int model::load_training_data(const dataset &ds)
{
int nrow, ncol;
nrow = ds.ins_num();
ncol = ds.fea_num();
if (nrow <= 0 || ncol < 1) {
ULIB_FATAL("invalid training data dimensions");
return -1;
}
if (nrow > FLAGS_max_num_examples)
nrow = FLAGS_max_num_examples;
if (alloc_training_data(nrow, ncol)) {
ULIB_FATAL("couldn't allocate training data");
return -1;
}
double tavg = 0;
double tvar = 0;
for (int i = 0; i < nrow; ++i) {
double t = ds.get_tgv(i);
tavg += t;
tvar += t*t;
gsl_vector_set(_tv, i, t);
for (int j = 0; j < ncol; ++j)
gsl_matrix_set(_fm, i, j, ds.get_fea(i, j));
}
_t_avg = tavg/nrow;
_t_std = sqrt(tvar/nrow - _t_avg*_t_avg);
return 0;
}
int model::predict(const dataset &tds, gsl_matrix **pp)
{
int ret = -1;
gsl_matrix *mat = NULL;
gsl_matrix *ptv = NULL;
gsl_matrix *km1 = NULL;
gsl_matrix *km2 = NULL;
gsl_matrix *res = NULL;
gsl_matrix *stm = NULL;
gsl_vector_view avg_col;
gsl_vector_view dv;
if (tds.ins_num() <= 0 || tds.fea_num() != (int)_col_mean->size) {
ULIB_FATAL("invalid test dimensions, (ins_num=%d,fea_num=%d)",
tds.ins_num(), tds.fea_num());
goto done;
}
mat = gsl_matrix_alloc(tds.ins_num(), tds.fea_num());
if (mat == NULL) {
ULIB_FATAL("couldn't allocate test feature matrix");
goto done;
}
ptv = gsl_matrix_alloc(tds.ins_num(), 2);
if (ptv == NULL) {
ULIB_FATAL("couldn't allocate prediction matrix");
goto done;
}
if (tds.get_matrix(mat)) {
ULIB_FATAL("couldn't get test matrix");
goto done;
}
dbg_print_mat(mat, "Test Matrix:");
zero_out_mat(mat);
norm_mat(mat);
dbg_print_mat(mat, "Normalized Test Matrix:");
km1 = comp_kern_mat(mat, _fm, _kern);
if (km1 == NULL) {
ULIB_FATAL("couldn't compute test1 kernel matrix");
goto done;
}
dbg_print_mat(km1, "Test Kernel Matrix:");
km2 = comp_kern_mat(mat, mat, _kern);
if (km2 == NULL) {
ULIB_FATAL("couldn't compute test2 kernel matrix");
goto done;
}
dbg_print_mat(km1, "Test Kernel Matrix:");
dv = gsl_matrix_diagonal(km2);
res = gsl_matrix_alloc(km1->size1, _ikm->size2);
if (res == NULL) {
ULIB_FATAL("couldn't allocate temporary matrix");
goto done;
}
stm = gsl_matrix_alloc(km2->size1, km2->size2);
if (stm == NULL) {
ULIB_FATAL("couldn't allocate std matrix");
goto done;
}
gsl_blas_dgemm(CblasNoTrans, CblasNoTrans, 1.0, km1, _ikm, 0.0, res);
gsl_blas_dgemm(CblasNoTrans, CblasTrans, 1.0, res, km1, 0.0, stm);
gsl_matrix_sub(km2, stm);
dbg_print_mat(res, "Predictive Matrix:");
avg_col = gsl_matrix_column(ptv, 0);
gsl_blas_dgemv(CblasNoTrans, 1.0, res, _tv, 0.0, &avg_col.vector);
gsl_vector_add_constant(&avg_col.vector, _t_avg);
gsl_matrix_scale(km2, _t_std*_t_std);
gsl_vector_add_constant(&dv.vector, _noise_var);
for (size_t i = 0; i < km2->size1; ++i)
gsl_matrix_set(ptv, i, 1, sqrt(gsl_vector_get(&dv.vector, i)));
*pp = ptv;
ptv = NULL;
ret = 0;
done:
gsl_matrix_free(mat);
gsl_matrix_free(ptv);
gsl_matrix_free(km1);
gsl_matrix_free(km2);
gsl_matrix_free(res);
gsl_matrix_free(stm);
return ret;
}
