public static PredictorModelOutput CombineOvaModels(IHostEnvironment env, CombineOvaPredictorModelsInput input)
{
Contracts.CheckValue(env, nameof(env));
var host = env.Register("CombineOvaModels");
host.CheckValue(input, nameof(input));
EntryPointUtils.CheckInputArgs(host, input);
host.CheckNonEmpty(input.ModelArray, nameof(input.ModelArray));
// Something tells me we should put normalization as part of macro expansion, but since i get
// subgraph instead of learner it's a bit tricky to get learner and decide should we add
// normalization node or not, plus everywhere in code we leave that reposnsibility to TransformModel.
var normalizedView = input.ModelArray[0].TransformModel.Apply(host, input.TrainingData);
using (var ch = host.Start("CombineOvaModels"))
{
var schema = normalizedView.Schema;
var label = TrainUtils.MatchNameOrDefaultOrNull(ch, schema, nameof(input.LabelColumn),
input.LabelColumn,
DefaultColumnNames.Label);
var feature = TrainUtils.MatchNameOrDefaultOrNull(ch, schema, nameof(input.FeatureColumn),
input.FeatureColumn, DefaultColumnNames.Features);
var weight = TrainUtils.MatchNameOrDefaultOrNull(ch, schema, nameof(input.WeightColumn),
input.WeightColumn, DefaultColumnNames.Weight);
var data = new RoleMappedData(normalizedView, label, feature, null, weight);
return(new PredictorModelOutput
{
PredictorModel = new PredictorModelImpl(env, data, input.TrainingData,
OvaModelParameters.Create(host, input.UseProbabilities,
input.ModelArray.Select(p => p.Predictor as IPredictorProducing <float>).ToArray()))
});
}
}
private protected override OvaModelParameters CreatePredictor()
{
Host.Check(TrainedEnsemble != null, "The predictor cannot be created before training is complete.");
Host.Assert(_numClass > 1, "Must know the number of classes before creating a predictor.");
Host.Assert(TrainedEnsemble.NumTrees % _numClass == 0, "Number of trees should be a multiple of number of classes.");
var innerArgs = LightGbmInterfaceUtils.JoinParameters(Options);
IPredictorProducing <float>[] predictors = new IPredictorProducing <float> [_tlcNumClass];
for (int i = 0; i < _tlcNumClass; ++i)
{
var pred = CreateBinaryPredictor(i, innerArgs);
var cali = new PlattCalibrator(Host, -0.5, 0);
predictors[i] = new FeatureWeightsCalibratedPredictor(Host, pred, cali);
}
string obj = (string)GetGbmParameters()["objective"];
if (obj == "multiclass")
{
return(OvaModelParameters.Create(Host, OvaModelParameters.OutputFormula.Softmax, predictors));
}
else
{
return(OvaModelParameters.Create(Host, predictors));
}
}
[ConditionalFact(typeof(Environment), nameof(Environment.Is64BitProcess))] // LightGBM is 64-bit only
public void MultiClassLightGBM()
{
var env = new MLContext(seed: 0);
var dataPath = GetDataPath(TestDatasets.iris.trainFilename);
var dataSource = new MultiFileSource(dataPath);
var ctx = new MulticlassClassificationContext(env);
var reader = TextLoaderStatic.CreateReader(env,
c => (label: c.LoadText(0), features: c.LoadFloat(1, 4)));
OvaModelParameters pred = null;
// With a custom loss function we no longer get calibrated predictions.
var est = reader.MakeNewEstimator()
.Append(r => (label: r.label.ToKey(), r.features))
.Append(r => (r.label, preds: ctx.Trainers.LightGbm(
r.label,
r.features, onFit: p => pred = p)));
var pipe = reader.Append(est);
Assert.Null(pred);
var model = pipe.Fit(dataSource);
Assert.NotNull(pred);
var data = model.Read(dataSource);
// Just output some data on the schema for fun.
var schema = data.AsDynamic.Schema;
for (int c = 0; c < schema.Count; ++c)
{
Console.WriteLine($"{schema[c].Name}, {schema[c].Type}");
}
var metrics = ctx.Evaluate(data, r => r.label, r => r.preds, 2);
Assert.True(metrics.LogLoss > 0);
Assert.True(metrics.TopKAccuracy > 0);
}
protected override MulticlassPredictionTransformer <OvaModelParameters> MakeTransformer(OvaModelParameters model, Schema trainSchema) => new MulticlassPredictionTransformer <OvaModelParameters>(Host, model, trainSchema, FeatureColumn.Name, LabelColumn.Name);
