forked from pjreddie/darknet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbatchnorm_layer.c
211 lines (180 loc) · 8.01 KB
/
batchnorm_layer.c
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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
#include "batchnorm_layer.h"
#include "blas.h"
#include <stdio.h>
layer make_batchnorm_layer(int batch, int w, int h, int c)
{
fprintf(stderr, "Batch Normalization Layer: %d x %d x %d image\n", w,h,c);
layer layer = {0};
layer.type = BATCHNORM;
layer.batch = batch;
layer.h = layer.out_h = h;
layer.w = layer.out_w = w;
layer.c = layer.out_c = c;
layer.output = calloc(h * w * c * batch, sizeof(float));
layer.delta = calloc(h * w * c * batch, sizeof(float));
layer.inputs = w*h*c;
layer.outputs = layer.inputs;
layer.scales = calloc(c, sizeof(float));
layer.scale_updates = calloc(c, sizeof(float));
int i;
for(i = 0; i < c; ++i){
layer.scales[i] = 1;
}
layer.mean = calloc(c, sizeof(float));
layer.variance = calloc(c, sizeof(float));
layer.rolling_mean = calloc(c, sizeof(float));
layer.rolling_variance = calloc(c, sizeof(float));
layer.forward = forward_batchnorm_layer;
layer.backward = backward_batchnorm_layer;
#ifdef GPU
layer.forward_gpu = forward_batchnorm_layer_gpu;
layer.backward_gpu = backward_batchnorm_layer_gpu;
layer.output_gpu = cuda_make_array(layer.output, h * w * c * batch);
layer.delta_gpu = cuda_make_array(layer.delta, h * w * c * batch);
layer.scales_gpu = cuda_make_array(layer.scales, c);
layer.scale_updates_gpu = cuda_make_array(layer.scale_updates, c);
layer.mean_gpu = cuda_make_array(layer.mean, c);
layer.variance_gpu = cuda_make_array(layer.variance, c);
layer.rolling_mean_gpu = cuda_make_array(layer.mean, c);
layer.rolling_variance_gpu = cuda_make_array(layer.variance, c);
layer.mean_delta_gpu = cuda_make_array(layer.mean, c);
layer.variance_delta_gpu = cuda_make_array(layer.variance, c);
layer.x_gpu = cuda_make_array(layer.output, layer.batch*layer.outputs);
layer.x_norm_gpu = cuda_make_array(layer.output, layer.batch*layer.outputs);
#endif
return layer;
}
void backward_scale_cpu(float *x_norm, float *delta, int batch, int n, int size, float *scale_updates)
{
int i,b,f;
for(f = 0; f < n; ++f){
float sum = 0;
for(b = 0; b < batch; ++b){
for(i = 0; i < size; ++i){
int index = i + size*(f + n*b);
sum += delta[index] * x_norm[index];
}
}
scale_updates[f] += sum;
}
}
void mean_delta_cpu(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta)
{
int i,j,k;
for(i = 0; i < filters; ++i){
mean_delta[i] = 0;
for (j = 0; j < batch; ++j) {
for (k = 0; k < spatial; ++k) {
int index = j*filters*spatial + i*spatial + k;
mean_delta[i] += delta[index];
}
}
mean_delta[i] *= (-1./sqrt(variance[i] + .00001f));
}
}
void variance_delta_cpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *variance_delta)
{
int i,j,k;
for(i = 0; i < filters; ++i){
variance_delta[i] = 0;
for(j = 0; j < batch; ++j){
for(k = 0; k < spatial; ++k){
int index = j*filters*spatial + i*spatial + k;
variance_delta[i] += delta[index]*(x[index] - mean[i]);
}
}
variance_delta[i] *= -.5 * pow(variance[i] + .00001f, (float)(-3./2.));
}
}
void normalize_delta_cpu(float *x, float *mean, float *variance, float *mean_delta, float *variance_delta, int batch, int filters, int spatial, float *delta)
{
int f, j, k;
for(j = 0; j < batch; ++j){
for(f = 0; f < filters; ++f){
for(k = 0; k < spatial; ++k){
int index = j*filters*spatial + f*spatial + k;
delta[index] = delta[index] * 1./(sqrt(variance[f]) + .00001f) + variance_delta[f] * 2. * (x[index] - mean[f]) / (spatial * batch) + mean_delta[f]/(spatial*batch);
}
}
}
}
void resize_batchnorm_layer(layer *layer, int w, int h)
{
fprintf(stderr, "Not implemented\n");
}
void forward_batchnorm_layer(layer l, network_state state)
{
if(l.type == BATCHNORM) copy_cpu(l.outputs*l.batch, state.input, 1, l.output, 1);
if(l.type == CONNECTED){
l.out_c = l.outputs;
l.out_h = l.out_w = 1;
}
if(state.train){
mean_cpu(l.output, l.batch, l.out_c, l.out_h*l.out_w, l.mean);
variance_cpu(l.output, l.mean, l.batch, l.out_c, l.out_h*l.out_w, l.variance);
scal_cpu(l.out_c, .9, l.rolling_mean, 1);
axpy_cpu(l.out_c, .1, l.mean, 1, l.rolling_mean, 1);
scal_cpu(l.out_c, .9, l.rolling_variance, 1);
axpy_cpu(l.out_c, .1, l.variance, 1, l.rolling_variance, 1);
copy_cpu(l.outputs*l.batch, l.output, 1, l.x, 1);
normalize_cpu(l.output, l.mean, l.variance, l.batch, l.out_c, l.out_h*l.out_w);
copy_cpu(l.outputs*l.batch, l.output, 1, l.x_norm, 1);
} else {
normalize_cpu(l.output, l.rolling_mean, l.rolling_variance, l.batch, l.out_c, l.out_h*l.out_w);
}
scale_bias(l.output, l.scales, l.batch, l.out_c, l.out_h*l.out_w);
}
void backward_batchnorm_layer(const layer l, network_state state)
{
backward_scale_cpu(l.x_norm, l.delta, l.batch, l.out_c, l.out_w*l.out_h, l.scale_updates);
scale_bias(l.delta, l.scales, l.batch, l.out_c, l.out_h*l.out_w);
mean_delta_cpu(l.delta, l.variance, l.batch, l.out_c, l.out_w*l.out_h, l.mean_delta);
variance_delta_cpu(l.x, l.delta, l.mean, l.variance, l.batch, l.out_c, l.out_w*l.out_h, l.variance_delta);
normalize_delta_cpu(l.x, l.mean, l.variance, l.mean_delta, l.variance_delta, l.batch, l.out_c, l.out_w*l.out_h, l.delta);
if(l.type == BATCHNORM) copy_cpu(l.outputs*l.batch, l.delta, 1, state.delta, 1);
}
#ifdef GPU
void pull_batchnorm_layer(layer l)
{
cuda_pull_array(l.scales_gpu, l.scales, l.c);
cuda_pull_array(l.rolling_mean_gpu, l.rolling_mean, l.c);
cuda_pull_array(l.rolling_variance_gpu, l.rolling_variance, l.c);
}
void push_batchnorm_layer(layer l)
{
cuda_push_array(l.scales_gpu, l.scales, l.c);
cuda_push_array(l.rolling_mean_gpu, l.rolling_mean, l.c);
cuda_push_array(l.rolling_variance_gpu, l.rolling_variance, l.c);
}
void forward_batchnorm_layer_gpu(layer l, network_state state)
{
if(l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, state.input, 1, l.output_gpu, 1);
if(l.type == CONNECTED){
l.out_c = l.outputs;
l.out_h = l.out_w = 1;
}
if (state.train) {
fast_mean_gpu(l.output_gpu, l.batch, l.out_c, l.out_h*l.out_w, l.mean_gpu);
fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.out_c, l.out_h*l.out_w, l.variance_gpu);
scal_ongpu(l.out_c, .99, l.rolling_mean_gpu, 1);
axpy_ongpu(l.out_c, .01, l.mean_gpu, 1, l.rolling_mean_gpu, 1);
scal_ongpu(l.out_c, .99, l.rolling_variance_gpu, 1);
axpy_ongpu(l.out_c, .01, l.variance_gpu, 1, l.rolling_variance_gpu, 1);
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_gpu, 1);
normalize_gpu(l.output_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.out_c, l.out_h*l.out_w);
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_norm_gpu, 1);
} else {
normalize_gpu(l.output_gpu, l.rolling_mean_gpu, l.rolling_variance_gpu, l.batch, l.out_c, l.out_h*l.out_w);
}
scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
}
void backward_batchnorm_layer_gpu(const layer l, network_state state)
{
backward_scale_gpu(l.x_norm_gpu, l.delta_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.scale_updates_gpu);
scale_bias_gpu(l.delta_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.mean_delta_gpu);
fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.variance_delta_gpu);
normalize_delta_gpu(l.x_gpu, l.mean_gpu, l.variance_gpu, l.mean_delta_gpu, l.variance_delta_gpu, l.batch, l.out_c, l.out_w*l.out_h, l.delta_gpu);
if(l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, l.delta_gpu, 1, state.delta, 1);
}
#endif