ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
4
//
5
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
6
// in compliance with the License. You may obtain a copy of the License at
7
//
8
// https://opensource.org/licenses/BSD-3-Clause
9
//
10
// Unless required by applicable law or agreed to in writing, software distributed
11
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
12
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
13
// specific language governing permissions and limitations under the License.
14

15
#include "batchnorm.h"
16

17
namespace ncnn {
18

19
BatchNorm::BatchNorm()
20
{
21
    one_blob_only = true;
22
    support_inplace = true;
23
}
24

25
int BatchNorm::load_param(const ParamDict& pd)
26
{
27
    channels = pd.get(0, 0);
28
    eps = pd.get(1, 0.f);
29

30
    return 0;
31
}
32

33
int BatchNorm::load_model(const ModelBin& mb)
34
{
35
    slope_data = mb.load(channels, 1);
36
    if (slope_data.empty())
37
        return -100;
38

39
    mean_data = mb.load(channels, 1);
40
    if (mean_data.empty())
41
        return -100;
42

43
    var_data = mb.load(channels, 1);
44
    if (var_data.empty())
45
        return -100;
46

47
    bias_data = mb.load(channels, 1);
48
    if (bias_data.empty())
49
        return -100;
50

51
    a_data.create(channels);
52
    if (a_data.empty())
53
        return -100;
54
    b_data.create(channels);
55
    if (b_data.empty())
56
        return -100;
57

58
    for (int i = 0; i < channels; i++)
59
    {
60
        float sqrt_var = sqrtf(var_data[i] + eps);
61
        if (sqrt_var == 0.f)
62
            sqrt_var = 0.0001f; // sanitize divide by zero
63
        a_data[i] = bias_data[i] - slope_data[i] * mean_data[i] / sqrt_var;
64
        b_data[i] = slope_data[i] / sqrt_var;
65
    }
66

67
    return 0;
68
}
69

70
int BatchNorm::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
71
{
72
    // a = bias - slope * mean / sqrt(var)
73
    // b = slope / sqrt(var)
74
    // value = b * value + a
75

76
    int dims = bottom_top_blob.dims;
77

78
    if (dims == 1)
79
    {
80
        int w = bottom_top_blob.w;
81

82
        float* ptr = bottom_top_blob;
83

84
        #pragma omp parallel for num_threads(opt.num_threads)
85
        for (int i = 0; i < w; i++)
86
        {
87
            ptr[i] = b_data[i] * ptr[i] + a_data[i];
88
        }
89
    }
90

91
    if (dims == 2)
92
    {
93
        int w = bottom_top_blob.w;
94
        int h = bottom_top_blob.h;
95

96
        #pragma omp parallel for num_threads(opt.num_threads)
97
        for (int i = 0; i < h; i++)
98
        {
99
            float* ptr = bottom_top_blob.row(i);
100
            float a = a_data[i];
101
            float b = b_data[i];
102

103
            for (int j = 0; j < w; j++)
104
            {
105
                ptr[j] = b * ptr[j] + a;
106
            }
107
        }
108
    }
109

110
    if (dims == 3 || dims == 4)
111
    {
112
        int w = bottom_top_blob.w;
113
        int h = bottom_top_blob.h;
114
        int d = bottom_top_blob.d;
115
        int c = bottom_top_blob.c;
116
        int size = w * h * d;
117

118
        #pragma omp parallel for num_threads(opt.num_threads)
119
        for (int q = 0; q < c; q++)
120
        {
121
            float* ptr = bottom_top_blob.channel(q);
122
            float a = a_data[q];
123
            float b = b_data[q];
124

125
            for (int i = 0; i < size; i++)
126
            {
127
                ptr[i] = b * ptr[i] + a;
128
            }
129
        }
130
    }
131

132
    return 0;
133
}
134

135
} // namespace ncnn
136