BayesNet/tests/TestFolding.cc

96 lines
4.0 KiB
C++

#include <catch2/catch_test_macros.hpp>
#include <catch2/catch_approx.hpp>
#include <catch2/generators/catch_generators.hpp>
#include "TestUtils.h"
#include "Folding.h"
TEST_CASE("KFold Test", "[Platform][KFold]")
{
// Initialize a KFold object with k=5 and a seed of 19.
std::string file_name = GENERATE("glass", "iris", "ecoli", "diabetes");
auto raw = RawDatasets(file_name, true);
int nFolds = 5;
platform::KFold kfold(nFolds, raw.nSamples, 19);
int number = raw.nSamples * (kfold.getNumberOfFolds() - 1) / kfold.getNumberOfFolds();
SECTION("Number of Folds")
{
REQUIRE(kfold.getNumberOfFolds() == nFolds);
}
SECTION("Fold Test")
{
// Test each fold's size and contents.
for (int i = 0; i < nFolds; ++i) {
auto [train_indices, test_indices] = kfold.getFold(i);
bool result = train_indices.size() == number || train_indices.size() == number + 1;
REQUIRE(result);
REQUIRE(train_indices.size() + test_indices.size() == raw.nSamples);
}
}
}
map<int, int> counts(std::vector<int> y, std::vector<int> indices)
{
map<int, int> result;
for (auto i = 0; i < indices.size(); ++i) {
result[y[indices[i]]]++;
}
return result;
}
TEST_CASE("StratifiedKFold Test", "[Platform][StratifiedKFold]")
{
// Initialize a StratifiedKFold object with k=3, using the y std::vector, and a seed of 17.
std::string file_name = GENERATE("glass", "iris", "ecoli", "diabetes");
int nFolds = GENERATE(3, 5, 10);
auto raw = RawDatasets(file_name, true);
platform::StratifiedKFold stratified_kfoldt(nFolds, raw.yt, 17);
platform::StratifiedKFold stratified_kfoldv(nFolds, raw.yv, 17);
int number = raw.nSamples * (stratified_kfoldt.getNumberOfFolds() - 1) / stratified_kfoldt.getNumberOfFolds();
SECTION("Stratified Number of Folds")
{
REQUIRE(stratified_kfoldt.getNumberOfFolds() == nFolds);
}
SECTION("Stratified Fold Test")
{
// Test each fold's size and contents.
auto counts = map<int, std::vector<int>>();
// Initialize the counts per Fold
for (int i = 0; i < nFolds; ++i) {
counts[i] = std::vector<int>(raw.classNumStates, 0);
}
// Check fold and compute counts of each fold
for (int fold = 0; fold < nFolds; ++fold) {
auto [train_indicest, test_indicest] = stratified_kfoldt.getFold(fold);
auto [train_indicesv, test_indicesv] = stratified_kfoldv.getFold(fold);
REQUIRE(train_indicest == train_indicesv);
REQUIRE(test_indicest == test_indicesv);
// In the worst case scenario, the number of samples in the training set is number + raw.classNumStates
// because in that fold can come one remainder sample from each class.
REQUIRE(train_indicest.size() <= number + raw.classNumStates);
// If the number of samples in any class is less than the number of folds, then the fold is faulty.
// and the number of samples in the training set + test set will be less than nSamples
if (!stratified_kfoldt.isFaulty()) {
REQUIRE(train_indicest.size() + test_indicest.size() == raw.nSamples);
} else {
REQUIRE(train_indicest.size() + test_indicest.size() <= raw.nSamples);
}
auto train_t = torch::tensor(train_indicest);
auto ytrain = raw.yt.index({ train_t });
// Check that the class labels have been equally assign to each fold
for (const auto& idx : train_indicest) {
counts[fold][raw.yt[idx].item<int>()]++;
}
}
// Test the fold counting of every class
for (int fold = 0; fold < nFolds; ++fold) {
for (int j = 1; j < nFolds - 1; ++j) {
for (int k = 0; k < raw.classNumStates; ++k) {
REQUIRE(abs(counts.at(fold).at(k) - counts.at(j).at(k)) <= 1);
}
}
}
}
}