void testErrors(const LWPR_Model *model, double *MSE, double *wMSE) {
int i,numPoints = 0;
double x[2],y[2],yp[2];
double conf[2],weight[2];
weight[0]=0.0;
weight[1]=0.0;
MSE[0]=MSE[1]=wMSE[0]=wMSE[1]=0.0;
for (x[1]=-1.0; x[1]<=1.01; x[1]+=0.05) {
for (x[0]=-1.0; x[0]<=1.01; x[0]+=0.05) {
y[0] = cross(x[0],x[1]);
y[1] = y[0] + 10;
lwpr_predict(model, x, 0.001, yp, conf, NULL);
for (i=0;i<2;i++) {
double diff = y[i]-yp[i];
double sig2 = conf[i]*conf[i];
weight[i] += 1.0/sig2;
MSE[i] += diff*diff;
wMSE[i] += diff*diff/sig2;
}
numPoints++;
}
}
for (i=0;i<2;i++) {
MSE[i]/=(double) numPoints;
wMSE[i]/=weight[i];
}
}
static PyObject *PyLWPR_predict(PyLWPR *self, PyObject *args) {
double cutoff = 0.0;
LWPR_Model *model = &(self->model);
PyArrayObject *x;
if (!PyArg_ParseTuple(args, "O!|d", &PyArray_Type, &x, &cutoff)) return NULL;
if (set_vector_from_array(model->nIn, self->extra_in, x)) return NULL;
lwpr_predict(model,self->extra_in, cutoff, self->extra_out, NULL, NULL);
return get_array_from_vector(model->nOut, self->extra_out);
}
static PyObject *PyLWPR_predict_conf(PyLWPR *self, PyObject *args) {
double cutoff = 0.0;
LWPR_Model *model = &(self->model);
PyArrayObject *x;
PyObject *o1,*o2,*result;
if (!PyArg_ParseTuple(args, "O!|d", &PyArray_Type, &x, &cutoff)) return NULL;
if (set_vector_from_array(model->nIn, self->extra_in, x)) return NULL;
lwpr_predict(model,self->extra_in, cutoff, self->extra_out, self->extra_out2, NULL);
o1 = get_array_from_vector(model->nOut, self->extra_out);
o2 = get_array_from_vector(model->nOut, self->extra_out2);
result = Py_BuildValue("(O,O)",o1,o2);
Py_DECREF(o1);
Py_DECREF(o2);
return result;
}
int main(int argc, char** argv)
{
// Instantiate a ModelManager:
ModelManager manager("Test LWPR");
// Parse command-line:
if (manager.parseCommandLine((const int)argc, (const char**)argv, "", 0, 0) == false)
return(1);
manager.start();
double x[2];
double y,yp;
double mse;
FILE *fp;
LWPR_Model model;
int i,j;
/* This allocates some memory and sets initial values
** Note that the model structure itself already exists (on the stack)
*/
lwpr_init_model(&model,2,1,"2D_Cross");
/* Set initial distance metric to 50*(identity matrix) */
lwpr_set_init_D_spherical(&model,50);
/* Set init_alpha to 250 in all elements */
lwpr_set_init_alpha(&model,250);
/* Set w_gen to 0.2 */
model.w_gen = 0.2;
/* See above definition, we either use srand() on Windows or srand48 everywhere else */
SEED_RAND();
for (j=0;j<20;j++) {
mse = 0.0;
for (i=0;i<1000;i++) {
x[0] = 2.0*URAND()-1.0;
x[1] = 2.0*URAND()-1.0;
y = cross(x[0],x[1]) + 0.1*URAND()-0.05;
/* Update the model with one sample
**
** x points to (x[0],x[1]) (input vector)
** &y points to y (output "vector")
** &yp points to yp (prediction "vector")
**
** If you are interested in maximum activation, call
** lwpr_update(&model, x, &y, &yp, &max_w);
*/
lwpr_update(&model, x, &y, &yp, NULL);
mse+=(y-yp)*(y-yp);
}
mse/=500;
printf("#Data = %d #RFS = %d MSE = %f\n",model.n_data, model.sub[0].numRFS, mse);
}
fp = fopen("output.txt","w");
mse = 0.0;
i=0;
for (x[1]=-1.0; x[1]<=1.01; x[1]+=0.05) {
for (x[0]=-1.0; x[0]<=1.01; x[0]+=0.05) {
y = cross(x[0],x[1]);
/* Use the model for predicting an output
**
** x points to (x[0],x[1]) (input vector)
** 0.001 is the cutoff value (clip Gaussian kernel)
** &yp points to yp (prediction "vector")
**
** If you are interested in confidence bounds or
** maximum activation, call
** lwpr_predict(&model, x, 0.001, &yp, &conf, &max_w);
*/
lwpr_predict(&model, x, 0.001, &yp, NULL, NULL);
mse += (y-yp)*(y-yp);
i++;
fprintf(fp,"%8.5f %8.5f %8.5f\n",x[0],x[1],yp);
}
fprintf(fp,"\n\n");
}
fclose(fp);
printf("MSE on test data (%d) = %f\n",i,mse/(double) i);
printf("\nTo view the output, start gnuplot, and type:\n");
printf(" splot \"output.txt\"\n\n");
/* Free the memory that was allocated for receptive fields etc.
** Note again that this does not free the LWPR_Model structure
** itself (but it exists on the stack, so it's automatically free'd) */
lwpr_free_model(&model);
// stop all our ModelComponents
manager.stop();
// all done!
return 0;
}
