-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtest_nnc4.c
425 lines (344 loc) · 16 KB
/
test_nnc4.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
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
/*----------------------------------------------------------------------
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License version 2 as
published by the Free Software Foundation.
Neural Network Description:
Input(28*28) |>> CONV3X3 (26x26x8_out) >>|>> MAXPOOL2X2 (13x13x8_out) >>|
>> CONV3X3 (11x11x32_out) >>|>> MAXPOOL2X2 (5x5x32_out) >>|<< nvCells (10_out)
Reference:
1. https://victorzhou.com/blog/intro-to-cnns-part-1/
https://victorzhou.com/blog/intro-to-cnns-part-2/
2. MNIST handwrriten digit database:
http://yann.lecun.com/exdb/mnist/
Note:
1. Input image data should be normalized to [0 1.0], or loss function
may return NAN or INF results. So as weights/bias values.
2. Filter number 8 OR 5? Seem 5 is faster and also same accurate.
TODO:
1. Looptest to check mem leakage.
Journal:
2023-08-09: Create the file.
Midas Zhou
知之者不如好之者好之者不如乐之者
-----------------------------------------------------------------------*/
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <stdbool.h>
#include <math.h>
#include <string.h>
#include <sys/mman.h> /* mmap */
#include <fcntl.h> /* O_RDONLY */
#include <sys/stat.h>
#include <time.h>
#include "nnc.h"
#include "actfs.h"
#define ERR_LIMIT 0.0025 //0.001
#define BUFFER_IMGDATA 0 /* 1---MNIST image data are buffered to train_imgdata/test_imgdata
Except MNIST label data, they are ALWAYS buffered to train_target[]/test_target[]
0---NO buffer, read imgdata from MMAP directly
*/
float instLrate=0.0025; /* OR 0.005, Instant learning rate */
int main(void)
{
int i,j,k;
int count=0;
int loop=0;
/* Train data buffer */
const int TRAIN_IMGTOTAL=5000; //20000;
/* Note: each 28x28 pixles, gray values[0 255] are normalized to [0 1.0] */
double *train_imgdata = NULL;
if(BUFFER_IMGDATA) {
train_imgdata=(double *)malloc(TRAIN_IMGTOTAL*28*28*sizeof(double));
if(train_imgdata==NULL) exit(1);
}
unsigned char train_target[TRAIN_IMGTOTAL]; /* digits 0~9 */
/* Test data buffer */
const int TEST_IMGTOTAL=5000;
/* Note: each 28x28 pixles, gray values[0 255] are normalized to [0 1.0] */
double *test_imgdata=NULL;
if(BUFFER_IMGDATA) {
test_imgdata= (double *)malloc(TEST_IMGTOTAL*28*28*sizeof(double));
if(test_imgdata==NULL) exit(1);
}
unsigned char test_target[TEST_IMGTOTAL]; /* digits 0~9 */
/* To link it to input_layer->pins */
double data_input[28*28]; /* Normalized to [0 1.0] */
double data_target[10]; /* one_hot target values. ONLY one '1' and others are '0'. */
/* Fliters */
int numFilters=8;
int numFiltersC2=32;
/* Training */
double err, batch_err, mean_err;
int nb=1; /* number of batches */
int bs=TRAIN_IMGTOTAL/nb; //500; /* batch size; <=TRAIN_IMGTOTAL input samples (numbers + teacher value) for each batch training */
bool gradient_checked=false;
/* Timer */
time_t t_start, t_end;
int hours,mins,secs;
time_t tm_t; /* time in seconds */
struct tm *tm_s; /* time in struct */
/* <<<<<<<<<<<<<<<<< Read data from MNIST database file >>>>>>>>>>>>> */
////////////* Read MNIST handwriten digit database *////////////
/* ** ---- Image data ---
*
[offset] [type] [value] [description]
0000 32bit integer 0x00000803(2051) magic number
0004 32bit integer 10000 number of images
0008 32bit integer 28 number of rows
0012 32bit integer 28 number of columns
(followings are all pixel data -----: pdata )
0016 unsigned byte ?? pixel
0017 unsigned byte ?? pixel
... ...
*/
/* ** ---- Label data ---
*
[offset] [type] [value] [description]
0000 32bit integer 0x00000801(2049) magic number
0004 32bit integer 10000 number of items
(followings are all label data -----: )
0008 unsigned byte ?? label
0009 unsigned byte ?? label
... ...
*/
const char *train_images_path="train-images.idx3-ubyte";
const char *train_labels_path="train-labels.idx1-ubyte";
int fd0, fd1;
struct stat sb;
unsigned int magic;
unsigned int imgcnt, labelcnt;
unsigned int nr,nc;
void *addr0, *addr1;
unsigned char *pdata0, *pdata1;
unsigned char *pTrainImg, *pTestImg;
/* ------------ Read in data for training ------------ */
/* M1. Open MNIST image data file */
//fd0=open("t10k-images.idx3-ubyte", O_RDONLY);
fd0=open(train_images_path, O_RDONLY);
if(fd0<0) {
printf("%s: Fail to open image data file!\n", __func__);
exit(1);
}
if(fstat(fd0, &sb)<0)
exit(1);
/* M2. MMAP image data file */
//addr0=mmap(NULL, 8*1024*1024, PROT_READ, MAP_PRIVATE, fd0, 0);
addr0=mmap(NULL, sb.st_size, PROT_READ, MAP_PRIVATE, fd0, 0);
if(addr0==MAP_FAILED) {
printf("%s: Fail to mmap file!\n", __func__);
perror("mmap");
close(fd0);
exit(1);
}
pdata0 = (unsigned char *)addr0;
/* M3. Read MNIST data header */
magic=be32toh( *(unsigned int*)(pdata0+0) );
imgcnt=be32toh( *(unsigned int*)(pdata0+4) );
nr=be32toh( *(unsigned int*)(pdata0+8) );
nc=be32toh( *(unsigned int*)(pdata0+12) );
printf("MNIST '%s': magic=%d, imgcnt=%d, nr=%d, nc=%d\n", train_images_path, magic, imgcnt, nr,nc);
/* Result: MNIST t10k-images: magic=2051, imgcnt=10000, nr=28, nc=28 */
/* M4. Assign pdata to the beginning of pixel data */
pdata0 +=16; //(unsigned char*)pdata+16;
pdata0 +=0*28*28; /* Skip some images, they may be errors. */
pTrainImg = pdata0;
pTestImg = pdata0+TRAIN_IMGTOTAL*28*28;
if(BUFFER_IMGDATA) {
/* M5. Copy into train_imgdata[] */
for(i=0; i<TRAIN_IMGTOTAL*28*28; i++)
train_imgdata[i] = (double)pdata0[i] / 255.0;
/* M6. Copy into test_imgdata[] */
for(i=0; i<TEST_IMGTOTAL*28*28; i++) {
test_imgdata[i] = (double)pdata0[TRAIN_IMGTOTAL*28*28+i] / 255.0;
//test_imgdata[i] = (double)pdata[i];
}
}
/* MA1. Open MNIST label data file. */
//fd1=open("t10k-labels.idx1-ubyte", O_RDONLY);
fd1=open(train_labels_path, O_RDONLY);
if(fd1<0) {
printf("%s: Fail to open label data file!\n", __func__);
exit(1);
}
if(fstat(fd1, &sb)<0)
exit(1);
/* MA2. MMAP file */
//addr1=mmap(NULL, 10*1024, PROT_READ, MAP_PRIVATE, fd1, 0);
addr1=mmap(NULL, sb.st_size, PROT_READ, MAP_PRIVATE, fd1, 0);
if(addr1==MAP_FAILED) {
printf("%s: Fail to mmap file!\n", __func__);
perror("mmap");
close(fd1);
exit(1);
}
pdata1 = (unsigned char *)addr1;
/* MA3. Read MNIST data header */
magic=be32toh( *(unsigned int*)(pdata1+0) );
labelcnt=be32toh( *(unsigned int*)(pdata1+4) );
printf("MNIST '%s': magic=%d, labelcnt=%d\n", train_labels_path, magic, labelcnt);
/* MA4. Assign pdata to the beginning of label data */
pdata1 +=8;
pdata1 +=0*1; /* Skip some images, they may be errors. */
/* MA5. Copy into train_target[] */
for(i=0; i<TRAIN_IMGTOTAL; i++) {
train_target[i] = pdata1[i];
//printf("train_target[%d]: %d\n", i, train_target[i]);
}
/* MA6. Copy into test_target[] */
for(i=0; i<TEST_IMGTOTAL; i++) {
test_target[i] = pdata1[TRAIN_IMGTOTAL+i];
// printf("test_target[%d]: %d\n", i, train_target[i]);
}
////////////* END: Read MNIST handwriten digit database *//////////
/* <<<<<<<<<<<<<<<<< Create CNN(Convolution Neural Network) >>>>>>>>>>>>> */
printf("Create CNN model...\n");
/* 1. Create an input CONV3X3 Layer: (numFilters, numChannels, w, h, double *din, withBias) */
CONV3X3 *conv3x3=new_conv3x3(numFilters, 1, 28,28, data_input, true);
conv3x3->transfunc = func_ReLU;
NVLAYER *conv_layer=new_nvlayer(0, NULL, false); /* An empty nvlayer to hold conv3x3 */
conv_layer->conv3x3=conv3x3;
/* 2. Create a MAXPOOL2X2 Layer */ /* (CONV3X3 *pinconv3x3, numFilters, imw, imh, double **din) */
//MAXPOOL2X2 *maxpool2x2 =new_maxpool2x2( conv3x3, 8, conv3x3->ow, conv3x3->oh, NULL); /* If conv3x3, other's are ignored */
MAXPOOL2X2 *maxpool2x2 =new_maxpool2x2( conv3x3, 0, 0, 0, NULL); /* If conv3x3, other's are ignored */
NVLAYER *maxpool_layer=new_nvlayer(0, NULL, false); /* An empty nvlayer to hold conv3x3 */
maxpool_layer->maxpool2x2=maxpool2x2;
/* 2A. Create a CONV3X3 Layer: (numFilters, numChannels, w, h, double *din, withBias) */
CONV3X3 *conv3x3A=new_conv3x3(numFiltersC2, maxpool2x2->nf, maxpool2x2->ow, maxpool2x2->oh, &maxpool2x2->douts[0][0], true);
conv3x3->transfunc = func_ReLU;
conv3x3A->prederr = maxpool2x2->derr; /* Set prederr for backpropagation */
NVLAYER *convA_layer=new_nvlayer(0, NULL, false); /* An empty nvlayer to hold conv3x3 */
convA_layer->conv3x3=conv3x3A;
/* 3. Create a MAXPOOL2X2 Layer */ /* (CONV3X3 *pinconv3x3, numFilters, imw, imh, double **din) */
MAXPOOL2X2 *maxpool2x2A =new_maxpool2x2( conv3x3A, 0, 0, 0, NULL); /* If conv3x3, other's are ignored */
NVLAYER *maxpoolA_layer=new_nvlayer(0, NULL, false); /* An empty nvlayer to hold conv3x3 */
maxpoolA_layer->maxpool2x2=maxpool2x2A;
/* 3A. Create the output nvcell layer */
/* (nin, NVCELL * const *incells, double *din, double *dw, double bias, double (*transfer)() ) */
NVCELL *output_tempcell=new_nvcell(maxpool2x2A->nf*maxpool2x2A->ow*maxpool2x2A->oh, NULL, &maxpool2x2A->douts[0][0], NULL, 0, NULL); //func_R$
NVLAYER *output_layer=new_nvlayer(10, output_tempcell, true); /* true for transfunc defined */
output_layer->transfunc = func_softmax;
/* 4. Create an nerve net */
NVNET *nnet=new_nvnet(5); /* 5 layers inside */
nnet->nvlayers[0]=conv_layer;
nnet->nvlayers[1]=maxpool_layer;
nnet->nvlayers[2]=convA_layer;
nnet->nvlayers[3]=maxpoolA_layer;
nnet->nvlayers[4]=output_layer;
/* 5. Init params */
nvnet_init_params(nnet);
/* <<<<<<<<<<<<<<<<< CNN Training Process >>>>>>>>>>>>> */
/* 6. Set learning_rate and momentum friction */
nnc_set_mfrict(0.1);
/* NOTE:
* NO MORE USE if call nvnet_update_params()/nvnet_mmtupdate_params() later.
* Call nvnet_update_params(nnet, learn_rate), OR nvnet_mmtupdate_params(nnet, learn_rate, mfrict)
*/
/* 7. Init err value */
mean_err=10;
/* 7a. Start timing */
t_start=time(NULL);
printf("NN model starts training ...\n");
/* 8. Batch training */
count=0;
gradient_checked=false;
while( count<10 || (mean_err > ERR_LIMIT && count<3000 ) )
{
/* 8.1 Reset batch err */
batch_err=0.0;
/* 8.2 batch learning */
for(nb=0; nb< TRAIN_IMGTOTAL/bs; nb++) {
/* Run all samples in the batch */
for(i=0; i<bs; i++) {
//printf("\n === %dth_train, batch item %d/%d ===\n", count+1, i+1, bs);
/* 8.2.R1. update ONE sample for input: data_input, data_target */
#if !BUFFER_IMGDATA /* --- CORSS_CHECK T.1 --- */
for(k=0; k<28*28; k++)
data_input[k]=pTrainImg[(nb*bs+i)*(28*28)+k]/255.0; /* From mmap */
#else /* --- CORSS_CHECK T.1 --- */
conv_layer->conv3x3->din = train_imgdata+(nb*bs+i)*(28*28); /* From pointer train_imgdata */
#endif
for(k=0; k<10; k++)
data_target[k] = (k==train_target[nb*bs+i] ? 1.0 : 0.0);
/* 8.2.R2. nvnet feed forward, accumlate err values for batch error
NOTICE: The output layer has transfunc defined, as func_softmax().
*/
err = nvnet_feed_forward(nnet, data_target, func_lossCrossEntropy); //func_lossMSE);
if(isnan(err) || isinf(err) ) { /* If NAN */
printf("Return err is nan or inf! Too big learn_rate? or input data unnormalized?\n");
exit(1);
}
batch_err +=err;
//printf("nvnet_feed_forward get err=%f\n", err);
/* 8.2.R3. nvnet feed backward, update cell->derrs */
nvnet_feed_backward(nnet);
/* 8.2.R4. update params after feedback(backpropagation) computation */
//nvnet_mmtupdate_params(nnet, 0.002); //0.01);
nvnet_update_params(nnet, instLrate); /* ---0.01 learn_rate */
} /* for(i) */
} /* for(nb) */
/* 8.3 Mean err for batch training */
mean_err = batch_err/(nb*bs);
/* 8.4 Count training */
count++;
time(&tm_t);
tm_s=localtime(&tm_t);
printf("Epoch %d: samples=%d, mean_err=%0.8f [%02d:%02d:%02d]\n",count, TRAIN_IMGTOTAL, mean_err,
tm_s->tm_hour,tm_s->tm_min,tm_s->tm_sec);
}
/* 8a. End timing */
t_end=time(NULL);
secs=difftime(t_end, t_start);
hours=secs/3600;
mins=(secs-hours*3600)/60;
secs -= 3600*hours+60*mins;
/* 9. Print results */
printf(" %dth training, mean_err=%0.8f \n",count, mean_err);
printf("Finish %d times batch learning! time eplapsed: %02d:%02d:%02d \n",count, hours, mins, secs);
/* <<<<<<<<<<<<<<<<< Test CNN Model >>>>>>>>>>>>> */
int errcnt=0;
printf("\n----------- Test learned NN Model -----------\n");
for(i=0; i<TEST_IMGTOTAL; i++)
{
/* T1. Update data_input,data_target */
#if !BUFFER_IMGDATA /* --- CORSS_CHECK 8.2.R1 --- */
for(k=0; k<28*28; k++)
data_input[k]=pTestImg[i*(28*28)+k]/255.0; /* From mmap */
#else /* --- CORSS_CHECK 8.2.R1 --- */
conv_layer->conv3x3->din = test_imgdata+i*(28*28); /* Pointer to test_imgdata */
#endif
for(k=0; k<10; k++)
data_target[k] = (k==test_target[i] ? 1.0 : 0.0);
/* T2. CNN feed forward to predict */
err = nvnet_feed_forward(nnet, data_target, func_lossCrossEntropy); //func_lossMSE);
if( isnan(err) || isinf(err) ) { /* If NAN */
printf("Return err is nan or inf! Too big learn_rate?\n");
exit(1);
}
/* T3. Print result */
printf("[%d] Output: ", i);
for(k=0; k< output_layer->nc ; k++)
printf("%f, ",output_layer->douts[k]);
printf(" ");
printf("Target: %d ", test_target[i]);
if(output_layer->douts[ test_target[i] ] >0.5)
printf(" OK\n");
else {
printf(" Fails\n");
errcnt++;
}
}
printf("Train samples: %d, train time: %02d:%02d:%02d, Test samples: %d\n", TRAIN_IMGTOTAL, hours, mins, secs, TEST_IMGTOTAL);
printf("Err/Total: %d/%d Accuracy: %.2f%%\n", errcnt, TEST_IMGTOTAL, 100.0*(1.0-1.0*errcnt/TEST_IMGTOTAL));
/* Free data and nvcell/nvnet */
free(train_imgdata);
free(test_imgdata);
free_nvcell(output_tempcell);
free_nvnet(nnet); /* free nvnet also free its nvlayers and nvcells inside */
/* Unmap and close */
munmap(addr0, 8*1024*1024);
close(fd0);
munmap(addr1, 8*1024);
close(fd1);
return 0;
}