Tong Zhang
-
Efficient and fast face recognition
-
Each function of the convolution layer is parameter tunable，such as padding
-
Optimization by matrix multiplication, compilation optimization and various other means.
-
Arm and X86 support.
- Modify CMakeLists.txt according to comments
- Input path
void addzero(Matrix *matrix, int padding);
void convolution(Matrix *matrix1, Matrix matrix2, Matrix *ans, int stride, float *bias, int anschannel);
void maxpool(const Matrix *matrix1, int size, Matrix *ans);
void Relu(Matrix *matrix);
- Transfer into Matrix
#pragma omp parallel for schedule(dynamic)
for (si = 0; si < matrix1->getSize() + 1 - matrix2.getSize(); si += stride) {
for (sj = 0; sj < matrix1->getSize() + 1 - matrix2.getSize(); sj += stride) {
for (int c = 0; c < matrix1->getChannel(); c++) {
for (int i = 0; i < matrix2.getSize(); i++) {
for (int j = 0; j < matrix2.getSize(); j++) {
data[pl++] = matrix1->getData()[c * matrix1->getSize() * matrix1->getSize() +
(si + i) * matrix1->getSize() + sj + j];
}
}
}
}
}
- Multiply the Matrix and add bias.
Matrix temp = matrix2 * (*matrix1);
ans->setData(temp.getData());
pl = 0;
for (sc = 0; sc < ans->getChannel(); sc++) {
for (si = 0; si < size * size; si++) {
temp.getData()[pl++] += bias[sc];
}
}
Change the order of i,j,k
#pragma omp parallel for schedule(dynamic)
for (int k = 0; k < matrix->getSize(); k++) {
for (int j = 0; j < matrix->getSize(); j++) {
for (int i = 0; i < matrix->getChannel(); i++) {
if (matrix->getData()[i * matrix->getSize() * matrix->getSize() + j * matrix->getSize() + k] < 0)
matrix->getData()[i * matrix->getSize() * matrix->getSize() + j * matrix->getSize() + k] = 0;
}
}
}
- In the convolution operation, the kernel is regarded as a window. Every time the window is moved, nine numbers are multiplied by it, and the nine numbers are straightened into one row, and the row-major order is used to make access to memory faster.
- Optimize matrix multiplication by AVX, OMP, block（cache locality） (OMP basically does not work. It may be caused by the fact that the amount of data is too small)
-
The order of loop i, j, k is reversed , so as to make the pointer jump as small as possible
-
Variables are transferred using pointers
-
Optimized by compiler.
- Before optimization
