Multiple outputs - classification and regression

DL4J
1

I have model with 2 outputs. One output is for classification problem - gender classification and other output is for age regression. I cant run RegressionEvaluation and Evaluation in same time.

EvaluativeListener(@NonNull MultiDataSetIterator iterator, int frequency, @NonNull InvocationType type,
IEvaluation… evaluations)
This listener receive array of IEvaluation, but if you add regression and classification evaluation it will break in runtime when it start evaluations…

ImageRecordReader imageRecordReaderAge = new ImageRecordReader(height, width, channels, labelMakerAge);
imageRecordReaderAge.initialize(dataAge, getImageTransform(isTraining));
ImageRecordReader imageRecordReaderGender = new ImageRecordReader(height, width, channels, labelMakerGender);
imageRecordReaderGender.initialize(dataGender, getImageTransform(isTraining));

    MultiDataSetIterator iterator = new RecordReaderMultiDataSetIterator.Builder(batchSize)
            .addReader("inputAge", imageRecordReaderAge)
            .addReader("inputGender", imageRecordReaderGender)
            .addInput("inputAge", 0, 0)
            .addInput("inputGender", 0, 0)
            .addOutput("inputAge", 1, 1) 
            .addOutput("inputGender", 1, 1)  
    iterator.setPreProcessor(dataNormalization);

    RegressionEvaluation regressionEvaluation = new RegressionEvaluation();
    //Evaluation evaluation = new Evaluation(labels);
    return new EvaluativeListener(dataSetIterator, frequency, type, regressionEvaluation);

Also EarlyStoppingConfiguration receive only one scoreCalculator and it must be DataSetLossCalculator.

Is posible to monitor both evaluations in same time?

2

Any solution for this?

3

@lavajaw those evaluations were built with 1 output in mind. I would suggest setting up 2 separate objects and accumulate them for each one. Use a listener as a hook for setting up the pre processing.
Those evaluate objects can be updated with the .eval(…) methods:

You will just have to do the tracking yourself.

4

@agibsonccc I will do, thanks.