120 lines
4.1 KiB
C++
120 lines
4.1 KiB
C++
#include "core.h"
|
|
#include <cmath>
|
|
#include <cstdlib>
|
|
|
|
#define USE_PARALLEL // Çàêîììåíòèðóéòå, ÷òîáû îòêëþ÷èòü ïàðàëëåëèçì
|
|
#define MAX_THREADS 0 // Óêàæèòå ÷èñëî ÿäåð (0 — èñïîëüçîâàòü âñå äîñòóïíûå)
|
|
|
|
#ifdef USE_PARALLEL
|
|
#include <omp.h>
|
|
#define OMP_PARALLEL _Pragma("omp parallel for")
|
|
#define OMP_SET_THREADS() { if (MAX_THREADS > 0) omp_set_num_threads(MAX_THREADS); }
|
|
#else
|
|
#define OMP_PARALLEL
|
|
#define OMP_SET_THREADS()
|
|
#endif
|
|
|
|
|
|
NeuralNetwork::NeuralNetwork(LayerStructure_t layers[], int count) : numLayers(count) {
|
|
for (int i = 0; i < count; i++) sizes.push_back(layers[i].size);
|
|
for (int i = 0; i < count - 1; i++) {
|
|
std::vector<std::vector<double>> layerW;
|
|
double scale = sqrt(2.0 / sizes[i]);
|
|
for (int j = 0; j < sizes[i+1]; j++) {
|
|
std::vector<double> nodeW;
|
|
for (int k = 0; k < sizes[i]; k++)
|
|
nodeW.push_back(((double)rand()/RAND_MAX * 2 - 1) * scale);
|
|
layerW.push_back(nodeW);
|
|
}
|
|
weights.push_back(layerW);
|
|
biases.push_back(std::vector<double>(sizes[i+1], 0.0));
|
|
}
|
|
}
|
|
|
|
std::vector<double> NeuralNetwork::feedForward(const std::vector<double>& input) {
|
|
OMP_SET_THREADS(); // Ïðèìåíÿåì ëèìèò ÿäåð
|
|
|
|
outputs.clear();
|
|
outputs.push_back(input);
|
|
|
|
std::vector<double> curr = input;
|
|
|
|
for (int i = 0; i < numLayers - 1; i++) {
|
|
// Çàðàíåå ãîòîâèì âåêòîð íóæíîãî ðàçìåðà äëÿ òåêóùåãî ñëîÿ
|
|
std::vector<double> next(sizes[i + 1]);
|
|
|
|
// Ðàñïàðàëëåëèâàåì ðàñ÷åò êàæäîãî íåéðîíà â ñëîå
|
|
OMP_PARALLEL
|
|
for (int j = 0; j < sizes[i + 1]; j++) {
|
|
double sum = biases[i][j];
|
|
|
|
// Âíóòðåííèé öèêë îáû÷íî îñòàâëÿåì ïîñëåäîâàòåëüíûì,
|
|
// òàê êàê ñîçäàíèå ïîòîêîâ çäåñü äàñò áîëüøå òîðìîçîâ, ÷åì ïîëüçû
|
|
for (int k = 0; k < (int)curr.size(); k++) {
|
|
sum += curr[k] * weights[i][j][k];
|
|
}
|
|
|
|
// Ñèãìîèäà. Òåïåðü ïèøåì ïî èíäåêñó j — ýòî áåçîïàñíî äëÿ ïîòîêîâ
|
|
next[j] = 1.0 / (1.0 + exp(-sum));
|
|
}
|
|
|
|
curr = next;
|
|
outputs.push_back(curr);
|
|
}
|
|
return curr;
|
|
}
|
|
|
|
|
|
double NeuralNetwork::train(const std::vector<double>& input, const std::vector<double>& target, double lr) {
|
|
// 0. Óñòàíàâëèâàåì êîëè÷åñòâî ïîòîêîâ (åñëè âêëþ÷åíî â define)
|
|
OMP_SET_THREADS();
|
|
|
|
// 1. Ïðÿìîé ïðîõîä (ïîëó÷àåì ïðåäñêàçàíèå)
|
|
std::vector<double> pred = feedForward(input);
|
|
|
|
std::vector<std::vector<double>> errors(numLayers);
|
|
errors[numLayers - 1].resize(sizes[numLayers - 1]);
|
|
|
|
double totalErr = 0;
|
|
|
|
// 2. Âû÷èñëåíèå îøèáêè íà âûõîäíîì ñëîå
|
|
for (int i = 0; i < sizes[numLayers - 1]; i++) {
|
|
double e = target[i] - pred[i];
|
|
// Ïðîèçâîäíàÿ ôóíêöèè àêòèâàöèè (ñèãìîèäû): pred * (1 - pred)
|
|
errors[numLayers - 1][i] = e * pred[i] * (1.0 - pred[i]);
|
|
totalErr += e * e;
|
|
}
|
|
|
|
// 3. Îáðàòíîå ðàñïðîñòðàíåíèå îøèáêè (Backpropagation) ïî ñêðûòûì ñëîÿì
|
|
for (int i = numLayers - 2; i > 0; i--) {
|
|
errors[i].resize(sizes[i]);
|
|
|
|
// Ïàðàëëåëèì âû÷èñëåíèÿ äëÿ êàæäîãî íåéðîíà â ñëîå
|
|
OMP_PARALLEL
|
|
for (int j = 0; j < sizes[i]; j++) {
|
|
double e = 0;
|
|
for (int k = 0; k < sizes[i + 1]; k++) {
|
|
e += errors[i + 1][k] * weights[i][k][j];
|
|
}
|
|
// outputs[i][j] — ýòî ñîõðàíåííûé ðåçóëüòàò àêòèâàöèè ýòîãî íåéðîíà èç feedForward
|
|
errors[i][j] = e * outputs[i][j] * (1.0 - outputs[i][j]);
|
|
}
|
|
}
|
|
|
|
// 4. Îáíîâëåíèå âåñîâ è ñìåùåíèé (Biases)
|
|
for (int i = 0; i < numLayers - 1; i++) {
|
|
|
|
// Ïàðàëëåëèì îáíîâëåíèå âåñîâ ñëåäóþùåãî ñëîÿ
|
|
OMP_PARALLEL
|
|
for (int j = 0; j < sizes[i + 1]; j++) {
|
|
for (int k = 0; k < sizes[i]; k++) {
|
|
// Ãðàäèåíòíûé ñïóñê: ïðèáàâëÿåì (lr * îøèáêà * âõîäíîé ñèãíàë)
|
|
weights[i][j][k] += lr * errors[i + 1][j] * outputs[i][k];
|
|
}
|
|
biases[i][j] += lr * errors[i + 1][j];
|
|
}
|
|
}
|
|
|
|
return totalErr;
|
|
}
|