1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
|
#include <algorithm>
#include <vector>
#include "caffe/common_layers.hpp"
#include "caffe/layer.hpp"
#include "caffe/util/math_functions.hpp"
namespace caffe {
template <typename Dtype>
void MVNLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
const Dtype* bottom_data = bottom[0]->gpu_data();
Dtype* top_data = top[0]->mutable_gpu_data();
int num;
if (this->layer_param_.mvn_param().across_channels())
num = bottom[0]->num();
else
num = bottom[0]->num() * bottom[0]->channels();
int dim = bottom[0]->count() / num;
if (this->layer_param_.mvn_param().normalize_variance()) {
// put the squares of bottom into temp_
caffe_gpu_powx(bottom[0]->count(), bottom_data, Dtype(2),
temp_.mutable_gpu_data());
// computes variance using var(X) = E(X^2) - (EX)^2
caffe_gpu_gemv<Dtype>(CblasNoTrans, num, dim, 1. / dim, bottom_data,
sum_multiplier_.gpu_data(), 0., mean_.mutable_gpu_data()); // EX
caffe_gpu_gemv<Dtype>(CblasNoTrans, num, dim, 1. / dim, temp_.gpu_data(),
sum_multiplier_.gpu_data(), 0.,
variance_.mutable_gpu_data()); // E(X^2)
caffe_gpu_powx(mean_.count(), mean_.gpu_data(), Dtype(2),
temp_.mutable_gpu_data()); // (EX)^2
caffe_gpu_sub(mean_.count(), variance_.gpu_data(), temp_.gpu_data(),
variance_.mutable_gpu_data()); // variance
// do mean and variance normalization
// subtract mean
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, dim, 1, -1.,
mean_.gpu_data(), sum_multiplier_.gpu_data(), 0.,
temp_.mutable_gpu_data());
caffe_gpu_add(temp_.count(), bottom_data, temp_.gpu_data(), top_data);
// normalize variance
caffe_gpu_powx(variance_.count(), variance_.gpu_data(), Dtype(0.5),
variance_.mutable_gpu_data());
caffe_gpu_add_scalar(variance_.count(), eps_, variance_.mutable_gpu_data());
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, dim, 1, 1.,
variance_.gpu_data(), sum_multiplier_.gpu_data(), 0.,
temp_.mutable_gpu_data());
caffe_gpu_div(temp_.count(), top_data, temp_.gpu_data(), top_data);
} else {
caffe_gpu_gemv<Dtype>(CblasNoTrans, num, dim, 1. / dim, bottom_data,
sum_multiplier_.gpu_data(), 0., mean_.mutable_gpu_data()); // EX
// subtract mean
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, dim, 1, -1.,
mean_.gpu_data(), sum_multiplier_.gpu_data(), 0.,
temp_.mutable_gpu_data());
caffe_gpu_add(temp_.count(), bottom_data, temp_.gpu_data(), top_data);
}
}
template <typename Dtype>
void MVNLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down,
const vector<Blob<Dtype>*>& bottom) {
const Dtype* top_diff = top[0]->gpu_diff();
const Dtype* top_data = top[0]->gpu_data();
const Dtype* bottom_data = bottom[0]->gpu_data();
Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
int num;
if (this->layer_param_.mvn_param().across_channels())
num = bottom[0]->num();
else
num = bottom[0]->num() * bottom[0]->channels();
int dim = bottom[0]->count() / num;
if (this->layer_param_.mvn_param().normalize_variance()) {
caffe_gpu_mul(temp_.count(), top_data, top_diff, bottom_diff);
caffe_gpu_gemv<Dtype>(CblasNoTrans, num, dim, 1., bottom_diff,
sum_multiplier_.gpu_data(), 0., mean_.mutable_gpu_data());
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, dim, 1, 1.,
mean_.gpu_data(), sum_multiplier_.gpu_data(), 0.,
bottom_diff);
caffe_gpu_mul(temp_.count(), top_data, bottom_diff, bottom_diff);
caffe_gpu_gemv<Dtype>(CblasNoTrans, num, dim, 1., top_diff,
sum_multiplier_.gpu_data(), 0., mean_.mutable_gpu_data());
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, dim, 1, 1.,
mean_.gpu_data(), sum_multiplier_.gpu_data(), 1.,
bottom_diff);
caffe_gpu_axpby(temp_.count(), Dtype(1), top_diff, Dtype(-1. / dim),
bottom_diff);
// put the squares of bottom into temp_
caffe_gpu_powx(temp_.count(), bottom_data, Dtype(2),
temp_.mutable_gpu_data());
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, dim, 1, 1.,
variance_.gpu_data(), sum_multiplier_.gpu_data(), 0.,
temp_.mutable_gpu_data());
caffe_gpu_div(temp_.count(), bottom_diff, temp_.gpu_data(), bottom_diff);
} else {
caffe_gpu_gemv<Dtype>(CblasNoTrans, num, dim, 1. / dim, top_diff,
sum_multiplier_.gpu_data(), 0., mean_.mutable_gpu_data());
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, dim, 1, -1.,
mean_.gpu_data(), sum_multiplier_.gpu_data(), 0.,
temp_.mutable_gpu_data());
caffe_gpu_add(temp_.count(), top_diff, temp_.gpu_data(), bottom_diff);
}
}
INSTANTIATE_LAYER_GPU_FUNCS(MVNLayer);
} // namespace caffe
|
