IGNITE-11040: [ML] Add parser for Spark Random forest regressor
authorzaleslaw <zaleslaw.sin@gmail.com>
Fri, 1 Feb 2019 10:10:59 +0000 (13:10 +0300)
committerYury Babak <ybabak@gridgain.com>
Fri, 1 Feb 2019 10:10:59 +0000 (13:10 +0300)
This closes #5997

15 files changed:
examples/src/main/java/org/apache/ignite/examples/ml/inference/spark/modelparser/RandomForestRegressionFromSparkExample.java [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/data/._SUCCESS.crc [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/data/.part-00000-06273895-4b81-4a77-823e-dfd32d1560eb-c000.snappy.parquet.crc [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/data/_SUCCESS [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/data/part-00000-06273895-4b81-4a77-823e-dfd32d1560eb-c000.snappy.parquet [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/metadata/._SUCCESS.crc [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/metadata/.part-00000.crc [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/metadata/_SUCCESS [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/metadata/part-00000 [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/._SUCCESS.crc [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/.part-00000-6844c864-0418-4554-b496-ef7584098d6d-c000.snappy.parquet.crc [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/_SUCCESS [new file with mode: 0644]
examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/part-00000-6844c864-0418-4554-b496-ef7584098d6d-c000.snappy.parquet [new file with mode: 0644]
modules/ml/spark-model-parser/src/main/java/org/apache/ignite/ml/sparkmodelparser/SparkModelParser.java
modules/ml/spark-model-parser/src/main/java/org/apache/ignite/ml/sparkmodelparser/SupportedSparkModels.java

diff --git a/examples/src/main/java/org/apache/ignite/examples/ml/inference/spark/modelparser/RandomForestRegressionFromSparkExample.java b/examples/src/main/java/org/apache/ignite/examples/ml/inference/spark/modelparser/RandomForestRegressionFromSparkExample.java
new file mode 100644 (file)
index 0000000..8c4a737
--- /dev/null
@@ -0,0 +1,91 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.ignite.examples.ml.inference.spark.modelparser;
+
+import java.io.FileNotFoundException;
+import javax.cache.Cache;
+import org.apache.ignite.Ignite;
+import org.apache.ignite.IgniteCache;
+import org.apache.ignite.Ignition;
+import org.apache.ignite.cache.query.QueryCursor;
+import org.apache.ignite.cache.query.ScanQuery;
+import org.apache.ignite.examples.ml.tutorial.TitanicUtils;
+import org.apache.ignite.ml.composition.ModelsComposition;
+import org.apache.ignite.ml.math.functions.IgniteBiFunction;
+import org.apache.ignite.ml.math.primitives.vector.Vector;
+import org.apache.ignite.ml.math.primitives.vector.VectorUtils;
+import org.apache.ignite.ml.sparkmodelparser.SparkModelParser;
+import org.apache.ignite.ml.sparkmodelparser.SupportedSparkModels;
+
+/**
+ * Run Random Forest regression model loaded from snappy.parquet file.
+ * The snappy.parquet file was generated by Spark MLLib model.write.overwrite().save(..) operator.
+ * <p>
+ * You can change the test data used in this example and re-run it to explore this algorithm further.</p>
+ */
+public class RandomForestRegressionFromSparkExample {
+    /** Path to Spark Random Forest regression model. */
+    public static final String SPARK_MDL_PATH = "examples/src/main/resources/models/spark/serialized/rfreg/data" +
+        "/part-00000-06273895-4b81-4a77-823e-dfd32d1560eb-c000.snappy.parquet";
+
+    /** Run example. */
+    public static void main(String[] args) throws FileNotFoundException {
+        System.out.println();
+        System.out.println(">>> Random Forest regression model loaded from Spark through serialization over partitioned dataset usage example started.");
+        // Start ignite grid.
+        try (Ignite ignite = Ignition.start("examples/config/example-ignite.xml")) {
+            System.out.println(">>> Ignite grid started.");
+
+            IgniteCache<Integer, Object[]> dataCache = TitanicUtils.readPassengers(ignite);
+
+            IgniteBiFunction<Integer, Object[], Vector> featureExtractor = (k, v) -> {
+                double[] data = new double[] {(double)v[0], (double)v[1], (double)v[5], (double)v[6]};
+                data[0] = Double.isNaN(data[0]) ? 0 : data[0];
+                data[1] = Double.isNaN(data[1]) ? 0 : data[1];
+                data[2] = Double.isNaN(data[2]) ? 0 : data[2];
+                data[3] = Double.isNaN(data[3]) ? 0 : data[3];
+                return VectorUtils.of(data);
+            };
+
+            IgniteBiFunction<Integer, Object[], Double> lbExtractor = (k, v) -> (double)v[4];
+
+            ModelsComposition mdl = (ModelsComposition)SparkModelParser.parse(
+                SPARK_MDL_PATH,
+                SupportedSparkModels.RANDOM_FOREST_REGRESSION
+            );
+
+            System.out.println(">>> Random Forest regression model: " + mdl);
+
+            System.out.println(">>> ---------------------------------");
+            System.out.println(">>> | Prediction\t| Ground Truth\t|");
+            System.out.println(">>> ---------------------------------");
+
+            try (QueryCursor<Cache.Entry<Integer, Object[]>> observations = dataCache.query(new ScanQuery<>())) {
+                for (Cache.Entry<Integer, Object[]> observation : observations) {
+                    Vector inputs = featureExtractor.apply(observation.getKey(), observation.getValue());
+                    double groundTruth = lbExtractor.apply(observation.getKey(), observation.getValue());
+                    double prediction = mdl.predict(inputs);
+
+                    System.out.printf(">>> | %.4f\t\t| %.4f\t\t|\n", prediction, groundTruth);
+                }
+            }
+
+            System.out.println(">>> ---------------------------------");
+        }
+    }
+}
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/data/._SUCCESS.crc b/examples/src/main/resources/models/spark/serialized/rfreg/data/._SUCCESS.crc
new file mode 100644 (file)
index 0000000..3b7b044
Binary files /dev/null and b/examples/src/main/resources/models/spark/serialized/rfreg/data/._SUCCESS.crc differ
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/data/.part-00000-06273895-4b81-4a77-823e-dfd32d1560eb-c000.snappy.parquet.crc b/examples/src/main/resources/models/spark/serialized/rfreg/data/.part-00000-06273895-4b81-4a77-823e-dfd32d1560eb-c000.snappy.parquet.crc
new file mode 100644 (file)
index 0000000..808d407
Binary files /dev/null and b/examples/src/main/resources/models/spark/serialized/rfreg/data/.part-00000-06273895-4b81-4a77-823e-dfd32d1560eb-c000.snappy.parquet.crc differ
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/data/_SUCCESS b/examples/src/main/resources/models/spark/serialized/rfreg/data/_SUCCESS
new file mode 100644 (file)
index 0000000..e69de29
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/data/part-00000-06273895-4b81-4a77-823e-dfd32d1560eb-c000.snappy.parquet b/examples/src/main/resources/models/spark/serialized/rfreg/data/part-00000-06273895-4b81-4a77-823e-dfd32d1560eb-c000.snappy.parquet
new file mode 100644 (file)
index 0000000..d76c5e8
Binary files /dev/null and b/examples/src/main/resources/models/spark/serialized/rfreg/data/part-00000-06273895-4b81-4a77-823e-dfd32d1560eb-c000.snappy.parquet differ
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/metadata/._SUCCESS.crc b/examples/src/main/resources/models/spark/serialized/rfreg/metadata/._SUCCESS.crc
new file mode 100644 (file)
index 0000000..3b7b044
Binary files /dev/null and b/examples/src/main/resources/models/spark/serialized/rfreg/metadata/._SUCCESS.crc differ
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/metadata/.part-00000.crc b/examples/src/main/resources/models/spark/serialized/rfreg/metadata/.part-00000.crc
new file mode 100644 (file)
index 0000000..58877b3
Binary files /dev/null and b/examples/src/main/resources/models/spark/serialized/rfreg/metadata/.part-00000.crc differ
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/metadata/_SUCCESS b/examples/src/main/resources/models/spark/serialized/rfreg/metadata/_SUCCESS
new file mode 100644 (file)
index 0000000..e69de29
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/metadata/part-00000 b/examples/src/main/resources/models/spark/serialized/rfreg/metadata/part-00000
new file mode 100644 (file)
index 0000000..dc41ce1
--- /dev/null
@@ -0,0 +1 @@
+{"class":"org.apache.spark.ml.regression.RandomForestRegressionModel","timestamp":1548250373346,"sparkVersion":"2.2.0","uid":"rfr_42762d520f2e","paramMap":{"maxMemoryInMB":256,"subsamplingRate":1.0,"featuresCol":"features","predictionCol":"prediction","impurity":"variance","seed":235498149,"minInfoGain":0.0,"numTrees":50,"labelCol":"age","checkpointInterval":10,"maxDepth":8,"cacheNodeIds":false,"minInstancesPerNode":1,"featureSubsetStrategy":"auto","maxBins":32},"numFeatures":4,"numTrees":50}
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/._SUCCESS.crc b/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/._SUCCESS.crc
new file mode 100644 (file)
index 0000000..3b7b044
Binary files /dev/null and b/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/._SUCCESS.crc differ
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/.part-00000-6844c864-0418-4554-b496-ef7584098d6d-c000.snappy.parquet.crc b/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/.part-00000-6844c864-0418-4554-b496-ef7584098d6d-c000.snappy.parquet.crc
new file mode 100644 (file)
index 0000000..ad6428f
Binary files /dev/null and b/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/.part-00000-6844c864-0418-4554-b496-ef7584098d6d-c000.snappy.parquet.crc differ
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/_SUCCESS b/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/_SUCCESS
new file mode 100644 (file)
index 0000000..e69de29
diff --git a/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/part-00000-6844c864-0418-4554-b496-ef7584098d6d-c000.snappy.parquet b/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/part-00000-6844c864-0418-4554-b496-ef7584098d6d-c000.snappy.parquet
new file mode 100644 (file)
index 0000000..7ab3347
Binary files /dev/null and b/examples/src/main/resources/models/spark/serialized/rfreg/treesMetadata/part-00000-6844c864-0418-4554-b496-ef7584098d6d-c000.snappy.parquet differ
index 8a66a3c..cc54848 100644 (file)
@@ -32,6 +32,7 @@ import org.apache.ignite.ml.IgniteModel;
 import org.apache.ignite.ml.clustering.kmeans.KMeansModel;
 import org.apache.ignite.ml.composition.ModelsComposition;
 import org.apache.ignite.ml.composition.boosting.GDBTrainer;
+import org.apache.ignite.ml.composition.predictionsaggregator.MeanValuePredictionsAggregator;
 import org.apache.ignite.ml.composition.predictionsaggregator.OnMajorityPredictionsAggregator;
 import org.apache.ignite.ml.composition.predictionsaggregator.WeightedPredictionsAggregator;
 import org.apache.ignite.ml.inference.Model;
@@ -89,12 +90,26 @@ public class SparkModelParser {
                 return loadKMeansModel(ignitePathToMdl);
             case DECISION_TREE_REGRESSION:
                 return loadDecisionTreeRegressionModel(ignitePathToMdl);
+            case RANDOM_FOREST_REGRESSION:
+                return loadRandomForestRegressionModel(ignitePathToMdl);
             default:
                 throw new UnsupportedSparkModelException(ignitePathToMdl);
         }
     }
 
     /**
+     * Load Random Forest Regression model.
+     *
+     * @param pathToMdl Path to model.
+     */
+    private static Model loadRandomForestRegressionModel(String pathToMdl) {
+        final List<IgniteModel<Vector, Double>> models = parseTreesForRandomForestAlgorithm(pathToMdl);
+        if (models == null)
+            return null;
+        return new ModelsComposition(models, new MeanValuePredictionsAggregator());
+    }
+
+    /**
      * Load Decision Tree Regression model.
      *
      * @param pathToMdl Path to model.
@@ -103,6 +118,11 @@ public class SparkModelParser {
         return loadDecisionTreeModel(pathToMdl);
     }
 
+    /**
+     * Load K-Means model.
+     *
+     * @param pathToMdl Path to model.
+     */
     private static Model loadKMeansModel(String pathToMdl) {
         Vector[] centers = null;
 
@@ -251,6 +271,18 @@ public class SparkModelParser {
      * @param pathToMdl Path to model.
      */
     private static Model loadRandomForestModel(String pathToMdl) {
+        final List<IgniteModel<Vector, Double>> models = parseTreesForRandomForestAlgorithm(pathToMdl);
+        if (models == null)
+            return null;
+        return new ModelsComposition(models, new OnMajorityPredictionsAggregator());
+    }
+
+    /**
+     * Parse trees from file for common Random Forest ensemble.
+     *
+     * @param pathToMdl Path to model.
+     */
+    private static List<IgniteModel<Vector, Double>> parseTreesForRandomForestAlgorithm(String pathToMdl) {
         try (ParquetFileReader r = ParquetFileReader.open(HadoopInputFile.fromPath(new Path(pathToMdl), new Configuration()))) {
             PageReadStore pages;
 
@@ -280,11 +312,9 @@ public class SparkModelParser {
                     }
                 }
             }
-
-            final List<IgniteModel<Vector, Double>> models = new ArrayList<>();
+            List<IgniteModel<Vector, Double>> models = new ArrayList<>();
             nodesByTreeId.forEach((key, nodes) -> models.add(buildDecisionTreeModel(nodes)));
-
-            return new ModelsComposition(models, new OnMajorityPredictionsAggregator());
+            return models;
         }
         catch (IOException e) {
             System.out.println("Error reading parquet file.");
index 4b203fe..26d0394 100644 (file)
@@ -44,6 +44,9 @@ public enum SupportedSparkModels {
     /** Decision tree regression. */
     DECISION_TREE_REGRESSION,
 
+    /** Random forest regression. */
+    RANDOM_FOREST_REGRESSION,
+
     /**
      * Gradient boosted trees.
      * NOTE: support binary classification only with raw labels 0 and 1