Merge pull request #907 from RobAltena/master

Update examples

Author: AlexDBlack

.gitignore

@@ -61,3 +61,4 @@ Word2vec-index/
 !tutorials/*.json
 *end.model
 
+arbiterExample/

dl4j-examples/src/main/java/org/deeplearning4j/examples/arbiter/BasicHyperparameterOptimizationExample.java

@@ -1,4 +1,4 @@
-/*******************************************************************************
+/* *****************************************************************************
  * Copyright (c) 2015-2019 Skymind, Inc.
  *
  * This program and the accompanying materials are made available under the
@@ -42,11 +42,11 @@
 import org.deeplearning4j.arbiter.task.MultiLayerNetworkTaskCreator;
 import org.deeplearning4j.arbiter.ui.listener.ArbiterStatusListener;
 import org.deeplearning4j.datasets.iterator.impl.MnistDataSetIterator;
-import org.deeplearning4j.eval.Evaluation;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
 import org.deeplearning4j.nn.weights.WeightInit;
 import org.deeplearning4j.ui.api.UIServer;
 import org.deeplearning4j.ui.storage.FileStatsStorage;
+import org.nd4j.evaluation.classification.Evaluation.Metric;
 import org.nd4j.linalg.activations.Activation;
 import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
 import org.nd4j.linalg.lossfunctions.LossFunctions;
@@ -114,14 +114,16 @@ public static void main(String[] args) throws Exception {
         //     This will result in examples being saved to arbiterExample/0/, arbiterExample/1/, arbiterExample/2/, ...
         String baseSaveDirectory = "arbiterExample/";
         File f = new File(baseSaveDirectory);
-        if (f.exists()) f.delete();
+        if (f.exists()) //noinspection ResultOfMethodCallIgnored
+            f.delete();
+        //noinspection ResultOfMethodCallIgnored
         f.mkdir();
         ResultSaver modelSaver = new FileModelSaver(baseSaveDirectory);
 
         // (d) What are we actually trying to optimize?
         //     In this example, let's use classification accuracy on the test set
         //     See also ScoreFunctions.testSetF1(), ScoreFunctions.testSetRegression(regressionValue) etc
-        ScoreFunction scoreFunction = new EvaluationScoreFunction(Evaluation.Metric.ACCURACY);
+        ScoreFunction scoreFunction = new EvaluationScoreFunction(Metric.ACCURACY);
 
 
         // (e) When should we stop searching? Specify this with termination conditions

dl4j-examples/src/main/java/org/deeplearning4j/examples/arbiter/genetic/BaseGeneticHyperparameterOptimizationExample.java

@@ -1,4 +1,4 @@
-/*******************************************************************************
+/* *****************************************************************************
  * Copyright (c) 2015-2019 Skymind, Inc.
  *
  * This program and the accompanying materials are made available under the
@@ -25,9 +25,10 @@
 import org.deeplearning4j.arbiter.optimize.generator.genetic.population.PopulationModel;
 import org.deeplearning4j.arbiter.optimize.runner.IOptimizationRunner;
 import org.deeplearning4j.arbiter.scoring.impl.EvaluationScoreFunction;
-import org.deeplearning4j.eval.Evaluation;
 import org.deeplearning4j.nn.graph.ComputationGraph;
+import org.nd4j.evaluation.classification.Evaluation.Metric;
 
+import java.io.IOException;
 import java.util.List;
 
 /**
@@ -38,23 +39,13 @@
  *
  * @author Alexandre Boulanger
  */
-
 public class BaseGeneticHyperparameterOptimizationExample {
 
-    public static void main(String[] args) throws Exception {
-
-        ComputationGraphSpace cgs = GeneticSearchExampleConfiguration.GetGraphConfiguration();
-
-        EvaluationScoreFunction scoreFunction = new EvaluationScoreFunction(Evaluation.Metric.F1);
-
-        // This is where we create the GeneticSearchCandidateGenerator with its default behavior:
-        //  - a population that fits 30 candidates and is culled back to 20 when it overflows
-        //  - new candidates are generated with a probability of 85% of being the result of breeding (a k-point crossover with 1 to 4 points)
-        //  - the new candidate have a probability of 0.5% of sustaining a random mutation on one of its genes.
-        GeneticSearchCandidateGenerator candidateGenerator = new GeneticSearchCandidateGenerator.Builder(cgs, scoreFunction).build();
-
-        // Let's have a listener to print the population size after each evaluation.
-        PopulationModel populationModel = candidateGenerator.getPopulationModel();
+    /**
+     * Common code used by two Arbiter examples.
+     */
+    public static void run(PopulationModel populationModel, GeneticSearchCandidateGenerator candidateGenerator,
+                           EvaluationScoreFunction scoreFunction) throws IOException {
         populationModel.addListener(new ExamplePopulationListener());
 
         IOptimizationRunner runner = GeneticSearchExampleConfiguration.BuildRunner(candidateGenerator, scoreFunction);
@@ -80,13 +71,32 @@ public static void main(String[] args) throws Exception {
         System.out.println(bestModel.getConfiguration().toJson());
     }
 
+    public static void main(String[] args) throws Exception {
+
+        ComputationGraphSpace cgs = GeneticSearchExampleConfiguration.GetGraphConfiguration();
+
+        EvaluationScoreFunction scoreFunction = new EvaluationScoreFunction(Metric.F1);
+
+        // This is where we create the GeneticSearchCandidateGenerator with its default behavior:
+        //  - a population that fits 30 candidates and is culled back to 20 when it overflows
+        //  - new candidates are generated with a probability of 85% of being the result of breeding (a k-point crossover with 1 to 4 points)
+        //  - the new candidate have a probability of 0.5% of sustaining a random mutation on one of its genes.
+        GeneticSearchCandidateGenerator candidateGenerator = new GeneticSearchCandidateGenerator.Builder(cgs, scoreFunction).build();
+
+        // Let's have a listener to print the population size after each evaluation.
+        PopulationModel populationModel = candidateGenerator.getPopulationModel();
+        populationModel.addListener(new ExamplePopulationListener());
+        run(populationModel, candidateGenerator, scoreFunction);
+    }
+
     public static class ExamplePopulationListener implements PopulationListener {
 
+        @SuppressWarnings("OptionalGetWithoutIsPresent")
         @Override
         public void onChanged(List<Chromosome> population) {
             double best = population.get(0).getFitness();
             double average = population.stream()
-                .mapToDouble(c -> c.getFitness())
+                .mapToDouble(Chromosome::getFitness)
                 .average()
                 .getAsDouble();
             System.out.println(String.format("\nPopulation size is %1$s, best score is %2$s, average score is %3$s", population.size(), best, average));

dl4j-examples/src/main/java/org/deeplearning4j/examples/arbiter/genetic/CustomGeneticHyperparameterOptimizationExample.java

@@ -1,4 +1,4 @@
-/*******************************************************************************
+/* *****************************************************************************
  * Copyright (c) 2015-2019 Skymind, Inc.
  *
  * This program and the accompanying materials are made available under the
@@ -20,25 +20,19 @@
 import org.apache.commons.math3.random.RandomGenerator;
 import org.apache.commons.math3.random.SynchronizedRandomGenerator;
 import org.deeplearning4j.arbiter.ComputationGraphSpace;
-import org.deeplearning4j.arbiter.optimize.api.OptimizationResult;
-import org.deeplearning4j.arbiter.optimize.api.saving.ResultReference;
 import org.deeplearning4j.arbiter.optimize.generator.GeneticSearchCandidateGenerator;
-import org.deeplearning4j.arbiter.optimize.generator.genetic.Chromosome;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.crossover.CrossoverOperator;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.crossover.KPointCrossover;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.crossover.parentselection.TwoParentSelection;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.culling.CullOperator;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.culling.LeastFitCullOperator;
-import org.deeplearning4j.arbiter.optimize.generator.genetic.population.PopulationListener;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.population.PopulationModel;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.selection.GeneticSelectionOperator;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.selection.SelectionOperator;
-import org.deeplearning4j.arbiter.optimize.runner.IOptimizationRunner;
 import org.deeplearning4j.arbiter.scoring.impl.EvaluationScoreFunction;
-import org.deeplearning4j.eval.Evaluation;
-import org.deeplearning4j.nn.graph.ComputationGraph;
+import org.nd4j.evaluation.classification.Evaluation.Metric;
 
-import java.util.List;
+import static org.deeplearning4j.examples.arbiter.genetic.BaseGeneticHyperparameterOptimizationExample.run;
 
 /**
  * In this hyperparameter optimization example, we change the default behavior of the genetic candidate generator.
@@ -55,7 +49,7 @@ public static void main(String[] args) throws Exception {
 
         ComputationGraphSpace cgs = GeneticSearchExampleConfiguration.GetGraphConfiguration();
 
-        EvaluationScoreFunction scoreFunction = new EvaluationScoreFunction(Evaluation.Metric.F1);
+        EvaluationScoreFunction scoreFunction = new EvaluationScoreFunction(Metric.F1);
 
         // The ExampleCullOperator extends the default cull operator (least fit) to include an artificial predator.
         CullOperator cullOperator = new ExampleCullOperator();
@@ -85,30 +79,8 @@ public static void main(String[] args) throws Exception {
             .build();
 
         // Let's have a listener to print the population size after each evaluation.
-        populationModel.addListener(new ExamplePopulationListener());
-
-        IOptimizationRunner runner = GeneticSearchExampleConfiguration.BuildRunner(candidateGenerator, scoreFunction);
-
-        //Start the hyperparameter optimization
-        runner.execute();
-
-        //Print out some basic stats regarding the optimization procedure
-        String s = "Best score: " + runner.bestScore() + "\n" +
-            "Index of model with best score: " + runner.bestScoreCandidateIndex() + "\n" +
-            "Number of configurations evaluated: " + runner.numCandidatesCompleted() + "\n";
-        System.out.println(s);
-
-
-        //Get all results, and print out details of the best result:
-        int indexOfBestResult = runner.bestScoreCandidateIndex();
-        List<ResultReference> allResults = runner.getResults();
-
-        OptimizationResult bestResult = allResults.get(indexOfBestResult).getResult();
-        ComputationGraph bestModel = (ComputationGraph) bestResult.getResultReference().getResultModel();
-
-        System.out.println("\n\nConfiguration of best model:\n");
-        System.out.println(bestModel.getConfiguration().toJson());
-
+        populationModel.addListener(new BaseGeneticHyperparameterOptimizationExample.ExamplePopulationListener());
+        run(populationModel, candidateGenerator, scoreFunction);
     }
 
     // This is an example of a custom behavior for the genetic algorithm. We force one of the parent to be one of the
@@ -158,17 +130,4 @@ public void cullPopulation() {
             System.out.println(String.format("Randomly removed %1$s candidate(s).", preyCount));
         }
     }
-
-    public static class ExamplePopulationListener implements PopulationListener {
-
-        @Override
-        public void onChanged(List<Chromosome> population) {
-            double best = population.get(0).getFitness();
-            double average = population.stream()
-                .mapToDouble(Chromosome::getFitness)
-                .average()
-                .getAsDouble();
-            System.out.println(String.format("\nPopulation size is %1$s, best score is %2$s, average score is %3$s", population.size(), best, average));
-        }
-    }
 }

dl4j-examples/src/main/java/org/deeplearning4j/examples/arbiter/genetic/GeneticSearchExampleConfiguration.java

@@ -1,4 +1,4 @@
-/*******************************************************************************
+/* *****************************************************************************
  * Copyright (c) 2015-2019 Skymind, Inc.
  *
  * This program and the accompanying materials are made available under the
@@ -49,22 +49,20 @@
 import java.io.File;
 import java.util.Properties;
 
-public class GeneticSearchExampleConfiguration {
+class GeneticSearchExampleConfiguration {
 
-    public static ComputationGraphSpace GetGraphConfiguration() {
+    static ComputationGraphSpace GetGraphConfiguration() {
         int inputSize = 784;
         int outputSize = 47;
 
         // First, we setup the hyperspace parameters. These are the values which will change, breed and mutate
         // while attempting to find the best candidate.
-        DiscreteParameterSpace<Activation> activationSpace = new DiscreteParameterSpace(new Activation[] {
-            Activation.ELU,
+        DiscreteParameterSpace<Activation> activationSpace = new DiscreteParameterSpace<>(Activation.ELU,
             Activation.RELU,
             Activation.LEAKYRELU,
             Activation.TANH,
             Activation.SELU,
-            Activation.HARDSIGMOID
-        });
+            Activation.HARDSIGMOID);
         IntegerParameterSpace[] layersParametersSpace = new IntegerParameterSpace[] {
             new IntegerParameterSpace(outputSize, inputSize),
             new IntegerParameterSpace(outputSize, inputSize),