|
|
53488e |
|
|
|
53488e |
#include <ctime></ctime>
|
|
|
53488e |
#include <cstdlib></cstdlib>
|
|
|
53488e |
#include <cstdio></cstdio>
|
|
|
53488e |
|
|
|
53488e |
#include <chrono></chrono>
|
|
|
53488e |
#include <algorithm></algorithm>
|
|
|
53488e |
|
|
|
53488e |
#include "nnlayer3.mt.inc.cpp"
|
|
|
53488e |
#include "tga.inc.cpp"
|
|
|
53488e |
|
|
|
53488e |
|
|
|
53488e |
long long timeUs() {
|
|
|
53488e |
static std::chrono::steady_clock::time_point begin = std::chrono::steady_clock::now();
|
|
|
53488e |
return (long long)std::chrono::duration_cast<std::chrono::microseconds>( std::chrono::steady_clock::now() - begin ).count();</std::chrono::microseconds>
|
|
|
53488e |
}
|
|
|
53488e |
|
|
|
53488e |
|
|
|
53488e |
void imgTrain(Layer &l, const char *datafile, int size, const char *outfile, int blockSize, int blocksCount, Real trainRatio, int threads) {
|
|
|
53488e |
Layer &fl = l.front();
|
|
|
53488e |
Layer &bl = l.back();
|
|
|
53488e |
|
|
|
53488e |
assert(!l.prev);
|
|
|
53488e |
assert(datafile);
|
|
|
53488e |
assert(size > 0);
|
|
|
53488e |
assert(fl.countNeurons() == size);
|
|
|
53488e |
assert(bl.countNeurons() == size);
|
|
|
53488e |
|
|
|
53488e |
assert(blockSize > 0);
|
|
|
53488e |
assert(blocksCount > 0);
|
|
|
53488e |
assert(trainRatio > 0);
|
|
|
53488e |
assert(threads > 0);
|
|
|
53488e |
|
|
|
53488e |
FILE *f = fopen(datafile, "rb");
|
|
|
53488e |
if (!f)
|
|
|
53488e |
{ printf("cannot open file: %s\n", datafile); return; }
|
|
|
53488e |
fseeko64(f, 0, SEEK_END);
|
|
|
53488e |
long long fsize = ftello64(f);
|
|
|
53488e |
int xCount = (int)(fsize/size);
|
|
|
53488e |
if (xCount <= 0)
|
|
|
53488e |
{ printf("no tests in file: %s\n", datafile); return; }
|
|
|
53488e |
|
|
|
53488e |
printf("allocate %lld bytes for tests\n", ((long long)blockSize + 1)*size);
|
|
|
53488e |
|
|
|
53488e |
int *block = new int[blockSize*2];
|
|
|
53488e |
int *shuffle = block + blockSize;
|
|
|
53488e |
unsigned char *blockData = new unsigned char[(blockSize + 1)*size];
|
|
|
53488e |
unsigned char *blockResData = blockData + blockSize*size;
|
|
|
53488e |
bool err = false;
|
|
|
53488e |
|
|
|
53488e |
for(int j = 0; j < blockSize; ++j)
|
|
|
53488e |
shuffle[j] = j;
|
|
|
53488e |
|
|
|
53488e |
TrainMT tmt;
|
|
|
53488e |
tmt.layer = &fl;
|
|
|
53488e |
tmt.dataX = blockData;
|
|
|
53488e |
tmt.dataY = blockData;
|
|
|
53488e |
tmt.strideX = tmt.strideY = size;
|
|
|
53488e |
tmt.shuffle = shuffle;
|
|
|
53488e |
tmt.count = blockSize;
|
|
|
53488e |
tmt.trainRatio = trainRatio;
|
|
|
53488e |
|
|
|
53488e |
printf("training %d (%d x %d blocks), tests: %d, ratio: %f:\n", blocksCount*blockSize, blocksCount, blockSize, xCount, trainRatio);
|
|
|
53488e |
|
|
|
53488e |
long long t0 = timeUs();
|
|
|
53488e |
for(int i = 0; i < blocksCount; ++i) {
|
|
|
53488e |
for(int j = 0; j < blockSize; ++j) {
|
|
|
53488e |
block[j] = rand()%xCount;
|
|
|
53488e |
std::swap(shuffle[j], shuffle[rand()%blockSize]);
|
|
|
53488e |
}
|
|
|
53488e |
std::sort(block, block + blockSize);
|
|
|
53488e |
|
|
|
53488e |
for(int j = 0; j < blockSize; ++j) {
|
|
|
53488e |
fseeko64(f, block[j]*(long long)size, SEEK_SET);
|
|
|
53488e |
if (!fread(blockData + j*size, size, 1, f))
|
|
|
53488e |
{ printf("cannot read data from file: %s\n", datafile); err = true; break; }
|
|
|
53488e |
}
|
|
|
53488e |
if (err) break;
|
|
|
53488e |
|
|
|
53488e |
//printf(" next data block loaded\n");
|
|
|
53488e |
|
|
|
53488e |
long long t = timeUs();
|
|
|
53488e |
double res = tmt.train(threads);
|
|
|
53488e |
long long dt = timeUs() - t;
|
|
|
53488e |
|
|
|
53488e |
if (outfile && !fl.saveAll(outfile))
|
|
|
53488e |
{ printf("cannot save neural network weights to file: %s\n", outfile); err = true; break; }
|
|
|
53488e |
|
|
|
53488e |
unsigned char *data = blockResData;
|
|
|
53488e |
for(Neuron *ibn = bl.neurons, *e = ibn + size; ibn < e; ++ibn, ++data) {
|
|
|
53488e |
Real v = (ibn->v - 0.25)*2;
|
|
|
53488e |
*data = v < 0 ? 0u : v > 1 ? 255u : (unsigned char)(v * 255.999);
|
|
|
53488e |
}
|
|
|
53488e |
tgaSave("data/output/sampleX.tga", blockData + shuffle[blockSize-1]*size, 256, 256, 3);
|
|
|
53488e |
tgaSave("data/output/sampleY.tga", blockResData, 256, 256, 3);
|
|
|
53488e |
|
|
|
53488e |
long long t1 = timeUs();
|
|
|
53488e |
long long dt0 = t1 - t0;
|
|
|
53488e |
t0 = t1;
|
|
|
53488e |
|
|
|
53488e |
printf("%4d: total: %6d, avg result: %f, time: %f + %f = %f\n", i+1, (i+1)*blockSize, res, (dt0-dt)*0.000001, dt*0.000001, dt0*0.000001);
|
|
|
53488e |
|
|
|
53488e |
}
|
|
|
53488e |
|
|
|
53488e |
delete[] block;
|
|
|
53488e |
delete[] blockData;
|
|
|
53488e |
|
|
|
53488e |
printf("finished\n");
|
|
|
53488e |
}
|
|
|
53488e |
|
|
|
53488e |
|
|
|
53488e |
int main() {
|
|
|
53488e |
srand(time(NULL));
|
|
|
53488e |
|
|
|
53488e |
const char *datafile = "data/img256-data.bin";
|
|
|
53488e |
const char *outfile = "data/output/img256-weights.bin";
|
|
|
53488e |
|
|
|
53488e |
printf("create neural network\n");
|
|
|
53488e |
|
|
|
53488e |
Layer l(nullptr, 256, 256, 3);
|
|
|
53488e |
//new Layer(&l, 128, 128, 3, 6);
|
|
|
53488e |
//new Layer(&l, 64, 64, 3, 8);
|
|
|
53488e |
//new Layer(&l, 32, 32, 3, 11);
|
|
|
53488e |
//new Layer(&l, 16, 16, 4, 16);
|
|
|
53488e |
//new Layer(&l, 16, 16, 4, 16);
|
|
|
53488e |
//new Layer(&l, 16, 16, 4, 16);
|
|
|
53488e |
//new Layer(&l, 32, 32, 3, 11);
|
|
|
53488e |
//new Layer(&l, 64, 64, 3, 8);
|
|
|
53488e |
//new Layer(&l, 128, 128, 3, 6);
|
|
|
53488e |
new Layer(&l, 256, 256, 3, 4);
|
|
|
53488e |
new Layer(&l, 256, 256, 3, 4);
|
|
|
53488e |
new Layer(&l, 256, 256, 3, 4);
|
|
|
53488e |
|
|
|
53488e |
printf(" neurons: %d, links %d, memSize: %llu\n", l.totalNeurons(), l.totalLinks(), (unsigned long long)l.totalMemSize());
|
|
|
53488e |
|
|
|
53488e |
if (outfile) {
|
|
|
53488e |
printf("try load previously saved network\n");
|
|
|
53488e |
l.loadAll(outfile);
|
|
|
53488e |
}
|
|
|
53488e |
|
|
|
53488e |
printf("train\n");
|
|
|
53488e |
imgTrain(l, datafile, l.countNeurons(), outfile, 1000, 10000, 0.1, 4);
|
|
|
53488e |
|
|
|
53488e |
return 0;
|
|
|
53488e |
}
|
|
|
53488e |
|