use of org.apache.ignite.ml.nn.architecture.MLPArchitecture in project ignite by apache.
the class MLPGroupTrainerTest method doTestXOR.
/**
* Test training of 'xor' by {@link MLPGroupUpdateTrainer}.
*/
private <U extends Serializable> void doTestXOR(UpdatesStrategy<? super MultilayerPerceptron, U> stgy) {
int samplesCnt = 1000;
Matrix xorInputs = new DenseLocalOnHeapMatrix(new double[][] { { 0.0, 0.0 }, { 0.0, 1.0 }, { 1.0, 0.0 }, { 1.0, 1.0 } }, StorageConstants.ROW_STORAGE_MODE).transpose();
Matrix xorOutputs = new DenseLocalOnHeapMatrix(new double[][] { { 0.0 }, { 1.0 }, { 1.0 }, { 0.0 } }, StorageConstants.ROW_STORAGE_MODE).transpose();
MLPArchitecture conf = new MLPArchitecture(2).withAddedLayer(10, true, Activators.RELU).withAddedLayer(1, false, Activators.SIGMOID);
IgniteCache<Integer, LabeledVector<Vector, Vector>> cache = LabeledVectorsCache.createNew(ignite);
String cacheName = cache.getName();
Random rnd = new Random(12345L);
try (IgniteDataStreamer<Integer, LabeledVector<Vector, Vector>> streamer = ignite.dataStreamer(cacheName)) {
streamer.perNodeBufferSize(10000);
for (int i = 0; i < samplesCnt; i++) {
int col = Math.abs(rnd.nextInt()) % 4;
streamer.addData(i, new LabeledVector<>(xorInputs.getCol(col), xorOutputs.getCol(col)));
}
}
int totalCnt = 30;
int failCnt = 0;
double maxFailRatio = 0.3;
MLPGroupUpdateTrainer<U> trainer = MLPGroupUpdateTrainer.getDefault(ignite).withSyncPeriod(3).withTolerance(0.001).withMaxGlobalSteps(100).withUpdateStrategy(stgy);
for (int i = 0; i < totalCnt; i++) {
MLPGroupUpdateTrainerCacheInput trainerInput = new MLPGroupUpdateTrainerCacheInput(conf, new RandomInitializer(new Random(123L + i)), 6, cache, 10, new Random(123L + i));
MultilayerPerceptron mlp = trainer.train(trainerInput);
Matrix predict = mlp.apply(xorInputs);
Tracer.showAscii(predict);
X.println(xorOutputs.getRow(0).minus(predict.getRow(0)).kNorm(2) + "");
failCnt += TestUtils.checkIsInEpsilonNeighbourhoodBoolean(xorOutputs.getRow(0), predict.getRow(0), 5E-1) ? 0 : 1;
}
double failRatio = (double) failCnt / totalCnt;
System.out.println("Fail percentage: " + (failRatio * 100) + "%.");
assertTrue(failRatio < maxFailRatio);
}
use of org.apache.ignite.ml.nn.architecture.MLPArchitecture in project ignite by apache.
the class MLPTest method paramsCountTest.
/**
* Test parameters count works well.
*/
@Test
public void paramsCountTest() {
int inputSize = 10;
int layerWithBiasNeuronsCnt = 13;
int layerWithoutBiasNeuronsCnt = 17;
MLPArchitecture conf = new MLPArchitecture(inputSize).withAddedLayer(layerWithBiasNeuronsCnt, true, Activators.SIGMOID).withAddedLayer(layerWithoutBiasNeuronsCnt, false, Activators.SIGMOID);
Assert.assertEquals(layerWithBiasNeuronsCnt * inputSize + layerWithBiasNeuronsCnt + (layerWithoutBiasNeuronsCnt * layerWithBiasNeuronsCnt), conf.parametersCount());
}
use of org.apache.ignite.ml.nn.architecture.MLPArchitecture in project ignite by apache.
the class MLPTest method testDifferentiation.
/**
* Test differentiation.
*/
@Test
public void testDifferentiation() {
int inputSize = 2;
int firstLayerNeuronsCnt = 1;
double w10 = 0.1;
double w11 = 0.2;
MLPArchitecture conf = new MLPArchitecture(inputSize).withAddedLayer(firstLayerNeuronsCnt, false, Activators.SIGMOID);
MultilayerPerceptron mlp = new MultilayerPerceptron(conf);
mlp.setWeight(1, 0, 0, w10);
mlp.setWeight(1, 1, 0, w11);
double x0 = 1.0;
double x1 = 3.0;
Matrix inputs = new DenseLocalOnHeapMatrix(new double[][] { { x0, x1 } }).transpose();
double ytt = 1.0;
Matrix truth = new DenseLocalOnHeapMatrix(new double[][] { { ytt } }).transpose();
Vector grad = mlp.differentiateByParameters(LossFunctions.MSE, inputs, truth);
// Let yt be y ground truth value.
// d/dw1i [(yt - sigma(w10 * x0 + w11 * x1))^2] =
// 2 * (yt - sigma(w10 * x0 + w11 * x1)) * (-1) * (sigma(w10 * x0 + w11 * x1)) * (1 - sigma(w10 * x0 + w11 * x1)) * xi =
// let z = sigma(w10 * x0 + w11 * x1)
// - 2* (yt - z) * (z) * (1 - z) * xi.
IgniteTriFunction<Double, Vector, Vector, Vector> partialDer = (yt, w, x) -> {
Double z = Activators.SIGMOID.apply(w.dot(x));
return x.copy().map(xi -> -2 * (yt - z) * z * (1 - z) * xi);
};
Vector weightsVec = mlp.weights(1).getRow(0);
Tracer.showAscii(weightsVec);
Vector trueGrad = partialDer.apply(ytt, weightsVec, inputs.getCol(0));
Tracer.showAscii(trueGrad);
Tracer.showAscii(grad);
Assert.assertEquals(mlp.architecture().parametersCount(), grad.size());
Assert.assertEquals(trueGrad, grad);
}
use of org.apache.ignite.ml.nn.architecture.MLPArchitecture in project ignite by apache.
the class MLPTest method testSimpleMLPPrediction.
/**
* Tests that MLP with 2 layer, 1 neuron in each layer and weight equal to 1 is equivalent to sigmoid function.
*/
@Test
public void testSimpleMLPPrediction() {
MLPArchitecture conf = new MLPArchitecture(1).withAddedLayer(1, false, Activators.SIGMOID);
MultilayerPerceptron mlp = new MultilayerPerceptron(conf, new MLPConstInitializer(1));
int input = 2;
Matrix predict = mlp.apply(new DenseLocalOnHeapMatrix(new double[][] { { input } }));
Assert.assertEquals(predict, new DenseLocalOnHeapMatrix(new double[][] { { Activators.SIGMOID.apply(input) } }));
}
use of org.apache.ignite.ml.nn.architecture.MLPArchitecture in project ignite by apache.
the class MLPGroupTrainerExample method main.
/**
* Executes example.
*
* @param args Command line arguments, none required.
*/
public static void main(String[] args) throws InterruptedException {
// 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 IgniteThread, we must work with SparseDistributedMatrix inside IgniteThread
// because we create ignite cache internally.
IgniteThread igniteThread = new IgniteThread(ignite.configuration().getIgniteInstanceName(), MLPGroupTrainerExample.class.getSimpleName(), () -> {
int samplesCnt = 10000;
Matrix xorInputs = new DenseLocalOnHeapMatrix(new double[][] { { 0.0, 0.0 }, { 0.0, 1.0 }, { 1.0, 0.0 }, { 1.0, 1.0 } }, StorageConstants.ROW_STORAGE_MODE).transpose();
Matrix xorOutputs = new DenseLocalOnHeapMatrix(new double[][] { { 0.0 }, { 1.0 }, { 1.0 }, { 0.0 } }, StorageConstants.ROW_STORAGE_MODE).transpose();
MLPArchitecture conf = new MLPArchitecture(2).withAddedLayer(10, true, Activators.RELU).withAddedLayer(1, false, Activators.SIGMOID);
IgniteCache<Integer, LabeledVector<Vector, Vector>> cache = LabeledVectorsCache.createNew(ignite);
String cacheName = cache.getName();
Random rnd = new Random(12345L);
try (IgniteDataStreamer<Integer, LabeledVector<Vector, Vector>> streamer = ignite.dataStreamer(cacheName)) {
streamer.perNodeBufferSize(100);
for (int i = 0; i < samplesCnt; i++) {
int col = Math.abs(rnd.nextInt()) % 4;
streamer.addData(i, new LabeledVector<>(xorInputs.getCol(col), xorOutputs.getCol(col)));
}
}
int totalCnt = 100;
int failCnt = 0;
MLPGroupUpdateTrainer<RPropParameterUpdate> trainer = MLPGroupUpdateTrainer.getDefault(ignite).withSyncPeriod(3).withTolerance(0.001).withMaxGlobalSteps(20);
for (int i = 0; i < totalCnt; i++) {
MLPGroupUpdateTrainerCacheInput trainerInput = new MLPGroupUpdateTrainerCacheInput(conf, new RandomInitializer(rnd), 6, cache, 10);
MultilayerPerceptron mlp = trainer.train(trainerInput);
Matrix predict = mlp.apply(xorInputs);
System.out.println(">>> Prediction data at step " + i + " of total " + totalCnt + ":");
Tracer.showAscii(predict);
System.out.println("Difference estimate: " + xorOutputs.getRow(0).minus(predict.getRow(0)).kNorm(2));
failCnt += closeEnough(xorOutputs.getRow(0), predict.getRow(0)) ? 0 : 1;
}
double failRatio = (double) failCnt / totalCnt;
System.out.println("\n>>> Fail percentage: " + (failRatio * 100) + "%.");
System.out.println("\n>>> Distributed multilayer perceptron example completed.");
});
igniteThread.start();
igniteThread.join();
}
}
Aggregations