private void Train(ExperimentModifier settings)
{
var setup = MulticlassExperimentSettingsHelper.SetupExperiment(_mLContext, settings, settings.Paths, settings.ForPrs);
try
{
// Start experiment
var textLoader = _mLContext.Data.CreateTextLoader(setup.columnInference.TextLoaderOptions);
var paths = settings.Paths;
// train once:
var experimentResult = MulticlassExperimentHelper.Train(
_mLContext, settings.LabelColumnName, setup.experimentSettings, new MulticlassExperimentProgressHandler(), paths, textLoader);
// train twice
var refitModel = MulticlassExperimentHelper.Retrain(experimentResult,
"refit model",
new MultiFileSource(paths.TrainPath, paths.ValidatePath),
paths.ValidatePath,
paths.FittedModelPath, textLoader, _mLContext);
// final train:
refitModel = MulticlassExperimentHelper.Retrain(_mLContext, experimentResult, setup.columnInference, paths);
}
catch (Exception ex)
{
Trace.WriteLine(ex);
}
}
public static (ColumnInferenceResults columnInference, MulticlassExperimentSettings experimentSettings) SetupExperiment(
MLContext mlContext, ExperimentModifier st, DataFilePaths paths, bool forPrs)
{
var columnInference = InferColumns(mlContext, paths.TrainPath, st.LabelColumnName);
var columnInformation = columnInference.ColumnInformation;
st.ColumnSetup(columnInformation, forPrs);
var experimentSettings = new MulticlassExperimentSettings();
st.TrainerSetup(experimentSettings.Trainers);
experimentSettings.MaxExperimentTimeInSeconds = st.ExperimentTime;
var cts = new System.Threading.CancellationTokenSource();
experimentSettings.CancellationToken = cts.Token;
// Set the cache directory to null.
// This will cause all models produced by AutoML to be kept in memory
// instead of written to disk after each run, as AutoML is training.
// (Please note: for an experiment on a large dataset, opting to keep all
// models trained by AutoML in memory could cause your system to run out
// of memory.)
experimentSettings.CacheDirectory = new DirectoryInfo(Path.GetTempPath());
experimentSettings.OptimizingMetric = MulticlassClassificationMetric.MicroAccuracy;
return(columnInference, experimentSettings);
}
public void Train(DataFilePaths files, bool forPrs)
{
var stopWatch = Stopwatch.StartNew();
var st = new ExperimentModifier(files, forPrs);
Train(st);
stopWatch.Stop();
Trace.WriteLine($"Done creating model in {stopWatch.ElapsedMilliseconds}ms");
}
