Files
Platform/src/grid/GridSearch.cpp

241 lines
9.7 KiB
C++

#include <iostream>
#include <cstddef>
#include <torch/torch.h>
#include <folding.hpp>
#include "main/Models.h"
#include "common/Paths.h"
#include "common/Colors.h"
#include "common/Utils.h"
#include "GridSearch.h"
namespace platform {
GridSearch::GridSearch(struct ConfigGrid& config) : config(config)
{
if (config.smooth_strategy == "ORIGINAL")
smooth_type = bayesnet::Smoothing_t::ORIGINAL;
else if (config.smooth_strategy == "LAPLACE")
smooth_type = bayesnet::Smoothing_t::LAPLACE;
else if (config.smooth_strategy == "CESTNIK")
smooth_type = bayesnet::Smoothing_t::CESTNIK;
else {
std::cerr << "GridSearch: Unknown smoothing strategy: " << config.smooth_strategy << std::endl;
exit(1);
}
}
json GridSearch::loadResults()
{
std::ifstream file(Paths::grid_output(config.model));
if (file.is_open()) {
return json::parse(file);
}
return json();
}
std::vector<std::string> GridSearch::filterDatasets(Datasets& datasets) const
{
// Load datasets
auto datasets_names = datasets.getNames();
if (config.continue_from != NO_CONTINUE()) {
// Continue previous execution:
if (std::find(datasets_names.begin(), datasets_names.end(), config.continue_from) == datasets_names.end()) {
throw std::invalid_argument("Dataset " + config.continue_from + " not found");
}
// Remove datasets already processed
std::vector<string>::iterator it = datasets_names.begin();
while (it != datasets_names.end()) {
if (*it != config.continue_from) {
it = datasets_names.erase(it);
} else {
if (config.only)
++it;
else
break;
}
}
}
// Exclude datasets
for (const auto& name : config.excluded) {
auto dataset = name.get<std::string>();
auto it = std::find(datasets_names.begin(), datasets_names.end(), dataset);
if (it == datasets_names.end()) {
throw std::invalid_argument("Dataset " + dataset + " already excluded or doesn't exist!");
}
datasets_names.erase(it);
}
return datasets_names;
}
json GridSearch::build_tasks_mpi(int rank)
{
auto tasks = json::array();
auto grid = GridData(Paths::grid_input(config.model));
auto datasets = Datasets(false, Paths::datasets());
auto all_datasets = datasets.getNames();
auto datasets_names = filterDatasets(datasets);
for (int idx_dataset = 0; idx_dataset < datasets_names.size(); ++idx_dataset) {
auto dataset = datasets_names[idx_dataset];
for (const auto& seed : config.seeds) {
auto combinations = grid.getGrid(dataset);
for (int n_fold = 0; n_fold < config.n_folds; n_fold++) {
json task = {
{ "dataset", dataset },
{ "idx_dataset", idx_dataset},
{ "seed", seed },
{ "fold", n_fold},
};
tasks.push_back(task);
}
}
}
// Shuffle the array so heavy datasets are eas ier spread across the workers
std::mt19937 g{ 271 }; // Use fixed seed to obtain the same shuffle
std::shuffle(tasks.begin(), tasks.end(), g);
std::cout << "* Number of tasks: " << tasks.size() << std::endl;
std::cout << separator << std::flush;
for (int i = 0; i < tasks.size(); ++i) {
if ((i + 1) % 10 == 0)
std::cout << separator;
else
std::cout << (i + 1) % 10;
}
std::cout << separator << std::endl << separator << std::flush;
return tasks;
}
void GridSearch::go(struct ConfigMPI& config_mpi)
{
/*
* Each task is a json object with the following structure:
* {
* "dataset": "dataset_name",
* "idx_dataset": idx_dataset, // used to identify the dataset in the results
* // this index is relative to the list of used datasets in the actual run not to the whole datasets list
* "seed": # of seed to use,
* "fold": # of fold to process
* }
*
* This way a task consists in process all combinations of hyperparameters for a dataset, seed and fold
*
* The overall process consists in these steps:
* 0. Create the MPI result type & tasks
* 0.1 Create the MPI result type
* 0.2 Manager creates the tasks
* 1. Manager will broadcast the tasks to all the processes
* 1.1 Broadcast the number of tasks
* 1.2 Broadcast the length of the following string
* 1.2 Broadcast the tasks as a char* string
* 2a. Producer delivers the tasks to the consumers
* 2a.1 Producer will loop to send all the tasks to the consumers and receive the results
* 2a.2 Producer will send the end message to all the consumers
* 2b. Consumers process the tasks and send the results to the producer
* 2b.1 Consumers announce to the producer that they are ready to receive a task
* 2b.2 Consumers receive the task from the producer and process it
* 2b.3 Consumers send the result to the producer
* 3. Manager select the bests scores for each dataset
* 3.1 Loop thru all the results obtained from each outer fold (task) and select the best
* 3.2 Save the results
*/
//
// 0.1 Create the MPI result type
//
Task_Result result;
int tasks_size;
MPI_Datatype MPI_Result;
MPI_Datatype type[5] = { MPI_UNSIGNED, MPI_UNSIGNED, MPI_INT, MPI_DOUBLE, MPI_DOUBLE };
int blocklen[5] = { 1, 1, 1, 1, 1 };
MPI_Aint disp[5];
disp[0] = offsetof(Task_Result, idx_dataset);
disp[1] = offsetof(Task_Result, idx_combination);
disp[2] = offsetof(Task_Result, n_fold);
disp[3] = offsetof(Task_Result, score);
disp[4] = offsetof(Task_Result, time);
MPI_Type_create_struct(5, blocklen, disp, type, &MPI_Result);
MPI_Type_commit(&MPI_Result);
//
// 0.2 Manager creates the tasks
//
char* msg;
json tasks;
if (config_mpi.rank == config_mpi.manager) {
timer.start();
tasks = build_tasks_mpi(config_mpi.rank);
auto tasks_str = tasks.dump();
tasks_size = tasks_str.size();
msg = new char[tasks_size + 1];
strcpy(msg, tasks_str.c_str());
}
//
// 1. Manager will broadcast the tasks to all the processes
//
MPI_Bcast(&tasks_size, 1, MPI_INT, config_mpi.manager, MPI_COMM_WORLD);
if (config_mpi.rank != config_mpi.manager) {
msg = new char[tasks_size + 1];
}
MPI_Bcast(msg, tasks_size + 1, MPI_CHAR, config_mpi.manager, MPI_COMM_WORLD);
tasks = json::parse(msg);
delete[] msg;
auto env = platform::DotEnv();
auto datasets = Datasets(config.discretize, Paths::datasets(), env.get("discretize_algo"));
if (config_mpi.rank == config_mpi.manager) {
//
// 2a. Producer delivers the tasks to the consumers
//
auto datasets_names = filterDatasets(datasets);
json all_results = mpi_search_producer(datasets_names, tasks, config_mpi, MPI_Result);
std::cout << separator << std::endl;
//
// 3. Manager select the bests sccores for each dataset
//
auto results = initializeResults();
select_best_results_folds(results, all_results, config.model);
//
// 3.2 Save the results
//
save(results);
} else {
//
// 2b. Consumers process the tasks and send the results to the producer
//
mpi_search_consumer(datasets, tasks, config, config_mpi, MPI_Result);
}
}
json GridSearch::initializeResults()
{
// Load previous results if continue is set
json results;
if (config.continue_from != NO_CONTINUE()) {
if (!config.quiet)
std::cout << "* Loading previous results" << std::endl;
try {
std::ifstream file(Paths::grid_output(config.model));
if (file.is_open()) {
results = json::parse(file);
results = results["results"];
}
}
catch (const std::exception& e) {
std::cerr << "* There were no previous results" << std::endl;
std::cerr << "* Initizalizing new results" << std::endl;
results = json();
}
}
return results;
}
void GridSearch::save(json& results)
{
std::ofstream file(Paths::grid_output(config.model));
json output = {
{ "model", config.model },
{ "score", config.score },
{ "discretize", config.discretize },
{ "stratified", config.stratified },
{ "n_folds", config.n_folds },
{ "seeds", config.seeds },
{ "date", get_date() + " " + get_time()},
{ "nested", config.nested},
{ "platform", config.platform },
{ "duration", timer.getDurationString(true)},
{ "results", results }
};
file << output.dump(4);
}
} /* namespace platform */