void MLP::modifyWeights(const integer &exampleIndex, const realnumber &learningRate)
{
modifyDelta(m_input.col(exampleIndex), m_output.col(exampleIndex), 0);
for (integer j = m_last; j >= 0; --j)
m_layers[j] +=
learningRate
* m_Delta[j]
* addBias( (j>0) ? run(exampleIndex, j-1) : m_input.col(exampleIndex) ).transpose();
}
void myConvKernel_naive()
{
float *filterOutput_buf = (float*) _mm_malloc(sizeof(float) * outputSize, 512);
assert(filterOutput_buf != NULL);
memset(outputPlanes, 0, outputSize * nOutputPlanes);
#pragma omp parallel
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int ioHeight_spos = BLOCK_LOW(tid, nthreads, ioHeight);
int ioHeight_epos = BLOCK_LOW(tid + 1, nthreads, ioHeight);
int oS_spos = ioHeight_spos * ioWidth;
int oS_size = (ioHeight_epos - ioHeight_spos) * ioWidth;
for (int opIndex = 0; opIndex < nOutputPlanes; opIndex++)
{
float *filterOutput = filterOutput_buf;
float *outputPlane = outputPlanes + opIndex * outputSize;
for (int ipIndex = 0; ipIndex < nInputPlanes; ipIndex++)
{
int wMatIndex = nInputPlanes * opIndex + ipIndex;
float *inputPlane = inputPlanes + ipIndex * paddedInSize;
float *weightMatrix = weights + wMatIndex * wSize;
convolve3x3withPad(
inputPlane, filterOutput, weightMatrix,
ioHeight_spos, ioHeight_epos
);
addVec(oS_size, filterOutput + oS_spos, outputPlane + oS_spos);
}
}
#pragma omp barrier
#pragma omp for
for (int opIndex = 0; opIndex < nOutputPlanes; opIndex++)
{
int wMatIndex = nInputPlanes * opIndex;
float *outputPlane = outputPlanes + opIndex * outputSize;
addBias(outputSize, (float)(biases[opIndex]), outputPlane);
scaleIfLessThanX(outputSize, outputPlane, 0.0, 0.1);
}
}
_mm_free(filterOutput_buf);
}
