1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
|
/* The kernel log likelihood is somewhat convoluted, to retain numeric precision. This
test compares the result to the direct approach, in the case of a few data sets
that are small enough that we don't have to worry about underflow.
*/
#include <apop.h>
long double sum_of_parts(apop_data *d1, apop_data *target){
apop_data *row, *trow;
long double p=1;
for(int i=0; (trow=Apop_r(target, i)) && trow->vector; i++){
long double tprob=0;
for(int j=0; (row=Apop_r(d1, j)) && row->vector; j++)
tprob+= apop_p(trow, apop_model_set_parameters(apop_normal, *row->vector->data, 1));
p *=tprob/d1->vector->size;
}
return p;
}
void go(apop_data *d1, apop_data *d2){
apop_model *k = apop_model_set_settings(apop_kernel_density, .base_data=d1);
assert(fabs(apop_p(d2, k)- sum_of_parts(d1, d2)) < 1e-5);
apop_model *test_copying = apop_model_copy(k);
assert(fabs(apop_p(d2, test_copying)- sum_of_parts(d1, d2)) < 1e-8);
apop_model_free(k);
apop_model_free(test_copying);
}
int main(){
apop_data *d1= apop_data_falloc((4), 2,4,6,8);
go(d1, apop_data_falloc((4), 1,3,5,7));
go(d1, apop_data_falloc((4), 1,1,1,1));
apop_data *d2= apop_data_falloc((4, 4, 1), 2,1.1, 4,2.2, 6,3.1, 8,0);
go(d2, apop_data_falloc((4, 4, 1), 1, 0, 3,0, 5, 0, 7, 0));
}
|
