/*
* Fill an svm_problem struct. problem->x will be malloc'd.
*/
void set_problem(struct svm_problem *problem, char *X, char *Y, char *sample_weight, npy_intp *dims, int kernel_type)
{
if (problem == NULL) return;
problem->l = (int) dims[0]; /* number of samples */
problem->y = (double *) Y;
problem->x = dense_to_libsvm((double *) X, dims); /* implicit call to malloc */
problem->W = (double *) sample_weight;
}
/*
* Create and return a svm_problem struct. It is up to the user to free resulting
* structure.
*/
struct svm_problem * set_problem(char *X, char *Y, npy_intp *dims, int kernel_type)
{
struct svm_problem *problem;
problem = (struct svm_problem *) malloc(sizeof(struct svm_problem));
if (problem == NULL) return NULL;
problem->l = (int) dims[0]; /* number of samples */
problem->y = (double *) Y;
problem->x = dense_to_libsvm((double *) X, dims);
if (problem->x == NULL) {
free(problem);
return NULL;
}
return problem;
}
int copy_predict_values(char *predict, struct svm_model *model,
npy_intp *predict_dims, char *dec_values, int nr_class)
{
npy_intp i;
struct svm_node *predict_nodes;
predict_nodes = dense_to_libsvm((double *) predict, predict_dims);
if (predict_nodes == NULL)
return -1;
for(i=0; i<predict_dims[0]; ++i) {
svm_predict_values(model, &predict_nodes[i],
((double *) dec_values) + i*nr_class);
}
free(predict_nodes);
return 0;
}
int copy_predict_proba(char *predict, struct svm_model *model, npy_intp *predict_dims,
char *dec_values)
{
npy_intp i, n, m;
struct svm_node *predict_nodes;
n = predict_dims[0];
m = (npy_intp) model->nr_class;
predict_nodes = dense_to_libsvm((double *) predict, predict_dims);
if (predict_nodes == NULL)
return -1;
for(i=0; i<n; ++i) {
svm_predict_probability(model, &predict_nodes[i],
((double *) dec_values) + i*m);
}
free(predict_nodes);
return 0;
}
/*
* Predict using model.
*
* It will return -1 if we run out of memory.
*/
int copy_predict(char *predict, struct svm_model *model, npy_intp *predict_dims,
char *dec_values)
{
double *t = (double *) dec_values;
struct svm_node *predict_nodes;
npy_intp i;
predict_nodes = dense_to_libsvm((double *) predict, predict_dims);
if (predict_nodes == NULL)
return -1;
for(i=0; i<predict_dims[0]; ++i) {
*t = svm_predict(model, &predict_nodes[i]);
++t;
}
free(predict_nodes);
return 0;
}
/*
* Create and return an instance of svm_model.
*
* The copy of model->sv_coef should be straightforward, but
* unfortunately to represent a matrix numpy and libsvm use different
* approaches, so it requires some iteration.
*
* Possible issue: on 64 bits, the number of columns that numpy can
* store is a long, but libsvm enforces this number (model->l) to be
* an int, so we might have numpy matrices that do not fit into libsvm's
* data structure.
*
*/
struct svm_model *set_model(struct svm_parameter *param, int nr_class,
char *SV, npy_intp *SV_dims,
char *support, npy_intp *support_dims,
npy_intp *sv_coef_strides,
char *sv_coef, char *rho, char *nSV, char *label,
char *probA, char *probB)
{
struct svm_model *model;
double *dsv_coef = (double *) sv_coef;
int i, m;
m = nr_class * (nr_class-1)/2;
if ((model = malloc(sizeof(struct svm_model))) == NULL)
goto model_error;
if ((model->nSV = malloc(nr_class * sizeof(int))) == NULL)
goto nsv_error;
if ((model->label = malloc(nr_class * sizeof(int))) == NULL)
goto label_error;
if ((model->sv_coef = malloc((nr_class-1)*sizeof(double *))) == NULL)
goto sv_coef_error;
if ((model->rho = malloc( m * sizeof(double))) == NULL)
goto rho_error;
model->nr_class = nr_class;
model->param = *param;
model->l = (int) support_dims[0];
if (param->kernel_type == PRECOMPUTED) {
if ((model->SV = malloc ((model->l) * sizeof(struct svm_node))) == NULL)
goto SV_error;
for (i=0; i<model->l; ++i) {
model->SV[i].ind = ((int *) support)[i];
model->SV[i].values = NULL;
}
} else {
model->SV = dense_to_libsvm((double *) SV, SV_dims);
}
/*
* regression and one-class does not use nSV, label.
* TODO: does this provoke memory leaks (we just malloc'ed them)?
*/
if (param->svm_type < 2) {
memcpy(model->nSV, nSV, model->nr_class * sizeof(int));
memcpy(model->label, label, model->nr_class * sizeof(int));
}
for (i=0; i < model->nr_class-1; i++) {
model->sv_coef[i] = dsv_coef + i*(model->l);
}
for (i=0; i<m; ++i) {
(model->rho)[i] = -((double *) rho)[i];
}
/*
* just to avoid segfaults, these features are not wrapped but
* svm_destroy_model will try to free them.
*/
if (param->probability) {
if ((model->probA = malloc(m * sizeof(double))) == NULL)
goto probA_error;
memcpy(model->probA, probA, m * sizeof(double));
if ((model->probB = malloc(m * sizeof(double))) == NULL)
goto probB_error;
memcpy(model->probB, probB, m * sizeof(double));
} else {
model->probA = NULL;
model->probB = NULL;
}
/* We'll free SV ourselves */
model->free_sv = 0;
return model;
probB_error:
free(model->probA);
probA_error:
free(model->SV);
SV_error:
free(model->rho);
rho_error:
free(model->sv_coef);
sv_coef_error:
free(model->label);
label_error:
free(model->nSV);
nsv_error:
free(model);
model_error:
return NULL;
}
