SIGSEGV when device_type=cuda #91
Description
I compiled lightgbm with support for GPU and CUDA, and it correctly works (with both CPU and CUDA device type) when used from CLI reading datasets from csv files.
The same lightgbm library correctly works in java using lightgbm4j, for regression tasks, using "device_type=cpu".
Same Java library causes SIGSEGV error when "device_type=cuda".
Below my code:
package ai;
import com.microsoft.ml.lightgbm.PredictionType;
import io.github.metarank.lightgbm4j.LGBMBooster;
import io.github.metarank.lightgbm4j.LGBMDataset;
import io.github.metarank.lightgbm4j.LGBMException;
import java.util.Random;
public class ParametricRegressionExample {
public static void main(String[] args) {
try {
// Generazione del dataset parametrico
int numSamples = 10000; // Numero di campioni
int numFeatures = 3; // Numero di caratteristiche
double slope = 2.0f; // Pendenza della regressione
double intercept = 5.0f; // Intercetta
double noiseLevel = 0.5f; // Livello di rumore casuale
double[] data = generateData(numSamples, numFeatures);
float[] labels = generateLabels(data, numSamples, numFeatures, slope, intercept, noiseLevel);
// Creazione del dataset LightGBM
LGBMDataset dataset = LGBMDataset.createFromMat(data, numSamples, numFeatures, true, "", null);
dataset.setField("label", labels);
// Configurazione dei parametri per il modello di regressione
String parameters = "device_type=cuda\n"
+ "task=train\n"
+ "boosting_type=dart\n"
+ "objective=regression\n"
+ "metric=l2\n"
+ "metric_freq=1\n"
+ "is_training_metric=true\n"
+ "max_bin=255\n"
+ "data=train_data.txt\n"
+ "valid_data=test_data.txt\n"
+ "num_trees=65\n"
+ "learning_rate=0.05\n"
+ "num_leaves=35\n"
+ "device_type=cpu\n"
+ "tree_learner=serial\n"
+ "feature_fraction=0.6\n"
+ "bagging_freq=5\n"
+ "bagging_fraction=0.8\n"
+ "min_data_in_leaf=200\n"
+ "min_sum_hessian_in_leaf=10.0\n"
+ "use_two_round_loading=false\n"
+ "is_save_binary_file=false\n"
+ "output_model=path/to/model\n"
+ "num_machines=1";
// Inizializzazione del booster per l'addestramento
LGBMBooster booster = LGBMBooster.create(dataset, parameters);
// Numero di iterazioni per l'addestramento
int numIterations = 100;
for (int i = 0; i < numIterations; i++) {
booster.updateOneIter();
System.out.println("Iterazione " + (i + 1) + " completata.");
}
// Predizione su nuovi dati
double[] newData = {2.0f, 3.0f, 4.0f};
double[] predictions = booster.predictForMat(newData, 1, numFeatures, true, PredictionType.C_API_PREDICT_NORMAL);
System.out.println("Predizioni per nuovo dato: " + predictions[0]);
// Chiusura delle risorse
booster.close();
dataset.close();
} catch (LGBMException e) {
System.err.println("Errore durante l'esecuzione di LightGBM: " + e.getMessage());
}
}
private static double[] generateData(int numSamples, int numFeatures) {
Random random = new Random();
double[] data = new double[numSamples * numFeatures];
for (int i = 0; i < data.length; i++) {
data[i] = (double) random.nextFloat() * 10;
}
return data;
}
private static float[] generateLabels(double[] data, int numSamples, int numFeatures, double slope, double intercept, double noiseLevel) {
Random random = new Random();
float[] labels = new float[numSamples];
for (int i = 0; i < numSamples; i++) {
double sum = intercept;
for (int j = 0; j < numFeatures; j++) {
sum += slope * data[i * numFeatures + j];
}
labels[i] = (float) (sum + (random.nextGaussian() * noiseLevel));
}
return labels;
}
}