Loading a model saved via python does not appear to save the "missing" parameter.

[cpr-pcurry]

I have a xgboost model that I have saved in Python (using xgboost == 0.8), with the parameter "missing" = -999.
However, when I load the model in SharpLearning and attempt a regression test on the training features, I see a 8% decrease in performance. This goes away when I remove the missing values (-999) from the training data.

This suggests to me that the "missing" parameter is not being loaded correctly by SharpLearning. Is the "missing" parameter supported currently?

Thanks for your help!

[cpr-pcurry]

I even tried adding -999 / "NaN" / "NAN" / "nan" as a missing value using SetParameter on the booster.

            var booster = new Booster(modelFile);
            booster.SetParameter("missing", "NAN");
            
            // Loads XGBoost model, verifies that it all works.
            using (var loadedModel = new ClassificationXGBoostModel(booster))
            {

                foreach (string[] row in rows)
                {
                    // extract label and features
                    double label = double.Parse(row.Last());
                    // Predict expects a 2d array
                    double[] features = row.Take(row.Length - 1).Select(double.Parse).ToArray();

                    // set all values that are -999 to nan
                    for (int i = 0; i < features.Length; i++)
                    {
                        if (features[i] == -999)
                        {
                            features[i] = double.NaN;
                        }
                    }
                    var prediction = loadedModel.Predict(features);
                    if (prediction == label)
                    {
                        correctPredictions++;
                    }
                    predictions++;
                }
            }

and converting all -999 to double.NaN. However, I still get the error:
XGBoost.lib.DllFailException: [12:52:03] C:\dev\libs\xgboost\src\c_api\c_api.cc:366: Check failed: nan_missing There are NAN in the matrix, however, you did not set missing=NAN

[mdabros]

Hi @cpr-pcurry,

SharpLearning just wraps the booster class from https://github.com/PicNet/XGBoost.Net so all loading is handled by that class.
You can see the load from file method below. It creates and loads the booster directly via native calls to the XGBoost library. I would assume the "missing" parameter is loaded as part of the model.

https://github.com/PicNet/XGBoost.Net/blob/master/XGBoost/lib/Booster.cs#L35

    public Booster(string fileName, int silent = 1)
    {
        IntPtr tempPtr;
        var newBooster = XGBOOST_NATIVE_METHODS.XGBoosterCreate(null, 0,out tempPtr); 
        var output = XGBOOST_NATIVE_METHODS.XGBoosterLoadModel(tempPtr, fileName);
        if (output == -1) throw new DllFailException(XGBOOST_NATIVE_METHODS.XGBGetLastError());
        handle = tempPtr;
    }

To test if the original issue (with different results with missings included) has to do with some of the conversions being done in SharpLearning when parsing in the feature vector to the model, you can try the above example just using XGBoost.Net. You need to use the DMatix type as input for the booster model. Something like this:

            // Parse feature vector (as jagged array)
            var floatObservation = new float[][]
            using (var data = new DMatrix(floatObservation))
            {
                // This contains the probabilies for each class if classification.
               var prediction = m_model.Predict(data);
            }

Hope this can help you in debugging the issue.