Train an integer instead of a classification

#1

My question is how can I train an integer (STOCK_FK) instead of the classification (.classification(finalSchema.getIndexOfColumn(“STOCK_FK”), 3))?..

    Schema schema = new Schema.Builder()
            .addColumnsInteger("ID", "HZW_FK")
            .addColumnInteger("STOCK_FK")
            .addColumnsInteger("DISPO_MONTH", "DISPO_YEAR")
            .addColumnDouble("DISPO_AMOUNT")
            .build();

    DataAnalysis analysis = AnalyzeLocal.analyze(schema, recordReader);
    HtmlAnalysis.createHtmlAnalysisFile(analysis, new File("C:/Disposition/analysis.html"));

    TransformProcess transformProcess = new TransformProcess.Builder(schema)
            .removeColumns("ID", "DISPO_MONTH", "DISPO_YEAR")
            .normalize("HZW_FK", Normalize.MinMax, analysis)
            .normalize("STOCK_FK", Normalize.MinMax, analysis)
            .normalize("DISPO_AMOUNT", Normalize.Log2MeanExcludingMin, analysis)
            .build();

    Schema finalSchema = transformProcess.getFinalSchema();

    TransformProcessRecordReader trainRecordReader = new TransformProcessRecordReader(new CSVRecordReader(), transformProcess);
    trainRecordReader.initialize(inputSplit);

    int batchSize = 30;
    RecordReaderDataSetIterator trainIterator = new RecordReaderDataSetIterator.Builder(trainRecordReader, batchSize)
    		.classification(finalSchema.getIndexOfColumn("STOCK_FK"), 3)
            .build();


    MultiLayerConfiguration config = new NeuralNetConfiguration.Builder()
            .seed(0xC0FFEE)
            .weightInit(WeightInit.XAVIER)
            .activation(Activation.TANH)
            .updater(new Adam.Builder().learningRate(0.001).build())
            .l2(0.0000316)
            .list(
                    new DenseLayer.Builder().nOut(25).build(),
                    new DenseLayer.Builder().nOut(25).build(),
                    new DenseLayer.Builder().nOut(25).build(),
                    new DenseLayer.Builder().nOut(25).build(),
                    new DenseLayer.Builder().nOut(25).build(),
                    new OutputLayer.Builder(new LossMCXENT()).nOut(3).activation(Activation.SOFTMAX).build()
            )
            .setInputType(InputType.feedForward(finalSchema.numColumns() - 1))
            .build();

    MultiLayerNetwork model = new MultiLayerNetwork(config);
    model.init();

    UIServer uiServer = UIServer.getInstance();
    StatsStorage statsStorage = new InMemoryStatsStorage();
    uiServer.attach(statsStorage);

    model.addListeners(new ScoreIterationListener(1));
    model.addListeners(new StatsListener(statsStorage, 250));

    model.fit(trainIterator, 59);

    TransformProcessRecordReader testRecordReader = new TransformProcessRecordReader(new CSVRecordReader(), transformProcess);
    testRecordReader.initialize( new FileSplit(new File("C:/Disposition/Test/")));
    RecordReaderDataSetIterator testIterator = new RecordReaderDataSetIterator.Builder(testRecordReader, batchSize)
    		.classification(finalSchema.getIndexOfColumn("STOCK_FK"), 3)  
            .build();

    Evaluation evaluate = model.evaluate(testIterator);
    System.out.println(evaluate.stats());
    System.out.println("MCC: "+evaluate.matthewsCorrelation(EvaluationAveraging.Macro));

    File modelSave = new File("C:/Disposition/model.bin");
    model.save(modelSave);
    ModelSerializer.addObjectToFile(modelSave, "dataanalysis", analysis.toJson());
    ModelSerializer.addObjectToFile(modelSave, "schema", finalSchema.toJson());
}

}

#2

You can’t train an integer, but you can train a float.

As I already said in a previous answer, you change the classification part of the record reader to be regression.

And then you change the output layer to use LossMSE with Identity activation.

And obviously you will want to apply regression evaluation instead of the normal classification regression.

#3

like new OutputLayer.Builder(new LossMSE()).nOut(3).activation(Activation.IDENTITY).build()?

#4

Yes, if you actually build it from single layers instead of using a list like you did in your original post.

But I guess you will not have 3 outputs in this case, but instead just a single one.

#5

thank you and .evaluateRegression(finalSchema.getIndexOfColumn(“STOCK”), 1)?

#6

I’ve linked the javadoc for it, as you can see it expects a dataset iterator. Please, take the time to read the things that get linked. You are talking to a human not to a compiler.

#7

thank you very much I got a solution!