use of org.apache.ignite.ml.nn.MLPTrainer in project ignite by apache.
the class MLPTrainerExample method main.
/**
* Executes example.
*
* @param args Command line arguments, none required.
*/
public static void main(String[] args) {
// IMPL NOTE based on MLPGroupTrainerTest#testXOR
System.out.println(">>> Distributed multilayer perceptron example started.");
// Start ignite grid.
try (Ignite ignite = Ignition.start("examples/config/example-ignite.xml")) {
System.out.println(">>> Ignite grid started.");
// Create cache with training data.
CacheConfiguration<Integer, LabeledVector<double[]>> trainingSetCfg = new CacheConfiguration<>();
trainingSetCfg.setName("TRAINING_SET");
trainingSetCfg.setAffinity(new RendezvousAffinityFunction(false, 10));
IgniteCache<Integer, LabeledVector<double[]>> trainingSet = null;
try {
trainingSet = ignite.createCache(trainingSetCfg);
// Fill cache with training data.
trainingSet.put(0, new LabeledVector<>(VectorUtils.of(0, 0), new double[] { 0 }));
trainingSet.put(1, new LabeledVector<>(VectorUtils.of(0, 1), new double[] { 1 }));
trainingSet.put(2, new LabeledVector<>(VectorUtils.of(1, 0), new double[] { 1 }));
trainingSet.put(3, new LabeledVector<>(VectorUtils.of(1, 1), new double[] { 0 }));
// Define a layered architecture.
MLPArchitecture arch = new MLPArchitecture(2).withAddedLayer(10, true, Activators.RELU).withAddedLayer(1, false, Activators.SIGMOID);
// Define a neural network trainer.
MLPTrainer<SimpleGDParameterUpdate> trainer = new MLPTrainer<>(arch, LossFunctions.MSE, new UpdatesStrategy<>(new SimpleGDUpdateCalculator(0.1), SimpleGDParameterUpdate.SUM_LOCAL, SimpleGDParameterUpdate.AVG), 3000, 4, 50, 123L);
// Train neural network and get multilayer perceptron model.
MultilayerPerceptron mlp = trainer.fit(ignite, trainingSet, new LabeledDummyVectorizer<>());
int totalCnt = 4;
int failCnt = 0;
// Calculate score.
for (int i = 0; i < 4; i++) {
LabeledVector<double[]> pnt = trainingSet.get(i);
Matrix predicted = mlp.predict(new DenseMatrix(new double[][] { { pnt.features().get(0), pnt.features().get(1) } }));
double predictedVal = predicted.get(0, 0);
double lbl = pnt.label()[0];
System.out.printf(">>> key: %d\t\t predicted: %.4f\t\tlabel: %.4f\n", i, predictedVal, lbl);
failCnt += Math.abs(predictedVal - lbl) < 0.5 ? 0 : 1;
}
double failRatio = (double) failCnt / totalCnt;
System.out.println("\n>>> Fail percentage: " + (failRatio * 100) + "%.");
System.out.println("\n>>> Distributed multilayer perceptron example completed.");
} finally {
trainingSet.destroy();
}
} finally {
System.out.flush();
}
}
use of org.apache.ignite.ml.nn.MLPTrainer in project ignite by apache.
the class StackingTest method testSimpleStack.
/**
* Tests simple stack training.
*/
@Test
public void testSimpleStack() {
StackedDatasetTrainer<Vector, Vector, Double, LinearRegressionModel, Double> trainer = new StackedDatasetTrainer<>();
UpdatesStrategy<SmoothParametrized, SimpleGDParameterUpdate> updatesStgy = new UpdatesStrategy<>(new SimpleGDUpdateCalculator(0.2), SimpleGDParameterUpdate.SUM_LOCAL, SimpleGDParameterUpdate.AVG);
MLPArchitecture arch = new MLPArchitecture(2).withAddedLayer(10, true, Activators.RELU).withAddedLayer(1, false, Activators.SIGMOID);
MLPTrainer<SimpleGDParameterUpdate> trainer1 = new MLPTrainer<>(arch, LossFunctions.MSE, updatesStgy, 3000, 10, 50, 123L);
// Convert model trainer to produce Vector -> Vector model
DatasetTrainer<AdaptableDatasetModel<Vector, Vector, Matrix, Matrix, MultilayerPerceptron>, Double> mlpTrainer = AdaptableDatasetTrainer.of(trainer1).beforeTrainedModel((Vector v) -> new DenseMatrix(v.asArray(), 1)).afterTrainedModel((Matrix mtx) -> mtx.getRow(0)).withConvertedLabels(VectorUtils::num2Arr);
final double factor = 3;
StackedModel<Vector, Vector, Double, LinearRegressionModel> mdl = trainer.withAggregatorTrainer(new LinearRegressionLSQRTrainer().withConvertedLabels(x -> x * factor)).addTrainer(mlpTrainer).withAggregatorInputMerger(VectorUtils::concat).withSubmodelOutput2VectorConverter(IgniteFunction.identity()).withVector2SubmodelInputConverter(IgniteFunction.identity()).withOriginalFeaturesKept(IgniteFunction.identity()).withEnvironmentBuilder(TestUtils.testEnvBuilder()).fit(getCacheMock(xor), parts, new DoubleArrayVectorizer<Integer>().labeled(Vectorizer.LabelCoordinate.LAST));
assertEquals(0.0 * factor, mdl.predict(VectorUtils.of(0.0, 0.0)), 0.3);
assertEquals(1.0 * factor, mdl.predict(VectorUtils.of(0.0, 1.0)), 0.3);
assertEquals(1.0 * factor, mdl.predict(VectorUtils.of(1.0, 0.0)), 0.3);
assertEquals(0.0 * factor, mdl.predict(VectorUtils.of(1.0, 1.0)), 0.3);
}
use of org.apache.ignite.ml.nn.MLPTrainer in project ignite by apache.
the class LinearRegressionSGDTrainer method updateModel.
/**
* {@inheritDoc}
*/
@Override
protected <K, V> LinearRegressionModel updateModel(LinearRegressionModel mdl, DatasetBuilder<K, V> datasetBuilder, Preprocessor<K, V> extractor) {
assert updatesStgy != null;
IgniteFunction<Dataset<EmptyContext, SimpleLabeledDatasetData>, MLPArchitecture> archSupplier = dataset -> {
int cols = dataset.compute(data -> {
if (data.getFeatures() == null)
return null;
return data.getFeatures().length / data.getRows();
}, (a, b) -> {
if (a == null)
return b == null ? 0 : b;
if (b == null)
return a;
return b;
});
MLPArchitecture architecture = new MLPArchitecture(cols);
architecture = architecture.withAddedLayer(1, true, Activators.LINEAR);
return architecture;
};
MLPTrainer<?> trainer = new MLPTrainer<>(archSupplier, LossFunctions.MSE, updatesStgy, maxIterations, batchSize, locIterations, seed);
IgniteFunction<LabeledVector<Double>, LabeledVector<double[]>> func = lv -> new LabeledVector<>(lv.features(), new double[] { lv.label() });
PatchedPreprocessor<K, V, Double, double[]> patchedPreprocessor = new PatchedPreprocessor<>(func, extractor);
MultilayerPerceptron mlp = Optional.ofNullable(mdl).map(this::restoreMLPState).map(m -> trainer.update(m, datasetBuilder, patchedPreprocessor)).orElseGet(() -> trainer.fit(datasetBuilder, patchedPreprocessor));
double[] p = mlp.parameters().getStorage().data();
return new LinearRegressionModel(new DenseVector(Arrays.copyOf(p, p.length - 1)), p[p.length - 1]);
}
use of org.apache.ignite.ml.nn.MLPTrainer in project ignite by apache.
the class LogisticRegressionSGDTrainer method updateModel.
/**
* {@inheritDoc}
*/
@Override
protected <K, V> LogisticRegressionModel updateModel(LogisticRegressionModel mdl, DatasetBuilder<K, V> datasetBuilder, Preprocessor<K, V> extractor) {
IgniteFunction<Dataset<EmptyContext, SimpleLabeledDatasetData>, MLPArchitecture> archSupplier = dataset -> {
Integer cols = dataset.compute(data -> {
if (data.getFeatures() == null)
return null;
return data.getFeatures().length / data.getRows();
}, (a, b) -> {
// If both are null then zero will be propagated, no good.
if (a == null)
return b;
return a;
});
if (cols == null)
throw new IllegalStateException("Cannot train on empty dataset");
MLPArchitecture architecture = new MLPArchitecture(cols);
architecture = architecture.withAddedLayer(1, true, Activators.SIGMOID);
return architecture;
};
MLPTrainer<?> trainer = new MLPTrainer<>(archSupplier, LossFunctions.L2, updatesStgy, maxIterations, batchSize, locIterations, seed).withEnvironmentBuilder(envBuilder);
MultilayerPerceptron mlp;
IgniteFunction<LabeledVector<Double>, LabeledVector<double[]>> func = lv -> new LabeledVector<>(lv.features(), new double[] { lv.label() });
PatchedPreprocessor<K, V, Double, double[]> patchedPreprocessor = new PatchedPreprocessor<>(func, extractor);
if (mdl != null) {
mlp = restoreMLPState(mdl);
mlp = trainer.update(mlp, datasetBuilder, patchedPreprocessor);
} else
mlp = trainer.fit(datasetBuilder, patchedPreprocessor);
double[] params = mlp.parameters().getStorage().data();
return new LogisticRegressionModel(new DenseVector(Arrays.copyOf(params, params.length - 1)), params[params.length - 1]);
}
Aggregations