@Override
protected MRPipeline createPipeline() throws IOException {
JobStepConfig stepConfig = getConfig();
ClusterSettings settings = ClusterSettings.create(ConfigUtils.getDefaultConfig());
String instanceDir = stepConfig.getInstanceDir();
int generationID = stepConfig.getGenerationID();
int iteration = stepConfig.getIteration();
String prefix = Namespaces.getInstanceGenerationPrefix(instanceDir, generationID);
String outputKey = prefix + String.format("sketch/%d/", iteration);
if (!validOutputPath(outputKey)) {
return null;
}
// get normalized vectors
String inputKey = prefix + "normalized/";
MRPipeline p = createBasicPipeline(DistanceToClosestFn.class);
AvroType<Pair<Integer, RealVector>> inputType = Avros.pairs(Avros.ints(), MLAvros.vector());
PCollection<Pair<Integer, RealVector>> in = p.read(avroInput(inputKey, inputType));
// either create or load the set of currently chosen k-sketch vectors
// they are stored in a KSketchIndex object
DistanceToClosestFn<RealVector> distanceToClosestFn;
UpdateIndexFn updateIndexFn;
if (iteration
== 1) { // Iteration 1 is the first real iteration; iteration 0 contains initial state
KSketchIndex index = createInitialIndex(settings, in);
distanceToClosestFn = new DistanceToClosestFn<>(index);
updateIndexFn = new UpdateIndexFn(index);
} else {
// Get the index location from the previous iteration
String previousIndexKey = prefix + String.format("sketch/%d/", iteration - 1);
distanceToClosestFn = new DistanceToClosestFn<>(previousIndexKey);
updateIndexFn = new UpdateIndexFn(previousIndexKey);
}
// compute distance of each vector in dataset to closest vector in k-sketch
PTable<Integer, Pair<RealVector, Double>> weighted =
in.parallelDo(
"computeDistances",
distanceToClosestFn,
Avros.tableOf(Avros.ints(), Avros.pairs(MLAvros.vector(), Avros.doubles())));
// run weighted reservoir sampling on the vector to select another group of
// settings.getSketchPoints()
// to add to the k-sketch
PTable<Integer, RealVector> kSketchSample =
ReservoirSampling.groupedWeightedSample(
weighted, settings.getSketchPoints(), RandomManager.getRandom());
// update the KSketchIndex with the newly-chosen vectors
kSketchSample
.parallelDo("updateIndex", updateIndexFn, Serializables.avro(KSketchIndex.class))
.write(avroOutput(outputKey));
return p;
}
@Override
protected MRPipeline createPipeline() throws IOException {
JobStepConfig stepConfig = getConfig();
Config config = ConfigUtils.getDefaultConfig();
ClusterSettings clusterSettings = ClusterSettings.create(config);
String instanceDir = stepConfig.getInstanceDir();
long generationID = stepConfig.getGenerationID();
String prefix = Namespaces.getInstanceGenerationPrefix(instanceDir, generationID);
String outputKey = prefix + "weighted/";
if (!validOutputPath(outputKey)) {
return null;
}
String indexKey = prefix + "sketch/" + clusterSettings.getSketchIterations();
String inputKey = prefix + "normalized/";
MRPipeline p = createBasicPipeline(ClosestSketchVectorFn.class);
// first I compute the weight of each k-sketch vector, i.e., Voronoi partition
// I aggregate all together and persist on disk
// PCollection<ClosestSketchVectorData> weights = inputPairs(p, inputKey, MLAvros.vector())
PCollection<ClosestSketchVectorData> weights =
PTables.asPTable(
inputPairs(p, inputKey, MLAvros.vector())
.parallelDo(
"computingSketchVectorWeights",
new ClosestSketchVectorFn<RealVector>(indexKey, clusterSettings),
Avros.pairs(Avros.ints(), Avros.reflects(ClosestSketchVectorData.class))))
.groupByKey(1)
.combineValues(new ClosestSketchVectorAggregator(clusterSettings))
.values()
.write(avroOutput(outputKey + "kSketchVectorWeights/"));
// this "pipeline" takes a single ClosestSketchVectorData and returns weighted vectors
// could be done outside MapReduce, but that would require me to materialize the
// ClosestSketchVectorData
weights
.parallelDo(
"generatingWeightedSketchVectors",
new WeightVectorsFn(indexKey),
KMeansTypes.FOLD_WEIGHTED_VECTOR)
.write(avroOutput(outputKey + "weightedKSketchVectors/"));
return p;
}