Using SharpLearning.XGBoost

Native dependency and disposable models

SharpLearning.XGBoost depends on the native gradient boosting library xgboost. This means that the models produced by the learners in SharpLearning.XGBoost has a native dependency and implements the IDisposable interface. So when using an XGBoost model, RegressionXGBoostModel or ClassificationXGBoostModel, make sure to call the Dispose or place the model in a using statement like shown below:

var learner = new RegressionXGBoostLearner();

using (var model = learner.Learn(observations, targets))
{
    var predictions = model.Predict(observations);
}

The using statement is especially important if several models are learned in a loop, for instance when tuning hyperparameters. If the models are not disposed after each iteration, the previous models will not be collected, and the loop will leak memory.

If an XGBoost model is a member of another type, make sure to implement the IDisposable interface on the type, and call the dispose method on the model:

public void Dispose()
{
    if (m_model != null)
    {
        m_model.Dispose();
    }
}

Saving and loading models

Since the model from XGBoost, RegressionXGBoostModel or ClassificationXGBoostModel, rely on the native implementation, the regular serialization methods (GenericBinarySerializer and GenericXmlDataContractSerializer) does not work with the XGBoost models. Therefore the only way to save an XGBoost model is by calling the Save` method directly on the model. The save method takes as argument the filepath of the model to save:

var learner = new RegressionXGBoostLearner();
// Saves XGBoost model.
using (var model = learner.Learn(observations, targets))
{
    model.Save("C:\model.xgb");
}

Likewise the model is loaded by using the static Load method directly on the model class, for instance:

// Loads XGBoost model.
using (var loadedModel = RegressionXGBoostModel.Load("model.xgb"))
{
    var predictions = loadedModel.Predict(observations);
}

Training with GPU

Future work

In the future, it might be possible to convert models from SharpLearning.XGBoost to models from SharpLearning.GradientBoost, to add support for the general serialization methods, and to avoid native dependencies when using the models. But this is currently not possible