use of org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer in project deeplearning4j by deeplearning4j.
the class TestPreProcessedData method testPreprocessedData.
@Test
public void testPreprocessedData() {
//Test _loading_ of preprocessed data
int dataSetObjSize = 5;
int batchSizePerExecutor = 10;
String path = FilenameUtils.concat(System.getProperty("java.io.tmpdir"), "dl4j_testpreprocdata");
File f = new File(path);
if (f.exists())
f.delete();
f.mkdir();
DataSetIterator iter = new IrisDataSetIterator(5, 150);
int i = 0;
while (iter.hasNext()) {
File f2 = new File(FilenameUtils.concat(path, "data" + (i++) + ".bin"));
iter.next().save(f2);
}
MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().updater(Updater.RMSPROP).optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT).iterations(1).list().layer(0, new org.deeplearning4j.nn.conf.layers.DenseLayer.Builder().nIn(4).nOut(3).activation(Activation.TANH).build()).layer(1, new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(LossFunctions.LossFunction.MCXENT).nIn(3).nOut(3).activation(Activation.SOFTMAX).build()).pretrain(false).backprop(true).build();
SparkDl4jMultiLayer sparkNet = new SparkDl4jMultiLayer(sc, conf, new ParameterAveragingTrainingMaster.Builder(numExecutors(), dataSetObjSize).batchSizePerWorker(batchSizePerExecutor).averagingFrequency(1).repartionData(Repartition.Always).build());
sparkNet.setCollectTrainingStats(true);
sparkNet.fit("file:///" + path.replaceAll("\\\\", "/"));
SparkTrainingStats sts = sparkNet.getSparkTrainingStats();
//4 'fits' per averaging (4 executors, 1 averaging freq); 10 examples each -> 40 examples per fit. 150/40 = 3 averagings (round down); 3*4 = 12
int expNumFits = 12;
//Unfortunately: perfect partitioning isn't guaranteed by SparkUtils.balancedRandomSplit (esp. if original partitions are all size 1
// which appears to be occurring at least some of the time), but we should get close to what we expect...
assertTrue(Math.abs(expNumFits - sts.getValue("ParameterAveragingWorkerFitTimesMs").size()) < 3);
assertEquals(3, sts.getValue("ParameterAveragingMasterMapPartitionsTimesMs").size());
}
use of org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer in project deeplearning4j by deeplearning4j.
the class TestKryoWarning method doTestMLN.
private static void doTestMLN(SparkConf sparkConf) {
JavaSparkContext sc = new JavaSparkContext(sparkConf);
try {
MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().list().layer(0, new OutputLayer.Builder().nIn(10).nOut(10).build()).pretrain(false).backprop(true).build();
TrainingMaster tm = new ParameterAveragingTrainingMaster.Builder(1).build();
SparkDl4jMultiLayer sml = new SparkDl4jMultiLayer(sc, conf, tm);
} finally {
sc.stop();
}
}
use of org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer in project deeplearning4j by deeplearning4j.
the class TestCompareParameterAveragingSparkVsSingleMachine method testAverageEveryStepCNN.
@Test
public void testAverageEveryStepCNN() {
//Idea: averaging every step with SGD (SGD updater + optimizer) is mathematically identical to doing the learning
// on a single machine for synchronous distributed training
//BUT: This is *ONLY* the case if all workers get an identical number of examples. This won't be the case if
// we use RDD.randomSplit (which is what occurs if we use .fit(JavaRDD<DataSet> on a data set that needs splitting),
// which might give a number of examples that isn't divisible by number of workers (like 39 examples on 4 executors)
//This is also ONLY the case using SGD updater
int miniBatchSizePerWorker = 10;
int nWorkers = 4;
for (boolean saveUpdater : new boolean[] { true, false }) {
JavaSparkContext sc = getContext(nWorkers);
try {
//Do training locally, for 3 minibatches
int[] seeds = { 1, 2, 3 };
MultiLayerNetwork net = new MultiLayerNetwork(getConfCNN(12345, Updater.SGD));
net.init();
INDArray initialParams = net.params().dup();
for (int i = 0; i < seeds.length; i++) {
DataSet ds = getOneDataSetCNN(miniBatchSizePerWorker * nWorkers, seeds[i]);
if (!saveUpdater)
net.setUpdater(null);
net.fit(ds);
}
INDArray finalParams = net.params().dup();
//Do training on Spark with one executor, for 3 separate minibatches
ParameterAveragingTrainingMaster tm = new ParameterAveragingTrainingMaster.Builder(1).averagingFrequency(1).batchSizePerWorker(miniBatchSizePerWorker).saveUpdater(saveUpdater).workerPrefetchNumBatches(0).rddTrainingApproach(RDDTrainingApproach.Export).build();
SparkDl4jMultiLayer sparkNet = new SparkDl4jMultiLayer(sc, getConfCNN(12345, Updater.SGD), tm);
sparkNet.setCollectTrainingStats(true);
INDArray initialSparkParams = sparkNet.getNetwork().params().dup();
for (int i = 0; i < seeds.length; i++) {
List<DataSet> list = getOneDataSetAsIndividalExamplesCNN(miniBatchSizePerWorker * nWorkers, seeds[i]);
JavaRDD<DataSet> rdd = sc.parallelize(list);
sparkNet.fit(rdd);
}
System.out.println(sparkNet.getSparkTrainingStats().statsAsString());
INDArray finalSparkParams = sparkNet.getNetwork().params().dup();
System.out.println("Initial (Local) params: " + Arrays.toString(initialParams.data().asFloat()));
System.out.println("Initial (Spark) params: " + Arrays.toString(initialSparkParams.data().asFloat()));
System.out.println("Final (Local) params: " + Arrays.toString(finalParams.data().asFloat()));
System.out.println("Final (Spark) params: " + Arrays.toString(finalSparkParams.data().asFloat()));
assertArrayEquals(initialParams.data().asFloat(), initialSparkParams.data().asFloat(), 1e-8f);
assertArrayEquals(finalParams.data().asFloat(), finalSparkParams.data().asFloat(), 1e-6f);
double sparkScore = sparkNet.getScore();
assertTrue(sparkScore > 0.0);
assertEquals(net.score(), sparkScore, 1e-3);
} finally {
sc.stop();
}
}
}
use of org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer in project deeplearning4j by deeplearning4j.
the class TestSparkMultiLayerParameterAveraging method testSmallAmountOfData.
@Test
public void testSmallAmountOfData() {
//Idea: Test spark training where some executors don't get any data
//in this case: by having fewer examples (2 DataSets) than executors (local[*])
MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().updater(Updater.RMSPROP).optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT).iterations(1).list().layer(0, new org.deeplearning4j.nn.conf.layers.DenseLayer.Builder().nIn(nIn).nOut(3).activation(Activation.TANH).build()).layer(1, new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(LossFunctions.LossFunction.MSE).nIn(3).nOut(nOut).activation(Activation.SOFTMAX).build()).build();
SparkDl4jMultiLayer sparkNet = new SparkDl4jMultiLayer(sc, conf, new ParameterAveragingTrainingMaster(true, numExecutors(), 1, 10, 1, 0));
Nd4j.getRandom().setSeed(12345);
DataSet d1 = new DataSet(Nd4j.rand(1, nIn), Nd4j.rand(1, nOut));
DataSet d2 = new DataSet(Nd4j.rand(1, nIn), Nd4j.rand(1, nOut));
JavaRDD<DataSet> rddData = sc.parallelize(Arrays.asList(d1, d2));
sparkNet.fit(rddData);
}
use of org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer in project deeplearning4j by deeplearning4j.
the class TestSparkMultiLayerParameterAveraging method testEvaluation.
@Test
public void testEvaluation() {
SparkDl4jMultiLayer sparkNet = getBasicNetwork();
MultiLayerNetwork netCopy = sparkNet.getNetwork().clone();
Evaluation evalExpected = new Evaluation();
INDArray outLocal = netCopy.output(input, Layer.TrainingMode.TEST);
evalExpected.eval(labels, outLocal);
Evaluation evalActual = sparkNet.evaluate(sparkData);
assertEquals(evalExpected.accuracy(), evalActual.accuracy(), 1e-3);
assertEquals(evalExpected.f1(), evalActual.f1(), 1e-3);
assertEquals(evalExpected.getNumRowCounter(), evalActual.getNumRowCounter(), 1e-3);
assertMapEquals(evalExpected.falseNegatives(), evalActual.falseNegatives());
assertMapEquals(evalExpected.falsePositives(), evalActual.falsePositives());
assertMapEquals(evalExpected.trueNegatives(), evalActual.trueNegatives());
assertMapEquals(evalExpected.truePositives(), evalActual.truePositives());
assertEquals(evalExpected.precision(), evalActual.precision(), 1e-3);
assertEquals(evalExpected.recall(), evalActual.recall(), 1e-3);
assertEquals(evalExpected.getConfusionMatrix(), evalActual.getConfusionMatrix());
}
Aggregations