void
closestColumnsAndDistances(
const MappedMatrix& inMatrix,
const MappedColumnVector& inVector,
DistanceFunction& inMetric,
RandomAccessIterator ioFirst,
RandomAccessIterator ioLast) {
ReverseLexicographicComparator<
typename std::iterator_traits<RandomAccessIterator>::value_type>
comparator;
std::fill(ioFirst, ioLast,
std::make_tuple(0, std::numeric_limits<double>::infinity()));
for (Index i = 0; i < inMatrix.cols(); ++i) {
double currentDist
= AnyType_cast<double>(
inMetric(MappedColumnVector(inMatrix.col(i)), inVector)
);
// outIndicesAndDistances is a heap, so the first element is maximal
if (currentDist < std::get<1>(*ioFirst)) {
// Unfortunately, the STL does not have a decrease-key function,
// so we are wasting a bit of performance here
std::pop_heap(ioFirst, ioLast, comparator);
*(ioLast - 1) = std::make_tuple(i, currentDist);
std::push_heap(ioFirst, ioLast, comparator);
}
}
std::sort_heap(ioFirst, ioLast, comparator);
}
AnyType
get_row_from_2d_array::run(AnyType & args) {
MappedMatrix input = args[0].getAs<MappedMatrix>();
int index = args[1].getAs<int>() - 1; // database index starts from 1
if (index < 0 or index >= input.cols()) {
std::stringstream err_msg;
err_msg << "Out-of-bound index: " << index << " >= " << input.cols();
throw std::runtime_error(err_msg.str());
}
MutableNativeColumnVector ret(this->allocateArray<double>(input.rows()));
ret = input.col(static_cast<Index>(index));
return ret;
}
AnyType matrix_vec_mult_in_mem_2d::run(AnyType & args){
MappedColumnVector vec = args[0].getAs<MappedColumnVector>();
MappedMatrix mat = args[1].getAs<MappedMatrix>();
// Note mat is constructed in the column-first order
// which means that mat is actually transposed
if(vec.size() != mat.cols()){
throw std::invalid_argument(
"dimensions mismatch: vec.size() != matrix.rows()");
};
// trans(vec) * trans(mat) = mat * vec
Matrix r = mat * vec;
ColumnVector v = r.col(0);
return v;
}
AnyType matrix_vec_mult_in_mem_1d::run(AnyType & args){
MappedColumnVector vec1 = args[0].getAs<MappedColumnVector>();
// matrix stored as a 1-d array
MappedColumnVector vec2 = args[1].getAs<MappedColumnVector>();
if(vec2.size() % vec1.size() != 0){
throw std::invalid_argument(
"dimensions mismatch: matrix.size() is not multiples of vec.size()");
};
MappedMatrix mat;
// the rebinding happens in column-major
mat.rebind(vec2.memoryHandle(), vec1.size(), vec2.size()/vec1.size());
Matrix r = trans(mat) * vec1;
ColumnVector v = r.col(0);
return v;
}
std::tuple<Index, double>
closestColumnAndDistance(
const MappedMatrix& inMatrix,
const MappedColumnVector& inVector,
DistanceFunction& inMetric) {
Index closestColumn = 0;
double minDist = std::numeric_limits<double>::infinity();
for (Index i = 0; i < inMatrix.cols(); ++i) {
double currentDist
= AnyType_cast<double>(
inMetric(MappedColumnVector(inMatrix.col(i)), inVector)
);
if (currentDist < minDist) {
closestColumn = i;
minDist = currentDist;
}
}
return std::tuple<Index, double>(closestColumn, minDist);
}
/*
* Permute each categorical variable and predict
*/
AnyType
rf_cat_imp_score::run(AnyType &args) {
if (args[0].isNull() || args[7].isNull()) { return Null(); }
Tree dt = args[0].getAs<ByteString>();
MutableNativeIntegerVector cat_features;
NativeColumnVector con_features;
try {
if (args[1].isNull()){
// no cat features
return Null();
}
else {
MutableNativeIntegerVector xx_cat = args[1].getAs<MutableNativeIntegerVector>();
cat_features.rebind(xx_cat.memoryHandle(), xx_cat.size());
}
if (args[2].isNull()){
con_features.rebind(this->allocateArray<double>(0));
}
else {
NativeColumnVector xx_con = args[2].getAs<NativeColumnVector>();
con_features.rebind(xx_con.memoryHandle(), xx_con.size());
}
} catch (const ArrayWithNullException &e) {
// not expect to reach here
// if max_surr = 0, nulls are filtered
// otherwise, mapped to -1 or NaN
return Null();
}
MappedIntegerVector cat_n_levels = args[3].getAs<MappedIntegerVector>();
int n_permutations = args[4].getAs<int>();
double y = args[5].getAs<double>();
bool is_classification = args[6].getAs<bool>();
MappedMatrix distributions = args[7].getAs<MappedMatrix>();
// returning
MutableNativeColumnVector permuted_predictions(
this->allocateArray<double>(cat_n_levels.size()));
// permute each and predict
NativeRandomNumberGenerator generator;
for (int p = 0; p < n_permutations; p ++) {
for (Index i = 0; i < cat_n_levels.size(); i ++) {
int orig_i = cat_features(i);
discrete_distribution<> ddist(distributions.col(i).data(),
distributions.col(i).data() + cat_n_levels(i) + 1);
variate_generator<NativeRandomNumberGenerator, discrete_distribution<> >
rvt(generator, ddist);
cat_features(i) = rvt() - 1;
// calling NativeIntegerVector for a const cast
// see EigenIntegration_impl.hpp in ports for details
double prediction = dt.predict_response(
NativeIntegerVector(cat_features.memoryHandle()), con_features);
double score = 0.;
if (is_classification) {
score = y - prediction < 1e-3 ? 1. : 0.;
} else {
score = - (y - prediction) * (y - prediction);
}
permuted_predictions(i) += score;
cat_features(i) = orig_i;
}
}
permuted_predictions /= n_permutations;
return permuted_predictions;
}
/*
* Permute each continuous variable and predict
*/
AnyType
rf_con_imp_score::run(AnyType &args) {
if (args[0].isNull() || args[7].isNull()) { return Null(); }
Tree dt = args[0].getAs<ByteString>();
NativeIntegerVector cat_features;
MutableNativeColumnVector con_features;
try {
if (args[1].isNull()){
// no cat features
cat_features.rebind(this->allocateArray<int>(0));
}
else {
NativeIntegerVector xx_cat = args[1].getAs<NativeIntegerVector>();
cat_features.rebind(xx_cat.memoryHandle(), xx_cat.size());
}
if (args[2].isNull()){
//no con features
return Null();
}
else {
MutableNativeColumnVector xx_con = args[2].getAs<MutableNativeColumnVector>();
con_features.rebind(xx_con.memoryHandle(), xx_con.size());
}
} catch (const ArrayWithNullException &e) {
// not expect to reach here
// if max_surr = 0, nulls are filtered
// otherwise, mapped to -1 or NaN
return Null();
}
// con_splits size = num_con_features x num_bins
// When num_con_features = 0, the input will be an empty string that is read
// as a ByteString
ConSplitsResult<RootContainer> splits_results = args[3].getAs<ByteString>();
int n_permutations = args[4].getAs<int>();
double y = args[5].getAs<double>();
bool is_classification = args[6].getAs<bool>();
MappedMatrix distributions = args[7].getAs<MappedMatrix>();
// returning
MutableNativeColumnVector permuted_predictions(
this->allocateArray<double>(con_features.size()));
// permute each and predict
NativeRandomNumberGenerator generator;
for (int p = 0; p < n_permutations; p ++) {
for (Index i = 0; i < con_features.size(); i ++) {
double orig_i = con_features(i);
discrete_distribution<> ddist(distributions.col(i).data(),
distributions.col(i).data() + distributions.rows());
variate_generator<NativeRandomNumberGenerator, discrete_distribution<> >
rvt(generator, ddist);
int outcome = rvt();
if (outcome == 0) {
con_features(i) = std::numeric_limits<double>::quiet_NaN();
} else if (outcome == static_cast<int>(distributions.rows()) - 1) {
// bin value that is larger than the last separator (last value in con_splits)
con_features(i) = splits_results.con_splits(i, outcome-2) + 1.;
} else {
con_features(i) = splits_results.con_splits(i, outcome-1);
}
// calling NativeColumnVector for a const cast
// see EigenIntegration_impl.hpp in ports for details
double prediction = dt.predict_response(
cat_features, NativeColumnVector(con_features.memoryHandle()));
double score = 0.;
if (is_classification) {
score = y - prediction < 1e-3 ? 1. : 0.;
} else {
score = - (y - prediction) * (y - prediction);
}
permuted_predictions(i) += score;
con_features(i) = orig_i;
}
}
permuted_predictions /= n_permutations;
return permuted_predictions;
}