#ifndef TRAIN_IMAGE_INC_CPP
#define TRAIN_IMAGE_INC_CPP
#include "train.inc.cpp"
#include "layer.simple.inc.cpp"
class TrainerImage: public Trainer {
protected:
std::vector<unsigned char> data;
std::vector<unsigned int> shuffle;
const char *datafile;
const char *outfile;
Layout ofl, obl;
Layout::List oflist, oblist;
int stride, count;
public:
TrainerImage(): stride(), count() { }
bool configure(const char *datafile, const char *outfile) {
this->datafile = datafile;
this->outfile = outfile;
}
data.clear();
FILE *f = fopen(filename, "rb");
if (!f)
return printf("cannot open file for read: %s\n", filename), false;
fseek(f, 0, SEEK_END);
size_t fs = ftello(f);
fseek(f, 0, SEEK_SET);
data.resize(fs, 0);
if (!fread(data.data(), fs, 1, f))
return printf("cannot read from file: %s\n", filename), fclose(f), data.clear(), false;
fclose(f);
return true;
}
void imgTrain(Layer &l, const char *datafile, int size, const char *outfile, double trainRatio, int count) {
Layer &bl = l.back();
assert(!l.prev);
assert(datafile);
assert(count > 0 && size > 0);
assert(l.size == size);
assert(bl.size == size);
int blockSize = 1000;//1024*1024*1024/size;
assert(blockSize > 0);
FILE *f = fopen(datafile, "rb");
if (!f)
{ printf("cannot open file: %s\n", datafile); return; }
fseeko64(f, 0, SEEK_END);
long long fsize = ftello64(f);
int xCount = (int)(fsize/size);
if (xCount <= 0)
{ printf("no tests in file: %s\n", datafile); return; }
int *block = new int[blockSize*2];
int *shuffle = block + blockSize;
double *results = new double[blockSize];
unsigned char *blockData = new unsigned char[(blockSize + 1)*size];
unsigned char *blockResData = blockData + blockSize*size;
bool err = false;
for(int j = 0; j < blockSize; ++j)
{ shuffle[j] = j; results[j] = 0; }
int blocksCount = (count - 1)/blockSize + 1;
printf("training %d (%d x %d blocks), tests: %d, ratio: %f:\n", blocksCount*blockSize, blocksCount, blockSize, xCount, trainRatio);
double avgSum = 0;
for(int i = 0; i < blocksCount; ++i) {
for(int j = 0; j < blockSize; ++j) {
block[j] = rand()%xCount;
std::swap(shuffle[i], shuffle[rand()%blockSize]);
}
std::sort(block, block + blockSize);
for(int j = 0; j < blockSize; ++j) {
fseeko64(f, block[j]*(long long)size, SEEK_SET);
if (!fread(blockData + j*size, size, 1, f))
{ printf("cannot read data from file: %s\n", datafile); err = true; break; }
}
if (err) break;
printf(" next data block loaded\n");
double sumQ = 0;
for(int j = 0; j < blockSize; ++j) {
unsigned char *data = blockData + shuffle[j]*size;
for(double *ia = l.a, *e = ia + l.size; ia < e; ++ia, ++data)
*ia = *data/255.0;
double firstQ = 0, q = 0;
for(int repeat = 0; repeat < 1; ++repeat) {
l.pass();
for(double *ia = l.a, *iba = bl.a, *ibda = bl.da, *e = ia + l.size; ia < e; ++ia, ++iba, ++ibda) {
double d = *ia - *iba;
*ibda = d;
q += d*d;
}
q /= size;
if (!repeat) firstQ = q;
bl.backpass(trainRatio);
}
sumQ += firstQ;
avgSum += firstQ - results[j];
results[j] = firstQ;
int avgCnt = i ? blockSize : j + 1;
printf(" %4d: total: %6d, avg result: %f, last result: %f -> %f\n", j+1, i*blockSize+j+1, avgSum/avgCnt, firstQ, q);
}
printf("%4d: total: %6d, avg result: %f\n", i+1, (i+1)*blockSize, sumQ/blockSize);
if (outfile && !l.save(outfile))
{ printf("cannot save neural network weights to file: %s\n", outfile); err = true; break; }
unsigned char *data = blockResData;
for(double *iba = bl.a, *e = iba + bl.size; iba < e; ++iba, ++data)
*data = (unsigned char)(*iba*255.999);
tgaSave("data/output/sampleX.tga", blockData + shuffle[blockSize-1]*size, 256, 256, 3);
tgaSave("data/output/sampleY.tga", blockResData, 256, 256, 3);
}
delete[] block;
delete[] results;
delete[] blockData;
printf("finished\n");
}
protected:
bool prepare() override {
ofl = optimizeLayoutSimple(fl->layout);
obl = optimizeLayoutSimple(bl->layout);
assert(ofl && obl);
assert(ofl.getActiveCount() == obl.getActiveCount());
ofl.split(oflist, threadsCount);
obl.split(oblist, threadsCount);
stride = ofl.getActiveCount() + 1;
count = data.size()/stride;
if (count <= 0) return false;
shuffle.resize(count);
for(int i = 0; i < count; ++i)
shuffle[i] = i;
return true;
}
bool prepareBlock() override {
int cnt = itersPerBlock > count ? count : itersPerBlock;
for(int i = 0; i < cnt; ++i) {
int j = rand()%count;
if (i != j) std::swap(shuffle[i], shuffle[j]);
}
return true;
}
void loadData(Barrier &barrier, int, int iter) override {
struct I: public Iter {
typedef const unsigned char* Type;
static inline void iter4(Neuron &n, Type d, AccumType&) { n.v = *d/(NeuronReal)255; }
};
const unsigned char *id = data.data() + shuffle[iter%count]*stride;
iterateNeurons2<I>(oflist[barrier.tid], ofl, fl->neurons, id);
}
AccumReal verifyDataMain(int, int iter) override {
struct I: public Iter {
typedef int Type;
struct AccumType { int ri, mi; NeuronReal m; };
static inline void iter4(Neuron &n, Type d, AccumType &a) {
NeuronReal v1 = d == a.ri;
NeuronReal v0 = n.v;
n.d *= v1 - v0;
if (a.m < v0) { a.m = v0; a.mi = d; }
}
};
I::AccumType a = { data[ (shuffle[iter%count] + 1)*stride - 1 ], 0, 0 };
iterateNeurons2<I>(obl, obl, bl->neurons, 0, 1, &a);
return a.mi != a.ri;
}
};
#endif