inline
mxArray* to_mxarray(const array_nd<array_nd<T> >& a)
{
mxArray* mx = mxCreateCellArray(a.ndims(),
reinterpret_cast<const mwSize*>(a.sizes().get_base_ptr()));
size_t len = a.length();
for(size_t i = 0; i < len; i++) {
mxSetCell(mx, mwIndex(i), to_mxarray(a[i]));
}
return mx;
}
static void
predict_one (const mxArray* xtic,
const mxArray* oas,
const mxArray* filtmat,
mwIndex node,
mwIndex example,
const float* mybias,
double* yptr)
{
const mwIndex* fmir = mxGetIr (filtmat);
const mwIndex* fmjc = mxGetJc (filtmat);
const mwIndex* xir = mxGetIr (xtic);
const mwIndex* xjc = mxGetJc (xtic);
const double* xs = mxGetPr (xtic);
mwSize nexamples = mxGetN (xtic);
mwSize nfeatures = mxGetM (oas);
mwSize ncandidates = mxGetN (oas);
const float* oasptr = (const float*) mxGetPr (oas);
mwIndex argmax[N]; std::fill (argmax, argmax + N, mwIndex (0));
float max[N]; std::fill (max, max + N, -FLT_MAX);
for (mwIndex fmk = fmjc[node]; fmk < fmjc[node+1]; ++fmk) {
mwIndex candidate = fmir[fmk];
const float* oascandidate = oasptr + (nfeatures * candidate);
float pred = 0;
for (mwIndex hk = xjc[example]; hk < xjc[example+1]; ++hk) {
mwIndex feature = xir[hk];
pred += oascandidate[feature] * xs[hk];
}
if (mybias) {
pred += mybias[fmk - fmjc[node]];
}
mwIndex arg = 1 + candidate;
for (unsigned int n = 0; n < N; ++n) {
if (pred > max[n]) {
std::swap (argmax[n], arg);
std::swap (max[n], pred);
}
}
}
for (unsigned int n = 0; n < N; ++n) {
yptr[example + n * nexamples] = argmax[n];
}
}
inline
void assign(array_nd<array<T> >& a, const mxArray* mx)
{
if(mx::type_class<T>() == mxUNKNOWN_CLASS)
throw etype();
if(mxGetClassID(mx) != mx::type_class<T>())
throw etype();
size_array sz(mxGetNumberOfDimensions(mx),
reinterpret_cast<const size_t*>(mxGetDimensions(mx)));
a.resize(sz);
size_t len = a.length();
for(size_t i = 0; i < len; i++) {
assign(a[i], mxGetCell(mx, mwIndex(i)));
}
}
