forked from attractivechaos/sann
-
Notifications
You must be signed in to change notification settings - Fork 0
/
sfnn.c
136 lines (125 loc) · 4.91 KB
/
sfnn.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
#include <stdlib.h>
#include <string.h>
#include <assert.h>
#include <float.h>
#include <math.h>
#include "sann_priv.h"
int sfnn_n_par(int n_layers, const int32_t *n_neurons)
{
int k, n_par = 0;
for (k = 1; k < n_layers; ++k)
n_par += n_neurons[k] * (n_neurons[k-1] + 1);
return n_par;
}
#define sfnn_par2ptr(type_t, n_layers, n_neurons, _t, _w, _b) do { \
int _k; \
type_t _q = (_t); \
if ((_w) == 0) (_w) = (type_t*)calloc((n_layers), sizeof(type_t)); \
if ((_b) == 0) (_b) = (type_t*)calloc((n_layers), sizeof(type_t)); \
for (_k = 1; _k < (n_layers); ++_k) { \
(_b)[_k] = _q, _q += (n_neurons)[_k]; \
(_w)[_k] = _q, _q += (n_neurons)[_k] * (n_neurons)[_k-1]; \
} \
} while (0)
sfnn_buf_t *sfnn_buf_init(int n_layers, const int32_t *n_neurons, cfloat_p t)
{
int k, sum_neurons = 0;
sfnn_buf_t *b;
float *p;
b = (sfnn_buf_t*)calloc(1, sizeof(sfnn_buf_t));
b->out = (float**)calloc(n_layers * 5, sizeof(float*));
b->deriv = b->out + n_layers;
b->delta = b->deriv + n_layers;
b->dw = b->delta + n_layers;
b->db = b->dw + n_layers;
sfnn_par2ptr(cfloat_p, n_layers, n_neurons, t, b->w, b->b);
for (k = 1; k < n_layers; ++k)
sum_neurons += n_neurons[k];
p = b->buf = (float*)calloc(n_neurons[0] + sum_neurons * 3, sizeof(float));
b->out[0] = p, p += n_neurons[0]; // ->deriv[0] and ->delta[0] are not allocated
for (k = 1; k < n_layers; ++k) {
b->out[k] = p, p += n_neurons[k];
b->deriv[k] = p, p += n_neurons[k];
b->delta[k] = p, p += n_neurons[k];
}
return b;
}
void sfnn_buf_destroy(sfnn_buf_t *b)
{
free(b->w); free(b->b); free(b->out); free(b->buf); free(b);
}
void sfnn_core_forward(int n_layers, const int32_t *n_neurons, const int32_t *af, float r_in, float r_hidden, cfloat_p t, cfloat_p x, sfnn_buf_t *b)
{
int i, j, k;
float q[2] = { 1.0f / (1.0f - r_in), 1.0f / (1.0f - r_hidden) };
if (r_in > 0.0f && r_in < 1.0f) {
for (i = 0; i < n_neurons[0]; ++i)
b->out[0][i] = sann_drand() < r_in? 0.0f : x[i];
} else memcpy(b->out[0], x, n_neurons[0] * sizeof(float));
for (k = 1; k < n_layers; ++k) {
sann_activate_f func = sann_get_af(af[k-1]);
for (j = 0; j < n_neurons[k]; ++j)
if (k < n_layers - 1 && r_hidden > 0.0f && sann_drand() < r_hidden)
b->out[k][j] = b->deriv[k][j] = 0.0f;
else b->out[k][j] = func(q[k>1] * sann_sdot(n_neurons[k-1], b->w[k] + j * n_neurons[k-1], b->out[k-1]) + b->b[k][j], &b->deriv[k][j]);
}
}
void sfnn_core_backward(int n_layers, const int32_t *n_neurons, float r_in, float r_hidden, cfloat_p y, float *g, sfnn_buf_t *b)
{
int i, j, k;
float q[2] = { 1.0f / (1.0f - r_in), 1.0f / (1.0f - r_hidden) };
for (j = 0, k = n_layers - 1; j < n_neurons[k]; ++j) // calculate delta[] at the output layer
b->delta[k][j] = b->out[k][j] - y[j];
for (k = n_layers - 1; k > 1; --k) { // calculate delta[k-1]
memset(b->delta[k-1], 0, n_neurons[k-1] * sizeof(float));
for (j = 0; j < n_neurons[k]; ++j)
sann_saxpy(n_neurons[k-1], q[1] * b->delta[k][j], b->w[k] + j * n_neurons[k-1], b->delta[k-1]);
for (i = 0; i < n_neurons[k-1]; ++i)
b->delta[k-1][i] *= b->deriv[k-1][i];
}
sfnn_par2ptr(float_p, n_layers, n_neurons, g, b->dw, b->db);
for (k = 1; k < n_layers; ++k) { // update gradiant
sann_saxpy(n_neurons[k], 1., b->delta[k], b->db[k]);
for (j = 0; j < n_neurons[k]; ++j)
sann_saxpy(n_neurons[k-1], q[k>1] * b->delta[k][j], b->out[k-1], b->dw[k] + j * n_neurons[k-1]);
}
}
void sfnn_core_backprop(int n_layers, const int32_t *n_neurons, const int32_t *af, float r_in, float r_hidden, cfloat_p t, cfloat_p x, cfloat_p y, float *g, sfnn_buf_t *b)
{
assert(af[n_layers-2] == SANN_AF_SIGM);
sfnn_core_forward(n_layers, n_neurons, af, r_in, r_hidden, t, x, b);
sfnn_core_backward(n_layers, n_neurons, r_in, r_hidden, y, g, b);
}
void sfnn_core_randpar(int n_layers, const int32_t *n_neurons, float *t)
{
float **b = 0, **w = 0;
int k, j, iset = 0;
double gset;
sfnn_par2ptr(float_p, n_layers, n_neurons, t, w, b);
for (k = 1; k < n_layers; ++k) {
float t;
int tmp = n_neurons[k-1] * n_neurons[k];
t = sqrt(n_neurons[k-1] + 1);
for (j = 0; j < n_neurons[k]; ++j)
b[k][j] = sann_normal(&iset, &gset) / t;
for (j = 0; j < tmp; ++j)
w[k][j] = sann_normal(&iset, &gset) / t;
}
free(b); free(w);
}
void sfnn_core_jacobian(int n_layers, const int32_t *n_neurons, const int32_t *af, cfloat_p t, cfloat_p x, int w, float *d, sfnn_buf_t *b)
{
int i, j, k;
sfnn_core_forward(n_layers, n_neurons, af, 0.0f, 0.0f, t, x, b);
memset(b->delta[n_layers-1], 0, n_neurons[n_layers-1] * sizeof(float));
b->delta[n_layers-1][w] = 1.0f;
for (k = n_layers - 1; k > 1; --k) { // calculate delta[k-1]
memset(b->delta[k-1], 0, n_neurons[k-1] * sizeof(float));
for (j = 0; j < n_neurons[k]; ++j)
sann_saxpy(n_neurons[k-1], b->delta[k][j], b->w[k] + j * n_neurons[k-1], b->delta[k-1]);
for (i = 0; i < n_neurons[k-1]; ++i)
b->delta[k-1][i] *= b->deriv[k-1][i];
}
for (j = 0; j < n_neurons[1]; ++j)
sann_saxpy(n_neurons[0], b->delta[1][j], b->w[1] + j * n_neurons[0], d);
}