Java prediction method for Autoencoder POJO missing

[Roberto Rösler]

Hi,

I exported a trained autoencoder model as POJO and now want to use it to predict new cases. However, it seems there is no implementation for a prediction method

Exception in thread "main" java.lang.RuntimeException: Unimplemented 55  at
hex.genmodel.easy.EasyPredictModelWrapper.predictAutoEncoder(EasyPredictModelWrapper.java:103)at
main.main(main.java:71)

or see https://github.com/h2oai/h2o-3/blob/master/h2o-genmodel/src/main/java/hex/genmodel/easy/EasyPredictModelWrapper.java).

I'm really curious about that because which value has the POJO if I'm not able to make any predictions with it?

 

Best regards,

 

Roberto

[Tom Kraljevic]

Hi Roberto,

The generated Deep Learning POJO does have the autoencoder inside.

There just isn’t an Easy wrapper for it yet.  Users of it so far used the low-level POJO directly.

We would be interested in your feedback in terms of what that Easy API should look like.

(Basically, I’d love to see you write some code using the imaginary results of predictAutoEncoder().)

Thanks,

Tom

--
You received this message because you are subscribed to the Google Groups "H2O Open Source Scalable Machine Learning - h2ostream" group.
To unsubscribe from this group and stop receiving emails from it, send an email to h2ostream+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Roberto Rösler]

Hi Tom,

 

your right, the POJO contains a score method, but I was hoping that there will be also an Easy-API wrapper that can be used with it. Would result in a more comfortable and unified API.

I'm not a Java programmer, but I would expect that the resulting "AutoEncoderModelPrediction" object afterwards give me access on (1) the reconstructed values (like with the predict function from the R and Python API), (2) the per-feature reconstruction error and (3) the average reconstruction error ( (2) and (3) in a similar way like the anomaly method from the R/Python-API).

So I edited my test code on the "AutoEncoderModelPrediction" class like this:

public class AutoEncoderModelPrediction extends AbstractPrediction {
    public double[] predictions;
    public double[] feature;
    public double[] reconstrunctionError;
    public double averageReconstructionError;
}

and on EasyPredictModelWrapper (just update the predictAutoEncoder method) like this:

  public AutoEncoderModelPrediction  predictAutoEncoder(RowData data) throws PredictException {
    double[] preds = preamble(ModelCategory.AutoEncoder, data);
   
    // save predictions
    AutoEncoderModelPrediction  p = new AutoEncoderModelPrediction();
    p.predictions = preds;
   
    // save raw data
    double[] rawData = new double[m.nfeatures()];
    setToNaN(rawData);
    fillRawData(data, rawData);
    p.feature = rawData;
   
    //calculate and reconstruction error
    double[] reconstrunctionError = new double [rawData.length];
    for (int i = 0; i < reconstrunctionError.length; i++) {
        reconstrunctionError[i] = Math.pow(rawData[i] - preds[i],2);
    }
    p.reconstrunctionError = reconstrunctionError;
   
    //calculate mean squared error
    double sum = 0;
    for (int i = 0; i < reconstrunctionError.length; i++) {
        sum = sum + reconstrunctionError[i];
    }
    p.averageReconstructionError = sum/reconstrunctionError.length;
   
    return p;
  }

Now I can use something like this inside my little test program (the main java from your POJO documentation) to get the result for a specific feature vector:

        AutoEncoderModelPrediction p = model.predictAutoEncoder(row);
        System.out.println("Got predictions ...");
        System.out.println(Arrays.toString(p.predictions));
       
        System.out.println("\nRaw data ...");
        System.out.println(Arrays.toString(p.feature));
       
        System.out.println("\nReconstruction error ...");
        System.out.println(Arrays.toString(p.reconstrunctionError));
       
        System.out.println("\nAverage reconstruction error ...");
        System.out.println(p.averageReconstructionError);

Will similar functionality find their way into your product? 

Best

Roberto

[Tom Kraljevic]

Thanks for the feedback Roberto.

This jira captures the request.

https://0xdata.atlassian.net/browse/PUBDEV-2232

Tom