softmax_layer.c
3.49 KB
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
#include "softmax_layer.h"
#include "blas.h"
#include "cuda.h"
#include <float.h>
#include <math.h>
#include <stdlib.h>
#include <stdio.h>
#include <assert.h>
softmax_layer make_softmax_layer(int batch, int inputs, int groups)
{
assert(inputs%groups == 0);
fprintf(stderr, "softmax %4d\n", inputs);
softmax_layer l = {0};
l.type = SOFTMAX;
l.batch = batch;
l.groups = groups;
l.inputs = inputs;
l.outputs = inputs;
l.loss = calloc(inputs*batch, sizeof(float));
l.output = calloc(inputs*batch, sizeof(float));
l.delta = calloc(inputs*batch, sizeof(float));
l.cost = calloc(1, sizeof(float));
l.forward = forward_softmax_layer;
l.backward = backward_softmax_layer;
#ifdef GPU
l.forward_gpu = forward_softmax_layer_gpu;
l.backward_gpu = backward_softmax_layer_gpu;
l.output_gpu = cuda_make_array(l.output, inputs*batch);
l.loss_gpu = cuda_make_array(l.loss, inputs*batch);
l.delta_gpu = cuda_make_array(l.delta, inputs*batch);
#endif
return l;
}
void forward_softmax_layer(const softmax_layer l, network net)
{
if(l.softmax_tree){
int i;
int count = 0;
for (i = 0; i < l.softmax_tree->groups; ++i) {
int group_size = l.softmax_tree->group_size[i];
softmax_cpu(net.input + count, group_size, l.batch, l.inputs, 1, 0, 1, l.temperature, l.output + count);
count += group_size;
}
} else {
softmax_cpu(net.input, l.inputs/l.groups, l.batch, l.inputs, l.groups, l.inputs/l.groups, 1, l.temperature, l.output);
}
if(net.truth && !l.noloss){
softmax_x_ent_cpu(l.batch*l.inputs, l.output, net.truth, l.delta, l.loss);
l.cost[0] = sum_array(l.loss, l.batch*l.inputs);
}
}
void backward_softmax_layer(const softmax_layer l, network net)
{
axpy_cpu(l.inputs*l.batch, 1, l.delta, 1, net.delta, 1);
}
#ifdef GPU
void pull_softmax_layer_output(const softmax_layer layer)
{
cuda_pull_array(layer.output_gpu, layer.output, layer.inputs*layer.batch);
}
void forward_softmax_layer_gpu(const softmax_layer l, network net)
{
if(l.softmax_tree){
softmax_tree(net.input_gpu, 1, l.batch, l.inputs, l.temperature, l.output_gpu, *l.softmax_tree);
/*
int i;
int count = 0;
for (i = 0; i < l.softmax_tree->groups; ++i) {
int group_size = l.softmax_tree->group_size[i];
softmax_gpu(net.input_gpu + count, group_size, l.batch, l.inputs, 1, 0, 1, l.temperature, l.output_gpu + count);
count += group_size;
}
*/
} else {
if(l.spatial){
softmax_gpu(net.input_gpu, l.c, l.batch*l.c, l.inputs/l.c, l.w*l.h, 1, l.w*l.h, 1, l.output_gpu);
}else{
softmax_gpu(net.input_gpu, l.inputs/l.groups, l.batch, l.inputs, l.groups, l.inputs/l.groups, 1, l.temperature, l.output_gpu);
}
}
if(net.truth && !l.noloss){
softmax_x_ent_gpu(l.batch*l.inputs, l.output_gpu, net.truth_gpu, l.delta_gpu, l.loss_gpu);
if(l.softmax_tree){
mask_gpu(l.batch*l.inputs, l.delta_gpu, SECRET_NUM, net.truth_gpu, 0);
mask_gpu(l.batch*l.inputs, l.loss_gpu, SECRET_NUM, net.truth_gpu, 0);
}
cuda_pull_array(l.loss_gpu, l.loss, l.batch*l.inputs);
l.cost[0] = sum_array(l.loss, l.batch*l.inputs);
}
}
void backward_softmax_layer_gpu(const softmax_layer layer, network net)
{
axpy_gpu(layer.batch*layer.inputs, 1, layer.delta_gpu, 1, net.delta_gpu, 1);
}
#endif