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
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
|
#ifndef TH_GENERIC_FILE
#define TH_GENERIC_FILE "generic/BatchNormalization.c"
#else
void THNN_(BatchNormalization_updateOutput)(
THNNState *state, THTensor *input, THTensor *output,
THTensor *weight, THTensor *bias,
THTensor *running_mean, THTensor *running_var,
THTensor *save_mean, THTensor *save_std,
bool train, double momentum, double eps)
{
THTensor_(resizeAs)(output, input);
long nInput = THTensor_(size)(input, 1);
long f;
ptrdiff_t n = THTensor_(nElement)(input) / nInput;
#pragma omp parallel for
for (f = 0; f < nInput; ++f) {
THTensor *in = THTensor_(newSelect)(input, 1, f);
THTensor *out = THTensor_(newSelect)(output, 1, f);
real mean, invstd;
if (train) {
// compute mean per input
accreal sum = 0;
TH_TENSOR_APPLY(real, in, sum += *in_data;);
mean = (real) sum / n;
THTensor_(set1d)(save_mean, f, (real) mean);
// compute variance per input
sum = 0;
TH_TENSOR_APPLY(real, in,
sum += (*in_data - mean) * (*in_data - mean););
if (sum == 0 && eps == 0.0) {
invstd = 0;
} else {
invstd = (real) (1 / sqrt(sum/n + eps));
}
THTensor_(set1d)(save_std, f, (real) invstd);
// update running averages
THTensor_(set1d)(running_mean, f,
(real) (momentum * mean + (1 - momentum) * THTensor_(get1d)(running_mean, f)));
accreal unbiased_var = sum / (n - 1);
THTensor_(set1d)(running_var, f,
(real) (momentum * unbiased_var + (1 - momentum) * THTensor_(get1d)(running_var, f)));
} else {
mean = THTensor_(get1d)(running_mean, f);
invstd = 1 / sqrt(THTensor_(get1d)(running_var, f) + eps);
}
// compute output
real w = weight ? THTensor_(get1d)(weight, f) : 1;
real b = bias ? THTensor_(get1d)(bias, f) : 0;
TH_TENSOR_APPLY2(real, in, real, out,
*out_data = (real) (((*in_data - mean) * invstd) * w + b););
THTensor_(free)(out);
THTensor_(free)(in);
}
}
void THNN_(BatchNormalization_backward)(
THNNState *state, THTensor *input, THTensor *gradOutput, THTensor *gradInput,
THTensor *gradWeight, THTensor *gradBias, THTensor *weight,
THTensor *running_mean, THTensor *running_var,
THTensor *save_mean, THTensor *save_std,
bool train, double scale, double eps)
{
THNN_CHECK_SHAPE(input, gradOutput);
long nInput = THTensor_(size)(input, 1);
long f;
ptrdiff_t n = THTensor_(nElement)(input) / nInput;
#pragma omp parallel for
for (f = 0; f < nInput; ++f) {
THTensor *in = THTensor_(newSelect)(input, 1, f);
THTensor *gradOut = THTensor_(newSelect)(gradOutput, 1, f);
real w = weight ? THTensor_(get1d)(weight, f) : 1;
real mean, invstd;
if (train) {
mean = THTensor_(get1d)(save_mean, f);
invstd = THTensor_(get1d)(save_std, f);
} else {
mean = THTensor_(get1d)(running_mean, f);
invstd = 1 / sqrt(THTensor_(get1d)(running_var, f) + eps);
}
// sum over all gradOutput in feature plane
accreal sum = 0;
TH_TENSOR_APPLY(real, gradOut, sum += *gradOut_data;);
// dot product of the Q(X) and gradOuput
accreal dotp = 0;
TH_TENSOR_APPLY2(real, in, real, gradOut,
dotp += (*in_data - mean) * (*gradOut_data););
if (gradInput) {
THTensor_(resizeAs)(gradInput, input);
THTensor *gradIn = THTensor_(newSelect)(gradInput, 1, f);
if (train) {
// when in training mode
// Q(X) = X - E[x] ; i.e. input centered to zero mean
// Y = Q(X) / σ ; i.e. BN output before weight and bias
// dL/dX = (Q(dL/dY) - dot(Y, dL/dY) * Y) / σ * w
// projection of gradOutput on to output scaled by std
real k = (real) dotp * invstd * invstd / n;
TH_TENSOR_APPLY2(real, gradIn, real, in,
*gradIn_data = (*in_data - mean) * k;);
accreal gradMean = sum / n;
TH_TENSOR_APPLY2(real, gradIn, real, gradOut,
*gradIn_data = (*gradOut_data - gradMean - *gradIn_data) * invstd * w;);
} else {
// when in evaluation mode
// Q(X) = X - running_mean ; i.e. input centered to zero mean
// Y = Q(X) / running_std ; i.e. BN output before weight and bias
// dL/dX = w / running_std
TH_TENSOR_APPLY2(real, gradIn, real, gradOut,
*gradIn_data = *gradOut_data * invstd * w;);
}
THTensor_(free)(gradIn);
}
if (gradWeight) {
real val = THTensor_(get1d)(gradWeight, f);
THTensor_(set1d)(gradWeight, f, val + scale * dotp * invstd);
}
if (gradBias) {
real val = THTensor_(get1d)(gradBias, f);
THTensor_(set1d)(gradBias, f, val + scale * sum);
}
THTensor_(free)(gradOut);
THTensor_(free)(in);
}
}
#endif
|
