/* Draws a histogram with N bars where N = dym_rows(freqs)
Each bar is compartmented into K subbars, on top of each other, where
K = dym_cols(freqs) and the j'th subcomponent of the i'th bar has
height = dym_ref(freqs,i,j) and is on top of the j-1'th subcomponent.
The i'th bar is centered on x=xlo + (i+0.5) * (xhi - xlo) / num_bars */
void ongr_plot_hist(frame *fr,ongr *on,dym *freqs,bool ratio)
{
int i;
int num_bars = dym_rows(freqs);
int num_classes = dym_cols(freqs);
double xlo = on -> x_axis.lo;
double xhi = on -> x_axis.hi;
double bar_width = 0.9 * (xhi - xlo) / num_bars;
for ( i = 0 ; i < num_bars ; i++ )
{
double xmid = xlo + (i + 0.5) * (xhi - xlo) / num_bars;
double x1 = xmid - bar_width/2;
double x2 = xmid + bar_width/2;
int j;
double sumy = 0.0;
double total_sum_y = real_max(1e-5,dym_sum_row(freqs,i));
for ( j = 0 ; j < num_classes ; j++ )
{
double y1 = sumy;
double dy = dym_ref(freqs,i,j) / ((ratio) ? (total_sum_y/100.0) : 1.0);
double y2 = sumy + dy;
int amut_col = color_code_to_ag_color(j);
ongr_colored_bordered_rectangle(fr,on,x1,y1,x2,y2,amut_col);
sumy = y2;
}
}
}
/* Creates two new dyms, both with the same number of columns as the
input x. *r_test will have "num_test_rows" rows and *r_train will
have "dym_rows(x) - num_test_rows" rows. *r_test will consist of
"num_test_rows" rows of x selected at random without replacement.
*r_train will contain the rest. The rows will appear in the same order
as they did in the original. */
void break_dym_into_train_and_test(dym *x,int num_test_rows,
dym **r_train,dym **r_test)
{
ivec *train_rows;
ivec *test_rows;
make_train_and_test_rows(/* train_and_test_rows = */ NULL /* denoting ALL */,
dym_rows(x),num_test_rows,&train_rows,&test_rows);
*r_train = mk_dym_from_subset_of_rows(x,train_rows);
*r_test = mk_dym_from_subset_of_rows(x,test_rows);
free_ivec(train_rows);
free_ivec(test_rows);
}
void lr_compute_n_from_dym( const dym *M, double b0, dyv *b, dyv *n)
{
int numrows, numgood, row, j;
double sum;
numrows = dym_rows( M);
numgood = dyv_size(b);
for (row=0; row < numrows; ++row) {
sum = 0.0;
for (j=0; j<numgood; ++j) sum += dym_ref( M, row, j) * dyv_ref( b, j);
sum += b0;
dyv_set( n, row, sum);
}
return;
}
double lr_deviance_from_dym_b( const dym *M, dyv *y, double b0, dyv *b)
{
int numrows;
double dev;
dyv *n, *u;
numrows = dym_rows( M);
n = mk_dyv( numrows);
u = mk_dyv( numrows);
lr_compute_n_from_dym( M, b0, b, n);
lr_compute_u_from_n( n, u);
dev = lr_deviance_basic( y, u);
free_dyv( n);
free_dyv( u);
return dev;
}
lr_train *mk_lr_train_from_dym( dym *factors, dyv *outputs, lr_options *opts)
{
/* Set rows to NULL if you want all rows from ds to be used. */
int numrows, numatts;
lr_train *lrt;
numrows = dym_rows( factors);
numatts = dym_cols( factors)+1; /* Number of factors including constant. */
/* Create lr lrt structure. */
lrt = AM_MALLOC(lr_train);
/* Copy in opts. */
lrt->opts = mk_copy_lr_options( opts);
/* Assign factors and outputs into lr structure. */
lrt->X = NULL;
lrt->M = factors;
/* Outputs. */
lrt->y = mk_copy_dyv( outputs);
if (!dyv_is_binary( outputs)) {
my_error( "mk_lr_train: Error: outputs are not binary.\n");
}
/* Set log likelihood of saturated model. */
lrt->likesat = 0.0;
/* Initialize remainder of lr struct */
lrt->numatts = numatts;
lrt->numrows = numrows;
/* Create lr_state member. */
lrt->lrs = mk_lr_state( lrt, opts);
/* Now that the structure is complete, update n and u to prepare for
iterations. */
lr_train_update_n(lrt);
lr_train_update_u(lrt);
return lrt;
}
