
1994 |   bottom: "conv3_7_1x1_increase"
1995 |   top: "conv3_7"
1996 |   eltwise_param {
1997 |     operation: SUM
1998 |   }
1999 | }
2000 | layer {
2001 |   name: "conv3_7/relu"
2002 |   type: "ReLU"
2003 |   bottom: "conv3_7"
2004 |   top: "conv3_7"
2005 | }
2006 | layer {
2007 |   name: "conv3_8_1x1_reduce"
2008 |   type: "Convolution"
2009 |   bottom: "conv3_7"
2010 |   top: "conv3_8_1x1_reduce"
2011 |   convolution_param {
2012 |     num_output: 128
2013 |     bias_term: false
2014 |     pad: 0
2015 |     kernel_size: 1
2016 |     stride: 1
2017 |   }
2018 | }
2019 | layer {
2020 |   name: "conv3_8_1x1_reduce/bn"
2021 |   type: "BatchNorm"
2022 |   bottom: "conv3_8_1x1_reduce"
2023 |   top: "conv3_8_1x1_reduce"
2024 |   batch_norm_param {
2025 |     use_global_stats: true
2026 |   }
2027 | }
2028 | layer {
2029 |   name: "conv3_8_1x1_reduce/scale"
2030 |   type: "Scale"
2031 |   bottom: "conv3_8_1x1_reduce"
2032 |   top: "conv3_8_1x1_reduce"
2033 |   scale_param {
2034 |     bias_term: true
2035 |   }
2036 | }
2037 | layer {
2038 |   name: "conv3_8_1x1_reduce/relu"
2039 |   type: "ReLU"
2040 |   bottom: "conv3_8_1x1_reduce"
2041 |   top: "conv3_8_1x1_reduce"
2042 | }
2043 | layer {
2044 |   name: "conv3_8_3x3"
2045 |   type: "Convolution"
2046 |   bottom: "conv3_8_1x1_reduce"
2047 |   top: "conv3_8_3x3"
2048 |   convolution_param {
2049 |     num_output: 128
2050 |     bias_term: false
2051 |     pad: 1
2052 |     kernel_size: 3
2053 |     stride: 1
2054 |   }
2055 | }
2056 | layer {
2057 |   name: "conv3_8_3x3/bn"
2058 |   type: "BatchNorm"
2059 |   bottom: "conv3_8_3x3"
2060 |   top: "conv3_8_3x3"
2061 |   batch_norm_param {
2062 |     use_global_stats: true
2063 |   }
2064 | }
2065 | layer {
2066 |   name: "conv3_8_3x3/scale"
2067 |   type: "Scale"
2068 |   bottom: "conv3_8_3x3"
2069 |   top: "conv3_8_3x3"
2070 |   scale_param {
2071 |     bias_term: true
2072 |   }
2073 | }
2074 | layer {
2075 |   name: "conv3_8_3x3/relu"
2076 |   type: "ReLU"
2077 |   bottom: "conv3_8_3x3"
2078 |   top: "conv3_8_3x3"
2079 | }
2080 | layer {
2081 |   name: "conv3_8_1x1_increase"
2082 |   type: "Convolution"
2083 |   bottom: "conv3_8_3x3"
2084 |   top: "conv3_8_1x1_increase"
2085 |   convolution_param {
2086 |     num_output: 512
2087 |     bias_term: false
2088 |     pad: 0
2089 |     kernel_size: 1
2090 |     stride: 1
2091 |   }
2092 | }
2093 | layer {
2094 |   name: "conv3_8_1x1_increase/bn"
2095 |   type: "BatchNorm"
2096 |   bottom: "conv3_8_1x1_increase"
2097 |   top: "conv3_8_1x1_increase"
2098 |   batch_norm_param {
2099 |     use_global_stats: true
2100 |   }
2101 | }
2102 | layer {
2103 |   name: "conv3_8_1x1_increase/scale"
2104 |   type: "Scale"
2105 |   bottom: "conv3_8_1x1_increase"
2106 |   top: "conv3_8_1x1_increase"
2107 |   scale_param {
2108 |     bias_term: true
2109 |   }
2110 | }
2111 | layer {
2112 |   name: "conv3_8"
2113 |   type: "Eltwise"
2114 |   bottom: "conv3_7"
2115 |   bottom: "conv3_8_1x1_increase"
2116 |   top: "conv3_8"
2117 |   eltwise_param {
2118 |     operation: SUM
2119 |   }
2120 | }
2121 | layer {
2122 |   name: "conv3_8/relu"
2123 |   type: "ReLU"
2124 |   bottom: "conv3_8"
2125 |   top: "conv3_8"
2126 | }
2127 | layer {
2128 |   name: "conv3_9_1x1_reduce"
2129 |   type: "Convolution"
2130 |   bottom: "conv3_8"
2131 |   top: "conv3_9_1x1_reduce"
2132 |   convolution_param {
2133 |     num_output: 128
2134 |     bias_term: false
2135 |     pad: 0
2136 |     kernel_size: 1
2137 |     stride: 1
2138 |   }
2139 | }
2140 | layer {
2141 |   name: "conv3_9_1x1_reduce/bn"
2142 |   type: "BatchNorm"
2143 |   bottom: "conv3_9_1x1_reduce"
2144 |   top: "conv3_9_1x1_reduce"
2145 |   batch_norm_param {
2146 |     use_global_stats: true
2147 |   }
2148 | }
2149 | layer {
2150 |   name: "conv3_9_1x1_reduce/scale"
2151 |   type: "Scale"
2152 |   bottom: "conv3_9_1x1_reduce"
2153 |   top: "conv3_9_1x1_reduce"
2154 |   scale_param {
2155 |     bias_term: true
2156 |   }
2157 | }
2158 | layer {
2159 |   name: "conv3_9_1x1_reduce/relu"
2160 |   type: "ReLU"
2161 |   bottom: "conv3_9_1x1_reduce"
2162 |   top: "conv3_9_1x1_reduce"
2163 | }
2164 | layer {
2165 |   name: "conv3_9_3x3"
2166 |   type: "Convolution"
2167 |   bottom: "conv3_9_1x1_reduce"
2168 |   top: "conv3_9_3x3"
2169 |   convolution_param {
2170 |     num_output: 128
2171 |     bias_term: false
2172 |     pad: 1
2173 |     kernel_size: 3
2174 |     stride: 1
2175 |   }
2176 | }
2177 | layer {
2178 |   name: "conv3_9_3x3/bn"
2179 |   type: "BatchNorm"
2180 |   bottom: "conv3_9_3x3"
2181 |   top: "conv3_9_3x3"
2182 |   batch_norm_param {
2183 |     use_global_stats: true
2184 |   }
2185 | }
2186 | layer {
2187 |   name: "conv3_9_3x3/scale"
2188 |   type: "Scale"
2189 |   bottom: "conv3_9_3x3"
2190 |   top: "conv3_9_3x3"
2191 |   scale_param {
2192 |     bias_term: true
2193 |   }
2194 | }
2195 | layer {
2196 |   name: "conv3_9_3x3/relu"
2197 |   type: "ReLU"
2198 |   bottom: "conv3_9_3x3"
2199 |   top: "conv3_9_3x3"
2200 | }
2201 | layer {
2202 |   name: "conv3_9_1x1_increase"
2203 |   type: "Convolution"
2204 |   bottom: "conv3_9_3x3"
2205 |   top: "conv3_9_1x1_increase"
2206 |   convolution_param {
2207 |     num_output: 512
2208 |     bias_term: false
2209 |     pad: 0
2210 |     kernel_size: 1
2211 |     stride: 1
2212 |   }
2213 | }
2214 | layer {
2215 |   name: "conv3_9_1x1_increase/bn"
2216 |   type: "BatchNorm"
2217 |   bottom: "conv3_9_1x1_increase"
2218 |   top: "conv3_9_1x1_increase"
2219 |   batch_norm_param {
2220 |     use_global_stats: true
2221 |   }
2222 | }
2223 | layer {
2224 |   name: "conv3_9_1x1_increase/scale"
2225 |   type: "Scale"
2226 |   bottom: "conv3_9_1x1_increase"
2227 |   top: "conv3_9_1x1_increase"
2228 |   scale_param {
2229 |     bias_term: true
2230 |   }
2231 | }
2232 | layer {
2233 |   name: "conv3_9"
2234 |   type: "Eltwise"
2235 |   bottom: "conv3_8"
2236 |   bottom: "conv3_9_1x1_increase"
2237 |   top: "conv3_9"
2238 |   eltwise_param {
2239 |     operation: SUM
2240 |   }
2241 | }
2242 | layer {
2243 |   name: "conv3_9/relu"
2244 |   type: "ReLU"
2245 |   bottom: "conv3_9"
2246 |   top: "conv3_9"
2247 | }
2248 | layer {
2249 |   name: "conv3_10_1x1_reduce"
2250 |   type: "Convolution"
2251 |   bottom: "conv3_9"
2252 |   top: "conv3_10_1x1_reduce"
2253 |   convolution_param {
2254 |     num_output: 128
2255 |     bias_term: false
2256 |     pad: 0
2257 |     kernel_size: 1
2258 |     stride: 1
2259 |   }
2260 | }
2261 | layer {
2262 |   name: "conv3_10_1x1_reduce/bn"
2263 |   type: "BatchNorm"
2264 |   bottom: "conv3_10_1x1_reduce"
2265 |   top: "conv3_10_1x1_reduce"
2266 |   batch_norm_param {
2267 |     use_global_stats: true
2268 |   }
2269 | }
2270 | layer {
2271 |   name: "conv3_10_1x1_reduce/scale"
2272 |   type: "Scale"
2273 |   bottom: "conv3_10_1x1_reduce"
2274 |   top: "conv3_10_1x1_reduce"
2275 |   scale_param {
2276 |     bias_term: true
2277 |   }
2278 | }
2279 | layer {
2280 |   name: "conv3_10_1x1_reduce/relu"
2281 |   type: "ReLU"
2282 |   bottom: "conv3_10_1x1_reduce"
2283 |   top: "conv3_10_1x1_reduce"
2284 | }
2285 | layer {
2286 |   name: "conv3_10_3x3"
2287 |   type: "Convolution"
2288 |   bottom: "conv3_10_1x1_reduce"
2289 |   top: "conv3_10_3x3"
2290 |   convolution_param {
2291 |     num_output: 128
2292 |     bias_term: false
2293 |     pad: 1
2294 |     kernel_size: 3
2295 |     stride: 1
2296 |   }
2297 | }
2298 | layer {
2299 |   name: "conv3_10_3x3/bn"
2300 |   type: "BatchNorm"
2301 |   bottom: "conv3_10_3x3"
2302 |   top: "conv3_10_3x3"
2303 |   batch_norm_param {
2304 |     use_global_stats: true
2305 |   }
2306 | }
2307 | layer {
2308 |   name: "conv3_10_3x3/scale"
2309 |   type: "Scale"
2310 |   bottom: "conv3_10_3x3"
2311 |   top: "conv3_10_3x3"
2312 |   scale_param {
2313 |     bias_term: true
2314 |   }
2315 | }
2316 | layer {
2317 |   name: "conv3_10_3x3/relu"
2318 |   type: "ReLU"
2319 |   bottom: "conv3_10_3x3"
2320 |   top: "conv3_10_3x3"
2321 | }
2322 | layer {
2323 |   name: "conv3_10_1x1_increase"
2324 |   type: "Convolution"
2325 |   bottom: "conv3_10_3x3"
2326 |   top: "conv3_10_1x1_increase"
2327 |   convolution_param {
2328 |     num_output: 512
2329 |     bias_term: false
2330 |     pad: 0
2331 |     kernel_size: 1
2332 |     stride: 1
2333 |   }
2334 | }
2335 | layer {
2336 |   name: "conv3_10_1x1_increase/bn"
2337 |   type: "BatchNorm"
2338 |   bottom: "conv3_10_1x1_increase"
2339 |   top: "conv3_10_1x1_increase"
2340 |   batch_norm_param {
2341 |     use_global_stats: true
2342 |   }
2343 | }
2344 | layer {
2345 |   name: "conv3_10_1x1_increase/scale"
2346 |   type: "Scale"
2347 |   bottom: "conv3_10_1x1_increase"
2348 |   top: "conv3_10_1x1_increase"
2349 |   scale_param {
2350 |     bias_term: true
2351 |   }
2352 | }
2353 | layer {
2354 |   name: "conv3_10"
2355 |   type: "Eltwise"
2356 |   bottom: "conv3_9"
2357 |   bottom: "conv3_10_1x1_increase"
2358 |   top: "conv3_10"
2359 |   eltwise_param {
2360 |     operation: SUM
2361 |   }
2362 | }
2363 | layer {
2364 |   name: "conv3_10/relu"
2365 |   type: "ReLU"
2366 |   bottom: "conv3_10"
2367 |   top: "conv3_10"
2368 | }
2369 | layer {
2370 |   name: "conv3_11_1x1_reduce"
2371 |   type: "Convolution"
2372 |   bottom: "conv3_10"
2373 |   top: "conv3_11_1x1_reduce"
2374 |   convolution_param {
2375 |     num_output: 128
2376 |     bias_term: false
2377 |     pad: 0
2378 |     kernel_size: 1
2379 |     stride: 1
2380 |   }
2381 | }
2382 | layer {
2383 |   name: "conv3_11_1x1_reduce/bn"
2384 |   type: "BatchNorm"
2385 |   bottom: "conv3_11_1x1_reduce"
2386 |   top: "conv3_11_1x1_reduce"
2387 |   batch_norm_param {
2388 |     use_global_stats: true
2389 |   }
2390 | }
2391 | layer {
2392 |   name: "conv3_11_1x1_reduce/scale"
2393 |   type: "Scale"
2394 |   bottom: "conv3_11_1x1_reduce"
2395 |   top: "conv3_11_1x1_reduce"
2396 |   scale_param {
2397 |     bias_term: true
2398 |   }
2399 | }
2400 | layer {
2401 |   name: "conv3_11_1x1_reduce/relu"
2402 |   type: "ReLU"
2403 |   bottom: "conv3_11_1x1_reduce"
2404 |   top: "conv3_11_1x1_reduce"
2405 | }
2406 | layer {
2407 |   name: "conv3_11_3x3"
2408 |   type: "Convolution"
2409 |   bottom: "conv3_11_1x1_reduce"
2410 |   top: "conv3_11_3x3"
2411 |   convolution_param {
2412 |     num_output: 128
2413 |     bias_term: false
2414 |     pad: 1
2415 |     kernel_size: 3
2416 |     stride: 1
2417 |   }
2418 | }
2419 | layer {
2420 |   name: "conv3_11_3x3/bn"
2421 |   type: "BatchNorm"
2422 |   bottom: "conv3_11_3x3"
2423 |   top: "conv3_11_3x3"
2424 |   batch_norm_param {
2425 |     use_global_stats: true
2426 |   }
2427 | }
2428 | layer {
2429 |   name: "conv3_11_3x3/scale"
2430 |   type: "Scale"
2431 |   bottom: "conv3_11_3x3"
2432 |   top: "conv3_11_3x3"
2433 |   scale_param {
2434 |     bias_term: true
2435 |   }
2436 | }
2437 | layer {
2438 |   name: "conv3_11_3x3/relu"
2439 |   type: "ReLU"
2440 |   bottom: "conv3_11_3x3"
2441 |   top: "conv3_11_3x3"
2442 | }
2443 | layer {
2444 |   name: "conv3_11_1x1_increase"
2445 |   type: "Convolution"
2446 |   bottom: "conv3_11_3x3"
2447 |   top: "conv3_11_1x1_increase"
2448 |   convolution_param {
2449 |     num_output: 512
2450 |     bias_term: false
2451 |     pad: 0
2452 |     kernel_size: 1
2453 |     stride: 1
2454 |   }
2455 | }
2456 | layer {
2457 |   name: "conv3_11_1x1_increase/bn"
2458 |   type: "BatchNorm"
2459 |   bottom: "conv3_11_1x1_increase"
2460 |   top: "conv3_11_1x1_increase"
2461 |   batch_norm_param {
2462 |     use_global_stats: true
2463 |   }
2464 | }
2465 | layer {
2466 |   name: "conv3_11_1x1_increase/scale"
2467 |   type: "Scale"
2468 |   bottom: "conv3_11_1x1_increase"
2469 |   top: "conv3_11_1x1_increase"
2470 |   scale_param {
2471 |     bias_term: true
2472 |   }
2473 | }
2474 | layer {
2475 |   name: "conv3_11"
2476 |   type: "Eltwise"
2477 |   bottom: "conv3_10"
2478 |   bottom: "conv3_11_1x1_increase"
2479 |   top: "conv3_11"
2480 |   eltwise_param {
2481 |     operation: SUM
2482 |   }
2483 | }
2484 | layer {
2485 |   name: "conv3_11/relu"
2486 |   type: "ReLU"
2487 |   bottom: "conv3_11"
2488 |   top: "conv3_11"
2489 | }
2490 | layer {
2491 |   name: "conv3_12_1x1_reduce"
2492 |   type: "Convolution"
2493 |   bottom: "conv3_11"
2494 |   top: "conv3_12_1x1_reduce"
2495 |   convolution_param {
2496 |     num_output: 128
2497 |     bias_term: false
2498 |     pad: 0
2499 |     kernel_size: 1
2500 |     stride: 1
2501 |   }
2502 | }
2503 | layer {
2504 |   name: "conv3_12_1x1_reduce/bn"
2505 |   type: "BatchNorm"
2506 |   bottom: "conv3_12_1x1_reduce"
2507 |   top: "conv3_12_1x1_reduce"
2508 |   batch_norm_param {
2509 |     use_global_stats: true
2510 |   }
2511 | }
2512 | layer {
2513 |   name: "conv3_12_1x1_reduce/scale"
2514 |   type: "Scale"
2515 |   bottom: "conv3_12_1x1_reduce"
2516 |   top: "conv3_12_1x1_reduce"
2517 |   scale_param {
2518 |     bias_term: true
2519 |   }
2520 | }
2521 | layer {
2522 |   name: "conv3_12_1x1_reduce/relu"
2523 |   type: "ReLU"
2524 |   bottom: "conv3_12_1x1_reduce"
2525 |   top: "conv3_12_1x1_reduce"
2526 | }
2527 | layer {
2528 |   name: "conv3_12_3x3"
2529 |   type: "Convolution"
2530 |   bottom: "conv3_12_1x1_reduce"
2531 |   top: "conv3_12_3x3"
2532 |   convolution_param {
2533 |     num_output: 128
2534 |     bias_term: false
2535 |     pad: 1
2536 |     kernel_size: 3
2537 |     stride: 1
2538 |   }
2539 | }
2540 | layer {
2541 |   name: "conv3_12_3x3/bn"
2542 |   type: "BatchNorm"
2543 |   bottom: "conv3_12_3x3"
2544 |   top: "conv3_12_3x3"
2545 |   batch_norm_param {
2546 |     use_global_stats: true
2547 |   }
2548 | }
2549 | layer {
2550 |   name: "conv3_12_3x3/scale"
2551 |   type: "Scale"
2552 |   bottom: "conv3_12_3x3"
2553 |   top: "conv3_12_3x3"
2554 |   scale_param {
2555 |     bias_term: true
2556 |   }
2557 | }
2558 | layer {
2559 |   name: "conv3_12_3x3/relu"
2560 |   type: "ReLU"
2561 |   bottom: "conv3_12_3x3"
2562 |   top: "conv3_12_3x3"
2563 | }
2564 | layer {
2565 |   name: "conv3_12_1x1_increase"
2566 |   type: "Convolution"
2567 |   bottom: "conv3_12_3x3"
2568 |   top: "conv3_12_1x1_increase"
2569 |   convolution_param {
2570 |     num_output: 512
2571 |     bias_term: false
2572 |     pad: 0
2573 |     kernel_size: 1
2574 |     stride: 1
2575 |   }
2576 | }
2577 | layer {
2578 |   name: "conv3_12_1x1_increase/bn"
2579 |   type: "BatchNorm"
2580 |   bottom: "conv3_12_1x1_increase"
2581 |   top: "conv3_12_1x1_increase"
2582 |   batch_norm_param {
2583 |     use_global_stats: true
2584 |   }
2585 | }
2586 | layer {
2587 |   name: "conv3_12_1x1_increase/scale"
2588 |   type: "Scale"
2589 |   bottom: "conv3_12_1x1_increase"
2590 |   top: "conv3_12_1x1_increase"
2591 |   scale_param {
2592 |     bias_term: true
2593 |   }
2594 | }
2595 | layer {
2596 |   name: "conv3_12"
2597 |   type: "Eltwise"
2598 |   bottom: "conv3_11"
2599 |   bottom: "conv3_12_1x1_increase"
2600 |   top: "conv3_12"
2601 |   eltwise_param {
2602 |     operation: SUM
2603 |   }
2604 | }
2605 | layer {
2606 |   name: "conv3_12/relu"
2607 |   type: "ReLU"
2608 |   bottom: "conv3_12"
2609 |   top: "conv3_12"
2610 | }
2611 | layer {
2612 |   name: "conv4_1_1x1_reduce"
2613 |   type: "Convolution"
2614 |   bottom: "conv3_12"
2615 |   top: "conv4_1_1x1_reduce"
2616 |   convolution_param {
2617 |     num_output: 256
2618 |     bias_term: false
2619 |     pad: 0
2620 |     kernel_size: 1
2621 |     stride: 1
2622 |   }
2623 | }
2624 | layer {
2625 |   name: "conv4_1_1x1_reduce/bn"
2626 |   type: "BatchNorm"
2627 |   bottom: "conv4_1_1x1_reduce"
2628 |   top: "conv4_1_1x1_reduce"
2629 |   batch_norm_param {
2630 |     use_global_stats: true
2631 |   }
2632 | }
2633 | layer {
2634 |   name: "conv4_1_1x1_reduce/scale"
2635 |   type: "Scale"
2636 |   bottom: "conv4_1_1x1_reduce"
2637 |   top: "conv4_1_1x1_reduce"
2638 |   scale_param {
2639 |     bias_term: true
2640 |   }
2641 | }
2642 | layer {
2643 |   name: "conv4_1_1x1_reduce/relu"
2644 |   type: "ReLU"
2645 |   bottom: "conv4_1_1x1_reduce"
2646 |   top: "conv4_1_1x1_reduce"
2647 | }
2648 | layer {
2649 |   name: "conv4_1_3x3"
2650 |   type: "Convolution"
2651 |   bottom: "conv4_1_1x1_reduce"
2652 |   top: "conv4_1_3x3"
2653 |   convolution_param {
2654 |     num_output: 256
2655 |     bias_term: false
2656 |     pad: 1
2657 |     kernel_size: 3
2658 |     stride: 2
2659 |   }
2660 | }
2661 | layer {
2662 |   name: "conv4_1_3x3/bn"
2663 |   type: "BatchNorm"
2664 |   bottom: "conv4_1_3x3"
2665 |   top: "conv4_1_3x3"
2666 |   batch_norm_param {
2667 |     use_global_stats: true
2668 |   }
2669 | }
2670 | layer {
2671 |   name: "conv4_1_3x3/scale"
2672 |   type: "Scale"
2673 |   bottom: "conv4_1_3x3"
2674 |   top: "conv4_1_3x3"
2675 |   scale_param {
2676 |     bias_term: true
2677 |   }
2678 | }
2679 | layer {
2680 |   name: "conv4_1_3x3/relu"
2681 |   type: "ReLU"
2682 |   bottom: "conv4_1_3x3"
2683 |   top: "conv4_1_3x3"
2684 | }
2685 | layer {
2686 |   name: "conv4_1_1x1_increase"
2687 |   type: "Convolution"
2688 |   bottom: "conv4_1_3x3"
2689 |   top: "conv4_1_1x1_increase"
2690 |   convolution_param {
2691 |     num_output: 1024
2692 |     bias_term: false
2693 |     pad: 0
2694 |     kernel_size: 1
2695 |     stride: 1
2696 |   }
2697 | }
2698 | layer {
2699 |   name: "conv4_1_1x1_increase/bn"
2700 |   type: "BatchNorm"
2701 |   bottom: "conv4_1_1x1_increase"
2702 |   top: "conv4_1_1x1_increase"
2703 |   batch_norm_param {
2704 |     use_global_stats: true
2705 |   }
2706 | }
2707 | layer {
2708 |   name: "conv4_1_1x1_increase/scale"
2709 |   type: "Scale"
2710 |   bottom: "conv4_1_1x1_increase"
2711 |   top: "conv4_1_1x1_increase"
2712 |   scale_param {
2713 |     bias_term: true
2714 |   }
2715 | }
2716 | layer {
2717 |   name: "conv4_1_1x1_proj"
2718 |   type: "Convolution"
2719 |   bottom: "conv3_12"
2720 |   top: "conv4_1_1x1_proj"
2721 |   convolution_param {
2722 |     num_output: 1024
2723 |     bias_term: false
2724 |     pad: 0
2725 |     kernel_size: 1
2726 |     stride: 2
2727 |   }
2728 | }
2729 | layer {
2730 |   name: "conv4_1_1x1_proj/bn"
2731 |   type: "BatchNorm"
2732 |   bottom: "conv4_1_1x1_proj"
2733 |   top: "conv4_1_1x1_proj"
2734 |   batch_norm_param {
2735 |     use_global_stats: true
2736 |   }
2737 | }
2738 | layer {
2739 |   name: "conv4_1_1x1_proj/scale"
2740 |   type: "Scale"
2741 |   bottom: "conv4_1_1x1_proj"
2742 |   top: "conv4_1_1x1_proj"
2743 |   scale_param {
2744 |     bias_term: true
2745 |   }
2746 | }
2747 | layer {
2748 |   name: "conv4_1"
2749 |   type: "Eltwise"
2750 |   bottom: "conv4_1_1x1_proj"
2751 |   bottom: "conv4_1_1x1_increase"
2752 |   top: "conv4_1"
2753 |   eltwise_param {
2754 |     operation: SUM
2755 |   }
2756 | }
2757 | layer {
2758 |   name: "conv4_1/relu"
2759 |   type: "ReLU"
2760 |   bottom: "conv4_1"
2761 |   top: "conv4_1"
2762 | }
2763 | layer {
2764 |   name: "conv4_2_1x1_reduce"
2765 |   type: "Convolution"
2766 |   bottom: "conv4_1"
2767 |   top: "conv4_2_1x1_reduce"
2768 |   convolution_param {
2769 |     num_output: 256
2770 |     bias_term: false
2771 |     pad: 0
2772 |     kernel_size: 1
2773 |     stride: 1
2774 |   }
2775 | }
2776 | layer {
2777 |   name: "conv4_2_1x1_reduce/bn"
2778 |   type: "BatchNorm"
2779 |   bottom: "conv4_2_1x1_reduce"
2780 |   top: "conv4_2_1x1_reduce"
2781 |   batch_norm_param {
2782 |     use_global_stats: true
2783 |   }
2784 | }
2785 | layer {
2786 |   name: "conv4_2_1x1_reduce/scale"
2787 |   type: "Scale"
2788 |   bottom: "conv4_2_1x1_reduce"
2789 |   top: "conv4_2_1x1_reduce"
2790 |   scale_param {
2791 |     bias_term: true
2792 |   }
2793 | }
2794 | layer {
2795 |   name: "conv4_2_1x1_reduce/relu"
2796 |   type: "ReLU"
2797 |   bottom: "conv4_2_1x1_reduce"
2798 |   top: "conv4_2_1x1_reduce"
2799 | }
2800 | layer {
2801 |   name: "conv4_2_3x3"
2802 |   type: "Convolution"
2803 |   bottom: "conv4_2_1x1_reduce"
2804 |   top: "conv4_2_3x3"
2805 |   convolution_param {
2806 |     num_output: 256
2807 |     bias_term: false
2808 |     pad: 1
2809 |     kernel_size: 3
2810 |     stride: 1
2811 |   }
2812 | }
2813 | layer {
2814 |   name: "conv4_2_3x3/bn"
2815 |   type: "BatchNorm"
2816 |   bottom: "conv4_2_3x3"
2817 |   top: "conv4_2_3x3"
2818 |   batch_norm_param {
2819 |     use_global_stats: true
2820 |   }
2821 | }
2822 | layer {
2823 |   name: "conv4_2_3x3/scale"
2824 |   type: "Scale"
2825 |   bottom: "conv4_2_3x3"
2826 |   top: "conv4_2_3x3"
2827 |   scale_param {
2828 |     bias_term: true
2829 |   }
2830 | }
2831 | layer {
2832 |   name: "conv4_2_3x3/relu"
2833 |   type: "ReLU"
2834 |   bottom: "conv4_2_3x3"
2835 |   top: "conv4_2_3x3"
2836 | }
2837 | layer {
2838 |   name: "conv4_2_1x1_increase"
2839 |   type: "Convolution"
2840 |   bottom: "conv4_2_3x3"
2841 |   top: "conv4_2_1x1_increase"
2842 |   convolution_param {
2843 |     num_output: 1024
2844 |     bias_term: false
2845 |     pad: 0
2846 |     kernel_size: 1
2847 |     stride: 1
2848 |   }
2849 | }
2850 | layer {
2851 |   name: "conv4_2_1x1_increase/bn"
2852 |   type: "BatchNorm"
2853 |   bottom: "conv4_2_1x1_increase"
2854 |   top: "conv4_2_1x1_increase"
2855 |   batch_norm_param {
2856 |     use_global_stats: true
2857 |   }
2858 | }
2859 | layer {
2860 |   name: "conv4_2_1x1_increase/scale"
2861 |   type: "Scale"
2862 |   bottom: "conv4_2_1x1_increase"
2863 |   top: "conv4_2_1x1_increase"
2864 |   scale_param {
2865 |     bias_term: true
2866 |   }
2867 | }
2868 | layer {
2869 |   name: "conv4_2"
2870 |   type: "Eltwise"
2871 |   bottom: "conv4_1"
2872 |   bottom: "conv4_2_1x1_increase"
2873 |   top: "conv4_2"
2874 |   eltwise_param {
2875 |     operation: SUM
2876 |   }
2877 | }
2878 | layer {
2879 |   name: "conv4_2/relu"
2880 |   type: "ReLU"
2881 |   bottom: "conv4_2"
2882 |   top: "conv4_2"
2883 | }
2884 | layer {
2885 |   name: "conv4_3_1x1_reduce"
2886 |   type: "Convolution"
2887 |   bottom: "conv4_2"
2888 |   top: "conv4_3_1x1_reduce"
2889 |   convolution_param {
2890 |     num_output: 256
2891 |     bias_term: false
2892 |     pad: 0
2893 |     kernel_size: 1
2894 |     stride: 1
2895 |   }
2896 | }
2897 | layer {
2898 |   name: "conv4_3_1x1_reduce/bn"
2899 |   type: "BatchNorm"
2900 |   bottom: "conv4_3_1x1_reduce"
2901 |   top: "conv4_3_1x1_reduce"
2902 |   batch_norm_param {
2903 |     use_global_stats: true
2904 |   }
2905 | }
2906 | layer {
2907 |   name: "conv4_3_1x1_reduce/scale"
2908 |   type: "Scale"
2909 |   bottom: "conv4_3_1x1_reduce"
2910 |   top: "conv4_3_1x1_reduce"
2911 |   scale_param {
2912 |     bias_term: true
2913 |   }
2914 | }
2915 | layer {
2916 |   name: "conv4_3_1x1_reduce/relu"
2917 |   type: "ReLU"
2918 |   bottom: "conv4_3_1x1_reduce"
2919 |   top: "conv4_3_1x1_reduce"
2920 | }
2921 | layer {
2922 |   name: "conv4_3_3x3"
2923 |   type: "Convolution"
2924 |   bottom: "conv4_3_1x1_reduce"
2925 |   top: "conv4_3_3x3"
2926 |   convolution_param {
2927 |     num_output: 256
2928 |     bias_term: false
2929 |     pad: 1
2930 |     kernel_size: 3
2931 |     stride: 1
2932 |   }
2933 | }
2934 | layer {
2935 |   name: "conv4_3_3x3/bn"
2936 |   type: "BatchNorm"
2937 |   bottom: "conv4_3_3x3"
2938 |   top: "conv4_3_3x3"
2939 |   batch_norm_param {
2940 |     use_global_stats: true
2941 |   }
2942 | }
2943 | layer {
2944 |   name: "conv4_3_3x3/scale"
2945 |   type: "Scale"
2946 |   bottom: "conv4_3_3x3"
2947 |   top: "conv4_3_3x3"
2948 |   scale_param {
2949 |     bias_term: true
2950 |   }
2951 | }
2952 | layer {
2953 |   name: "conv4_3_3x3/relu"
2954 |   type: "ReLU"
2955 |   bottom: "conv4_3_3x3"
2956 |   top: "conv4_3_3x3"
2957 | }
2958 | layer {
2959 |   name: "conv4_3_1x1_increase"
2960 |   type: "Convolution"
2961 |   bottom: "conv4_3_3x3"
2962 |   top: "conv4_3_1x1_increase"
2963 |   convolution_param {
2964 |     num_output: 1024
2965 |     bias_term: false
2966 |     pad: 0
2967 |     kernel_size: 1
2968 |     stride: 1
2969 |   }
2970 | }
2971 | layer {
2972 |   name: "conv4_3_1x1_increase/bn"
2973 |   type: "BatchNorm"
2974 |   bottom: "conv4_3_1x1_increase"
2975 |   top: "conv4_3_1x1_increase"
2976 |   batch_norm_param {
2977 |     use_global_stats: true
2978 |   }
2979 | }
2980 | layer {
2981 |   name: "conv4_3_1x1_increase/scale"
2982 |   type: "Scale"
2983 |   bottom: "conv4_3_1x1_increase"
2984 |   top: "conv4_3_1x1_increase"
2985 |   scale_param {
2986 |     bias_term: true
2987 |   }
2988 | }
2989 | layer {
2990 |   name: "conv4_3"
2991 |   type: "Eltwise"
2992 |   bottom: "conv4_2"
2993 |   bottom: "conv4_3_1x1_increase"
2994 |   top: "conv4_3"
2995 |   eltwise_param {
2996 |     operation: SUM
2997 |   }
2998 | }
2999 | layer {
3000 |   name: "conv4_3/relu"
3001 |   type: "ReLU"
3002 |   bottom: "conv4_3"
3003 |   top: "conv4_3"
3004 | }
3005 | layer {
3006 |   name: "conv4_4_1x1_reduce"
3007 |   type: "Convolution"
3008 |   bottom: "conv4_3"
3009 |   top: "conv4_4_1x1_reduce"
3010 |   convolution_param {
3011 |     num_output: 256
3012 |     bias_term: false
3013 |     pad: 0
3014 |     kernel_size: 1
3015 |     stride: 1
3016 |   }
3017 | }
3018 | layer {
3019 |   name: "conv4_4_1x1_reduce/bn"
3020 |   type: "BatchNorm"
3021 |   bottom: "conv4_4_1x1_reduce"
3022 |   top: "conv4_4_1x1_reduce"
3023 |   batch_norm_param {
3024 |     use_global_stats: true
3025 |   }
3026 | }
3027 | layer {
3028 |   name: "conv4_4_1x1_reduce/scale"
3029 |   type: "Scale"
3030 |   bottom: "conv4_4_1x1_reduce"
3031 |   top: "conv4_4_1x1_reduce"
3032 |   scale_param {
3033 |     bias_term: true
3034 |   }
3035 | }
3036 | layer {
3037 |   name: "conv4_4_1x1_reduce/relu"
3038 |   type: "ReLU"
3039 |   bottom: "conv4_4_1x1_reduce"
3040 |   top: "conv4_4_1x1_reduce"
3041 | }
3042 | layer {
3043 |   name: "conv4_4_3x3"
3044 |   type: "Convolution"
3045 |   bottom: "conv4_4_1x1_reduce"
3046 |   top: "conv4_4_3x3"
3047 |   convolution_param {
3048 |     num_output: 256
3049 |     bias_term: false
3050 |     pad: 1
3051 |     kernel_size: 3
3052 |     stride: 1
3053 |   }
3054 | }
3055 | layer {
3056 |   name: "conv4_4_3x3/bn"
3057 |   type: "BatchNorm"
3058 |   bottom: "conv4_4_3x3"
3059 |   top: "conv4_4_3x3"
3060 |   batch_norm_param {
3061 |     use_global_stats: true
3062 |   }
3063 | }
3064 | layer {
3065 |   name: "conv4_4_3x3/scale"
3066 |   type: "Scale"
3067 |   bottom: "conv4_4_3x3"
3068 |   top: "conv4_4_3x3"
3069 |   scale_param {
3070 |     bias_term: true
3071 |   }
3072 | }
3073 | layer {
3074 |   name: "conv4_4_3x3/relu"
3075 |   type: "ReLU"
3076 |   bottom: "conv4_4_3x3"
3077 |   top: "conv4_4_3x3"
3078 | }
3079 | layer {
3080 |   name: "conv4_4_1x1_increase"
3081 |   type: "Convolution"
3082 |   bottom: "conv4_4_3x3"
3083 |   top: "conv4_4_1x1_increase"
3084 |   convolution_param {
3085 |     num_output: 1024
3086 |     bias_term: false
3087 |     pad: 0
3088 |     kernel_size: 1
3089 |     stride: 1
3090 |   }
3091 | }
3092 | layer {
3093 |   name: "conv4_4_1x1_increase/bn"
3094 |   type: "BatchNorm"
3095 |   bottom: "conv4_4_1x1_increase"
3096 |   top: "conv4_4_1x1_increase"
3097 |   batch_norm_param {
3098 |     use_global_stats: true
3099 |   }
3100 | }
3101 | layer {
3102 |   name: "conv4_4_1x1_increase/scale"
3103 |   type: "Scale"
3104 |   bottom: "conv4_4_1x1_increase"
3105 |   top: "conv4_4_1x1_increase"
3106 |   scale_param {
3107 |     bias_term: true
3108 |   }
3109 | }
3110 | layer {
3111 |   name: "conv4_4"
3112 |   type: "Eltwise"
3113 |   bottom: "conv4_3"
3114 |   bottom: "conv4_4_1x1_increase"
3115 |   top: "conv4_4"
3116 |   eltwise_param {
3117 |     operation: SUM
3118 |   }
3119 | }
3120 | layer {
3121 |   name: "conv4_4/relu"
3122 |   type: "ReLU"
3123 |   bottom: "conv4_4"
3124 |   top: "conv4_4"
3125 | }
3126 | layer {
3127 |   name: "conv4_5_1x1_reduce"
3128 |   type: "Convolution"
3129 |   bottom: "conv4_4"
3130 |   top: "conv4_5_1x1_reduce"
3131 |   convolution_param {
3132 |     num_output: 256
3133 |     bias_term: false
3134 |     pad: 0
3135 |     kernel_size: 1
3136 |     stride: 1
3137 |   }
3138 | }
3139 | layer {
3140 |   name: "conv4_5_1x1_reduce/bn"
3141 |   type: "BatchNorm"
3142 |   bottom: "conv4_5_1x1_reduce"
3143 |   top: "conv4_5_1x1_reduce"
3144 |   batch_norm_param {
3145 |     use_global_stats: true
3146 |   }
3147 | }
3148 | layer {
3149 |   name: "conv4_5_1x1_reduce/scale"
3150 |   type: "Scale"
3151 |   bottom: "conv4_5_1x1_reduce"
3152 |   top: "conv4_5_1x1_reduce"
3153 |   scale_param {
3154 |     bias_term: true
3155 |   }
3156 | }
3157 | layer {
3158 |   name: "conv4_5_1x1_reduce/relu"
3159 |   type: "ReLU"
3160 |   bottom: "conv4_5_1x1_reduce"
3161 |   top: "conv4_5_1x1_reduce"
3162 | }
3163 | layer {
3164 |   name: "conv4_5_3x3"
3165 |   type: "Convolution"
3166 |   bottom: "conv4_5_1x1_reduce"
3167 |   top: "conv4_5_3x3"
3168 |   convolution_param {
3169 |     num_output: 256
3170 |     bias_term: false
3171 |     pad: 1
3172 |     kernel_size: 3
3173 |     stride: 1
3174 |   }
3175 | }
3176 | layer {
3177 |   name: "conv4_5_3x3/bn"
3178 |   type: "BatchNorm"
3179 |   bottom: "conv4_5_3x3"
3180 |   top: "conv4_5_3x3"
3181 |   batch_norm_param {
3182 |     use_global_stats: true
3183 |   }
3184 | }
3185 | layer {
3186 |   name: "conv4_5_3x3/scale"
3187 |   type: "Scale"
3188 |   bottom: "conv4_5_3x3"
3189 |   top: "conv4_5_3x3"
3190 |   scale_param {
3191 |     bias_term: true
3192 |   }
3193 | }
3194 | layer {
3195 |   name: "conv4_5_3x3/relu"
3196 |   type: "ReLU"
3197 |   bottom: "conv4_5_3x3"
3198 |   top: "conv4_5_3x3"
3199 | }
3200 | layer {
3201 |   name: "conv4_5_1x1_increase"
3202 |   type: "Convolution"
3203 |   bottom: "conv4_5_3x3"
3204 |   top: "conv4_5_1x1_increase"
3205 |   convolution_param {
3206 |     num_output: 1024
3207 |     bias_term: false
3208 |     pad: 0
3209 |     kernel_size: 1
3210 |     stride: 1
3211 |   }
3212 | }
3213 | layer {
3214 |   name: "conv4_5_1x1_increase/bn"
3215 |   type: "BatchNorm"
3216 |   bottom: "conv4_5_1x1_increase"
3217 |   top: "conv4_5_1x1_increase"
3218 |   batch_norm_param {
3219 |     use_global_stats: true
3220 |   }
3221 | }
3222 | layer {
3223 |   name: "conv4_5_1x1_increase/scale"
3224 |   type: "Scale"
3225 |   bottom: "conv4_5_1x1_increase"
3226 |   top: "conv4_5_1x1_increase"
3227 |   scale_param {
3228 |     bias_term: true
3229 |   }
3230 | }
3231 | layer {
3232 |   name: "conv4_5"
3233 |   type: "Eltwise"
3234 |   bottom: "conv4_4"
3235 |   bottom: "conv4_5_1x1_increase"
3236 |   top: "conv4_5"
3237 |   eltwise_param {
3238 |     operation: SUM
3239 |   }
3240 | }
3241 | layer {
3242 |   name: "conv4_5/relu"
3243 |   type: "ReLU"
3244 |   bottom: "conv4_5"
3245 |   top: "conv4_5"
3246 | }
3247 | layer {
3248 |   name: "conv4_6_1x1_reduce"
3249 |   type: "Convolution"
3250 |   bottom: "conv4_5"
3251 |   top: "conv4_6_1x1_reduce"
3252 |   convolution_param {
3253 |     num_output: 256
3254 |     bias_term: false
3255 |     pad: 0
3256 |     kernel_size: 1
3257 |     stride: 1
3258 |   }
3259 | }
3260 | layer {
3261 |   name: "conv4_6_1x1_reduce/bn"
3262 |   type: "BatchNorm"
3263 |   bottom: "conv4_6_1x1_reduce"
3264 |   top: "conv4_6_1x1_reduce"
3265 |   batch_norm_param {
3266 |     use_global_stats: true
3267 |   }
3268 | }
3269 | layer {
3270 |   name: "conv4_6_1x1_reduce/scale"
3271 |   type: "Scale"
3272 |   bottom: "conv4_6_1x1_reduce"
3273 |   top: "conv4_6_1x1_reduce"
3274 |   scale_param {
3275 |     bias_term: true
3276 |   }
3277 | }
3278 | layer {
3279 |   name: "conv4_6_1x1_reduce/relu"
3280 |   type: "ReLU"
3281 |   bottom: "conv4_6_1x1_reduce"
3282 |   top: "conv4_6_1x1_reduce"
3283 | }
3284 | layer {
3285 |   name: "conv4_6_3x3"
3286 |   type: "Convolution"
3287 |   bottom: "conv4_6_1x1_reduce"
3288 |   top: "conv4_6_3x3"
3289 |   convolution_param {
3290 |     num_output: 256
3291 |     bias_term: false
3292 |     pad: 1
3293 |     kernel_size: 3
3294 |     stride: 1
3295 |   }
3296 | }
3297 | layer {
3298 |   name: "conv4_6_3x3/bn"
3299 |   type: "BatchNorm"
3300 |   bottom: "conv4_6_3x3"
3301 |   top: "conv4_6_3x3"
3302 |   batch_norm_param {
3303 |     use_global_stats: true
3304 |   }
3305 | }
3306 | layer {
3307 |   name: "conv4_6_3x3/scale"
3308 |   type: "Scale"
3309 |   bottom: "conv4_6_3x3"
3310 |   top: "conv4_6_3x3"
3311 |   scale_param {
3312 |     bias_term: true
3313 |   }
3314 | }
3315 | layer {
3316 |   name: "conv4_6_3x3/relu"
3317 |   type: "ReLU"
3318 |   bottom: "conv4_6_3x3"
3319 |   top: "conv4_6_3x3"
3320 | }
3321 | layer {
3322 |   name: "conv4_6_1x1_increase"
3323 |   type: "Convolution"
3324 |   bottom: "conv4_6_3x3"
3325 |   top: "conv4_6_1x1_increase"
3326 |   convolution_param {
3327 |     num_output: 1024
3328 |     bias_term: false
3329 |     pad: 0
3330 |     kernel_size: 1
3331 |     stride: 1
3332 |   }
3333 | }
3334 | layer {
3335 |   name: "conv4_6_1x1_increase/bn"
3336 |   type: "BatchNorm"
3337 |   bottom: "conv4_6_1x1_increase"
3338 |   top: "conv4_6_1x1_increase"
3339 |   batch_norm_param {
3340 |     use_global_stats: true
3341 |   }
3342 | }
3343 | layer {
3344 |   name: "conv4_6_1x1_increase/scale"
3345 |   type: "Scale"
3346 |   bottom: "conv4_6_1x1_increase"
3347 |   top: "conv4_6_1x1_increase"
3348 |   scale_param {
3349 |     bias_term: true
3350 |   }
3351 | }
3352 | layer {
3353 |   name: "conv4_6"
3354 |   type: "Eltwise"
3355 |   bottom: "conv4_5"
3356 |   bottom: "conv4_6_1x1_increase"
3357 |   top: "conv4_6"
3358 |   eltwise_param {
3359 |     operation: SUM
3360 |   }
3361 | }
3362 | layer {
3363 |   name: "conv4_6/relu"
3364 |   type: "ReLU"
3365 |   bottom: "conv4_6"
3366 |   top: "conv4_6"
3367 | }
3368 | layer {
3369 |   name: "conv4_7_1x1_reduce"
3370 |   type: "Convolution"
3371 |   bottom: "conv4_6"
3372 |   top: "conv4_7_1x1_reduce"
3373 |   convolution_param {
3374 |     num_output: 256
3375 |     bias_term: false
3376 |     pad: 0
3377 |     kernel_size: 1
3378 |     stride: 1
3379 |   }
3380 | }
3381 | layer {
3382 |   name: "conv4_7_1x1_reduce/bn"
3383 |   type: "BatchNorm"
3384 |   bottom: "conv4_7_1x1_reduce"
3385 |   top: "conv4_7_1x1_reduce"
3386 |   batch_norm_param {
3387 |     use_global_stats: true
3388 |   }
3389 | }
3390 | layer {
3391 |   name: "conv4_7_1x1_reduce/scale"
3392 |   type: "Scale"
3393 |   bottom: "conv4_7_1x1_reduce"
3394 |   top: "conv4_7_1x1_reduce"
3395 |   scale_param {
3396 |     bias_term: true
3397 |   }
3398 | }
3399 | layer {
3400 |   name: "conv4_7_1x1_reduce/relu"
3401 |   type: "ReLU"
3402 |   bottom: "conv4_7_1x1_reduce"
3403 |   top: "conv4_7_1x1_reduce"
3404 | }
3405 | layer {
3406 |   name: "conv4_7_3x3"
3407 |   type: "Convolution"
3408 |   bottom: "conv4_7_1x1_reduce"
3409 |   top: "conv4_7_3x3"
3410 |   convolution_param {
3411 |     num_output: 256
3412 |     bias_term: false
3413 |     pad: 1
3414 |     kernel_size: 3
3415 |     stride: 1
3416 |   }
3417 | }
3418 | layer {
3419 |   name: "conv4_7_3x3/bn"
3420 |   type: "BatchNorm"
3421 |   bottom: "conv4_7_3x3"
3422 |   top: "conv4_7_3x3"
3423 |   batch_norm_param {
3424 |     use_global_stats: true
3425 |   }
3426 | }
3427 | layer {
3428 |   name: "conv4_7_3x3/scale"
3429 |   type: "Scale"
3430 |   bottom: "conv4_7_3x3"
3431 |   top: "conv4_7_3x3"
3432 |   scale_param {
3433 |     bias_term: true
3434 |   }
3435 | }
3436 | layer {
3437 |   name: "conv4_7_3x3/relu"
3438 |   type: "ReLU"
3439 |   bottom: "conv4_7_3x3"
3440 |   top: "conv4_7_3x3"
3441 | }
3442 | layer {
3443 |   name: "conv4_7_1x1_increase"
3444 |   type: "Convolution"
3445 |   bottom: "conv4_7_3x3"
3446 |   top: "conv4_7_1x1_increase"
3447 |   convolution_param {
3448 |     num_output: 1024
3449 |     bias_term: false
3450 |     pad: 0
3451 |     kernel_size: 1
3452 |     stride: 1
3453 |   }
3454 | }
3455 | layer {
3456 |   name: "conv4_7_1x1_increase/bn"
3457 |   type: "BatchNorm"
3458 |   bottom: "conv4_7_1x1_increase"
3459 |   top: "conv4_7_1x1_increase"
3460 |   batch_norm_param {
3461 |     use_global_stats: true
3462 |   }
3463 | }
3464 | layer {
3465 |   name: "conv4_7_1x1_increase/scale"
3466 |   type: "Scale"
3467 |   bottom: "conv4_7_1x1_increase"
3468 |   top: "conv4_7_1x1_increase"
3469 |   scale_param {
3470 |     bias_term: true
3471 |   }
3472 | }
3473 | layer {
3474 |   name: "conv4_7"
3475 |   type: "Eltwise"
3476 |   bottom: "conv4_6"
3477 |   bottom: "conv4_7_1x1_increase"
3478 |   top: "conv4_7"
3479 |   eltwise_param {
3480 |     operation: SUM
3481 |   }
3482 | }
3483 | layer {
3484 |   name: "conv4_7/relu"
3485 |   type: "ReLU"
3486 |   bottom: "conv4_7"
3487 |   top: "conv4_7"
3488 | }
3489 | layer {
3490 |   name: "conv4_8_1x1_reduce"
3491 |   type: "Convolution"
3492 |   bottom: "conv4_7"
3493 |   top: "conv4_8_1x1_reduce"
3494 |   convolution_param {
3495 |     num_output: 256
3496 |     bias_term: false
3497 |     pad: 0
3498 |     kernel_size: 1
3499 |     stride: 1
3500 |   }
3501 | }
3502 | layer {
3503 |   name: "conv4_8_1x1_reduce/bn"
3504 |   type: "BatchNorm"
3505 |   bottom: "conv4_8_1x1_reduce"
3506 |   top: "conv4_8_1x1_reduce"
3507 |   batch_norm_param {
3508 |     use_global_stats: true
3509 |   }
3510 | }
3511 | layer {
3512 |   name: "conv4_8_1x1_reduce/scale"
3513 |   type: "Scale"
3514 |   bottom: "conv4_8_1x1_reduce"
3515 |   top: "conv4_8_1x1_reduce"
3516 |   scale_param {
3517 |     bias_term: true
3518 |   }
3519 | }
3520 | layer {
3521 |   name: "conv4_8_1x1_reduce/relu"
3522 |   type: "ReLU"
3523 |   bottom: "conv4_8_1x1_reduce"
3524 |   top: "conv4_8_1x1_reduce"
3525 | }
3526 | layer {
3527 |   name: "conv4_8_3x3"
3528 |   type: "Convolution"
3529 |   bottom: "conv4_8_1x1_reduce"
3530 |   top: "conv4_8_3x3"
3531 |   convolution_param {
3532 |     num_output: 256
3533 |     bias_term: false
3534 |     pad: 1
3535 |     kernel_size: 3
3536 |     stride: 1
3537 |   }
3538 | }
3539 | layer {
3540 |   name: "conv4_8_3x3/bn"
3541 |   type: "BatchNorm"
3542 |   bottom: "conv4_8_3x3"
3543 |   top: "conv4_8_3x3"
3544 |   batch_norm_param {
3545 |     use_global_stats: true
3546 |   }
3547 | }
3548 | layer {
3549 |   name: "conv4_8_3x3/scale"
3550 |   type: "Scale"
3551 |   bottom: "conv4_8_3x3"
3552 |   top: "conv4_8_3x3"
3553 |   scale_param {
3554 |     bias_term: true
3555 |   }
3556 | }
3557 | layer {
3558 |   name: "conv4_8_3x3/relu"
3559 |   type: "ReLU"
3560 |   bottom: "conv4_8_3x3"
3561 |   top: "conv4_8_3x3"
3562 | }
3563 | layer {
3564 |   name: "conv4_8_1x1_increase"
3565 |   type: "Convolution"
3566 |   bottom: "conv4_8_3x3"
3567 |   top: "conv4_8_1x1_increase"
3568 |   convolution_param {
3569 |     num_output: 1024
3570 |     bias_term: false
3571 |     pad: 0
3572 |     kernel_size: 1
3573 |     stride: 1
3574 |   }
3575 | }
3576 | layer {
3577 |   name: "conv4_8_1x1_increase/bn"
3578 |   type: "BatchNorm"
3579 |   bottom: "conv4_8_1x1_increase"
3580 |   top: "conv4_8_1x1_increase"
3581 |   batch_norm_param {
3582 |     use_global_stats: true
3583 |   }
3584 | }
3585 | layer {
3586 |   name: "conv4_8_1x1_increase/scale"
3587 |   type: "Scale"
3588 |   bottom: "conv4_8_1x1_increase"
3589 |   top: "conv4_8_1x1_increase"
3590 |   scale_param {
3591 |     bias_term: true
3592 |   }
3593 | }
3594 | layer {
3595 |   name: "conv4_8"
3596 |   type: "Eltwise"
3597 |   bottom: "conv4_7"
3598 |   bottom: "conv4_8_1x1_increase"
3599 |   top: "conv4_8"
3600 |   eltwise_param {
3601 |     operation: SUM
3602 |   }
3603 | }
3604 | layer {
3605 |   name: "conv4_8/relu"
3606 |   type: "ReLU"
3607 |   bottom: "conv4_8"
3608 |   top: "conv4_8"
3609 | }
3610 | layer {
3611 |   name: "conv4_9_1x1_reduce"
3612 |   type: "Convolution"
3613 |   bottom: "conv4_8"
3614 |   top: "conv4_9_1x1_reduce"
3615 |   convolution_param {
3616 |     num_output: 256
3617 |     bias_term: false
3618 |     pad: 0
3619 |     kernel_size: 1
3620 |     stride: 1
3621 |   }
3622 | }
3623 | layer {
3624 |   name: "conv4_9_1x1_reduce/bn"
3625 |   type: "BatchNorm"
3626 |   bottom: "conv4_9_1x1_reduce"
3627 |   top: "conv4_9_1x1_reduce"
3628 |   batch_norm_param {
3629 |     use_global_stats: true
3630 |   }
3631 | }
3632 | layer {
3633 |   name: "conv4_9_1x1_reduce/scale"
3634 |   type: "Scale"
3635 |   bottom: "conv4_9_1x1_reduce"
3636 |   top: "conv4_9_1x1_reduce"
3637 |   scale_param {
3638 |     bias_term: true
3639 |   }
3640 | }
3641 | layer {
3642 |   name: "conv4_9_1x1_reduce/relu"
3643 |   type: "ReLU"
3644 |   bottom: "conv4_9_1x1_reduce"
3645 |   top: "conv4_9_1x1_reduce"
3646 | }
3647 | layer {
3648 |   name: "conv4_9_3x3"
3649 |   type: "Convolution"
3650 |   bottom: "conv4_9_1x1_reduce"
3651 |   top: "conv4_9_3x3"
3652 |   convolution_param {
3653 |     num_output: 256
3654 |     bias_term: false
3655 |     pad: 1
3656 |     kernel_size: 3
3657 |     stride: 1
3658 |   }
3659 | }
3660 | layer {
3661 |   name: "conv4_9_3x3/bn"
3662 |   type: "BatchNorm"
3663 |   bottom: "conv4_9_3x3"
3664 |   top: "conv4_9_3x3"
3665 |   batch_norm_param {
3666 |     use_global_stats: true
3667 |   }
3668 | }
3669 | layer {
3670 |   name: "conv4_9_3x3/scale"
3671 |   type: "Scale"
3672 |   bottom: "conv4_9_3x3"
3673 |   top: "conv4_9_3x3"
3674 |   scale_param {
3675 |     bias_term: true
3676 |   }
3677 | }
3678 | layer {
3679 |   name: "conv4_9_3x3/relu"
3680 |   type: "ReLU"
3681 |   bottom: "conv4_9_3x3"
3682 |   top: "conv4_9_3x3"
3683 | }
3684 | layer {
3685 |   name: "conv4_9_1x1_increase"
3686 |   type: "Convolution"
3687 |   bottom: "conv4_9_3x3"
3688 |   top: "conv4_9_1x1_increase"
3689 |   convolution_param {
3690 |     num_output: 1024
3691 |     bias_term: false
3692 |     pad: 0
3693 |     kernel_size: 1
3694 |     stride: 1
3695 |   }
3696 | }
3697 | layer {
3698 |   name: "conv4_9_1x1_increase/bn"
3699 |   type: "BatchNorm"
3700 |   bottom: "conv4_9_1x1_increase"
3701 |   top: "conv4_9_1x1_increase"
3702 |   batch_norm_param {
3703 |     use_global_stats: true
3704 |   }
3705 | }
3706 | layer {
3707 |   name: "conv4_9_1x1_increase/scale"
3708 |   type: "Scale"
3709 |   bottom: "conv4_9_1x1_increase"
3710 |   top: "conv4_9_1x1_increase"
3711 |   scale_param {
3712 |     bias_term: true
3713 |   }
3714 | }
3715 | layer {
3716 |   name: "conv4_9"
3717 |   type: "Eltwise"
3718 |   bottom: "conv4_8"
3719 |   bottom: "conv4_9_1x1_increase"
3720 |   top: "conv4_9"
3721 |   eltwise_param {
3722 |     operation: SUM
3723 |   }
3724 | }
3725 | layer {
3726 |   name: "conv4_9/relu"
3727 |   type: "ReLU"
3728 |   bottom: "conv4_9"
3729 |   top: "conv4_9"
3730 | }
3731 | layer {
3732 |   name: "conv4_10_1x1_reduce"
3733 |   type: "Convolution"
3734 |   bottom: "conv4_9"
3735 |   top: "conv4_10_1x1_reduce"
3736 |   convolution_param {
3737 |     num_output: 256
3738 |     bias_term: false
3739 |     pad: 0
3740 |     kernel_size: 1
3741 |     stride: 1
3742 |   }
3743 | }
3744 | layer {
3745 |   name: "conv4_10_1x1_reduce/bn"
3746 |   type: "BatchNorm"
3747 |   bottom: "conv4_10_1x1_reduce"
3748 |   top: "conv4_10_1x1_reduce"
3749 |   batch_norm_param {
3750 |     use_global_stats: true
3751 |   }
3752 | }
3753 | layer {
3754 |   name: "conv4_10_1x1_reduce/scale"
3755 |   type: "Scale"
3756 |   bottom: "conv4_10_1x1_reduce"
3757 |   top: "conv4_10_1x1_reduce"
3758 |   scale_param {
3759 |     bias_term: true
3760 |   }
3761 | }
3762 | layer {
3763 |   name: "conv4_10_1x1_reduce/relu"
3764 |   type: "ReLU"
3765 |   bottom: "conv4_10_1x1_reduce"
3766 |   top: "conv4_10_1x1_reduce"
3767 | }
3768 | layer {
3769 |   name: "conv4_10_3x3"
3770 |   type: "Convolution"
3771 |   bottom: "conv4_10_1x1_reduce"
3772 |   top: "conv4_10_3x3"
3773 |   convolution_param {
3774 |     num_output: 256
3775 |     bias_term: false
3776 |     pad: 1
3777 |     kernel_size: 3
3778 |     stride: 1
3779 |   }
3780 | }
3781 | layer {
3782 |   name: "conv4_10_3x3/bn"
3783 |   type: "BatchNorm"
3784 |   bottom: "conv4_10_3x3"
3785 |   top: "conv4_10_3x3"
3786 |   batch_norm_param {
3787 |     use_global_stats: true
3788 |   }
3789 | }
3790 | layer {
3791 |   name: "conv4_10_3x3/scale"
3792 |   type: "Scale"
3793 |   bottom: "conv4_10_3x3"
3794 |   top: "conv4_10_3x3"
3795 |   scale_param {
3796 |     bias_term: true
3797 |   }
3798 | }
3799 | layer {
3800 |   name: "conv4_10_3x3/relu"
3801 |   type: "ReLU"
3802 |   bottom: "conv4_10_3x3"
3803 |   top: "conv4_10_3x3"
3804 | }
3805 | layer {
3806 |   name: "conv4_10_1x1_increase"
3807 |   type: "Convolution"
3808 |   bottom: "conv4_10_3x3"
3809 |   top: "conv4_10_1x1_increase"
3810 |   convolution_param {
3811 |     num_output: 1024
3812 |     bias_term: false
3813 |     pad: 0
3814 |     kernel_size: 1
3815 |     stride: 1
3816 |   }
3817 | }
3818 | layer {
3819 |   name: "conv4_10_1x1_increase/bn"
3820 |   type: "BatchNorm"
3821 |   bottom: "conv4_10_1x1_increase"
3822 |   top: "conv4_10_1x1_increase"
3823 |   batch_norm_param {
3824 |     use_global_stats: true
3825 |   }
3826 | }
3827 | layer {
3828 |   name: "conv4_10_1x1_increase/scale"
3829 |   type: "Scale"
3830 |   bottom: "conv4_10_1x1_increase"
3831 |   top: "conv4_10_1x1_increase"
3832 |   scale_param {
3833 |     bias_term: true
3834 |   }
3835 | }
3836 | layer {
3837 |   name: "conv4_10"
3838 |   type: "Eltwise"
3839 |   bottom: "conv4_9"
3840 |   bottom: "conv4_10_1x1_increase"
3841 |   top: "conv4_10"
3842 |   eltwise_param {
3843 |     operation: SUM
3844 |   }
3845 | }
3846 | layer {
3847 |   name: "conv4_10/relu"
3848 |   type: "ReLU"
3849 |   bottom: "conv4_10"
3850 |   top: "conv4_10"
3851 | }
3852 | layer {
3853 |   name: "conv4_11_1x1_reduce"
3854 |   type: "Convolution"
3855 |   bottom: "conv4_10"
3856 |   top: "conv4_11_1x1_reduce"
3857 |   convolution_param {
3858 |     num_output: 256
3859 |     bias_term: false
3860 |     pad: 0
3861 |     kernel_size: 1
3862 |     stride: 1
3863 |   }
3864 | }
3865 | layer {
3866 |   name: "conv4_11_1x1_reduce/bn"
3867 |   type: "BatchNorm"
3868 |   bottom: "conv4_11_1x1_reduce"
3869 |   top: "conv4_11_1x1_reduce"
3870 |   batch_norm_param {
3871 |     use_global_stats: true
3872 |   }
3873 | }
3874 | layer {
3875 |   name: "conv4_11_1x1_reduce/scale"
3876 |   type: "Scale"
3877 |   bottom: "conv4_11_1x1_reduce"
3878 |   top: "conv4_11_1x1_reduce"
3879 |   scale_param {
3880 |     bias_term: true
3881 |   }
3882 | }
3883 | layer {
3884 |   name: "conv4_11_1x1_reduce/relu"
3885 |   type: "ReLU"
3886 |   bottom: "conv4_11_1x1_reduce"
3887 |   top: "conv4_11_1x1_reduce"
3888 | }
3889 | layer {
3890 |   name: "conv4_11_3x3"
3891 |   type: "Convolution"
3892 |   bottom: "conv4_11_1x1_reduce"
3893 |   top: "conv4_11_3x3"
3894 |   convolution_param {
3895 |     num_output: 256
3896 |     bias_term: false
3897 |     pad: 1
3898 |     kernel_size: 3
3899 |     stride: 1
3900 |   }
3901 | }
3902 | layer {
3903 |   name: "conv4_11_3x3/bn"
3904 |   type: "BatchNorm"
3905 |   bottom: "conv4_11_3x3"
3906 |   top: "conv4_11_3x3"
3907 |   batch_norm_param {
3908 |     use_global_stats: true
3909 |   }
3910 | }
3911 | layer {
3912 |   name: "conv4_11_3x3/scale"
3913 |   type: "Scale"
3914 |   bottom: "conv4_11_3x3"
3915 |   top: "conv4_11_3x3"
3916 |   scale_param {
3917 |     bias_term: true
3918 |   }
3919 | }
3920 | layer {
3921 |   name: "conv4_11_3x3/relu"
3922 |   type: "ReLU"
3923 |   bottom: "conv4_11_3x3"
3924 |   top: "conv4_11_3x3"
3925 | }
3926 | layer {
3927 |   name: "conv4_11_1x1_increase"
3928 |   type: "Convolution"
3929 |   bottom: "conv4_11_3x3"
3930 |   top: "conv4_11_1x1_increase"
3931 |   convolution_param {
3932 |     num_output: 1024
3933 |     bias_term: false
3934 |     pad: 0
3935 |     kernel_size: 1
3936 |     stride: 1
3937 |   }
3938 | }
3939 | layer {
3940 |   name: "conv4_11_1x1_increase/bn"
3941 |   type: "BatchNorm"
3942 |   bottom: "conv4_11_1x1_increase"
3943 |   top: "conv4_11_1x1_increase"
3944 |   batch_norm_param {
3945 |     use_global_stats: true
3946 |   }
3947 | }
3948 | layer {
3949 |   name: "conv4_11_1x1_increase/scale"
3950 |   type: "Scale"
3951 |   bottom: "conv4_11_1x1_increase"
3952 |   top: "conv4_11_1x1_increase"
3953 |   scale_param {
3954 |     bias_term: true
3955 |   }
3956 | }
3957 | layer {
3958 |   name: "conv4_11"
3959 |   type: "Eltwise"
3960 |   bottom: "conv4_10"
3961 |   bottom: "conv4_11_1x1_increase"
3962 |   top: "conv4_11"
3963 |   eltwise_param {
3964 |     operation: SUM
3965 |   }
3966 | }
3967 | layer {
3968 |   name: "conv4_11/relu"
3969 |   type: "ReLU"
3970 |   bottom: "conv4_11"
3971 |   top: "conv4_11"
3972 | }
3973 | layer {
3974 |   name: "conv4_12_1x1_reduce"
3975 |   type: "Convolution"
3976 |   bottom: "conv4_11"
3977 |   top: "conv4_12_1x1_reduce"
3978 |   convolution_param {
3979 |     num_output: 256
3980 |     bias_term: false
3981 |     pad: 0
3982 |     kernel_size: 1
3983 |     stride: 1
3984 |   }
3985 | }
3986 | layer {
3987 |   name: "conv4_12_1x1_reduce/bn"
3988 |   type: "BatchNorm"
3989 |   bottom: "conv4_12_1x1_reduce"
3990 |   top: "conv4_12_1x1_reduce"
3991 |   batch_norm_param {
3992 |     use_global_stats: true
3993 |   }
3994 | }
3995 | layer {
3996 |   name: "conv4_12_1x1_reduce/scale"
3997 |   type: "Scale"
3998 |   bottom: "conv4_12_1x1_reduce"
3999 |   top: "conv4_12_1x1_reduce"
4000 |   scale_param {
4001 |     bias_term: true
4002 |   }
4003 | }
4004 | layer {
4005 |   name: "conv4_12_1x1_reduce/relu"
4006 |   type: "ReLU"
4007 |   bottom: "conv4_12_1x1_reduce"
4008 |   top: "conv4_12_1x1_reduce"
4009 | }
4010 | layer {
4011 |   name: "conv4_12_3x3"
4012 |   type: "Convolution"
4013 |   bottom: "conv4_12_1x1_reduce"
4014 |   top: "conv4_12_3x3"
4015 |   convolution_param {
4016 |     num_output: 256
4017 |     bias_term: false
4018 |     pad: 1
4019 |     kernel_size: 3
4020 |     stride: 1
4021 |   }
4022 | }
4023 | layer {
4024 |   name: "conv4_12_3x3/bn"
4025 |   type: "BatchNorm"
4026 |   bottom: "conv4_12_3x3"
4027 |   top: "conv4_12_3x3"
4028 |   batch_norm_param {
4029 |     use_global_stats: true
4030 |   }
4031 | }
4032 | layer {
4033 |   name: "conv4_12_3x3/scale"
4034 |   type: "Scale"
4035 |   bottom: "conv4_12_3x3"
4036 |   top: "conv4_12_3x3"
4037 |   scale_param {
4038 |     bias_term: true
4039 |   }
4040 | }
4041 | layer {
4042 |   name: "conv4_12_3x3/relu"
4043 |   type: "ReLU"
4044 |   bottom: "conv4_12_3x3"
4045 |   top: "conv4_12_3x3"
4046 | }
4047 | layer {
4048 |   name: "conv4_12_1x1_increase"
4049 |   type: "Convolution"
4050 |   bottom: "conv4_12_3x3"
4051 |   top: "conv4_12_1x1_increase"
4052 |   convolution_param {
4053 |     num_output: 1024
4054 |     bias_term: false
4055 |     pad: 0
4056 |     kernel_size: 1
4057 |     stride: 1
4058 |   }
4059 | }
4060 | layer {
4061 |   name: "conv4_12_1x1_increase/bn"
4062 |   type: "BatchNorm"
4063 |   bottom: "conv4_12_1x1_increase"
4064 |   top: "conv4_12_1x1_increase"
4065 |   batch_norm_param {
4066 |     use_global_stats: true
4067 |   }
4068 | }
4069 | layer {
4070 |   name: "conv4_12_1x1_increase/scale"
4071 |   type: "Scale"
4072 |   bottom: "conv4_12_1x1_increase"
4073 |   top: "conv4_12_1x1_increase"
4074 |   scale_param {
4075 |     bias_term: true
4076 |   }
4077 | }
4078 | layer {
4079 |   name: "conv4_12"
4080 |   type: "Eltwise"
4081 |   bottom: "conv4_11"
4082 |   bottom: "conv4_12_1x1_increase"
4083 |   top: "conv4_12"
4084 |   eltwise_param {
4085 |     operation: SUM
4086 |   }
4087 | }
4088 | layer {
4089 |   name: "conv4_12/relu"
4090 |   type: "ReLU"
4091 |   bottom: "conv4_12"
4092 |   top: "conv4_12"
4093 | }
4094 | layer {
4095 |   name: "conv4_13_1x1_reduce"
4096 |   type: "Convolution"
4097 |   bottom: "conv4_12"
4098 |   top: "conv4_13_1x1_reduce"
4099 |   convolution_param {
4100 |     num_output: 256
4101 |     bias_term: false
4102 |     pad: 0
4103 |     kernel_size: 1
4104 |     stride: 1
4105 |   }
4106 | }
4107 | layer {
4108 |   name: "conv4_13_1x1_reduce/bn"
4109 |   type: "BatchNorm"
4110 |   bottom: "conv4_13_1x1_reduce"
4111 |   top: "conv4_13_1x1_reduce"
4112 |   batch_norm_param {
4113 |     use_global_stats: true
4114 |   }
4115 | }
4116 | layer {
4117 |   name: "conv4_13_1x1_reduce/scale"
4118 |   type: "Scale"
4119 |   bottom: "conv4_13_1x1_reduce"
4120 |   top: "conv4_13_1x1_reduce"
4121 |   scale_param {
4122 |     bias_term: true
4123 |   }
4124 | }
4125 | layer {
4126 |   name: "conv4_13_1x1_reduce/relu"
4127 |   type: "ReLU"
4128 |   bottom: "conv4_13_1x1_reduce"
4129 |   top: "conv4_13_1x1_reduce"
4130 | }
4131 | layer {
4132 |   name: "conv4_13_3x3"
4133 |   type: "Convolution"
4134 |   bottom: "conv4_13_1x1_reduce"
4135 |   top: "conv4_13_3x3"
4136 |   convolution_param {
4137 |     num_output: 256
4138 |     bias_term: false
4139 |     pad: 1
4140 |     kernel_size: 3
4141 |     stride: 1
4142 |   }
4143 | }
4144 | layer {
4145 |   name: "conv4_13_3x3/bn"
4146 |   type: "BatchNorm"
4147 |   bottom: "conv4_13_3x3"
4148 |   top: "conv4_13_3x3"
4149 |   batch_norm_param {
4150 |     use_global_stats: true
4151 |   }
4152 | }
4153 | layer {
4154 |   name: "conv4_13_3x3/scale"
4155 |   type: "Scale"
4156 |   bottom: "conv4_13_3x3"
4157 |   top: "conv4_13_3x3"
4158 |   scale_param {
4159 |     bias_term: true
4160 |   }
4161 | }
4162 | layer {
4163 |   name: "conv4_13_3x3/relu"
4164 |   type: "ReLU"
4165 |   bottom: "conv4_13_3x3"
4166 |   top: "conv4_13_3x3"
4167 | }
4168 | layer {
4169 |   name: "conv4_13_1x1_increase"
4170 |   type: "Convolution"
4171 |   bottom: "conv4_13_3x3"
4172 |   top: "conv4_13_1x1_increase"
4173 |   convolution_param {
4174 |     num_output: 1024
4175 |     bias_term: false
4176 |     pad: 0
4177 |     kernel_size: 1
4178 |     stride: 1
4179 |   }
4180 | }
4181 | layer {
4182 |   name: "conv4_13_1x1_increase/bn"
4183 |   type: "BatchNorm"
4184 |   bottom: "conv4_13_1x1_increase"
4185 |   top: "conv4_13_1x1_increase"
4186 |   batch_norm_param {
4187 |     use_global_stats: true
4188 |   }
4189 | }
4190 | layer {
4191 |   name: "conv4_13_1x1_increase/scale"
4192 |   type: "Scale"
4193 |   bottom: "conv4_13_1x1_increase"
4194 |   top: "conv4_13_1x1_increase"
4195 |   scale_param {
4196 |     bias_term: true
4197 |   }
4198 | }
4199 | layer {
4200 |   name: "conv4_13"
4201 |   type: "Eltwise"
4202 |   bottom: "conv4_12"
4203 |   bottom: "conv4_13_1x1_increase"
4204 |   top: "conv4_13"
4205 |   eltwise_param {
4206 |     operation: SUM
4207 |   }
4208 | }
4209 | layer {
4210 |   name: "conv4_13/relu"
4211 |   type: "ReLU"
4212 |   bottom: "conv4_13"
4213 |   top: "conv4_13"
4214 | }
4215 | layer {
4216 |   name: "conv4_14_1x1_reduce"
4217 |   type: "Convolution"
4218 |   bottom: "conv4_13"
4219 |   top: "conv4_14_1x1_reduce"
4220 |   convolution_param {
4221 |     num_output: 256
4222 |     bias_term: false
4223 |     pad: 0
4224 |     kernel_size: 1
4225 |     stride: 1
4226 |   }
4227 | }
4228 | layer {
4229 |   name: "conv4_14_1x1_reduce/bn"
4230 |   type: "BatchNorm"
4231 |   bottom: "conv4_14_1x1_reduce"
4232 |   top: "conv4_14_1x1_reduce"
4233 |   batch_norm_param {
4234 |     use_global_stats: true
4235 |   }
4236 | }
4237 | layer {
4238 |   name: "conv4_14_1x1_reduce/scale"
4239 |   type: "Scale"
4240 |   bottom: "conv4_14_1x1_reduce"
4241 |   top: "conv4_14_1x1_reduce"
4242 |   scale_param {
4243 |     bias_term: true
4244 |   }
4245 | }
4246 | layer {
4247 |   name: "conv4_14_1x1_reduce/relu"
4248 |   type: "ReLU"
4249 |   bottom: "conv4_14_1x1_reduce"
4250 |   top: "conv4_14_1x1_reduce"
4251 | }
4252 | layer {
4253 |   name: "conv4_14_3x3"
4254 |   type: "Convolution"
4255 |   bottom: "conv4_14_1x1_reduce"
4256 |   top: "conv4_14_3x3"
4257 |   convolution_param {
4258 |     num_output: 256
4259 |     bias_term: false
4260 |     pad: 1
4261 |     kernel_size: 3
4262 |     stride: 1
4263 |   }
4264 | }
4265 | layer {
4266 |   name: "conv4_14_3x3/bn"
4267 |   type: "BatchNorm"
4268 |   bottom: "conv4_14_3x3"
4269 |   top: "conv4_14_3x3"
4270 |   batch_norm_param {
4271 |     use_global_stats: true
4272 |   }
4273 | }
4274 | layer {
4275 |   name: "conv4_14_3x3/scale"
4276 |   type: "Scale"
4277 |   bottom: "conv4_14_3x3"
4278 |   top: "conv4_14_3x3"
4279 |   scale_param {
4280 |     bias_term: true
4281 |   }
4282 | }
4283 | layer {
4284 |   name: "conv4_14_3x3/relu"
4285 |   type: "ReLU"
4286 |   bottom: "conv4_14_3x3"
4287 |   top: "conv4_14_3x3"
4288 | }
4289 | layer {
4290 |   name: "conv4_14_1x1_increase"
4291 |   type: "Convolution"
4292 |   bottom: "conv4_14_3x3"
4293 |   top: "conv4_14_1x1_increase"
4294 |   convolution_param {
4295 |     num_output: 1024
4296 |     bias_term: false
4297 |     pad: 0
4298 |     kernel_size: 1
4299 |     stride: 1
4300 |   }
4301 | }
4302 | layer {
4303 |   name: "conv4_14_1x1_increase/bn"
4304 |   type: "BatchNorm"
4305 |   bottom: "conv4_14_1x1_increase"
4306 |   top: "conv4_14_1x1_increase"
4307 |   batch_norm_param {
4308 |     use_global_stats: true
4309 |   }
4310 | }
4311 | layer {
4312 |   name: "conv4_14_1x1_increase/scale"
4313 |   type: "Scale"
4314 |   bottom: "conv4_14_1x1_increase"
4315 |   top: "conv4_14_1x1_increase"
4316 |   scale_param {
4317 |     bias_term: true
4318 |   }
4319 | }
4320 | layer {
4321 |   name: "conv4_14"
4322 |   type: "Eltwise"
4323 |   bottom: "conv4_13"
4324 |   bottom: "conv4_14_1x1_increase"
4325 |   top: "conv4_14"
4326 |   eltwise_param {
4327 |     operation: SUM
4328 |   }
4329 | }
4330 | layer {
4331 |   name: "conv4_14/relu"
4332 |   type: "ReLU"
4333 |   bottom: "conv4_14"
4334 |   top: "conv4_14"
4335 | }
4336 | layer {
4337 |   name: "conv4_15_1x1_reduce"
4338 |   type: "Convolution"
4339 |   bottom: "conv4_14"
4340 |   top: "conv4_15_1x1_reduce"
4341 |   convolution_param {
4342 |     num_output: 256
4343 |     bias_term: false
4344 |     pad: 0
4345 |     kernel_size: 1
4346 |     stride: 1
4347 |   }
4348 | }
4349 | layer {
4350 |   name: "conv4_15_1x1_reduce/bn"
4351 |   type: "BatchNorm"
4352 |   bottom: "conv4_15_1x1_reduce"
4353 |   top: "conv4_15_1x1_reduce"
4354 |   batch_norm_param {
4355 |     use_global_stats: true
4356 |   }
4357 | }
4358 | layer {
4359 |   name: "conv4_15_1x1_reduce/scale"
4360 |   type: "Scale"
4361 |   bottom: "conv4_15_1x1_reduce"
4362 |   top: "conv4_15_1x1_reduce"
4363 |   scale_param {
4364 |     bias_term: true
4365 |   }
4366 | }
4367 | layer {
4368 |   name: "conv4_15_1x1_reduce/relu"
4369 |   type: "ReLU"
4370 |   bottom: "conv4_15_1x1_reduce"
4371 |   top: "conv4_15_1x1_reduce"
4372 | }
4373 | layer {
4374 |   name: "conv4_15_3x3"
4375 |   type: "Convolution"
4376 |   bottom: "conv4_15_1x1_reduce"
4377 |   top: "conv4_15_3x3"
4378 |   convolution_param {
4379 |     num_output: 256
4380 |     bias_term: false
4381 |     pad: 1
4382 |     kernel_size: 3
4383 |     stride: 1
4384 |   }
4385 | }
4386 | layer {
4387 |   name: "conv4_15_3x3/bn"
4388 |   type: "BatchNorm"
4389 |   bottom: "conv4_15_3x3"
4390 |   top: "conv4_15_3x3"
4391 |   batch_norm_param {
4392 |     use_global_stats: true
4393 |   }
4394 | }
4395 | layer {
4396 |   name: "conv4_15_3x3/scale"
4397 |   type: "Scale"
4398 |   bottom: "conv4_15_3x3"
4399 |   top: "conv4_15_3x3"
4400 |   scale_param {
4401 |     bias_term: true
4402 |   }
4403 | }
4404 | layer {
4405 |   name: "conv4_15_3x3/relu"
4406 |   type: "ReLU"
4407 |   bottom: "conv4_15_3x3"
4408 |   top: "conv4_15_3x3"
4409 | }
4410 | layer {
4411 |   name: "conv4_15_1x1_increase"
4412 |   type: "Convolution"
4413 |   bottom: "conv4_15_3x3"
4414 |   top: "conv4_15_1x1_increase"
4415 |   convolution_param {
4416 |     num_output: 1024
4417 |     bias_term: false
4418 |     pad: 0
4419 |     kernel_size: 1
4420 |     stride: 1
4421 |   }
4422 | }
4423 | layer {
4424 |   name: "conv4_15_1x1_increase/bn"
4425 |   type: "BatchNorm"
4426 |   bottom: "conv4_15_1x1_increase"
4427 |   top: "conv4_15_1x1_increase"
4428 |   batch_norm_param {
4429 |     use_global_stats: true
4430 |   }
4431 | }
4432 | layer {
4433 |   name: "conv4_15_1x1_increase/scale"
4434 |   type: "Scale"
4435 |   bottom: "conv4_15_1x1_increase"
4436 |   top: "conv4_15_1x1_increase"
4437 |   scale_param {
4438 |     bias_term: true
4439 |   }
4440 | }
4441 | layer {
4442 |   name: "conv4_15"
4443 |   type: "Eltwise"
4444 |   bottom: "conv4_14"
4445 |   bottom: "conv4_15_1x1_increase"
4446 |   top: "conv4_15"
4447 |   eltwise_param {
4448 |     operation: SUM
4449 |   }
4450 | }
4451 | layer {
4452 |   name: "conv4_15/relu"
4453 |   type: "ReLU"
4454 |   bottom: "conv4_15"
4455 |   top: "conv4_15"
4456 | }
4457 | layer {
4458 |   name: "conv4_16_1x1_reduce"
4459 |   type: "Convolution"
4460 |   bottom: "conv4_15"
4461 |   top: "conv4_16_1x1_reduce"
4462 |   convolution_param {
4463 |     num_output: 256
4464 |     bias_term: false
4465 |     pad: 0
4466 |     kernel_size: 1
4467 |     stride: 1
4468 |   }
4469 | }
4470 | layer {
4471 |   name: "conv4_16_1x1_reduce/bn"
4472 |   type: "BatchNorm"
4473 |   bottom: "conv4_16_1x1_reduce"
4474 |   top: "conv4_16_1x1_reduce"
4475 |   batch_norm_param {
4476 |     use_global_stats: true
4477 |   }
4478 | }
4479 | layer {
4480 |   name: "conv4_16_1x1_reduce/scale"
4481 |   type: "Scale"
4482 |   bottom: "conv4_16_1x1_reduce"
4483 |   top: "conv4_16_1x1_reduce"
4484 |   scale_param {
4485 |     bias_term: true
4486 |   }
4487 | }
4488 | layer {
4489 |   name: "conv4_16_1x1_reduce/relu"
4490 |   type: "ReLU"
4491 |   bottom: "conv4_16_1x1_reduce"
4492 |   top: "conv4_16_1x1_reduce"
4493 | }
4494 | layer {
4495 |   name: "conv4_16_3x3"
4496 |   type: "Convolution"
4497 |   bottom: "conv4_16_1x1_reduce"
4498 |   top: "conv4_16_3x3"
4499 |   convolution_param {
4500 |     num_output: 256
4501 |     bias_term: false
4502 |     pad: 1
4503 |     kernel_size: 3
4504 |     stride: 1
4505 |   }
4506 | }
4507 | layer {
4508 |   name: "conv4_16_3x3/bn"
4509 |   type: "BatchNorm"
4510 |   bottom: "conv4_16_3x3"
4511 |   top: "conv4_16_3x3"
4512 |   batch_norm_param {
4513 |     use_global_stats: true
4514 |   }
4515 | }
4516 | layer {
4517 |   name: "conv4_16_3x3/scale"
4518 |   type: "Scale"
4519 |   bottom: "conv4_16_3x3"
4520 |   top: "conv4_16_3x3"
4521 |   scale_param {
4522 |     bias_term: true
4523 |   }
4524 | }
4525 | layer {
4526 |   name: "conv4_16_3x3/relu"
4527 |   type: "ReLU"
4528 |   bottom: "conv4_16_3x3"
4529 |   top: "conv4_16_3x3"
4530 | }
4531 | layer {
4532 |   name: "conv4_16_1x1_increase"
4533 |   type: "Convolution"
4534 |   bottom: "conv4_16_3x3"
4535 |   top: "conv4_16_1x1_increase"
4536 |   convolution_param {
4537 |     num_output: 1024
4538 |     bias_term: false
4539 |     pad: 0
4540 |     kernel_size: 1
4541 |     stride: 1
4542 |   }
4543 | }
4544 | layer {
4545 |   name: "conv4_16_1x1_increase/bn"
4546 |   type: "BatchNorm"
4547 |   bottom: "conv4_16_1x1_increase"
4548 |   top: "conv4_16_1x1_increase"
4549 |   batch_norm_param {
4550 |     use_global_stats: true
4551 |   }
4552 | }
4553 | layer {
4554 |   name: "conv4_16_1x1_increase/scale"
4555 |   type: "Scale"
4556 |   bottom: "conv4_16_1x1_increase"
4557 |   top: "conv4_16_1x1_increase"
4558 |   scale_param {
4559 |     bias_term: true
4560 |   }
4561 | }
4562 | layer {
4563 |   name: "conv4_16"
4564 |   type: "Eltwise"
4565 |   bottom: "conv4_15"
4566 |   bottom: "conv4_16_1x1_increase"
4567 |   top: "conv4_16"
4568 |   eltwise_param {
4569 |     operation: SUM
4570 |   }
4571 | }
4572 | layer {
4573 |   name: "conv4_16/relu"
4574 |   type: "ReLU"
4575 |   bottom: "conv4_16"
4576 |   top: "conv4_16"
4577 | }
4578 | layer {
4579 |   name: "conv4_17_1x1_reduce"
4580 |   type: "Convolution"
4581 |   bottom: "conv4_16"
4582 |   top: "conv4_17_1x1_reduce"
4583 |   convolution_param {
4584 |     num_output: 256
4585 |     bias_term: false
4586 |     pad: 0
4587 |     kernel_size: 1
4588 |     stride: 1
4589 |   }
4590 | }
4591 | layer {
4592 |   name: "conv4_17_1x1_reduce/bn"
4593 |   type: "BatchNorm"
4594 |   bottom: "conv4_17_1x1_reduce"
4595 |   top: "conv4_17_1x1_reduce"
4596 |   batch_norm_param {
4597 |     use_global_stats: true
4598 |   }
4599 | }
4600 | layer {
4601 |   name: "conv4_17_1x1_reduce/scale"
4602 |   type: "Scale"
4603 |   bottom: "conv4_17_1x1_reduce"
4604 |   top: "conv4_17_1x1_reduce"
4605 |   scale_param {
4606 |     bias_term: true
4607 |   }
4608 | }
4609 | layer {
4610 |   name: "conv4_17_1x1_reduce/relu"
4611 |   type: "ReLU"
4612 |   bottom: "conv4_17_1x1_reduce"
4613 |   top: "conv4_17_1x1_reduce"
4614 | }
4615 | layer {
4616 |   name: "conv4_17_3x3"
4617 |   type: "Convolution"
4618 |   bottom: "conv4_17_1x1_reduce"
4619 |   top: "conv4_17_3x3"
4620 |   convolution_param {
4621 |     num_output: 256
4622 |     bias_term: false
4623 |     pad: 1
4624 |     kernel_size: 3
4625 |     stride: 1
4626 |   }
4627 | }
4628 | layer {
4629 |   name: "conv4_17_3x3/bn"
4630 |   type: "BatchNorm"
4631 |   bottom: "conv4_17_3x3"
4632 |   top: "conv4_17_3x3"
4633 |   batch_norm_param {
4634 |     use_global_stats: true
4635 |   }
4636 | }
4637 | layer {
4638 |   name: "conv4_17_3x3/scale"
4639 |   type: "Scale"
4640 |   bottom: "conv4_17_3x3"
4641 |   top: "conv4_17_3x3"
4642 |   scale_param {
4643 |     bias_term: true
4644 |   }
4645 | }
4646 | layer {
4647 |   name: "conv4_17_3x3/relu"
4648 |   type: "ReLU"
4649 |   bottom: "conv4_17_3x3"
4650 |   top: "conv4_17_3x3"
4651 | }
4652 | layer {
4653 |   name: "conv4_17_1x1_increase"
4654 |   type: "Convolution"
4655 |   bottom: "conv4_17_3x3"
4656 |   top: "conv4_17_1x1_increase"
4657 |   convolution_param {
4658 |     num_output: 1024
4659 |     bias_term: false
4660 |     pad: 0
4661 |     kernel_size: 1
4662 |     stride: 1
4663 |   }
4664 | }
4665 | layer {
4666 |   name: "conv4_17_1x1_increase/bn"
4667 |   type: "BatchNorm"
4668 |   bottom: "conv4_17_1x1_increase"
4669 |   top: "conv4_17_1x1_increase"
4670 |   batch_norm_param {
4671 |     use_global_stats: true
4672 |   }
4673 | }
4674 | layer {
4675 |   name: "conv4_17_1x1_increase/scale"
4676 |   type: "Scale"
4677 |   bottom: "conv4_17_1x1_increase"
4678 |   top: "conv4_17_1x1_increase"
4679 |   scale_param {
4680 |     bias_term: true
4681 |   }
4682 | }
4683 | layer {
4684 |   name: "conv4_17"
4685 |   type: "Eltwise"
4686 |   bottom: "conv4_16"
4687 |   bottom: "conv4_17_1x1_increase"
4688 |   top: "conv4_17"
4689 |   eltwise_param {
4690 |     operation: SUM
4691 |   }
4692 | }
4693 | layer {
4694 |   name: "conv4_17/relu"
4695 |   type: "ReLU"
4696 |   bottom: "conv4_17"
4697 |   top: "conv4_17"
4698 | }
4699 | layer {
4700 |   name: "conv4_18_1x1_reduce"
4701 |   type: "Convolution"
4702 |   bottom: "conv4_17"
4703 |   top: "conv4_18_1x1_reduce"
4704 |   convolution_param {
4705 |     num_output: 256
4706 |     bias_term: false
4707 |     pad: 0
4708 |     kernel_size: 1
4709 |     stride: 1
4710 |   }
4711 | }
4712 | layer {
4713 |   name: "conv4_18_1x1_reduce/bn"
4714 |   type: "BatchNorm"
4715 |   bottom: "conv4_18_1x1_reduce"
4716 |   top: "conv4_18_1x1_reduce"
4717 |   batch_norm_param {
4718 |     use_global_stats: true
4719 |   }
4720 | }
4721 | layer {
4722 |   name: "conv4_18_1x1_reduce/scale"
4723 |   type: "Scale"
4724 |   bottom: "conv4_18_1x1_reduce"
4725 |   top: "conv4_18_1x1_reduce"
4726 |   scale_param {
4727 |     bias_term: true
4728 |   }
4729 | }
4730 | layer {
4731 |   name: "conv4_18_1x1_reduce/relu"
4732 |   type: "ReLU"
4733 |   bottom: "conv4_18_1x1_reduce"
4734 |   top: "conv4_18_1x1_reduce"
4735 | }
4736 | layer {
4737 |   name: "conv4_18_3x3"
4738 |   type: "Convolution"
4739 |   bottom: "conv4_18_1x1_reduce"
4740 |   top: "conv4_18_3x3"
4741 |   convolution_param {
4742 |     num_output: 256
4743 |     bias_term: false
4744 |     pad: 1
4745 |     kernel_size: 3
4746 |     stride: 1
4747 |   }
4748 | }
4749 | layer {
4750 |   name: "conv4_18_3x3/bn"
4751 |   type: "BatchNorm"
4752 |   bottom: "conv4_18_3x3"
4753 |   top: "conv4_18_3x3"
4754 |   batch_norm_param {
4755 |     use_global_stats: true
4756 |   }
4757 | }
4758 | layer {
4759 |   name: "conv4_18_3x3/scale"
4760 |   type: "Scale"
4761 |   bottom: "conv4_18_3x3"
4762 |   top: "conv4_18_3x3"
4763 |   scale_param {
4764 |     bias_term: true
4765 |   }
4766 | }
4767 | layer {
4768 |   name: "conv4_18_3x3/relu"
4769 |   type: "ReLU"
4770 |   bottom: "conv4_18_3x3"
4771 |   top: "conv4_18_3x3"
4772 | }
4773 | layer {
4774 |   name: "conv4_18_1x1_increase"
4775 |   type: "Convolution"
4776 |   bottom: "conv4_18_3x3"
4777 |   top: "conv4_18_1x1_increase"
4778 |   convolution_param {
4779 |     num_output: 1024
4780 |     bias_term: false
4781 |     pad: 0
4782 |     kernel_size: 1
4783 |     stride: 1
4784 |   }
4785 | }
4786 | layer {
4787 |   name: "conv4_18_1x1_increase/bn"
4788 |   type: "BatchNorm"
4789 |   bottom: "conv4_18_1x1_increase"
4790 |   top: "conv4_18_1x1_increase"
4791 |   batch_norm_param {
4792 |     use_global_stats: true
4793 |   }
4794 | }
4795 | layer {
4796 |   name: "conv4_18_1x1_increase/scale"
4797 |   type: "Scale"
4798 |   bottom: "conv4_18_1x1_increase"
4799 |   top: "conv4_18_1x1_increase"
4800 |   scale_param {
4801 |     bias_term: true
4802 |   }
4803 | }
4804 | layer {
4805 |   name: "conv4_18"
4806 |   type: "Eltwise"
4807 |   bottom: "conv4_17"
4808 |   bottom: "conv4_18_1x1_increase"
4809 |   top: "conv4_18"
4810 |   eltwise_param {
4811 |     operation: SUM
4812 |   }
4813 | }
4814 | layer {
4815 |   name: "conv4_18/relu"
4816 |   type: "ReLU"
4817 |   bottom: "conv4_18"
4818 |   top: "conv4_18"
4819 | }
4820 | layer {
4821 |   name: "conv4_19_1x1_reduce"
4822 |   type: "Convolution"
4823 |   bottom: "conv4_18"
4824 |   top: "conv4_19_1x1_reduce"
4825 |   convolution_param {
4826 |     num_output: 256
4827 |     bias_term: false
4828 |     pad: 0
4829 |     kernel_size: 1
4830 |     stride: 1
4831 |   }
4832 | }
4833 | layer {
4834 |   name: "conv4_19_1x1_reduce/bn"
4835 |   type: "BatchNorm"
4836 |   bottom: "conv4_19_1x1_reduce"
4837 |   top: "conv4_19_1x1_reduce"
4838 |   batch_norm_param {
4839 |     use_global_stats: true
4840 |   }
4841 | }
4842 | layer {
4843 |   name: "conv4_19_1x1_reduce/scale"
4844 |   type: "Scale"
4845 |   bottom: "conv4_19_1x1_reduce"
4846 |   top: "conv4_19_1x1_reduce"
4847 |   scale_param {
4848 |     bias_term: true
4849 |   }
4850 | }
4851 | layer {
4852 |   name: "conv4_19_1x1_reduce/relu"
4853 |   type: "ReLU"
4854 |   bottom: "conv4_19_1x1_reduce"
4855 |   top: "conv4_19_1x1_reduce"
4856 | }
4857 | layer {
4858 |   name: "conv4_19_3x3"
4859 |   type: "Convolution"
4860 |   bottom: "conv4_19_1x1_reduce"
4861 |   top: "conv4_19_3x3"
4862 |   convolution_param {
4863 |     num_output: 256
4864 |     bias_term: false
4865 |     pad: 1
4866 |     kernel_size: 3
4867 |     stride: 1
4868 |   }
4869 | }
4870 | layer {
4871 |   name: "conv4_19_3x3/bn"
4872 |   type: "BatchNorm"
4873 |   bottom: "conv4_19_3x3"
4874 |   top: "conv4_19_3x3"
4875 |   batch_norm_param {
4876 |     use_global_stats: true
4877 |   }
4878 | }
4879 | layer {
4880 |   name: "conv4_19_3x3/scale"
4881 |   type: "Scale"
4882 |   bottom: "conv4_19_3x3"
4883 |   top: "conv4_19_3x3"
4884 |   scale_param {
4885 |     bias_term: true
4886 |   }
4887 | }
4888 | layer {
4889 |   name: "conv4_19_3x3/relu"
4890 |   type: "ReLU"
4891 |   bottom: "conv4_19_3x3"
4892 |   top: "conv4_19_3x3"
4893 | }
4894 | layer {
4895 |   name: "conv4_19_1x1_increase"
4896 |   type: "Convolution"
4897 |   bottom: "conv4_19_3x3"
4898 |   top: "conv4_19_1x1_increase"
4899 |   convolution_param {
4900 |     num_output: 1024
4901 |     bias_term: false
4902 |     pad: 0
4903 |     kernel_size: 1
4904 |     stride: 1
4905 |   }
4906 | }
4907 | layer {
4908 |   name: "conv4_19_1x1_increase/bn"
4909 |   type: "BatchNorm"
4910 |   bottom: "conv4_19_1x1_increase"
4911 |   top: "conv4_19_1x1_increase"
4912 |   batch_norm_param {
4913 |     use_global_stats: true
4914 |   }
4915 | }
4916 | layer {
4917 |   name: "conv4_19_1x1_increase/scale"
4918 |   type: "Scale"
4919 |   bottom: "conv4_19_1x1_increase"
4920 |   top: "conv4_19_1x1_increase"
4921 |   scale_param {
4922 |     bias_term: true
4923 |   }
4924 | }
4925 | layer {
4926 |   name: "conv4_19"
4927 |   type: "Eltwise"
4928 |   bottom: "conv4_18"
4929 |   bottom: "conv4_19_1x1_increase"
4930 |   top: "conv4_19"
4931 |   eltwise_param {
4932 |     operation: SUM
4933 |   }
4934 | }
4935 | layer {
4936 |   name: "conv4_19/relu"
4937 |   type: "ReLU"
4938 |   bottom: "conv4_19"
4939 |   top: "conv4_19"
4940 | }
4941 | layer {
4942 |   name: "conv4_20_1x1_reduce"
4943 |   type: "Convolution"
4944 |   bottom: "conv4_19"
4945 |   top: "conv4_20_1x1_reduce"
4946 |   convolution_param {
4947 |     num_output: 256
4948 |     bias_term: false
4949 |     pad: 0
4950 |     kernel_size: 1
4951 |     stride: 1
4952 |   }
4953 | }
4954 | layer {
4955 |   name: "conv4_20_1x1_reduce/bn"
4956 |   type: "BatchNorm"
4957 |   bottom: "conv4_20_1x1_reduce"
4958 |   top: "conv4_20_1x1_reduce"
4959 |   batch_norm_param {
4960 |     use_global_stats: true
4961 |   }
4962 | }
4963 | layer {
4964 |   name: "conv4_20_1x1_reduce/scale"
4965 |   type: "Scale"
4966 |   bottom: "conv4_20_1x1_reduce"
4967 |   top: "conv4_20_1x1_reduce"
4968 |   scale_param {
4969 |     bias_term: true
4970 |   }
4971 | }
4972 | layer {
4973 |   name: "conv4_20_1x1_reduce/relu"
4974 |   type: "ReLU"
4975 |   bottom: "conv4_20_1x1_reduce"
4976 |   top: "conv4_20_1x1_reduce"
4977 | }
4978 | layer {
4979 |   name: "conv4_20_3x3"
4980 |   type: "Convolution"
4981 |   bottom: "conv4_20_1x1_reduce"
4982 |   top: "conv4_20_3x3"
4983 |   convolution_param {
4984 |     num_output: 256
4985 |     bias_term: false
4986 |     pad: 1
4987 |     kernel_size: 3
4988 |     stride: 1
4989 |   }
4990 | }
4991 | layer {
4992 |   name: "conv4_20_3x3/bn"
4993 |   type: "BatchNorm"
4994 |   bottom: "conv4_20_3x3"
4995 |   top: "conv4_20_3x3"
4996 |   batch_norm_param {
4997 |     use_global_stats: true
4998 |   }
4999 | }
5000 | layer {
5001 |   name: "conv4_20_3x3/scale"
5002 |   type: "Scale"
5003 |   bottom: "conv4_20_3x3"
5004 |   top: "conv4_20_3x3"
5005 |   scale_param {
5006 |     bias_term: true
5007 |   }
5008 | }
5009 | layer {
5010 |   name: "conv4_20_3x3/relu"
5011 |   type: "ReLU"
5012 |   bottom: "conv4_20_3x3"
5013 |   top: "conv4_20_3x3"
5014 | }
5015 | layer {
5016 |   name: "conv4_20_1x1_increase"
5017 |   type: "Convolution"
5018 |   bottom: "conv4_20_3x3"
5019 |   top: "conv4_20_1x1_increase"
5020 |   convolution_param {
5021 |     num_output: 1024
5022 |     bias_term: false
5023 |     pad: 0
5024 |     kernel_size: 1
5025 |     stride: 1
5026 |   }
5027 | }
5028 | layer {
5029 |   name: "conv4_20_1x1_increase/bn"
5030 |   type: "BatchNorm"
5031 |   bottom: "conv4_20_1x1_increase"
5032 |   top: "conv4_20_1x1_increase"
5033 |   batch_norm_param {
5034 |     use_global_stats: true
5035 |   }
5036 | }
5037 | layer {
5038 |   name: "conv4_20_1x1_increase/scale"
5039 |   type: "Scale"
5040 |   bottom: "conv4_20_1x1_increase"
5041 |   top: "conv4_20_1x1_increase"
5042 |   scale_param {
5043 |     bias_term: true
5044 |   }
5045 | }
5046 | layer {
5047 |   name: "conv4_20"
5048 |   type: "Eltwise"
5049 |   bottom: "conv4_19"
5050 |   bottom: "conv4_20_1x1_increase"
5051 |   top: "conv4_20"
5052 |   eltwise_param {
5053 |     operation: SUM
5054 |   }
5055 | }
5056 | layer {
5057 |   name: "conv4_20/relu"
5058 |   type: "ReLU"
5059 |   bottom: "conv4_20"
5060 |   top: "conv4_20"
5061 | }
5062 | layer {
5063 |   name: "conv4_21_1x1_reduce"
5064 |   type: "Convolution"
5065 |   bottom: "conv4_20"
5066 |   top: "conv4_21_1x1_reduce"
5067 |   convolution_param {
5068 |     num_output: 256
5069 |     bias_term: false
5070 |     pad: 0
5071 |     kernel_size: 1
5072 |     stride: 1
5073 |   }
5074 | }
5075 | layer {
5076 |   name: "conv4_21_1x1_reduce/bn"
5077 |   type: "BatchNorm"
5078 |   bottom: "conv4_21_1x1_reduce"
5079 |   top: "conv4_21_1x1_reduce"
5080 |   batch_norm_param {
5081 |     use_global_stats: true
5082 |   }
5083 | }
5084 | layer {
5085 |   name: "conv4_21_1x1_reduce/scale"
5086 |   type: "Scale"
5087 |   bottom: "conv4_21_1x1_reduce"
5088 |   top: "conv4_21_1x1_reduce"
5089 |   scale_param {
5090 |     bias_term: true
5091 |   }
5092 | }
5093 | layer {
5094 |   name: "conv4_21_1x1_reduce/relu"
5095 |   type: "ReLU"
5096 |   bottom: "conv4_21_1x1_reduce"
5097 |   top: "conv4_21_1x1_reduce"
5098 | }
5099 | layer {
5100 |   name: "conv4_21_3x3"
5101 |   type: "Convolution"
5102 |   bottom: "conv4_21_1x1_reduce"
5103 |   top: "conv4_21_3x3"
5104 |   convolution_param {
5105 |     num_output: 256
5106 |     bias_term: false
5107 |     pad: 1
5108 |     kernel_size: 3
5109 |     stride: 1
5110 |   }
5111 | }
5112 | layer {
5113 |   name: "conv4_21_3x3/bn"
5114 |   type: "BatchNorm"
5115 |   bottom: "conv4_21_3x3"
5116 |   top: "conv4_21_3x3"
5117 |   batch_norm_param {
5118 |     use_global_stats: true
5119 |   }
5120 | }
5121 | layer {
5122 |   name: "conv4_21_3x3/scale"
5123 |   type: "Scale"
5124 |   bottom: "conv4_21_3x3"
5125 |   top: "conv4_21_3x3"
5126 |   scale_param {
5127 |     bias_term: true
5128 |   }
5129 | }
5130 | layer {
5131 |   name: "conv4_21_3x3/relu"
5132 |   type: "ReLU"
5133 |   bottom: "conv4_21_3x3"
5134 |   top: "conv4_21_3x3"
5135 | }
5136 | layer {
5137 |   name: "conv4_21_1x1_increase"
5138 |   type: "Convolution"
5139 |   bottom: "conv4_21_3x3"
5140 |   top: "conv4_21_1x1_increase"
5141 |   convolution_param {
5142 |     num_output: 1024
5143 |     bias_term: false
5144 |     pad: 0
5145 |     kernel_size: 1
5146 |     stride: 1
5147 |   }
5148 | }
5149 | layer {
5150 |   name: "conv4_21_1x1_increase/bn"
5151 |   type: "BatchNorm"
5152 |   bottom: "conv4_21_1x1_increase"
5153 |   top: "conv4_21_1x1_increase"
5154 |   batch_norm_param {
5155 |     use_global_stats: true
5156 |   }
5157 | }
5158 | layer {
5159 |   name: "conv4_21_1x1_increase/scale"
5160 |   type: "Scale"
5161 |   bottom: "conv4_21_1x1_increase"
5162 |   top: "conv4_21_1x1_increase"
5163 |   scale_param {
5164 |     bias_term: true
5165 |   }
5166 | }
5167 | layer {
5168 |   name: "conv4_21"
5169 |   type: "Eltwise"
5170 |   bottom: "conv4_20"
5171 |   bottom: "conv4_21_1x1_increase"
5172 |   top: "conv4_21"
5173 |   eltwise_param {
5174 |     operation: SUM
5175 |   }
5176 | }
5177 | layer {
5178 |   name: "conv4_21/relu"
5179 |   type: "ReLU"
5180 |   bottom: "conv4_21"
5181 |   top: "conv4_21"
5182 | }
5183 | layer {
5184 |   name: "conv4_22_1x1_reduce"
5185 |   type: "Convolution"
5186 |   bottom: "conv4_21"
5187 |   top: "conv4_22_1x1_reduce"
5188 |   convolution_param {
5189 |     num_output: 256
5190 |     bias_term: false
5191 |     pad: 0
5192 |     kernel_size: 1
5193 |     stride: 1
5194 |   }
5195 | }
5196 | layer {
5197 |   name: "conv4_22_1x1_reduce/bn"
5198 |   type: "BatchNorm"
5199 |   bottom: "conv4_22_1x1_reduce"
5200 |   top: "conv4_22_1x1_reduce"
5201 |   batch_norm_param {
5202 |     use_global_stats: true
5203 |   }
5204 | }
5205 | layer {
5206 |   name: "conv4_22_1x1_reduce/scale"
5207 |   type: "Scale"
5208 |   bottom: "conv4_22_1x1_reduce"
5209 |   top: "conv4_22_1x1_reduce"
5210 |   scale_param {
5211 |     bias_term: true
5212 |   }
5213 | }
5214 | layer {
5215 |   name: "conv4_22_1x1_reduce/relu"
5216 |   type: "ReLU"
5217 |   bottom: "conv4_22_1x1_reduce"
5218 |   top: "conv4_22_1x1_reduce"
5219 | }
5220 | layer {
5221 |   name: "conv4_22_3x3"
5222 |   type: "Convolution"
5223 |   bottom: "conv4_22_1x1_reduce"
5224 |   top: "conv4_22_3x3"
5225 |   convolution_param {
5226 |     num_output: 256
5227 |     bias_term: false
5228 |     pad: 1
5229 |     kernel_size: 3
5230 |     stride: 1
5231 |   }
5232 | }
5233 | layer {
5234 |   name: "conv4_22_3x3/bn"
5235 |   type: "BatchNorm"
5236 |   bottom: "conv4_22_3x3"
5237 |   top: "conv4_22_3x3"
5238 |   batch_norm_param {
5239 |     use_global_stats: true
5240 |   }
5241 | }
5242 | layer {
5243 |   name: "conv4_22_3x3/scale"
5244 |   type: "Scale"
5245 |   bottom: "conv4_22_3x3"
5246 |   top: "conv4_22_3x3"
5247 |   scale_param {
5248 |     bias_term: true
5249 |   }
5250 | }
5251 | layer {
5252 |   name: "conv4_22_3x3/relu"
5253 |   type: "ReLU"
5254 |   bottom: "conv4_22_3x3"
5255 |   top: "conv4_22_3x3"
5256 | }
5257 | layer {
5258 |   name: "conv4_22_1x1_increase"
5259 |   type: "Convolution"
5260 |   bottom: "conv4_22_3x3"
5261 |   top: "conv4_22_1x1_increase"
5262 |   convolution_param {
5263 |     num_output: 1024
5264 |     bias_term: false
5265 |     pad: 0
5266 |     kernel_size: 1
5267 |     stride: 1
5268 |   }
5269 | }
5270 | layer {
5271 |   name: "conv4_22_1x1_increase/bn"
5272 |   type: "BatchNorm"
5273 |   bottom: "conv4_22_1x1_increase"
5274 |   top: "conv4_22_1x1_increase"
5275 |   batch_norm_param {
5276 |     use_global_stats: true
5277 |   }
5278 | }
5279 | layer {
5280 |   name: "conv4_22_1x1_increase/scale"
5281 |   type: "Scale"
5282 |   bottom: "conv4_22_1x1_increase"
5283 |   top: "conv4_22_1x1_increase"
5284 |   scale_param {
5285 |     bias_term: true
5286 |   }
5287 | }
5288 | layer {
5289 |   name: "conv4_22"
5290 |   type: "Eltwise"
5291 |   bottom: "conv4_21"
5292 |   bottom: "conv4_22_1x1_increase"
5293 |   top: "conv4_22"
5294 |   eltwise_param {
5295 |     operation: SUM
5296 |   }
5297 | }
5298 | layer {
5299 |   name: "conv4_22/relu"
5300 |   type: "ReLU"
5301 |   bottom: "conv4_22"
5302 |   top: "conv4_22"
5303 | }
5304 | layer {
5305 |   name: "conv4_23_1x1_reduce"
5306 |   type: "Convolution"
5307 |   bottom: "conv4_22"
5308 |   top: "conv4_23_1x1_reduce"
5309 |   convolution_param {
5310 |     num_output: 256
5311 |     bias_term: false
5312 |     pad: 0
5313 |     kernel_size: 1
5314 |     stride: 1
5315 |   }
5316 | }
5317 | layer {
5318 |   name: "conv4_23_1x1_reduce/bn"
5319 |   type: "BatchNorm"
5320 |   bottom: "conv4_23_1x1_reduce"
5321 |   top: "conv4_23_1x1_reduce"
5322 |   batch_norm_param {
5323 |     use_global_stats: true
5324 |   }
5325 | }
5326 | layer {
5327 |   name: "conv4_23_1x1_reduce/scale"
5328 |   type: "Scale"
5329 |   bottom: "conv4_23_1x1_reduce"
5330 |   top: "conv4_23_1x1_reduce"
5331 |   scale_param {
5332 |     bias_term: true
5333 |   }
5334 | }
5335 | layer {
5336 |   name: "conv4_23_1x1_reduce/relu"
5337 |   type: "ReLU"
5338 |   bottom: "conv4_23_1x1_reduce"
5339 |   top: "conv4_23_1x1_reduce"
5340 | }
5341 | layer {
5342 |   name: "conv4_23_3x3"
5343 |   type: "Convolution"
5344 |   bottom: "conv4_23_1x1_reduce"
5345 |   top: "conv4_23_3x3"
5346 |   convolution_param {
5347 |     num_output: 256
5348 |     bias_term: false
5349 |     pad: 1
5350 |     kernel_size: 3
5351 |     stride: 1
5352 |   }
5353 | }
5354 | layer {
5355 |   name: "conv4_23_3x3/bn"
5356 |   type: "BatchNorm"
5357 |   bottom: "conv4_23_3x3"
5358 |   top: "conv4_23_3x3"
5359 |   batch_norm_param {
5360 |     use_global_stats: true
5361 |   }
5362 | }
5363 | layer {
5364 |   name: "conv4_23_3x3/scale"
5365 |   type: "Scale"
5366 |   bottom: "conv4_23_3x3"
5367 |   top: "conv4_23_3x3"
5368 |   scale_param {
5369 |     bias_term: true
5370 |   }
5371 | }
5372 | layer {
5373 |   name: "conv4_23_3x3/relu"
5374 |   type: "ReLU"
5375 |   bottom: "conv4_23_3x3"
5376 |   top: "conv4_23_3x3"
5377 | }
5378 | layer {
5379 |   name: "conv4_23_1x1_increase"
5380 |   type: "Convolution"
5381 |   bottom: "conv4_23_3x3"
5382 |   top: "conv4_23_1x1_increase"
5383 |   convolution_param {
5384 |     num_output: 1024
5385 |     bias_term: false
5386 |     pad: 0
5387 |     kernel_size: 1
5388 |     stride: 1
5389 |   }
5390 | }
5391 | layer {
5392 |   name: "conv4_23_1x1_increase/bn"
5393 |   type: "BatchNorm"
5394 |   bottom: "conv4_23_1x1_increase"
5395 |   top: "conv4_23_1x1_increase"
5396 |   batch_norm_param {
5397 |     use_global_stats: true
5398 |   }
5399 | }
5400 | layer {
5401 |   name: "conv4_23_1x1_increase/scale"
5402 |   type: "Scale"
5403 |   bottom: "conv4_23_1x1_increase"
5404 |   top: "conv4_23_1x1_increase"
5405 |   scale_param {
5406 |     bias_term: true
5407 |   }
5408 | }
5409 | layer {
5410 |   name: "conv4_23"
5411 |   type: "Eltwise"
5412 |   bottom: "conv4_22"
5413 |   bottom: "conv4_23_1x1_increase"
5414 |   top: "conv4_23"
5415 |   eltwise_param {
5416 |     operation: SUM
5417 |   }
5418 | }
5419 | layer {
5420 |   name: "conv4_23/relu"
5421 |   type: "ReLU"
5422 |   bottom: "conv4_23"
5423 |   top: "conv4_23"
5424 | }
5425 | layer {
5426 |   name: "conv5_1_1x1_reduce"
5427 |   type: "Convolution"
5428 |   bottom: "conv4_23"
5429 |   top: "conv5_1_1x1_reduce"
5430 |   convolution_param {
5431 |     num_output: 512
5432 |     bias_term: false
5433 |     pad: 0
5434 |     kernel_size: 1
5435 |     stride: 1
5436 |   }
5437 | }
5438 | layer {
5439 |   name: "conv5_1_1x1_reduce/bn"
5440 |   type: "BatchNorm"
5441 |   bottom: "conv5_1_1x1_reduce"
5442 |   top: "conv5_1_1x1_reduce"
5443 |   batch_norm_param {
5444 |     use_global_stats: true
5445 |   }
5446 | }
5447 | layer {
5448 |   name: "conv5_1_1x1_reduce/scale"
5449 |   type: "Scale"
5450 |   bottom: "conv5_1_1x1_reduce"
5451 |   top: "conv5_1_1x1_reduce"
5452 |   scale_param {
5453 |     bias_term: true
5454 |   }
5455 | }
5456 | layer {
5457 |   name: "conv5_1_1x1_reduce/relu"
5458 |   type: "ReLU"
5459 |   bottom: "conv5_1_1x1_reduce"
5460 |   top: "conv5_1_1x1_reduce"
5461 | }
5462 | layer {
5463 |   name: "conv5_1_3x3"
5464 |   type: "Convolution"
5465 |   bottom: "conv5_1_1x1_reduce"
5466 |   top: "conv5_1_3x3"
5467 |   convolution_param {
5468 |     num_output: 512
5469 |     bias_term: false
5470 |     pad: 1
5471 |     kernel_size: 3
5472 |     stride: 2
5473 |   }
5474 | }
5475 | layer {
5476 |   name: "conv5_1_3x3/bn"
5477 |   type: "BatchNorm"
5478 |   bottom: "conv5_1_3x3"
5479 |   top: "conv5_1_3x3"
5480 |   batch_norm_param {
5481 |     use_global_stats: true
5482 |   }
5483 | }
5484 | layer {
5485 |   name: "conv5_1_3x3/scale"
5486 |   type: "Scale"
5487 |   bottom: "conv5_1_3x3"
5488 |   top: "conv5_1_3x3"
5489 |   scale_param {
5490 |     bias_term: true
5491 |   }
5492 | }
5493 | layer {
5494 |   name: "conv5_1_3x3/relu"
5495 |   type: "ReLU"
5496 |   bottom: "conv5_1_3x3"
5497 |   top: "conv5_1_3x3"
5498 | }
5499 | layer {
5500 |   name: "conv5_1_1x1_increase"
5501 |   type: "Convolution"
5502 |   bottom: "conv5_1_3x3"
5503 |   top: "conv5_1_1x1_increase"
5504 |   convolution_param {
5505 |     num_output: 2048
5506 |     bias_term: false
5507 |     pad: 0
5508 |     kernel_size: 1
5509 |     stride: 1
5510 |   }
5511 | }
5512 | layer {
5513 |   name: "conv5_1_1x1_increase/bn"
5514 |   type: "BatchNorm"
5515 |   bottom: "conv5_1_1x1_increase"
5516 |   top: "conv5_1_1x1_increase"
5517 |   batch_norm_param {
5518 |     use_global_stats: true
5519 |   }
5520 | }
5521 | layer {
5522 |   name: "conv5_1_1x1_increase/scale"
5523 |   type: "Scale"
5524 |   bottom: "conv5_1_1x1_increase"
5525 |   top: "conv5_1_1x1_increase"
5526 |   scale_param {
5527 |     bias_term: true
5528 |   }
5529 | }
5530 | layer {
5531 |   name: "conv5_1_1x1_proj"
5532 |   type: "Convolution"
5533 |   bottom: "conv4_23"
5534 |   top: "conv5_1_1x1_proj"
5535 |   convolution_param {
5536 |     num_output: 2048
5537 |     bias_term: false
5538 |     pad: 0
5539 |     kernel_size: 1
5540 |     stride: 2
5541 |   }
5542 | }
5543 | layer {
5544 |   name: "conv5_1_1x1_proj/bn"
5545 |   type: "BatchNorm"
5546 |   bottom: "conv5_1_1x1_proj"
5547 |   top: "conv5_1_1x1_proj"
5548 |   batch_norm_param {
5549 |     use_global_stats: true
5550 |   }
5551 | }
5552 | layer {
5553 |   name: "conv5_1_1x1_proj/scale"
5554 |   type: "Scale"
5555 |   bottom: "conv5_1_1x1_proj"
5556 |   top: "conv5_1_1x1_proj"
5557 |   scale_param {
5558 |     bias_term: true
5559 |   }
5560 | }
5561 | layer {
5562 |   name: "conv5_1"
5563 |   type: "Eltwise"
5564 |   bottom: "conv5_1_1x1_proj"
5565 |   bottom: "conv5_1_1x1_increase"
5566 |   top: "conv5_1"
5567 |   eltwise_param {
5568 |     operation: SUM
5569 |   }
5570 | }
5571 | layer {
5572 |   name: "conv5_1/relu"
5573 |   type: "ReLU"
5574 |   bottom: "conv5_1"
5575 |   top: "conv5_1"
5576 | }
5577 | layer {
5578 |   name: "conv5_2_1x1_reduce"
5579 |   type: "Convolution"
5580 |   bottom: "conv5_1"
5581 |   top: "conv5_2_1x1_reduce"
5582 |   convolution_param {
5583 |     num_output: 512
5584 |     bias_term: false
5585 |     pad: 0
5586 |     kernel_size: 1
5587 |     stride: 1
5588 |   }
5589 | }
5590 | layer {
5591 |   name: "conv5_2_1x1_reduce/bn"
5592 |   type: "BatchNorm"
5593 |   bottom: "conv5_2_1x1_reduce"
5594 |   top: "conv5_2_1x1_reduce"
5595 |   batch_norm_param {
5596 |     use_global_stats: true
5597 |   }
5598 | }
5599 | layer {
5600 |   name: "conv5_2_1x1_reduce/scale"
5601 |   type: "Scale"
5602 |   bottom: "conv5_2_1x1_reduce"
5603 |   top: "conv5_2_1x1_reduce"
5604 |   scale_param {
5605 |     bias_term: true
5606 |   }
5607 | }
5608 | layer {
5609 |   name: "conv5_2_1x1_reduce/relu"
5610 |   type: "ReLU"
5611 |   bottom: "conv5_2_1x1_reduce"
5612 |   top: "conv5_2_1x1_reduce"
5613 | }
5614 | layer {
5615 |   name: "conv5_2_3x3"
5616 |   type: "Convolution"
5617 |   bottom: "conv5_2_1x1_reduce"
5618 |   top: "conv5_2_3x3"
5619 |   convolution_param {
5620 |     num_output: 512
5621 |     bias_term: false
5622 |     pad: 1
5623 |     kernel_size: 3
5624 |     stride: 1
5625 |   }
5626 | }
5627 | layer {
5628 |   name: "conv5_2_3x3/bn"
5629 |   type: "BatchNorm"
5630 |   bottom: "conv5_2_3x3"
5631 |   top: "conv5_2_3x3"
5632 |   batch_norm_param {
5633 |     use_global_stats: true
5634 |   }
5635 | }
5636 | layer {
5637 |   name: "conv5_2_3x3/scale"
5638 |   type: "Scale"
5639 |   bottom: "conv5_2_3x3"
5640 |   top: "conv5_2_3x3"
5641 |   scale_param {
5642 |     bias_term: true
5643 |   }
5644 | }
5645 | layer {
5646 |   name: "conv5_2_3x3/relu"
5647 |   type: "ReLU"
5648 |   bottom: "conv5_2_3x3"
5649 |   top: "conv5_2_3x3"
5650 | }
5651 | layer {
5652 |   name: "conv5_2_1x1_increase"
5653 |   type: "Convolution"
5654 |   bottom: "conv5_2_3x3"
5655 |   top: "conv5_2_1x1_increase"
5656 |   convolution_param {
5657 |     num_output: 2048
5658 |     bias_term: false
5659 |     pad: 0
5660 |     kernel_size: 1
5661 |     stride: 1
5662 |   }
5663 | }
5664 | layer {
5665 |   name: "conv5_2_1x1_increase/bn"
5666 |   type: "BatchNorm"
5667 |   bottom: "conv5_2_1x1_increase"
5668 |   top: "conv5_2_1x1_increase"
5669 |   batch_norm_param {
5670 |     use_global_stats: true
5671 |   }
5672 | }
5673 | layer {
5674 |   name: "conv5_2_1x1_increase/scale"
5675 |   type: "Scale"
5676 |   bottom: "conv5_2_1x1_increase"
5677 |   top: "conv5_2_1x1_increase"
5678 |   scale_param {
5679 |     bias_term: true
5680 |   }
5681 | }
5682 | layer {
5683 |   name: "conv5_2"
5684 |   type: "Eltwise"
5685 |   bottom: "conv5_1"
5686 |   bottom: "conv5_2_1x1_increase"
5687 |   top: "conv5_2"
5688 |   eltwise_param {
5689 |     operation: SUM
5690 |   }
5691 | }
5692 | layer {
5693 |   name: "conv5_2/relu"
5694 |   type: "ReLU"
5695 |   bottom: "conv5_2"
5696 |   top: "conv5_2"
5697 | }
5698 | layer {
5699 |   name: "conv5_3_1x1_reduce"
5700 |   type: "Convolution"
5701 |   bottom: "conv5_2"
5702 |   top: "conv5_3_1x1_reduce"
5703 |   convolution_param {
5704 |     num_output: 512
5705 |     bias_term: false
5706 |     pad: 0
5707 |     kernel_size: 1
5708 |     stride: 1
5709 |   }
5710 | }
5711 | layer {
5712 |   name: "conv5_3_1x1_reduce/bn"
5713 |   type: "BatchNorm"
5714 |   bottom: "conv5_3_1x1_reduce"
5715 |   top: "conv5_3_1x1_reduce"
5716 |   batch_norm_param {
5717 |     use_global_stats: true
5718 |   }
5719 | }
5720 | layer {
5721 |   name: "conv5_3_1x1_reduce/scale"
5722 |   type: "Scale"
5723 |   bottom: "conv5_3_1x1_reduce"
5724 |   top: "conv5_3_1x1_reduce"
5725 |   scale_param {
5726 |     bias_term: true
5727 |   }
5728 | }
5729 | layer {
5730 |   name: "conv5_3_1x1_reduce/relu"
5731 |   type: "ReLU"
5732 |   bottom: "conv5_3_1x1_reduce"
5733 |   top: "conv5_3_1x1_reduce"
5734 | }
5735 | layer {
5736 |   name: "conv5_3_3x3"
5737 |   type: "Convolution"
5738 |   bottom: "conv5_3_1x1_reduce"
5739 |   top: "conv5_3_3x3"
5740 |   convolution_param {
5741 |     num_output: 512
5742 |     bias_term: false
5743 |     pad: 1
5744 |     kernel_size: 3
5745 |     stride: 1
5746 |   }
5747 | }
5748 | layer {
5749 |   name: "conv5_3_3x3/bn"
5750 |   type: "BatchNorm"
5751 |   bottom: "conv5_3_3x3"
5752 |   top: "conv5_3_3x3"
5753 |   batch_norm_param {
5754 |     use_global_stats: true
5755 |   }
5756 | }
5757 | layer {
5758 |   name: "conv5_3_3x3/scale"
5759 |   type: "Scale"
5760 |   bottom: "conv5_3_3x3"
5761 |   top: "conv5_3_3x3"
5762 |   scale_param {
5763 |     bias_term: true
5764 |   }
5765 | }
5766 | layer {
5767 |   name: "conv5_3_3x3/relu"
5768 |   type: "ReLU"
5769 |   bottom: "conv5_3_3x3"
5770 |   top: "conv5_3_3x3"
5771 | }
5772 | layer {
5773 |   name: "conv5_3_1x1_increase"
5774 |   type: "Convolution"
5775 |   bottom: "conv5_3_3x3"
5776 |   top: "conv5_3_1x1_increase"
5777 |   convolution_param {
5778 |     num_output: 2048
5779 |     bias_term: false
5780 |     pad: 0
5781 |     kernel_size: 1
5782 |     stride: 1
5783 |   }
5784 | }
5785 | layer {
5786 |   name: "conv5_3_1x1_increase/bn"
5787 |   type: "BatchNorm"
5788 |   bottom: "conv5_3_1x1_increase"
5789 |   top: "conv5_3_1x1_increase"
5790 |   batch_norm_param {
5791 |     use_global_stats: true
5792 |   }
5793 | }
5794 | layer {
5795 |   name: "conv5_3_1x1_increase/scale"
5796 |   type: "Scale"
5797 |   bottom: "conv5_3_1x1_increase"
5798 |   top: "conv5_3_1x1_increase"
5799 |   scale_param {
5800 |     bias_term: true
5801 |   }
5802 | }
5803 | layer {
5804 |   name: "conv5_3"
5805 |   type: "Eltwise"
5806 |   bottom: "conv5_2"
5807 |   bottom: "conv5_3_1x1_increase"
5808 |   top: "conv5_3"
5809 |   eltwise_param {
5810 |     operation: SUM
5811 |   }
5812 | }
5813 | layer {
5814 |   name: "conv5_3/relu"
5815 |   type: "ReLU"
5816 |   bottom: "conv5_3"
5817 |   top: "conv5_3"
5818 | }
5819 | layer {
5820 |   name: "pool5/7x7_s1"
5821 |   type: "Pooling"
5822 |   bottom: "conv5_3"
5823 |   top: "pool5/7x7_s1"
5824 |   pooling_param {
5825 |     pool: AVE
5826 |     kernel_size: 7
5827 |     stride: 1
5828 |   }
5829 | }
5830 | layer {
5831 |   name: "classifier"
5832 |   type: "InnerProduct"
5833 |   bottom: "pool5/7x7_s1"
5834 |   top: "classifier"
5835 |   inner_product_param {
5836 |     num_output: 365
5837 |   }
5838 | }
5839 | layer {
5840 |   name: "prob"
5841 |   type: "Softmax"
5842 |   bottom: "classifier"
5843 |   top: "prob"
5844 | }
5845 | 


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