summaryrefslogtreecommitdiff
path: root/neural.c
blob: 817bb50b478cdf58bed1c9f6482c31b69cbd57ac (plain)
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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <stdbool.h>
#include <string.h>
#include <math.h>
#include <limits.h>
#include <time.h>

#include "cleanbench.h"
#include "randnum.h"


/********************************
** BACK PROPAGATION NEURAL NET **
*********************************
** This code is a modified version of the code
** that was submitted to BYTE Magazine by
** Maureen Caudill.  It accomanied an article
** that I CANNOT NOW RECALL.
** The author's original heading/comment was
** as follows:
**
**  Backpropagation Network
**  Written by Maureen Caudill
**  in Think C 4.0 on a Macintosh
**
**  (c) Maureen Caudill 1988-1991
**  This network will accept 5x7 input patterns
**  and produce 8 bit output patterns.
**  The source code may be copied or modified without restriction,
**  but no fee may be charged for its use.
**
** ++++++++++++++
** I have modified the code so that it will work
** on systems other than a Macintosh -- RG
*/

/*
**  LOOP_MAX
**
** This constant sets the max number of loops through the neural
** net that the system will attempt before giving up.  This
** is not a critical constant.  You can alter it if your system
** has sufficient horsepower.
*/
#define LOOP_MAX  500000L

#define ERR -1
#define MAXPATS 10              /* max number of patterns in data file */
#define IN_X_SIZE 5             /* number of neurodes/row of input layer */
#define IN_Y_SIZE 7             /* number of neurodes/col of input layer */
#define IN_SIZE 35              /* equals IN_X_SIZE*IN_Y_SIZE */
#define MID_SIZE 8              /* number of neurodes in middle layer */
#define OUT_SIZE 8              /* number of neurodes in output layer */
#define MARGIN 0.1              /* how near to 1,0 do we have to come to stop? */
#define BETA 0.09               /* beta learning constant */
#define ALPHA 0.09              /* momentum term constant */
#define STOP 0.1                /* when worst_error less than STOP, training is done */

double  mid_wts[MID_SIZE][IN_SIZE];     /* middle layer weights */
double  out_wts[OUT_SIZE][MID_SIZE];    /* output layer weights */
double  mid_out[MID_SIZE];              /* middle layer output */
double  out_out[OUT_SIZE];              /* output layer output */
double  mid_error[MID_SIZE];            /* middle layer errors */
double  out_error[OUT_SIZE];            /* output layer errors */
double  mid_wt_change[MID_SIZE][IN_SIZE]; /* storage for last wt change */
double  out_wt_change[OUT_SIZE][MID_SIZE]; /* storage for last wt change */
double  in_pats[MAXPATS][IN_SIZE];      /* input patterns */
double  out_pats[MAXPATS][OUT_SIZE];    /* desired output patterns */
double  tot_out_error[MAXPATS];         /* measure of whether net is done */
double  out_wt_cum_change[OUT_SIZE][MID_SIZE]; /* accumulated wt changes */
double  mid_wt_cum_change[MID_SIZE][IN_SIZE];  /* accumulated wt changes */

double  worst_error; /* worst error each pass through the data */
double  average_error; /* average error each pass through the data */
double  avg_out_error[MAXPATS]; /* average error each pattern */

int iteration_count;    /* number of passes thru network so far */
int numpats;            /* number of patterns in data file */
int numpasses;          /* number of training passes through data file */
int learned;            /* flag--if true, network has learned all patterns */

static clock_t DoNNetIteration(unsigned long nloops);
static void do_mid_forward(int patt);
static void do_out_forward();
static void do_forward_pass(int patt);
static void do_out_error(int patt);
static void worst_pass_error();
static void do_mid_error();
static void adjust_out_wts();
static void adjust_mid_wts();
static void do_back_pass(int patt);
static void move_wt_changes();
static int check_out_error();
static void zero_changes();
static void randomize_wts();
static int read_data_file();

/***********
** DoNNet **
************
** Perform the neural net benchmark.
** Note that this benchmark is one of the few that
** requires an input file.  That file is "NNET.DAT" and
** should be on the local directory (from which the
** benchmark program in launched).
*/
double
DoNNET(void)
{
        clock_t         total_time = 0;
        int             iterations = 0;
        static bool     is_adjusted = false;
        static int      loops = 0;

        /*
        ** Init random number generator.
        ** NOTE: It is important that the random number generator
        **  be re-initialized for every pass through this test.
        **  The NNET algorithm uses the random number generator
        **  to initialize the net.  Results are sensitive to
        **  the initial neural net state.
        */
        randnum((int32_t)3);

        /*
        ** Read in the input and output patterns.  We'll do this
        ** only once here at the beginning.  These values don't
        ** change once loaded.
        */
        if(read_data_file()!=0) {
                exit(1);
        }

        /*
        ** See if we need to perform self adjustment loop.
        */
        if (is_adjusted == false) {
                is_adjusted = true;
                /*
                ** Do self-adjustment.  This involves initializing the
                ** # of loops and increasing the loop count until we
                ** get a number of loops that we can use.
                */
                do {
                        randnum(3);
                        ++loops;
                } while ((DoNNetIteration(loops) <= MINIMUM_TICKS) && (loops < LOOP_MAX));
        }

        do {
                randnum((int32_t)3);    /* Gotta do this for Neural Net */
                total_time += DoNNetIteration(loops);
                iterations += loops;
        } while (total_time < MINIMUM_SECONDS * CLOCKS_PER_SEC);

        return (double)(iterations * CLOCKS_PER_SEC) / total_time;
}

/********************
** DoNNetIteration **
*********************
** Do a single iteration of the neural net benchmark.
** By iteration, we mean a "learning" pass.
*/
static clock_t
DoNNetIteration(unsigned long nloops)
{
        clock_t start, stop;
int patt;

/*
** Run nloops learning cycles.  Notice that, counted with
** the learning cycle is the weight randomization and
** zeroing of changes.  This should reduce clock jitter,
** since we don't have to stop and start the clock for
** each iteration.
*/
        start = clock();
while(nloops--)
{
	randomize_wts();
	zero_changes();
	iteration_count=1;
	learned = false;
	numpasses = 0;
	while (!learned)
	{
		for (patt=0; patt<numpats; patt++)
		{
			worst_error = 0.0;      /* reset this every pass through data */
			move_wt_changes();      /* move last pass's wt changes to momentum array */
			do_forward_pass(patt);
			do_back_pass(patt);
			iteration_count++;
		}
		numpasses ++;
		learned = check_out_error();
	}
}
        stop = clock();

        return stop - start;
}

/*************************
** do_mid_forward(patt) **
**************************
** Process the middle layer's forward pass
** The activation of middle layer's neurode is the weighted
** sum of the inputs from the input pattern, with sigmoid
** function applied to the inputs.
**/
static void  do_mid_forward(int patt)
{
double  sum;
int     neurode, i;

for (neurode=0;neurode<MID_SIZE; neurode++)
{
	sum = 0.0;
	for (i=0; i<IN_SIZE; i++)
	{       /* compute weighted sum of input signals */
		sum += mid_wts[neurode][i]*in_pats[patt][i];
	}
	/*
	** apply sigmoid function f(x) = 1/(1+exp(-x)) to weighted sum
	*/
	sum = 1.0/(1.0+exp(-sum));
	mid_out[neurode] = sum;
}
return;
}

/*********************
** do_out_forward() **
**********************
** process the forward pass through the output layer
** The activation of the output layer is the weighted sum of
** the inputs (outputs from middle layer), modified by the
** sigmoid function.
**/
static void  do_out_forward()
{
double sum;
int neurode, i;

for (neurode=0; neurode<OUT_SIZE; neurode++)
{
	sum = 0.0;
	for (i=0; i<MID_SIZE; i++)
	{       /*
		** compute weighted sum of input signals
		** from middle layer
		*/
		sum += out_wts[neurode][i]*mid_out[i];
	}
	/*
	** Apply f(x) = 1/(1+exp(-x)) to weighted input
	*/
	sum = 1.0/(1.0+exp(-sum));
	out_out[neurode] = sum;
}
}

/**********************
** do_forward_pass() **
***********************
** control function for the forward pass through the network
**/
static void  do_forward_pass(int patt)
{
do_mid_forward(patt);   /* process forward pass, middle layer */
do_out_forward();       /* process forward pass, output layer */
}

/***********************
** do_out_error(patt) **
************************
** Compute the error for the output layer neurodes.
** This is simply Desired - Actual.
**/
static void do_out_error(int patt)
{
int neurode;
double error,tot_error, sum;

tot_error = 0.0;
sum = 0.0;
for (neurode=0; neurode<OUT_SIZE; neurode++)
{
	out_error[neurode] = out_pats[patt][neurode] - out_out[neurode];
	/*
	** while we're here, also compute magnitude
	** of total error and worst error in this pass.
	** We use these to decide if we are done yet.
	*/
	error = out_error[neurode];
	if (error <0.0)
	{
		sum += -error;
		if (-error > tot_error)
			tot_error = -error; /* worst error this pattern */
	}
	else
	{
		sum += error;
		if (error > tot_error)
			tot_error = error; /* worst error this pattern */
	}
}
avg_out_error[patt] = sum/OUT_SIZE;
tot_out_error[patt] = tot_error;
return;
}

/***********************
** worst_pass_error() **
************************
** Find the worst and average error in the pass and save it
**/
static void  worst_pass_error()
{
double error,sum;

int i;

error = 0.0;
sum = 0.0;
for (i=0; i<numpats; i++)
{
	if (tot_out_error[i] > error) error = tot_out_error[i];
	sum += avg_out_error[i];
}
worst_error = error;
average_error = sum/numpats;
return;
}

/*******************
** do_mid_error() **
********************
** Compute the error for the middle layer neurodes
** This is based on the output errors computed above.
** Note that the derivative of the sigmoid f(x) is
**        f'(x) = f(x)(1 - f(x))
** Recall that f(x) is merely the output of the middle
** layer neurode on the forward pass.
**/
static void do_mid_error()
{
double sum;
int neurode, i;

for (neurode=0; neurode<MID_SIZE; neurode++)
{
	sum = 0.0;
	for (i=0; i<OUT_SIZE; i++)
		sum += out_wts[i][neurode]*out_error[i];

	/*
	** apply the derivative of the sigmoid here
	** Because of the choice of sigmoid f(I), the derivative
	** of the sigmoid is f'(I) = f(I)(1 - f(I))
	*/
	mid_error[neurode] = mid_out[neurode]*(1-mid_out[neurode])*sum;
}
return;
}

/*********************
** adjust_out_wts() **
**********************
** Adjust the weights of the output layer.  The error for
** the output layer has been previously propagated back to
** the middle layer.
** Use the Delta Rule with momentum term to adjust the weights.
**/
static void adjust_out_wts()
{
int weight, neurode;
double learn,delta,alph;

learn = BETA;
alph  = ALPHA;
for (neurode=0; neurode<OUT_SIZE; neurode++)
{
	for (weight=0; weight<MID_SIZE; weight++)
	{
		/* standard delta rule */
		delta = learn * out_error[neurode] * mid_out[weight];

		/* now the momentum term */
		delta += alph * out_wt_change[neurode][weight];
		out_wts[neurode][weight] += delta;

		/* keep track of this pass's cum wt changes for next pass's momentum */
		out_wt_cum_change[neurode][weight] += delta;
	}
}
return;
}

/*************************
** adjust_mid_wts(patt) **
**************************
** Adjust the middle layer weights using the previously computed
** errors.
** We use the Generalized Delta Rule with momentum term
**/
static void adjust_mid_wts(int patt)
{
int weight, neurode;
double learn,alph,delta;

learn = BETA;
alph  = ALPHA;
for (neurode=0; neurode<MID_SIZE; neurode++)
{
	for (weight=0; weight<IN_SIZE; weight++)
	{
		/* first the basic delta rule */
		delta = learn * mid_error[neurode] * in_pats[patt][weight];

		/* with the momentum term */
		delta += alph * mid_wt_change[neurode][weight];
		mid_wts[neurode][weight] += delta;

		/* keep track of this pass's cum wt changes for next pass's momentum */
		mid_wt_cum_change[neurode][weight] += delta;
	}
}
return;
}

/*******************
** do_back_pass() **
********************
** Process the backward propagation of error through network.
**/
void  do_back_pass(int patt)
{

do_out_error(patt);
do_mid_error();
adjust_out_wts();
adjust_mid_wts(patt);

return;
}


/**********************
** move_wt_changes() **
***********************
** Move the weight changes accumulated last pass into the wt-change
** array for use by the momentum term in this pass. Also zero out
** the accumulating arrays after the move.
**/
static void move_wt_changes()
{
int i,j;

for (i = 0; i<MID_SIZE; i++)
	for (j = 0; j<IN_SIZE; j++)
	{
		mid_wt_change[i][j] = mid_wt_cum_change[i][j];
		/*
		** Zero it out for next pass accumulation.
		*/
		mid_wt_cum_change[i][j] = 0.0;
	}

for (i = 0; i<OUT_SIZE; i++)
	for (j=0; j<MID_SIZE; j++)
	{
		out_wt_change[i][j] = out_wt_cum_change[i][j];
		out_wt_cum_change[i][j] = 0.0;
	}

return;
}

/**********************
** check_out_error() **
***********************
** Check to see if the error in the output layer is below
** MARGIN*OUT_SIZE for all output patterns.  If so, then
** assume the network has learned acceptably well.  This
** is simply an arbitrary measure of how well the network
** has learned -- many other standards are possible.
**/
static int check_out_error()
{
int result,i,error;

result  = true;
error   = false;
worst_pass_error();     /* identify the worst error in this pass */

/*
#ifdef DEBUG
printf("\n Iteration # %d",iteration_count);
#endif
*/
for (i=0; i<numpats; i++)
{
/*      printf("\n Error pattern %d:   Worst: %8.3f; Average: %8.3f",
	  i+1,tot_out_error[i], avg_out_error[i]);
	fprintf(outfile,
	 "\n Error pattern %d:   Worst: %8.3f; Average: %8.3f",
	 i+1,tot_out_error[i]);
*/

	if (worst_error >= STOP) result = false;
	if (tot_out_error[i] >= 16.0) error = true;
}

if (error) result = ERR;


#ifdef DEBUG
/* printf("\n Error this pass thru data:   Worst: %8.3f; Average: %8.3f",
 worst_error,average_error);
*/
/* fprintf(outfile,
 "\n Error this pass thru data:   Worst: %8.3f; Average: %8.3f",
  worst_error, average_error); */
#endif

return(result);
}


/*******************
** zero_changes() **
********************
** Zero out all the wt change arrays
**/
static void zero_changes()
{
int i,j;

for (i = 0; i<MID_SIZE; i++)
{
	for (j=0; j<IN_SIZE; j++)
	{
		mid_wt_change[i][j] = 0.0;
		mid_wt_cum_change[i][j] = 0.0;
	}
}

for (i = 0; i< OUT_SIZE; i++)
{
	for (j=0; j<MID_SIZE; j++)
	{
		out_wt_change[i][j] = 0.0;
		out_wt_cum_change[i][j] = 0.0;
	}
}
return;
}


/********************
** randomize_wts() **
*********************
** Intialize the weights in the middle and output layers to
** random values between -0.25..+0.25
** Function rand() returns a value between 0 and 32767.
**
** NOTE: Had to make alterations to how the random numbers were
** created.  -- RG.
**/
static void randomize_wts()
{
int neurode,i;
double value;

/*
** Following not used int benchmark version -- RG
**
**        printf("\n Please enter a random number seed (1..32767):  ");
**        scanf("%d", &i);
**        srand(i);
*/

for (neurode = 0; neurode<MID_SIZE; neurode++)
{
	for(i=0; i<IN_SIZE; i++)
	{
	        /* value=(double)abs_randwc(100000L); */
		value=(double)abs_randwc((int32_t)100000);
		value=value/(double)100000.0 - (double) 0.5;
		mid_wts[neurode][i] = value/2;
	}
}
for (neurode=0; neurode<OUT_SIZE; neurode++)
{
	for(i=0; i<MID_SIZE; i++)
	{
	        /* value=(double)abs_randwc(100000L); */
		value=(double)abs_randwc((int32_t)100000);
		value=value/(double)10000.0 - (double) 0.5;
		out_wts[neurode][i] = value/2;
	}
}

return;
}


/*********************
** read_data_file() **
**********************
** Read in the input data file and store the patterns in
** in_pats and out_pats.
** The format for the data file is as follows:
**
** line#   data expected
** -----   ------------------------------
** 1               In-X-size,in-y-size,out-size
** 2               number of patterns in file
** 3               1st X row of 1st input pattern
** 4..             following rows of 1st input pattern pattern
**                 in-x+2  y-out pattern
**                                 1st X row of 2nd pattern
**                 etc.
**
** Each row of data is separated by commas or spaces.
** The data is expected to be ascii text corresponding to
** either a +1 or a 0.
**
** Sample input for a 1-pattern file (The comments to the
** right may NOT be in the file unless more sophisticated
** parsing of the input is done.):
**
** 5,7,8                      input is 5x7 grid, output is 8 bits
** 1                          one pattern in file
** 0,1,1,1,0                  beginning of pattern for "O"
** 1,0,0,0,1
** 1,0,0,0,1
** 1,0,0,0,1
** 1,0,0,0,1
** 1,0,0,0,0
** 0,1,1,1,0
** 0,1,0,0,1,1,1,1            ASCII code for "O" -- 0100 1111
**
** Clearly, this simple scheme can be expanded or enhanced
** any way you like.
**
** Returns -1 if any file error occurred, otherwise 0.
**/
static int read_data_file()
{
/*
** The Neural Net test requires an input data file.
** The name is specified here.
*/
const char *inpath="NNET.DAT";

FILE *infile;

int xinsize,yinsize,youtsize;
int patt, element, i, row;
int vals_read;
int val1,val2,val3,val4,val5,val6,val7,val8;

/* printf("\n Opening and retrieving data from file."); */

infile = fopen(inpath, "r");
if (infile == NULL)
{
	printf("\n CPU:NNET--error in opening file!");
	return -1 ;
}
vals_read =fscanf(infile,"%d  %d  %d",&xinsize,&yinsize,&youtsize);
if (vals_read != 3)
{
	printf("\n CPU:NNET -- Should read 3 items in line one; did read %d",vals_read);
	return -1;
}
vals_read=fscanf(infile,"%d",&numpats);
if (vals_read !=1)
{
	printf("\n CPU:NNET -- Should read 1 item in line 2; did read %d",vals_read);
	return -1;
}
if (numpats > MAXPATS)
	numpats = MAXPATS;

for (patt=0; patt<numpats; patt++)
{
	element = 0;
	for (row = 0; row<yinsize; row++)
	{
		vals_read = fscanf(infile,"%d  %d  %d  %d  %d",
			&val1, &val2, &val3, &val4, &val5);
		if (vals_read != 5)
		{
			printf ("\n CPU:NNET -- failure in reading input!");
			return -1;
		}
		element=row*xinsize;

		in_pats[patt][element] = (double) val1; element++;
		in_pats[patt][element] = (double) val2; element++;
		in_pats[patt][element] = (double) val3; element++;
		in_pats[patt][element] = (double) val4; element++;
		in_pats[patt][element] = (double) val5; element++;
	}
	for (i=0;i<IN_SIZE; i++)
	{
		if (in_pats[patt][i] >= 0.9)
			in_pats[patt][i] = 0.9;
		if (in_pats[patt][i] <= 0.1)
			in_pats[patt][i] = 0.1;
	}
	element = 0;
	vals_read = fscanf(infile,"%d  %d  %d  %d  %d  %d  %d  %d",
		&val1, &val2, &val3, &val4, &val5, &val6, &val7, &val8);

	out_pats[patt][element] = (double) val1; element++;
	out_pats[patt][element] = (double) val2; element++;
	out_pats[patt][element] = (double) val3; element++;
	out_pats[patt][element] = (double) val4; element++;
	out_pats[patt][element] = (double) val5; element++;
	out_pats[patt][element] = (double) val6; element++;
	out_pats[patt][element] = (double) val7; element++;
	out_pats[patt][element] = (double) val8; element++;
}

fclose(infile);
return(0);
}