-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathnn.hpp
1065 lines (877 loc) · 44.9 KB
/
nn.hpp
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
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#ifndef _nn_h
#define _nn_h
#define ennVersion "(1,2,0)"
#define PATHTOKERNALFILE "//home//parallella//Work//nnP//nn.cl"
#define PATHTOCLDEFSFILE "//home//parallella//Work//nnP//cldefs.inc"
#define CORECOUNT 16
using namespace std ;
#include <sys/stat.h> // POSIX only
#include <sstream>
#include <cstdlib>
#include <math.h>
#include <vector>
#include <time.h> // only used for the random seed generator
#include "nnFile.hpp"
#include <stdcl.h>
const int maxNodes = 64;
const int maxLayers = 3; // no longer used
enum layer_modifier { BIAS_NODE, TRANSITION_SIGMOID, TRANSITION_LINEAR, TRANSITION_BINARY };
enum node_modifier { INPUT_BINARY, INPUT_UNIFORM, INPUT_BIPOLAR };
struct nodeData
{
public:
node_modifier inputType; /// not used right now other than in randomise()
float p; /// P(x=1) = p
bool pIsOneHalf;
float nodeValue;
// float bias;
// vector<float> * incomingWeights;
};
struct layerData
{
public:
unsigned int nodeCount;
bool hasBiasNode;
layer_modifier transition;
vector<nodeData> * nodeInfo;
};
typedef void (*funcRunCallback)(const int, void *);
typedef void (*funcTrainCallback)(void *);
typedef void (*funcTestCallback)(const int, vector<float>*, vector<float>*, vector<float>*, vector<float>*, void *);
class nn
{
public:
nn(int inputLayerWidth, int hiddenLayerWidth, int outputLayerWidth, string & newName, float learningRateParam = 0.1)
/*
* Create a new network with
* inputLayerWidth input nodes,
* hiddenLayerWidth hidden nodes,
* outputLayerWidth output nodes and
* learningRateParam as the learning rate with
* newName as the network name
*/
{
vector<unsigned int> widths(3); // only while the number of layers is restricted to 3
widths[0] = inputLayerWidth;
widths[1] = hiddenLayerWidth;
widths[2] = outputLayerWidth;
setNetworkTopology(&widths);
networkName = newName;
clLearningRate = (cl_float)learningRateParam;
randomise();
majorVersion = minorVersion = revision = 0;
networkName = newName;
}
nn(vector<unsigned int>* networkTopo, string & newName, float learningRate = 0.1)
{
setNetworkTopology(networkTopo);
networkName = newName;
clLearningRate = (cl_float)learningRate;
randomise();
majorVersion = minorVersion = revision = 0;
networkName = newName;
}
// nn(int layerCount, int* layerWidths, float learningRateParam, string & newName)
nn(NNFile * newFile)
/*
* Reconstruct a network from a saved file with the wrapper newFile
*/
{
newFile->readInFile((void*)this);
};
~nn()
/*
* The network object destructor.
*
* ALWAYS make sure you call this function.
*
*/
{
int i;
unsigned int j;
if (clInputLayer) clfree((void*)clInputLayer);
if (clOutputLayer) clfree((void*)clOutputLayer);
if (clNodeBiases) clfree((void*)clNodeBiases);
if (clWeights) clfree((void*)clWeights);
if (clLayerWidths) clfree((void*)clLayerWidths);
if (clOutputError) clfree((void*)clOutputError);
// not not a pointer delete errorVector;
for (i=0; i<layerCount; i++)
{
for (j=0; j<(*layers)[i].nodeCount; i++)
delete (*layers)[i].nodeInfo;
}
delete layers;
}
// operation
void run(NNFile * inFile, funcRunCallback runComplete = NULL)
/*
* Run the contents of the data file wrapped by inFile calling the runComplete callback
* for each line.
*
* typedef void (*funcRunCallback)(const int index, void * thisNetwork);
* index is the index of the row that has just been run
* theNetwork is a void pointer to this object.
*
* Call ((nn*)theNetwork)->runResult(vector<float>* existingVector) to retrieve the result
*
*/
{
runCallback = runComplete; // Cannot pass callback via readInFile yet
inFile->readInFile((void*)this);
}
void run(vector<float> * inputVector, funcRunCallback runComplete = NULL, const int index = 0)
/*
* Pass inputVector to the input layer and trigger it to execute the network logic. Call the
* runComplete callback if it is not NULL.
*
* typedef void (*funcRunCallback)(const int index, void * thisNetwork);
* index is the index of the row that has just been run
* theNetwork is a void pointer to this object.
*
* Call ((nn*)theNetwork)->runResult(vector<float>* existingVector) to retrieve the result
*
* Note: this version is not multi threaded so waitForActivation and blockTillValue do nothing
*/
{
unsigned int i;
void * openHandle;
cl_kernel krn;
clndrange_t ndr;
// char strInfo[128];
// CONTEXT * pCon = stdcpu; // the cpu context !! all the storage exists in the atdacc context so this will not work as is
CONTEXT * pCon = stdacc;
cl_float * clDebug;
clDebug = (cl_float*)clmalloc(pCon, 2048*sizeof(float), 0);
for(i=0;i<2048;i++) clDebug[i]=-1000;
for (i=0; i < (*layers)[0].nodeCount; i++)
clInputLayer[i] = (*inputVector)[i];
/// openHandle = clopen(pCon, 0, CLLD_NOW); /// linked in version - the elf file must be linked into the executable at link time
writeDefsFile();
openHandle = clopen(pCon, PATHTOKERNALFILE, CLLD_NOW); /// JIT compile from file version
/// appendDefsToKernalString(); //TODO
/// openHandle = clsopen(pCon, str_k_forward, CLLD_NOW); /// string version (not done yet)
/// Get the handle to the kernel
krn = clsym(pCon, openHandle, "k_forward", CLLD_NOW);
// clGetKernelInfo(krn, CL_KERNEL_FUNCTION_NAME, sizeof(strInfo), strInfo, NULL);
ndr = clndrange_init1d(0, 16, 16); // get the core count from a cl call
/// transfer the inputdata biases and wieghts to the acc using clsync(,,, C_MEM_DEVICE|CL_EVENT_NOWAIT)
clmsync(pCon, 0, clOutputLayer, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
clmsync(pCon, 0, clInputLayer, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
clmsync(pCon, 0, clNodeBiases, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
clmsync(pCon, 0, clWeights, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
clmsync(pCon, 0, clDebug, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
//cout << "Calling clforka\n";
clforka(pCon, 0, krn, &ndr, CL_EVENT_NOWAIT,
clInputLayer,
clNodeBiases,
clWeights,
clOutputLayer,
clDebug);
//cout << "Transferring memory contents from the Epiphany using clmsync\n";
clmsync(pCon, 0, clOutputLayer, CL_MEM_HOST|CL_EVENT_NOWAIT);
clmsync(pCon, 0, clDebug, CL_MEM_HOST|CL_EVENT_NOWAIT);
clflush(pCon, 0, 0);
clwait(pCon, 0, CL_ALL_EVENT);
/// test
i=0;
if (clDebug[i] >= -1000) /// if we have put anything in the debug buffer
{
filebuf fbuf;
fbuf.open(".//nn.csv", std::ios::out);
ostream fout(&fbuf);
fout.precision(12);
while ((clDebug[i] > -999) && (i<2048))
{
if (clDebug[i] > 999)
fout << "\n";
else
fout << clDebug[i] << ",";
i++;
}
fout.flush();
fbuf.close();
}
if (runComplete != NULL)
{
runComplete(index, (void*)this);
}
else
{
if (runCallback != NULL) // this is the storage var for call backs that have been passed in when running from a file (saves having to pass the callback via networkFile class)
{
runCallback(index, (void*)this);
}
}
}
void run(vector<float> * inputVector, vector<float> * outputVector)
/*
* Run a single input vector and return the result. This call is designed to
* run synchronously.
*/
{
run(inputVector);
// wait for the result
runResult(outputVector);
}
vector<float> * runResult(vector<float> * outputVector)
/*
* Set and return outputVector from the last run.
*
* Call this quickly - I'm not sure how long it will be before the result is written
* over by the next output.
*
*/
{
unsigned int outI;
// float fl;
for (outI = 0; outI < layerNWidth(); outI++)
{
// (*outputVector)[outI] = (*layers)[2].nodeInfo->operator[](outI).nodeValue; // copy the contents
(*outputVector)[outI] = clOutputLayer[outI];
// fl = clOutputLayer[outI];
}
return outputVector;
}
void train(NNFile * trFile, funcTrainCallback trComplete = NULL)
/*
* Train the network using the training set in the file wrapped by trFile. Call the trComplete callback once
* when training is complete.
*
* typedef void (*funcTrainCallback)(void * nnObj); passes an anomymous pointer to this object back via the callback
*
* call ((nn*)nnObj)->trainingError(vector<float>* existingVector); to retrieve the most recent training error vector
*
*/
{
// call train with each vector
// unsigned int i;
//
// for (i=0; i < trFile->inputLines(); i++)
// train(trFile->inputSet(i), trFile->outputSet(i)); // don't pass the call back because we only want it called at the end not after each training set
trFile->readInFile((void*)this, true);
// block til complete
incrementRevision();
if (trComplete != NULL)
trComplete((void*)this);
}
void train(vector<float> * inputVector, vector<float> * desiredVector, funcTrainCallback trComplete = NULL)
/*
* Train the network with the single pair, inputVector and desiredVector. Call the trComplete callback if it is not NULL
* when training is complete.
*
* typedef void (*funcTrainCallback)(void * nnObj); passes an anomymous pointer to this object back via the callback
*
* call ((nn*)nnObj)->trainingError(vector<float>* existingVector); to retrieve the most recent training error vector
*
*/
{
try
{
unsigned int i;
void * openHandle;
cl_kernel krn;
clndrange_t ndr;
// char strInfo[128];
// CONTEXT * pCon = stdcpu; // the cpu context !! all the storage exists in the atdacc context so this will not work as is
CONTEXT * pCon = stdacc;
cl_float * clDesiredOutput = (cl_float*)clmalloc(pCon, desiredVector->size() * sizeof(float), 0);
cl_float * clDebug;
clDebug = (cl_float*)clmalloc(pCon, 2048*sizeof(float), 0);
for(i=0;i<2048;i++) clDebug[i]=-1000;
clWeightDeltas = (cl_float*)clmalloc(pCon, totalWeights*sizeof(float), 0); // temporary: space for core's to share incoming weight deltas
for (i=0; i < (*layers)[0].nodeCount; i++)
clInputLayer[i] = (*inputVector)[i];
for (i=0; i < (*layers)[layerCount-1].nodeCount; i++)
clDesiredOutput[i] = (*desiredVector)[i];
openHandle = clopen(pCon, 0, CLLD_NOW); /// linked in version - the elf file must be linked into the executable at link time
writeDefsFile();
/// openHandle = clopen(pCon, PATHTOKERNALFILE, CLLD_NOW); /// JIT compile from file version
/// appendDefsToKernalString(); //TODO
/// openHandle = clsopen(pCon, str_k_forward, CLLD_NOW); /// string version (not done yet)
/// Get the handle to the kernel
krn = clsym(pCon, openHandle, "k_train", CLLD_NOW);
// clGetKernelInfo(krn, CL_KERNEL_FUNCTION_NAME, sizeof(strInfo), strInfo, NULL);
ndr = clndrange_init1d(0, 16, 16); // get the core count from a cl call
/// transfer the inputdata biases and wieghts to the acc using clsync(,,, C_MEM_DEVICE|CL_EVENT_NOWAIT)
clmsync(pCon, 0, clInputLayer, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
clmsync(pCon, 0, clDesiredOutput, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
clmsync(pCon, 0, clNodeBiases, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
clmsync(pCon, 0, clWeights, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
clmsync(pCon, 0, clOutputError, CL_MEM_DEVICE|CL_EVENT_NOWAIT); /// not sure if I have to sync this one here
clmsync(pCon, 0, clDebug, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
clmsync(pCon, 0, clWeightDeltas, CL_MEM_DEVICE|CL_EVENT_NOWAIT);
//cout << "Calling clforka\n";
clforka(pCon, 0, krn, &ndr, CL_EVENT_NOWAIT,
clInputLayer,
clDesiredOutput,
clNodeBiases,
clWeights,
clOutputError,
clLearningRate,
clWeightDeltas,
clDebug);
//cout << "Transferring memory contents from the Epiphany using clmsync\n";
clflush(pCon, 0, 0);
clwait(pCon, 0, CL_ALL_EVENT);
clmsync(pCon, 0, clOutputError, CL_MEM_HOST|CL_EVENT_NOWAIT); /// The final output error
clmsync(pCon, 0, clWeights, CL_MEM_HOST|CL_EVENT_NOWAIT); /// The modified weights
clmsync(pCon, 0, clNodeBiases, CL_MEM_HOST|CL_EVENT_NOWAIT); /// The modified biases
clmsync(pCon, 0, clWeightDeltas, CL_MEM_HOST|CL_EVENT_NOWAIT); // testing
clmsync(pCon, 0, clDebug, CL_MEM_HOST|CL_EVENT_NOWAIT); // testing
clflush(pCon, 0, 0);
clwait(pCon, 0, CL_ALL_EVENT);
/// test
i=0;
if (clDebug[i] >= -1000) /// if we have put anything in the debug buffer
{
filebuf fbuf;
fbuf.open(".//nn.csv", std::ios::out);
ostream fout(&fbuf);
fout.precision(12);
while ((clDebug[i] > -999) && (i<2048))
{
if (clDebug[i] > 999)
{
/// a value greater tha 999 is to insert a \n in the output but print the value just in case it gets in ther by mistake
fout << clDebug[i];
fout << "\n";
}
else
fout << clDebug[i] << ",";
i++;
}
fout.flush();
fbuf.close();
}
if (trComplete != NULL)
trComplete((void*)this);
}
catch (internal_Error & iErr)
{
cout << iErr.mesg;// << " last error:" << iErr.lastError;
}
hasChanged = true;
}
status_t trainingError(vector<float> * errorVector)
/*
* Return the most recent error vector generated by the most recent training set.
*
* Note: the errorVector must exist and be the right size
*
*/
{
unsigned int i;
unsigned int outputNodeCount = layerNWidth();
if (errorVector->size() != outputNodeCount)
return FAILURE;
else
for (i = 0; i < outputNodeCount; i++)
(*errorVector)[i] = clOutputError[i];
return SUCCESS;
}
void test(NNFile * testFile, funcTestCallback testComplete = NULL)
/*
* Run the data component of the training file inside the wrapper testFile and
* compare the output generated by the network to the desired output. Calculate the difference.
*
* The call back function funcTestCallback is called once for every line in the input file and has the following form:
*
* typedef void (*funcTestCallback)(const int index, vector<float>* inputVector, vector<float>* desiredOutput, vector<float>* outputVector, vector<float>* errorVector, void * thisObject);
* index: the row number in the file
* inputVector: a pointer to the test data vector
* desiredOutput: a pointer to the desired output vector
* outputVector: a pointer to the actual output from the net
* errorVector: a pointer to a vector containing the desired minus the actual output
* thisObject: an anonymous pointer to this object
*
*/
{
testFile->readInFile((void*)this, false); // false to indicate that the file is NOT a training file
}
void test(const int index, vector<float> * inputVector, vector<float> * desiredOutput, funcTestCallback testComplete = NULL)
/*
* Test a single input vector and compare the result with the givine output vector. Then
* compare the output generated by the network to the desired output. Calculate the difference.
*
* The call back function funcTestCallback is called once and has the following form:
*
* typedef void (*funcTestCallback)(const int index, vector<float>* inputVector, vector<float>* desiredOutput, vector<float>* outputVector, vector<float>* errorVector, void * thisObject);
* index: the row number in the file
* inputVector: a pointer to the test data vector
* desiredOutput: a pointer to the desired output vector
* outputVector: a pointer to the actual output from the net
* errorVector: a pointer to a vector containing the desired minus the actual output
* thisObject: an anonymous pointer to this object
*
*/
{
size_t i;
vector<float> outputVec((*layers)[2].nodeCount);
// run
run(inputVector);
// block til value
// compare
// theOutputLayer->returnOutputVector(&outputVec);
for (i = 0; i != errorVector.size(); i++)
errorVector[i] = outputVec[i] - (*desiredOutput)[i];
if (testComplete != NULL)
testComplete(index, inputVector, desiredOutput, &outputVec, &errorVector, (void*)this);
}
void randomise()
/*
* Randomise the weights and biases in the network thereby restarting the training cycle from a different place.
*/
{
unsigned int layer;
float linkWeightVectorLength;
unsigned int faninToNode; // the number of incoming links
unsigned int firstLink, lastLink;
unsigned int linkIndex, nodeI;
unsigned int nodeIndex = 0;
int newRand;
float numerator, denominator;
float weight, weightMax;
node_modifier input_type;
bool pEqualsOneHalf;
float p;
srand(time(NULL));
for (layer = 1; layer < (unsigned int)layerCount; layer++)
{
// set the weight to a random number
// pEqualsOneHalf == true assumes p == 0.5
// for uniform inputs p is the upper most positive value expected
faninToNode = (*layers)[layer-1].nodeCount;
for (nodeI=0; nodeI<(*layers)[layer].nodeCount; nodeI++)
{
input_type = (*layers)[layer].nodeInfo->operator[](nodeI).inputType;
pEqualsOneHalf = (*layers)[layer].nodeInfo->operator[](nodeI).pIsOneHalf;
p = (*layers)[layer].nodeInfo->operator[](nodeI).p;
if ((input_type == INPUT_BINARY) && pEqualsOneHalf)
{
numerator = (float)5.1;
denominator = sqrt((float)faninToNode);
}
else if ((input_type == INPUT_BINARY) && !pEqualsOneHalf)
{
numerator = (float)2.55;
denominator = sqrt((float)faninToNode * p * (1 - p));
}
else if ((input_type == INPUT_BIPOLAR) && pEqualsOneHalf)
{
numerator = (float)2.55;
denominator = sqrt((float)faninToNode);
}
else if ((input_type == INPUT_BIPOLAR) && !pEqualsOneHalf)
{
numerator = (float)1.28;
denominator = sqrt((float)faninToNode * p * (1 - p));
}
else if (input_type == INPUT_UNIFORM)
{
numerator = (float)4.4;
denominator = p * sqrt((float)faninToNode);
}
else
{
throw; // opps!
}
weightMax = numerator / denominator;
firstLink = clNodeWeightIndex[nodeIndex]; // the index into the weight array where this node's links start
if (((cl_int)layer == (layerCount - 1)) && (nodeI == (*layers)[layer].nodeCount) - 1) //ie this is the verly lasy output nnode
lastLink = totalWeights;
else
lastLink = clNodeWeightIndex[nodeIndex+1]; // the index into the weight array where the next node's links start
linkWeightVectorLength = 0.0;
for(linkIndex=firstLink; linkIndex<lastLink; linkIndex++)
{
newRand = rand();
weight = (weightMax - ((float)newRand / ((float)RAND_MAX / (2 * weightMax))));
linkWeightVectorLength += weight * weight;
clWeights[linkIndex] = weight;
}
newRand = rand();
linkWeightVectorLength = sqrt(linkWeightVectorLength);
clNodeBiases[nodeIndex++] = linkWeightVectorLength - ((float)newRand / ((float)RAND_MAX / (2 * linkWeightVectorLength)));
}
}
hasChanged = true;
incrementMinorVersion();
}
// access
status_t saveTo(string * strPath)
/*
* Save the network to a file called <network Name>_<majorVersion>_<minorVersion>_<revision>.enn in the path supplied in string object strPath.
*
* Note: if you have the path name already as a C string call saveTo(const char *) rather than this function
*
*/
{
return saveTo(strPath->c_str());
}
status_t saveTo(const char * cstrPath)
/*
* Save the network to a file called <network Name>_<majorVersion>_<minorVersion>_<revision>.enn in the path supplied in C string cstrPath
*/
{
fstream * pFile;
status_t rVal;
char cstrPathFile[255]; // dumb
char cstrFileName[25]; // dumb
if (checkExists(cstrPath, false))
{
sprintf(cstrPathFile, "%s//%s", cstrPath, defaultName(cstrFileName));
pFile = new fstream();
pFile->open(cstrPathFile, ios::out);
rVal = saveTo(pFile);
pFile->close();
delete pFile;
return rVal;
}
else
throw format_Error(ENN_ERR_NON_FILE);
return SUCCESS;
}
status_t saveTo(fstream * pFile)
/*
* Save the network to the file stream pointed to by pFile. The name of the file will not be changed.
*/
{
string strContent;
status_t rVal;
rVal = saveOn(&strContent);
(*pFile) << strContent;
return rVal;
}
// save to disk
status_t saveOn(string * strOut)
/*
* save the net in the given existing string
*/
{
stringstream ss;
unsigned int layerI, nodeI, linkI;
unsigned int nodeIndex = 0;
unsigned int weightIndex = 0;
ss.precision(8);
ss << "version" << ennVersion << "\nname(" << networkName << "," << majorVersion << "," << minorVersion << "," << revision << ")\n" ;
//ss << "version(1,0,0)" << "\nname(" << networkName << "," << majorVersion << "," << minorVersion << "," << revision << ")\n" ; // compatabiity
//ss << "networkTopology(" << (*layers)[0].nodeCount << "," << (*layers)[1].nodeCount << "," << (*layers)[2].nodeCount << ")\n"; // compatability 3 layer network
ss << "networkTopology(" << layerCount << ";";
for (layerI=0 ; layerI< (unsigned int)(layerCount - 1); layerI++)
ss << (*layers)[layerI].nodeCount << ",";
ss << (*layers)[layerI].nodeCount << ")\n"; // finish with the )
ss << "learning(" << clLearningRate << "," << clTrainingMomentum << ")\n";
ss << "comment(link(layer, to node, from node, weight))\n";
ss << "comment(node(layer, node, bias))\n";
// call the detail storage process here
for (layerI = 1; layerI < (unsigned int)layerCount; layerI++)
{
ss << "comment(Storing layer:" << layerI <<")\n";
ss << "comment(TBD:layer modifiers)\n";
for (nodeI=0; nodeI<(*layers)[layerI].nodeCount; nodeI++)
{
for (linkI=0; linkI<(*layers)[layerI-1].nodeCount; linkI++)
ss << "link(" << layerI << "," << nodeI << "," << linkI << "," << clWeights[weightIndex++] << ")\n";
//ss << "link(" << (layerI - 1) << "," << linkI << "," << nodeI << "," << clWeights[weightIndex++] << ")\n"; // compatability mode
ss << "node(" << layerI << "," << nodeI << "," << clNodeBiases[nodeIndex++] << ")\n";
ss << "comment(TBD:node modifiers)\n";
}
}
hasChanged = false;
(*strOut) = ss.str();
return SUCCESS;
}
// Modify
status_t alter(int newIn, int newHidden, int newOut)
/*
* Alter the topology of the network to be
* newIn: the new number of input nodes
* newHidden: the new number of hidden nodes
* newOut: the new number of output nodes
*
* This will randomise the network and increment the major version resetting the minorVerions and revision
*
*/
{
// unsigned int layerNo = 0;
// delete theInputLayer;
// delete theHiddenLayer;
// delete theOutputLayer;
// net.setHiddenNodes(newHidden);
// net.setOutputNodes(newOut);
// net.setStandardInputNodes(newIn);
// theInputLayer = new inputLayer(net, layerNo++); // deleted in ~nn
// theHiddenLayer = new hiddenLayer(net, layerNo++); // deleted in ~nn
// theOutputLayer = new outputLayer(net, layerNo++); // deleted in ~nn
// theInputLayer->connectNodes(theHiddenLayer->nodeList());
// theHiddenLayer->connectNodes(theOutputLayer->nodeList());
randomise();
incrementMajorVersion();
hasChanged = true;
return SUCCESS;
}
// status_t alter(unsigned int layer, layer_modifier mod, bool boolAdd = true)
// /*
// * Alter a layer within the network. Currently you can only add or remove a bias node from layer zero (the input layer)
// *
// * This will randomise the network and increment the major version resetting the minorVerions and revision
// *
// */
// {
// network_description newNet;
//
// newNet = net; // keep all the old values
//
//// delete theInputLayer;
//// delete theHiddenLayer;
//// delete theOutputLayer;
//
// newNet.setInputLayerBiasNode(boolAdd);
//
// setup(newNet);
//
// randomise();
// incrementMajorVersion();
//
// hasChanged = true;
//
// return SUCCESS;
// }
char * defaultName(char * buffer)
/*
* Return the default name for the network, which is: <network Name>_<majorVersion>_<minorVersion>_<revision>.enn
*
* Note: the calling function must make sure that there is enough room in the buffer
*
*/
{
sprintf(buffer, "%s_%d_%d_%d.nn", networkName.c_str(), majorVersion, minorVersion, revision);
return buffer;
}
bool needsSaving() { return hasChanged; } // Return true if the network has changed since it was last saved.
// Build - callbacks for the networkFile that is reading in the netowrk from a .enn file
void setNetworkTopology(vector<unsigned int> * layerWidths)
{
cl_int i;
unsigned int j;
unsigned int nodeIndex = 0; // the index into the nodeWeightIndex array (which is flat unlike the layers->node structure)
int prevLayerNodeCount;
cl_int weightIndex = 0; // the index into weight array
layerCount = (cl_int)layerWidths->size();
// cout << "Layer widths:" << layerCount << " - ";
// for (i = 0; i<layerCount; i++)
// cout << (*layerWidths)[i] << " " ;
// cout << "\n";
layers = new vector<layerData>(layerCount);
setupLayer(&((*layers)[0]), (*layerWidths)[0], 0); // arg 2 is the previous layer width therefore 0 for the input layer
for (i=1; i<layerCount; i++)
setupLayer(&((*layers)[i]), (*layerWidths)[i], (*layerWidths)[i-1]);
clLayerWidths = (cl_int*) clmalloc(stdacc, layerCount * sizeof(cl_int), 0);
for (i=0; i<layerCount; i++)
clLayerWidths[i] = (*layers)[i].nodeCount;
clInputLayer = (cl_float*) clmalloc(stdacc, (size_t)layerZeroWidth() * sizeof(cl_float), 0);
clOutputLayer = (cl_float*) clmalloc(stdacc, (size_t)layerNWidth() * sizeof(cl_float), 0);
clOutputError = (cl_float*) clmalloc(stdacc, (size_t)layerNWidth() * sizeof(cl_float), 0);
//testing
for (i=0;i<(cl_int)layerNWidth();i++)
clOutputLayer[i] = -1.0;
//\\testing
totalWeights = 0;
nodeBiasArraySize = 0;
largestDerivedLayer = 0;
largestInputLayer = (*layerWidths)[0];
maxWeightsPerCore = 0;
totalDerivedNodes = (*layerWidths)[0]; /// input layer is copied to the derived value array to streamline forward and back passes
for (i=1; i < layerCount; i++)
{
totalWeights += (*layerWidths)[i] * (*layerWidths)[i-1];
nodeBiasArraySize += (*layerWidths)[i];
largestDerivedLayer = (largestDerivedLayer < (*layerWidths)[i]) ? (*layerWidths)[i] : largestDerivedLayer;
totalDerivedNodes += (*layerWidths)[i];
largestInputLayer = ((largestInputLayer < (*layerWidths)[i]) && (i != (layerCount - 1))) ? (*layerWidths)[i] : largestInputLayer;
maxWeightsPerCore += (((*layerWidths)[i] / CORECOUNT) + 1) * (*layerWidths)[i-1];
}
// cout << "total weights: " << totalWeights << " total Node Biases " << nodeBiasArraySize << "\n";
clWeights = (cl_float*) clmalloc(stdacc, totalWeights * sizeof(cl_float), 0);
clNodeBiases = (cl_float*) clmalloc(stdacc, nodeBiasArraySize * sizeof(cl_float), 0);
clNodeWeightIndex = (cl_int*) clmalloc(stdacc, nodeBiasArraySize * sizeof(cl_int), 0);
for (i=1; i < layerCount; i++)
{
prevLayerNodeCount = (*layers)[i-1].nodeCount;
for(j=0; j<(*layers)[i].nodeCount; j++)
{
clNodeWeightIndex[nodeIndex++] = weightIndex;
weightIndex += prevLayerNodeCount;
}
}
}
void setNodeBias(unsigned int layer, unsigned int node, float bias)
{
unsigned int i;
unsigned int offset = 0;
// layers->operator[](layer).nodeInfo->operator[](node).bias = bias;
for(i=1; i<layer; i++) /// bias array starts at 0 for layer 1
offset += clLayerWidths[i];
clNodeBiases[offset + node] = (cl_float)bias;
// cout << "b," << layer << "," << node << "," << bias << "," << offset << "," << "\n";
}
void setLinkWeight(unsigned int layer, unsigned int fromNode, unsigned int toNode, float weight)
{
unsigned int i;
unsigned int offset = 0;
// layers->operator[](layer).nodeInfo->operator[](toNode).incomingWeights->operator[](fromNode) = weight;
for(i=1; i < layer; i++)
{
offset += clLayerWidths[i] * clLayerWidths[i-1]; // gets us to the begining of the layer where the link goes
}
offset += clLayerWidths[i-1] * toNode;
offset += fromNode;
clWeights[offset] = (cl_float)weight;
// cout << "weight layer: " << layer << " from " << fromNode << " to " << toNode << " weight " << weight << " stored as " << clWeights[offset] << " \n";
}
void setName(string * name)
{
// cout << "name:" << (*name) << "\n";
networkName = *name;
}
void setVersion(unsigned int major, unsigned int minor, unsigned int revis)
{
// cout << "version " << major << " " << minor << " " << revis << "\n";
majorVersion = major;
minorVersion = minor;
revision = revis;
}
void setTrainingLearningRate(float learningRate)
{
// cout << "setting LR:" << learningRate << "\n";
clLearningRate = (cl_float)learningRate;
}
void setTrainingMomentum(float momentum)
{
// cout << "setting momentum:" << momentum << "\n";
clTrainingMomentum = (cl_float)momentum;
}
void setHasBiasNode(unsigned int layer, bool hasBiasNode)
{
// cout << "layer: " << layer << "bias node:" << hasBiasNode << "\n";
(*layers)[layer].hasBiasNode = hasBiasNode; // the node is added at setup time and is only used if hasBiasNode is true
}
void setNodeModifier(unsigned int layer, unsigned int node, node_modifier mod)
{
/// not implemented yet
throw;
}
// Access
unsigned int layerZeroWidth()
{
return (*layers)[0].nodeCount;
}
unsigned int layerNWidth()
{
return (*layers)[layerCount-1].nodeCount;
}
// Setup
private:
void incrementRevision() { revision++; }
void incrementMinorVersion() { minorVersion++; revision = 0; }
void incrementMajorVersion() { majorVersion++; minorVersion = revision = 0; }
void setupLayer(layerData * layer, unsigned int width, unsigned int previousLayerWidth)
{
unsigned int nodeI;
// cout << "layer width: " << width << " prev:" << previousLayerWidth << "\n";
layer->nodeCount = width;
layer->nodeInfo = new vector<nodeData>(width + 1); // add the space now for the bias node
layer->transition = TRANSITION_SIGMOID;
layer->hasBiasNode = false;
for(nodeI = 0; nodeI < width; nodeI++)
{
// cout << "adding node: " << nodeI << " content\n";
layer->nodeInfo->operator[](nodeI).inputType = INPUT_UNIFORM;
layer->nodeInfo->operator[](nodeI).p = 0.5;
layer->nodeInfo->operator[](nodeI).pIsOneHalf = true;
if (previousLayerWidth != 0) // previous == 0 indicates that the layer is the input layer therefore does not need any incoming weights
{
// layer->nodeInfo->operator[](nodeI).incomingWeights = new vector<float>(previousLayerWidth + 1); // add one in case the previous layer has a bias node
// cout << "space for incoming weights: " << layer->nodeInfo->operator[](nodeI).incomingWeights->size() << "\n";
}
// else
// cout << "no incoming weight space allocated\n";
//layer->nodeInfo->operator[](nodeI).bias = rand();
}
// and set up the bias node in case it gets switched on later
// cout << "adding bias node content\n";
layer->nodeInfo->operator[](nodeI).inputType = INPUT_BINARY;
layer->nodeInfo->operator[](nodeI).p = 1;
layer->nodeInfo->operator[](nodeI).pIsOneHalf = false;
layer->nodeInfo->operator[](nodeI).nodeValue = 1;
}
// Other
bool checkExists(const char * fileName, bool boolShouldBeFile = true)
{
struct stat fileAtt;
if (stat(fileName, &fileAtt) != 0)
return false;
else
if (boolShouldBeFile)
return S_ISREG(fileAtt.st_mode);
else
return S_ISDIR(fileAtt.st_mode);
}
void writeDefsFile()
{
fstream * pFile;
cl_int i;
if (checkExists(PATHTOCLDEFSFILE, true))