public static (IDataView[] trainDatasets, IDataView[] validationDatasets) CrossValSplit(MLContext context,
IDataView trainData, uint numFolds, string samplingKeyColumn)
{
var originalColumnNames = trainData.Schema.Select(c => c.Name);
var splits = context.Data.CrossValidationSplit(trainData, (int)numFolds, samplingKeyColumnName: samplingKeyColumn);
var trainDatasets = new List <IDataView>();
var validationDatasets = new List <IDataView>();
foreach (var split in splits)
{
if (DatasetDimensionsUtil.IsDataViewEmpty(split.TrainSet) ||
DatasetDimensionsUtil.IsDataViewEmpty(split.TestSet))
{
continue;
}
var trainDataset = DropAllColumnsExcept(context, split.TrainSet, originalColumnNames);
var validationDataset = DropAllColumnsExcept(context, split.TestSet, originalColumnNames);
trainDatasets.Add(trainDataset);
validationDatasets.Add(validationDataset);
}
if (!trainDatasets.Any())
{
throw new InvalidOperationException("All cross validation folds have empty train or test data. " +
"Try increasing the number of rows provided in training data, or lowering specified number of " +
"cross validation folds.");
}
return(trainDatasets.ToArray(), validationDatasets.ToArray());
}
private static void ValidateTrainData(IDataView trainData, ColumnInformation columnInformation)
{
if (trainData == null)
{
throw new ArgumentNullException(nameof(trainData), "Training data cannot be null");
}
if (DatasetDimensionsUtil.IsDataViewEmpty(trainData))
{
throw new ArgumentException("Training data has 0 rows", nameof(trainData));
}
foreach (var column in trainData.Schema)
{
if (column.Name == DefaultColumnNames.Features && column.Type.GetItemType() != NumberDataViewType.Single)
{
throw new ArgumentException($"{DefaultColumnNames.Features} column must be of data type {NumberDataViewType.Single}", nameof(trainData));
}
if ((column.Name != columnInformation.LabelColumnName &&
column.Name != columnInformation.UserIdColumnName &&
column.Name != columnInformation.ItemIdColumnName &&
column.Name != columnInformation.GroupIdColumnName)
&&
column.Type.GetItemType() != BooleanDataViewType.Instance &&
column.Type.GetItemType() != NumberDataViewType.Single &&
column.Type.GetItemType() != TextDataViewType.Instance)
{
throw new ArgumentException($"Only supported feature column types are " +
$"{BooleanDataViewType.Instance}, {NumberDataViewType.Single}, and {TextDataViewType.Instance}. " +
$"Please change the feature column {column.Name} of type {column.Type} to one of " +
$"the supported types.", nameof(trainData));
}
}
}
private static void ValidateValidationData(IDataView trainData, IDataView validationData)
{
if (validationData == null)
{
return;
}
if (DatasetDimensionsUtil.IsDataViewEmpty(validationData))
{
throw new ArgumentException("Validation data has 0 rows", nameof(validationData));
}
const string schemaMismatchError = "Training data and validation data schemas do not match.";
if (trainData.Schema.Count != validationData.Schema.Count)
{
throw new ArgumentException($"{schemaMismatchError} Train data has '{trainData.Schema.Count}' columns," +
$"and validation data has '{validationData.Schema.Count}' columns.", nameof(validationData));
}
foreach (var trainCol in trainData.Schema)
{
var validCol = validationData.Schema.GetColumnOrNull(trainCol.Name);
if (validCol == null)
{
throw new ArgumentException($"{schemaMismatchError} Column '{trainCol.Name}' exsits in train data, but not in validation data.", nameof(validationData));
}
if (trainCol.Type != validCol.Value.Type)
{
throw new ArgumentException($"{schemaMismatchError} Column '{trainCol.Name}' is of type {trainCol.Type} in train data, and type " +
$"{validCol.Value.Type} in validation data.", nameof(validationData));
}
}
}
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);
}
private static void ValidateValidationData(IDataView trainData, IDataView validationData)
{
if (validationData == null)
{
return;
}
if (DatasetDimensionsUtil.IsDataViewEmpty(validationData))
{
throw new ArgumentException("Validation data has 0 rows", nameof(validationData));
}
const string schemaMismatchError = "Training data and validation data schemas do not match.";
if (trainData.Schema.Count(c => !c.IsHidden) != validationData.Schema.Count(c => !c.IsHidden))
{
throw new ArgumentException($"{schemaMismatchError} Train data has '{trainData.Schema.Count}' columns," +
$"and validation data has '{validationData.Schema.Count}' columns.", nameof(validationData));
}
// Validate that every active column in the train data corresponds to an active column in the validation data.
// (Indirectly, since we asserted above that the train and validation data have the same number of active columns, this also
// ensures the reverse -- that every active column in the validation data corresponds to an active column in the train data.)
foreach (var trainCol in trainData.Schema)
{
if (trainCol.IsHidden)
{
continue;
}
var validCol = validationData.Schema.GetColumnOrNull(trainCol.Name);
if (validCol == null)
{
throw new ArgumentException($"{schemaMismatchError} Column '{trainCol.Name}' exists in train data, but not in validation data.", nameof(validationData));
}
if (trainCol.Type != validCol.Value.Type)
{
throw new ArgumentException($"{schemaMismatchError} Column '{trainCol.Name}' is of type {trainCol.Type} in train data, and type " +
$"{validCol.Value.Type} in validation data.", nameof(validationData));
}
}
}
/// <summary>
/// Executes an AutoML experiment.
/// </summary>
/// <param name="trainData">The training data to be used by the AutoML experiment.</param>
/// <param name="columnInformation">Column information for the dataset.</param>
/// <param name="preFeaturizer">Pre-featurizer that AutoML will apply to the data during an
/// experiment. (The pre-featurizer will be fit only on the training data split to produce a
/// trained transform. Then, the trained transform will be applied to both the training
/// data split and corresponding validation data split.)</param>
/// <param name="progressHandler">A user-defined object that implements
/// the <see cref="IProgress{T}"/> interface. AutoML will invoke the method
/// <see cref="IProgress{T}.Report(T)"/> after each model it produces during the
/// course of the experiment.
/// </param>
/// <returns>The experiment result.</returns>
/// <remarks>
/// Depending on the size of your data, the AutoML experiment could take a long time to execute.
/// </remarks>
public ExperimentResult <TMetrics> Execute(IDataView trainData, ColumnInformation columnInformation,
IEstimator <ITransformer> preFeaturizer = null, IProgress <RunDetail <TMetrics> > progressHandler = null)
{
// 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);
if (rowCount < crossValRowCountThreshold)
{
const int numCrossValFolds = 10;
var splitResult = SplitUtil.CrossValSplit(Context, trainData, numCrossValFolds, columnInformation?.SamplingKeyColumnName);
return(ExecuteCrossValSummary(splitResult.trainDatasets, columnInformation, splitResult.validationDatasets, preFeaturizer, progressHandler));
}
else
{
var splitResult = SplitUtil.TrainValidateSplit(Context, trainData, columnInformation?.SamplingKeyColumnName);
return(ExecuteTrainValidate(splitResult.trainData, columnInformation, splitResult.validationData, preFeaturizer, progressHandler));
}
}
public static ColumnDimensions[] CalcColumnDimensions(MLContext context, IDataView data, PurposeInference.Column[] purposes)
{
data = context.Data.TakeRows(data, MaxRowsToRead);
var colDimensions = new ColumnDimensions[data.Schema.Count];
for (var i = 0; i < data.Schema.Count; i++)
{
var column = data.Schema[i];
var purpose = purposes[i];
// default column dimensions
int? cardinality = null;
bool?hasMissing = null;
var itemType = column.Type.GetItemType();
// If categorical text feature, calculate cardinality
if (itemType.IsText() && purpose.Purpose == ColumnPurpose.CategoricalFeature)
{
cardinality = DatasetDimensionsUtil.GetTextColumnCardinality(data, column);
}
// If numeric feature, discover missing values
if (itemType == NumberDataViewType.Single)
{
hasMissing = column.Type.IsVector() ?
DatasetDimensionsUtil.HasMissingNumericVector(data, column) :
DatasetDimensionsUtil.HasMissingNumericSingleValue(data, column);
}
colDimensions[i] = new ColumnDimensions(cardinality, hasMissing);
}
return(colDimensions);
}