public static MultiModelPipeline Append(this IEstimator <ITransformer> estimator, params SweepableEstimator[] estimators)
{
var sweepableEstimator = new SweepableEstimator((context, parameter) => estimator, new SearchSpace.SearchSpace());
var multiModelPipeline = new MultiModelPipeline().Append(sweepableEstimator).Append(estimators);
return(multiModelPipeline);
}
public PipelineProposer(AutoMLExperimentSettings settings)
{
// this cost is used to initialize eci when started, the smaller the number, the less cost this trainer will use at start, and more likely it will be
// picked.
_estimatorCost = new Dictionary <EstimatorType, double>()
{
{ EstimatorType.LightGbmRegression, 0.788 },
{ EstimatorType.FastTreeRegression, 0.382 },
{ EstimatorType.FastForestRegression, 0.374 },
{ EstimatorType.SdcaRegression, 0.566 },
{ EstimatorType.FastTreeTweedieRegression, 0.401 },
{ EstimatorType.LbfgsPoissonRegressionRegression, 4.73 },
{ EstimatorType.FastForestOva, 4.283 },
{ EstimatorType.FastTreeOva, 3.701 },
{ EstimatorType.LightGbmMulti, 4.765 },
{ EstimatorType.SdcaMaximumEntropyMulti, 10.129 },
{ EstimatorType.SdcaLogisticRegressionOva, 13.16 },
{ EstimatorType.LbfgsMaximumEntropyMulti, 7.980 },
{ EstimatorType.LbfgsLogisticRegressionOva, 11.513 },
{ EstimatorType.LightGbmBinary, 4.765 },
{ EstimatorType.FastTreeBinary, 3.701 },
{ EstimatorType.FastForestBinary, 4.283 },
{ EstimatorType.SdcaLogisticRegressionBinary, 13.16 },
{ EstimatorType.LbfgsLogisticRegressionBinary, 11.513 },
{ EstimatorType.ForecastBySsa, 1 },
{ EstimatorType.ImageClassificationMulti, 1 },
{ EstimatorType.MatrixFactorization, 1 },
};
_rand = new Random(settings.Seed ?? 0);
_multiModelPipeline = null;
}
public static MultiModelPipeline Append(this SweepableEstimator estimator, params SweepableEstimator[] estimators)
{
var multiModelPipeline = new MultiModelPipeline();
multiModelPipeline = multiModelPipeline.Append(estimator);
return(multiModelPipeline.Append(estimators));
}
public override ExperimentResult <BinaryClassificationMetrics> Execute(IDataView trainData, ColumnInformation columnInformation, IEstimator <ITransformer> preFeaturizer = null, IProgress <RunDetail <BinaryClassificationMetrics> > progressHandler = null)
{
var label = columnInformation.LabelColumnName;
_experiment.SetEvaluateMetric(Settings.OptimizingMetric, label);
_experiment.SetTrainingTimeInSeconds(Settings.MaxExperimentTimeInSeconds);
// Cross val threshold for # of dataset rows --
// If dataset has < threshold # of rows, use cross val.
// Else, run experiment using train-validate split.
const int crossValRowCountThreshold = 15000;
var rowCount = DatasetDimensionsUtil.CountRows(trainData, crossValRowCountThreshold);
// TODO
// split cross validation result according to sample key as well.
if (rowCount < crossValRowCountThreshold)
{
const int numCrossValFolds = 10;
_experiment.SetDataset(trainData, numCrossValFolds);
}
else
{
var splitData = Context.Data.TrainTestSplit(trainData);
_experiment.SetDataset(splitData.TrainSet, splitData.TestSet);
}
MultiModelPipeline pipeline = new MultiModelPipeline();
if (preFeaturizer != null)
{
pipeline = pipeline.Append(preFeaturizer);
}
pipeline = pipeline.Append(Context.Auto().Featurizer(trainData, columnInformation, Features))
.Append(Context.Auto().BinaryClassification(label, Features));
_experiment.SetPipeline(pipeline);
var monitor = new BinaryClassificationTrialResultMonitor();
monitor.OnTrialCompleted += (o, e) =>
{
var detail = ToRunDetail(e);
progressHandler?.Report(detail);
};
_experiment.SetMonitor(monitor);
_experiment.Run();
var runDetails = monitor.RunDetails.Select(e => ToRunDetail(e));
var bestRun = ToRunDetail(monitor.BestRun);
var result = new ExperimentResult <BinaryClassificationMetrics>(runDetails, bestRun);
return(result);
}
public static MultiModelPipeline Append(this SweepableEstimatorPipeline pipeline, params SweepableEstimator[] estimators)
{
var multiModelPipeline = new MultiModelPipeline();
foreach (var estimator in pipeline.Estimators)
{
multiModelPipeline = multiModelPipeline.Append(estimator);
}
return(multiModelPipeline.Append(estimators));
}
public AutoMLExperiment SetPipeline(SweepableEstimatorPipeline pipeline)
{
var res = new MultiModelPipeline();
foreach (var e in pipeline.Estimators)
{
res = res.Append(e);
}
SetPipeline(res);
return(this);
}
/// <summary>
/// Create a single featurize pipeline according to <paramref name="columnInformation"/>. This function will collect all columns in <paramref name="columnInformation"/>,
/// featurizing them using <see cref="CatalogFeaturizer(string[], string[])"/>, <see cref="NumericFeaturizer(string[], string[])"/> or <see cref="TextFeaturizer(string, string)"/>. And combine
/// them into a single feature column as output.
/// </summary>
/// <param name="data">input data.</param>
/// <param name="columnInformation">column information.</param>
/// <param name="outputColumnName">output feature column.</param>
/// <returns>A <see cref="MultiModelPipeline"/> for featurization.</returns>
public MultiModelPipeline Featurizer(IDataView data, ColumnInformation columnInformation, string outputColumnName = "Features")
{
Contracts.CheckValue(data, nameof(data));
Contracts.CheckValue(columnInformation, nameof(columnInformation));
var columnPurposes = PurposeInference.InferPurposes(this._context, data, columnInformation);
var textFeatures = columnPurposes.Where(c => c.Purpose == ColumnPurpose.TextFeature);
var numericFeatures = columnPurposes.Where(c => c.Purpose == ColumnPurpose.NumericFeature);
var catalogFeatures = columnPurposes.Where(c => c.Purpose == ColumnPurpose.CategoricalFeature);
var textFeatureColumnNames = textFeatures.Select(c => data.Schema[c.ColumnIndex].Name).ToArray();
var numericFeatureColumnNames = numericFeatures.Select(c => data.Schema[c.ColumnIndex].Name).ToArray();
var catalogFeatureColumnNames = catalogFeatures.Select(c => data.Schema[c.ColumnIndex].Name).ToArray();
var pipeline = new MultiModelPipeline();
if (numericFeatureColumnNames.Length > 0)
{
pipeline = pipeline.Append(this.NumericFeaturizer(numericFeatureColumnNames, numericFeatureColumnNames));
}
if (catalogFeatureColumnNames.Length > 0)
{
pipeline = pipeline.Append(this.CatalogFeaturizer(catalogFeatureColumnNames, catalogFeatureColumnNames));
}
foreach (var textColumn in textFeatureColumnNames)
{
pipeline = pipeline.Append(this.TextFeaturizer(textColumn, textColumn));
}
var option = new ConcatOption
{
InputColumnNames = textFeatureColumnNames.Concat(numericFeatureColumnNames).Concat(catalogFeatureColumnNames).ToArray(),
OutputColumnName = outputColumnName,
};
if (option.InputColumnNames.Length > 0)
{
pipeline = pipeline.Append(SweepableEstimatorFactory.CreateConcatenate(option));
}
return(pipeline);
}
public override CrossValidationExperimentResult <BinaryClassificationMetrics> Execute(IDataView trainData, uint numberOfCVFolds, ColumnInformation columnInformation = null, IEstimator <ITransformer> preFeaturizer = null, IProgress <CrossValidationRunDetail <BinaryClassificationMetrics> > progressHandler = null)
{
var label = columnInformation.LabelColumnName;
_experiment.SetEvaluateMetric(Settings.OptimizingMetric, label);
_experiment.SetTrainingTimeInSeconds(Settings.MaxExperimentTimeInSeconds);
_experiment.SetDataset(trainData, (int)numberOfCVFolds);
MultiModelPipeline pipeline = new MultiModelPipeline();
if (preFeaturizer != null)
{
pipeline = pipeline.Append(preFeaturizer);
}
pipeline = pipeline.Append(Context.Auto().Featurizer(trainData, columnInformation, "__Features__"))
.Append(Context.Auto().BinaryClassification(label, featureColumnName: Features));
_experiment.SetPipeline(pipeline);
var monitor = new BinaryClassificationTrialResultMonitor();
monitor.OnTrialCompleted += (o, e) =>
{
var runDetails = ToCrossValidationRunDetail(e);
progressHandler?.Report(runDetails);
};
_experiment.SetMonitor(monitor);
_experiment.Run();
var runDetails = monitor.RunDetails.Select(e => ToCrossValidationRunDetail(e));
var bestResult = ToCrossValidationRunDetail(monitor.BestRun);
var result = new CrossValidationExperimentResult <BinaryClassificationMetrics>(runDetails, bestResult);
return(result);
}
public TrialSettings Propose(TrialSettings settings)
{
_multiModelPipeline = settings.ExperimentSettings.Pipeline;
_learnerInitialCost = _multiModelPipeline.PipelineIds.ToDictionary(kv => kv, kv => GetEstimatedCostForPipeline(kv, _multiModelPipeline));
var pipelineIds = _multiModelPipeline.PipelineIds;
if (_eci == null)
{
// initialize eci with the estimated cost and always start from pipeline which has lowest cost.
_eci = pipelineIds.ToDictionary(kv => kv, kv => GetEstimatedCostForPipeline(kv, _multiModelPipeline));
settings.Schema = _eci.OrderBy(kv => kv.Value).First().Key;
}
else
{
var probabilities = pipelineIds.Select(id => _eci[id]).ToArray();
probabilities = ArrayMath.Inverse(probabilities);
probabilities = ArrayMath.Normalize(probabilities);
// sample
var randdouble = _rand.NextDouble();
var sum = 0.0;
// selected pipeline id index
int i;
for (i = 0; i != pipelineIds.Length; ++i)
{
sum += ((double[])probabilities)[i];
if (sum > randdouble)
{
break;
}
}
settings.Schema = pipelineIds[i];
}
settings.Pipeline = _multiModelPipeline.BuildSweepableEstimatorPipeline(settings.Schema);
return(settings);
}
private double GetEstimatedCostForPipeline(string kv, MultiModelPipeline multiModelPipeline)
{
var entity = Entity.FromExpression(kv);
var estimatorTypes = entity.ValueEntities().Where(v => v is StringEntity s && s.Value != "Nil")
.Select(v =>
{
var s = v as StringEntity;
var estimator = multiModelPipeline.Estimators[s.Value];
return(estimator.EstimatorType);
});
var res = 1;
foreach (var estimatorType in estimatorTypes)
{
if (_estimatorCost.ContainsKey(estimatorType))
{
return(_estimatorCost[estimatorType]);
}
}
return(res);
}
public AutoMLExperiment SetPipeline(MultiModelPipeline pipeline)
{
_settings.Pipeline = pipeline;
return(this);
}
