STATIC MYBOOL finalize_scaling(lprec *lp, REAL *scaledelta)
{
int i;
/* Check if we should equilibrate */
if(is_scalemode(lp, SCALE_EQUILIBRATE) && !is_scaletype(lp, SCALE_CURTISREID)) {
int oldmode;
oldmode = lp->scalemode;
lp->scalemode = SCALE_LINEAR + SCALE_EXTREME;
scale(lp, scaledelta);
lp->scalemode = oldmode;
}
/* Check if we should prevent rounding errors */
if(is_scalemode(lp, SCALE_POWER2)) {
REAL *scalars;
if(scaledelta == NULL)
scalars = lp->scalars;
else
scalars = scaledelta;
for(i = 0; i <= lp->sum; i++)
scalars[i] = roundPower2(scalars[i]);
}
/* Then transfer the scalars to the model's data */
return( scale_rows(lp, scaledelta) && scale_columns(lp, scaledelta) );
}
template <typename T, typename X> bool scaler<T, X>::scale_with_log_balance() {
T balance = get_balance();
T balance_before_scaling = balance;
// todo : analyze the scale order : rows-columns, or columns-rows. Iterate if needed
for (int i = 0; i < 10; i++) {
scale_rows();
scale_columns();
T nb = get_balance();
if (nb < T(0.9) * balance) {
balance = nb;
} else {
balance = nb;
break;
}
}
return balance <= balance_before_scaling;
}
void scale_columns (
tensor& out,
const tensor& m,
const tensor& v
)
{
DLIB_CASSERT(have_same_dimensions(out,m));
DLIB_CASSERT(is_vector(v));
if (m.size() == 0 && v.size() == 0)
return;
DLIB_CASSERT(m.size() != 0);
DLIB_CASSERT(m.size()/m.num_samples() == v.size());
#ifdef DLIB_USE_CUDA
cuda::scale_columns(out, m, v);
#else
DLIB_CASSERT(false, "shouldn't be called right now");
out = scale_columns(mat(m), mat(v));
#endif
}
