caffe中的BatchNorm层
在训练一个小的分类网络时，发现加上BatchNorm层之后的检索效果相对于之前，效果会有提升，因此将该网络结构记录在这里，供以后查阅使用：
添加该层之前：
1 layer {
2 name: "conv1"
3 type: "Convolution"
4 bottom: "data"
5 top: "conv1"
6 param {
7 lr_mult: 1
8 decay_mult: 1
9 }
10 param {
11 lr_mult: 2
12 decay_mult: 0
13 }
14 convolution_param {
15 num_output: 64
16 kernel_size: 7
17 stride: 2
18 weight_filler {
19 type: "gaussian"
20 std: 0.01
21 }
22 bias_filler {
23 type: "constant"
24 value: 0
25 }
26 }
27 }
28 layer {
29 name: "relu1"
30 type: "ReLU"
31 bottom: "conv1"
32 top: "conv1"
33 }
34 layer {
35 name: "pool1"
36 type: "Pooling"
37 bottom: "conv1"
38 top: "pool1"
39 pooling_param {
40 pool: MAX
41 kernel_size: 3
42 stride: 2
43 }
44 }
45 layer {
46 name: "norm1"
47 type: "LRN"
48 bottom: "pool1"
49 top: "norm1"
50 lrn_param {
51 local_size: 5
52 alpha: 0.0001
53 beta: 0.75
54 }
55 }
56 layer {
57 name: "conv2"
58 type: "Convolution"
59 bottom: "norm1"
60 top: "conv2"
61 param {
62 lr_mult: 1
63 decay_mult: 1
64 }
65 param {
66 lr_mult: 2
67 decay_mult: 0
68 }
69 convolution_param {
70 num_output: 128
71 pad: 2
72 kernel_size: 5
73 stride: 2
74 group: 2
75 weight_filler {
76 type: "gaussian"
77 std: 0.01
78 }
79 bias_filler {
80 type: "constant"
81 value: 1
82 }
83 }
84 }
85 layer {
86 name: "relu2"
87 type: "ReLU"
88 bottom: "conv2"
89 top: "conv2"
90 }
91 layer {
92 name: "pool2"
93 type: "Pooling"
94 bottom: "conv2"
95 top: "pool2"
96 pooling_param {
97 pool: MAX
98 kernel_size: 3
99 stride: 2
100 }
101 }
102 layer {
103 name: "norm2"
104 type: "LRN"
105 bottom: "pool2"
106 top: "norm2"
107 lrn_param {
108 local_size: 5
109 alpha: 0.0001
110 beta: 0.75
111 }
112 }
113 layer {
114 name: "conv3"
115 type: "Convolution"
116 bottom: "norm2"
117 top: "conv3"
118 param {
119 lr_mult: 1
120 decay_mult: 1
121 }
122 param {
123 lr_mult: 2
124 decay_mult: 0
125 }
126 convolution_param {
127 num_output: 192
128 pad: 1
129 stride: 2
130 kernel_size: 3
131 weight_filler {
132 type: "gaussian"
133 std: 0.01
134 }
135 bias_filler {
136 type: "constant"
137 value: 0
138 }
139 }
140 }
141 layer {
142 name: "relu3"
143 type: "ReLU"
144 bottom: "conv3"
145 top: "conv3"
146 }
147 layer {
148 name: "conv4"
149 type: "Convolution"
150 bottom: "conv3"
151 top: "conv4"
152 param {
153 lr_mult: 1
154 decay_mult: 1
155 }
156 param {
157 lr_mult: 2
158 decay_mult: 0
159 }
160 convolution_param {
161 num_output: 192
162 pad: 1
163 kernel_size: 3
164 group: 2
165 weight_filler {
166 type: "gaussian"
167 std: 0.01
168 }
169 bias_filler {
170 type: "constant"
171 value: 1
172 }
173 }
174 }
175 layer {
176 name: "relu4"
177 type: "ReLU"
178 bottom: "conv4"
179 top: "conv4"
180 }
181 layer {
182 name: "conv5"
183 type: "Convolution"
184 bottom: "conv4"
185 top: "conv5"
186 param {
187 lr_mult: 1
188 decay_mult: 1
189 }
190 param {
191 lr_mult: 2
192 decay_mult: 0
193 }
194 convolution_param {
195 num_output: 128
196 pad: 1
197 stride: 2
198 kernel_size: 3
199 group: 2
200 weight_filler {
201 type: "gaussian"
202 std: 0.01
203 }
204 bias_filler {
205 type: "constant"
206 value: 1
207 }
208 }
209 }
210 layer {
211 name: "relu5"
212 type: "ReLU"
213 bottom: "conv5"
214 top: "conv5"
215 }
216 layer {
217 name: "pool5"
218 type: "Pooling"
219 bottom: "conv5"
220 top: "pool5"
221 pooling_param {
222 pool: MAX
223 kernel_size: 2
224 stride: 1
225 }
226 }
227 layer {
228 name: "fc6_srx"
229 type: "InnerProduct"
230 bottom: "pool5"
231 top: "fc6"
232 param {
233 lr_mult: 1
234 decay_mult: 1
235 }
236 param {
237 lr_mult: 2
238 decay_mult: 0
239 }
240 inner_product_param {
241 num_output: 768
242 weight_filler {
243 type: "gaussian"
244 std: 0.005
245 }
246 bias_filler {
247 type: "constant"
248 value: 1
249 }
250 }
251 }
252 layer {
253 name: "relu7"
254 type: "ReLU"
255 bottom: "fc6"
256 top: "fc6"
257 }
258 layer {
259 name: "drop7"
260 type: "Dropout"
261 bottom: "fc6"
262 top: "fc6"
263 dropout_param {
264 dropout_ratio: 0.5
265 }
266 }
267 layer {
268 name: "fc7_srx"
269 type: "InnerProduct"
270 bottom: "fc6"
271 top: "fc7"
272 param {
273 lr_mult: 1
274 decay_mult: 1
275 }
276 param {
277 lr_mult: 2
278 decay_mult: 0
279 }
280 inner_product_param {
281 num_output: 5275
282 weight_filler {
283 type: "gaussian"
284 std: 0.01
285 }
286 bias_filler {
287 type: "constant"
288 value: 0
289 }
290 }
291 }
292 layer{
293 name: "loss"
294 type: "SoftmaxWithLoss"
295 top: "SoftmaxWithLoss"
296 bottom: "fc7"
297 bottom: "label"
298 include {
299 phase: TRAIN
300 }
301 }
302 layer {
303 name: "accuracy"
304 type: "Accuracy"
305 bottom: "fc7"
306 bottom: "label"
307 top: "accuracy"
308 include {
309 phase: TEST
310 }
311 }添加该层之后：
1 layer {
2 name: "conv1"
3 type: "Convolution"
4 bottom: "data"
5 top: "conv1"
6 param {
7 lr_mult: 1
8 decay_mult: 1
9 }
10 param {
11 lr_mult: 2
12 decay_mult: 0
13 }
14 convolution_param {
15 num_output: 64
16 kernel_size: 7
17 stride: 2
18 weight_filler {
19 type: "gaussian"
20 std: 0.01
21 }
22 bias_filler {
23 type: "constant"
24 value: 0
25 }
26 }
27 }
28 ##############
29 layer {
30 bottom: "conv1"
31 top: "conv1"
32 name: "bn1"
33 type: "BatchNorm"
34 param {
35 lr_mult: 0
36 }
37 param {
38 lr_mult: 0
39 }
40 param {
41 lr_mult: 0
42 }
43 }
44 ##############
45 layer {
46 name: "relu1"
47 type: "ReLU"
48 bottom: "conv1"
49 top: "conv1"
50 }
51 layer {
52 name: "pool1"
53 type: "Pooling"
54 bottom: "conv1"
55 top: "pool1"
56 pooling_param {
57 pool: MAX
58 kernel_size: 3
59 stride: 2
60 }
61 }
62
63 layer {
64 name: "conv2"
65 type: "Convolution"
66 bottom: "pool1"
67 top: "conv2"
68 param {
69 lr_mult: 1
70 decay_mult: 1
71 }
72 param {
73 lr_mult: 2
74 decay_mult: 0
75 }
76 convolution_param {
77 num_output: 128
78 pad: 2
79 kernel_size: 5
80 stride: 2
81 group: 2
82 weight_filler {
83 type: "gaussian"
84 std: 0.01
85 }
86 bias_filler {
87 type: "constant"
88 value: 1
89 }
90 }
91 }
92 ##############
93 layer {
94 bottom: "conv2"
95 top: "conv2"
96 name: "bn2"
97 type: "BatchNorm"
98 param {
99 lr_mult: 0
100 }
101 param {
102 lr_mult: 0
103 }
104 param {
105 lr_mult: 0
106 }
107 }
108 ##############
109 layer {
110 name: "relu2"
111 type: "ReLU"
112 bottom: "conv2"
113 top: "conv2"
114 }
115 layer {
116 name: "pool2"
117 type: "Pooling"
118 bottom: "conv2"
119 top: "pool2"
120 pooling_param {
121 pool: MAX
122 kernel_size: 3
123 stride: 2
124 }
125 }
126
127 layer {
128 name: "conv3"
129 type: "Convolution"
130 bottom: "pool2"
131 top: "conv3"
132 param {
133 lr_mult: 1
134 decay_mult: 1
135 }
136 param {
137 lr_mult: 2
138 decay_mult: 0
139 }
140 convolution_param {
141 num_output: 192
142 pad: 1
143 stride: 2
144 kernel_size: 3
145 weight_filler {
146 type: "gaussian"
147 std: 0.01
148 }
149 bias_filler {
150 type: "constant"
151 value: 0
152 }
153 }
154 }
155 ##############
156 layer {
157 bottom: "conv3"
158 top: "conv3"
159 name: "bn3"
160 type: "BatchNorm"
161 param {
162 lr_mult: 0
163 }
164 param {
165 lr_mult: 0
166 }
167 param {
168 lr_mult: 0
169 }
170 }
171 ##############
172 layer {
173 name: "relu3"
174 type: "ReLU"
175 bottom: "conv3"
176 top: "conv3"
177 }
178 layer {
179 name: "conv4"
180 type: "Convolution"
181 bottom: "conv3"
182 top: "conv4"
183 param {
184 lr_mult: 1
185 decay_mult: 1
186 }
187 param {
188 lr_mult: 2
189 decay_mult: 0
190 }
191 convolution_param {
192 num_output: 192
193 pad: 1
194 kernel_size: 3
195 group: 2
196 weight_filler {
197 type: "gaussian"
198 std: 0.01
199 }
200 bias_filler {
201 type: "constant"
202 value: 1
203 }
204 }
205 }
206 ##############
207 layer {
208 bottom: "conv4"
209 top: "conv4"
210 name: "bn4"
211 type: "BatchNorm"
212 param {
213 lr_mult: 0
214 }
215 param {
216 lr_mult: 0
217 }
218 param {
219 lr_mult: 0
220 }
221 }
222 ##############
223 layer {
224 name: "relu4"
225 type: "ReLU"
226 bottom: "conv4"
227 top: "conv4"
228 }
229 layer {
230 name: "conv5"
231 type: "Convolution"
232 bottom: "conv4"
233 top: "conv5"
234 param {
235 lr_mult: 1
236 decay_mult: 1
237 }
238 param {
239 lr_mult: 2
240 decay_mult: 0
241 }
242 convolution_param {
243 num_output: 128
244 pad: 1
245 stride: 2
246 kernel_size: 3
247 group: 2
248 weight_filler {
249 type: "gaussian"
250 std: 0.01
251 }
252 bias_filler {
253 type: "constant"
254 value: 1
255 }
256 }
257 }
258 ##############
259 layer {
260 bottom: "conv5"
261 top: "conv5"
262 name: "bn5"
263 type: "BatchNorm"
264 param {
265 lr_mult: 0
266 }
267 param {
268 lr_mult: 0
269 }
270 param {
271 lr_mult: 0
272 }
273 }
274 ##############
275 layer {
276 name: "relu5"
277 type: "ReLU"
278 bottom: "conv5"
279 top: "conv5"
280 }
281 layer {
282 name: "pool5"
283 type: "Pooling"
284 bottom: "conv5"
285 top: "pool5"
286 pooling_param {
287 pool: MAX
288 kernel_size: 2
289 stride: 1
290 }
291 }
292 layer {
293 name: "fc6_srx"
294 type: "InnerProduct"
295 bottom: "pool5"
296 top: "fc6"
297 param {
298 lr_mult: 5
299 decay_mult: 2
300 }
301 param {
302 lr_mult: 8
303 decay_mult: 0
304 }
305 inner_product_param {
306 num_output: 768
307 weight_filler {
308 type: "gaussian"
309 std: 0.005
310 }
311 bias_filler {
312 type: "constant"
313 value: 1
314 }
315 }
316 }
317 layer {
318 name: "relu7"
319 type: "ReLU"
320 bottom: "fc6"
321 top: "fc6"
322 }
323 layer {
324 name: "drop7"
325 type: "Dropout"
326 bottom: "fc6"
327 top: "fc6"
328 dropout_param {
329 dropout_ratio: 0.5
330 }
331 }
332 layer {
333 name: "fc7_srx"
334 type: "InnerProduct"
335 bottom: "fc6"
336 top: "fc7"
337 param {
338 lr_mult: 5
339 decay_mult: 2
340 }
341 param {
342 lr_mult: 8
343 decay_mult: 0
344 }
345 inner_product_param {
346 num_output: 5275
347 weight_filler {
348 type: "gaussian"
349 std: 0.01
350 }
351 bias_filler {
352 type: "constant"
353 value: 0
354 }
355 }
356 }
357 layer{
358 name: "loss"
359 type: "SoftmaxWithLoss"
360 top: "SoftmaxWithLoss"
361 bottom: "fc7"
362 bottom: "label"
363 include {
364 phase: TRAIN
365 }
366 }
367 layer {
368 name: "accuracy"
369 type: "Accuracy"
370 bottom: "fc7"
371 bottom: "label"
372 top: "accuracy"
373 include {
374 phase: TEST
375 }
376 }