372 lines
17 KiB
C++
372 lines
17 KiB
C++
#include <filesystem>
|
|
#include "data.hpp"
|
|
#include "toolbox.hpp"
|
|
#include "config.hpp"
|
|
#include "toolbox_unit_test.hpp"
|
|
#include "ViolaJones.hpp"
|
|
#include "ViolaJones_device.hpp"
|
|
|
|
#if GPU_BOOSTED
|
|
#include "gpu_unit_test.hpp"
|
|
#define LABEL "GPU"
|
|
#else
|
|
#define LABEL "CPU"
|
|
#endif
|
|
|
|
/**
|
|
* @brief Execute the preprocessing phase
|
|
*
|
|
* The preprocessing phase consist of the following steps :
|
|
* - Load the dataset
|
|
* - Calculate features
|
|
* - Calculate integral images
|
|
* - Apply features to images
|
|
* - Calculate argsort of the featured images
|
|
*
|
|
* @return std::tuple<np::Array<int32_t>, np::Array<uint16_t>, np::Array<uint8_t>, np::Array<int32_t>, np::Array<uint8_t>> Tuple containing in order : training features, training features sorted indexes, training labels, testing features, testing labels
|
|
*/
|
|
std::tuple<np::Array<int32_t>, np::Array<uint16_t>, np::Array<uint8_t>, np::Array<int32_t>, np::Array<uint8_t>> preprocessing(void) {
|
|
// Creating state saver folders if they don't exist already
|
|
if (SAVE_STATE)
|
|
for (const char* const folder_name : { "models", "out" })
|
|
std::filesystem::create_directory(folder_name);
|
|
|
|
const std::chrono::system_clock::time_point preproc_timestamp = perf_counter_ns();
|
|
const std::array<int32_t, 3> preproc_gaps = { 49, -18, 29 };
|
|
header(preproc_gaps, { "Preprocessing", "Time spent (ns)", "Formatted time spent" });
|
|
|
|
const auto [ X_train, y_train, X_test, y_test ] = state_saver<uint8_t, 4>("Loading sets", preproc_gaps[0], { "X_train", "y_train", "X_test", "y_test" },
|
|
FORCE_REDO, SAVE_STATE, OUT_DIR, load_datasets);
|
|
|
|
#if __DEBUG
|
|
printf("X_train\n");
|
|
print(X_train.shape);
|
|
print(X_train, { IDX_INSPECT });
|
|
printf("X_test\n");
|
|
print(X_test.shape);
|
|
print(X_test, { IDX_INSPECT });
|
|
printf("y_train\n");
|
|
print(y_train.shape);
|
|
print(y_train, { IDX_INSPECT, IDX_INSPECT + IDX_INSPECT_OFFSET });
|
|
printf("y_test\n");
|
|
print(y_test.shape);
|
|
print(y_test, { IDX_INSPECT, IDX_INSPECT + IDX_INSPECT_OFFSET });
|
|
#endif
|
|
|
|
const np::Array<uint8_t> feats = state_saver<uint8_t>("Building features", preproc_gaps[0], "feats",
|
|
FORCE_REDO, SAVE_STATE, OUT_DIR, build_features, X_train.shape[1], X_train.shape[2]);
|
|
|
|
#if __DEBUG
|
|
printf("feats\n");
|
|
print(feats.shape);
|
|
print_feat(feats, { IDX_INSPECT });
|
|
#endif
|
|
|
|
const np::Array<uint32_t> X_train_ii = state_saver<uint32_t>("Converting training set to integral images (" LABEL ")", preproc_gaps[0], "X_train_ii_" LABEL,
|
|
FORCE_REDO, SAVE_STATE, OUT_DIR, set_integral_image, X_train);
|
|
const np::Array<uint32_t> X_test_ii = state_saver<uint32_t>("Converting testing set to integral images (" LABEL ")", preproc_gaps[0], "X_test_ii_" LABEL,
|
|
FORCE_REDO, SAVE_STATE, OUT_DIR, set_integral_image, X_test);
|
|
|
|
#if __DEBUG
|
|
printf("X_train_ii\n");
|
|
print(X_train_ii.shape);
|
|
print(X_train_ii, { IDX_INSPECT });
|
|
printf("X_test_ii\n");
|
|
print(X_test_ii.shape);
|
|
print(X_test_ii, { IDX_INSPECT });
|
|
#endif
|
|
|
|
const np::Array<int32_t> X_train_feat = state_saver<int32_t>("Applying features to training set (" LABEL ")", preproc_gaps[0], "X_train_feat_" LABEL,
|
|
FORCE_REDO, SAVE_STATE, OUT_DIR, apply_features, feats, X_train_ii);
|
|
const np::Array<int32_t> X_test_feat = state_saver<int32_t>("Applying features to testing set (" LABEL ")", preproc_gaps[0], "X_test_feat_" LABEL,
|
|
FORCE_REDO, SAVE_STATE, OUT_DIR, apply_features, feats, X_test_ii);
|
|
|
|
#if __DEBUG
|
|
printf("X_train_feat\n");
|
|
print(X_train_feat.shape);
|
|
print(X_train_feat, { IDX_INSPECT, IDX_INSPECT + IDX_INSPECT_OFFSET });
|
|
printf("X_test_feat\n");
|
|
print(X_test_feat.shape);
|
|
print(X_test_feat, { IDX_INSPECT, IDX_INSPECT + IDX_INSPECT_OFFSET });
|
|
#endif
|
|
|
|
// const np::Array<int32_t> indices = state_saver<int32_t>("Selecting best features", preproc_gaps[0], "indices", select_percentile, X_train_feat, d.y_train);
|
|
|
|
#if __DEBUG
|
|
// print_feature(indices);
|
|
#endif
|
|
|
|
const np::Array<uint16_t> X_train_feat_argsort = state_saver<uint16_t>("Precalculating training set argsort (" LABEL ")", preproc_gaps[0], "X_train_feat_argsort_" LABEL,
|
|
FORCE_REDO, SAVE_STATE, OUT_DIR, argsort_2d, X_train_feat);
|
|
|
|
#if __DEBUG
|
|
printf("X_train_feat_argsort\n");
|
|
print(X_train_feat_argsort.shape);
|
|
print(X_train_feat_argsort, { IDX_INSPECT, IDX_INSPECT + IDX_INSPECT_OFFSET });
|
|
#endif
|
|
|
|
const np::Array<uint16_t> X_test_feat_argsort = state_saver<uint16_t>("Precalculating testing set argsort (" LABEL ")", preproc_gaps[0], "X_test_feat_argsort_" LABEL,
|
|
FORCE_REDO, SAVE_STATE, OUT_DIR, argsort_2d, X_test_feat);
|
|
|
|
#if __DEBUG
|
|
printf("X_test_feat_argsort\n");
|
|
print(X_test_feat_argsort.shape);
|
|
print(X_test_feat_argsort, { IDX_INSPECT, IDX_INSPECT + IDX_INSPECT_OFFSET });
|
|
#endif
|
|
const long long time_spent = duration_ns(perf_counter_ns() - preproc_timestamp);
|
|
formatted_line(preproc_gaps, "├", "┼", "─", "┤");
|
|
formatted_row(preproc_gaps, { "Preprocessing summary", thousand_sep(time_spent).c_str(), format_time_ns(time_spent).c_str() });
|
|
footer(preproc_gaps);
|
|
return { X_train_feat, X_train_feat_argsort, y_train, X_test_feat, y_test };
|
|
}
|
|
|
|
/**
|
|
* @brief Train the weak classifiers.
|
|
*
|
|
* @param X_train_feat Training images
|
|
* @param X_train_feat_argsort Sorted indexes of the training images features
|
|
* @param y_train Training labels
|
|
* @return List of trained models
|
|
*/
|
|
std::array<std::array<np::Array<float64_t>, 2>, TS.size()> train(const np::Array<int32_t>& X_train_feat, const np::Array<uint16_t>& X_train_feat_argsort, const np::Array<uint8_t>& y_train) noexcept {
|
|
const std::chrono::system_clock::time_point training_timestamp = perf_counter_ns();
|
|
const std::array<int32_t, 3> training_gaps = { 26, -18, 29 };
|
|
header(training_gaps, { "Training", "Time spent (ns)", "Formatted time spent" });
|
|
|
|
std::array<std::array<np::Array<float64_t>, 2>, TS.size()> models;
|
|
|
|
size_t i = 0;
|
|
for (const size_t T : TS) {
|
|
char title[BUFFER_SIZE] = { 0 };
|
|
char alphas_title[BUFFER_SIZE] = { 0 };
|
|
char final_classifiers_title[BUFFER_SIZE] = { 0 };
|
|
snprintf(title, BUFFER_SIZE, "ViolaJones T = %-4lu (%s)", T, LABEL);
|
|
snprintf(alphas_title, BUFFER_SIZE, "alphas_%lu_%s", T, LABEL);
|
|
snprintf(final_classifiers_title, BUFFER_SIZE, "final_classifiers_%lu_%s", T, LABEL);
|
|
|
|
const auto [ alphas, final_classifiers ] = state_saver<float64_t, 2>(title, training_gaps[0], { alphas_title, final_classifiers_title },
|
|
FORCE_REDO, SAVE_STATE, MODEL_DIR, train_viola_jones, T, X_train_feat, X_train_feat_argsort, y_train);
|
|
#if __DEBUG
|
|
printf("alphas\n");
|
|
print(alphas);
|
|
printf("final_classifiers\n");
|
|
print(final_classifiers);
|
|
#endif
|
|
models[i++] = { alphas, final_classifiers };
|
|
}
|
|
const long long time_spent = duration_ns(perf_counter_ns() - training_timestamp);
|
|
formatted_line(training_gaps, "├", "┼", "─", "┤");
|
|
formatted_row(training_gaps, { "Training summary", thousand_sep(time_spent).c_str(), format_time_ns(time_spent).c_str() });
|
|
footer(training_gaps);
|
|
|
|
return models;
|
|
}
|
|
|
|
/**
|
|
* @brief Benchmark the trained classifiers on the training and testing sets.
|
|
*
|
|
* @param models List of trained models
|
|
* @param X_train_feat Training features
|
|
* @param y_train Training labels
|
|
* @param X_test_feat Testing features
|
|
* @param y_test Testing labels
|
|
*/
|
|
void testing_and_evaluating(const std::array<std::array<np::Array<float64_t>, 2>, TS.size()>& models, const np::Array<int32_t>& X_train_feat, const np::Array<uint8_t>& y_train, const np::Array<int32_t>& X_test_feat, const np::Array<uint8_t>& y_test) {
|
|
const std::array<int32_t, 5> testing_gaps = { 26, -19, 24, -19, 24 };
|
|
header(testing_gaps, { "Testing", "Time spent (ns) (E)", "Formatted time spent (E)", "Time spent (ns) (T)", "Formatted time spent (T)" });
|
|
std::array<std::array<float64_t, 8>, TS.size()> results;
|
|
|
|
size_t i = 0;
|
|
long long total_train_timestamp = 0;
|
|
long long total_test_timestamp = 0;
|
|
for (const auto& [ alphas, final_classifiers ] : models) {
|
|
char title[BUFFER_SIZE] = { 0 };
|
|
snprintf(title, BUFFER_SIZE, "ViolaJones T = %-4i (%s)", TS[i], LABEL);
|
|
|
|
std::chrono::system_clock::time_point start = perf_counter_ns();
|
|
const np::Array<uint8_t> y_pred_train = classify_viola_jones(alphas, final_classifiers, X_train_feat);
|
|
const long long t_pred_train = duration_ns(perf_counter_ns() - start);
|
|
total_train_timestamp += t_pred_train;
|
|
const float64_t e_acc = accuracy_score(y_train, y_pred_train);
|
|
const float64_t e_f1 = f1_score(y_train, y_pred_train);
|
|
float64_t e_FN, e_FP;
|
|
std::tie(std::ignore, e_FN, e_FP, std::ignore) = confusion_matrix(y_train, y_pred_train);
|
|
|
|
start = perf_counter_ns();
|
|
const np::Array<uint8_t> y_pred_test = classify_viola_jones(alphas, final_classifiers, X_test_feat);
|
|
const long long t_pred_test = duration_ns(perf_counter_ns() - start);
|
|
total_test_timestamp += t_pred_test;
|
|
const float64_t t_acc = accuracy_score(y_test, y_pred_test);
|
|
const float64_t t_f1 = f1_score(y_test, y_pred_test);
|
|
float64_t t_FN, t_FP;
|
|
std::tie(std::ignore, t_FN, t_FP, std::ignore) = confusion_matrix(y_test, y_pred_test);
|
|
results[i++] = { e_acc, e_f1, e_FN, e_FP, t_acc, t_f1, t_FN, t_FP };
|
|
|
|
formatted_row(testing_gaps, { title, thousand_sep(t_pred_train).c_str(), format_time_ns(t_pred_train).c_str(), thousand_sep(t_pred_test).c_str(), format_time_ns(t_pred_test).c_str() });
|
|
}
|
|
formatted_line(testing_gaps, "├", "┼", "─", "┤");
|
|
formatted_row(testing_gaps, { "Testing summary", thousand_sep(total_train_timestamp).c_str(), format_time_ns(total_train_timestamp).c_str(), thousand_sep(total_test_timestamp).c_str(), format_time_ns(total_test_timestamp).c_str() });
|
|
footer(testing_gaps);
|
|
|
|
const std::array<int32_t, 9> evaluating_gaps = { 19, -7, -6, -6, -6, -7, -6, -6, -6 };
|
|
header(evaluating_gaps, { "Evaluating", "ACC (E)", "F1 (E)", "FN (E)", "FP (E)", "ACC (T)", "F1 (T)", "FN (T)", "FP (T)"});
|
|
|
|
i = 0;
|
|
for (const size_t T : TS) {
|
|
char title[BUFFER_SIZE] = { 0 };
|
|
snprintf(title, BUFFER_SIZE, "ViolaJones T = %-4lu", T);
|
|
const auto [e_acc, e_f1, e_FN, e_FP, t_acc, t_f1, t_FN, t_FP] = results[i++];
|
|
printf("│ %-19s │ %'6.2f%% │ %'6.2f │ %'6.0f │ %'6.0f │ %6.2f%% │ %'6.2f │ %'6.0f │ %'6.0f │\n", title, e_acc * 100, e_f1, e_FN, e_FP, t_acc * 100, t_f1, t_FN, t_FP);
|
|
}
|
|
footer(evaluating_gaps);
|
|
}
|
|
|
|
/**
|
|
* @brief Test if the each result is equals to other devices.
|
|
*
|
|
* Given ViolaJones is a fully deterministic algorithm. The results, regardless the device, should be the same,
|
|
* this function check this assertion.
|
|
*/
|
|
void unit_test(void) {
|
|
const std::chrono::system_clock::time_point unit_timestamp = perf_counter_ns();
|
|
const std::array<int32_t, 4> unit_gaps = { 37, -10, -18, 29};
|
|
header(unit_gaps, { "Unit testing", "Test state", "Time spent (ns)", "Formatted time spent" });
|
|
|
|
char title[BUFFER_SIZE] = { 0 };
|
|
char tmp_title[BUFFER_SIZE / 2] = { 0 };
|
|
char file_cpu[BUFFER_SIZE] = { 0 };
|
|
char file_gpu[BUFFER_SIZE] = { 0 };
|
|
uint64_t n_total = 0, n_success = 0;
|
|
|
|
const auto test_fnc = [&unit_gaps, &n_total, &n_success](const char* const title, const auto& fnc) noexcept {
|
|
++n_total;
|
|
const std::chrono::system_clock::time_point start = perf_counter_ns();
|
|
const bool state = fnc();
|
|
const long long time_spent = duration_ns(perf_counter_ns() - start);
|
|
if(state){
|
|
formatted_row(unit_gaps, { title, "Passed", thousand_sep(time_spent).c_str(), format_time_ns(time_spent).c_str() });
|
|
++n_success;
|
|
} else
|
|
formatted_row(unit_gaps, { title, "Failed", thousand_sep(time_spent).c_str(), format_time_ns(time_spent).c_str() });
|
|
};
|
|
|
|
for (const char* const label : { "train", "test" }) {
|
|
snprintf(file_cpu, BUFFER_SIZE, OUT_DIR "/X_%s_ii_CPU.bin", label);
|
|
snprintf(file_gpu, BUFFER_SIZE, OUT_DIR "/X_%s_ii_GPU.bin", label);
|
|
if (std::filesystem::exists(file_cpu) && std::filesystem::exists(file_gpu)) {
|
|
snprintf(tmp_title, BUFFER_SIZE / 2, "X_%s_ii", label);
|
|
snprintf(title, BUFFER_SIZE, "%-22s - CPU vs GPU", tmp_title);
|
|
test_fnc(title, [&file_cpu, &file_gpu]{
|
|
const np::Array<uint32_t> X_train_ii_cpu = load<uint32_t>(file_cpu);
|
|
const np::Array<uint32_t> X_train_ii_gpu = load<uint32_t>(file_gpu);
|
|
return unit_test_cpu_vs_gpu<uint32_t>(X_train_ii_cpu, X_train_ii_gpu);
|
|
});
|
|
}
|
|
snprintf(file_cpu, BUFFER_SIZE, OUT_DIR "/X_%s_feat_CPU.bin", label);
|
|
snprintf(file_gpu, BUFFER_SIZE, OUT_DIR "/X_%s_feat_GPU.bin", label);
|
|
uint8_t feat = 0;
|
|
char file_feat[BUFFER_SIZE] = { 0 };
|
|
if (std::filesystem::exists(file_cpu)) {
|
|
strncpy(file_feat, file_cpu, BUFFER_SIZE);
|
|
feat = 1;
|
|
} else if (std::filesystem::exists(file_gpu)) {
|
|
strncpy(file_feat, file_gpu, BUFFER_SIZE);
|
|
feat = 2;
|
|
}
|
|
if (feat != 0) {
|
|
const np::Array<int32_t> X_feat = load<int32_t>(file_feat);
|
|
snprintf(file_gpu, BUFFER_SIZE, feat == 1 ? OUT_DIR "/X_%s_feat_GPU.bin" : OUT_DIR "/X_%s_feat_CPU.bin", label);
|
|
if (std::filesystem::exists(file_gpu)) {
|
|
snprintf(tmp_title, BUFFER_SIZE / 2, "X_%s_feat", label);
|
|
snprintf(title, BUFFER_SIZE, "%-22s - CPU vs GPU", tmp_title);
|
|
test_fnc(title, [&X_feat, &file_gpu]{
|
|
const np::Array<int32_t> X_feat_aux = load<int32_t>(file_gpu);
|
|
return unit_test_cpu_vs_gpu<int32_t>(X_feat, X_feat_aux);
|
|
});
|
|
}
|
|
snprintf(file_cpu, BUFFER_SIZE, OUT_DIR "/X_%s_feat_argsort_CPU.bin", label);
|
|
np::Array<uint16_t> X_feat_argsort_cpu;
|
|
uint8_t loaded = 0;
|
|
if (std::filesystem::exists(file_cpu)) {
|
|
++loaded;
|
|
snprintf(tmp_title, BUFFER_SIZE / 2, "X_%s_feat_argsort", label);
|
|
snprintf(title, BUFFER_SIZE, "%-22s - CPU argsort", tmp_title);
|
|
test_fnc(title, [&X_feat, &X_feat_argsort_cpu, &file_cpu]{
|
|
X_feat_argsort_cpu = load<uint16_t>(file_cpu);
|
|
return unit_test_argsort_2d<int32_t>(X_feat, X_feat_argsort_cpu);
|
|
});
|
|
}
|
|
snprintf(file_gpu, BUFFER_SIZE, OUT_DIR "/X_%s_feat_argsort_GPU.bin", label);
|
|
np::Array<uint16_t> X_feat_argsort_gpu;
|
|
if (std::filesystem::exists(file_gpu)) {
|
|
++loaded;
|
|
snprintf(tmp_title, BUFFER_SIZE / 2, "X_%s_feat_argsort", label);
|
|
snprintf(title, BUFFER_SIZE, "%-22s - GPU argsort", tmp_title);
|
|
test_fnc(title, [&X_feat, &X_feat_argsort_gpu, &file_gpu]{
|
|
X_feat_argsort_gpu = load<uint16_t>(file_gpu);
|
|
return unit_test_argsort_2d<int32_t>(X_feat, X_feat_argsort_gpu);
|
|
});
|
|
}
|
|
if (loaded == 2){
|
|
snprintf(tmp_title, BUFFER_SIZE / 2, "X_%s_feat_argsort", label);
|
|
snprintf(title, BUFFER_SIZE, "%-22s - CPU vs GPU", tmp_title);
|
|
test_fnc(title, [&X_feat_argsort_cpu, &X_feat_argsort_gpu]{ return unit_test_cpu_vs_gpu<uint16_t>(X_feat_argsort_cpu, X_feat_argsort_gpu); });
|
|
}
|
|
}
|
|
}
|
|
|
|
for (const size_t T : TS)
|
|
for (const char* const label : { "alphas", "final_classifiers" }) {
|
|
snprintf(file_cpu, BUFFER_SIZE, MODEL_DIR "/%s_%lu_CPU.bin", label, T);
|
|
snprintf(file_gpu, BUFFER_SIZE, MODEL_DIR "/%s_%lu_GPU.bin", label, T);
|
|
if (std::filesystem::exists(file_cpu) && std::filesystem::exists(file_gpu)){
|
|
snprintf(tmp_title, BUFFER_SIZE / 2, "%s_%ld", label, T);
|
|
snprintf(title, BUFFER_SIZE, "%-22s - CPU vs GPU", tmp_title);
|
|
test_fnc(title, [&file_cpu, &file_gpu]{
|
|
const np::Array<float64_t> cpu = load<float64_t>(file_cpu);
|
|
const np::Array<float64_t> gpu = load<float64_t>(file_gpu);
|
|
return unit_test_cpu_vs_gpu<float64_t>(cpu, gpu);
|
|
});
|
|
}
|
|
}
|
|
|
|
const long long time_spent = duration_ns(perf_counter_ns() - unit_timestamp);
|
|
|
|
if (n_total == 0)
|
|
formatted_row(unit_gaps, { "Unit testing summary", "No files", thousand_sep(time_spent).c_str(), format_time_ns(time_spent).c_str() });
|
|
else {
|
|
snprintf(title, BUFFER_SIZE, "%ld/%ld", n_success, n_total);
|
|
formatted_line(unit_gaps, "├", "┼", "─", "┤");
|
|
formatted_row(unit_gaps, { "Unit testing summary", title, thousand_sep(time_spent).c_str(), format_time_ns(time_spent).c_str() });
|
|
}
|
|
footer(unit_gaps);
|
|
}
|
|
|
|
int32_t main(void){
|
|
setlocale(LC_NUMERIC, ""); // Allow proper number display
|
|
|
|
const std::chrono::system_clock::time_point unit_timestamp = perf_counter_ns();
|
|
const std::array<int32_t, 3> unit_gaps = { 27, -18, 29 };
|
|
header(unit_gaps, { "Unit testing", "Time spent (ns)", "Formatted time spent" });
|
|
#if GPU_BOOSTED
|
|
benchmark_function_void("Testing GPU capabilities 1D", unit_gaps[0], test_working, 50000);
|
|
benchmark_function_void("Testing GPU capabilities 2D", unit_gaps[0], test_working_2d, 200, 500);
|
|
benchmark_function_void("Testing GPU capabilities 3D", unit_gaps[0], test_working_3d, 30, 40, 500);
|
|
#endif
|
|
benchmark_function_void("Testing format_time", unit_gaps[0], format_time_test);
|
|
benchmark_function_void("Testing format_time_ns", unit_gaps[0], format_time_ns_test);
|
|
benchmark_function_void("Testing format_byte_size", unit_gaps[0], format_byte_size_test);
|
|
benchmark_function_void("Testing thousand_sep", unit_gaps[0], thousand_sep_test);
|
|
const long long time_spent = duration_ns(perf_counter_ns() - unit_timestamp);
|
|
formatted_line(unit_gaps, "├", "┼", "─", "┤");
|
|
formatted_row(unit_gaps, { "Unit testing summary", thousand_sep(time_spent).c_str(), format_time_ns(time_spent).c_str() });
|
|
footer(unit_gaps);
|
|
|
|
const auto [ X_train_feat, X_train_feat_argsort, y_train, X_test_feat, y_test ] = preprocessing();
|
|
const std::array<std::array<np::Array<float64_t>, 2>, TS.size()> models = train(X_train_feat, X_train_feat_argsort, y_train);
|
|
testing_and_evaluating(models, X_train_feat, y_train, X_test_feat, y_test);
|
|
unit_test();
|
|
|
|
return EXIT_SUCCESS;
|
|
}
|