1 // Copyright 2016 SoundAI Technology Co., Ltd. (author: Charles Wang)
2 //
3 // Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
4 // in compliance with the License. You may obtain a copy of the License at
5 //
6 // https://opensource.org/licenses/BSD-3-Clause
7 //
8 // Unless required by applicable law or agreed to in writing, software distributed
9 // under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
10 // CONDITIONS OF ANY KIND, either express or implied. See the License for the
11 // specific language governing permissions and limitations under the License.
12
13 #include "statisticspooling.h"
14
15 #include <float.h>
16 #include <limits.h>
17 #include <math.h>
18
19 namespace ncnn {
20
StatisticsPooling()21 StatisticsPooling::StatisticsPooling()
22 {
23 one_blob_only = true;
24 support_inplace = false;
25 }
26
load_param(const ParamDict & pd)27 int StatisticsPooling::load_param(const ParamDict& pd)
28 {
29 include_stddev = pd.get(0, 0);
30
31 return 0;
32 }
33
forward(const Mat & bottom_blob,Mat & top_blob,const Option & opt) const34 int StatisticsPooling::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
35 {
36 int w = bottom_blob.w;
37 int h = bottom_blob.h;
38 int channels = bottom_blob.c;
39 int size = w * h;
40 size_t elemsize = bottom_blob.elemsize;
41
42 int out_channels = channels;
43 if (include_stddev)
44 {
45 out_channels *= 2;
46 }
47
48 top_blob.create(out_channels, elemsize, opt.blob_allocator);
49
50 #pragma omp parallel for num_threads(opt.num_threads)
51 for (int q = 0; q < channels; q++)
52 {
53 const float* ptr = bottom_blob.channel(q);
54
55 float mean = 0.f;
56 for (int i = 0; i < size; i++)
57 {
58 mean += ptr[i];
59 }
60 top_blob[q] = mean / w / h;
61 }
62
63 #pragma omp parallel for num_threads(opt.num_threads)
64 for (int q = channels; q < out_channels; q++)
65 {
66 const float* ptr = bottom_blob.channel(q - channels);
67
68 float std = 0.f;
69 for (int i = 0; i < size; i++)
70 {
71 std += pow((ptr[i] - top_blob[q - channels]), 2);
72 }
73 top_blob[q] = sqrt(std / w / h);
74 }
75
76 return 0;
77 }
78
79 } // namespace ncnn
80