moo_fann 0.1.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +7 -0
- data/ext/moo_fann/config.h +8 -0
- data/ext/moo_fann/doublefann.c +30 -0
- data/ext/moo_fann/doublefann.h +33 -0
- data/ext/moo_fann/extconf.rb +25 -0
- data/ext/moo_fann/fann.c +1803 -0
- data/ext/moo_fann/fann.h +613 -0
- data/ext/moo_fann/fann_activation.h +144 -0
- data/ext/moo_fann/fann_augment.h +133 -0
- data/ext/moo_fann/fann_cascade.c +1048 -0
- data/ext/moo_fann/fann_cascade.h +557 -0
- data/ext/moo_fann/fann_data.h +824 -0
- data/ext/moo_fann/fann_error.c +210 -0
- data/ext/moo_fann/fann_error.h +165 -0
- data/ext/moo_fann/fann_internal.h +152 -0
- data/ext/moo_fann/fann_io.c +802 -0
- data/ext/moo_fann/fann_io.h +100 -0
- data/ext/moo_fann/fann_train.c +1047 -0
- data/ext/moo_fann/fann_train.h +1310 -0
- data/ext/moo_fann/fann_train_data.c +1243 -0
- data/ext/moo_fann/moo_fann.c +1768 -0
- data/ext/moo_fann/ruby_compat.h +12 -0
- data/lib/moo_fann.rb +30 -0
- data/lib/moo_fann/version.rb +30 -0
- metadata +75 -0
@@ -0,0 +1,100 @@
|
|
1
|
+
/*
|
2
|
+
Fast Artificial Neural Network Library (fann)
|
3
|
+
Copyright (C) 2003-2012 Steffen Nissen (sn@leenissen.dk)
|
4
|
+
|
5
|
+
This library is free software; you can redistribute it and/or
|
6
|
+
modify it under the terms of the GNU Lesser General Public
|
7
|
+
License as published by the Free Software Foundation; either
|
8
|
+
version 2.1 of the License, or (at your option) any later version.
|
9
|
+
|
10
|
+
This library is distributed in the hope that it will be useful,
|
11
|
+
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
12
|
+
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
13
|
+
Lesser General Public License for more details.
|
14
|
+
|
15
|
+
You should have received a copy of the GNU Lesser General Public
|
16
|
+
License along with this library; if not, write to the Free Software
|
17
|
+
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
|
18
|
+
*/
|
19
|
+
|
20
|
+
#ifndef __fann_io_h__
|
21
|
+
#define __fann_io_h__
|
22
|
+
|
23
|
+
/* Section: FANN File Input/Output
|
24
|
+
|
25
|
+
It is possible to save an entire ann to a file with <fann_save> for future loading with <fann_create_from_file>.
|
26
|
+
*/
|
27
|
+
|
28
|
+
/* Group: File Input and Output */
|
29
|
+
|
30
|
+
/* Function: fann_create_from_file
|
31
|
+
|
32
|
+
Constructs a backpropagation neural network from a configuration file, which have been saved by <fann_save>.
|
33
|
+
|
34
|
+
See also:
|
35
|
+
<fann_save>, <fann_save_to_fixed>
|
36
|
+
|
37
|
+
This function appears in FANN >= 1.0.0.
|
38
|
+
*/
|
39
|
+
FANN_EXTERNAL struct fann *FANN_API fann_create_from_file(const char *configuration_file);
|
40
|
+
|
41
|
+
|
42
|
+
/* Function: fann_save
|
43
|
+
|
44
|
+
Save the entire network to a configuration file.
|
45
|
+
|
46
|
+
The configuration file contains all information about the neural network and enables
|
47
|
+
<fann_create_from_file> to create an exact copy of the neural network and all of the
|
48
|
+
parameters associated with the neural network.
|
49
|
+
|
50
|
+
These three parameters (<fann_set_callback>, <fann_set_error_log>,
|
51
|
+
<fann_set_user_data>) are *NOT* saved to the file because they cannot safely be
|
52
|
+
ported to a different location. Also temporary parameters generated during training
|
53
|
+
like <fann_get_MSE> is not saved.
|
54
|
+
|
55
|
+
Return:
|
56
|
+
The function returns 0 on success and -1 on failure.
|
57
|
+
|
58
|
+
See also:
|
59
|
+
<fann_create_from_file>, <fann_save_to_fixed>
|
60
|
+
|
61
|
+
This function appears in FANN >= 1.0.0.
|
62
|
+
*/
|
63
|
+
FANN_EXTERNAL int FANN_API fann_save(struct fann *ann, const char *configuration_file);
|
64
|
+
|
65
|
+
|
66
|
+
/* Function: fann_save_to_fixed
|
67
|
+
|
68
|
+
Saves the entire network to a configuration file.
|
69
|
+
But it is saved in fixed point format no matter which
|
70
|
+
format it is currently in.
|
71
|
+
|
72
|
+
This is usefull for training a network in floating points,
|
73
|
+
and then later executing it in fixed point.
|
74
|
+
|
75
|
+
The function returns the bit position of the fix point, which
|
76
|
+
can be used to find out how accurate the fixed point network will be.
|
77
|
+
A high value indicates high precision, and a low value indicates low
|
78
|
+
precision.
|
79
|
+
|
80
|
+
A negative value indicates very low precision, and a very
|
81
|
+
strong possibility for overflow.
|
82
|
+
(the actual fix point will be set to 0, since a negative
|
83
|
+
fix point does not make sence).
|
84
|
+
|
85
|
+
Generally, a fix point lower than 6 is bad, and should be avoided.
|
86
|
+
The best way to avoid this, is to have less connections to each neuron,
|
87
|
+
or just less neurons in each layer.
|
88
|
+
|
89
|
+
The fixed point use of this network is only intended for use on machines that
|
90
|
+
have no floating point processor, like an iPAQ. On normal computers the floating
|
91
|
+
point version is actually faster.
|
92
|
+
|
93
|
+
See also:
|
94
|
+
<fann_create_from_file>, <fann_save>
|
95
|
+
|
96
|
+
This function appears in FANN >= 1.0.0.
|
97
|
+
*/
|
98
|
+
FANN_EXTERNAL int FANN_API fann_save_to_fixed(struct fann *ann, const char *configuration_file);
|
99
|
+
|
100
|
+
#endif
|
@@ -0,0 +1,1047 @@
|
|
1
|
+
/*
|
2
|
+
Fast Artificial Neural Network Library (fann)
|
3
|
+
Copyright (C) 2003-2012 Steffen Nissen (sn@leenissen.dk)
|
4
|
+
|
5
|
+
This library is free software; you can redistribute it and/or
|
6
|
+
modify it under the terms of the GNU Lesser General Public
|
7
|
+
License as published by the Free Software Foundation; either
|
8
|
+
version 2.1 of the License, or (at your option) any later version.
|
9
|
+
|
10
|
+
This library is distributed in the hope that it will be useful,
|
11
|
+
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
12
|
+
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
13
|
+
Lesser General Public License for more details.
|
14
|
+
|
15
|
+
You should have received a copy of the GNU Lesser General Public
|
16
|
+
License along with this library; if not, write to the Free Software
|
17
|
+
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
|
18
|
+
*/
|
19
|
+
|
20
|
+
#include <stdio.h>
|
21
|
+
#include <stdlib.h>
|
22
|
+
#include <stdarg.h>
|
23
|
+
#include <string.h>
|
24
|
+
#include <math.h>
|
25
|
+
|
26
|
+
#include "config.h"
|
27
|
+
#include "fann.h"
|
28
|
+
|
29
|
+
/*#define DEBUGTRAIN*/
|
30
|
+
|
31
|
+
#ifndef FIXEDFANN
|
32
|
+
/* INTERNAL FUNCTION
|
33
|
+
Calculates the derived of a value, given an activation function
|
34
|
+
and a steepness
|
35
|
+
*/
|
36
|
+
fann_type fann_activation_derived(unsigned int activation_function,
|
37
|
+
fann_type steepness, fann_type value, fann_type sum)
|
38
|
+
{
|
39
|
+
switch (activation_function)
|
40
|
+
{
|
41
|
+
case FANN_LINEAR:
|
42
|
+
case FANN_LINEAR_PIECE:
|
43
|
+
case FANN_LINEAR_PIECE_SYMMETRIC:
|
44
|
+
return (fann_type) fann_linear_derive(steepness, value);
|
45
|
+
case FANN_SIGMOID:
|
46
|
+
case FANN_SIGMOID_STEPWISE:
|
47
|
+
value = fann_clip(value, 0.01f, 0.99f);
|
48
|
+
return (fann_type) fann_sigmoid_derive(steepness, value);
|
49
|
+
case FANN_SIGMOID_SYMMETRIC:
|
50
|
+
case FANN_SIGMOID_SYMMETRIC_STEPWISE:
|
51
|
+
value = fann_clip(value, -0.98f, 0.98f);
|
52
|
+
return (fann_type) fann_sigmoid_symmetric_derive(steepness, value);
|
53
|
+
case FANN_GAUSSIAN:
|
54
|
+
/* value = fann_clip(value, 0.01f, 0.99f); */
|
55
|
+
return (fann_type) fann_gaussian_derive(steepness, value, sum);
|
56
|
+
case FANN_GAUSSIAN_SYMMETRIC:
|
57
|
+
/* value = fann_clip(value, -0.98f, 0.98f); */
|
58
|
+
return (fann_type) fann_gaussian_symmetric_derive(steepness, value, sum);
|
59
|
+
case FANN_ELLIOT:
|
60
|
+
value = fann_clip(value, 0.01f, 0.99f);
|
61
|
+
return (fann_type) fann_elliot_derive(steepness, value, sum);
|
62
|
+
case FANN_ELLIOT_SYMMETRIC:
|
63
|
+
value = fann_clip(value, -0.98f, 0.98f);
|
64
|
+
return (fann_type) fann_elliot_symmetric_derive(steepness, value, sum);
|
65
|
+
case FANN_SIN_SYMMETRIC:
|
66
|
+
return (fann_type) fann_sin_symmetric_derive(steepness, sum);
|
67
|
+
case FANN_COS_SYMMETRIC:
|
68
|
+
return (fann_type) fann_cos_symmetric_derive(steepness, sum);
|
69
|
+
case FANN_SIN:
|
70
|
+
return (fann_type) fann_sin_derive(steepness, sum);
|
71
|
+
case FANN_COS:
|
72
|
+
return (fann_type) fann_cos_derive(steepness, sum);
|
73
|
+
case FANN_THRESHOLD:
|
74
|
+
fann_error(NULL, FANN_E_CANT_TRAIN_ACTIVATION);
|
75
|
+
}
|
76
|
+
return 0;
|
77
|
+
}
|
78
|
+
|
79
|
+
/* INTERNAL FUNCTION
|
80
|
+
Calculates the activation of a value, given an activation function
|
81
|
+
and a steepness
|
82
|
+
*/
|
83
|
+
fann_type fann_activation(struct fann * ann, unsigned int activation_function, fann_type steepness,
|
84
|
+
fann_type value)
|
85
|
+
{
|
86
|
+
value = fann_mult(steepness, value);
|
87
|
+
fann_activation_switch(activation_function, value, value);
|
88
|
+
return value;
|
89
|
+
}
|
90
|
+
|
91
|
+
/* Trains the network with the backpropagation algorithm.
|
92
|
+
*/
|
93
|
+
FANN_EXTERNAL void FANN_API fann_train(struct fann *ann, fann_type * input,
|
94
|
+
fann_type * desired_output)
|
95
|
+
{
|
96
|
+
fann_run(ann, input);
|
97
|
+
|
98
|
+
fann_compute_MSE(ann, desired_output);
|
99
|
+
|
100
|
+
fann_backpropagate_MSE(ann);
|
101
|
+
|
102
|
+
fann_update_weights(ann);
|
103
|
+
}
|
104
|
+
#endif
|
105
|
+
|
106
|
+
|
107
|
+
/* INTERNAL FUNCTION
|
108
|
+
Helper function to update the MSE value and return a diff which takes symmetric functions into account
|
109
|
+
*/
|
110
|
+
fann_type fann_update_MSE(struct fann *ann, struct fann_neuron* neuron, fann_type neuron_diff)
|
111
|
+
{
|
112
|
+
float neuron_diff2;
|
113
|
+
|
114
|
+
switch (neuron->activation_function)
|
115
|
+
{
|
116
|
+
case FANN_LINEAR_PIECE_SYMMETRIC:
|
117
|
+
case FANN_THRESHOLD_SYMMETRIC:
|
118
|
+
case FANN_SIGMOID_SYMMETRIC:
|
119
|
+
case FANN_SIGMOID_SYMMETRIC_STEPWISE:
|
120
|
+
case FANN_ELLIOT_SYMMETRIC:
|
121
|
+
case FANN_GAUSSIAN_SYMMETRIC:
|
122
|
+
case FANN_SIN_SYMMETRIC:
|
123
|
+
case FANN_COS_SYMMETRIC:
|
124
|
+
neuron_diff /= (fann_type)2.0;
|
125
|
+
break;
|
126
|
+
case FANN_THRESHOLD:
|
127
|
+
case FANN_LINEAR:
|
128
|
+
case FANN_SIGMOID:
|
129
|
+
case FANN_SIGMOID_STEPWISE:
|
130
|
+
case FANN_GAUSSIAN:
|
131
|
+
case FANN_GAUSSIAN_STEPWISE:
|
132
|
+
case FANN_ELLIOT:
|
133
|
+
case FANN_LINEAR_PIECE:
|
134
|
+
case FANN_SIN:
|
135
|
+
case FANN_COS:
|
136
|
+
break;
|
137
|
+
}
|
138
|
+
|
139
|
+
#ifdef FIXEDFANN
|
140
|
+
neuron_diff2 =
|
141
|
+
(neuron_diff / (float) ann->multiplier) * (neuron_diff / (float) ann->multiplier);
|
142
|
+
#else
|
143
|
+
neuron_diff2 = (float) (neuron_diff * neuron_diff);
|
144
|
+
#endif
|
145
|
+
|
146
|
+
ann->MSE_value += neuron_diff2;
|
147
|
+
|
148
|
+
/*printf("neuron_diff %f = (%f - %f)[/2], neuron_diff2=%f, sum=%f, MSE_value=%f, num_MSE=%d\n", neuron_diff, *desired_output, neuron_value, neuron_diff2, last_layer_begin->sum, ann->MSE_value, ann->num_MSE); */
|
149
|
+
if(fann_abs(neuron_diff) >= ann->bit_fail_limit)
|
150
|
+
{
|
151
|
+
ann->num_bit_fail++;
|
152
|
+
}
|
153
|
+
|
154
|
+
return neuron_diff;
|
155
|
+
}
|
156
|
+
|
157
|
+
/* Tests the network.
|
158
|
+
*/
|
159
|
+
FANN_EXTERNAL fann_type *FANN_API fann_test(struct fann *ann, fann_type * input,
|
160
|
+
fann_type * desired_output)
|
161
|
+
{
|
162
|
+
fann_type neuron_value;
|
163
|
+
fann_type *output_begin = fann_run(ann, input);
|
164
|
+
fann_type *output_it;
|
165
|
+
const fann_type *output_end = output_begin + ann->num_output;
|
166
|
+
fann_type neuron_diff;
|
167
|
+
struct fann_neuron *output_neuron = (ann->last_layer - 1)->first_neuron;
|
168
|
+
|
169
|
+
/* calculate the error */
|
170
|
+
for(output_it = output_begin; output_it != output_end; output_it++)
|
171
|
+
{
|
172
|
+
neuron_value = *output_it;
|
173
|
+
|
174
|
+
neuron_diff = (*desired_output - neuron_value);
|
175
|
+
|
176
|
+
neuron_diff = fann_update_MSE(ann, output_neuron, neuron_diff);
|
177
|
+
|
178
|
+
desired_output++;
|
179
|
+
output_neuron++;
|
180
|
+
|
181
|
+
ann->num_MSE++;
|
182
|
+
}
|
183
|
+
|
184
|
+
return output_begin;
|
185
|
+
}
|
186
|
+
|
187
|
+
/* get the mean square error.
|
188
|
+
*/
|
189
|
+
FANN_EXTERNAL float FANN_API fann_get_MSE(struct fann *ann)
|
190
|
+
{
|
191
|
+
if(ann->num_MSE)
|
192
|
+
{
|
193
|
+
return ann->MSE_value / (float) ann->num_MSE;
|
194
|
+
}
|
195
|
+
else
|
196
|
+
{
|
197
|
+
return 0;
|
198
|
+
}
|
199
|
+
}
|
200
|
+
|
201
|
+
FANN_EXTERNAL unsigned int FANN_API fann_get_bit_fail(struct fann *ann)
|
202
|
+
{
|
203
|
+
return ann->num_bit_fail;
|
204
|
+
}
|
205
|
+
|
206
|
+
/* reset the mean square error.
|
207
|
+
*/
|
208
|
+
FANN_EXTERNAL void FANN_API fann_reset_MSE(struct fann *ann)
|
209
|
+
{
|
210
|
+
/*printf("resetMSE %d %f\n", ann->num_MSE, ann->MSE_value);*/
|
211
|
+
ann->num_MSE = 0;
|
212
|
+
ann->MSE_value = 0;
|
213
|
+
ann->num_bit_fail = 0;
|
214
|
+
}
|
215
|
+
|
216
|
+
#ifndef FIXEDFANN
|
217
|
+
|
218
|
+
/* INTERNAL FUNCTION
|
219
|
+
compute the error at the network output
|
220
|
+
(usually, after forward propagation of a certain input vector, fann_run)
|
221
|
+
the error is a sum of squares for all the output units
|
222
|
+
also increments a counter because MSE is an average of such errors
|
223
|
+
|
224
|
+
After this train_errors in the output layer will be set to:
|
225
|
+
neuron_value_derived * (desired_output - neuron_value)
|
226
|
+
*/
|
227
|
+
void fann_compute_MSE(struct fann *ann, fann_type * desired_output)
|
228
|
+
{
|
229
|
+
fann_type neuron_value, neuron_diff, *error_it = 0, *error_begin = 0;
|
230
|
+
struct fann_neuron *last_layer_begin = (ann->last_layer - 1)->first_neuron;
|
231
|
+
const struct fann_neuron *last_layer_end = last_layer_begin + ann->num_output;
|
232
|
+
const struct fann_neuron *first_neuron = ann->first_layer->first_neuron;
|
233
|
+
|
234
|
+
/* if no room allocated for the error variabels, allocate it now */
|
235
|
+
if(ann->train_errors == NULL)
|
236
|
+
{
|
237
|
+
ann->train_errors = (fann_type *) calloc(ann->total_neurons, sizeof(fann_type));
|
238
|
+
if(ann->train_errors == NULL)
|
239
|
+
{
|
240
|
+
fann_error((struct fann_error *) ann, FANN_E_CANT_ALLOCATE_MEM);
|
241
|
+
return;
|
242
|
+
}
|
243
|
+
}
|
244
|
+
else
|
245
|
+
{
|
246
|
+
/* clear the error variabels */
|
247
|
+
memset(ann->train_errors, 0, (ann->total_neurons) * sizeof(fann_type));
|
248
|
+
}
|
249
|
+
error_begin = ann->train_errors;
|
250
|
+
|
251
|
+
#ifdef DEBUGTRAIN
|
252
|
+
printf("\ncalculate errors\n");
|
253
|
+
#endif
|
254
|
+
/* calculate the error and place it in the output layer */
|
255
|
+
error_it = error_begin + (last_layer_begin - first_neuron);
|
256
|
+
|
257
|
+
for(; last_layer_begin != last_layer_end; last_layer_begin++)
|
258
|
+
{
|
259
|
+
neuron_value = last_layer_begin->value;
|
260
|
+
neuron_diff = *desired_output - neuron_value;
|
261
|
+
|
262
|
+
neuron_diff = fann_update_MSE(ann, last_layer_begin, neuron_diff);
|
263
|
+
|
264
|
+
if(ann->train_error_function)
|
265
|
+
{ /* TODO make switch when more functions */
|
266
|
+
if(neuron_diff < -.9999999)
|
267
|
+
neuron_diff = -17.0;
|
268
|
+
else if(neuron_diff > .9999999)
|
269
|
+
neuron_diff = 17.0;
|
270
|
+
else
|
271
|
+
neuron_diff = (fann_type) log((1.0 + neuron_diff) / (1.0 - neuron_diff));
|
272
|
+
}
|
273
|
+
|
274
|
+
*error_it = fann_activation_derived(last_layer_begin->activation_function,
|
275
|
+
last_layer_begin->activation_steepness, neuron_value,
|
276
|
+
last_layer_begin->sum) * neuron_diff;
|
277
|
+
|
278
|
+
desired_output++;
|
279
|
+
error_it++;
|
280
|
+
|
281
|
+
ann->num_MSE++;
|
282
|
+
}
|
283
|
+
}
|
284
|
+
|
285
|
+
/* INTERNAL FUNCTION
|
286
|
+
Propagate the error backwards from the output layer.
|
287
|
+
|
288
|
+
After this the train_errors in the hidden layers will be:
|
289
|
+
neuron_value_derived * sum(outgoing_weights * connected_neuron)
|
290
|
+
*/
|
291
|
+
void fann_backpropagate_MSE(struct fann *ann)
|
292
|
+
{
|
293
|
+
fann_type tmp_error;
|
294
|
+
unsigned int i;
|
295
|
+
struct fann_layer *layer_it;
|
296
|
+
struct fann_neuron *neuron_it, *last_neuron;
|
297
|
+
struct fann_neuron **connections;
|
298
|
+
|
299
|
+
fann_type *error_begin = ann->train_errors;
|
300
|
+
fann_type *error_prev_layer;
|
301
|
+
fann_type *weights;
|
302
|
+
const struct fann_neuron *first_neuron = ann->first_layer->first_neuron;
|
303
|
+
const struct fann_layer *second_layer = ann->first_layer + 1;
|
304
|
+
struct fann_layer *last_layer = ann->last_layer;
|
305
|
+
|
306
|
+
/* go through all the layers, from last to first.
|
307
|
+
* And propagate the error backwards */
|
308
|
+
for(layer_it = last_layer - 1; layer_it > second_layer; --layer_it)
|
309
|
+
{
|
310
|
+
last_neuron = layer_it->last_neuron;
|
311
|
+
|
312
|
+
/* for each connection in this layer, propagate the error backwards */
|
313
|
+
if(ann->connection_rate >= 1)
|
314
|
+
{
|
315
|
+
if(ann->network_type == FANN_NETTYPE_LAYER)
|
316
|
+
{
|
317
|
+
error_prev_layer = error_begin + ((layer_it - 1)->first_neuron - first_neuron);
|
318
|
+
}
|
319
|
+
else
|
320
|
+
{
|
321
|
+
error_prev_layer = error_begin;
|
322
|
+
}
|
323
|
+
|
324
|
+
for(neuron_it = layer_it->first_neuron; neuron_it != last_neuron; neuron_it++)
|
325
|
+
{
|
326
|
+
|
327
|
+
tmp_error = error_begin[neuron_it - first_neuron];
|
328
|
+
weights = ann->weights + neuron_it->first_con;
|
329
|
+
for(i = neuron_it->last_con - neuron_it->first_con; i--;)
|
330
|
+
{
|
331
|
+
/*printf("i = %d\n", i);
|
332
|
+
* printf("error_prev_layer[%d] = %f\n", i, error_prev_layer[i]);
|
333
|
+
* printf("weights[%d] = %f\n", i, weights[i]); */
|
334
|
+
error_prev_layer[i] += tmp_error * weights[i];
|
335
|
+
}
|
336
|
+
}
|
337
|
+
}
|
338
|
+
else
|
339
|
+
{
|
340
|
+
for(neuron_it = layer_it->first_neuron; neuron_it != last_neuron; neuron_it++)
|
341
|
+
{
|
342
|
+
|
343
|
+
tmp_error = error_begin[neuron_it - first_neuron];
|
344
|
+
weights = ann->weights + neuron_it->first_con;
|
345
|
+
connections = ann->connections + neuron_it->first_con;
|
346
|
+
for(i = neuron_it->last_con - neuron_it->first_con; i--;)
|
347
|
+
{
|
348
|
+
error_begin[connections[i] - first_neuron] += tmp_error * weights[i];
|
349
|
+
}
|
350
|
+
}
|
351
|
+
}
|
352
|
+
|
353
|
+
/* then calculate the actual errors in the previous layer */
|
354
|
+
error_prev_layer = error_begin + ((layer_it - 1)->first_neuron - first_neuron);
|
355
|
+
last_neuron = (layer_it - 1)->last_neuron;
|
356
|
+
|
357
|
+
for(neuron_it = (layer_it - 1)->first_neuron; neuron_it != last_neuron; neuron_it++)
|
358
|
+
{
|
359
|
+
*error_prev_layer *= fann_activation_derived(neuron_it->activation_function,
|
360
|
+
neuron_it->activation_steepness, neuron_it->value, neuron_it->sum);
|
361
|
+
error_prev_layer++;
|
362
|
+
}
|
363
|
+
|
364
|
+
}
|
365
|
+
}
|
366
|
+
|
367
|
+
/* INTERNAL FUNCTION
|
368
|
+
Update weights for incremental training
|
369
|
+
*/
|
370
|
+
void fann_update_weights(struct fann *ann)
|
371
|
+
{
|
372
|
+
struct fann_neuron *neuron_it, *last_neuron, *prev_neurons;
|
373
|
+
fann_type tmp_error, delta_w, *weights;
|
374
|
+
struct fann_layer *layer_it;
|
375
|
+
unsigned int i;
|
376
|
+
unsigned int num_connections;
|
377
|
+
|
378
|
+
/* store some variabels local for fast access */
|
379
|
+
const float learning_rate = ann->learning_rate;
|
380
|
+
const float learning_momentum = ann->learning_momentum;
|
381
|
+
struct fann_neuron *first_neuron = ann->first_layer->first_neuron;
|
382
|
+
struct fann_layer *first_layer = ann->first_layer;
|
383
|
+
const struct fann_layer *last_layer = ann->last_layer;
|
384
|
+
fann_type *error_begin = ann->train_errors;
|
385
|
+
fann_type *deltas_begin, *weights_deltas;
|
386
|
+
|
387
|
+
/* if no room allocated for the deltas, allocate it now */
|
388
|
+
if(ann->prev_weights_deltas == NULL)
|
389
|
+
{
|
390
|
+
ann->prev_weights_deltas =
|
391
|
+
(fann_type *) calloc(ann->total_connections_allocated, sizeof(fann_type));
|
392
|
+
if(ann->prev_weights_deltas == NULL)
|
393
|
+
{
|
394
|
+
fann_error((struct fann_error *) ann, FANN_E_CANT_ALLOCATE_MEM);
|
395
|
+
return;
|
396
|
+
}
|
397
|
+
}
|
398
|
+
|
399
|
+
#ifdef DEBUGTRAIN
|
400
|
+
printf("\nupdate weights\n");
|
401
|
+
#endif
|
402
|
+
deltas_begin = ann->prev_weights_deltas;
|
403
|
+
prev_neurons = first_neuron;
|
404
|
+
for(layer_it = (first_layer + 1); layer_it != last_layer; layer_it++)
|
405
|
+
{
|
406
|
+
#ifdef DEBUGTRAIN
|
407
|
+
printf("layer[%d]\n", layer_it - first_layer);
|
408
|
+
#endif
|
409
|
+
last_neuron = layer_it->last_neuron;
|
410
|
+
if(ann->connection_rate >= 1)
|
411
|
+
{
|
412
|
+
if(ann->network_type == FANN_NETTYPE_LAYER)
|
413
|
+
{
|
414
|
+
prev_neurons = (layer_it - 1)->first_neuron;
|
415
|
+
}
|
416
|
+
for(neuron_it = layer_it->first_neuron; neuron_it != last_neuron; neuron_it++)
|
417
|
+
{
|
418
|
+
tmp_error = error_begin[neuron_it - first_neuron] * learning_rate;
|
419
|
+
num_connections = neuron_it->last_con - neuron_it->first_con;
|
420
|
+
weights = ann->weights + neuron_it->first_con;
|
421
|
+
weights_deltas = deltas_begin + neuron_it->first_con;
|
422
|
+
for(i = 0; i != num_connections; i++)
|
423
|
+
{
|
424
|
+
delta_w = tmp_error * prev_neurons[i].value + learning_momentum * weights_deltas[i];
|
425
|
+
weights[i] += delta_w ;
|
426
|
+
weights_deltas[i] = delta_w;
|
427
|
+
}
|
428
|
+
}
|
429
|
+
}
|
430
|
+
else
|
431
|
+
{
|
432
|
+
for(neuron_it = layer_it->first_neuron; neuron_it != last_neuron; neuron_it++)
|
433
|
+
{
|
434
|
+
tmp_error = error_begin[neuron_it - first_neuron] * learning_rate;
|
435
|
+
num_connections = neuron_it->last_con - neuron_it->first_con;
|
436
|
+
weights = ann->weights + neuron_it->first_con;
|
437
|
+
weights_deltas = deltas_begin + neuron_it->first_con;
|
438
|
+
for(i = 0; i != num_connections; i++)
|
439
|
+
{
|
440
|
+
delta_w = tmp_error * prev_neurons[i].value + learning_momentum * weights_deltas[i];
|
441
|
+
weights[i] += delta_w;
|
442
|
+
weights_deltas[i] = delta_w;
|
443
|
+
}
|
444
|
+
}
|
445
|
+
}
|
446
|
+
}
|
447
|
+
}
|
448
|
+
|
449
|
+
/* INTERNAL FUNCTION
|
450
|
+
Update slopes for batch training
|
451
|
+
layer_begin = ann->first_layer+1 and layer_end = ann->last_layer-1
|
452
|
+
will update all slopes.
|
453
|
+
|
454
|
+
*/
|
455
|
+
void fann_update_slopes_batch(struct fann *ann, struct fann_layer *layer_begin,
|
456
|
+
struct fann_layer *layer_end)
|
457
|
+
{
|
458
|
+
struct fann_neuron *neuron_it, *last_neuron, *prev_neurons, **connections;
|
459
|
+
fann_type tmp_error;
|
460
|
+
unsigned int i, num_connections;
|
461
|
+
|
462
|
+
/* store some variabels local for fast access */
|
463
|
+
struct fann_neuron *first_neuron = ann->first_layer->first_neuron;
|
464
|
+
fann_type *error_begin = ann->train_errors;
|
465
|
+
fann_type *slope_begin, *neuron_slope;
|
466
|
+
|
467
|
+
/* if no room allocated for the slope variabels, allocate it now */
|
468
|
+
if(ann->train_slopes == NULL)
|
469
|
+
{
|
470
|
+
ann->train_slopes =
|
471
|
+
(fann_type *) calloc(ann->total_connections_allocated, sizeof(fann_type));
|
472
|
+
if(ann->train_slopes == NULL)
|
473
|
+
{
|
474
|
+
fann_error((struct fann_error *) ann, FANN_E_CANT_ALLOCATE_MEM);
|
475
|
+
return;
|
476
|
+
}
|
477
|
+
}
|
478
|
+
|
479
|
+
if(layer_begin == NULL)
|
480
|
+
{
|
481
|
+
layer_begin = ann->first_layer + 1;
|
482
|
+
}
|
483
|
+
|
484
|
+
if(layer_end == NULL)
|
485
|
+
{
|
486
|
+
layer_end = ann->last_layer - 1;
|
487
|
+
}
|
488
|
+
|
489
|
+
slope_begin = ann->train_slopes;
|
490
|
+
|
491
|
+
#ifdef DEBUGTRAIN
|
492
|
+
printf("\nupdate slopes\n");
|
493
|
+
#endif
|
494
|
+
|
495
|
+
prev_neurons = first_neuron;
|
496
|
+
|
497
|
+
for(; layer_begin <= layer_end; layer_begin++)
|
498
|
+
{
|
499
|
+
#ifdef DEBUGTRAIN
|
500
|
+
printf("layer[%d]\n", layer_begin - ann->first_layer);
|
501
|
+
#endif
|
502
|
+
last_neuron = layer_begin->last_neuron;
|
503
|
+
if(ann->connection_rate >= 1)
|
504
|
+
{
|
505
|
+
if(ann->network_type == FANN_NETTYPE_LAYER)
|
506
|
+
{
|
507
|
+
prev_neurons = (layer_begin - 1)->first_neuron;
|
508
|
+
}
|
509
|
+
|
510
|
+
for(neuron_it = layer_begin->first_neuron; neuron_it != last_neuron; neuron_it++)
|
511
|
+
{
|
512
|
+
tmp_error = error_begin[neuron_it - first_neuron];
|
513
|
+
neuron_slope = slope_begin + neuron_it->first_con;
|
514
|
+
num_connections = neuron_it->last_con - neuron_it->first_con;
|
515
|
+
for(i = 0; i != num_connections; i++)
|
516
|
+
{
|
517
|
+
neuron_slope[i] += tmp_error * prev_neurons[i].value;
|
518
|
+
}
|
519
|
+
}
|
520
|
+
}
|
521
|
+
else
|
522
|
+
{
|
523
|
+
for(neuron_it = layer_begin->first_neuron; neuron_it != last_neuron; neuron_it++)
|
524
|
+
{
|
525
|
+
tmp_error = error_begin[neuron_it - first_neuron];
|
526
|
+
neuron_slope = slope_begin + neuron_it->first_con;
|
527
|
+
num_connections = neuron_it->last_con - neuron_it->first_con;
|
528
|
+
connections = ann->connections + neuron_it->first_con;
|
529
|
+
for(i = 0; i != num_connections; i++)
|
530
|
+
{
|
531
|
+
neuron_slope[i] += tmp_error * connections[i]->value;
|
532
|
+
}
|
533
|
+
}
|
534
|
+
}
|
535
|
+
}
|
536
|
+
}
|
537
|
+
|
538
|
+
/* INTERNAL FUNCTION
|
539
|
+
Clears arrays used for training before a new training session.
|
540
|
+
Also creates the arrays that do not exist yet.
|
541
|
+
*/
|
542
|
+
void fann_clear_train_arrays(struct fann *ann)
|
543
|
+
{
|
544
|
+
unsigned int i;
|
545
|
+
fann_type delta_zero;
|
546
|
+
|
547
|
+
/* if no room allocated for the slope variabels, allocate it now
|
548
|
+
* (calloc clears mem) */
|
549
|
+
if(ann->train_slopes == NULL)
|
550
|
+
{
|
551
|
+
ann->train_slopes =
|
552
|
+
(fann_type *) calloc(ann->total_connections_allocated, sizeof(fann_type));
|
553
|
+
if(ann->train_slopes == NULL)
|
554
|
+
{
|
555
|
+
fann_error((struct fann_error *) ann, FANN_E_CANT_ALLOCATE_MEM);
|
556
|
+
return;
|
557
|
+
}
|
558
|
+
}
|
559
|
+
else
|
560
|
+
{
|
561
|
+
memset(ann->train_slopes, 0, (ann->total_connections_allocated) * sizeof(fann_type));
|
562
|
+
}
|
563
|
+
|
564
|
+
/* if no room allocated for the variabels, allocate it now */
|
565
|
+
if(ann->prev_steps == NULL)
|
566
|
+
{
|
567
|
+
ann->prev_steps = (fann_type *) malloc(ann->total_connections_allocated * sizeof(fann_type));
|
568
|
+
if(ann->prev_steps == NULL)
|
569
|
+
{
|
570
|
+
fann_error((struct fann_error *) ann, FANN_E_CANT_ALLOCATE_MEM);
|
571
|
+
return;
|
572
|
+
}
|
573
|
+
}
|
574
|
+
|
575
|
+
if(ann->training_algorithm == FANN_TRAIN_RPROP)
|
576
|
+
{
|
577
|
+
delta_zero = ann->rprop_delta_zero;
|
578
|
+
|
579
|
+
for(i = 0; i < ann->total_connections_allocated; i++)
|
580
|
+
ann->prev_steps[i] = delta_zero;
|
581
|
+
}
|
582
|
+
else
|
583
|
+
{
|
584
|
+
memset(ann->prev_steps, 0, (ann->total_connections_allocated) * sizeof(fann_type));
|
585
|
+
}
|
586
|
+
|
587
|
+
/* if no room allocated for the variabels, allocate it now */
|
588
|
+
if(ann->prev_train_slopes == NULL)
|
589
|
+
{
|
590
|
+
ann->prev_train_slopes =
|
591
|
+
(fann_type *) calloc(ann->total_connections_allocated, sizeof(fann_type));
|
592
|
+
if(ann->prev_train_slopes == NULL)
|
593
|
+
{
|
594
|
+
fann_error((struct fann_error *) ann, FANN_E_CANT_ALLOCATE_MEM);
|
595
|
+
return;
|
596
|
+
}
|
597
|
+
}
|
598
|
+
else
|
599
|
+
{
|
600
|
+
memset(ann->prev_train_slopes, 0, (ann->total_connections_allocated) * sizeof(fann_type));
|
601
|
+
}
|
602
|
+
}
|
603
|
+
|
604
|
+
/* INTERNAL FUNCTION
|
605
|
+
Update weights for batch training
|
606
|
+
*/
|
607
|
+
void fann_update_weights_batch(struct fann *ann, unsigned int num_data, unsigned int first_weight,
|
608
|
+
unsigned int past_end)
|
609
|
+
{
|
610
|
+
fann_type *train_slopes = ann->train_slopes;
|
611
|
+
fann_type *weights = ann->weights;
|
612
|
+
const float epsilon = ann->learning_rate / num_data;
|
613
|
+
unsigned int i = first_weight;
|
614
|
+
|
615
|
+
for(; i != past_end; i++)
|
616
|
+
{
|
617
|
+
weights[i] += train_slopes[i] * epsilon;
|
618
|
+
train_slopes[i] = 0.0;
|
619
|
+
}
|
620
|
+
}
|
621
|
+
|
622
|
+
/* INTERNAL FUNCTION
|
623
|
+
The quickprop training algorithm
|
624
|
+
*/
|
625
|
+
void fann_update_weights_quickprop(struct fann *ann, unsigned int num_data,
|
626
|
+
unsigned int first_weight, unsigned int past_end)
|
627
|
+
{
|
628
|
+
fann_type *train_slopes = ann->train_slopes;
|
629
|
+
fann_type *weights = ann->weights;
|
630
|
+
fann_type *prev_steps = ann->prev_steps;
|
631
|
+
fann_type *prev_train_slopes = ann->prev_train_slopes;
|
632
|
+
|
633
|
+
fann_type w, prev_step, slope, prev_slope, next_step;
|
634
|
+
|
635
|
+
float epsilon = ann->learning_rate / num_data;
|
636
|
+
float decay = ann->quickprop_decay; /*-0.0001;*/
|
637
|
+
float mu = ann->quickprop_mu; /*1.75; */
|
638
|
+
float shrink_factor = (float) (mu / (1.0 + mu));
|
639
|
+
|
640
|
+
unsigned int i = first_weight;
|
641
|
+
|
642
|
+
for(; i != past_end; i++)
|
643
|
+
{
|
644
|
+
w = weights[i];
|
645
|
+
prev_step = prev_steps[i];
|
646
|
+
slope = train_slopes[i] + decay * w;
|
647
|
+
prev_slope = prev_train_slopes[i];
|
648
|
+
next_step = 0.0;
|
649
|
+
|
650
|
+
/* The step must always be in direction opposite to the slope. */
|
651
|
+
if(prev_step > 0.001)
|
652
|
+
{
|
653
|
+
/* If last step was positive... */
|
654
|
+
if(slope > 0.0) /* Add in linear term if current slope is still positive. */
|
655
|
+
next_step += epsilon * slope;
|
656
|
+
|
657
|
+
/*If current slope is close to or larger than prev slope... */
|
658
|
+
if(slope > (shrink_factor * prev_slope))
|
659
|
+
next_step += mu * prev_step; /* Take maximum size negative step. */
|
660
|
+
else
|
661
|
+
next_step += prev_step * slope / (prev_slope - slope); /* Else, use quadratic estimate. */
|
662
|
+
}
|
663
|
+
else if(prev_step < -0.001)
|
664
|
+
{
|
665
|
+
/* If last step was negative... */
|
666
|
+
if(slope < 0.0) /* Add in linear term if current slope is still negative. */
|
667
|
+
next_step += epsilon * slope;
|
668
|
+
|
669
|
+
/* If current slope is close to or more neg than prev slope... */
|
670
|
+
if(slope < (shrink_factor * prev_slope))
|
671
|
+
next_step += mu * prev_step; /* Take maximum size negative step. */
|
672
|
+
else
|
673
|
+
next_step += prev_step * slope / (prev_slope - slope); /* Else, use quadratic estimate. */
|
674
|
+
}
|
675
|
+
else /* Last step was zero, so use only linear term. */
|
676
|
+
next_step += epsilon * slope;
|
677
|
+
|
678
|
+
/*
|
679
|
+
if(next_step > 1000 || next_step < -1000)
|
680
|
+
{
|
681
|
+
printf("quickprop[%d] weight=%f, slope=%f, prev_slope=%f, next_step=%f, prev_step=%f\n",
|
682
|
+
i, weights[i], slope, prev_slope, next_step, prev_step);
|
683
|
+
|
684
|
+
if(next_step > 1000)
|
685
|
+
next_step = 1000;
|
686
|
+
else
|
687
|
+
next_step = -1000;
|
688
|
+
}
|
689
|
+
*/
|
690
|
+
|
691
|
+
/* update global data arrays */
|
692
|
+
prev_steps[i] = next_step;
|
693
|
+
|
694
|
+
w += next_step;
|
695
|
+
|
696
|
+
if(w > 1500)
|
697
|
+
weights[i] = 1500;
|
698
|
+
else if(w < -1500)
|
699
|
+
weights[i] = -1500;
|
700
|
+
else
|
701
|
+
weights[i] = w;
|
702
|
+
|
703
|
+
/*weights[i] = w;*/
|
704
|
+
|
705
|
+
prev_train_slopes[i] = slope;
|
706
|
+
train_slopes[i] = 0.0;
|
707
|
+
}
|
708
|
+
}
|
709
|
+
|
710
|
+
/* INTERNAL FUNCTION
|
711
|
+
The iRprop- algorithm
|
712
|
+
*/
|
713
|
+
void fann_update_weights_irpropm(struct fann *ann, unsigned int first_weight, unsigned int past_end)
|
714
|
+
{
|
715
|
+
fann_type *train_slopes = ann->train_slopes;
|
716
|
+
fann_type *weights = ann->weights;
|
717
|
+
fann_type *prev_steps = ann->prev_steps;
|
718
|
+
fann_type *prev_train_slopes = ann->prev_train_slopes;
|
719
|
+
|
720
|
+
fann_type prev_step, slope, prev_slope, next_step, same_sign;
|
721
|
+
|
722
|
+
float increase_factor = ann->rprop_increase_factor; /*1.2; */
|
723
|
+
float decrease_factor = ann->rprop_decrease_factor; /*0.5; */
|
724
|
+
float delta_min = ann->rprop_delta_min; /*0.0; */
|
725
|
+
float delta_max = ann->rprop_delta_max; /*50.0; */
|
726
|
+
|
727
|
+
unsigned int i = first_weight;
|
728
|
+
|
729
|
+
for(; i != past_end; i++)
|
730
|
+
{
|
731
|
+
prev_step = fann_max(prev_steps[i], (fann_type) 0.0001); /* prev_step may not be zero because then the training will stop */
|
732
|
+
slope = train_slopes[i];
|
733
|
+
prev_slope = prev_train_slopes[i];
|
734
|
+
|
735
|
+
same_sign = prev_slope * slope;
|
736
|
+
|
737
|
+
if(same_sign >= 0.0)
|
738
|
+
next_step = fann_min(prev_step * increase_factor, delta_max);
|
739
|
+
else
|
740
|
+
{
|
741
|
+
next_step = fann_max(prev_step * decrease_factor, delta_min);
|
742
|
+
slope = 0;
|
743
|
+
}
|
744
|
+
|
745
|
+
if(slope < 0)
|
746
|
+
{
|
747
|
+
weights[i] -= next_step;
|
748
|
+
if(weights[i] < -1500)
|
749
|
+
weights[i] = -1500;
|
750
|
+
}
|
751
|
+
else
|
752
|
+
{
|
753
|
+
weights[i] += next_step;
|
754
|
+
if(weights[i] > 1500)
|
755
|
+
weights[i] = 1500;
|
756
|
+
}
|
757
|
+
|
758
|
+
/*if(i == 2){
|
759
|
+
* printf("weight=%f, slope=%f, next_step=%f, prev_step=%f\n", weights[i], slope, next_step, prev_step);
|
760
|
+
* } */
|
761
|
+
|
762
|
+
/* update global data arrays */
|
763
|
+
prev_steps[i] = next_step;
|
764
|
+
prev_train_slopes[i] = slope;
|
765
|
+
train_slopes[i] = 0.0;
|
766
|
+
}
|
767
|
+
}
|
768
|
+
|
769
|
+
/* INTERNAL FUNCTION
|
770
|
+
The SARprop- algorithm
|
771
|
+
*/
|
772
|
+
void fann_update_weights_sarprop(struct fann *ann, unsigned int epoch, unsigned int first_weight, unsigned int past_end)
|
773
|
+
{
|
774
|
+
fann_type *train_slopes = ann->train_slopes;
|
775
|
+
fann_type *weights = ann->weights;
|
776
|
+
fann_type *prev_steps = ann->prev_steps;
|
777
|
+
fann_type *prev_train_slopes = ann->prev_train_slopes;
|
778
|
+
|
779
|
+
fann_type prev_step, slope, prev_slope, next_step = 0, same_sign;
|
780
|
+
|
781
|
+
/* These should be set from variables */
|
782
|
+
float increase_factor = ann->rprop_increase_factor; /*1.2; */
|
783
|
+
float decrease_factor = ann->rprop_decrease_factor; /*0.5; */
|
784
|
+
/* TODO: why is delta_min 0.0 in iRprop? SARPROP uses 1x10^-6 (Braun and Riedmiller, 1993) */
|
785
|
+
float delta_min = 0.000001f;
|
786
|
+
float delta_max = ann->rprop_delta_max; /*50.0; */
|
787
|
+
float weight_decay_shift = ann->sarprop_weight_decay_shift; /* ld 0.01 = -6.644 */
|
788
|
+
float step_error_threshold_factor = ann->sarprop_step_error_threshold_factor; /* 0.1 */
|
789
|
+
float step_error_shift = ann->sarprop_step_error_shift; /* ld 3 = 1.585 */
|
790
|
+
float T = ann->sarprop_temperature;
|
791
|
+
float MSE = fann_get_MSE(ann);
|
792
|
+
float RMSE = (float)sqrt(MSE);
|
793
|
+
|
794
|
+
unsigned int i = first_weight;
|
795
|
+
|
796
|
+
|
797
|
+
/* for all weights; TODO: are biases included? */
|
798
|
+
for(; i != past_end; i++)
|
799
|
+
{
|
800
|
+
/* TODO: confirm whether 1x10^-6 == delta_min is really better */
|
801
|
+
prev_step = fann_max(prev_steps[i], (fann_type) 0.000001); /* prev_step may not be zero because then the training will stop */
|
802
|
+
/* calculate SARPROP slope; TODO: better as new error function? (see SARPROP paper)*/
|
803
|
+
slope = -train_slopes[i] - weights[i] * (fann_type)fann_exp2(-T * epoch + weight_decay_shift);
|
804
|
+
|
805
|
+
/* TODO: is prev_train_slopes[i] 0.0 in the beginning? */
|
806
|
+
prev_slope = prev_train_slopes[i];
|
807
|
+
|
808
|
+
same_sign = prev_slope * slope;
|
809
|
+
|
810
|
+
if(same_sign > 0.0)
|
811
|
+
{
|
812
|
+
next_step = fann_min(prev_step * increase_factor, delta_max);
|
813
|
+
/* TODO: are the signs inverted? see differences between SARPROP paper and iRprop */
|
814
|
+
if (slope < 0.0)
|
815
|
+
weights[i] += next_step;
|
816
|
+
else
|
817
|
+
weights[i] -= next_step;
|
818
|
+
}
|
819
|
+
else if(same_sign < 0.0)
|
820
|
+
{
|
821
|
+
if(prev_step < step_error_threshold_factor * MSE)
|
822
|
+
next_step = prev_step * decrease_factor + (float)rand() / RAND_MAX * RMSE * (fann_type)fann_exp2(-T * epoch + step_error_shift);
|
823
|
+
else
|
824
|
+
next_step = fann_max(prev_step * decrease_factor, delta_min);
|
825
|
+
|
826
|
+
slope = 0.0;
|
827
|
+
}
|
828
|
+
else
|
829
|
+
{
|
830
|
+
if(slope < 0.0)
|
831
|
+
weights[i] += prev_step;
|
832
|
+
else
|
833
|
+
weights[i] -= prev_step;
|
834
|
+
}
|
835
|
+
|
836
|
+
|
837
|
+
/*if(i == 2){
|
838
|
+
* printf("weight=%f, slope=%f, next_step=%f, prev_step=%f\n", weights[i], slope, next_step, prev_step);
|
839
|
+
* } */
|
840
|
+
|
841
|
+
/* update global data arrays */
|
842
|
+
prev_steps[i] = next_step;
|
843
|
+
prev_train_slopes[i] = slope;
|
844
|
+
train_slopes[i] = 0.0;
|
845
|
+
}
|
846
|
+
}
|
847
|
+
|
848
|
+
#endif
|
849
|
+
|
850
|
+
FANN_GET_SET(enum fann_train_enum, training_algorithm)
|
851
|
+
FANN_GET_SET(float, learning_rate)
|
852
|
+
|
853
|
+
FANN_EXTERNAL void FANN_API fann_set_activation_function_hidden(struct fann *ann,
|
854
|
+
enum fann_activationfunc_enum activation_function)
|
855
|
+
{
|
856
|
+
struct fann_neuron *last_neuron, *neuron_it;
|
857
|
+
struct fann_layer *layer_it;
|
858
|
+
struct fann_layer *last_layer = ann->last_layer - 1; /* -1 to not update the output layer */
|
859
|
+
|
860
|
+
for(layer_it = ann->first_layer + 1; layer_it != last_layer; layer_it++)
|
861
|
+
{
|
862
|
+
last_neuron = layer_it->last_neuron;
|
863
|
+
for(neuron_it = layer_it->first_neuron; neuron_it != last_neuron; neuron_it++)
|
864
|
+
{
|
865
|
+
neuron_it->activation_function = activation_function;
|
866
|
+
}
|
867
|
+
}
|
868
|
+
}
|
869
|
+
|
870
|
+
FANN_EXTERNAL struct fann_layer* FANN_API fann_get_layer(struct fann *ann, int layer)
|
871
|
+
{
|
872
|
+
if(layer <= 0 || layer >= (ann->last_layer - ann->first_layer))
|
873
|
+
{
|
874
|
+
fann_error((struct fann_error *) ann, FANN_E_INDEX_OUT_OF_BOUND, layer);
|
875
|
+
return NULL;
|
876
|
+
}
|
877
|
+
|
878
|
+
return ann->first_layer + layer;
|
879
|
+
}
|
880
|
+
|
881
|
+
FANN_EXTERNAL struct fann_neuron* FANN_API fann_get_neuron_layer(struct fann *ann, struct fann_layer* layer, int neuron)
|
882
|
+
{
|
883
|
+
if(neuron >= (layer->last_neuron - layer->first_neuron))
|
884
|
+
{
|
885
|
+
fann_error((struct fann_error *) ann, FANN_E_INDEX_OUT_OF_BOUND, neuron);
|
886
|
+
return NULL;
|
887
|
+
}
|
888
|
+
|
889
|
+
return layer->first_neuron + neuron;
|
890
|
+
}
|
891
|
+
|
892
|
+
FANN_EXTERNAL struct fann_neuron* FANN_API fann_get_neuron(struct fann *ann, unsigned int layer, int neuron)
|
893
|
+
{
|
894
|
+
struct fann_layer *layer_it = fann_get_layer(ann, layer);
|
895
|
+
if(layer_it == NULL)
|
896
|
+
return NULL;
|
897
|
+
return fann_get_neuron_layer(ann, layer_it, neuron);
|
898
|
+
}
|
899
|
+
|
900
|
+
FANN_EXTERNAL enum fann_activationfunc_enum FANN_API
|
901
|
+
fann_get_activation_function(struct fann *ann, int layer, int neuron)
|
902
|
+
{
|
903
|
+
struct fann_neuron* neuron_it = fann_get_neuron(ann, layer, neuron);
|
904
|
+
if (neuron_it == NULL)
|
905
|
+
{
|
906
|
+
return (enum fann_activationfunc_enum)-1; /* layer or neuron out of bounds */
|
907
|
+
}
|
908
|
+
else
|
909
|
+
{
|
910
|
+
return neuron_it->activation_function;
|
911
|
+
}
|
912
|
+
}
|
913
|
+
|
914
|
+
FANN_EXTERNAL void FANN_API fann_set_activation_function(struct fann *ann,
|
915
|
+
enum fann_activationfunc_enum
|
916
|
+
activation_function,
|
917
|
+
int layer,
|
918
|
+
int neuron)
|
919
|
+
{
|
920
|
+
struct fann_neuron* neuron_it = fann_get_neuron(ann, layer, neuron);
|
921
|
+
if(neuron_it == NULL)
|
922
|
+
return;
|
923
|
+
|
924
|
+
neuron_it->activation_function = activation_function;
|
925
|
+
}
|
926
|
+
|
927
|
+
FANN_EXTERNAL void FANN_API fann_set_activation_function_layer(struct fann *ann,
|
928
|
+
enum fann_activationfunc_enum
|
929
|
+
activation_function,
|
930
|
+
int layer)
|
931
|
+
{
|
932
|
+
struct fann_neuron *last_neuron, *neuron_it;
|
933
|
+
struct fann_layer *layer_it = fann_get_layer(ann, layer);
|
934
|
+
|
935
|
+
if(layer_it == NULL)
|
936
|
+
return;
|
937
|
+
|
938
|
+
last_neuron = layer_it->last_neuron;
|
939
|
+
for(neuron_it = layer_it->first_neuron; neuron_it != last_neuron; neuron_it++)
|
940
|
+
{
|
941
|
+
neuron_it->activation_function = activation_function;
|
942
|
+
}
|
943
|
+
}
|
944
|
+
|
945
|
+
|
946
|
+
FANN_EXTERNAL void FANN_API fann_set_activation_function_output(struct fann *ann,
|
947
|
+
enum fann_activationfunc_enum activation_function)
|
948
|
+
{
|
949
|
+
struct fann_neuron *last_neuron, *neuron_it;
|
950
|
+
struct fann_layer *last_layer = ann->last_layer - 1;
|
951
|
+
|
952
|
+
last_neuron = last_layer->last_neuron;
|
953
|
+
for(neuron_it = last_layer->first_neuron; neuron_it != last_neuron; neuron_it++)
|
954
|
+
{
|
955
|
+
neuron_it->activation_function = activation_function;
|
956
|
+
}
|
957
|
+
}
|
958
|
+
|
959
|
+
FANN_EXTERNAL void FANN_API fann_set_activation_steepness_hidden(struct fann *ann,
|
960
|
+
fann_type steepness)
|
961
|
+
{
|
962
|
+
struct fann_neuron *last_neuron, *neuron_it;
|
963
|
+
struct fann_layer *layer_it;
|
964
|
+
struct fann_layer *last_layer = ann->last_layer - 1; /* -1 to not update the output layer */
|
965
|
+
|
966
|
+
for(layer_it = ann->first_layer + 1; layer_it != last_layer; layer_it++)
|
967
|
+
{
|
968
|
+
last_neuron = layer_it->last_neuron;
|
969
|
+
for(neuron_it = layer_it->first_neuron; neuron_it != last_neuron; neuron_it++)
|
970
|
+
{
|
971
|
+
neuron_it->activation_steepness = steepness;
|
972
|
+
}
|
973
|
+
}
|
974
|
+
}
|
975
|
+
|
976
|
+
FANN_EXTERNAL fann_type FANN_API
|
977
|
+
fann_get_activation_steepness(struct fann *ann, int layer, int neuron)
|
978
|
+
{
|
979
|
+
struct fann_neuron* neuron_it = fann_get_neuron(ann, layer, neuron);
|
980
|
+
if(neuron_it == NULL)
|
981
|
+
{
|
982
|
+
return -1; /* layer or neuron out of bounds */
|
983
|
+
}
|
984
|
+
else
|
985
|
+
{
|
986
|
+
return neuron_it->activation_steepness;
|
987
|
+
}
|
988
|
+
}
|
989
|
+
|
990
|
+
FANN_EXTERNAL void FANN_API fann_set_activation_steepness(struct fann *ann,
|
991
|
+
fann_type steepness,
|
992
|
+
int layer,
|
993
|
+
int neuron)
|
994
|
+
{
|
995
|
+
struct fann_neuron* neuron_it = fann_get_neuron(ann, layer, neuron);
|
996
|
+
if(neuron_it == NULL)
|
997
|
+
return;
|
998
|
+
|
999
|
+
neuron_it->activation_steepness = steepness;
|
1000
|
+
}
|
1001
|
+
|
1002
|
+
FANN_EXTERNAL void FANN_API fann_set_activation_steepness_layer(struct fann *ann,
|
1003
|
+
fann_type steepness,
|
1004
|
+
int layer)
|
1005
|
+
{
|
1006
|
+
struct fann_neuron *last_neuron, *neuron_it;
|
1007
|
+
struct fann_layer *layer_it = fann_get_layer(ann, layer);
|
1008
|
+
|
1009
|
+
if(layer_it == NULL)
|
1010
|
+
return;
|
1011
|
+
|
1012
|
+
last_neuron = layer_it->last_neuron;
|
1013
|
+
for(neuron_it = layer_it->first_neuron; neuron_it != last_neuron; neuron_it++)
|
1014
|
+
{
|
1015
|
+
neuron_it->activation_steepness = steepness;
|
1016
|
+
}
|
1017
|
+
}
|
1018
|
+
|
1019
|
+
FANN_EXTERNAL void FANN_API fann_set_activation_steepness_output(struct fann *ann,
|
1020
|
+
fann_type steepness)
|
1021
|
+
{
|
1022
|
+
struct fann_neuron *last_neuron, *neuron_it;
|
1023
|
+
struct fann_layer *last_layer = ann->last_layer - 1;
|
1024
|
+
|
1025
|
+
last_neuron = last_layer->last_neuron;
|
1026
|
+
for(neuron_it = last_layer->first_neuron; neuron_it != last_neuron; neuron_it++)
|
1027
|
+
{
|
1028
|
+
neuron_it->activation_steepness = steepness;
|
1029
|
+
}
|
1030
|
+
}
|
1031
|
+
|
1032
|
+
FANN_GET_SET(enum fann_errorfunc_enum, train_error_function)
|
1033
|
+
FANN_GET_SET(fann_callback_type, callback)
|
1034
|
+
FANN_GET_SET(float, quickprop_decay)
|
1035
|
+
FANN_GET_SET(float, quickprop_mu)
|
1036
|
+
FANN_GET_SET(float, rprop_increase_factor)
|
1037
|
+
FANN_GET_SET(float, rprop_decrease_factor)
|
1038
|
+
FANN_GET_SET(float, rprop_delta_min)
|
1039
|
+
FANN_GET_SET(float, rprop_delta_max)
|
1040
|
+
FANN_GET_SET(float, rprop_delta_zero)
|
1041
|
+
FANN_GET_SET(float, sarprop_weight_decay_shift)
|
1042
|
+
FANN_GET_SET(float, sarprop_step_error_threshold_factor)
|
1043
|
+
FANN_GET_SET(float, sarprop_step_error_shift)
|
1044
|
+
FANN_GET_SET(float, sarprop_temperature)
|
1045
|
+
FANN_GET_SET(enum fann_stopfunc_enum, train_stop_function)
|
1046
|
+
FANN_GET_SET(fann_type, bit_fail_limit)
|
1047
|
+
FANN_GET_SET(float, learning_momentum)
|