public Predictors <TOutput> Train(Datasets data
, Func <int, double> learningRateSchedule = null // optional: learning rate as a function of iteration (zero-based)
)
{
// For multi class, the number of labels is required.
if (!(PredictionKind != PredictionKind.MultiClassClassification || Objective.NumClass > 1))
{
throw new Exception("LightGBM requires the number of classes to be specified in the parameters for multi-class classification.");
}
if (PredictionKind == PredictionKind.Ranking)
{
if (data.Training.GetGroups() == null)
{
throw new Exception("Require Groups training data for ObjectiveType.LambdaRank");
}
if (data.Validation != null && data.Validation.GetGroups() == null)
{
throw new Exception("Require Groups validation data for ObjectiveType.LambdaRank");
}
}
TrainMetrics.Clear();
ValidMetrics.Clear();
Booster?.Dispose();
Booster = null;
Datasets = data;
var args = GetParameters(data);
Booster = Train(args, data.Training, data.Validation, TrainMetrics, ValidMetrics, learningRateSchedule);
(var model, var argsout) = Booster.GetModel();
TrainedEnsemble = model;
FeatureCount = data.Training.NumFeatures;
// check parameter strings
if (learningRateSchedule != null)
{
argsout.Learning.LearningRate = args.Learning.LearningRate;
}
var strIn = args.ToString();
var strOut = argsout.ToString();
if (strIn != strOut)
{
throw new Exception($"Parameters differ:\n{strIn}\n{strOut}");
}
var managed = CreateManagedPredictor();
var native = CreateNativePredictor();
return(new Predictors <TOutput>(managed, native));
}
/// <summary>
/// Adds the specified train loss.
/// </summary>
/// <param name="trainLoss">The train loss.</param>
/// <param name="trainMetric">The train metric.</param>
/// <param name="valLoss">The value loss.</param>
/// <param name="valMetric">The value metric.</param>
public void Add(List <double> trainLoss, List <double> trainMetric, List <double> valLoss, List <double> valMetric)
{
TrainLosses.Add(trainLoss.Average());
TrainMetrics.Add(trainMetric.Average());
if (valLoss.Count > 0)
{
ValLosses.Add(valLoss.Average());
}
if (valMetric.Count > 0)
{
ValMetrics.Add(valMetric.Average());
}
}
/// <summary>
/// Continue training current model, optionally with a new training dataset.
/// </summary>
/// <param name="data"></param>
/// <param name="learningRateSchedule"></param>
/// <returns></returns>
public Predictors <TOutput> ContinueTraining(Dataset trainingData = null
, Func <int, double> learningRateSchedule = null // optional: learning rate as a function of iteration (zero-based)
)
{
if (Booster == null)
{
throw new Exception("No existing booster to train.");
}
if (Datasets == null)
{
throw new Exception("No existing data sets.");
}
if (trainingData != null)
{
Datasets.Training = trainingData;
Booster.ResetTrainingData(trainingData);
}
// For multi class, the number of labels is required.
if (!(PredictionKind != PredictionKind.MultiClassClassification || Objective.NumClass > 1))
{
throw new Exception("LightGBM requires the number of classes to be specified in the parameters for multi-class classification.");
}
if (PredictionKind == PredictionKind.Ranking)
{
if (Datasets.Training.GetGroups() == null)
{
throw new Exception("Require Groups training data for ObjectiveType.LambdaRank");
}
if (Datasets.Validation != null && Datasets.Validation.GetGroups() == null)
{
throw new Exception("Require Groups validation data for ObjectiveType.LambdaRank");
}
}
// NOTE: existing metrics cleared
TrainMetrics.Clear();
ValidMetrics.Clear();
var args = GetParameters(Datasets);
// TODO: HOW TO RESET VALIDATION DATA???
Train(args, Booster, (Datasets.Validation != null), TrainMetrics, ValidMetrics, learningRateSchedule);
(var model, var argsout) = Booster.GetModel();
TrainedEnsemble = model;
FeatureCount = Datasets.Training.NumFeatures;
// check parameter strings
if (learningRateSchedule != null)
{
argsout.Learning.LearningRate = args.Learning.LearningRate;
}
// if both ForceColWise and ForceRowWise are false on the input, LightGBM appears to set one of them to be true on the output?
if (!args.Learning.ForceColWise && !args.Learning.ForceRowWise)
{
argsout.Learning.ForceColWise = false;
argsout.Learning.ForceRowWise = false;
}
// for some reason this parameter is not returned in the output model
argsout.Objective.MetricFreq = args.Objective.MetricFreq;
var strIn = args.ToString();
var strOut = argsout.ToString();
if (strIn != strOut)
{
throw new Exception($"Parameters differ:\n{strIn}\n{strOut}");
}
var managed = CreateManagedPredictor();
var native = CreateNativePredictor();
return(new Predictors <TOutput>(managed, native));
}
